merge: bring upstream v0.3.0 (72 commits) into Ai4Sci fork

Merged upstream/main (418abca, release v0.3.0) into our fork on a
dedicated branch. 21 conflicting files resolved; main worktree untouched.

Resolution policy and key decisions:
- Keep Ai4Sci runtime endpoints, durable dispatch, workspace scopes and
  the HITL/DynamicReview approval chain (approval path is product-critical).
- Adopt upstream model registry (llm/registry.py): our 136 model entries
  are a strict subset of upstream's 180, so dropping our inline table
  loses nothing and gains 44 new models.
- Adopt upstream native EvoChatDeepSeek; drop our obsolete
  _patch_deepseek_reasoning_passback monkey patch.
- Keep our six patches.py additions, ported onto upstream's new
  _OpenAICompatContent class: stable tool-call ids, tool-history
  sanitization, drop_reasoning_metadata, empty-SSE keepalive,
  extracted-document-text patch, _has_assistant_tool_protocol.
- Keep our skill-budget middleware path (skills=None) instead of passing
  skills through, to avoid double loading.
- Keep sanitized error labels (_safe_error_label) while adopting
  upstream's injected MiddlewareEventSink for fallback narration.
- Keep port 3076 and the LANGGRAPH_SERVER_URL override; adopt upstream's
  host/probe-host handling and CONFIG_DRIFT_SINCE_LAUNCH.
- Adopt upstream dependency stack: deepagents 0.7.6, langchain-quickjs
  0.3.7, langgraph-api 0.14; keep our extra deps (rfc8785, pillow,
  firecrawl-anydoc, nest-asyncio).
- Align call sites with upstream APIs: create_tool_selector_middleware
  now takes events= instead of track_stream_selection=.

.env.example

@@ -18,6 +18,9 @@ KIMI_API_KEY=                 # kimi.com/code (Kimi 代码计划)
 # Aggregator platforms (optional)
 SILICONFLOW_API_KEY=          # siliconflow.cn
 OPENROUTER_API_KEY=           # openrouter.ai
+REQUESTY_API_KEY=             # requesty.ai
+ATLASCLOUD_API_KEY=           # atlascloud.ai
+NOVITA_API_KEY=               # novita.ai
 
 # Custom endpoints (optional)
 CUSTOM_OPENAI_API_KEY=       # OpenAI-compatible endpoint

.github/assets/badge-pypi-dark.svg

@@ -5,5 +5,5 @@
   <rect x="54" y="5" width="62" height="24" rx="6" fill="#1565c0"/>
   <text x="85" y="22" text-anchor="middle"
         font-family="Inter, -apple-system, system-ui, sans-serif"
-        font-size="13" font-weight="700" fill="#ffffff">v0.2.2</text>
+        font-size="13" font-weight="700" fill="#ffffff">v0.3.0</text>
 </svg>

.github/assets/badge-pypi-light.svg

@@ -5,5 +5,5 @@
   <rect x="54" y="5" width="62" height="24" rx="6" fill="#2563eb"/>
   <text x="85" y="22" text-anchor="middle"
         font-family="Inter, -apple-system, system-ui, sans-serif"
-        font-size="13" font-weight="700" fill="#ffffff">v0.2.2</text>
+        font-size="13" font-weight="700" fill="#ffffff">v0.3.0</text>
 </svg>

.github/workflows/build.yml

@@ -11,7 +11,9 @@ jobs:
     timeout-minutes: 10
     steps:
       - uses: actions/checkout@v5
-      - uses: astral-sh/setup-uv@v6
+      # Full tag required — setup-uv dropped major/minor tags in v8.0.0, so
+      # `@v9` does not resolve. See the note in lint.yml.
+      - uses: astral-sh/setup-uv@v9.0.0
         with:
           python-version: "3.11"
           cache-dependency-glob: "**/pyproject.toml"

.github/workflows/docker.yml

@@ -3,7 +3,10 @@ name: Docker
 on:
   push:
     branches: ["main"]
-    tags: ["v*"]
+  # Version images build when a GitHub Release is published — the same event
+  # that triggers the PyPI upload (publish.yml), so the two channels stay in sync.
+  release:
+    types: [published]
   pull_request:
     paths:
       - "Dockerfile"
@@ -32,6 +35,19 @@ jobs:
     steps:
       - uses: actions/checkout@93cb6efe18208431cddfb8368fd83d5badbf9bfd  # v5.0.1
 
+      # Same guard as publish.yml — a release whose tag mismatches pyproject
+      # must not publish versioned images either.
+      - name: Guard — package version must match the release tag
+        if: github.event_name == 'release'
+        run: |
+          VERSION=$(grep -m1 '^version = ' pyproject.toml | sed -E 's/^version = "(.*)"/\1/')
+          TAG="${GITHUB_REF_NAME#v}"
+          echo "pyproject version: $VERSION | release tag: $TAG"
+          if [ "$VERSION" != "$TAG" ]; then
+            echo "::error::pyproject version ($VERSION) does not match release tag ($TAG); refusing to publish images."
+            exit 1
+          fi
+
       - uses: docker/setup-qemu-action@c7c53464625b32c7a7e944ae62b3e17d2b600130  # v3.7.0
       - uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f  # v3.12.0
 

.github/workflows/lint.yml

@@ -11,7 +11,11 @@ jobs:
     timeout-minutes: 5
     steps:
       - uses: actions/checkout@v5
-      - uses: astral-sh/setup-uv@v6
+      # Pinned to a full tag on purpose: setup-uv stopped publishing major /
+      # minor tags in v8.0.0 (supply-chain hardening), so `@v9` does not exist
+      # and would fail the job. Releases are immutable, so the tag is as safe
+      # as a SHA. v7 is where the action moved off the deprecated node20.
+      - uses: astral-sh/setup-uv@v9.0.0
         with:
           python-version: "3.11"
           cache-dependency-glob: "**/pyproject.toml"

.github/workflows/publish.yml

@@ -0,0 +1,67 @@
+name: Publish to PyPI
+
+# Publishes EvoScientist to PyPI via OpenID Connect (trusted publishing) — no
+# API token or password involved. Fires when a GitHub Release is published
+# (i.e. after `gh release create vX.Y.Z`), builds the sdist + wheel, guards
+# that the package version matches the release tag, then uploads with a
+# short-lived OIDC token.
+#
+# One-time PyPI setup (Manage project -> Publishing -> Add a new publisher):
+#   Owner:            EvoScientist
+#   Repository:       EvoScientist
+#   Workflow name:    publish.yml
+#   Environment name: pypi
+on:
+  release:
+    types: [published]
+  # Manual re-run escape hatch — dispatch it from the release tag; the version
+  # guard rejects any non-tag ref.
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  publish:
+    name: Build and publish to PyPI
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    environment:
+      name: pypi
+      url: https://pypi.org/project/EvoScientist/
+    permissions:
+      id-token: write  # required to mint the OIDC token PyPI verifies
+    steps:
+      # Actions in this job are pinned to full commit SHAs (like docker.yml):
+      # it holds id-token: write and PyPI publishing authority.
+      - uses: actions/checkout@93cb6efe18208431cddfb8368fd83d5badbf9bfd  # v5.0.1
+        with:
+          persist-credentials: false
+      - uses: astral-sh/setup-uv@c771a70e6277c0a99b617c7a806ffedaca235ff9  # v9.0.0
+        with:
+          python-version: "3.11"
+          # No shared cache in the publishing job — build from clean sources only.
+          enable-cache: false
+
+      - name: Build sdist + wheel
+        run: uv build
+
+      - name: Guard — package version must match the release tag
+        run: |
+          if [ "${GITHUB_REF_TYPE}" != "tag" ]; then
+            echo "::error::This workflow must run from a version tag (got ${GITHUB_REF_TYPE} '${GITHUB_REF_NAME}'); dispatch it from the release tag."
+            exit 1
+          fi
+          VERSION=$(grep -m1 '^version = ' pyproject.toml | sed -E 's/^version = "(.*)"/\1/')
+          TAG="${GITHUB_REF_NAME#v}"
+          echo "pyproject version: $VERSION | release tag: $TAG"
+          if [ "$VERSION" != "$TAG" ]; then
+            echo "::error::pyproject version ($VERSION) does not match release tag ($TAG); refusing to publish."
+            exit 1
+          fi
+
+      - name: Twine metadata check
+        run: uvx twine check dist/*
+
+      - name: Publish to PyPI (trusted publishing)
+        uses: pypa/gh-action-pypi-publish@dc37677b2e1c63e2034f94d8a5b11f265b73ba33  # v1.14.2

.github/workflows/test.yml

@@ -21,11 +21,13 @@ jobs:
     runs-on: ${{ matrix.os }}
     steps:
       - uses: actions/checkout@v5
-      - uses: astral-sh/setup-uv@v6
+      # Full tag required — setup-uv dropped major/minor tags in v8.0.0, so
+      # `@v9` does not resolve. See the note in lint.yml.
+      - uses: astral-sh/setup-uv@v9.0.0
         with:
           python-version: ${{ matrix.python-version }}
           cache-dependency-glob: "**/pyproject.toml"
       - name: Install dependencies
-        run: uv sync --dev
+        run: uv sync --dev --extra all-channels
       - name: Run pytest
         run: uv run pytest -v --timeout=30

EvoScientist/EvoScientist.py

@@ -23,7 +23,11 @@ from collections.abc import Sequence
 from pathlib import Path
 from typing import TYPE_CHECKING
 
-from langchain.agents.middleware import AgentMiddleware, HumanInTheLoopMiddleware
+from langchain.agents.middleware import (
+    AgentMiddleware,
+    HumanInTheLoopMiddleware,
+    TodoListMiddleware,
+)
 
 from . import paths as _paths_mod
 from .config import (
@@ -42,6 +46,9 @@ logging.getLogger("deepagents.middleware.skills").setLevel(logging.ERROR)
 if TYPE_CHECKING:
     from langgraph.graph.state import CompiledStateGraph
 
+    from .middleware.events import MiddlewareEventSink
+    from .runtime import AsyncRuntime
+
 # =============================================================================
 # Constants
 # =============================================================================
@@ -50,6 +57,15 @@ SUBAGENTS_CONFIG = Path(__file__).parent / "subagents"
 SKILLS_DIR = str(Path(__file__).parent / "skills")
 DEFAULT_SKILL_SOURCES = ("/skills/",)
 
+# Tools requiring human approval on attended agents (deepagents 0.7.0 ships a
+# recursive `delete` FS tool that would otherwise bypass the execute blocklist).
+HITL_INTERRUPT_ON: dict[str, bool] = {
+    "execute": True,
+    "run_in_background": True,
+    "schedule_task": True,
+    "delete": True,
+}
+
 # =============================================================================
 # Lazy state — initialized on first use, not at import time
 # =============================================================================
@@ -245,7 +261,11 @@ def _load_mcp_config_once() -> tuple[str, dict]:
     return sig, cfg
 
 
-def _load_mcp_tools_cached(on_progress=None) -> dict[str, list]:
+def _load_mcp_tools_cached(
+    on_progress=None,
+    *,
+    runtime: "AsyncRuntime | None" = None,
+) -> dict[str, list]:
     """Load MCP tools with config-aware caching.
 
     Args:
@@ -266,7 +286,11 @@ def _load_mcp_tools_cached(on_progress=None) -> dict[str, list]:
     if _MCP_TOOLS_CACHE_KEY == cfg_key and _MCP_TOOLS_CACHE_VALUE is not None:
         return {k: list(v) for k, v in _MCP_TOOLS_CACHE_VALUE.items()}
 
-    loaded = load_mcp_tools(config=cfg, on_progress=on_progress)
+    loaded = load_mcp_tools(
+        config=cfg,
+        on_progress=on_progress,
+        runtime=runtime,
+    )
     _MCP_TOOLS_CACHE_KEY = cfg_key
     _MCP_TOOLS_CACHE_VALUE = {k: list(v) for k, v in loaded.items()}
     return {k: list(v) for k, v in loaded.items()}
@@ -316,6 +340,7 @@ def _inject_subagent_middleware(
         RecoverableToolEffectMiddleware,
         RepetitiveToolCallGuardMiddleware,
         ToolErrorHandlerMiddleware,
+        ToolHistoryRepairMiddleware,
         ToolProtocolGuardMiddleware,
         create_context_editing_middleware,
         create_memory_lifecycle_middleware,
@@ -366,12 +391,15 @@ def _inject_subagent_middleware(
                 max_consecutive_errors=max_consecutive_tool_errors,
             ),
             ToolProtocolGuardMiddleware(),
+            # Sync subagents replay their own history to strict providers too.
+            ToolHistoryRepairMiddleware(),
             # Subagents share the main agent's model: use the threaded
             # ``chat_model`` on the pure path, else defer to the factory's
             # ``_ensure_chat_model()`` fallback (when ``chat_model=None``).
             create_context_editing_middleware(chat_model),
             create_runtime_context_middleware(),
             ToolErrorHandlerMiddleware(),
+            TodoListMiddleware(),
             ContextOverflowMapperMiddleware(),
         ]
         if memory_controls.memory_enabled:
@@ -443,8 +471,46 @@ def _apply_budgeted_skill_context(kwargs: dict, backend) -> dict:
     return updated
 
 
+def _fold_expert_subagents(subs: list[dict], tool_registry: dict) -> None:
+    """Append expert-skill sub-agent specs to ``subs``, guarding names.
+
+    Each installed expert skill becomes an in-process sub-agent entry so
+    the main agent's ``task`` tool (and the QuickJS ``task()`` global) can
+    dispatch to it in-turn by name. The same experts independently get a
+    background reach via ``build_expert_async_subagent_specs``; the two
+    reaches land on separate tool schemas, so sharing the name is safe.
+
+    Skips (with a warning) any expert whose ``name`` collides with a
+    subagent already in ``subs`` or with ``general-purpose``. The reserved
+    name matters because ``_ensure_general_purpose_subagent`` runs right
+    after this and early-returns when it sees the slot occupied — an expert
+    named ``general-purpose`` would silently take the slot and deepagents'
+    default subagent prompt would never reach the agent.
+    """
+    from deepagents.middleware.subagents import GENERAL_PURPOSE_SUBAGENT
+
+    from .subagents.expert_container import build_expert_subagent_specs
+
+    logger = logging.getLogger(__name__)
+    taken = {s.get("name") for s in subs} | {GENERAL_PURPOSE_SUBAGENT["name"]}
+    for spec in build_expert_subagent_specs(tool_registry=tool_registry):
+        name = spec["name"]
+        if name in taken:
+            logger.warning(
+                "Expert skill %r collides with an existing sub-agent name; skipping.",
+                name,
+            )
+            continue
+        taken.add(name)
+        subs.append(spec)
+
+
 def _maybe_swap_async_subagents(
-    subs: list, middleware: list | None = None, *, cfg=None
+    subs: list,
+    middleware: list | None = None,
+    *,
+    tool_registry: dict | None = None,
+    cfg=None,
 ) -> list:
     """Replace ``_async``-flagged sub-agents with ``AsyncSubAgent`` specs when enabled.
 
@@ -460,17 +526,24 @@ def _maybe_swap_async_subagents(
     Adding a new async sub-agent requires no change here — flip
     ``async: true`` in its yaml and create the matching deployment graph.
 
-    All return paths strip the internal ``_async`` field from sub-agent dicts
-    before handoff, since deepagents may schema-validate the kwarg.
+    YAML tool names stay in the internal ``_tool_names`` field until this
+    decision point. In-process specs resolve them against ``tool_registry``;
+    swapped remote specs discard them because their graph factory resolves
+    tools in its own process. All return paths strip internal fields before
+    handoff, since deepagents may schema-validate the kwargs.
 
     When async subagents are actually swapped in and ``middleware`` is provided,
     appends ``AsyncWatcherMiddleware`` so launches spawn an
     ``async_notifier`` watcher.
     """
+    from .utils import resolve_subagent_tools
+
     cfg = cfg if cfg is not None else _ensure_config()
+    tool_registry = tool_registry or {}
     if not getattr(cfg, "enable_async_subagents", False):
-        # Async fully disabled — strip the internal flag before handoff.
+        # Async fully disabled: every spec will run in-process.
         for s in subs:
+            resolve_subagent_tools(s, tool_registry)
             s.pop("_async", None)
         return subs
 
@@ -484,9 +557,9 @@ def _maybe_swap_async_subagents(
             "enable_async_subagents=true but langgraph dev is not reachable; "
             "falling back to in-process sync delegation for all sub-agents."
         )
-        # Strip the internal ``_async`` flag (carried from ``load_subagents``)
-        # before sub-agents reach deepagents — it's never a deepagents key.
+        # Every spec falls back to in-process execution.
         for s in subs:
+            resolve_subagent_tools(s, tool_registry)
             s.pop("_async", None)
         return subs
 
@@ -498,14 +571,18 @@ def _maybe_swap_async_subagents(
 
     if not async_specs:
         for s in subs:
+            resolve_subagent_tools(s, tool_registry)
             s.pop("_async", None)
         return subs
 
     from deepagents import AsyncSubAgent
 
-    from .langgraph_dev.sdk import configured_langgraph_dev_url
+    from .langgraph_dev.sdk import langgraph_dev_url
 
-    runtime_url = configured_langgraph_dev_url()
+    # Self-dispatch target. Resolved through ``langgraph_dev_url`` so it tracks
+    # both ``langgraph_dev_port`` and ``langgraph_dev_host`` — a wildcard bind
+    # maps back to loopback, a pinned interface is honored verbatim.
+    dev_url = langgraph_dev_url(cfg)
     out = []
     agent_specs: dict[str, AsyncSubAgent] = {}
     # MCP tools routed to async sub-agents (via ``expose_to: <name>`` in
@@ -520,19 +597,22 @@ def _maybe_swap_async_subagents(
                 name=name,
                 description=async_specs[name],
                 graph_id=name,
-                url=runtime_url,
+                url=dev_url,
             )
             agent_specs[name] = spec
             out.append(spec)
         else:
-            # Strip the internal flag before handoff to deepagents.
+            resolve_subagent_tools(s, tool_registry)
             s.pop("_async", None)
             out.append(s)
 
     if agent_specs and middleware is not None:
+        from .cli import async_notifier
         from .middleware.async_watcher import AsyncWatcherMiddleware
 
-        middleware.append(AsyncWatcherMiddleware(agent_specs))
+        # Composition root wires the concrete notifier port into the middleware;
+        # the middleware itself never imports the CLI layer.
+        middleware.append(AsyncWatcherMiddleware(agent_specs, notifier=async_notifier))
 
     # Forward the CLI's live (model, provider) into deepagents'
     # start/update_async_task tool calls so the deployed graph can
@@ -546,6 +626,138 @@ def _maybe_swap_async_subagents(
     return out
 
 
+def _route_async_specs_through_evo_middleware(
+    subs: list, base_middleware: list, *, cfg=None
+) -> list:
+    """Move ``AsyncSubAgent`` specs from ``subs`` into ``EvoAsyncSubAgentMiddleware``.
+
+    Deepagents' ``create_deep_agent`` auto-composes the vanilla
+    ``AsyncSubAgentMiddleware`` when it sees ``graph_id``-carrying entries
+    in ``subagents=``. We need our payload-aware subclass to handle those
+    (see ``EvoScientist/middleware/expert_async_subagent.py`` for the
+    upstream-workaround rationale). To prevent the auto-composition and
+    route all async dispatch through our subclass, we strip AsyncSubAgent
+    specs from ``subs`` here and hand them to our middleware.
+
+    Also folds in ``AsyncSubAgent`` specs for installed expert skills —
+    all pointing at the shared ``expert-container-async`` graph, marked
+    ``is_expert=True`` so the middleware requires a payload with
+    ``skill_name``.
+
+    The completion watcher (``AsyncWatcherMiddleware``) is found or created
+    before the middleware so the middleware's resolve-on-miss start tool can
+    hold the watcher's agent dict by reference — see the wiring block below.
+
+    Returns:
+        ``subs`` with ``graph_id``-carrying entries removed. Safe to pass
+        as ``create_deep_agent(subagents=...)`` — the async-auto-compose
+        branch is skipped for empty async lists.
+    """
+    from .middleware.async_watcher import AsyncWatcherMiddleware
+    from .middleware.expert_async_subagent import EvoAsyncSubAgentMiddleware
+    from .subagents.expert_container_async import build_expert_async_subagent_specs
+
+    cfg = cfg if cfg is not None else _ensure_config()
+
+    async_specs = [s for s in subs if "graph_id" in s]
+    sync_subs = [s for s in subs if "graph_id" not in s]
+    expert_specs = build_expert_async_subagent_specs(cfg=cfg)
+    async_specs.extend(expert_specs)
+
+    # Find or create the completion watcher BEFORE constructing
+    # ``EvoAsyncSubAgentMiddleware``: the middleware's resolve-on-miss start
+    # tool must hold the watcher's agent dict by reference, so an expert
+    # discovered mid-session lands in the dispatch table and the watcher in
+    # one step. Without the watcher update, dispatch succeeds but the
+    # watcher's ``get_async(agent_name)`` raises KeyError inside its
+    # ``try/except`` and the completion notification silently never fires.
+    watcher_agents: dict | None = None
+    watcher = next(
+        (m for m in base_middleware if isinstance(m, AsyncWatcherMiddleware)),
+        None,
+    )
+    if watcher is None:
+        # No YAML async subagents were registered, so ``_maybe_swap`` did
+        # not install the watcher. Install it now so experts still get
+        # completion notifications. Appended before the middleware's
+        # index-0 insert below, which yields the same final order as
+        # append-after-insert: ``[EvoAsync..., ..., watcher]``.
+        if expert_specs:
+            from .cli import async_notifier
+
+            watcher = AsyncWatcherMiddleware(
+                {s["name"]: s for s in expert_specs},
+                notifier=async_notifier,
+            )
+            base_middleware.append(watcher)
+    elif expert_specs:
+        # Extend AsyncWatcherMiddleware's client cache with expert specs so
+        # start_async_task launches for experts spawn a completion watcher —
+        # otherwise the watcher's ``get_async(agent_name)`` KeyErrors on the
+        # expert name, no notification is enqueued, and the main agent never
+        # learns the task finished. ``_maybe_swap_async_subagents`` above only
+        # populates the watcher with YAML-defined async subagents
+        # (writing-agent, data-analysis-agent, scheduler); this hook folds in
+        # the experts too.
+        #
+        # The mutation reaches through two layers of private state:
+        # ``AsyncWatcherMiddleware._clients`` (our own) and
+        # ``_ClientCache._agents`` (upstream deepagents). If upstream ever
+        # renames ``_agents`` or wraps it in an immutable snapshot, the
+        # ``.update(...)`` below silently lands on nothing — expert
+        # completion nudges then stop firing without a diagnostic surface.
+        # Convert that silent-drop into a grep-able error line and leave
+        # ``watcher_agents`` unset; expert dispatches still work (without
+        # completion notifications and without mid-session resolution into
+        # the watcher) until upstream drift is fixed.
+        if not hasattr(watcher._clients, "_agents"):
+            logging.getLogger(__name__).error(
+                "AsyncWatcherMiddleware._clients has no `_agents` slot — "
+                "deepagents internal renamed; expert completion "
+                "notifications will not fire until the extension hook is "
+                "updated to the new attribute name."
+            )
+            watcher = None
+        else:
+            watcher._clients._agents.update({s["name"]: s for s in expert_specs})
+    # The second ``hasattr`` is not redundant with the one in the ``elif``
+    # above: that check only runs when ``expert_specs`` is non-empty. When a
+    # pre-existing watcher has an empty expert set, this is the only guard
+    # standing between an upstream rename of ``_ClientCache._agents`` and an
+    # AttributeError that would kill agent construction — without it, the
+    # drift degrades to "no completion nudges" instead of crashing.
+    if watcher is not None and hasattr(watcher._clients, "_agents"):
+        watcher_agents = watcher._clients._agents
+
+    if async_specs:
+        # ``_maybe_swap_async_subagents`` installs the model-passthrough patch
+        # only when the yaml-async spec list is non-empty. An expert-only setup
+        # (no ``writing-agent`` / ``data-analysis-agent`` / ``scheduler`` in
+        # yaml) would otherwise miss the patch entirely, so we install it here
+        # too. Idempotent — the shared ``_model_passthrough_patched`` flag
+        # guards against double-patching.
+        from .llm.patches import _patch_deepagents_model_passthrough
+
+        _patch_deepagents_model_passthrough()
+
+        # Prepend rather than append so the ``## Async subagents`` prompt
+        # section stays in the stable prefix. Appending pushes it past the
+        # volatile memory tail, invalidating the cached prefix on every
+        # memory change.
+        base_middleware.insert(
+            0,
+            EvoAsyncSubAgentMiddleware(
+                async_subagents=async_specs,
+                watcher_agents=watcher_agents,
+                # The construction cfg, so resolve-on-miss specs the same
+                # langgraph_dev_port the construction-time specs used instead
+                # of re-reading config from disk at dispatch time.
+                cfg=cfg,
+            ),
+        )
+    return sync_subs
+
+
 def _build_base_kwargs(
     base_backend, base_middleware, *, cfg=None, chat_model=None, workspace_dir=None
 ):
@@ -562,13 +774,22 @@ def _build_base_kwargs(
 
     subs = load_subagents(
         SUBAGENTS_CONFIG,
-        tool_registry=tool_registry,
     )
+    _fold_expert_subagents(subs, tool_registry)
     _ensure_general_purpose_subagent(subs)
     _inject_subagent_middleware(
         subs, workspace_dir=workspace_dir, cfg=cfg, chat_model=chat_model
     )
-    subs = _maybe_swap_async_subagents(subs, base_middleware, cfg=cfg)
+    subs = _maybe_swap_async_subagents(
+        subs,
+        base_middleware,
+        tool_registry=tool_registry,
+        cfg=cfg,
+    )
+    # Route AsyncSubAgent specs (both standard and expert) through
+    # EvoAsyncSubAgentMiddleware so the payload-aware start_async_task tool
+    # replaces upstream's non-parameterisable one.
+    subs = _route_async_specs_through_evo_middleware(subs, base_middleware, cfg=cfg)
     return {
         "name": "EvoScientist",
         "model": chat_model if chat_model is not None else _ensure_chat_model(),
@@ -589,6 +810,7 @@ def load_mcp_and_build_kwargs(
     cfg=None,
     chat_model=None,
     workspace_dir=None,
+    runtime: "AsyncRuntime | None" = None,
 ):
     """Load MCP tools (cached by config) and build agent kwargs.
 
@@ -607,7 +829,10 @@ def load_mcp_and_build_kwargs(
     from .utils import load_subagents
 
     cfg = cfg if cfg is not None else _ensure_config()
-    mcp_by_agent = _load_mcp_tools_cached(on_progress=on_mcp_progress)
+    mcp_by_agent = _load_mcp_tools_cached(
+        on_progress=on_mcp_progress,
+        runtime=runtime,
+    )
     if not mcp_by_agent:
         return _build_base_kwargs(
             base_backend,
@@ -671,8 +896,8 @@ def load_mcp_and_build_kwargs(
 
     subs = load_subagents(
         SUBAGENTS_CONFIG,
-        tool_registry=registry,
     )
+    _fold_expert_subagents(subs, registry)
 
     _ensure_general_purpose_subagent(subs)
     _inject_subagent_middleware(
@@ -707,7 +932,16 @@ def load_mcp_and_build_kwargs(
 
     # Swap selected sub-agents to AsyncSubAgent (must happen AFTER MCP injection
     # since async sub-agents are remote graphs that load their own tools).
-    subs = _maybe_swap_async_subagents(subs, base_middleware, cfg=cfg)
+    subs = _maybe_swap_async_subagents(
+        subs,
+        base_middleware,
+        tool_registry=registry,
+        cfg=cfg,
+    )
+    # Mirror the base path: route AsyncSubAgent specs through
+    # EvoAsyncSubAgentMiddleware so the payload-aware start_async_task tool
+    # is the one composed into the main agent.
+    subs = _route_async_specs_through_evo_middleware(subs, base_middleware, cfg=cfg)
 
     return {
         "name": "EvoScientist",
@@ -726,8 +960,19 @@ def load_mcp_and_build_kwargs(
 # =============================================================================
 
 
-def _get_legacy_backend():
-    """Build the deployment-root backend used outside Web full deploy."""
+def _get_legacy_backend(
+    *, guard_dangerous: bool | None = None, refuse_delete: bool = False
+):
+    """Build the deployment-root backend used outside Web full deploy.
+
+    ``guard_dangerous`` — when ``None`` (default) follows ``cfg.auto_approve``;
+    the two research async sub-agent graphs (``writing-agent`` /
+    ``data-analysis-agent``) pass ``True`` because their remote thread has no
+    approval path at all (see ``subagents/_factory._GUARDED_ASYNC_SUBAGENTS``).
+    ``refuse_delete`` — the same two async graphs pass ``True`` so the recursive
+    ``delete`` FS tool is refused and relayed to the orchestrator for approval,
+    rather than deleting unattended.
+    """
     from deepagents.backends import CompositeBackend
 
     from .backends import (
@@ -737,6 +982,8 @@ def _get_legacy_backend():
     )
 
     cfg = _ensure_config()
+    if guard_dangerous is None:
+        guard_dangerous = cfg.auto_approve
     workspace_dir = str(_paths_mod.WORKSPACE_ROOT)
     set_active_workspace(workspace_dir)
     memory_dir = str(_paths_mod.MEMORIES_DIR)
@@ -750,6 +997,8 @@ def _get_legacy_backend():
         virtual_mode=True,
         timeout=cfg.sandbox_execute_timeout,
         dangerous=cfg.dangerous_mode,
+        guard_dangerous=guard_dangerous,
+        refuse_delete=refuse_delete,
     )
     sk_backend = MergedSkillsBackend(
         primary_dir=user_skills_dir,
@@ -769,10 +1018,14 @@ def _get_legacy_backend():
     )
 
 
-def _get_default_backend():
+def _get_default_backend(
+    *, guard_dangerous: bool | None = None, refuse_delete: bool = False
+):
     """Use Origin's conversation-scoped backend for Web full deploy."""
     if os.environ.get("EVOSCIENTIST_DEPLOY_MODE", "").lower() != "full":
-        return _get_legacy_backend()
+        return _get_legacy_backend(
+            guard_dangerous=guard_dangerous, refuse_delete=refuse_delete
+        )
     from .workspace_scope import create_workspace_backend_factory
 
     cfg = _ensure_config()
@@ -800,6 +1053,7 @@ def _get_default_middleware(
     enable_scheduler: bool | None = None,
     enable_memory_workers: bool | None = None,
     install_subagent_guard: bool = False,
+    events: "MiddlewareEventSink | None" = None,
 ):
     """Build the default middleware list.
 
@@ -819,6 +1073,11 @@ def _get_default_middleware(
             (avoids writing module globals on the pure path).
         memory_source_agent: Attribution name for profile/observation writes.
             Async sub-agent factories pass their deployed agent name here.
+        events: Frontend/session-supplied event sink. Middleware report
+            tool-selection events and model-fallback notices to it.
+            Defaults to the current stream run's sink for main agents; async
+            sub-agent stacks are always forced to ``NoOpSink`` (they must not
+            drive the main-agent widgets).
     """
     from .middleware import (
         DEFAULT_REPETITIVE_TOOL_CALL_THRESHOLD,
@@ -831,7 +1090,9 @@ def _get_default_middleware(
         RecoverableToolEffectMiddleware,
         RepetitiveToolCallGuardMiddleware,
         ToolErrorHandlerMiddleware,
+        ToolHistoryRepairMiddleware,
         ToolProtocolGuardMiddleware,
+        create_active_team_middleware,
         create_code_interpreter_middleware,
         create_context_editing_middleware,
         create_memory_lifecycle_middleware,
@@ -842,6 +1103,13 @@ def _get_default_middleware(
         default_memory_scheduler,
         load_fallback_chain,
     )
+    from .middleware.events import NO_OP_SINK, RunScopedEventSink
+
+    # Subagent stacks never drive the main-agent frontend widgets; force the
+    # no-op sink there regardless of what the caller passed. Main stacks built
+    # without an explicit frontend/session sink report into the active stream
+    # run's sink, preserving selector suppression for headless local runs.
+    events = NO_OP_SINK if for_async_subagent else (events or RunScopedEventSink())
 
     cfg = cfg if cfg is not None else _ensure_config()
     repetitive_tool_call_threshold = getattr(
@@ -882,6 +1150,8 @@ def _get_default_middleware(
             MemoryObservationTarget.AGENT
         ),
         "memory_scheduler": memory_scheduler,
+        # First-contact intro: main agent only, and never in unattended runs.
+        "enable_profile_bootstrap": not for_async_subagent and not bool(cfg.auto_mode),
     }
     if memory_max_inline_profile_chars is not None:
         memory_kwargs["max_inline_profile_chars"] = memory_max_inline_profile_chars
@@ -914,7 +1184,10 @@ def _get_default_middleware(
             else {}
         ),
         model=resolved_selector_model,
-        track_stream_selection=not for_async_subagent,
+        # A frontend sink enables the streaming selection lifecycle; subagent /
+        # headless stacks stay silent (upstream replaced track_stream_selection
+        # with the events sink).
+        events=events if not for_async_subagent else None,
     )
     mw = [
         # Outermost — catches provider-SDK exceptions from the model
@@ -924,14 +1197,22 @@ def _get_default_middleware(
         ErrorNormalizationMiddleware(),
         RecoverableMeteringMiddleware(),
         RecoverableToolEffectMiddleware(),
+        ToolHistoryRepairMiddleware(),
         create_context_editing_middleware(model),
-        *([ModelFallbackMiddleware()] if enable_legacy_model_fallback else []),
+        *(
+            [ModelFallbackMiddleware(events=events)]
+            if enable_legacy_model_fallback
+            else []
+        ),
         RepetitiveToolCallGuardMiddleware(
             threshold=repetitive_tool_call_threshold,
             max_consecutive_errors=max_consecutive_tool_errors,
         ),
         ContextOverflowMapperMiddleware(),
         ToolErrorHandlerMiddleware(),
+        # deepagents 0.7.0 dropped TodoListMiddleware from its defaults;
+        # EXPERIMENT_WORKFLOW planning and the todo UI pipeline require it.
+        TodoListMiddleware(),
         *selector_middlewares,
         ToolProtocolGuardMiddleware(),
         # Interpreter prompt must land before runtime/memory context, so this
@@ -969,13 +1250,32 @@ def _get_default_middleware(
 
         mw.insert(0, AskUserMiddleware())
 
+    # Expert prompt for the main agent — injects the ## Experts concept every
+    # turn (plus the invited-expert list when experts are invited). Inserted
+    # AFTER AskUser so it sits ahead of AskUser in the stack and runs first,
+    # landing its block right after ## Skills System (experts mirror skills).
+    # Main agent only: a running expert graph must not inject the expert prompt
+    # into its own baked-in persona.
+    if not for_async_subagent:
+        mw.insert(0, create_active_team_middleware())
+
     # Background-process tools (run_in_background / check_process / stop_process /
     # list_processes) — main agent only. Async sub-agents run on langgraph-dev and
     # must not spawn local OS processes.
     if not for_async_subagent and enable_background_execution:
+        from .cli import async_notifier
         from .middleware.background import BackgroundExecutionMiddleware
 
-        mw.append(BackgroundExecutionMiddleware())
+        # Inject the notifier port + the assembly-time dangerous-mode policy
+        # (agents rebuild on config change, so the captured flag never staler
+        # than the agent it lives on).
+        mw.append(
+            BackgroundExecutionMiddleware(
+                async_notifier,
+                dangerous=cfg.dangerous_mode,
+                guard_dangerous=cfg.auto_approve,
+            )
+        )
 
     if install_subagent_guard:
         mw.append(DisableSubagentToolMiddleware())
@@ -983,6 +1283,19 @@ def _get_default_middleware(
     return mw
 
 
+def _build_hitl_interrupt_on(*, auto_approve: bool) -> dict[str, bool] | None:
+    """Return :data:`HITL_INTERRUPT_ON` for ``create_deep_agent``, or ``None``
+    when the user opted out (``auto_approve`` / ``auto_mode`` /
+    ``dangerous_mode``) so nothing is armed and unattended runs never pause.
+    Passing it to ``create_deep_agent`` (not ``HumanInTheLoopMiddleware``) lets
+    declarative sub-agents inherit it while ``AsyncSubAgent`` specs do not — so
+    async agents can't hang on an approval nobody can deliver.
+    """
+    if auto_approve:
+        return None
+    return dict(HITL_INTERRUPT_ON)
+
+
 def _get_default_agent():
     """Build the default agent (no checkpointer) on first access.
 
@@ -1082,6 +1395,7 @@ def _get_default_agent():
 
         _EvoScientist_agent = create_deep_agent(
             **kwargs,
+            interrupt_on=_build_hitl_interrupt_on(auto_approve=cfg.auto_approve),
         ).with_config({"recursion_limit": cfg.recursion_limit})
     return _EvoScientist_agent
 
@@ -1145,6 +1459,8 @@ def create_cli_agent(
     main_agent_route_middleware: AgentMiddleware | None = None,
     execution_profile=None,
     agent_model_set=None,
+    events: "MiddlewareEventSink | None" = None,
+    runtime: "AsyncRuntime | None" = None,
 ) -> "CompiledStateGraph":
     """Create agent with checkpointer for CLI multi-turn support.
 
@@ -1187,6 +1503,8 @@ def create_cli_agent(
             after ConfigurableModelMiddleware and before tool selection. When
             provided, EvoScientist's legacy model fallback is disabled for the
             top-level agent so the host is the only fallback authority.
+        runtime: Optional application-scoped runtime for synchronous MCP tool
+            discovery. Direct callers get a scoped runtime when omitted.
     """
     import os as _os
 
@@ -1267,6 +1585,7 @@ def create_cli_agent(
             virtual_mode=True,
             timeout=cfg.sandbox_execute_timeout,
             dangerous=cfg.dangerous_mode,
+            guard_dangerous=cfg.auto_approve,
         )
     sk_backend = MergedSkillsBackend(
         primary_dir=_usr_skills_dir,
@@ -1318,6 +1637,7 @@ def create_cli_agent(
                 bool(profile.memory_workers) if profile is not None else None
             ),
             install_subagent_guard=(profile is not None and not profile.subagents),
+            events=events,
         )
     )
     from .middleware import ProviderContextMediaMiddleware
@@ -1383,6 +1703,7 @@ def create_cli_agent(
         cfg=cfg,
         chat_model=chat_model,
         workspace_dir=workspace_dir,
+        runtime=runtime,
     )
     if not enable_subagents:
         kwargs = {**kwargs, "subagents": []}
@@ -1420,4 +1741,5 @@ def create_cli_agent(
     return create_deep_agent(
         **kwargs,
         checkpointer=checkpointer,
+        interrupt_on=_build_hitl_interrupt_on(auto_approve=cfg.auto_approve),
     ).with_config({"recursion_limit": cfg.recursion_limit})

EvoScientist/backends.py

@@ -4,13 +4,21 @@ import os
 import posixpath
 import re
 import shlex
+import signal
+import subprocess
 import sys
+import threading
+import time
 import uuid
+from dataclasses import dataclass
+from enum import StrEnum
 from pathlib import Path
+from typing import TYPE_CHECKING
 
 from deepagents.backends import FilesystemBackend, LocalShellBackend
 from deepagents.backends.protocol import (
     BackendProtocol,
+    DeleteResult,
     EditResult,
     ExecuteResponse,
     FileDownloadResponse,
@@ -23,6 +31,10 @@ from deepagents.backends.protocol import (
 from filelock import FileLock
 
 from . import paths
+from .cancellation import current_cancel_event
+
+if TYPE_CHECKING:
+    from langgraph.types import Command
 
 # Reproduced here to dodge a circular import from .EvoScientist (the canonical
 # SKILLS_DIR constant).
@@ -69,6 +81,87 @@ BLOCKED_COMMANDS = [
 ]
 
 
+_active_shell_processes_lock = threading.RLock()
+_active_shell_processes: dict[threading.Event, set[subprocess.Popen[str]]] = {}
+_PROCESS_DRAIN_GRACE_SECONDS = 1.0
+
+
+def _terminate_process_tree(process: subprocess.Popen[str]) -> None:
+    """Force-stop a shell and its descendants without waiting for reaping."""
+    # A completed Popen has already reaped its PID, which the OS may reuse.
+    # Inspect the recorded state rather than calling poll(): an exited but
+    # unreaped shell can still have live descendants in its process group.
+    if process.returncode is not None:
+        return
+
+    try:
+        if os.name == "nt":
+            # CREATE_NEW_PROCESS_GROUP alone does not make terminate() recursive.
+            # taskkill is the native way to stop the complete descendant tree.
+            subprocess.run(
+                ["taskkill", "/PID", str(process.pid), "/T", "/F"],
+                check=False,
+                capture_output=True,
+                timeout=5,
+            )
+        else:
+            os.killpg(process.pid, signal.SIGKILL)
+    except (OSError, subprocess.SubprocessError):
+        try:
+            process.kill()
+        except OSError:
+            pass
+
+
+def _stop_collecting_process_output(process: subprocess.Popen[str]) -> None:
+    """Close inherited pipes and reap *process* without blocking the caller."""
+    for pipe in (process.stdout, process.stderr):
+        if pipe is not None:
+            try:
+                pipe.close()
+            except OSError:
+                pass
+
+    if process.poll() is None:
+        threading.Thread(target=process.wait, daemon=True).start()
+
+
+def cancel_active_shell_processes(event: threading.Event) -> None:
+    """Terminate every active shell command associated with *event*."""
+    with _active_shell_processes_lock:
+        processes = tuple(_active_shell_processes.get(event, ()))
+    for process in processes:
+        _terminate_process_tree(process)
+
+
+def _register_shell_process(
+    event: threading.Event | None,
+    process: subprocess.Popen[str],
+) -> None:
+    if event is None:
+        return
+    with _active_shell_processes_lock:
+        _active_shell_processes.setdefault(event, set()).add(process)
+        cancel_now = event.is_set()
+    if cancel_now:
+        _terminate_process_tree(process)
+
+
+def _unregister_shell_process(
+    event: threading.Event | None,
+    process: subprocess.Popen[str],
+) -> None:
+    if event is None:
+        return
+    with _active_shell_processes_lock:
+        processes = _active_shell_processes.get(event)
+        if processes is None:
+            return
+        processes.discard(process)
+        if not processes:
+            _active_shell_processes.pop(event, None)
+
+
 def _shell_token_spans(command: str) -> list[dict[str, object]]:
     """Tokenize enough shell syntax to find quoted SSH remote commands.
 
@@ -88,7 +181,11 @@ def _shell_token_spans(command: str) -> list[dict[str, object]]:
         if ch in "`();|&":
             return ch
         if ch in "<>":
-            if index + 1 < n and command[index + 1] == ch:
+            # `>>`/`<<` and `>|` (force-clobber redirect) are single redirection
+            # operators, NOT a pipe — the trailing `|` must not read as a boundary.
+            if index + 1 < n and (
+                command[index + 1] == ch or (ch == ">" and command[index + 1] == "|")
+            ):
                 return command[index : index + 2]
             return ch
         if ch.isdigit():
@@ -97,7 +194,9 @@ def _shell_token_spans(command: str) -> list[dict[str, object]]:
                 j += 1
             if j < n and command[j] in "<>":
                 end = j + 1
-                if end < n and command[end] in ("&", command[j]):
+                # `2>&1`, `2>>`, and `2>|` (fd force-clobber) are single
+                # redirection operators — the trailing `|` is not a pipe.
+                if end < n and command[end] in ("&", "|", command[j]):
                     end += 1
                 return command[index:end]
         return None
@@ -159,6 +258,175 @@ def _shell_token_spans(command: str) -> list[dict[str, object]]:
     return tokens
 
 
+# Commands that are dangerous as the RIGHT-HAND SIDE of a pipe (they consume
+# piped data as code or ship it off-box). Everything else piping is normal.
+_PIPE_NETWORKING_RHS = frozenset(
+    {
+        "nc",
+        "ncat",
+        "netcat",
+        "ssh",
+        "curl",
+        "wget",
+        "telnet",
+        "socat",
+        "scp",
+        "sftp",
+        "rsync",
+        "ftp",
+    }
+)
+_PIPE_INTERPRETER_RHS = frozenset(
+    {
+        "sh",
+        "bash",
+        "zsh",
+        "dash",
+        "ash",
+        "ksh",
+        "fish",
+        "python",
+        "python2",
+        "python3",
+        "node",
+        "bun",
+        "deno",
+        "ruby",
+        "perl",
+        "php",
+        "lua",
+        "iex",
+        "elixir",
+    }
+)
+_PIPE_DANGEROUS_RHS = _PIPE_INTERPRETER_RHS | _PIPE_NETWORKING_RHS
+
+
+def check_dangerous_command(command: str) -> str | None:
+    """Return a reason if *command* pipes output into an interpreter or a
+    network tool, else ``None``.
+
+    Deliberately narrow: this guards indirect prompt injection (the agent
+    ingests untrusted web content and could be induced to run
+    ``curl … | bash``). Everyday research shell — pipes into ``grep``/``head``,
+    redirects, ``python -c``, ``..``/``~`` paths — is NOT flagged here.
+    Workspace confinement stays in :func:`validate_command`.
+
+    Only the token immediately after the pipe is inspected, so wrapper
+    commands like ``env bash``, ``xargs bash``, or ``timeout 5 bash`` are
+    not detected — this is a known limitation, not a bug to fix here.
+    """
+    after_pipe = False
+    for token in _shell_token_spans(command):
+        if token.get("type") == "op":
+            value = token.get("value")
+            if value == "|":
+                after_pipe = True
+            elif value == "&" and after_pipe:
+                # `|&` (pipe stdout+stderr) tokenizes as `|` then `&`;
+                # keep the pipe context open across the `&`.
+                pass
+            else:
+                after_pipe = False
+            continue
+        if after_pipe:
+            base = str(token.get("value", "")).split("/")[-1]
+            # strip trailing version digits: python3.11 -> python, lua5.4 -> lua
+            normalized = re.sub(r"[0-9.]+$", "", base) or base
+            if base in _PIPE_DANGEROUS_RHS or normalized in _PIPE_DANGEROUS_RHS:
+                kind = (
+                    "networking tool"
+                    if base in _PIPE_NETWORKING_RHS
+                    or normalized in _PIPE_NETWORKING_RHS
+                    else "interpreter"
+                )
+                return f"pipes output into {kind} '{base}'"
+            after_pipe = False
+    return None
+
+
+class ActionDecision(StrEnum):
+    """Outcome of the shell-action policy."""
+
+    APPROVE = "approve"
+    REJECT = "reject"
+    PROMPT = "prompt"
+
+
+@dataclass(frozen=True)
+class ActionVerdict:
+    """A decision plus the reason to show the user or feed back to the agent."""
+
+    decision: ActionDecision
+    reason: str = ""
+
+
+def resolve_action_decision(
+    command: str,
+    *,
+    auto_approve: bool = False,
+    dangerous_mode: bool = False,
+    allow_list: list[str] | None = None,
+) -> ActionVerdict:
+    """Single source of truth for approve / reject / prompt.
+
+    Precedence:
+      1. ``dangerous_mode`` — the user asked for full power; run everything.
+      2. dangerous detection — pipe into interpreter/network.
+      3. ``auto_approve`` — opt-out means *never prompt*: approve, or reject
+         a dangerous command with a reason the agent can act on.
+      4. ``allow_list`` — case-sensitive match on a whole command or a
+         command-plus-space prefix; blank entries are ignored. Every segment of
+         a chain (``a; b``, ``a && b``, ``a | b``) must match, so an allow-listed
+         prefix cannot carry a non-listed command in behind it.
+    """
+    if dangerous_mode:
+        return ActionVerdict(ActionDecision.APPROVE)
+
+    reason = check_dangerous_command(command)
+
+    if auto_approve:
+        if reason:
+            return ActionVerdict(ActionDecision.REJECT, reason)
+        return ActionVerdict(ActionDecision.APPROVE)
+
+    if reason:
+        return ActionVerdict(ActionDecision.PROMPT, reason)
+
+    if allow_list:
+        prefixes = [p.strip() for p in allow_list if p.strip()]
+        # Match on a token boundary so allow-listing `ls` does not also approve
+        # `lsof` (case-sensitive, like the shell). Require EVERY segment of a
+        # chain to match, so `ls; rm -rf x` cannot ride in on an allow-listed
+        # `ls`. ``None`` means an unparseable construct (substitution/newline) —
+        # decline rather than risk approving a hidden command.
+        segments = _split_command_segments(command)
+        if segments is not None:
+            segments = segments or [command.strip()]
+            if prefixes and all(
+                any(seg == p or seg.startswith(p + " ") for p in prefixes)
+                for seg in segments
+            ):
+                return ActionVerdict(ActionDecision.APPROVE)
+
+    return ActionVerdict(ActionDecision.PROMPT)
+
+
+def build_hitl_resume(interrupt_id: str, decisions: list[dict]) -> "Command":
+    """Build a HITL resume Command keyed by interrupt_id.
+
+    Keying by id (not the flat ``{"decisions": …}``) is REQUIRED whenever the
+    graph has more than one pending interrupt — parallel sub-agents that each
+    call ``execute`` do exactly that, and a flat resume raises
+    ``RuntimeError: When there are multiple pending interrupts …``. Resuming a
+    single id resolves that interrupt and re-parks the rest (they re-emit on the
+    next stream), so callers drain them one at a time. Safe for N=1 too.
+    """
+    from langgraph.types import Command
+
+    return Command(resume={interrupt_id: {"decisions": decisions}})
+
+
 _SSH_OPTIONS_WITH_VALUE = {
     "-B",
     "-b",
@@ -401,6 +669,40 @@ def _split_shell_commands(command: str) -> list[str]:
     return base_commands
 
 
+def _split_command_segments(command: str) -> list[str] | None:
+    """Split a compound command into raw segment strings on command boundaries.
+
+    Quote-aware (via ``_shell_token_spans``). Boundaries are ``;`` ``&&`` ``||``
+    ``|`` ``&`` and grouping; redirections are not boundaries. Lets the allow-list
+    clear a chain only when *every* segment is allow-listed, not just the leading
+    one (``ls; rm -rf x`` must not ride in on an allow-listed ``ls``).
+
+    Returns ``None`` when the command contains a construct this small tokenizer
+    cannot safely reason about — command substitution (``$(...)`` or backticks,
+    which run a hidden command even inside double quotes) or a newline separator —
+    so the caller declines to allow-list it rather than approve a hidden command.
+    Deliberately a substring over-approximation: a literal/quoted ``$(``, backtick,
+    or newline also declines (a safe extra prompt, never a bypass). Quote/escape
+    awareness is intentionally not attempted — that fragility caused the original
+    chaining gap.
+    """
+    if "$(" in command or "`" in command or "\n" in command or "\r" in command:
+        return None
+    boundaries = {"&&", "||", ";", "|", "&", "(", ")"}
+    segments: list[str] = []
+    seg_start = 0
+    for token in _shell_token_spans(command):
+        if token.get("type") == "op" and token.get("value") in boundaries:
+            seg = command[seg_start : int(token["start"])].strip()
+            if seg:
+                segments.append(seg)
+            seg_start = int(token["end"])
+    tail = command[seg_start:].strip()
+    if tail:
+        segments.append(tail)
+    return segments
+
+
 def _has_traversal_component(command: str) -> bool:
     """Check if command contains '..' as a path component (not substring)."""
     from pathlib import PurePosixPath
@@ -819,6 +1121,11 @@ class ReadOnlyFilesystemBackend(FilesystemBackend):
             for file_path, _ in files
         ]
 
+    def delete(self, file_path: str) -> DeleteResult:
+        return DeleteResult(
+            error="This directory is read-only. Delete operations are not permitted here."
+        )
+
 
 class MemoryFilesystemBackend(FilesystemBackend):
     """Filesystem backend for memory files with structured-write enforcement.
@@ -835,6 +1142,10 @@ class MemoryFilesystemBackend(FilesystemBackend):
         "Raw edits under /memories are limited to existing "
         "/memories/profile/... files. Use memory tools for observations."
     )
+    _RAW_DELETE_ERROR = (
+        "Deletes under /memories are blocked. Manage memory files through "
+        "memory tools instead."
+    )
 
     def __init__(
         self,
@@ -871,6 +1182,9 @@ class MemoryFilesystemBackend(FilesystemBackend):
             for file_path, _ in files
         ]
 
+    def delete(self, file_path: str) -> DeleteResult:
+        return DeleteResult(error=self._RAW_DELETE_ERROR)
+
 
 def build_memory_agent_backend(
     *,
@@ -1082,7 +1396,12 @@ class MergedSkillsBackend(BackendProtocol):
 
 
 def prepare_sandbox_command(
-    command: str, cwd: str | Path, *, virtual_mode: bool = True, dangerous: bool = False
+    command: str,
+    cwd: str | Path,
+    *,
+    virtual_mode: bool = True,
+    dangerous: bool = False,
+    guard_dangerous: bool = False,
 ) -> tuple[str, str | None]:
     """Normalize workspace paths in ``command`` and validate it for the sandbox.
 
@@ -1092,7 +1411,16 @@ def prepare_sandbox_command(
 
     Returns ``(prepared_command, error)``: ``error`` is a message string when the command
     is rejected (the caller must NOT run it), otherwise ``None``.
+
+    ``guard_dangerous`` (see :func:`check_dangerous_command`) does not see inside an SSH
+    remote payload: a dangerous pipe *inside* a quoted ``ssh host '...'`` argument is not
+    detected, because the quoted payload is a single opaque token. Piping *into* ``ssh``
+    itself (e.g. ``cat secret | ssh host x``) is detected — the check runs on the original,
+    unmasked command so the SSH-masking done below (which also replaces the literal ``ssh``
+    token) does not blind it.
     """
+    original_command = command
+
     ssh_error = _validate_ssh_remote_command_format(command)
     if ssh_error:
         return command, ssh_error
@@ -1128,6 +1456,19 @@ def prepare_sandbox_command(
     )
     if error:
         return command, error
+
+    # No interactive approval is reachable here (unattended main agent, or an
+    # async sub-agent on a remote thread), so refuse the narrow dangerous set
+    # with a reason the agent can act on rather than running it blind.
+    if guard_dangerous and not dangerous:
+        dangerous_reason = check_dangerous_command(original_command)
+        if dangerous_reason:
+            return _restore_spans(command, ssh_replacements), (
+                f"Command blocked: {dangerous_reason}. "
+                f"Rewrite it to avoid that, or request approval from the user "
+                f"(the orchestrator can re-issue it after approval)."
+            )
+
     return _restore_spans(command, ssh_replacements), None
 
 
@@ -1153,6 +1494,8 @@ class CustomSandboxBackend(LocalShellBackend):
         env: dict[str, str] | None = None,
         inherit_env: bool = True,
         dangerous: bool = False,
+        guard_dangerous: bool = False,
+        refuse_delete: bool = False,
     ):
         """
         Initialize custom sandbox backend.
@@ -1168,8 +1511,20 @@ class CustomSandboxBackend(LocalShellBackend):
                 paths anywhere on disk (no workspace confinement). Forces
                 ``virtual_mode=False`` and relaxes path validation while keeping
                 the privileged-command blocklist. Defaults to False.
+            guard_dangerous: Refuse the narrow dangerous-command set (see
+                :func:`check_dangerous_command`) outright, for contexts where
+                no interactive approval is reachable (unattended auto-approve
+                runs, async sub-agents). Bypassed when ``dangerous=True``.
+                Defaults to False.
+            refuse_delete: Refuse the recursive ``delete`` FS tool outright,
+                relaying an approval request to the orchestrator. Used for async
+                research sub-agents (writing / data-analysis) that have no
+                interactive approval path. Bypassed when ``dangerous=True``.
+                Defaults to False.
         """
         self._dangerous = dangerous
+        self._guard_dangerous = guard_dangerous
+        self._refuse_delete = refuse_delete
         if dangerous:
             # Real paths require the legacy (non-virtual) resolution path so the
             # parent backend returns absolute paths as-is.
@@ -1239,6 +1594,22 @@ class CustomSandboxBackend(LocalShellBackend):
 
         return super()._resolve_path(key)
 
+    _DELETE_APPROVAL_ERROR = (
+        "Delete blocked: needs approval. Report it to the orchestrator, which "
+        "can re-issue it after approval."
+    )
+
+    def delete(self, file_path: str) -> DeleteResult:
+        """Refuse ``delete`` for guarded async sub-agents (no approval path).
+
+        No ``adelete`` override is needed: the inherited ``BackendProtocol.adelete``
+        runs ``asyncio.to_thread(self.delete, ...)``, so async sub-agents reach
+        this refusal too.
+        """
+        if self._refuse_delete and not self._dangerous:
+            return DeleteResult(error=self._DELETE_APPROVAL_ERROR)
+        return super().delete(file_path)
+
     def execute(self, command: str, *, timeout: int | None = None) -> ExecuteResponse:
         """
         Execute shell command in sandbox environment.
@@ -1248,16 +1619,172 @@ class CustomSandboxBackend(LocalShellBackend):
         - Access to paths outside workspace
         - Dangerous system commands
 
-        Then delegates to LocalShellBackend.execute() for actual execution.
+        The validated command is handed to the owned process runner so
+        cancelling an agent turn can terminate the complete process tree.
         """
+        # Preserve LocalShellBackend's public validation contract.  This
+        # override cannot delegate execution to the base implementation because
+        # it must retain the Popen handle for cancellation, so validate before
+        # command preparation and process launch instead.
+        if not command or not isinstance(command, str):
+            return ExecuteResponse(
+                output="Error: Command must be a non-empty string.",
+                exit_code=1,
+                truncated=False,
+            )
+
         command, error = prepare_sandbox_command(
-            command, self.cwd, virtual_mode=self.virtual_mode, dangerous=self._dangerous
+            command,
+            self.cwd,
+            virtual_mode=self.virtual_mode,
+            dangerous=self._dangerous,
+            guard_dangerous=self._guard_dangerous,
         )
         if error:
             return ExecuteResponse(output=error, exit_code=1, truncated=False)
 
-        # Delegate to parent for subprocess execution
-        response = super().execute(command, timeout=timeout)
+        return self._execute_prepared_command(command, timeout=timeout)
+
+    def _execute_prepared_command(
+        self,
+        command: str,
+        *,
+        timeout: int | None = None,
+    ) -> ExecuteResponse:
+        """Execute an already validated command in an owned process group."""
+
+        effective_timeout = timeout if timeout is not None else self._default_timeout
+        if effective_timeout <= 0:
+            msg = f"timeout must be positive, got {effective_timeout}"
+            raise ValueError(msg)
+
+        cancel_event = current_cancel_event()
+        if cancel_event is not None and cancel_event.is_set():
+            return ExecuteResponse(
+                output="Command cancelled before execution.",
+                exit_code=130,
+                truncated=False,
+            )
+
+        process: subprocess.Popen[str] | None = None
+        termination_reason: str | None = None
+        output_abandoned = False
+        try:
+            process_options: dict[str, object] = {}
+            if os.name == "nt":
+                process_options["creationflags"] = subprocess.CREATE_NEW_PROCESS_GROUP
+            else:
+                process_options["start_new_session"] = True
+
+            process = subprocess.Popen(
+                command,
+                shell=True,
+                stdout=subprocess.PIPE,
+                stderr=subprocess.PIPE,
+                stdin=subprocess.DEVNULL,
+                text=True,
+                env=self._env,
+                cwd=str(self.cwd),
+                **process_options,
+            )
+            _register_shell_process(cancel_event, process)
+            deadline = time.monotonic() + effective_timeout
+            drain_deadline: float | None = None
+
+            while True:
+                now = time.monotonic()
+                if (
+                    termination_reason is None
+                    and cancel_event is not None
+                    and cancel_event.is_set()
+                ):
+                    termination_reason = "cancelled"
+                    _terminate_process_tree(process)
+                    drain_deadline = now + _PROCESS_DRAIN_GRACE_SECONDS
+                elif termination_reason is None and now >= deadline:
+                    termination_reason = "timed_out"
+                    _terminate_process_tree(process)
+                    drain_deadline = now + _PROCESS_DRAIN_GRACE_SECONDS
+
+                if drain_deadline is not None and now >= drain_deadline:
+                    _stop_collecting_process_output(process)
+                    stdout = stderr = ""
+                    output_abandoned = True
+                    break
+
+                communicate_deadline = (
+                    drain_deadline if drain_deadline is not None else deadline
+                )
+                try:
+                    stdout, stderr = process.communicate(
+                        timeout=max(
+                            0.01,
+                            min(0.1, communicate_deadline - time.monotonic()),
+                        )
+                    )
+                    break
+                except subprocess.TimeoutExpired:
+                    continue
+
+            if termination_reason == "timed_out":
+                if timeout is not None:
+                    timeout_output = (
+                        "Error: Command timed out after "
+                        f"{effective_timeout} seconds (custom timeout). The command "
+                        "may be stuck or require more time."
+                    )
+                else:
+                    timeout_output = (
+                        f"Error: Command timed out after {effective_timeout} seconds. "
+                        "For long-running commands, re-run using the timeout parameter."
+                    )
+                response = ExecuteResponse(
+                    output=timeout_output,
+                    exit_code=124,
+                    truncated=output_abandoned,
+                )
+            elif termination_reason == "cancelled" or (
+                cancel_event is not None and cancel_event.is_set()
+            ):
+                response = ExecuteResponse(
+                    output="Command cancelled.",
+                    exit_code=130,
+                    truncated=output_abandoned,
+                )
+            else:
+                output_parts = []
+                if stdout:
+                    output_parts.append(stdout)
+                if stderr:
+                    stderr_lines = stderr.strip().split("\n")
+                    output_parts.extend(f"[stderr] {line}" for line in stderr_lines)
+                output = "\n".join(output_parts) if output_parts else "<no output>"
+
+                truncated = False
+                if len(output) > self._max_output_bytes:
+                    output = output[: self._max_output_bytes]
+                    output += (
+                        f"\n\n... Output truncated at {self._max_output_bytes} bytes."
+                    )
+                    truncated = True
+                if process.returncode != 0:
+                    output = f"{output.rstrip()}\n\nExit code: {process.returncode}"
+                response = ExecuteResponse(
+                    output=output,
+                    exit_code=process.returncode,
+                    truncated=truncated,
+                )
+        except Exception as exc:
+            if process is not None:
+                _terminate_process_tree(process)
+            response = ExecuteResponse(
+                output=f"Error executing command ({type(exc).__name__}): {exc}",
+                exit_code=1,
+                truncated=False,
+            )
+        finally:
+            if process is not None:
+                _unregister_shell_process(cancel_event, process)
 
         # Enhance timeout errors with actionable recovery guidance
         if response.exit_code == 124:
@@ -1325,6 +1852,12 @@ class AutoskillProposalSandboxBackend(CustomSandboxBackend):
             for file_path, _ in files
         ]
 
+    def delete(self, file_path: str) -> DeleteResult:
+        return DeleteResult(
+            error="Deletes are blocked for AutoSkills. Manage proposal files "
+            "under /autoskill-proposals/ instead."
+        )
+
     def execute(self, command: str, *, timeout: int | None = None) -> ExecuteResponse:
         return super().execute(
             self._rewrite_autoskill_mount(command),

EvoScientist/cancellation.py

@@ -0,0 +1,27 @@
+"""Cancellation context shared by streaming frontends and blocking tools."""
+
+from __future__ import annotations
+
+import contextvars
+import threading
+from collections.abc import Iterator
+from contextlib import contextmanager
+
+_current_cancel_event: contextvars.ContextVar[threading.Event | None] = (
+    contextvars.ContextVar("evoscientist_cancel_event", default=None)
+)
+
+
+@contextmanager
+def bind_cancel_event(event: threading.Event) -> Iterator[None]:
+    """Make a stream's cancellation event visible to nested sync tool calls."""
+    token = _current_cancel_event.set(event)
+    try:
+        yield
+    finally:
+        _current_cancel_event.reset(token)
+
+
+def current_cancel_event() -> threading.Event | None:
+    """Return the cancellation event bound to the current agent run, if any."""
+    return _current_cancel_event.get()

EvoScientist/channels/base.py

@@ -7,20 +7,24 @@ This module defines the Channel interface that all messaging channels
 import asyncio
 import logging
 import re
+import threading
 from abc import ABC, abstractmethod
 from collections import OrderedDict
 from collections.abc import AsyncIterator, Awaitable, Callable
 from collections.abc import Callable as CallableABC
 from dataclasses import dataclass, field
-from datetime import datetime
+from datetime import UTC, datetime
+from email.utils import parsedate_to_datetime
 from pathlib import Path
 from typing import Any
 
 from ..paths import MEDIA_DIR
+from ..runtime import AsyncRuntime
 from .bus.events import InboundMessage, OutboundMessage
 from .capabilities import ChannelCapabilities
 from .debug import TraceMixin, debug_trace_enabled
 from .formatter import UnifiedFormatter
+from .interaction import is_slash_command
 from .plugin import ChannelMeta, ChannelPlugin
 
 _logger = logging.getLogger(__name__)
@@ -298,6 +302,8 @@ class Channel(TraceMixin, ChannelPlugin, ABC):
             maxsize=queue_maxsize
         )
         self._running = False
+        self._startup_event = threading.Event()
+        self._startup_error: str | None = None
 
         # Global tracing can be enabled via shared config/env even when
         # individual channel factories have not been updated yet.
@@ -741,68 +747,144 @@ class Channel(TraceMixin, ChannelPlugin, ABC):
 
     # ── Send retry abstraction ──────────────────────────────────────
 
-    _non_retryable_patterns: tuple[str, ...] = ()
+    # HTTP status codes that should never be retried. Listed explicitly
+    # rather than as a 4xx range: 408 and 425 are retryable by definition and
+    # 429 is handled by the rate-limit path.
+    _non_retryable_status_codes: tuple[int, ...] = (400, 401, 403, 404)
+
+    # Structured SDK error codes that should never be retried (e.g. Slack invalid_auth)
+    # Channel-specific message patterns (e.g. Feishu 10003, DingTalk 40014) are handled
+    # via _non_retryable_patterns in respective channel subclasses.
+    _non_retryable_error_codes: tuple[str, ...] = (
+        "invalid_auth",
+        "invalid_token",
+        "expired_token",
+        "token_expired",
+        "token_revoked",
+        "account_inactive",
+        "not_authed",
+        "no_permission",
+        "missing_scope",
+    )
+
+    _non_retryable_patterns: tuple[str, ...] = (
+        "unauthorized",
+        "forbidden",
+        "permission denied",
+        "invalid token",
+        "invalid api key",
+        "authentication failed",
+    )
     _rate_limit_patterns: tuple[str, ...] = ("429", "ratelimit")
     _rate_limit_delay: float = 1.0
 
     def _extract_retry_after(self, exc: Exception) -> float | None:
         """Extract retry-wait seconds from an exception.
 
-        Returns ``None`` to signal that the error is **not retryable**.
+        Returns a retry delay in seconds, or ``None`` when the error is
+        explicitly non-retryable.
 
         Pipeline:
-        1. SDK-provided ``retry_after`` attribute (Telegram / Slack SDKs).
-        2. HTTP ``Retry-After`` header via :meth:`_parse_retry_after_header`.
-        3. Non-retryable pattern match → ``None``.
-        4. Rate-limit pattern match → ``_rate_limit_delay``.
-        5. Default ``1.0`` s (generic transient-error retry).
+        1. Non-retryable detection → ``None``. Evaluates HTTP status codes
+           (e.g. 401, 403), structured SDK error codes (e.g. Slack
+           ``"invalid_auth"``), and message pattern matching
+           (e.g. ``"unauthorized"``, ``"forbidden"``).
+        2. Server-supplied delay via :meth:`_extract_retry_delay`
+           (httpx ``Retry-After``; channels override for their SDK).
+        3. Rate-limit pattern match → ``_rate_limit_delay``.
+        4. Default ``1.0`` s for generic transient errors.
 
         Channels can customize behavior declaratively via class attributes
-        ``_non_retryable_patterns``, ``_rate_limit_patterns``, and
-        ``_rate_limit_delay``, or override this method entirely.
+        ``_non_retryable_patterns``, ``_rate_limit_patterns``,
+        ``_non_retryable_status_codes``, ``_non_retryable_error_codes``,
+        and ``_rate_limit_delay``, or override this method entirely.
         """
-        # 1. SDK retry_after attribute
-        retry = getattr(exc, "retry_after", None)
-        if retry is not None:
-            return float(retry)
+        # 1. Non-retryable detection: evaluate status codes, structured SDK
+        #    error codes, and message patterns independently.
+        status_code = self._extract_status_code(exc)
+        if status_code is not None and status_code in self._non_retryable_status_codes:
+            return None
 
-        # 2. HTTP Retry-After header
-        header_val = self._parse_retry_after_header(exc)
-        if header_val is not None:
-            return header_val
+        sdk_error = self._extract_sdk_error_code(exc)
+        if sdk_error is not None and sdk_error in self._non_retryable_error_codes:
+            return None
 
         msg = str(exc).lower()
-
-        # 3. Non-retryable patterns
         if self._non_retryable_patterns and any(
             p in msg for p in self._non_retryable_patterns
         ):
             return None
 
-        # 4. Rate-limit patterns
+        # 2. Server-supplied delay
+        delay = self._extract_retry_delay(exc)
+        if delay is not None:
+            return delay
+
+        # 3. Rate-limit patterns
         if self._rate_limit_patterns and any(
             p in msg for p in self._rate_limit_patterns
         ):
             return self._rate_limit_delay
 
-        # 5. Default
+        # 4. Default: transient error, retry with the standard delay
         return 1.0
 
-    def _parse_retry_after_header(self, exc: Exception) -> float | None:
-        """Try to extract a ``Retry-After`` value from an HTTP response."""
-        resp = getattr(exc, "response", None)
-        if resp is None:
-            return None
-        headers = getattr(resp, "headers", None)
-        if not headers:
-            return None
-        raw = headers.get("Retry-After") or headers.get("retry-after")
-        if raw is None:
-            return None
+    def _extract_status_code(self, exc: Exception) -> int | None:
+        """Extract HTTP status from an httpx error.
+
+        Channels with other SDKs (e.g. ``SlackChannel``, ``DiscordChannel``)
+        override this method.
+        """
+        import httpx
+
+        if isinstance(exc, httpx.HTTPStatusError):
+            return exc.response.status_code
+
+        return None
+
+    def _extract_sdk_error_code(self, exc: Exception) -> str | None:
+        """Extract structured SDK error code string from an exception.
+
+        Plain HTTP carries no structured error code by default (returns ``None``).
+        Subclasses with specialized SDKs (e.g. ``SlackChannel``, ``DiscordChannel``)
+        override this method.
+        """
+        return None
+
+    def _extract_retry_delay(self, exc: Exception) -> float | None:
+        """Retry delay the server asked for, in seconds, or ``None``.
+
+        Base implementation reads the ``Retry-After`` header of an httpx
+        error.  Channels whose SDK reports the delay differently
+        (``SlackChannel``, ``TelegramChannel``, ``DiscordChannel``) override.
+        """
+        import httpx
+
+        if isinstance(exc, httpx.HTTPStatusError):
+            raw = exc.response.headers.get("retry-after")
+            if raw is not None:
+                return self._parse_retry_after(raw)
+        return None
+
+    @staticmethod
+    def _parse_retry_after(raw: str) -> float | None:
+        """Convert a ``Retry-After`` header value to seconds.
+
+        RFC 9110 allows either delay-seconds or an HTTP-date; a date is
+        returned as the non-negative number of seconds until it.  Unparseable
+        values yield ``None`` so the caller can fall back to its own delay.
+        """
         try:
             return float(raw)
+        except ValueError:
+            pass
+        try:
+            when = parsedate_to_datetime(raw)
         except (ValueError, TypeError):
             return None
+        if when.tzinfo is None:
+            when = when.replace(tzinfo=UTC)
+        return max(0.0, (when - datetime.now(UTC)).total_seconds())
 
     async def _send_with_retry(
         self,
@@ -923,34 +1005,27 @@ class Channel(TraceMixin, ChannelPlugin, ABC):
             return None
         return self._raw_to_inbound(current)
 
-    def _build_inbound(self, raw: RawIncoming) -> InboundMessage | None:
+    def _build_inbound(
+        self,
+        raw: RawIncoming,
+        *,
+        runtime: AsyncRuntime | None = None,
+    ) -> InboundMessage | None:
         """Run *raw* through inbound middlewares and convert to InboundMessage.
 
-        Synchronous wrapper around :meth:`_build_inbound_async`.  When an
-        event loop is already running, the coroutine is scheduled on that
-        loop via :func:`asyncio.run_coroutine_threadsafe` to avoid
-        thread-safety issues with middleware state (DedupCache,
-        GroupHistoryBuffer, etc.).
+        Compatibility wrapper for synchronous integrations. Internal channel
+        implementations should await :meth:`_build_inbound_async` on their
+        transport loop. A caller may provide its application runtime to reuse
+        that owner; otherwise a runtime is scoped to this call.
+
+        This method deliberately rejects callers already running an event
+        loop. Blocking such a loop while scheduling the coroutine back onto it
+        deadlocks; async callers must await :meth:`_build_inbound_async`.
         """
-        import asyncio
-
-        try:
-            loop = asyncio.get_running_loop()
-        except RuntimeError:
-            loop = None
-
-        if loop is not None and loop.is_running():
-            future = asyncio.run_coroutine_threadsafe(
-                self._build_inbound_async(raw),
-                loop,
-            )
-            return future.result()
-        else:
-            new_loop = asyncio.new_event_loop()
-            try:
-                return new_loop.run_until_complete(self._build_inbound_async(raw))
-            finally:
-                new_loop.close()
+        if runtime is None:
+            with AsyncRuntime(thread_name="evosci-channel-adapter-runtime") as owned:
+                return self._build_inbound(raw, runtime=owned)
+        return runtime.run_sync(lambda: self._build_inbound_async(raw))
 
     def _raw_to_inbound(self, raw: RawIncoming) -> InboundMessage | None:
         """Convert a RawIncoming to InboundMessage (pure transformation, no filtering).
@@ -1054,6 +1129,43 @@ class Channel(TraceMixin, ChannelPlugin, ABC):
         """Buffer *msg* with debounce, then publish to bus."""
         sender = msg.sender_id
 
+        if self._on_activity:
+            try:
+                self._on_activity(sender, "received")
+            except Exception:
+                pass
+
+        # Slash commands are control messages, not prompt fragments. Flush any
+        # prompt already waiting for this sender, then publish the command as
+        # its own message so either arrival order cannot newline-merge them.
+        if is_slash_command(msg.content) and self._bus:
+            # A flush removes itself from this mapping before awaiting the bus
+            # publish.  Therefore a task still present here has not detached
+            # its buffered payload yet and is safe to cancel; an in-flight,
+            # backpressured publish is deliberately left alone.
+            debounce_task = self._debounce_tasks.pop(sender, None)
+            if debounce_task is not None:
+                debounce_task.cancel()
+                try:
+                    await debounce_task
+                except asyncio.CancelledError:
+                    # Awaiting a cancelled child normally raises here with no
+                    # cancellation pending on this task.  If our caller also
+                    # cancelled queue_message(), preserve that outer signal.
+                    current = asyncio.current_task()
+                    if current is not None and current.cancelling() > 0:
+                        raise
+            try:
+                await self._process_buffered_messages(sender)
+            except Exception:
+                _logger.error(
+                    f"{self.name} buffered-prompt flush failed for {sender}; "
+                    "publishing the command anyway",
+                    exc_info=True,
+                )
+            await self._bus.publish_inbound(msg)
+            return
+
         if sender not in self._message_buffers:
             self._message_buffers[sender] = []
             self._message_metadata[sender] = msg.metadata
@@ -1066,12 +1178,6 @@ class Channel(TraceMixin, ChannelPlugin, ABC):
         if msg.media:
             self._message_media[sender].extend(msg.media)
 
-        if self._on_activity:
-            try:
-                self._on_activity(sender, "received")
-            except Exception:
-                pass
-
         if sender in self._debounce_tasks:
             self._debounce_tasks[sender].cancel()
 
@@ -1086,8 +1192,10 @@ class Channel(TraceMixin, ChannelPlugin, ABC):
             await asyncio.sleep(_w)
             try:
                 await self._process_buffered_messages(_s)
-            except Exception as e:
-                _logger.error(f"{self.name} debounce flush error for {_s}: {e}")
+            except Exception:
+                _logger.error(
+                    f"{self.name} debounce flush error for {_s}", exc_info=True
+                )
 
         self._debounce_tasks[sender] = asyncio.create_task(debounce_callback())
 
@@ -1167,16 +1275,25 @@ class Channel(TraceMixin, ChannelPlugin, ABC):
         """Run the channel with auto-reconnect (exponential backoff)."""
         backoff = 1.0
         max_backoff = 60.0
+        self._startup_event.clear()
+        self._startup_error = None
         self._running = True
         while self._running:
             try:
                 await self.start()
+                self._startup_error = None
+                self._startup_event.set()
                 backoff = 1.0
                 async for msg in self.receive():
                     await self.queue_message(msg)
             except asyncio.CancelledError:
+                if not self._startup_event.is_set():
+                    self._startup_error = "startup cancelled"
+                    self._startup_event.set()
                 break
             except ChannelError as e:
+                self._startup_error = str(e)
+                self._startup_event.set()
                 self._trace_event(
                     "channel_fatal_error",
                     error_type=type(e).__name__,
@@ -1204,6 +1321,10 @@ class Channel(TraceMixin, ChannelPlugin, ABC):
                 await asyncio.sleep(backoff)
                 backoff = min(backoff * 2, max_backoff)
 
+        if not self._startup_event.is_set():
+            self._startup_error = "channel stopped before startup completed"
+            self._startup_event.set()
+
     # ── Channel allow-list check ─────────────────────────────────────
 
     def is_channel_allowed(self, channel_id: str) -> bool:

EvoScientist/channels/bus/events.py

@@ -45,6 +45,7 @@ class OutboundMessage:
     reply_to: str | None = None
     media: list[str] = field(default_factory=list)
     metadata: dict[str, Any] = field(default_factory=dict)
+    failure_notice: str | None = None
 
     @property
     def recipient(self) -> str:

EvoScientist/channels/channel_manager.py

@@ -17,7 +17,7 @@ import logging
 import pkgutil
 import time
 from collections.abc import Callable
-from dataclasses import dataclass, field
+from dataclasses import dataclass, field, replace
 from datetime import datetime
 from pathlib import Path
 from typing import Any
@@ -29,6 +29,11 @@ from .plugin import ChannelPlugin
 
 logger = logging.getLogger(__name__)
 
+# Best-effort failure notices must never wedge the dispatcher on a hung send.
+_FAILURE_NOTICE_TIMEOUT = 15.0
+
+CHANNEL_STARTUP_PENDING_DETAIL = "starting (bus)"
+
 
 # ═════════════════════════════════════════════════════════════════════
 # Account management (formerly account.py)
@@ -741,6 +746,12 @@ class ChannelManager:
                     delivery_failed = True
             if not delivery_failed and (msg.content or msg.media):
                 drained += 1
+            elif delivery_failed:
+                await self._send_failure_notice(
+                    channel,
+                    msg,
+                    timeout=max(1.0, deadline - time.monotonic()),
+                )
         dropped = self.bus.outbound.qsize()
         if drained or dropped:
             logger.info(f"Outbound drain: {drained} sent, {dropped} dropped")
@@ -841,6 +852,50 @@ class ChannelManager:
 
     # ── outbound routing ──
 
+    def _record_outbound_failure(self, channel_name: str, error: str) -> None:
+        health = self._health.get(channel_name)
+        if health is None:
+            return
+        health.consecutive_failures += 1
+        health.total_failures += 1
+        health.last_failure_time = time.time()
+        health.last_failure_error = error
+
+    async def _send_failure_notice(
+        self,
+        channel: Channel,
+        msg: OutboundMessage,
+        *,
+        timeout: float | None = None,
+    ) -> None:
+        """Best-effort short notice when the real payload could not be sent."""
+        if not msg.failure_notice:
+            return
+        fallback = replace(
+            msg,
+            content=msg.failure_notice,
+            media=[],
+            failure_notice=None,
+        )
+        try:
+            coro = channel.send(fallback)
+            if timeout is not None:
+                coro = asyncio.wait_for(coro, timeout=timeout)
+            fallback_ok = await coro
+        except Exception as fallback_error:
+            logger.error(
+                "Error sending delivery failure notice to %s: %s",
+                msg.channel,
+                fallback_error,
+            )
+        else:
+            if not fallback_ok:
+                logger.error(
+                    "Error sending delivery failure notice to %s: "
+                    "send() returned False",
+                    msg.channel,
+                )
+
     async def _dispatch_outbound(self) -> None:
         """Route outbound messages from the bus to the correct channel."""
         logger.info("Outbound dispatcher started")
@@ -870,13 +925,20 @@ class ChannelManager:
                     msg = processed
 
                 delivery_failed = False
+                failure_error = "one or more outbound deliveries failed"
                 if msg.content:
-                    text_ok = await channel.send(msg)
-                    if not text_ok:
-                        logger.error(
-                            f"Error sending to {msg.channel}: send() returned False"
-                        )
+                    try:
+                        text_ok = await channel.send(msg)
+                    except Exception as e:
+                        logger.error(f"Error sending to {msg.channel}", exc_info=True)
+                        failure_error = str(e)
                         delivery_failed = True
+                    else:
+                        if not text_ok:
+                            logger.error(
+                                f"Error sending to {msg.channel}: send() returned False"
+                            )
+                            delivery_failed = True
 
                 for media_path in msg.media:
                     try:
@@ -892,11 +954,18 @@ class ChannelManager:
                             )
                             delivery_failed = True
                     except Exception as e:
-                        logger.error(f"Error sending media to {msg.channel}: {e}")
+                        logger.error(
+                            f"Error sending media to {msg.channel}", exc_info=True
+                        )
+                        failure_error = str(e)
                         delivery_failed = True
 
                 if delivery_failed:
-                    raise RuntimeError("one or more outbound deliveries failed")
+                    await self._send_failure_notice(
+                        channel, msg, timeout=_FAILURE_NOTICE_TIMEOUT
+                    )
+                    self._record_outbound_failure(msg.channel, failure_error)
+                    continue
 
                 # Success
                 health = self._health.get(msg.channel)
@@ -904,13 +973,12 @@ class ChannelManager:
                     health.consecutive_failures = 0
                     health.total_successes += 1
             except Exception as e:
-                logger.error(f"Error sending to {msg.channel}: {e}")
-                health = self._health.get(msg.channel)
-                if health is not None:
-                    health.consecutive_failures += 1
-                    health.total_failures += 1
-                    health.last_failure_time = time.monotonic()
-                    health.last_failure_error = str(e)
+                # Unexpected internal error (pipeline, bookkeeping) — the
+                # transport paths above handle their own failures.
+                logger.error(
+                    f"Outbound dispatch error for {msg.channel}", exc_info=True
+                )
+                self._record_outbound_failure(msg.channel, str(e))
 
     # ── per-account lifecycle ──
 
@@ -978,6 +1046,31 @@ class ChannelManager:
         """Return names of currently running channels."""
         return [name for name, ch in self._channels.items() if ch._running]
 
+    def startup_results(self, *, timeout: float = 0.0) -> list[tuple[str, bool, str]]:
+        """Return each channel's initial connection result.
+
+        The optional timeout is shared across all channels, which start
+        concurrently. Channels still connecting when it expires are reported
+        as starting rather than connected.
+        """
+        deadline = time.monotonic() + max(timeout, 0.0)
+        for channel in self._channels.values():
+            remaining = deadline - time.monotonic()
+            if remaining > 0 and not channel._startup_event.is_set():
+                channel._startup_event.wait(remaining)
+
+        results: list[tuple[str, bool, str]] = []
+        for name, channel in self._channels.items():
+            if not channel._startup_event.is_set():
+                results.append((name, False, CHANNEL_STARTUP_PENDING_DETAIL))
+            elif channel._startup_error:
+                results.append((name, False, f"failed: {channel._startup_error}"))
+            elif channel._running:
+                results.append((name, True, "connected (bus)"))
+            else:
+                results.append((name, False, "stopped during startup"))
+        return results
+
     def get_stats(self) -> dict:
         """Return summary stats for all channels."""
         return {

EvoScientist/channels/consumer.py

@@ -20,6 +20,17 @@ from ..gateway import GraphGateway, GraphRunInput, GraphTarget, RunRequest
 from .base import Channel
 from .bus import MessageBus
 from .bus.events import InboundMessage, OutboundMessage
+from .capabilities import ChannelCapabilities
+from .interaction import (
+    ASK_USER_TIMEOUT,
+    HITL_APPROVAL_TIMEOUT,
+    REJECTED_FEEDBACK,
+    ApprovalPolicy,
+    InteractionIO,
+    PendingReplyRegistry,
+    resolve_approval,
+    resolve_ask_user,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -28,10 +39,6 @@ T = TypeVar("T")
 _MAX_CHAT_LOCKS = 10_000
 _MAX_SESSIONS = 10_000
 _MAX_HITL_ROUNDS = 50
-_HITL_APPROVAL_TIMEOUT = 120.0  # seconds to wait for HITL approval reply
-_ASK_USER_TIMEOUT = (
-    300.0  # seconds to wait for ask_user reply (longer for thinking time)
-)
 
 
 @dataclass
@@ -108,120 +115,56 @@ def _join_subagent_text(buffers: dict[str, tuple[str, list[str]]]) -> str:
     return "\n\n".join(sections)
 
 
-def _should_auto_approve(action_requests: list[dict]) -> bool:
-    """Check if all action requests can be auto-approved via config.
+class _ConsumerIO(InteractionIO):
+    """:class:`InteractionIO` over the consumer's bus + reply registry.
 
-    Returns True if no manual approval is needed (config auto_approve,
-    non-execute tools, or shell_allow_list match).
+    Publishes prompts through ``bus.publish_outbound`` and blocks for
+    replies on the consumer's shared :class:`PendingReplyRegistry` — both
+    on the consumer's own event loop, so the engine runs natively async
+    here with no thread hand-off.
     """
-    if not action_requests:
+
+    def __init__(
+        self, consumer: InboundConsumer, msg: InboundMessage, session_key: str
+    ) -> None:
+        self._consumer = consumer
+        self._msg = msg
+        self._session_key = session_key
+        self._last_reply_message: InboundMessage | None = None
+        channel = consumer._get_channel(msg.channel)
+        self.capabilities = (
+            channel.capabilities if channel is not None else ChannelCapabilities()
+        )
+        self.base_metadata = msg.metadata
+
+    async def send(self, content: str, *, metadata: dict | None = None) -> bool:
+        await self._consumer.bus.publish_outbound(
+            OutboundMessage(
+                channel=self._msg.channel,
+                chat_id=self._msg.chat_id,
+                content=content,
+                metadata=metadata if metadata is not None else self._msg.metadata,
+            )
+        )
         return True
 
-    try:
-        from ..config.settings import HITL_SHELL_TOOLS, load_config
+    async def wait_reply(self, *, timeout: float) -> str | None:
+        reply = await self._consumer._reply_registry.wait_event(
+            self._session_key, timeout
+        )
+        if reply is None:
+            self._last_reply_message = None
+            return None
+        self._last_reply_message = (
+            reply.context if isinstance(reply.context, InboundMessage) else None
+        )
+        return reply.content
 
-        cfg = load_config()
-    except Exception:
-        return False  # fail-closed
-
-    if cfg.auto_approve:
-        return True
-
-    shell_allow_list = (
-        [s.strip() for s in cfg.shell_allow_list.split(",") if s.strip()]
-        if cfg.shell_allow_list
-        else []
-    )
-
-    for req in action_requests:
-        name = req.get("name", "")
-        if name not in HITL_SHELL_TOOLS:
-            continue
-        args = req.get("args", {})
-        command = args.get("command", "") if isinstance(args, dict) else ""
-        cmd = command.strip()
-        if not any(cmd.startswith(prefix) for prefix in shell_allow_list):
-            return False
-    return True
-
-
-def _format_approval_prompt(
-    action_requests: list[dict], *, with_buttons: bool = False
-) -> str:
-    """Format an approval prompt as a text message for channel users.
-
-    When *with_buttons* is True, the trailing "Reply: 1=Approve..."
-    instruction is dropped — the buttons replace the textual cue.
-    """
-    lines = ["\u26a0\ufe0f Approval Required\n"]
-    for i, req in enumerate(action_requests, 1):
-        name = req.get("name", "")
-        args = req.get("args", {})
-        if isinstance(args, dict):
-            command = args.get("command", args.get("path", ""))
-        else:
-            command = ""
-        if command:
-            lines.append(f"  {i}. {name}: {command}")
-        else:
-            lines.append(f"  {i}. {name}")
-    if not with_buttons:
-        lines.append("")
-        lines.append("Reply: 1=Approve, 2=Reject, 3=Approve all")
-        lines.append("(Auto-reject in 2 min if no reply)")
-    return "\n".join(lines)
-
-
-def _parse_approval_reply(text: str) -> str | None:
-    """Parse a channel user's reply as an approval decision.
-
-    Returns "approve", "reject", "auto", or None if not recognized.
-    """
-    t = text.strip().lower()
-    if t in ("1", "y", "yes", "approve", "ok"):
-        return "approve"
-    if t in ("2", "n", "no", "reject"):
-        return "reject"
-    if t in ("3", "a", "auto", "approve all"):
-        return "auto"
-    return None
-
-
-def _approval_prompt_metadata(
-    base_metadata: dict | None, *, with_buttons: bool
-) -> dict:
-    """Outbound metadata for the HITL approval prompt.
-
-    When *with_buttons* is True, attaches Approve/Reject/Auto buttons whose
-    values match ``_parse_approval_reply`` so a click flows through the same
-    path as a typed ``"1"``/``"2"``/``"3"`` reply.
-    """
-    metadata = dict(base_metadata or {})
-    if with_buttons:
-        metadata["buttons"] = [
-            {"text": "Approve", "value": "1", "type": "primary"},
-            {"text": "Reject", "value": "2", "type": "danger"},
-            {"text": "Approve all", "value": "3"},
-        ]
-    return metadata
-
-
-@dataclass
-class _PendingInterrupt:
-    """Stored state for a pending HITL interrupt awaiting channel user reply."""
-
-    thread_id: str
-    action_requests: list
-    event: asyncio.Event  # set when user replies
-    decision: str | None = None  # "approve", "reject", "auto"
-
-
-@dataclass
-class _PendingAskUserReply:
-    """Stored state for a pending ask_user question awaiting channel user reply."""
-
-    event: asyncio.Event  # set when user replies
-    reply: str | None = None  # raw reply text
+    def take_reply_context(self) -> InboundMessage | None:
+        """Consume the last inbound reply context captured by ``wait_reply``."""
+        msg = self._last_reply_message
+        self._last_reply_message = None
+        return msg
 
 
 class InboundConsumer:
@@ -310,12 +253,12 @@ class InboundConsumer:
         # Metrics
         self._metrics = ConsumerMetrics()
 
-        # HITL: pending interrupts per session_key, and auto-approve sessions
-        self._pending_interrupts: dict[str, _PendingInterrupt] = {}
-        self._auto_approve_sessions: set[str] = set()
-
-        # ask_user: pending reply per session_key
-        self._pending_ask_user_replies: dict[str, _PendingAskUserReply] = {}
+        # Interaction engine state: one reply registry (routes the next
+        # message from a chat into a waiting prompt) and one approval
+        # policy (config rule + session "Approve all" grants), shared by
+        # the ask_user and HITL flows via ``channels.interaction``.
+        self._reply_registry = PendingReplyRegistry()
+        self._approval_policy = ApprovalPolicy()
 
     async def _get_thread_id(self, sender_id: str) -> str:
         """Get or create a thread ID for the given sender.
@@ -428,8 +371,6 @@ class InboundConsumer:
             except Exception:
                 pass
 
-        channel = self._get_channel(msg.channel)
-        thread_id = await self._get_thread_id(msg.sender_id)
         session_key = msg.session_key  # "channel:chat_id"
 
         # Lazily create per-chat lock; evict stale locks when too many
@@ -440,29 +381,39 @@ class InboundConsumer:
 
         self._metrics.total_processed += 1
 
-        # ask_user: check if this message is a reply to a pending question.
-        # Must be checked BEFORE HITL approval — any text is a valid answer.
-        if session_key in self._pending_ask_user_replies:
-            pending_ask = self._pending_ask_user_replies[session_key]
-            pending_ask.reply = msg.content
-            pending_ask.event.set()
-            return  # consumed as ask_user answer
+        # Reply interception: if a prompt (ask_user question or HITL
+        # approval) is waiting on this chat, hand it this message instead
+        # of starting a fresh agent turn.  The engine parses it (stop /
+        # cancel / choice / approval grammar), so the registry only routes
+        # text plus the original inbound context — one path for both flows.
+        if self._reply_registry.try_resolve(session_key, msg.content, context=msg):
+            return
 
-        # HITL: check if this message is a reply to a pending approval
-        if session_key in self._pending_interrupts:
-            pending = self._pending_interrupts[session_key]
-            decision = _parse_approval_reply(msg.content)
-            if decision is not None:
-                pending.decision = decision
-                pending.event.set()
-                return  # don't process as a new agent message
-            # Unrecognized reply — treat as new message, cancel pending
-            pending.decision = "reject"
-            pending.event.set()
-            del self._pending_interrupts[session_key]
+        # Resolved only for real agent turns — a consumed prompt reply must
+        # not create a graph thread or touch the sender-session LRU.
+        channel = self._get_channel(msg.channel)
+        thread_id = await self._get_thread_id(msg.sender_id)
 
         async with self._chat_locks[session_key]:
-            await self._stream_with_hitl(msg, channel, thread_id, session_key)
+            refeed = await self._stream_with_hitl(msg, channel, thread_id, session_key)
+
+        # An unrecognized reply to a pending approval rejects the action and
+        # then becomes a new agent turn. The lock was released above, so the
+        # previous turn has fully unwound before the refeed turn acquires it.
+        # Loops in case the refeed turn hits another approval that is again
+        # answered with unparseable text.
+        while refeed is not None:
+            channel = self._get_channel(refeed.channel)
+            thread_id = await self._get_thread_id(refeed.sender_id)
+            session_key = refeed.session_key
+            if session_key not in self._chat_locks:
+                self._chat_locks[session_key] = asyncio.Lock()
+                if len(self._chat_locks) > _MAX_CHAT_LOCKS:
+                    self._evict_chat_locks()
+            async with self._chat_locks[session_key]:
+                refeed = await self._stream_with_hitl(
+                    refeed, channel, thread_id, session_key
+                )
 
     async def _stream_with_hitl(
         self,
@@ -470,8 +421,13 @@ class InboundConsumer:
         channel: Channel | None,
         thread_id: str,
         session_key: str,
-    ) -> None:
-        """Stream agent events with HITL interrupt handling."""
+    ) -> InboundMessage | None:
+        """Stream agent events with HITL interrupt handling.
+
+        Returns ``None`` normally.  When a pending approval is answered
+        with unrecognized text, returns the intercepted inbound reply so the
+        caller can refeed it as a new agent turn after this one unwinds.
+        """
         from langgraph.types import Command
 
         stream_input: GraphRunInput = msg.content
@@ -609,108 +565,42 @@ class InboundConsumer:
                     stream_input = Command(resume=result)
                     continue
 
-                # HITL: resolve the interrupt
+                # HITL: resolve the interrupt through the shared engine.
+                # ``resolve_approval`` handles session/config auto-approve,
+                # the approval prompt (with capability-driven buttons), the
+                # reply wait, parsing (incl. /stop), and feedback strings.
                 action_reqs = interrupt_data.get("action_requests", [])
-                n = len(action_reqs) or 1
-
-                # Session auto-approve (user previously chose "Approve all")
-                if session_key in self._auto_approve_sessions:
-                    stream_input = Command(
-                        resume={"decisions": [{"type": "approve"} for _ in range(n)]}
-                    )
-                    continue
-
-                # Config auto-approve (auto_approve, non-execute, allow_list)
-                if _should_auto_approve(action_reqs):
-                    stream_input = Command(
-                        resume={"decisions": [{"type": "approve"} for _ in range(n)]}
-                    )
-                    continue
-
-                # Needs user approval — send prompt to channel
-                has_buttons = (
-                    channel is not None and channel.capabilities.inline_buttons
+                io = _ConsumerIO(self, msg, session_key)
+                outcome = await resolve_approval(
+                    action_reqs,
+                    io,
+                    self._approval_policy,
+                    session_key,
+                    timeout=HITL_APPROVAL_TIMEOUT,
                 )
-                prompt_text = _format_approval_prompt(
-                    action_reqs, with_buttons=has_buttons
-                )
-                approval_metadata = _approval_prompt_metadata(
-                    msg.metadata, with_buttons=has_buttons
-                )
-                await self.bus.publish_outbound(
-                    OutboundMessage(
-                        channel=msg.channel,
-                        chat_id=msg.chat_id,
-                        content=prompt_text,
-                        metadata=approval_metadata,
-                    )
-                )
-
-                # Wait for user reply
-                pending = _PendingInterrupt(
-                    thread_id=thread_id,
-                    action_requests=action_reqs,
-                    event=asyncio.Event(),
-                )
-                self._pending_interrupts[session_key] = pending
-
-                timed_out = False
-                try:
-                    await asyncio.wait_for(
-                        pending.event.wait(),
-                        timeout=_HITL_APPROVAL_TIMEOUT,
-                    )
-                except TimeoutError:
-                    timed_out = True
-                finally:
-                    # Unregister BEFORE any further await so a late reply can't flip
-                    # the decision back to approve during the notification round-trip.
-                    self._pending_interrupts.pop(session_key, None)
-
-                if timed_out:
-                    # Reject on timeout (fail-closed; matches cli/channel.py). Decision
-                    # is a local constant, not pending.decision, so it can't be
-                    # overwritten by a late reply after we unregistered above.
-                    decision = "reject"
-                    await self.bus.publish_outbound(
-                        OutboundMessage(
-                            channel=msg.channel,
-                            chat_id=msg.chat_id,
-                            content="⏰ Approval timed out. Action rejected.",
-                            metadata=msg.metadata,
+                if outcome.unrecognized_reply is not None:
+                    # Serve-mode policy: an unrecognized reply rejects the
+                    # pending action, confirms with reject feedback, and is
+                    # then processed as a new agent turn. The refeed is
+                    # returned to ``_handle_message`` so chat-lock ordering
+                    # stays serialized.
+                    await io.send(REJECTED_FEEDBACK)
+                    # In this flow, the final wait_reply call is exactly the
+                    # unrecognized approval reply. ask_user does not read this.
+                    refeed_msg = io.take_reply_context()
+                    if refeed_msg is None:
+                        logger.warning(
+                            "Unrecognized approval reply had no inbound context; "
+                            "dropping refeed"
                         )
-                    )
-                else:
-                    decision = pending.decision or "reject"
+                    return refeed_msg
+                if outcome.decisions is None:
+                    return None  # reject / timeout / stop — end the turn
 
-                # Visible confirmation so the click/reply registers (QQ has no
-                # message recall API for C2C).  Only fires when the user
-                # actually responded — silent on timeout to avoid claiming
-                # the user approved when they just walked away.
-                if pending.event.is_set():
-                    feedback_text = {
-                        "approve": "\u2705 已批准",
-                        "auto": "\u2705 已批准（后续自动通过）",
-                        "reject": "\u274c 已拒绝",
-                    }.get(decision)
-                    if feedback_text:
-                        await self.bus.publish_outbound(
-                            OutboundMessage(
-                                channel=msg.channel,
-                                chat_id=msg.chat_id,
-                                content=feedback_text,
-                                metadata=msg.metadata,
-                            )
-                        )
+                from ..backends import build_hitl_resume
 
-                if decision == "reject":
-                    return
-
-                if decision == "auto":
-                    self._auto_approve_sessions.add(session_key)
-
-                stream_input = Command(
-                    resume={"decisions": [{"type": "approve"} for _ in range(n)]}
+                stream_input = build_hitl_resume(
+                    interrupt_data.get("interrupt_id"), outcome.decisions
                 )
                 # continue to next HITL round
 
@@ -773,150 +663,27 @@ class InboundConsumer:
 
     # ── ask_user helpers ──
 
-    async def _wait_for_ask_user_reply(
-        self,
-        session_key: str,
-        timeout: float,
-    ) -> str | None:
-        """Register a pending ask_user slot and wait for the user to reply.
-
-        Returns the raw reply text, or ``None`` on timeout.
-        """
-        pending = _PendingAskUserReply(event=asyncio.Event())
-        self._pending_ask_user_replies[session_key] = pending
-        try:
-            await asyncio.wait_for(pending.event.wait(), timeout=timeout)
-        except TimeoutError:
-            pass
-        finally:
-            self._pending_ask_user_replies.pop(session_key, None)
-        return pending.reply
-
     async def _resolve_ask_user(
         self,
         msg: InboundMessage,
         event_data: dict,
         session_key: str,
     ) -> dict:
-        """Handle an ask_user interrupt: send questions to channel, collect answers.
+        """Handle an ask_user interrupt via the shared engine.
 
-        Mirrors the logic of ``cli.channel.channel_ask_user_prompt`` but runs
-        fully async inside the consumer event loop.
+        Delegates the whole question/answer flow (prompt formatting, choice
+        + "Other" grammar, ``/stop`` handling) to
+        :func:`channels.interaction.resolve_ask_user` over a
+        :class:`_ConsumerIO` adapter, so serve mode and the CLI bridge
+        cannot drift.
 
         Returns a dict suitable for ``Command(resume=...)``:
         ``{"answers": [...], "status": "answered"}`` or
         ``{"status": "cancelled"}``.
         """
         questions = event_data.get("questions", [])
-        if not questions:
-            return {"answers": [], "status": "answered"}
-
-        total = len(questions)
-        answers: list[str] = []
-
-        for i, q in enumerate(questions):
-            q_text = q.get("question", "")
-            q_type = q.get("type", "text")
-            required = q.get("required", True)
-
-            # -- Format question header --
-            if total == 1:
-                header = "\u2753 Quick check-in from EvoScientist\n"
-            else:
-                header = f"\u2753 Question {i + 1}/{total}\n"
-
-            lines: list[str] = [header, f"{i + 1}. {q_text}"]
-            if not required:
-                lines[-1] += " (optional)"
-
-            if q_type == "multiple_choice":
-                choices = q.get("choices", [])
-                for j, choice in enumerate(choices):
-                    label = choice.get("value", str(choice))
-                    letter = chr(ord("A") + j)
-                    lines.append(f"   {letter}. {label}")
-                other_letter = chr(ord("A") + len(choices))
-                lines.append(f"   {other_letter}. Other")
-                letters = "/".join(chr(ord("A") + k) for k in range(len(choices) + 1))
-                lines.append(f"\nReply with a letter ({letters}), or 'cancel'.")
-            else:
-                skip_hint = " Leave empty to skip." if not required else ""
-                lines.append(f"\nReply with your answer, or 'cancel'.{skip_hint}")
-
-            # -- Send question --
-            await self.bus.publish_outbound(
-                OutboundMessage(
-                    channel=msg.channel,
-                    chat_id=msg.chat_id,
-                    content="\n".join(lines),
-                    metadata=msg.metadata,
-                )
-            )
-
-            # -- Wait for user reply --
-            reply = await self._wait_for_ask_user_reply(
-                session_key,
-                _ASK_USER_TIMEOUT,
-            )
-
-            if not reply:
-                await self.bus.publish_outbound(
-                    OutboundMessage(
-                        channel=msg.channel,
-                        chat_id=msg.chat_id,
-                        content="\u23f0 Response timed out.",
-                        metadata=msg.metadata,
-                    )
-                )
-                return {"status": "cancelled"}
-
-            raw = reply.strip()
-            if raw.lower() == "cancel":
-                return {"status": "cancelled"}
-
-            # -- Parse answer --
-            if q_type == "multiple_choice":
-                choices = q.get("choices", [])
-                other_letter = chr(ord("A") + len(choices))
-                if len(raw) == 1 and raw.upper() == other_letter:
-                    # "Other" selected — ask for free-form input
-                    await self.bus.publish_outbound(
-                        OutboundMessage(
-                            channel=msg.channel,
-                            chat_id=msg.chat_id,
-                            content="Please type your answer:",
-                            metadata=msg.metadata,
-                        )
-                    )
-                    other_reply = await self._wait_for_ask_user_reply(
-                        session_key,
-                        _ASK_USER_TIMEOUT,
-                    )
-                    if not other_reply:
-                        await self.bus.publish_outbound(
-                            OutboundMessage(
-                                channel=msg.channel,
-                                chat_id=msg.chat_id,
-                                content="\u23f0 Response timed out.",
-                                metadata=msg.metadata,
-                            )
-                        )
-                        return {"status": "cancelled"}
-                    if other_reply.strip().lower() == "cancel":
-                        return {"status": "cancelled"}
-                    answers.append(other_reply.strip())
-                elif len(raw) == 1 and raw.upper().isalpha():
-                    idx = ord(raw.upper()) - ord("A")
-                    if 0 <= idx < len(choices):
-                        answers.append(choices[idx].get("value", raw))
-                    else:
-                        answers.append(raw)
-                else:
-                    answers.append(raw)
-            else:
-                answers.append(raw)
-
-        return {"answers": answers, "status": "answered"}
+        io = _ConsumerIO(self, msg, session_key)
+        return await resolve_ask_user(questions, io, timeout=ASK_USER_TIMEOUT)
 
     # ── internal ──
 

EvoScientist/channels/dingtalk/channel.py

@@ -35,7 +35,11 @@ class DingTalkChannel(Channel, WebSocketMixin, TokenMixin):
     capabilities = DINGTALK_CAPS
     name = "dingtalk"
     _ready_attrs = ("_http_client", "_access_token")
-    _non_retryable_patterns = ("invalidauthentication", "forbidden", "40014")
+    _non_retryable_patterns = (
+        *Channel._non_retryable_patterns,
+        "invalidauthentication",
+        "40014",
+    )
     _mention_pattern = r"@\S+\s*"
     _mention_strip_count = 1
 

EvoScientist/channels/discord/channel.py

@@ -208,6 +208,25 @@ class DiscordChannel(Channel):
             return str(self._client.user.id)
         return None
 
+    # ── Retry error code extraction (override base) ─────────────────
+
+    def _extract_status_code(self, exc: Exception) -> int | None:
+        """Extract HTTP status code from discord.HTTPException or fallback to base."""
+        import discord
+
+        if isinstance(exc, discord.HTTPException):
+            return exc.status
+        return super()._extract_status_code(exc)
+
+    def _extract_retry_delay(self, exc: Exception) -> float | None:
+        """Honor ``discord.RateLimited``, raised when a 429 exceeds
+        ``max_ratelimit_timeout`` and discord.py stops retrying internally."""
+        import discord
+
+        if isinstance(exc, discord.RateLimited):
+            return exc.retry_after
+        return super()._extract_retry_delay(exc)
+
     # ── Inbound ─────────────────────────────────────────────────────
 
     async def _on_message(self, message) -> None:

EvoScientist/channels/email/channel.py

@@ -73,7 +73,12 @@ class EmailChannel(Channel, PollingMixin):
     name = "email"
 
     capabilities = EMAIL_CAPS
-    _non_retryable_patterns = ("auth", "login", "credential")
+    _non_retryable_patterns = (
+        *Channel._non_retryable_patterns,
+        "auth",
+        "login",
+        "credential",
+    )
 
     def __init__(self, config: EmailConfig):
         super().__init__(config)

EvoScientist/channels/email/probe.py

@@ -25,7 +25,7 @@ async def validate_email_imap(
 
     import asyncio
 
-    loop = asyncio.get_event_loop()
+    loop = asyncio.get_running_loop()
 
     def _check():
         try:
@@ -62,7 +62,7 @@ async def validate_email_smtp(
 
     import asyncio
 
-    loop = asyncio.get_event_loop()
+    loop = asyncio.get_running_loop()
 
     def _check():
         server = None

EvoScientist/channels/feishu/channel.py

@@ -257,6 +257,7 @@ class FeishuChannel(Channel, WebhookMixin, TokenMixin):
     name = "feishu"
     _ready_attrs = ("_http_client", "_access_token")
     _non_retryable_patterns = (
+        *Channel._non_retryable_patterns,
         "app_access_token is empty",  # invalid credentials
         "10003",  # invalid app_id
         "10014",  # invalid app_secret
@@ -828,8 +829,18 @@ class FeishuChannel(Channel, WebhookMixin, TokenMixin):
         except Exception:
             return web.Response(status=400)
 
-        # ── Decrypt if encrypt_key is configured ──
-        if self.config.encrypt_key and "encrypt" in body:
+        # When encryption is configured the inbound POST MUST carry an
+        # ``encrypt`` field. A plaintext body used to skip decryption and
+        # reach the agent directly, defeating the encryption setup (issue
+        # #392). Treat a missing ``encrypt`` field on an
+        # encryption-configured channel as an authentication failure.
+        if self.config.encrypt_key:
+            if not isinstance(body, dict) or "encrypt" not in body:
+                logger.warning(
+                    "Feishu event rejected: encrypt_key is configured but the "
+                    "body has no 'encrypt' field (possible signature bypass)"
+                )
+                return web.Response(status=403)
             try:
                 body = self._decrypt_event(body["encrypt"])
             except Exception:

EvoScientist/channels/imessage/rpc_client.py

@@ -150,7 +150,7 @@ class ImsgRpcClient:
             "params": params or {},
         }
 
-        future: asyncio.Future = asyncio.get_event_loop().create_future()
+        future: asyncio.Future = asyncio.get_running_loop().create_future()
         self._pending[request_id] = future
 
         line = json.dumps(payload) + "\n"

EvoScientist/channels/interaction.py

@@ -0,0 +1,563 @@
+"""Transport-agnostic HITL and ask_user interaction engine.
+
+The module defines the channel-side protocol shared by serve mode and the
+CLI/TUI bridge: prompt formatting, reply grammar, stop handling, approval
+policy, pending-reply routing, and the async engine coroutines for approval
+and ask_user flows. Drivers provide transport-specific IO through
+:class:`InteractionIO`.
+"""
+
+from __future__ import annotations
+
+import asyncio
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Protocol
+
+if TYPE_CHECKING:
+    from .capabilities import ChannelCapabilities
+
+# ── timeout constants ──────────────────────────────────────────────────
+# Per-flow defaults.  HITL approval is short (a yes/no gate); ask_user is
+# longer because the human may need thinking time.
+HITL_APPROVAL_TIMEOUT = 120.0  # seconds to wait for a HITL approval reply
+ASK_USER_TIMEOUT = 300.0  # seconds to wait for an ask_user reply
+
+# ── stop-command grammar ──────────────────────────────────────────
+# Checked before reply parsing in *both* flows so a `/stop` mid-prompt
+# always cancels instead of being captured as a literal answer.
+_STOP_COMMANDS = frozenset(("/stop", "/cancel"))
+
+# ── feedback strings ─────────────────────────────────────────
+# Visible confirmations so a click/reply registers on channels without a
+# message-recall API (e.g. QQ C2C).
+APPROVED_FEEDBACK = "✅ Approved"
+APPROVED_AUTO_FEEDBACK = "✅ Approved (auto-approving future actions)"
+REJECTED_FEEDBACK = "❌ Rejected"
+UNRECOGNIZED_FEEDBACK = "Unrecognized reply. Action rejected."
+APPROVAL_TIMEOUT_FEEDBACK = "⏰ Approval timed out. Action rejected."
+ASK_USER_TIMEOUT_FEEDBACK = "⏰ Response timed out."
+OTHER_PROMPT = "Please type your answer:"
+
+# ── stop / cancel helpers ──────────────────────────────────────────────
+
+
+def is_slash_command(text: str | None) -> bool:
+    """Whether inbound content is a slash command (a control message, not a
+    prompt fragment)."""
+    return (text or "").lstrip().startswith("/")
+
+
+def is_stop_command(content: str | None) -> bool:
+    """Whether incoming content is a stop/cancel slash command."""
+    return (content or "").strip().lower() in _STOP_COMMANDS
+
+
+def is_cancel_reply(content: str | None) -> bool:
+    """Whether a reply is the literal ``cancel`` sentinel (case-insensitive)."""
+    return (content or "").strip().lower() == "cancel"
+
+
+# ── approval reply grammar ─────────────────────────────────────────────
+
+
+def parse_approval_reply(text: str) -> str | None:
+    """Parse a channel user's reply as an approval decision.
+
+    Returns "approve", "reject", "auto", or None if not recognized.
+    """
+    t = text.strip().lower()
+    if t in ("1", "y", "yes", "approve", "ok"):
+        return "approve"
+    if t in ("2", "n", "no", "reject"):
+        return "reject"
+    if t in ("3", "a", "auto", "approve all"):
+        return "auto"
+    return None
+
+
+def approve_decisions(action_requests: list) -> list[dict]:
+    """Build the ``decisions`` payload that approves every action request.
+
+    Length matches ``action_requests`` (with a floor of 1, matching the
+    consumer's historical ``len(...) or 1`` so an empty request list still
+    yields a single approve — the shape ``Command(resume=...)`` expects).
+    """
+    n = len(action_requests) or 1
+    return [{"type": "approve"} for _ in range(n)]
+
+
+# ── approval prompt formatting ─────────────────────────────────────────
+
+
+def format_approval_prompt(
+    action_requests: list[dict], *, with_buttons: bool = False
+) -> str:
+    """Format an approval prompt as a text message for channel users.
+
+    When *with_buttons* is True, the trailing "Reply: 1=Approve..."
+    instruction is dropped — the buttons replace the textual cue.
+    """
+    lines = ["⚠️ Approval Required\n"]
+    for i, req in enumerate(action_requests, 1):
+        name = req.get("name", "")
+        args = req.get("args", {})
+        if isinstance(args, dict):
+            # deepagents 0.7.0's `delete` tool uses `file_path`, not
+            # `command`/`path` — without this fallback the prompt shows
+            # only "delete" with no target.
+            command = args.get("command", args.get("path", args.get("file_path", "")))
+        else:
+            command = ""
+        if command:
+            lines.append(f"  {i}. {name}: {command}")
+        else:
+            lines.append(f"  {i}. {name}")
+    if not with_buttons:
+        lines.append("")
+        lines.append("Reply: 1=Approve, 2=Reject, 3=Approve all")
+        lines.append("(Auto-reject in 2 min if no reply)")
+    return "\n".join(lines)
+
+
+def approval_prompt_metadata(base_metadata: dict | None, *, with_buttons: bool) -> dict:
+    """Outbound metadata for the HITL approval prompt.
+
+    When *with_buttons* is True, attaches Approve/Reject/Auto buttons whose
+    values match ``parse_approval_reply`` so a click flows through the same
+    path as a typed ``"1"``/``"2"``/``"3"`` reply.
+    """
+    metadata = dict(base_metadata or {})
+    if with_buttons:
+        metadata["buttons"] = [
+            {"text": "Approve", "value": "1", "type": "primary"},
+            {"text": "Reject", "value": "2", "type": "danger"},
+            {"text": "Approve all", "value": "3"},
+        ]
+    return metadata
+
+
+# ── ask_user question formatting & answer grammar ──────────────────────
+
+
+def _choice_value(choice: object, fallback: str = "") -> str:
+    """Normalize one ask_user choice to its display/answer string.
+
+    Choices arrive from model-produced tool args; the schema says dicts with
+    a ``value`` key, but nothing enforces that at runtime, so plain strings
+    (or anything else) must not crash the prompt.
+    """
+    if isinstance(choice, dict):
+        return str(choice.get("value", fallback or choice))
+    return str(choice)
+
+
+def format_question_prompt(question: dict, index: int, total: int) -> str:
+    """Format one ask_user *question* as a channel message.
+
+    *index* is 0-based; *total* is the number of questions in the batch.
+    """
+    q_text = question.get("question", "")
+    q_type = question.get("type", "text")
+    required = question.get("required", True)
+
+    if total == 1:
+        header = "❓ Quick check-in from EvoScientist\n"
+    else:
+        header = f"❓ Question {index + 1}/{total}\n"
+
+    lines: list[str] = [header, f"{index + 1}. {q_text}"]
+    if not required:
+        lines[-1] += " (optional)"
+
+    if q_type == "multiple_choice":
+        choices = question.get("choices", [])
+        for j, choice in enumerate(choices):
+            label = _choice_value(choice)
+            letter = chr(ord("A") + j)
+            lines.append(f"   {letter}. {label}")
+        other_letter = chr(ord("A") + len(choices))
+        lines.append(f"   {other_letter}. Other")
+        letters = "/".join(chr(ord("A") + k) for k in range(len(choices) + 1))
+        lines.append(f"\nReply with a letter ({letters}), or 'cancel'.")
+    else:
+        skip_hint = " Leave empty to skip." if not required else ""
+        lines.append(f"\nReply with your answer, or 'cancel'.{skip_hint}")
+    return "\n".join(lines)
+
+
+def parse_choice_answer(raw: str, choices: list) -> tuple[str, str | None]:
+    """Classify a multiple-choice reply.
+
+    Returns ``(kind, value)``:
+
+    * ``("other", None)`` — the "Other" letter was chosen; the caller must
+      run the free-form sub-flow (send :data:`OTHER_PROMPT`, wait again).
+    * ``("answer", value)`` — a resolved answer string (the chosen
+      choice's ``value``, or the raw text when it isn't a valid letter).
+    """
+    other_letter = chr(ord("A") + len(choices))
+    if len(raw) == 1 and raw.upper() == other_letter:
+        return ("other", None)
+    if len(raw) == 1 and raw.upper().isalpha():
+        idx = ord(raw.upper()) - ord("A")
+        if 0 <= idx < len(choices):
+            return ("answer", _choice_value(choices[idx], raw))
+        return ("answer", raw)
+    return ("answer", raw)
+
+
+# ── approval policy ────────────────────────────────────────────────────
+
+
+def config_auto_approve(action_requests: list[dict]) -> bool:
+    """Whether config rules alone clear every action request.
+
+    Returns True if no manual approval is needed via config: the global
+    ``auto_approve`` flag, non-execute tools, or every shell command
+    resolving to :attr:`~EvoScientist.backends.ActionDecision.APPROVE` via
+    :func:`~EvoScientist.backends.resolve_action_decision` (token-boundary
+    ``shell_allow_list`` match, dangerous commands never auto-cleared).
+    Fail-closed on config load errors.
+    """
+    if not action_requests:
+        return True
+
+    try:
+        from ..backends import ActionDecision, resolve_action_decision
+        from ..config.settings import (
+            HITL_ALWAYS_PROMPT_TOOLS,
+            HITL_SHELL_TOOLS,
+            load_config,
+        )
+
+        cfg = load_config()
+    except Exception:
+        return False  # fail-closed
+
+    shell_allow_list = (
+        [s.strip() for s in cfg.shell_allow_list.split(",") if s.strip()]
+        if cfg.shell_allow_list
+        else []
+    )
+
+    for req in action_requests:
+        if not isinstance(req, dict):
+            return False  # malformed request — never auto-clear
+        name = req.get("name", "")
+        if name in HITL_ALWAYS_PROMPT_TOOLS:
+            return False
+        if name not in HITL_SHELL_TOOLS:
+            continue
+        args = req.get("args", {})
+        command = args.get("command", "") if isinstance(args, dict) else ""
+        verdict = resolve_action_decision(
+            command,
+            auto_approve=cfg.auto_approve,
+            dangerous_mode=cfg.dangerous_mode,
+            allow_list=shell_allow_list,
+        )
+        if verdict.decision is not ActionDecision.APPROVE:
+            return False
+    return True
+
+
+class ApprovalPolicy:
+    """Auto-approve policy backed by config rules and session grants.
+
+    One instance is owned per process. The consumer keeps one on its event
+    loop; the CLI bridge keeps one on the bus loop.
+    """
+
+    def __init__(self) -> None:
+        self._granted_sessions: set[str] = set()
+
+    def is_session_granted(self, session_key: str) -> bool:
+        """Whether the user previously chose "Approve all" for this session."""
+        return session_key in self._granted_sessions
+
+    def grant_session(self, session_key: str) -> None:
+        """Record an "Approve all" grant for this session."""
+        self._granted_sessions.add(session_key)
+
+    def clear_sessions(self) -> None:
+        """Forget all session grants (test hygiene / session reset)."""
+        self._granted_sessions.clear()
+
+    def auto_decision(
+        self, session_key: str, action_requests: list[dict]
+    ) -> list[dict] | None:
+        """Return an approve-all ``decisions`` list if this can auto-resolve.
+
+        Auto-resolves when the session was granted "Approve all" or when
+        config rules clear every request; otherwise returns ``None`` and
+        the caller must prompt the user.
+        """
+        if self.is_session_granted(session_key) or config_auto_approve(action_requests):
+            return approve_decisions(action_requests)
+        return None
+
+
+# ── transport adapter + reply registry ─────────────────────────────────
+
+
+class InteractionIO(Protocol):
+    """One conversation partner on one channel chat.
+
+    A transport adapter: the engine coroutines below drive a human
+    interaction entirely through this interface, so the same protocol
+    logic runs over the consumer's async loop and over the CLI bus loop.
+
+    Attributes
+    ----------
+    capabilities:
+        The channel's :class:`ChannelCapabilities` — the engine reads
+        ``inline_buttons`` to decide whether to attach approval buttons.
+    base_metadata:
+        The default outbound metadata for this chat (echoed back on each
+        send unless the engine supplies richer metadata, e.g. buttons).
+    """
+
+    capabilities: ChannelCapabilities
+    base_metadata: dict | None
+
+    async def send(self, content: str, *, metadata: dict | None = None) -> bool:
+        """Send *content* to the user; return True on success."""
+        ...
+
+    async def wait_reply(self, *, timeout: float) -> str | None:
+        """Wait for the user's next reply; return None on timeout."""
+        ...
+
+
+@dataclass(frozen=True)
+class PendingReply:
+    """A pending prompt reply plus optional transport-specific context."""
+
+    content: str
+    context: object | None = None
+
+
+class PendingReplyRegistry:
+    """Route "the next message from this chat" into a waiting coroutine.
+
+    One instance per process (the consumer owns one on its loop; the CLI
+    bridge owns one on the bus loop).  ``register`` / ``wait`` are used by
+    an :class:`InteractionIO` adapter to block for a reply; the inbound
+    interception point calls ``try_resolve`` to hand a message to that
+    waiter instead of enqueuing it as a fresh turn.
+
+    Asyncio-based: register/resolve must happen on the same event loop.
+    """
+
+    def __init__(self) -> None:
+        self._pending: dict[str, asyncio.Future[PendingReply]] = {}
+
+    def register(self, session_key: str) -> asyncio.Future[PendingReply]:
+        """Create and store a future awaiting the next reply for *session_key*."""
+        loop = asyncio.get_running_loop()
+        # A stale waiter for the same chat should never linger; cancel it
+        # so its coroutine unwinds instead of hanging until timeout.
+        stale = self._pending.get(session_key)
+        if stale is not None and not stale.done():
+            stale.cancel()
+        fut: asyncio.Future[PendingReply] = loop.create_future()
+        self._pending[session_key] = fut
+        return fut
+
+    def try_resolve(
+        self,
+        session_key: str,
+        content: str,
+        *,
+        context: object | None = None,
+    ) -> bool:
+        """Deliver *content* to a pending waiter.  Returns True if consumed."""
+        fut = self._pending.get(session_key)
+        if fut is not None and not fut.done():
+            fut.set_result(PendingReply(content=content, context=context))
+            return True
+        return False
+
+    def discard(self, session_key: str) -> None:
+        """Drop any pending waiter for *session_key* (idempotent)."""
+        self._pending.pop(session_key, None)
+
+    async def wait(self, session_key: str, timeout: float) -> str | None:
+        """Register, await a reply for *timeout* seconds, then clean up.
+
+        Returns the reply text, or ``None`` on timeout / cancellation.
+        """
+        reply = await self.wait_event(session_key, timeout)
+        return reply.content if reply is not None else None
+
+    async def wait_event(self, session_key: str, timeout: float) -> PendingReply | None:
+        """Register, await a reply event, then clean up.
+
+        Returns the full reply envelope, or ``None`` on timeout / registry
+        cancellation. Cancellation of the task awaiting this method propagates.
+        """
+        fut = self.register(session_key)
+        try:
+            done, _pending = await asyncio.wait({fut}, timeout=timeout)
+            if not done:
+                fut.cancel()
+                return None
+            try:
+                return fut.result()
+            except asyncio.CancelledError:
+                return None
+        finally:
+            # Identity-safe: only drop *our* slot, never a newer waiter that
+            # re-registered on the same chat while we were unwinding.
+            if self._pending.get(session_key) is fut:
+                self._pending.pop(session_key, None)
+
+    def clear(self) -> None:
+        """Cancel and forget every pending waiter (shutdown / test hygiene)."""
+        for fut in self._pending.values():
+            if not fut.done():
+                fut.cancel()
+        self._pending.clear()
+
+    def __contains__(self, session_key: str) -> bool:
+        return session_key in self._pending
+
+
+# ── engine coroutines ──────────────────────────────────────────────────
+
+
+async def resolve_ask_user(
+    questions: list[dict], io: InteractionIO, *, timeout: float = ASK_USER_TIMEOUT
+) -> dict:
+    """Drive an ask_user interrupt to a resume payload.
+
+    Sends each question in turn, collects answers, and handles the choice
+    grammar (letters + the "Other" free-form sub-flow).  ``/stop`` and
+    ``cancel`` are checked *before* parsing every reply.
+
+    Returns a dict suitable for ``Command(resume=...)``:
+    ``{"answers": [...], "status": "answered"}`` or ``{"status": "cancelled"}``.
+    """
+    if not questions:
+        return {"answers": [], "status": "answered"}
+
+    total = len(questions)
+    answers: list[str] = []
+
+    for i, q in enumerate(questions):
+        if not await io.send(format_question_prompt(q, i, total)):
+            return {"status": "cancelled"}
+
+        reply = await io.wait_reply(timeout=timeout)
+        if reply is None:
+            await io.send(ASK_USER_TIMEOUT_FEEDBACK)
+            return {"status": "cancelled"}
+
+        raw = reply.strip()
+        required = q.get("required", True) is not False
+        if raw == "":
+            if required:
+                return {"status": "cancelled"}
+            answers.append("")
+            continue
+
+        if is_stop_command(raw) or is_cancel_reply(raw):
+            return {"status": "cancelled"}
+
+        if q.get("type", "text") == "multiple_choice":
+            choices = q.get("choices", [])
+            kind, value = parse_choice_answer(raw, choices)
+            if kind == "other":
+                if not await io.send(OTHER_PROMPT):
+                    return {"status": "cancelled"}
+                other = await io.wait_reply(timeout=timeout)
+                if other is None:
+                    await io.send(ASK_USER_TIMEOUT_FEEDBACK)
+                    return {"status": "cancelled"}
+                other_raw = other.strip()
+                if other_raw == "":
+                    if required:
+                        return {"status": "cancelled"}
+                    answers.append("")
+                    continue
+                if is_stop_command(other_raw) or is_cancel_reply(other_raw):
+                    return {"status": "cancelled"}
+                answers.append(other_raw)
+            else:
+                answers.append(value)
+        else:
+            answers.append(raw)
+
+    return {"answers": answers, "status": "answered"}
+
+
+@dataclass
+class ApprovalOutcome:
+    """Result of :func:`resolve_approval`.
+
+    ``decisions`` is the approve-all payload on approve/auto, or ``None``
+    when the action was declined (reject / timeout / stop / unrecognized).
+
+    ``unrecognized_reply`` carries the raw reply text when parsing failed.
+    The engine centralizes *parsing* but does not decide the transport
+    policy for unparseable text — the drivers do: the consumer rejects the
+    pending action and refeeds the text as a new agent turn (a channel user
+    who ignores the prompt and types a fresh instruction must not lose it);
+    the CLI bridge sends :data:`UNRECOGNIZED_FEEDBACK` and declines.
+    """
+
+    decisions: list[dict] | None = None
+    unrecognized_reply: str | None = None
+
+
+async def resolve_approval(
+    action_requests: list,
+    io: InteractionIO,
+    policy: ApprovalPolicy,
+    session_key: str,
+    *,
+    timeout: float = HITL_APPROVAL_TIMEOUT,
+) -> ApprovalOutcome:
+    """Drive a HITL approval interrupt to an :class:`ApprovalOutcome`.
+
+    Auto-resolves via *policy* (session grant or config rule) without
+    prompting.  Otherwise sends the approval prompt (with capability-driven
+    buttons), waits for a reply, and parses it.  ``/stop`` cancels silently
+    (it already got its own ack from the transport's stop fast-path).  An
+    unrecognized reply declines *without feedback* and hands the raw text
+    back to the driver via ``unrecognized_reply`` (see
+    :class:`ApprovalOutcome` for the per-driver policy).
+    """
+    auto = policy.auto_decision(session_key, action_requests)
+    if auto is not None:
+        return ApprovalOutcome(decisions=auto)
+
+    has_buttons = bool(io.capabilities.inline_buttons)
+    prompt = format_approval_prompt(action_requests, with_buttons=has_buttons)
+    metadata = approval_prompt_metadata(io.base_metadata, with_buttons=has_buttons)
+    if not await io.send(prompt, metadata=metadata):
+        return ApprovalOutcome()
+
+    reply = await io.wait_reply(timeout=timeout)
+    if reply is None:
+        await io.send(APPROVAL_TIMEOUT_FEEDBACK)
+        return ApprovalOutcome()
+
+    if is_stop_command(reply):
+        return ApprovalOutcome()
+
+    decision = parse_approval_reply(reply)
+    if decision == "auto":
+        policy.grant_session(session_key)
+        await io.send(APPROVED_AUTO_FEEDBACK)
+        return ApprovalOutcome(decisions=approve_decisions(action_requests))
+    if decision == "approve":
+        await io.send(APPROVED_FEEDBACK)
+        return ApprovalOutcome(decisions=approve_decisions(action_requests))
+    if decision == "reject":
+        await io.send(REJECTED_FEEDBACK)
+        return ApprovalOutcome()
+
+    # Unrecognized — decline and report the raw text; the driver chooses
+    # the feedback / refeed policy.
+    return ApprovalOutcome(unrecognized_reply=reply)

EvoScientist/channels/middleware.py

@@ -23,6 +23,7 @@ from typing import Any
 from .base import RawIncoming
 from .bus.events import InboundMessage, OutboundMessage
 from .debug import emit_debug_event_if
+from .interaction import is_slash_command
 
 _logger = logging.getLogger(__name__)
 
@@ -811,8 +812,21 @@ class MentionGatingMiddleware(InboundMiddleware):
                 policy=self.require_mention,
             )
             return None
-        # Strip mentions from group messages
-        if raw.is_group and self._strip_fn:
+        # A slash command's platform target belongs only to its first token;
+        # preserve mentions in its arguments. Ordinary group messages may
+        # still carry a bot mention elsewhere and use the full-message strip.
+        if self._strip_fn and is_slash_command(raw.text):
+            text = raw.text
+            token_start = len(text) - len(text.lstrip())
+            token_end = token_start
+            while token_end < len(text) and not text[token_end].isspace():
+                token_end += 1
+            stripped_token = self._strip_fn(text[token_start:token_end])
+            raw = dataclasses.replace(
+                raw,
+                text=text[:token_start] + stripped_token + text[token_end:],
+            )
+        elif self._strip_fn and raw.is_group:
             raw = dataclasses.replace(raw, text=self._strip_fn(raw.text))
         return raw
 
@@ -928,6 +942,12 @@ class GroupHistoryMiddleware(InboundMiddleware):
             # Don't drop here — let MentionGatingMiddleware handle that
             return raw
 
+        # Slash commands must remain the leading content so channel command
+        # dispatchers can recognize them. Keep buffered chatter for the next
+        # normal mentioned message instead of injecting it ahead of a command.
+        if is_slash_command(raw.text):
+            return raw
+
         # Mentioned: inject history context
         history_context = self._buffer.format_context(raw.chat_id)
         if history_context:

EvoScientist/channels/qq/channel.py

@@ -116,7 +116,7 @@ class QQChannel(Channel):
 
     capabilities = QQ_CAPS
     _ready_attrs = ("_client", "_running")
-    _non_retryable_patterns = ()
+    _non_retryable_patterns = Channel._non_retryable_patterns
     _mention_pattern = r"@\S+\s*"
     _mention_strip_count = 1
     _markdown_fallback_exc_types: ClassVar[tuple[type[Exception], ...]] = (
@@ -235,7 +235,7 @@ class QQChannel(Channel):
 
         Surfaces the click as an :class:`InboundMessage` whose ``content`` is
         the button's ``data`` verbatim — so a "1"/"approve"/… click flows
-        through ``_parse_approval_reply`` exactly like a typed reply.
+        through ``parse_approval_reply`` exactly like a typed reply.
 
         The click runs through inbound middleware (Dedup suppresses QQ
         retries) but is published directly to the bus so the per-sender
@@ -264,7 +264,7 @@ class QQChannel(Channel):
             triggering_msg_id = getattr(resolved, "message_id", "") or ""
 
             # QQ may serialize non-str values; coerce.  Fall back to button id
-            # when no data — same path as a typed reply via _parse_approval_reply.
+            # when no data — same path as a typed reply via parse_approval_reply.
             button_value = str(button_data) if button_data != "" else ""
             text = button_value or button_id
 
@@ -360,7 +360,7 @@ class QQChannel(Channel):
         plain_text = self._plain_formatter.format(raw_text)
         # Plain-text fallback can't carry a keyboard.  Append `value=label`
         # pairs so the user can still type "1"/"approve"/… instead of
-        # tapping (`_parse_approval_reply` accepts the same values).
+        # tapping (`parse_approval_reply` accepts the same values).
         if buttons:
             pairs = []
             for btn in buttons:

EvoScientist/channels/signal/channel.py

@@ -32,7 +32,11 @@ class SignalChannel(Channel):
     name = "signal"
 
     capabilities = SIGNAL_CAPS
-    _non_retryable_patterns = ("unregistered", "auth")
+    _non_retryable_patterns = (
+        *Channel._non_retryable_patterns,
+        "unregistered",
+        "auth",
+    )
 
     def __init__(self, config: SignalConfig):
         super().__init__(config)

EvoScientist/channels/signal/probe.py

@@ -22,7 +22,7 @@ async def validate_signal(
         return False, "phone_number is required"
 
     # Check signal-cli binary
-    loop = asyncio.get_event_loop()
+    loop = asyncio.get_running_loop()
 
     def _check():
         try:

EvoScientist/channels/slack/channel.py

@@ -19,6 +19,13 @@ class SlackConfig(BaseChannelConfig):
     text_chunk_limit: int = 4096
 
 
+def _slack_response_types() -> tuple[type, ...]:
+    from slack_sdk.web.async_slack_response import AsyncSlackResponse
+    from slack_sdk.web.slack_response import SlackResponse
+
+    return (SlackResponse, AsyncSlackResponse)
+
+
 class SlackChannel(Channel):
     """Slack channel using slack-sdk Socket Mode."""
 
@@ -195,6 +202,45 @@ class SlackChannel(Channel):
     def _get_bot_identifier(self) -> str | None:
         return getattr(self, "_bot_user_id", None)
 
+    # ── Retry error code extraction (override base) ─────────────────
+
+    def _extract_status_code(self, exc: Exception) -> int | None:
+        """Extract HTTP status code from SlackApiError or fallback to base."""
+        from slack_sdk.errors import SlackApiError
+
+        if isinstance(exc, SlackApiError) and isinstance(
+            exc.response, _slack_response_types()
+        ):
+            return exc.response.status_code
+        return super()._extract_status_code(exc)
+
+    def _extract_sdk_error_code(self, exc: Exception) -> str | None:
+        """Extract structured error code string from SlackApiError."""
+        from slack_sdk.errors import SlackApiError
+
+        if isinstance(exc, SlackApiError) and isinstance(
+            exc.response, _slack_response_types()
+        ):
+            error = exc.response.get("error")
+            return error.lower() if isinstance(error, str) else None
+        return super()._extract_sdk_error_code(exc)
+
+    def _extract_retry_delay(self, exc: Exception) -> float | None:
+        """Read Slack's ``Retry-After`` header from a SlackApiError.
+
+        ``SlackResponse.headers`` is a plain ``dict`` whose key casing depends
+        on the HTTP client, so match the key case-insensitively (the same
+        approach slack_sdk's own ``RateLimitErrorRetryHandler`` takes).
+        """
+        from slack_sdk.errors import SlackApiError
+
+        if isinstance(exc, SlackApiError):
+            for key, raw in exc.response.headers.items():
+                if key.lower() == "retry-after":
+                    return self._parse_retry_after(raw)
+            return None
+        return super()._extract_retry_delay(exc)
+
     # ── ACK Reactions ───────────────────────────────────────────────
 
     async def _send_ack_reaction(

EvoScientist/channels/standalone.py

@@ -26,6 +26,13 @@ from .debug import emit_debug_event
 logger = logging.getLogger(__name__)
 
 
+async def _create_standalone_agent():
+    """Construct the synchronous agent without blocking the channel loop."""
+    from ..EvoScientist import create_cli_agent
+
+    return await asyncio.to_thread(create_cli_agent)
+
+
 def _channel_trace_enabled(channel: Channel) -> bool:
     """Check if debug tracing is enabled on the channel."""
     try:
@@ -107,10 +114,12 @@ async def _async_main(
     consumer: InboundConsumer | None = None
     if use_agent:
         logger.info("Loading EvoScientist agent...")
-        from ..EvoScientist import create_cli_agent
         from ..gateway import create_runtime_gateways
 
-        agent = create_cli_agent()
+        # Agent construction performs synchronous MCP discovery through the
+        # owned-runtime bridge.  Keep it off this already-running channel loop
+        # (and avoid blocking channel health/startup work while it loads).
+        agent = await _create_standalone_agent()
         runtime_gateways = create_runtime_gateways()
         logger.info("Agent loaded")
 
@@ -151,7 +160,7 @@ async def _async_main(
         await channel.stop()
         await manager.stop_health()
 
-    loop = asyncio.get_event_loop()
+    loop = asyncio.get_running_loop()
     for sig in (signal.SIGINT, signal.SIGTERM):
         loop.add_signal_handler(
             sig,

EvoScientist/channels/telegram/channel.py

@@ -2,7 +2,7 @@
 
 import logging
 from dataclasses import dataclass
-from datetime import datetime
+from datetime import datetime, timedelta
 from pathlib import Path
 from typing import ClassVar
 
@@ -34,7 +34,11 @@ class TelegramChannel(Channel):
     capabilities = TELEGRAM_CAPS
     _typing_interval: float = 4.0
     _ready_attrs = ("_app",)
-    _non_retryable_patterns = ("parse", "can't parse")
+    _non_retryable_patterns = (
+        *Channel._non_retryable_patterns,
+        "parse",
+        "can't parse",
+    )
     _mention_pattern = r"(?i)@{bot_id}\s*"
 
     def __init__(self, config: TelegramConfig):
@@ -80,9 +84,7 @@ class TelegramChannel(Channel):
                 | filters.LOCATION
             )
 
-        self._app.add_handler(
-            MessageHandler(media_filter & ~filters.COMMAND, self._on_message)
-        )
+        self._app.add_handler(MessageHandler(media_filter, self._on_message))
 
         await self._app.initialize()
         # Cache bot username for @mention detection in groups
@@ -94,12 +96,17 @@ class TelegramChannel(Channel):
         logger.info("Telegram channel started (polling)")
 
     async def _cleanup(self) -> None:
-        if self._app:
-            if self._app.updater and self._app.updater.running:
-                await self._app.updater.stop()
-            await self._app.stop()
-            await self._app.shutdown()
-            logger.info("Telegram channel stopped")
+        app = self._app
+        self._app = None
+        if app is None:
+            return
+
+        if app.updater and app.updater.running:
+            await app.updater.stop()
+        if app.running:
+            await app.stop()
+        await app.shutdown()
+        logger.info("Telegram channel stopped")
 
     # ── Typing indicator (override base) ────────────────────────────
 
@@ -111,6 +118,21 @@ class TelegramChannel(Channel):
                 action="typing",
             )
 
+    # ── Retry delay extraction (override base) ─────────────────────
+
+    def _extract_retry_delay(self, exc: Exception) -> float | None:
+        """Honor Telegram flood control (``telegram.error.RetryAfter``).
+
+        ``retry_after`` is an ``int`` by default and a ``timedelta`` when the
+        ``PTB_TIMEDELTA`` opt-in is enabled.
+        """
+        from telegram.error import RetryAfter
+
+        if isinstance(exc, RetryAfter):
+            ra = exc.retry_after
+            return ra.total_seconds() if isinstance(ra, timedelta) else float(ra)
+        return super()._extract_retry_delay(exc)
+
     # ── Send (template method overrides) ──────────────────────────
 
     async def _send_chunk(self, chat_id, formatted_text, raw_text, reply_to, metadata):
@@ -161,6 +183,21 @@ class TelegramChannel(Channel):
     def _get_bot_identifier(self) -> str | None:
         return self._bot_username or None
 
+    @staticmethod
+    def _command_target(text: str) -> str | None:
+        """Return a Telegram command's target username.
+
+        An empty string represents a bare command; ``None`` means the message
+        is not command-shaped.
+        """
+        parts = text.lstrip().split(None, 1)
+        if not parts or not parts[0].startswith("/"):
+            return None
+        command_token = parts[0][1:]
+        if "@" not in command_token:
+            return ""
+        return command_token.rsplit("@", 1)[1].lower()
+
     async def _send_ack_reaction(
         self, chat_id: str, message_id: str, emoji: str = "👀"
     ) -> None:
@@ -202,10 +239,19 @@ class TelegramChannel(Channel):
 
         # Detect group and mention status for centralized gating
         is_group = message.chat.type in ("group", "supergroup")
-        was_mentioned = True  # DM default
-        if is_group and self._bot_username:
+        was_mentioned = not is_group
+        if is_group:
             text_check = (message.text or message.caption or "").lower()
-            was_mentioned = f"@{self._bot_username}" in text_check
+            command_target = self._command_target(text_check)
+            if command_target is not None:
+                # A bare command that Telegram delivered to this bot is
+                # actionable. Commands explicitly addressed to another bot
+                # must remain ignored.
+                was_mentioned = not command_target or (
+                    bool(self._bot_username) and command_target == self._bot_username
+                )
+            elif self._bot_username:
+                was_mentioned = f"@{self._bot_username}" in text_check
 
         content_parts: list[str] = []
         media_paths: list[str] = []

EvoScientist/channels/wechat/channel.py

@@ -337,9 +337,21 @@ class WeChatChannel(Channel, WebhookMixin, TokenMixin):
         logger.info(f"WeChat callback POST received, body length={len(body)}")
         xml_data = parse_xml(body)
 
-        # If encrypted, decrypt first
+        # If encryption is configured, the inbound POST MUST carry an
+        # <Encrypt> element and a matching msg_signature. An unsigned body
+        # used to fall through to _safe_process_message and reach the agent
+        # regardless of credentials, which made the encryption setup
+        # ineffective (issue #392). Treat a missing <Encrypt> on an
+        # encryption-configured channel as an authentication failure.
         encrypt = xml_data.get("Encrypt", "")
-        if encrypt and self._crypto:
+        if self._crypto:
+            if not encrypt:
+                logger.warning(
+                    "WeChat POST rejected: encryption is configured but the "
+                    "body has no <Encrypt> element (possible signature bypass)"
+                )
+                return web.Response(status=403)
+
             signature = request.query.get("msg_signature", "")
             timestamp = request.query.get("timestamp", "")
             nonce = request.query.get("nonce", "")

EvoScientist/cli/__init__.py

@@ -79,6 +79,11 @@ def main():
     warnings.filterwarnings(
         "ignore", message=".*type is unknown and inference may fail.*"
     )
+    # v3 streaming is a deliberate choice (#268), so its beta notice is noise.
+    # Matched by message, not category, to keep other beta warnings visible.
+    warnings.filterwarnings(
+        "ignore", message=".*v3 streaming protocol on Pregel is experimental.*"
+    )
     from ..config import load_config
     from .commands import _configure_logging
 

EvoScientist/cli/_app.py

@@ -59,6 +59,12 @@ sessions_app = typer.Typer(
 )
 app.add_typer(sessions_app, name="sessions")
 
+# Background langgraph dev server management — the explicit counterpart to
+# langgraph_dev_keepalive: a server that outlives its CLI needs a first-class
+# way to inspect and stop it.
+server_app = typer.Typer(help="Manage the background langgraph dev server")
+app.add_typer(server_app, name="server")
+
 # Configure subcommand group — re-run a single onboarding section.
 configure_app = typer.Typer(
     help=(

EvoScientist/cli/agent.py

@@ -10,6 +10,8 @@ from ..paths import new_run_dir
 if TYPE_CHECKING:
     from langgraph.graph.state import CompiledStateGraph
 
+    from ..runtime import AsyncRuntime
+
 
 def _shorten_path(path: str) -> str:
     """Shorten absolute path to relative path from current directory."""
@@ -69,6 +71,8 @@ def _load_agent(
     chat_model=None,
     *,
     on_mcp_progress=None,
+    events=None,
+    runtime: "AsyncRuntime | None" = None,
 ) -> "CompiledStateGraph":
     """Load the CLI agent with optional persistent checkpointer.
 
@@ -83,6 +87,7 @@ def _load_agent(
             selects the pure (no module-global write) build path.
         on_mcp_progress: Optional per-server MCP progress callback.
             Signature ``(event, server_name, detail) -> None``.
+        runtime: Optional application-scoped runtime used for MCP discovery.
     """
     from ..EvoScientist import create_cli_agent
 
@@ -92,4 +97,6 @@ def _load_agent(
         config=config,
         chat_model=chat_model,
         on_mcp_progress=on_mcp_progress,
+        events=events,
+        runtime=runtime,
     )

EvoScientist/cli/async_notifier.py

@@ -117,6 +117,71 @@ def _enqueue(notification: AsyncTaskNotification) -> None:
     q.put(notification)
 
 
+def enqueue_task_notification(notification: AsyncTaskNotification) -> None:
+    """Public :class:`~EvoScientist.middleware.notifier.NotifierPort` entry point.
+
+    Route a completed-task notification onto the consumer queue. Thin wrapper
+    over :func:`_enqueue` so middleware can enqueue without reaching into the
+    module's private symbols.
+    """
+    _enqueue(notification)
+
+
+def enqueue_bg_process_notification(
+    *,
+    task_id: str,
+    agent_name: str,
+    status: str,
+    prompt: str = "",
+    origin_cli_thread_id: str | None = None,
+) -> None:
+    """Build and enqueue a background-process completion notification.
+
+    :class:`~EvoScientist.middleware.notifier.NotifierPort` entry point used by
+    the background middleware so it never constructs the CLI-owned
+    :class:`AsyncTaskNotification` itself — the ``kind="bg-process"`` tag and the
+    UTC ``received_at`` timestamp are filled in here.
+    """
+    _enqueue(
+        AsyncTaskNotification(
+            task_id=task_id,
+            agent_name=agent_name,
+            status=status,
+            received_at=datetime.now(UTC).strftime("%Y-%m-%dT%H:%M:%SZ"),
+            prompt=prompt,
+            kind="bg-process",
+            origin_cli_thread_id=origin_cli_thread_id,
+        )
+    )
+
+
+def pre_cancel_watcher(task_id: str) -> None:
+    """Cancel a stale watcher for ``task_id`` before a new run replaces it.
+
+    ``update_async_task`` starts a new run on the same ``thread_id`` with
+    ``multitask_strategy="interrupt"``, which closes the old run's stream
+    cleanly. Without pre-cancellation the old watcher would observe that clean
+    exit and enqueue a stale "success" notification before the new spawn can
+    replace it. Cancellation propagates ``CancelledError`` (a ``BaseException``)
+    which the watcher's ``except Exception:`` does not catch, so ``_enqueue``
+    never runs for the cancelled watcher.
+
+    No-op when there is no live watcher; swallows any error (a failed
+    pre-cancel only risks one stale notification, never a crashed tool call).
+    """
+    try:
+        old = _watcher_by_thread.get(task_id)
+        if old is not None and not old.done():
+            old.cancel()
+    except Exception:
+        logger.warning(
+            "Pre-cancel of stale watcher for task %s failed; a stale success "
+            "notification may be enqueued",
+            task_id,
+            exc_info=True,
+        )
+
+
 def has_pending_notifications(current_thread_id: str | None = None) -> bool:
     """Cheap predicate for poller idle paths — true iff there's anything to consume.
 

EvoScientist/cli/channel.py

@@ -12,6 +12,7 @@ for the main thread to set a response via ``_set_channel_response()``.
 from __future__ import annotations
 
 import asyncio
+import concurrent.futures
 import logging
 import queue
 import threading
@@ -24,11 +25,25 @@ from typing import TYPE_CHECKING, Any
 from rich.panel import Panel
 from rich.text import Text
 
+from ..channels.capabilities import ChannelCapabilities
+from ..channels.interaction import (
+    ASK_USER_TIMEOUT,
+    HITL_APPROVAL_TIMEOUT,
+    UNRECOGNIZED_FEEDBACK,
+    ApprovalPolicy,
+    InteractionIO,
+    PendingReplyRegistry,
+    is_slash_command,
+    is_stop_command,
+    resolve_approval,
+    resolve_ask_user,
+)
 from ..commands.base import ChannelRuntime
 from ..stream.console import console
 
 if TYPE_CHECKING:
     from ..gateway import GraphGateway
+    from ..runtime import AsyncRuntime
 
 _channel_logger = logging.getLogger(__name__)
 
@@ -60,6 +75,9 @@ _message_queue: queue.Queue[ChannelMessage] = queue.Queue()
 # Pending responses:
 # main → bus (msg_id → {"future": Future[str], "loop": loop, "response": str|None})
 _pending_responses: dict[str, dict] = {}
+# Sentinel response: the command's output already reached the channel via the
+# command UI, so the bus consumer must not deliver a second message.
+COMMAND_OUTPUT_ALREADY_SENT = "__evosci-command-output-already-sent__"
 _response_lock = threading.Lock()
 
 _RESPONSE_TIMEOUT = 600.0
@@ -264,6 +282,7 @@ async def dispatch_channel_slash_command(
     await_agent_ready: Callable[[], Awaitable[Any]] | None = None,
     on_cmd_completed: Callable[..., Awaitable[None]] | None = None,
     channel_runtime: ChannelRuntime | None = None,
+    async_runtime: AsyncRuntime | None = None,
 ) -> bool:
     """Dispatch a slash command from a channel message.
 
@@ -313,7 +332,7 @@ async def dispatch_channel_slash_command(
         ``cli/interactive.py:1002-1030``.  Headless serve passes
         ``None`` since it cannot hot-swap its polling-loop agent.
     """
-    if not msg.content.strip().startswith("/"):
+    if not is_slash_command(msg.content):
         return False
 
     try:
@@ -330,6 +349,7 @@ async def dispatch_channel_slash_command(
             on_cmd_completed=on_cmd_completed,
             channel_runtime=channel_runtime,
             graph_gateway=graph_gateway,
+            async_runtime=async_runtime,
         )
     except Exception as exc:
         # Last-ditch safety: any uncaught exception from inside the
@@ -366,6 +386,7 @@ async def _dispatch_channel_slash_impl(
     await_agent_ready: Callable[[], Awaitable[Any]] | None,
     on_cmd_completed: Callable[..., Awaitable[None]] | None,
     channel_runtime: ChannelRuntime | None,
+    async_runtime: AsyncRuntime | None,
 ) -> bool:
     """Inner body of ``dispatch_channel_slash_command``.
 
@@ -378,10 +399,17 @@ async def _dispatch_channel_slash_impl(
     from ..commands.channel_ui import ChannelCommandUI
     from ..commands.manager import manager as cmd_manager
 
+    # The wrapper only forwards slash-prefixed content, so an unresolved
+    # parse is always an unknown command — answer instead of feeding a typo
+    # to the agent.
     parsed = cmd_manager.resolve(msg.content)
     if parsed is None:
-        # Unknown slash command — let the agent handle it (matches TUI).
-        return False
+        bad_cmd = msg.content.split(None, 1)[0]
+        _set_channel_response(
+            msg.msg_id,
+            f"Unknown command: {bad_cmd}\nType /help to see available commands.",
+        )
+        return True
     cmd, cmd_args = parsed
 
     agent_for_ctx = agent
@@ -407,6 +435,7 @@ async def _dispatch_channel_slash_impl(
         checkpointer=checkpointer,
         channel_runtime=channel_runtime,
         graph_gateway=graph_gateway,
+        async_runtime=async_runtime,
     )
 
     try:
@@ -418,8 +447,11 @@ async def _dispatch_channel_slash_impl(
 
     if cmd_executed:
         if ctx.command_error is not None:
-            details = ctx.command_error or "(no details)"
-            _set_channel_response(msg.msg_id, f"Command error: {details}")
+            if ui.sent_to_channel:
+                _set_channel_response(msg.msg_id, COMMAND_OUTPUT_ALREADY_SENT)
+            else:
+                details = ctx.command_error or "(no details)"
+                _set_channel_response(msg.msg_id, f"Command error: {details}")
             return True
 
         if on_cmd_completed is not None:
@@ -439,7 +471,12 @@ async def _dispatch_channel_slash_impl(
             f"[{msg.channel_type}: Executed command from {msg.sender}]",
             "dim",
         )
-        _set_channel_response(msg.msg_id, f"Command executed: {msg.content}")
+        if ui.sent_to_channel:
+            # The user already saw the command's own output — a second
+            # "Command executed" message is just noise.
+            _set_channel_response(msg.msg_id, COMMAND_OUTPUT_ALREADY_SENT)
+        else:
+            _set_channel_response(msg.msg_id, f"Command executed: {msg.content}")
         return True
 
     # ``cmd_manager.execute`` returned False (empty / unparseable input).
@@ -448,21 +485,99 @@ async def _dispatch_channel_slash_impl(
 
 
 # ---------------------------------------------------------------------------
-# HITL approval intercept: bus thread ⇄ main CLI thread
+# HITL / ask_user interaction bridge: bus loop ⇄ main CLI thread
 # ---------------------------------------------------------------------------
-# When the main thread needs HITL approval from a channel user, it registers
-# a pending HITL wait for (channel, chat_id).  The bus consumer checks this
-# BEFORE normal enqueue, so the next reply from that user is intercepted.
+# The interaction protocol itself (prompt formatting, reply grammar,
+# feedback, auto-approve policy) lives in ``channels.interaction``.  Here we
+# only bridge it: the whole engine coroutine runs on the bus loop via
+# ``run_coroutine_threadsafe`` while the calling (main / TUI) thread blocks
+# on the resulting future.  Replies are routed by a single asyncio-based
+# ``PendingReplyRegistry`` fed from the inbound interception point — the bus
+# consumer checks it BEFORE normal enqueue, so the next reply from that chat
+# is intercepted.
 
-_pending_hitl: dict[str, dict] = {}  # "channel:chat_id" -> {event, reply}
-_hitl_lock = threading.Lock()
-_hitl_auto_approve: set[str] = set()  # "channel:chat_id" keys with auto-approve
 
-_HITL_APPROVAL_TIMEOUT = 120.0  # seconds to wait for HITL approval reply
-_ASK_USER_TIMEOUT = (
-    300.0  # seconds to wait for ask_user reply (longer for thinking time)
-)
-_STOP_COMMANDS = frozenset(("/stop", "/cancel"))
+# Extra head-room on the outer ``.result()`` wait so the engine's own
+# per-flow timeout always fires first and returns a clean cancelled/None
+# instead of the bridge tearing the coroutine down mid-flight.
+_ENGINE_RESULT_SLACK = 30.0
+_ENGINE_CANCEL_SETTLE_TIMEOUT = 1.0
+# Send timeout inside the bridge IO adapter (kept per-flow-independent, as
+# the standalone consumer has no send timeout).
+_BRIDGE_SEND_TIMEOUT = 15.0
+_ASK_USER_WAITS_PER_QUESTION = 2
+_ASK_USER_SENDS_PER_QUESTION = 3
+_HITL_SENDS_PER_APPROVAL = 2
+
+# One reply registry + one approval policy for the whole bridge process,
+# both living on the bus loop (replacing the old ``_pending_hitl`` /
+# ``_hitl_lock`` / ``_hitl_auto_approve`` module globals).
+_reply_registry = PendingReplyRegistry()
+_approval_policy = ApprovalPolicy()
+
+
+class _BridgeIO(InteractionIO):
+    """:class:`InteractionIO` for the CLI bridge, running on the bus loop.
+
+    ``send`` publishes outbound (bounded by :data:`_BRIDGE_SEND_TIMEOUT`);
+    ``wait_reply`` blocks on the shared :data:`_reply_registry`.  Both run on
+    the bus loop because the engine coroutine is scheduled there via
+    ``run_coroutine_threadsafe`` — no per-message thread hop.
+    """
+
+    def __init__(
+        self,
+        bus: Any,
+        msg: ChannelMessage,
+        capabilities: ChannelCapabilities,
+        session_key: str,
+    ) -> None:
+        self._bus = bus
+        self._msg = msg
+        self.capabilities = capabilities
+        self.base_metadata = msg.metadata
+        self._session_key = session_key
+
+    async def send(self, content: str, *, metadata: dict | None = None) -> bool:
+        from ..channels.bus.events import OutboundMessage
+
+        try:
+            await asyncio.wait_for(
+                self._bus.publish_outbound(
+                    OutboundMessage(
+                        channel=self._msg.channel_type,
+                        chat_id=self._msg.chat_id,
+                        content=content,
+                        metadata=metadata
+                        if metadata is not None
+                        else self._msg.metadata or {},
+                    )
+                ),
+                timeout=_BRIDGE_SEND_TIMEOUT,
+            )
+            return True
+        except Exception as exc:
+            _channel_logger.debug("bridge send failed: %s", exc)
+            return False
+
+    async def wait_reply(self, *, timeout: float) -> str | None:
+        return await _reply_registry.wait(self._session_key, timeout)
+
+
+def _ask_user_result_timeout(question_count: int) -> float:
+    per_question = (
+        ASK_USER_TIMEOUT * _ASK_USER_WAITS_PER_QUESTION
+        + _BRIDGE_SEND_TIMEOUT * _ASK_USER_SENDS_PER_QUESTION
+    )
+    return per_question * question_count + _ENGINE_RESULT_SLACK
+
+
+def _hitl_result_timeout() -> float:
+    return (
+        HITL_APPROVAL_TIMEOUT
+        + _BRIDGE_SEND_TIMEOUT * _HITL_SENDS_PER_APPROVAL
+        + _ENGINE_RESULT_SLACK
+    )
 
 
 # ---------------------------------------------------------------------------
@@ -596,265 +711,132 @@ def publish_to_channel_origin(thread_id: str | None, content: str) -> bool:
     return True
 
 
-def _is_stop_command(content: str | None) -> bool:
-    """Whether incoming content is a stop/cancel slash command."""
-    return (content or "").strip().lower() in _STOP_COMMANDS
+def _run_engine_on_bus(coro, *, result_timeout: float, on_error):
+    """Run *coro* (an engine coroutine) on the bus loop and block for it.
 
+    Schedules the coroutine on ``_bus_loop`` via ``run_coroutine_threadsafe``
+    and waits up to *result_timeout* seconds for it (the outer bound is the
+    engine's own per-flow timeout plus slack, so the engine's timeout fires
+    first). Returns *on_error* (a zero-arg factory) on any failure.
+    """
+    bus_loop = _bus_loop
+    if bus_loop is None:
+        coro.close()
+        return on_error()
+    try:
+        fut = asyncio.run_coroutine_threadsafe(coro, bus_loop)
+    except Exception as exc:
+        coro.close()
+        _channel_logger.debug("interaction engine bridge failed: %s", exc)
+        return on_error()
 
-def _register_hitl_wait(channel_type: str, chat_id: str) -> threading.Event:
-    """Register a pending HITL wait.  Returns a threading.Event to block on."""
-    key = f"{channel_type}:{chat_id}"
-    event = threading.Event()
-    with _hitl_lock:
-        _pending_hitl[key] = {"event": event, "reply": None}
-    return event
-
-
-def _pop_hitl_reply(channel_type: str, chat_id: str) -> str | None:
-    """Pop and return the HITL reply (or None if not set)."""
-    key = f"{channel_type}:{chat_id}"
-    with _hitl_lock:
-        slot = _pending_hitl.pop(key, None)
-    return slot["reply"] if slot else None
-
-
-def _try_set_hitl_reply(channel_type: str, chat_id: str, content: str) -> bool:
-    """Try to intercept a message as a HITL reply.  Returns True if consumed."""
-    key = f"{channel_type}:{chat_id}"
-    with _hitl_lock:
-        slot = _pending_hitl.get(key)
-        if slot:
-            slot["reply"] = content
-            slot["event"].set()
-            return True
-    return False
+    try:
+        return fut.result(timeout=result_timeout)
+    except concurrent.futures.TimeoutError as exc:
+        fut.cancel()
+        try:
+            asyncio.run_coroutine_threadsafe(asyncio.sleep(0), bus_loop).result(
+                timeout=_ENGINE_CANCEL_SETTLE_TIMEOUT
+            )
+        except concurrent.futures.TimeoutError:
+            _channel_logger.debug("interaction engine cancellation did not settle")
+        except Exception as settle_exc:
+            _channel_logger.debug(
+                "interaction engine failed while settling cancellation: %s",
+                settle_exc,
+            )
+        _channel_logger.debug("interaction engine bridge timed out: %s", exc)
+        return on_error()
+    except Exception as exc:
+        _channel_logger.debug("interaction engine bridge failed: %s", exc)
+        return on_error()
 
 
 def channel_ask_user_prompt(
     ask_user_data: dict,
     msg: ChannelMessage | None = None,
 ) -> dict:
-    """Format ask_user questions and collect answers from a channel user.
+    """Collect answers to ask_user questions from a channel user.
 
-    If *msg* is provided, sends questions via the bus and waits for a reply.
-    Otherwise falls back to returning a cancelled result.
+    Thin bridge: runs :func:`channels.interaction.resolve_ask_user` on the
+    bus loop over a :class:`_BridgeIO` and blocks for the result.  Signature
+    and return shape are unchanged (callers in ``interactive.py`` /
+    ``commands.py`` / ``tui_interactive.py`` are untouched).
 
-    Returns:
-        ``{"answers": [...], "status": "answered"}`` or
-        ``{"status": "cancelled"}``.
+    Returns ``{"answers": [...], "status": "answered"}`` or
+    ``{"status": "cancelled"}``.
     """
-    from ..channels.bus.events import OutboundMessage
-
     questions = ask_user_data.get("questions", [])
     if not questions:
         return {"answers": [], "status": "answered"}
-
-    if msg is None or not msg.bus_ref:
+    if msg is None or not msg.bus_ref or _bus_loop is None:
         return {"status": "cancelled"}
 
-    bus_loop = _bus_loop
-    if not bus_loop:
-        return {"status": "cancelled"}
-
-    def _send(content: str) -> bool:
-        try:
-            asyncio.run_coroutine_threadsafe(
-                msg.bus_ref.publish_outbound(
-                    OutboundMessage(
-                        channel=msg.channel_type,
-                        chat_id=msg.chat_id,
-                        content=content,
-                        metadata=msg.metadata or {},
-                    )
-                ),
-                bus_loop,
-            ).result(timeout=15)
-            return True
-        except Exception as exc:
-            _channel_logger.debug("ask_user send failed: %s", exc)
-            return False
-
-    # Ask one question at a time (consistent with Rich CLI / TUI)
-    total = len(questions)
-    answers: list[str] = []
-
-    for i, q in enumerate(questions):
-        q_text = q.get("question", "")
-        q_type = q.get("type", "text")
-        required = q.get("required", True)
-
-        # Format single question
-        if total == 1:
-            header = "\u2753 Quick check-in from EvoScientist\n"
-        else:
-            header = f"\u2753 Question {i + 1}/{total}\n"
-
-        lines = [header, f"{i + 1}. {q_text}"]
-        if not required:
-            lines[-1] += " (optional)"
-
-        if q_type == "multiple_choice":
-            choices = q.get("choices", [])
-            for j, choice in enumerate(choices):
-                label = choice.get("value", str(choice))
-                letter = chr(ord("A") + j)
-                lines.append(f"   {letter}. {label}")
-            other_letter = chr(ord("A") + len(choices))
-            lines.append(f"   {other_letter}. Other")
-            lines.append(
-                f"\nReply with a letter ({'/'.join(chr(ord('A') + k) for k in range(len(choices) + 1))}), or 'cancel'."
-            )
-        else:
-            skip_hint = " Leave empty to skip." if not required else ""
-            lines.append(f"\nReply with your answer, or 'cancel'.{skip_hint}")
-
-        if not _send("\n".join(lines)):
-            return {"status": "cancelled"}
-
-        # Wait for reply
-        hitl_event = _register_hitl_wait(msg.channel_type, msg.chat_id)
-        replied = hitl_event.wait(timeout=_ASK_USER_TIMEOUT)
-        reply_text = _pop_hitl_reply(msg.channel_type, msg.chat_id)
-
-        if not replied or not reply_text:
-            _send("\u23f0 Response timed out.")
-            return {"status": "cancelled"}
-
-        raw = reply_text.strip()
-        if _is_stop_command(raw):
-            return {"status": "cancelled"}
-        if raw.lower() == "cancel":
-            return {"status": "cancelled"}
-
-        # Parse answer
-        if q_type == "multiple_choice":
-            choices = q.get("choices", [])
-            other_letter = chr(ord("A") + len(choices))
-            if len(raw) == 1 and raw.upper() == other_letter:
-                # Other selected — ask for free-form input
-                if not _send("Please type your answer:"):
-                    return {"status": "cancelled"}
-                hitl_event = _register_hitl_wait(msg.channel_type, msg.chat_id)
-                replied = hitl_event.wait(timeout=_ASK_USER_TIMEOUT)
-                other_text = _pop_hitl_reply(msg.channel_type, msg.chat_id)
-                if not replied or not other_text:
-                    _send("\u23f0 Response timed out.")
-                    return {"status": "cancelled"}
-                if _is_stop_command(other_text):
-                    return {"status": "cancelled"}
-                if other_text.strip().lower() == "cancel":
-                    return {"status": "cancelled"}
-                answers.append(other_text.strip())
-            elif len(raw) == 1 and raw.upper().isalpha():
-                idx = ord(raw.upper()) - ord("A")
-                if 0 <= idx < len(choices):
-                    answers.append(choices[idx].get("value", raw))
-                else:
-                    answers.append(raw)
-            else:
-                answers.append(raw)
-        else:
-            answers.append(raw)
-
-    return {"answers": answers, "status": "answered"}
+    # ask_user never uses buttons; a plain capability set suffices.
+    io = _BridgeIO(
+        msg.bus_ref, msg, ChannelCapabilities(), _channel_message_session_key(msg)
+    )
+    return _run_engine_on_bus(
+        resolve_ask_user(questions, io, timeout=ASK_USER_TIMEOUT),
+        result_timeout=_ask_user_result_timeout(len(questions)),
+        on_error=lambda: {"status": "cancelled"},
+    )
 
 
 def channel_hitl_prompt(
     action_requests: list,
     msg: ChannelMessage,
 ) -> list[dict] | None:
-    """Send HITL approval prompt to channel user and wait for reply.
+    """Resolve a HITL approval prompt with a channel user.
 
-    Blocking function — uses threading.Event.wait().  Safe to call from a
-    background thread (CLI channel processing or asyncio.to_thread in TUI).
+    Thin bridge: runs :func:`channels.interaction.resolve_approval` on the
+    bus loop over a :class:`_BridgeIO` and blocks for the result.  Signature
+    and return shape are unchanged (callers are untouched).  Safe to call
+    from a background thread (CLI channel processing / TUI ``to_thread``).
 
-    Returns approval decisions list on approve/auto, or None on reject/timeout.
+    Returns the approval decisions list on approve/auto, or None on
+    reject / unrecognized / timeout / stop.
     """
-    from ..channels.bus.events import OutboundMessage
-    from ..channels.consumer import (
-        _approval_prompt_metadata,
-        _format_approval_prompt,
-        _parse_approval_reply,
-    )
+    session_key = _channel_message_session_key(msg)
+    decisions = _approval_policy.auto_decision(session_key, action_requests)
+    if decisions is not None:
+        return decisions
 
-    # Check session auto-approve (set by a previous "3" reply)
-    session_key = f"{msg.channel_type}:{msg.chat_id}"
-    if session_key in _hitl_auto_approve:
-        return [{"type": "approve"} for _ in action_requests]
-
-    bus_loop = _bus_loop
-    if not (bus_loop and msg.bus_ref):
+    if not (_bus_loop and msg.bus_ref):
         _channel_logger.debug("HITL: no bus_loop or bus_ref, rejecting")
         return None
 
-    # Look up the channel instance so we can attach buttons when the channel
-    # supports `inline_buttons` (Feishu cards, QQ keyboards, …).
+    # Look up the channel instance so the engine can attach buttons when the
+    # channel supports `inline_buttons` (Feishu cards, QQ keyboards, …).
     channel_obj = (
         _manager.get_channel(msg.channel_type) if _manager is not None else None
     )
-    has_buttons = channel_obj is not None and channel_obj.capabilities.inline_buttons
-    approval_metadata = _approval_prompt_metadata(
-        msg.metadata, with_buttons=has_buttons
+    capabilities = (
+        channel_obj.capabilities if channel_obj is not None else ChannelCapabilities()
     )
+    io = _BridgeIO(msg.bus_ref, msg, capabilities, session_key)
 
-    def _send(content: str, *, metadata: dict | None = None) -> bool:
-        """Send a message to the channel user.  Returns True on success."""
-        try:
-            asyncio.run_coroutine_threadsafe(
-                msg.bus_ref.publish_outbound(
-                    OutboundMessage(
-                        channel=msg.channel_type,
-                        chat_id=msg.chat_id,
-                        content=content,
-                        metadata=metadata
-                        if metadata is not None
-                        else msg.metadata or {},
-                    )
-                ),
-                bus_loop,
-            ).result(timeout=15)
-            return True
-        except Exception as exc:
-            _channel_logger.debug("HITL send failed: %s", exc)
-            return False
+    async def _hitl_flow() -> list[dict] | None:
+        outcome = await resolve_approval(
+            action_requests,
+            io,
+            _approval_policy,
+            session_key,
+            timeout=HITL_APPROVAL_TIMEOUT,
+        )
+        if outcome.unrecognized_reply is not None:
+            # CLI-bridge policy: an unparseable reply declines with the
+            # explicit notice. Only the serve-mode consumer refeeds the
+            # text as a new turn.
+            await io.send(UNRECOGNIZED_FEEDBACK)
+            return None
+        return outcome.decisions
 
-    # 1. Send approval prompt
-    prompt_text = _format_approval_prompt(action_requests, with_buttons=has_buttons)
-    if not _send(prompt_text, metadata=approval_metadata):
-        return None
-
-    # 2. Wait for channel user's reply
-    hitl_event = _register_hitl_wait(msg.channel_type, msg.chat_id)
-    replied = hitl_event.wait(timeout=_HITL_APPROVAL_TIMEOUT)
-    reply_text = _pop_hitl_reply(msg.channel_type, msg.chat_id)
-
-    if not replied or not reply_text:
-        _send("\u23f0 Approval timed out. Action rejected.")
-        return None
-
-    if _is_stop_command(reply_text):
-        # `/stop` already got its own immediate ack from the bus fast-path.
-        # Treat it as a pure cancel signal here so we don't send a second,
-        # contradictory "Unrecognized reply" message.
-        return None
-
-    # 3. Parse decision
-    decision = _parse_approval_reply(reply_text)
-    if decision == "auto":
-        _hitl_auto_approve.add(session_key)
-        _send("\u2705 已批准（后续自动通过）")
-        return [{"type": "approve"} for _ in action_requests]
-    if decision == "approve":
-        _send("\u2705 已批准")
-        return [{"type": "approve"} for _ in action_requests]
-
-    feedback = (
-        "\u274c 已拒绝"
-        if decision == "reject"
-        else "Unrecognized reply. Action rejected."
+    return _run_engine_on_bus(
+        _hitl_flow(),
+        result_timeout=_hitl_result_timeout(),
+        on_error=lambda: None,
     )
-    _send(feedback)
-    return None
 
 
 # ---------------------------------------------------------------------------
@@ -866,6 +848,11 @@ _bus_loop: asyncio.AbstractEventLoop | None = None
 _bus_thread: threading.Thread | None = None
 
 
+def get_channel_startup_results() -> list[tuple[str, bool, str]]:
+    """Return the current channel startup snapshot without waiting."""
+    return _manager.startup_results() if _manager is not None else []
+
+
 def _channels_is_running(channel_type: str | None = None) -> bool:
     """Check whether channels are running."""
     if _manager is None:
@@ -896,7 +883,7 @@ def _channels_stop(
 
     if channel_type is None:
         # Stop everything
-        if _bus_loop and _manager:
+        if _bus_loop and _manager and not _bus_loop.is_closed():
             try:
                 future = asyncio.run_coroutine_threadsafe(
                     _manager.stop_all(),
@@ -935,7 +922,7 @@ def _start_channels_bus_mode(
     thread_id: str,
     *,
     send_thinking: bool | None = None,
-) -> None:
+) -> list[tuple[str, bool, str]]:
     """Start all channels in bus mode with MessageBus + ChannelManager.
 
     Creates a single event loop in a daemon thread running the bus,
@@ -968,6 +955,10 @@ def _start_channels_bus_mode(
             try:
                 await mgr.start_all()
             finally:
+                # ``start_all`` returns when all channel tasks terminate. This
+                # includes immediate fatal startup failures, so tear down the
+                # dispatcher and health server before closing the bus loop.
+                await mgr.stop_all()
                 consumer.cancel()
                 try:
                     await consumer
@@ -994,6 +985,8 @@ def _start_channels_bus_mode(
             break
         time.sleep(0.1)
 
+    return mgr.startup_results(timeout=2.0)
+
 
 def _add_channel_to_running_bus(
     channel_type: str,
@@ -1040,13 +1033,16 @@ async def _bus_inbound_consumer(bus, manager) -> None:
             except asyncio.CancelledError:
                 break
 
-            # /stop should preempt HITL interception so cancel works while
-            # waiting for approvals/questions.  If a HITL wait is pending,
-            # still release it so the blocking prompt can unwind immediately.
-            if _is_stop_command(msg.content):
-                if _try_set_hitl_reply(msg.channel, msg.chat_id, msg.content):
+            session_key = _channel_session_key(msg.channel, msg.chat_id)
+
+            # /stop should preempt interaction interception so cancel works
+            # while waiting for approvals/questions.  If a prompt wait is
+            # pending, still deliver /stop into it so the blocking engine
+            # unwinds immediately (it treats /stop as a clean cancel).
+            if is_stop_command(msg.content):
+                if _reply_registry.try_resolve(session_key, msg.content):
                     _channel_logger.info(
-                        f"[bus] stop request released HITL wait for "
+                        f"[bus] stop request released interaction wait for "
                         f"{msg.channel}:{msg.chat_id}"
                     )
                 _task = asyncio.create_task(_handle_bus_message(bus, manager, msg))
@@ -1054,10 +1050,12 @@ async def _bus_inbound_consumer(bus, manager) -> None:
                 _task.add_done_callback(_tasks.discard)
                 continue
 
-            # Check if this message is a HITL approval reply
-            if _try_set_hitl_reply(msg.channel, msg.chat_id, msg.content):
+            # Reply interception sits ahead of normal enqueue — if a prompt
+            # is waiting on this chat, the next message is its reply and
+            # must NOT be enqueued as a fresh agent turn.
+            if _reply_registry.try_resolve(session_key, msg.content):
                 _channel_logger.info(
-                    f"[bus] HITL reply from {msg.channel}:{msg.sender_id}: "
+                    f"[bus] interaction reply from {msg.channel}:{msg.sender_id}: "
                     f"{msg.content[:60]}"
                 )
                 continue
@@ -1085,7 +1083,7 @@ async def _handle_bus_message(bus, manager, msg) -> None:
     # Fast-path: /stop intercept. Handle on the bus task itself so we
     # don't deadlock behind the main-thread stream we're trying to
     # interrupt. No typing indicator, no queue entry.
-    if _is_stop_command(msg.content):
+    if is_stop_command(msg.content):
         cancelled_count, active_count = _cancel_channel_session(
             msg.channel, msg.chat_id
         )
@@ -1187,16 +1185,21 @@ async def _handle_bus_message(bus, manager, msg) -> None:
                 return
 
         response = _pop_channel_response(cm.msg_id) or "No response"
-        await bus.publish_outbound(
-            OutboundMessage(
-                channel=msg.channel,
-                chat_id=msg.chat_id,
-                content=response,
-                reply_to=msg.message_id or None,
-                metadata=msg.metadata,
+        if response != COMMAND_OUTPUT_ALREADY_SENT:
+            await bus.publish_outbound(
+                OutboundMessage(
+                    channel=msg.channel,
+                    chat_id=msg.chat_id,
+                    content=response,
+                    reply_to=msg.message_id or None,
+                    metadata=msg.metadata,
+                )
             )
-        )
-        manager.record_message(msg.channel, "sent")
+            manager.record_message(msg.channel, "sent")
+        else:
+            # The command UI published its own response before returning the
+            # sentinel, so account for that delivery without sending an ack.
+            manager.record_message(msg.channel, "sent")
     except asyncio.CancelledError:
         _pop_channel_response(cm.msg_id, cancel_pending=True)
         if _channel_request_state(cm.msg_id) != "active":
@@ -1245,7 +1248,7 @@ def _auto_start_channel(
     *,
     send_thinking: bool | None = None,
     runtime: ChannelRuntime | None = None,
-) -> None:
+) -> list[tuple[str, bool, str]]:
     """Start channels automatically from config (bus mode).
 
     Args:
@@ -1257,18 +1260,22 @@ def _auto_start_channel(
             is accepted for callers that don't yet pass one.
     """
     if not config.channel_enabled:
-        return
+        return []
 
-    _start_channels_bus_mode(
+    results = _start_channels_bus_mode(
         config,
         agent,
         thread_id,
         send_thinking=send_thinking,
     )
-    # Bind only after startup succeeds; a failure above must not leave
-    # a stale runtime binding pointing at channels that never started.
-    if runtime is not None:
+    # A channel that is still starting may connect later and needs the runtime
+    # binding. Immediate failures must not leave a stale binding behind.
+    from ..channels.channel_manager import CHANNEL_STARTUP_PENDING_DETAIL
+
+    has_active_channel = any(
+        ok or detail == CHANNEL_STARTUP_PENDING_DETAIL for _, ok, detail in results
+    )
+    if runtime is not None and has_active_channel:
         runtime.bind(agent, thread_id)
-    types = [t.strip() for t in config.channel_enabled.split(",") if t.strip()]
-    results = [(ct, True, "connected (bus)") for ct in types]
     _print_channel_panel(results)
+    return results

EvoScientist/cli/channel_sends.py

@@ -0,0 +1,112 @@
+"""Non-blocking bridge for streaming callbacks sent through a channel loop."""
+
+from __future__ import annotations
+
+import asyncio
+import concurrent.futures
+import logging
+import threading
+from collections.abc import Coroutine
+from typing import Any
+
+
+class PendingChannelSends:
+    """Schedule channel I/O without blocking the caller's event loop.
+
+    Streaming callbacks run on the owned async runtime, while channel clients
+    belong to the channel bus loop.  Submissions therefore only enqueue work;
+    the frontend settles the returned futures after streaming has unwound.
+    """
+
+    def __init__(
+        self,
+        loop: asyncio.AbstractEventLoop | None,
+        logger: logging.Logger,
+    ) -> None:
+        self._loop = loop
+        self._logger = logger
+        self._lock = threading.Lock()
+        self._pending: list[tuple[concurrent.futures.Future[Any], str, int]] = []
+        self._tail: concurrent.futures.Future[Any] | None = None
+
+    @staticmethod
+    def _close(coro: Coroutine[Any, Any, Any]) -> None:
+        coro.close()
+
+    async def _run_after(
+        self,
+        predecessor: concurrent.futures.Future[Any] | None,
+        coro: Coroutine[Any, Any, Any],
+    ) -> Any:
+        if predecessor is not None:
+            try:
+                await asyncio.shield(asyncio.wrap_future(predecessor))
+            except asyncio.CancelledError:
+                task = asyncio.current_task()
+                if task is not None and task.cancelling():
+                    self._close(coro)
+                    raise
+            except Exception:
+                pass
+        return await coro
+
+    def submit(
+        self,
+        coro: Coroutine[Any, Any, Any],
+        label: str,
+        timeout: int = 15,
+    ) -> None:
+        """Schedule one send and return immediately."""
+        if self._loop is None:
+            self._close(coro)
+            return
+        with self._lock:
+            ordered_coro = self._run_after(self._tail, coro)
+            try:
+                future = asyncio.run_coroutine_threadsafe(ordered_coro, self._loop)
+            except Exception as exc:
+                self._close(ordered_coro)
+                self._close(coro)
+                self._logger.debug("%s send failed: %s", label, exc)
+                return
+            self._tail = future
+            self._pending.append((future, label, timeout))
+
+    def _take_pending(
+        self,
+    ) -> list[tuple[concurrent.futures.Future[Any], str, int]]:
+        with self._lock:
+            pending = self._pending
+            self._pending = []
+        return pending
+
+    def settle(self) -> None:
+        """Wait for scheduled sends from a synchronous frontend thread."""
+        for future, label, timeout in self._take_pending():
+            try:
+                future.result(timeout=timeout)
+            except Exception as exc:
+                future.cancel()
+                self._logger.debug("%s send failed: %s", label, exc)
+
+    async def settle_async(self) -> None:
+        """Wait for scheduled sends without blocking the frontend loop."""
+        pending = self._take_pending()
+        try:
+            for future, label, timeout in pending:
+                try:
+                    await asyncio.wait_for(asyncio.wrap_future(future), timeout=timeout)
+                except TimeoutError as exc:
+                    future.cancel()
+                    self._logger.debug("%s send failed: %s", label, exc)
+                except asyncio.CancelledError as exc:
+                    task = asyncio.current_task()
+                    if task is not None and task.cancelling():
+                        raise
+                    self._logger.debug("%s send failed: %s", label, exc)
+                except Exception as exc:
+                    self._logger.debug("%s send failed: %s", label, exc)
+        except asyncio.CancelledError:
+            for future, _label, _timeout in pending:
+                future.cancel()
+            raise

EvoScientist/cli/commands.py

@@ -1,6 +1,5 @@
 """Typer command registrations — onboard, config, mcp, main callback."""
 
-import asyncio
 import logging
 import os
 import queue
@@ -12,22 +11,30 @@ from importlib.metadata import version as _pkg_version
 from pathlib import Path
 from typing import TYPE_CHECKING, Annotated, Any, cast
 
+import click
 import typer
 from rich.markup import escape
 from rich.table import Table
 
-from ..commands.base import ChannelRuntime, Command, CommandContext
+from ..commands.base import (
+    ChannelRuntime,
+    Command,
+    CommandContext,
+    active_teams_configurable_extra,
+)
 from ..gateway import (
     GraphGateway,
     GraphTarget,
     RunRequest,
-    RuntimeGateways,
-    create_runtime_gateways,
 )
 from ..llm.context_window import DEFAULT_CONTEXT_WINDOW_FALLBACK, resolve_context_window
 from ..paths import ensure_dirs, set_active_workspace, set_workspace_root
+from ..runtime import AsyncRuntime
 from ..stream.console import console
-from . import async_notifier
+from . import (
+    async_notifier,
+    server_cmd,  # noqa: F401 — registers `EvoSci server` commands
+)
 from ._app import app, channel_app, config_app, configure_app, mcp_app, sessions_app
 from ._constants import build_metadata
 from .agent import (
@@ -53,6 +60,7 @@ from .channel import (
     publish_to_channel_origin,
     remember_channel_origin,
 )
+from .channel_sends import PendingChannelSends
 from .mcp_ui import (
     _mcp_add_server_from_kwargs,
     _mcp_edit_server_fields,
@@ -65,6 +73,36 @@ if TYPE_CHECKING:
     from langgraph.graph.state import CompiledStateGraph
 
     from ..config import EvoScientistConfig
+    from ..gateway import RuntimeGateways
+
+
+_ASYNC_RUNTIME_META_KEY = "evoscientist.async_runtime"
+
+
+def _close_cli_async_runtime(runtime: AsyncRuntime) -> None:
+    """Close the owned runtime or surface a controlled CLI shutdown failure."""
+    try:
+        runtime.close()
+    except TimeoutError as exc:
+        click.echo(
+            f"Error: Async runtime shutdown did not complete: {exc}",
+            err=True,
+        )
+        raise click.exceptions.Exit(1) from None
+
+
+def _get_cli_async_runtime(ctx: typer.Context) -> AsyncRuntime:
+    """Return the application-scoped runtime owned by this CLI invocation."""
+    root = ctx.find_root()
+    runtime = root.meta.get(_ASYNC_RUNTIME_META_KEY)
+    if runtime is None:
+        runtime = AsyncRuntime()
+        root.meta[_ASYNC_RUNTIME_META_KEY] = runtime
+        root.call_on_close(lambda: _close_cli_async_runtime(runtime))
+    if not isinstance(runtime, AsyncRuntime):  # pragma: no cover - defensive
+        raise RuntimeError("CLI async runtime context is invalid")
+    return runtime
+
 
 # =============================================================================
 # Onboard command
@@ -73,6 +111,7 @@ if TYPE_CHECKING:
 
 @app.command()
 def onboard(
+    ctx: typer.Context,
     skip_validation: bool = typer.Option(
         False, "--skip-validation", help="Skip API key validation during setup"
     ),
@@ -201,7 +240,11 @@ def onboard(
             strict=non_interactive,
         )
 
-    _run_onboard_cli(skip_validation=skip_validation, prompter=prompter)
+    _run_onboard_cli(
+        skip_validation=skip_validation,
+        prompter=prompter,
+        runtime=_get_cli_async_runtime(ctx),
+    )
 
 
 # =============================================================================
@@ -243,11 +286,21 @@ def _run_onboard_cli(**kwargs: Any) -> None:
         raise typer.Exit(code=1) from exc
 
 
-def _configure_section(section: str, skip_validation: bool = False) -> None:
+def _configure_section(
+    section: str,
+    skip_validation: bool = False,
+    *,
+    runtime: AsyncRuntime | None = None,
+) -> None:
     """Run a single onboarding section, reusing the wizard's step logic."""
+    kwargs: dict[str, Any] = {
+        "skip_validation": skip_validation,
+        "only_sections": {section},
+    }
+    if runtime is not None:
+        kwargs["runtime"] = runtime
     _run_onboard_cli(
-        skip_validation=skip_validation,
-        only_sections={section},
+        **kwargs,
     )
 
 
@@ -325,9 +378,9 @@ def configure_latex():
 
 
 @configure_app.command("channels")
-def configure_channels():
+def configure_channels(ctx: typer.Context):
     """Re-run channels selection and per-channel configuration."""
-    _configure_section("channels")
+    _configure_section("channels", runtime=_get_cli_async_runtime(ctx))
 
 
 # =============================================================================
@@ -336,24 +389,17 @@ def configure_channels():
 
 
 @channel_app.command("setup")
-def channel_setup():
+def channel_setup(ctx: typer.Context):
     """Interactive channel configuration wizard.
 
     Guides you through selecting and configuring messaging channels
     (Telegram, Discord, or iMessage).
     """
-    import asyncio
-
-    try:
-        asyncio.get_event_loop()
-    except RuntimeError:
-        asyncio.set_event_loop(asyncio.new_event_loop())
-
     from ..config import load_config, save_config
     from ..config.onboard.channels import _step_channels
 
     config = load_config()
-    updates = _step_channels(config)
+    updates = _step_channels(config, runtime=_get_cli_async_runtime(ctx))
     if updates:
         for key, value in updates.items():
             setattr(config, key, value)
@@ -464,7 +510,13 @@ def _ensure_async_subagent_server(config: Any, *, workspace_dir: str) -> None:
     state would route async sub-agent calls to a process pinned to /A
     while the main agent runs in /B.
     """
-    from ..langgraph_dev.manager import WorkspaceMismatchError, ensure_langgraph_dev
+    from ..langgraph_dev.manager import (
+        _DEFAULT_HOST,
+        WorkspaceMismatchError,
+        _is_loopback_host,
+        ensure_langgraph_dev,
+        is_async_subagents_available,
+    )
 
     try:
         with console.status(
@@ -477,6 +529,32 @@ def _ensure_async_subagent_server(config: Any, *, workspace_dir: str) -> None:
         console.print(f"[red]{exc}[/red]")
         raise typer.Exit(1) from exc
 
+    from ..langgraph_dev import manager as _lg_manager
+
+    if _lg_manager.CONFIG_DRIFT_SINCE_LAUNCH:
+        console.print(
+            "[yellow]⚠ Config changed since the background agent server was "
+            "launched — async sub-agents still use the old settings. Apply "
+            "them with [bold]EvoSci server stop[/bold], then restart "
+            "EvoSci.[/yellow]"
+        )
+
+    # The backend is shared by every UI mode, so the exposure warning lives
+    # here, not just in deploy/WebUI. Gated on the server being up: warning
+    # about a bind that never happened would be worse than saying nothing.
+    bind_host = str(getattr(config, "langgraph_dev_host", _DEFAULT_HOST) or "").strip()
+    if (
+        bind_host
+        and not _is_loopback_host(bind_host)
+        and is_async_subagents_available()
+    ):
+        console.print(
+            "[bold white on red] ⚠ PUBLIC BIND [/bold white on red] "
+            f"[bold red]Agent server listening on {bind_host} — no auth, and "
+            f"the agent can run shell. Use --host 127.0.0.1 on untrusted "
+            f"networks.[/bold red]"
+        )
+
 
 def _reconcile_autoskill_schedule(config: Any, *, workspace_dir: str) -> None:
     """Best-effort reconciliation for EvoMemory's hidden AutoSkills cron."""
@@ -672,9 +750,6 @@ async def compact_conversation(
     Returns a structured ``CompactResult``.
     """
     from langchain_core.messages.utils import count_tokens_approximately
-    from langchain_core.runnables import RunnableConfig
-
-    config: RunnableConfig = {"configurable": {"thread_id": thread_id}}
 
     try:
         state_values = await graph_gateway.get_state_values(target, thread_id)
@@ -778,22 +853,18 @@ async def compact_conversation(
     # Generate summary (LLM call)
     summary = await middleware._acreate_summary(to_summarize)
 
-    # Inject thread_id into LangGraph contextvar so _get_thread_id() finds it
-    # (compact runs outside a runnable context, so get_config() would fail
-    # and the middleware would generate a random "session_xxx" filename instead
-    # of reusing the real thread_id).
-    from langgraph.config import var_child_runnable_config
-
-    _token = var_child_runnable_config.set(config)
+    # Reuse the persisted _summarization_session_id (or generate one) so
+    # history keeps appending to a single file; re-persisted below.
+    session_id = middleware._get_session_id(state_values)
 
     # Offload old messages to backend
     file_path: str | None = None
     try:
-        file_path = await middleware._aoffload_to_backend(backend, to_summarize)
+        file_path = await middleware._aoffload_to_backend(
+            backend, to_summarize, session_id
+        )
     except Exception:
         pass  # non-fatal — proceed without offloaded history
-    finally:
-        var_child_runnable_config.reset(_token)
 
     from langchain_core.messages import HumanMessage
 
@@ -839,7 +910,7 @@ async def compact_conversation(
     await graph_gateway.update_state_values(
         target,
         thread_id,
-        {"_summarization_event": new_event},
+        {"_summarization_event": new_event, "_summarization_session_id": session_id},
     )
 
     return CompactResult(
@@ -874,7 +945,8 @@ class ServeRuntimeState:
     thread_id: str
     workspace_dir: str | None
     config: "EvoScientistConfig | None"
-    runtime_gateways: RuntimeGateways
+    runtime_gateways: "RuntimeGateways"
+    async_runtime: AsyncRuntime
     resume_warning_thread_id: str | None = None
 
     def set_agent(
@@ -969,6 +1041,7 @@ async def _apply_serve_resume_state(
                 _load_agent,
                 workspace_dir=new_workspace,
                 config=effective_config,
+                runtime=runtime_state.async_runtime,
             )
             await _sync_background_agent_server_workspace(
                 effective_config,
@@ -1119,8 +1192,6 @@ def _serve_process_message(
     via the ``on_cmd_completed`` hook because the command mutates
     ``ctx.thread_id`` / ``ctx.workspace_dir`` directly.
     """
-    import asyncio
-
     from .channel import _bus_loop
     from .tui_runtime import run_streaming
 
@@ -1139,14 +1210,10 @@ def _serve_process_message(
 
     # -- channel callback helpers (same pattern as interactive.py) --
 
+    pending_channel_sends = PendingChannelSends(_bus_loop, _serve_logger)
+
     def _send_to_channel(coro, label: str, timeout: int = 15) -> None:
-        loop = _bus_loop
-        if not loop:
-            return
-        try:
-            asyncio.run_coroutine_threadsafe(coro, loop).result(timeout=timeout)
-        except Exception as e:
-            _serve_logger.debug(f"{label} send failed: {e}")
+        pending_channel_sends.submit(coro, label, timeout)
 
     def _send_thinking(thinking: str) -> None:
         ch = msg.channel_ref
@@ -1196,31 +1263,15 @@ def _serve_process_message(
     # commands like ``/evoskills`` actually execute in serve mode instead
     # of being fed to the LLM as a plain prompt.  ``await_agent_ready`` is
     # None because the agent is always loaded before the serve loop polls.
-    # Uses a dedicated event loop (not ``asyncio.run``) so SIGINT handling
-    # installed by ``serve()`` remains authoritative — ``asyncio.run``
-    # swaps ``signal.set_wakeup_fd`` and can leave it dangling on edge
-    # cases, which breaks Ctrl+C between messages.
-    # ``set_event_loop`` is needed because some downstream commands
-    # (e.g. ``/install-mcp``) call ``asyncio.get_event_loop()``, which
-    # raises ``RuntimeError`` on Python 3.12+ when the thread has no
-    # current loop set.  The prior loop (often ``None``) is restored in
-    # the ``finally`` below so subsequent messages start from a clean
-    # slate.  Loop creation lives inside the try so an exception between
-    # creation and ``set_event_loop`` still closes the loop.
+    # Slash commands run on the application-owned runtime. The main thread
+    # remains the signal owner while command coroutines share one stable loop.
     try:
-        _prev_loop: asyncio.AbstractEventLoop | None
-        try:
-            _prev_loop = asyncio.get_event_loop_policy().get_event_loop()
-        except RuntimeError:
-            _prev_loop = None
-        _slash_loop: asyncio.AbstractEventLoop | None = None
         _slash_handled = False
         _slash_error: Exception | None = None
         try:
-            _slash_loop = asyncio.new_event_loop()
-            asyncio.set_event_loop(_slash_loop)
-            _slash_handled = _slash_loop.run_until_complete(
-                dispatch_channel_slash_command(
+            async_runtime = runtime_state.async_runtime
+            _slash_handled = async_runtime.run_sync(
+                lambda: dispatch_channel_slash_command(
                     msg,
                     agent=runtime_state.agent,
                     thread_id=runtime_state.thread_id,
@@ -1245,15 +1296,12 @@ def _serve_process_message(
                     ),
                     channel_runtime=channel_runtime,
                     graph_gateway=runtime_gateways.graph_gateway,
+                    async_runtime=async_runtime,
                 )
             )
         except Exception as exc:
             _slash_error = exc
             _serve_logger.exception("Slash dispatch failed for %s", msg.channel_type)
-        finally:
-            if _slash_loop is not None:
-                _slash_loop.close()
-            asyncio.set_event_loop(_prev_loop)
 
         if _slash_error is not None:
             _set_channel_response(msg.msg_id, f"Command error: {_slash_error}")
@@ -1280,6 +1328,7 @@ def _serve_process_message(
                 show_thinking=show_thinking,
                 interactive=True,
                 metadata=meta,
+                configurable_extra=active_teams_configurable_extra(channel_runtime),
                 on_thinking=_send_thinking,
                 on_todo=_send_todo,
                 on_file_write=_send_media,
@@ -1287,11 +1336,13 @@ def _serve_process_message(
                 ask_user_prompt_fn=_ask_user_prompt,
                 cancel_scope=_channel_message_cancel_scope(msg),
                 gateway=runtime_gateways.graph_gateway,
+                runtime=runtime_state.async_runtime,
             )
         except Exception as e:
             response = f"Error: {e}"
             console.print(f"[red]Serve error: {e}[/red]")
 
+        pending_channel_sends.settle()
         _set_channel_response(msg.msg_id, response)
         console.print(f"[dim][{msg.channel_type}] Replied to {msg.sender}[/dim]")
     finally:
@@ -1309,6 +1360,7 @@ def _serve_drain_notifications(
     model: str | None,
     workspace_dir: str,
     show_thinking: bool,
+    channel_runtime: ChannelRuntime | None = None,
 ) -> None:
     """Drain the async-task notification queue in headless serve mode.
 
@@ -1340,7 +1392,9 @@ def _serve_drain_notifications(
                 show_thinking=show_thinking,
                 interactive=True,
                 metadata=meta,
+                configurable_extra=active_teams_configurable_extra(channel_runtime),
                 gateway=runtime_state.runtime_gateways.graph_gateway,
+                runtime=runtime_state.async_runtime,
             )
         except Exception as exc:
             _serve_logger.warning("Notification agent turn failed: %s", exc)
@@ -1378,25 +1432,28 @@ def _serve_drain_notifications(
             current_thread_id=runtime_state.thread_id,
         )
 
-    _notif_loop: _aio.AbstractEventLoop | None = None
     try:
-        _notif_loop = _aio.new_event_loop()
-        _notif_loop.run_until_complete(_consume())
+        runtime_state.async_runtime.run_sync(_consume)
     except Exception as exc:
         _serve_logger.warning("Notification drain failed: %s", exc)
-    finally:
-        if _notif_loop is not None:
-            _notif_loop.close()
 
 
 @app.command()
 def serve(
+    ctx: typer.Context,
     no_thinking: bool = typer.Option(
         False, "--no-thinking", help="Disable thinking relay to channels"
     ),
     workdir: str | None = typer.Option(
         None, "--workdir", help="Override workspace directory"
     ),
+    host: str | None = typer.Option(
+        None,
+        "--host",
+        help="Interface to bind the langgraph dev backend to (default: "
+        "langgraph_dev_host = 127.0.0.1). Pass 0.0.0.0 to reach it from "
+        "another machine — the backend has no auth.",
+    ),
     auto_approve: bool = typer.Option(
         False,
         "--auto-approve",
@@ -1431,6 +1488,9 @@ def serve(
     from ..config import apply_config_to_env, get_effective_config
 
     cli_overrides = {}
+    # serve starts no front-end, so only the backend bind applies here.
+    if host is not None and host.strip():
+        cli_overrides["langgraph_dev_host"] = host.strip()
     if auto_approve:
         cli_overrides["auto_approve"] = True
     if auto_mode:
@@ -1445,6 +1505,7 @@ def serve(
         cli_overrides["log_level"] = "DEBUG"
         cli_overrides["channel_debug_tracing"] = True
     config = get_effective_config(cli_overrides)
+    async_runtime = _get_cli_async_runtime(ctx)
     if debug:
         os.environ["EVOSCIENTIST_LOG_LEVEL"] = "DEBUG"
         os.environ["EVOSCIENTIST_CHANNEL_DEBUG_TRACING"] = "true"
@@ -1495,11 +1556,15 @@ def serve(
             f"[bold red]{DANGEROUS_BANNER_MESSAGE}[/bold red]"
         )
     console.print("[dim]Loading agent...[/dim]")
-    agent = _load_agent(workspace_dir=ws, config=config)
+    agent = _load_agent(workspace_dir=ws, config=config, runtime=async_runtime)
+
+    from ..gateway import create_runtime_gateways
 
     runtime_gateways = create_runtime_gateways()
-    tid = asyncio.run(
-        runtime_gateways.graph_gateway.create_thread(GraphTarget(workspace_dir=ws))
+    tid = async_runtime.run_sync(
+        lambda: runtime_gateways.graph_gateway.create_thread(
+            GraphTarget(workspace_dir=ws)
+        )
     )
 
     # Mutable runtime shared with _serve_process_message so channel slash
@@ -1511,6 +1576,7 @@ def serve(
         workspace_dir=ws,
         config=config,
         runtime_gateways=runtime_gateways,
+        async_runtime=async_runtime,
     )
 
     channel_runtime = ChannelRuntime(agent=agent, thread_id=tid)
@@ -1551,9 +1617,22 @@ def serve(
     import threading
 
     shutdown_event = threading.Event()
+    no_active_cancel_scope = object()
+    active_cancel_scope: str | object | None = no_active_cancel_scope
 
     def _handle_shutdown(signum: int, _frame: Any) -> None:
         shutdown_event.set()
+        # Cancelling the owned asyncio task is not enough when it is awaiting a
+        # blocking execute call: the executor thread and its isolated process
+        # group keep running until the matching stream event is set.  Request
+        # scope cancellation before KeyboardInterrupt unwinds message cleanup
+        # (which discards that scope).  SIGTERM also needs this to unblock the
+        # synchronous serve call so the poll loop can observe shutdown_event.
+        scope = active_cancel_scope
+        if scope is not no_active_cancel_scope:
+            from ..stream.display import request_stream_cancel
+
+            request_stream_cancel(cast(str | None, scope))
         # Fall back to Python's default SIGINT behavior (raises
         # KeyboardInterrupt) so blocking I/O inside ``run_streaming``
         # is still interrupted.  For SIGTERM there's no default that
@@ -1573,6 +1652,7 @@ def serve(
             if shutdown_event.is_set():
                 break
             if msg is not None:
+                active_cancel_scope = _channel_message_cancel_scope(msg)
                 try:
                     _serve_process_message(
                         msg,
@@ -1588,15 +1668,23 @@ def serve(
                 except KeyboardInterrupt:
                     shutdown_event.set()
                     break
+                finally:
+                    active_cancel_scope = no_active_cancel_scope
 
             # Poll notification queue when idle (no channel message was pending).
             if async_notifier.has_pending_notifications(runtime_state.thread_id):
-                _serve_drain_notifications(
-                    runtime_state=runtime_state,
-                    model=config.model,
-                    workspace_dir=ws,
-                    show_thinking=effective_channel_thinking,
-                )
+                # Notification turns use the default stream cancellation scope.
+                active_cancel_scope = None
+                try:
+                    _serve_drain_notifications(
+                        runtime_state=runtime_state,
+                        model=config.model,
+                        workspace_dir=ws,
+                        show_thinking=effective_channel_thinking,
+                        channel_runtime=channel_runtime,
+                    )
+                finally:
+                    active_cancel_scope = no_active_cancel_scope
     except KeyboardInterrupt:
         shutdown_event.set()
     finally:
@@ -1954,20 +2042,16 @@ def sessions_callback(ctx: typer.Context):
     so the bare command is informative rather than silent.
     """
     if ctx.invoked_subcommand is None:
-        sessions_stats()
+        sessions_stats(ctx)
 
 
 @sessions_app.command("stats")
-def sessions_stats():
+def sessions_stats(ctx: typer.Context):
     """Show DB size, thread count, total checkpoints, top heaviest threads."""
-    import asyncio
-
     from ..sessions import db_stats
 
-    try:
-        stats = asyncio.get_event_loop().run_until_complete(db_stats())
-    except RuntimeError:
-        stats = asyncio.new_event_loop().run_until_complete(db_stats())
+    runtime = _get_cli_async_runtime(ctx)
+    stats = runtime.run_sync(db_stats)
 
     table = Table(title="EvoScientist sessions DB", show_header=True)
     table.add_column("Metric", style="cyan")
@@ -2094,6 +2178,15 @@ def _main_callback(
         "--ui",
         help="UI backend: tui (default), cli, or webui.",
     ),
+    host: str | None = typer.Option(
+        None,
+        "--host",
+        help="Interface to bind servers to (default: 127.0.0.1 for both). "
+        "Sets langgraph_dev_host — the backend shared by every UI mode — and "
+        "webui_host (WebUI mode only). Applies to the default entry; the "
+        "serve and deploy subcommands take their own --host. Pass 0.0.0.0 to "
+        "reach both from another machine (the backend has no auth).",
+    ),
     output_format: str | None = typer.Option(
         None,
         "--output-format",
@@ -2108,6 +2201,8 @@ def _main_callback(
     if ctx.invoked_subcommand is not None:
         return
 
+    async_runtime = _get_cli_async_runtime(ctx)
+
     # Load and apply configuration
     from ..config import apply_config_to_env, get_effective_config
 
@@ -2152,6 +2247,11 @@ def _main_callback(
         cli_overrides["show_thinking"] = False
     if ui:
         cli_overrides["ui_backend"] = ui
+    if host is not None and host.strip():
+        # One flag drives both servers; the backend applies in EVERY UI mode
+        # (auto-started for tui/cli/serve too), webui_host only in WebUI mode.
+        cli_overrides["webui_host"] = host.strip()
+        cli_overrides["langgraph_dev_host"] = host.strip()
     if auto_approve:
         cli_overrides["auto_approve"] = True
     if effective_auto_mode:
@@ -2319,6 +2419,7 @@ def _main_callback(
         # Single-shot mode: wrap in persistent checkpointer
         import asyncio
 
+        from ..gateway import create_runtime_gateways
         from ..sessions import get_checkpointer
         from ..stream.json_sink import stream_json
         from .interactive import _wait_for_memory_workers_before_exit, cmd_run
@@ -2350,10 +2451,12 @@ def _main_callback(
                 else:
                     tid = await graph_gateway.create_thread()
                 console.print("[dim]Loading agent...[/dim]")
-                agent = _load_agent(
+                agent = await asyncio.to_thread(
+                    _load_agent,
                     workspace_dir=workspace_dir,
                     checkpointer=checkpointer,
                     config=config,
+                    runtime=async_runtime,
                 )
                 try:
                     if effective_output_format == "stream-json":
@@ -2382,26 +2485,47 @@ def _main_callback(
                             # matching the text path (cmd_run does this itself).
                             _wait_for_memory_workers_before_exit()
                     else:
-                        cmd_run(
-                            agent,
-                            prompt,
-                            thread_id=tid,
-                            show_thinking=show_thinking,
-                            workspace_dir=workspace_dir,
-                            model=config.model,
-                            ui_backend=config.ui_backend,
-                            runtime_gateways=runtime_gateways,
+                        stream_worker = asyncio.create_task(
+                            asyncio.to_thread(
+                                cmd_run,
+                                agent,
+                                prompt,
+                                thread_id=tid,
+                                show_thinking=show_thinking,
+                                workspace_dir=workspace_dir,
+                                model=config.model,
+                                ui_backend=config.ui_backend,
+                                runtime_gateways=runtime_gateways,
+                                async_runtime=async_runtime,
+                            )
                         )
+                        try:
+                            await asyncio.shield(stream_worker)
+                        except asyncio.CancelledError:
+                            from ..stream.display import request_stream_cancel
+                            from .tui_runtime import settle_cancelled_worker
+
+                            await settle_cancelled_worker(
+                                stream_worker,
+                                on_cancel=request_stream_cancel,
+                            )
+                            raise
                 finally:
+                    # Model failures can bypass middleware ``after_agent``
+                    # hooks. Close any remaining QuickJS workers while this
+                    # event loop is still available; their synchronous GC
+                    # fallback can deadlock during interpreter shutdown.
+                    from ..middleware.code_interpreter import (
+                        aclose_code_interpreters,
+                    )
+
+                    await aclose_code_interpreters()
                     try:
                         print_resume_hint(tid, console=console)
                     except Exception:
                         pass
 
-        import nest_asyncio
-
-        nest_asyncio.apply()
-        asyncio.get_event_loop().run_until_complete(_single_shot())
+        async_runtime.run_sync(_single_shot)
     else:
         from .interactive import cmd_interactive
 
@@ -2418,6 +2542,7 @@ def _main_callback(
             thread_id=thread_id,
             ui_backend=config.ui_backend,
             config=config,
+            async_runtime=async_runtime,
         )
 
 

EvoScientist/cli/interactive.py

@@ -4,8 +4,10 @@ import asyncio
 import logging
 import queue
 import random
+import signal
 import sys
-from collections.abc import Callable
+import threading
+from collections.abc import Awaitable, Callable
 from dataclasses import dataclass
 from datetime import datetime
 from typing import TYPE_CHECKING, Any
@@ -62,6 +64,7 @@ from .channel import (
     _set_channel_response,
     dispatch_channel_slash_command,
 )
+from .channel_sends import PendingChannelSends
 from .file_mentions import complete_file_mention, resolve_file_mentions
 from .rich_command_ui import RichCLICommandUI
 from .status_bar import (
@@ -83,7 +86,12 @@ from .status_bar import (
     make_usage_status_snapshot,
 )
 from .tui_interactive import run_textual_interactive
-from .tui_runtime import resolve_ui_backend, run_streaming
+from .tui_runtime import (
+    StreamCancellationTimeout,
+    resolve_ui_backend,
+    run_streaming,
+    run_streaming_async,
+)
 
 _MEMORY_WORKER_SHUTDOWN_WAIT_SECONDS = 120.0
 _MEMORY_WORKER_SHUTDOWN_POLL_SECONDS = 0.5
@@ -97,6 +105,8 @@ _background_tasks: set[asyncio.Task] = set()
 if TYPE_CHECKING:
     from langgraph.graph.state import CompiledStateGraph
 
+    from ..runtime import AsyncRuntime
+
 
 @dataclass(frozen=True, slots=True)
 class _StartupSession:
@@ -107,6 +117,15 @@ class _StartupSession:
     resumed: bool
 
 
+async def _run_serialized_turn(
+    turn_lock: asyncio.Lock,
+    operation: Callable[[], Awaitable[Any]],
+) -> Any:
+    """Run one session turn without overlapping another frontend source."""
+    async with turn_lock:
+        return await operation()
+
+
 # =============================================================================
 # Banner
 # =============================================================================
@@ -328,6 +347,47 @@ async def _resolve_startup_session(
 # =============================================================================
 
 
+async def _run_rich_cli_streaming_turn(**kwargs: Any) -> str:
+    """Run one Rich CLI turn with a fresh, turn-local SIGINT policy.
+
+    ``asyncio.run`` installs a SIGINT handler whose interrupt count lasts for
+    the lifetime of the runner.  The Rich CLI intentionally recovers after a
+    cancelled turn, so relying on that handler makes Ctrl+C on a later turn
+    look like the runner's second interrupt and raises ``KeyboardInterrupt``.
+
+    While a model turn is active, route the first Ctrl+C to a child task
+    instead.  Restoring the runner's handler after every turn keeps Ctrl+C at
+    the prompt unchanged and resets the force-quit boundary for the next turn.
+    A second Ctrl+C before the current turn settles remains a force quit.
+    """
+    stream_task = asyncio.create_task(
+        run_streaming_async(**kwargs, recover_on_cancel=True)
+    )
+
+    # Interactive CLI execution belongs on the main thread, but retaining the
+    # ordinary await makes this helper safe in embedded/test environments where
+    # Python does not permit installing process signal handlers.
+    if threading.current_thread() is not threading.main_thread():
+        return await stream_task
+
+    previous_sigint = signal.getsignal(signal.SIGINT)
+    interrupted = False
+
+    def _cancel_turn(signum: int, frame: Any) -> None:
+        nonlocal interrupted
+        if interrupted or stream_task.done():
+            signal.default_int_handler(signum, frame)
+            return
+        interrupted = True
+        stream_task.cancel()
+
+    signal.signal(signal.SIGINT, _cancel_turn)
+    try:
+        return await stream_task
+    finally:
+        signal.signal(signal.SIGINT, previous_sigint)
+
+
 def cmd_interactive(
     show_thinking: bool = True,
     channel_send_thinking: bool = True,
@@ -340,6 +400,7 @@ def cmd_interactive(
     thread_id: str | None = None,
     ui_backend: str = "cli",
     config=None,
+    async_runtime: "AsyncRuntime | None" = None,
 ) -> None:
     """Interactive conversation mode with streaming output.
 
@@ -358,15 +419,15 @@ def cmd_interactive(
         thread_id: Optional thread ID to resume a previous session
         ui_backend: UI backend ('cli' or 'tui')
     """
-    import nest_asyncio
-
-    nest_asyncio.apply()
-
     resolved_ui_backend = resolve_ui_backend(ui_backend, warn_fallback=True)
     if resolved_ui_backend == "tui":
         from functools import partial
 
-        load_agent = partial(_load_agent, config=config)
+        load_agent = partial(
+            _load_agent,
+            config=config,
+            runtime=async_runtime,
+        )
         run_textual_interactive(
             show_thinking=show_thinking,
             channel_send_thinking=channel_send_thinking,
@@ -380,6 +441,7 @@ def cmd_interactive(
             load_agent=load_agent,
             create_session_workspace=_create_session_workspace,
             config=config,
+            async_runtime=async_runtime,
         )
         return
 
@@ -419,7 +481,7 @@ def cmd_interactive(
         width = console.size.width
         console.print(Text("\u2500" * width, style="dim"))
 
-    from ..commands.base import ChannelRuntime
+    from ..commands.base import ChannelRuntime, active_teams_configurable_extra
 
     channel_runtime = ChannelRuntime()
 
@@ -448,7 +510,17 @@ def cmd_interactive(
         on_progress=_on_mcp_progress,
     )
 
-    runtime_gateways = create_runtime_gateways()
+    # One frontend event sink for the whole session — injected into the agent's
+    # middleware (write side) and the local gateway's streaming path (read side)
+    # so both share one owner. It survives agent rebuilds (/model, /new, MCP
+    # reload) because the session, not the agent, holds it.
+    from ..stream.sink import SessionEventSink
+
+    event_sink = SessionEventSink(
+        fallback_display=lambda text, style: console.print(text, style=style)
+    )
+
+    runtime_gateways = create_runtime_gateways(events=event_sink)
     graph_gateway = runtime_gateways.graph_gateway
     requested_thread_id = thread_id
 
@@ -486,6 +558,8 @@ def cmd_interactive(
             workspace_dir=state["workspace_dir"],
             checkpointer=checkpointer,
             config=config,
+            events=event_sink,
+            runtime=async_runtime,
         )
 
     async def _await_agent_ready() -> "CompiledStateGraph":
@@ -857,6 +931,8 @@ def cmd_interactive(
 
             # ---- Channel queue processing (bus → main thread) ----
 
+            turn_lock = asyncio.Lock()
+
             async def _process_channel_message(msg: ChannelMessage) -> None:
                 """Process a single channel message with real-time streaming.
 
@@ -894,17 +970,12 @@ def cmd_interactive(
                     console.print(rx)
                     _print_separator()
 
+                    pending_channel_sends = PendingChannelSends(
+                        _ch_mod._bus_loop, _channel_logger
+                    )
+
                     def _send_to_channel(coro, label: str, timeout: int = 15) -> None:
-                        """Schedule an async channel send on the bus loop."""
-                        loop = _ch_mod._bus_loop
-                        if not loop:
-                            return
-                        try:
-                            asyncio.run_coroutine_threadsafe(coro, loop).result(
-                                timeout=timeout
-                            )
-                        except Exception as e:
-                            _channel_logger.debug(f"{label} send failed: {e}")
+                        pending_channel_sends.submit(coro, label, timeout)
 
                     def _send_thinking_to_channel(thinking: str) -> None:
                         ch = msg.channel_ref
@@ -1018,6 +1089,7 @@ def cmd_interactive(
                         on_cmd_completed=_on_channel_cmd_completed,
                         channel_runtime=channel_runtime,
                         graph_gateway=runtime_gateways.graph_gateway,
+                        async_runtime=async_runtime,
                     )
                     if _slash_handled:
                         # A channel-issued /new or /resume rotates the thread
@@ -1036,7 +1108,7 @@ def cmd_interactive(
                         await _refresh_status_snapshot(
                             msg.content, reset_streaming_text=True
                         )
-                        response = run_streaming(
+                        response = await run_streaming_async(
                             ui_backend=state["ui_backend"],
                             agent=ready_agent,
                             message=msg.content,
@@ -1044,6 +1116,9 @@ def cmd_interactive(
                             show_thinking=show_thinking,
                             interactive=True,
                             metadata=meta,
+                            configurable_extra=active_teams_configurable_extra(
+                                channel_runtime
+                            ),
                             on_thinking=_send_thinking_to_channel,
                             on_todo=_send_todo_to_channel,
                             on_file_write=_send_media_to_channel,
@@ -1053,11 +1128,13 @@ def cmd_interactive(
                             status_footer_builder=_stream_status_footer,
                             cancel_scope=_ch_mod._channel_message_cancel_scope(msg),
                             gateway=runtime_gateways.graph_gateway,
+                            runtime=async_runtime,
                         )
                     except Exception as e:
                         response = f"Error: {e}"
                         console.print(f"[red]Channel error: {e}[/red]")
 
+                    await pending_channel_sends.settle_async()
                     _set_channel_response(msg.msg_id, response)
                     await _refresh_status_snapshot(reset_streaming_text=True)
 
@@ -1094,7 +1171,7 @@ def cmd_interactive(
                 meta = build_metadata(state["workspace_dir"], model)
                 await _refresh_status_snapshot(text, reset_streaming_text=True)
                 ready_agent = await _await_agent_ready()
-                response = run_streaming(
+                response = await run_streaming_async(
                     ui_backend=state["ui_backend"],
                     agent=ready_agent,
                     message=text,
@@ -1107,9 +1184,11 @@ def cmd_interactive(
                     show_thinking=show_thinking,
                     interactive=True,
                     metadata=meta,
+                    configurable_extra=active_teams_configurable_extra(channel_runtime),
                     on_stream_event=_handle_stream_status_event,
                     status_footer_builder=_stream_status_footer,
                     gateway=runtime_gateways.graph_gateway,
+                    runtime=async_runtime,
                 )
                 _notif_tid = target_thread_id or state["thread_id"]
                 if _ch_mod.publish_to_channel_origin(_notif_tid, response):
@@ -1165,7 +1244,10 @@ def cmd_interactive(
                     except queue.Empty:
                         msg = None
                     if msg is not None:
-                        await _process_channel_message(msg)
+                        await _run_serialized_turn(
+                            turn_lock,
+                            lambda _msg=msg: _process_channel_message(_msg),
+                        )
                         continue  # check queues again immediately
 
                     # Notification path (only when no channel message was pending).
@@ -1182,8 +1264,13 @@ def cmd_interactive(
                         try:
                             await async_notifier.consume_notifications(
                                 run_message=lambda text, notifs, _tid=current_tid: (
-                                    _inject_notification_message(
-                                        text, notifs, target_thread_id=_tid
+                                    _run_serialized_turn(
+                                        turn_lock,
+                                        lambda: _inject_notification_message(
+                                            text,
+                                            notifs,
+                                            target_thread_id=_tid,
+                                        ),
                                     )
                                 ),
                                 read_async_tasks_state=read_async_tasks_state,
@@ -1318,6 +1405,7 @@ def cmd_interactive(
                                 input_tokens_hint=state.get("status_last_input_tokens"),
                                 channel_runtime=channel_runtime,
                                 graph_gateway=runtime_gateways.graph_gateway,
+                                async_runtime=async_runtime,
                             )
                             await cmd_manager.execute(user_input, ctx)
 
@@ -1400,17 +1488,26 @@ def cmd_interactive(
                         await _refresh_status_snapshot(
                             message_to_send, reset_streaming_text=True
                         )
-                        run_streaming(
-                            ui_backend=state["ui_backend"],
-                            agent=ready_agent,
-                            message=message_to_send,
-                            thread_id=state["thread_id"],
-                            show_thinking=show_thinking,
-                            interactive=True,
-                            metadata=meta,
-                            on_stream_event=_handle_stream_status_event,
-                            status_footer_builder=_stream_status_footer,
-                            gateway=runtime_gateways.graph_gateway,
+                        await _run_serialized_turn(
+                            turn_lock,
+                            lambda _agent=ready_agent, _message=message_to_send, _thread_id=state["thread_id"], _meta=meta: (
+                                _run_rich_cli_streaming_turn(
+                                    ui_backend=state["ui_backend"],
+                                    agent=_agent,
+                                    message=_message,
+                                    thread_id=_thread_id,
+                                    show_thinking=show_thinking,
+                                    interactive=True,
+                                    metadata=_meta,
+                                    configurable_extra=active_teams_configurable_extra(
+                                        channel_runtime
+                                    ),
+                                    on_stream_event=_handle_stream_status_event,
+                                    status_footer_builder=_stream_status_footer,
+                                    gateway=runtime_gateways.graph_gateway,
+                                    runtime=async_runtime,
+                                )
+                            ),
                         )
                         await _refresh_status_snapshot(reset_streaming_text=True)
                         console.print()
@@ -1425,6 +1522,14 @@ def cmd_interactive(
                         console.print()
                         state["running"] = False
                         break
+                    except StreamCancellationTimeout as e:
+                        console.print(f"[red]{escape(str(e))}[/red]")
+                        console.print(
+                            "[dim]Exiting because the active turn could not be "
+                            "stopped safely.[/dim]"
+                        )
+                        state["running"] = False
+                        break
                     except Exception as e:
                         error_msg = str(e)
                         if (
@@ -1445,6 +1550,17 @@ def cmd_interactive(
                     await queue_task
                 except asyncio.CancelledError:
                     pass
+                try:
+                    from ..middleware.code_interpreter import (
+                        aclose_code_interpreters,
+                    )
+
+                    await aclose_code_interpreters()
+                except Exception:
+                    _channel_logger.debug(
+                        "code interpreter cleanup failed",
+                        exc_info=True,
+                    )
                 # Best-effort: guard so a DB lookup failure here can't
                 # shadow the original exception exiting _async_main_loop.
                 current_tid = state.get("thread_id")
@@ -1482,6 +1598,7 @@ def cmd_run(
     ui_backend: str = "cli",
     *,
     runtime_gateways: RuntimeGateways,
+    async_runtime: "AsyncRuntime | None" = None,
 ) -> None:
     """Single-shot execution with streaming display.
 
@@ -1515,6 +1632,7 @@ def cmd_run(
             interactive=False,
             metadata=meta,
             gateway=runtime_gateways.graph_gateway,
+            runtime=async_runtime,
         )
         _wait_for_memory_workers_before_exit()
     except Exception as e:
@@ -1527,7 +1645,13 @@ def cmd_run(
             raise typer.Exit(1) from e
         else:
             console.print(f"[red]Error: {e}[/red]")
-            raise
+            # This is the process boundary for single-shot text mode.  Letting
+            # provider exceptions escape makes Typer/Rich render the complete
+            # async exception chain after we already printed a concise error;
+            # large OpenAI/httpx chains can keep the CLI busy well after the
+            # resume hint is shown.  Convert the failure to Click's controlled
+            # exit signal while preserving the cause for programmatic callers.
+            raise typer.Exit(1) from e
 
 
 def _wait_for_memory_workers_before_exit(

EvoScientist/cli/server_cmd.py

@@ -0,0 +1,99 @@
+"""``EvoSci server`` — inspect and stop the background langgraph dev server.
+
+The explicit counterpart to ``langgraph_dev_keepalive``: an opt-in server
+that outlives its CLI needs an equally explicit way to see and stop it.
+"""
+
+from __future__ import annotations
+
+import sys
+
+from ..stream.console import console
+from ._app import server_app
+
+
+@server_app.command("status")
+def server_status() -> None:
+    """Show the background langgraph dev server's state."""
+    from ..config import get_effective_config
+    from ..langgraph_dev.manager import (
+        _DEFAULT_HOST,
+        _DEFAULT_PORT,
+        _pid_serves_port,
+        _read_workspace_sidecar,
+        is_langgraph_dev_running,
+    )
+
+    config = get_effective_config()
+    port = int(getattr(config, "langgraph_dev_port", _DEFAULT_PORT))
+    host = (
+        str(getattr(config, "langgraph_dev_host", _DEFAULT_HOST) or _DEFAULT_HOST)
+    ).strip() or _DEFAULT_HOST
+    running = is_langgraph_dev_running(port=port, host=host)
+    sidecar = _read_workspace_sidecar()
+
+    if not running and sidecar is None:
+        console.print("[dim]No background langgraph dev server is running.[/dim]")
+        return
+    state = "[green]running[/green]" if running else "[red]not responding[/red]"
+    console.print(f"[bold]langgraph dev[/bold] on port {port}: {state}")
+    if sidecar is not None:
+        console.print(f"  workspace: {sidecar.get('workspace')}")
+        pid = sidecar.get("pid")
+        if _pid_serves_port(pid, port):
+            console.print(f"  pid:       {pid}")
+        else:
+            console.print(
+                f"  pid:       {pid} [yellow](stale record — this pid does "
+                f"not serve port {port})[/yellow]"
+            )
+    elif running:
+        console.print(
+            "  [yellow]no sidecar — externally managed or pre-keepalive server[/yellow]"
+        )
+
+
+@server_app.command("stop")
+def server_stop() -> None:
+    """Stop the background langgraph dev server started by EvoSci."""
+    from ..config import get_effective_config
+    from ..langgraph_dev.manager import (
+        _DEFAULT_HOST,
+        _DEFAULT_PORT,
+        is_langgraph_dev_running,
+        stop_recorded_server,
+    )
+
+    pid = stop_recorded_server()
+    if pid is not None:
+        console.print(f"[green]✓[/green] Stopped langgraph dev (pid {pid}).")
+        return
+    config = get_effective_config()
+    port = int(getattr(config, "langgraph_dev_port", _DEFAULT_PORT))
+    host = (
+        str(getattr(config, "langgraph_dev_host", _DEFAULT_HOST) or _DEFAULT_HOST)
+    ).strip() or _DEFAULT_HOST
+    if is_langgraph_dev_running(port=port, host=host):
+        # A server without ownership records (crashed session, deleted state
+        # files) can't be verified as ours — refuse to guess, hand the user
+        # the manual path instead of a silent no-op.
+        console.print(
+            f"[yellow]⚠ A langgraph dev is still serving port {port}, but "
+            f"EvoSci has no ownership record for it, so it was not "
+            f"touched.[/yellow]"
+        )
+        if sys.platform == "win32":
+            manual = (
+                f'powershell "Get-NetTCPConnection -LocalPort {port} | '
+                f'Select-Object -ExpandProperty OwningProcess | Stop-Process"'
+            )
+        else:
+            manual = f"kill $(lsof -ti :{port})"
+        console.print(
+            f"[dim]If it is yours, stop it manually: [bold]{manual}[/bold][/dim]"
+        )
+    else:
+        console.print(
+            "[dim]No EvoSci-owned langgraph dev server to stop "
+            "(stale state, if any, was cleaned up).[/dim]"
+        )

EvoScientist/cli/tui_backends.py

@@ -4,11 +4,14 @@ from __future__ import annotations
 
 from collections.abc import Callable
 from dataclasses import dataclass
-from typing import Any, Protocol
+from typing import TYPE_CHECKING, Any, Protocol
 
 from ..gateway import GraphGateway
 from ..stream.display import _run_streaming
 
+if TYPE_CHECKING:
+    from ..runtime import AsyncRuntime
+
 
 class StreamingTUIBackend(Protocol):
     """Protocol for TUI backends that can render agent streaming output."""
@@ -29,10 +32,12 @@ class StreamingTUIBackend(Protocol):
         on_stream_event: Callable[[str, Any], Any] | None = None,
         status_footer_builder: Callable[[], Any] | None = None,
         metadata: dict | None = None,
+        configurable_extra: dict[str, Any] | None = None,
         hitl_prompt_fn: Callable[[list], list[dict] | None] | None = None,
         ask_user_prompt_fn: Callable[[dict], dict] | None = None,
         cancel_scope: str | None = None,
         gateway: GraphGateway,
+        runtime: AsyncRuntime | None = None,
     ) -> str:
         """Run streaming and return final response text."""
 
@@ -57,10 +62,12 @@ class RichStreamingBackend:
         on_stream_event: Callable[[str, Any], Any] | None = None,
         status_footer_builder: Callable[[], Any] | None = None,
         metadata: dict | None = None,
+        configurable_extra: dict[str, Any] | None = None,
         hitl_prompt_fn: Callable[[list], list[dict] | None] | None = None,
         ask_user_prompt_fn: Callable[[dict], dict] | None = None,
         cancel_scope: str | None = None,
         gateway: GraphGateway,
+        runtime: AsyncRuntime | None = None,
     ) -> str:
         return _run_streaming(
             agent=agent,
@@ -74,8 +81,10 @@ class RichStreamingBackend:
             on_stream_event=on_stream_event,
             status_footer_builder=status_footer_builder,
             metadata=metadata,
+            configurable_extra=configurable_extra,
             hitl_prompt_fn=hitl_prompt_fn,
             ask_user_prompt_fn=ask_user_prompt_fn,
             cancel_scope=cancel_scope,
             gateway=gateway,
+            runtime=runtime,
         )

EvoScientist/cli/tui_interactive.py

@@ -11,6 +11,7 @@ import logging
 import queue
 import random
 import sys
+import threading
 from collections.abc import Callable
 from dataclasses import dataclass
 from datetime import datetime
@@ -53,11 +54,11 @@ from .channel import (
     ChannelMessage,
     _auto_start_channel,
     _channels_is_running,
-    _channels_running_list,
     _channels_stop,
     _message_queue,
     _set_channel_response,
     dispatch_channel_slash_command,
+    get_channel_startup_results,
 )
 from .file_mentions import complete_file_mention, resolve_file_mentions
 from .history_suggester import HistorySuggester
@@ -77,6 +78,9 @@ from .status_bar import (
     make_usage_status_snapshot,
 )
 
+if TYPE_CHECKING:
+    from ..runtime import AsyncRuntime
+
 _channel_logger = logging.getLogger(__name__)
 
 if TYPE_CHECKING:
@@ -92,6 +96,45 @@ def _shorten_path(path: str) -> str:
     return _sp(path)
 
 
+async def _auto_start_channel_in_worker(
+    agent: Any,
+    thread_id: str,
+    config: Any,
+    *,
+    send_thinking: bool,
+    runtime: Any,
+    stop_requested: threading.Event,
+) -> list[tuple[str, bool, str]]:
+    """Run blocking channel startup without occupying the TUI event loop."""
+
+    def _start() -> list[tuple[str, bool, str]]:
+        try:
+            return _auto_start_channel(
+                agent,
+                thread_id,
+                config,
+                send_thinking=send_thinking,
+                runtime=runtime,
+            )
+        finally:
+            if stop_requested.is_set():
+                _channels_stop(runtime=runtime)
+
+    worker = asyncio.create_task(asyncio.to_thread(_start))
+    try:
+        return await asyncio.shield(worker)
+    except asyncio.CancelledError:
+        stop_requested.set()
+        try:
+            await worker
+        except Exception:
+            _channel_logger.debug(
+                "Channel startup worker failed during cancellation",
+                exc_info=True,
+            )
+        raise
+
+
 def _build_welcome_banner(
     *,
     thread_id: str,
@@ -220,6 +263,9 @@ async def _sync_tui_command_completion(
     cmd: Command,
 ) -> None:
     """Adopt successful command-side state changes back into the TUI app."""
+    if app._exiting:
+        return
+
     agent_swapped = ctx.agent is not None and ctx.agent is not original_agent
     if agent_swapped:
         from ..EvoScientist import _ensure_config
@@ -273,6 +319,165 @@ def _stopped_response_after_narration(
     return display_current, display_stopped, full_stopped
 
 
+# (kind, payload, item_index): kind is "header"/"sep"/"item"; payload is the
+# category name for headers or the candidate for items; item_index is the
+# candidate's position in the source list (-1 for non-item rows).
+_CompletionRow = tuple[str, Any, int]
+
+
+def _build_completion_rows(items: list[Any]) -> list[_CompletionRow]:
+    """Flatten completion candidates into render rows with category headers."""
+    rows: list[_CompletionRow] = []
+    last_cat = ""
+    for i, candidate in enumerate(items):
+        cat = getattr(candidate, "category", "")
+        if cat and cat != last_cat:
+            if last_cat:
+                rows.append(("sep", "", -1))
+            rows.append(("header", cat, -1))
+            last_cat = cat
+        rows.append(("item", candidate, i))
+    return rows
+
+
+def _window_completion_rows(
+    rows: list[_CompletionRow],
+    selected: int,
+    max_rows: int,
+) -> tuple[list[_CompletionRow], int, int]:
+    """Slice *rows* to a window of at most *max_rows* total display lines.
+
+    The window always contains the selected item (top of the list when
+    nothing is selected) and reserves one line per overflow indicator.
+    Returns ``(visible_rows, hidden_items_above, hidden_items_below)``.
+    """
+    max_rows = max(max_rows, 5)
+    if len(rows) <= max_rows:
+        return list(rows), 0, 0
+
+    sel_row = 0
+    if selected >= 0:
+        for r, (kind, _payload, idx) in enumerate(rows):
+            if kind == "item" and idx == selected:
+                sel_row = r
+                break
+
+    # Center the selection; centering keeps it clear of the indicator
+    # lines that replace the window's edge rows when content is clipped.
+    start = min(max(sel_row - max_rows // 2, 0), len(rows) - max_rows)
+    end = start + max_rows
+    content_start = start + (1 if start > 0 else 0)
+    content_end = end - (1 if end < len(rows) else 0)
+
+    above = sum(1 for kind, _p, _i in rows[:content_start] if kind == "item")
+    below = sum(1 for kind, _p, _i in rows[content_end:] if kind == "item")
+    return rows[content_start:content_end], above, below
+
+
+def _render_completion_text(items: list[Any], selected: int, max_rows: int) -> Text:
+    """Render the completion popup content bounded to *max_rows* lines."""
+    rows = _build_completion_rows(items)
+    visible, above, below = _window_completion_rows(rows, selected, max_rows)
+
+    # Blank separator lines are cosmetic — drop them at the window edges.
+    while visible and visible[0][0] == "sep":
+        visible = visible[1:]
+    while visible and visible[-1][0] == "sep":
+        visible = visible[:-1]
+
+    lines: list[Text] = []
+    if above:
+        lines.append(Text(f"   ↑ {above} more", style="dim italic"))
+    for kind, payload, idx in visible:
+        if kind == "sep":
+            lines.append(Text())
+        elif kind == "header":
+            lines.append(Text(f" {payload}", style="bold #6b7280"))
+        elif idx == selected:
+            lines.append(
+                Text.assemble(
+                    ("  ▸ ", "bold"),
+                    (f"{payload.text:<28}", "bold"),
+                    (payload.description, "bold"),
+                )
+            )
+        else:
+            lines.append(
+                Text.assemble(
+                    ("    ", "#888888"),
+                    (f"{payload.text:<28}", "#888888"),
+                    (payload.description, "#888888"),
+                )
+            )
+    if below:
+        lines.append(Text(f"   ↓ {below} more", style="dim italic"))
+    return Text("\n").join(lines)
+
+
+# Hard cap on popup lines so the popup never dwarfs the chat area
+# (mainstream CLI behavior); matches the pre-#354 max-height.
+_COMPLETION_MAX_VISIBLE_ROWS = 15
+# Rows kept free for the input row, status bar and a slice of chat. On
+# terminals shorter than ~17 rows the 5-row floor wins over this
+# reservation — a smaller popup would be unusable.
+_COMPLETION_RESERVED_ROWS = 12
+
+
+def _completion_row_budget(height: int) -> int:
+    """Popup line budget for a terminal of *height* rows."""
+    if height <= 0:
+        return _COMPLETION_MAX_VISIBLE_ROWS
+    return max(5, min(height - _COMPLETION_RESERVED_ROWS, _COMPLETION_MAX_VISIBLE_ROWS))
+
+
+# Textual converts rich Text to Content and drops rich no_wrap/overflow
+# attributes, so line cropping must be enforced here in CSS.
+_COMPLETIONS_CSS = """
+#completions {
+    display: none;
+    height: auto;
+    background: #1e1f26;
+    padding: 0 1;
+    border-bottom: solid #0284c7;
+    text-wrap: nowrap;
+    text-overflow: ellipsis;
+}
+"""
+
+
+def _normalize_chat_scroll(container: Any) -> None:
+    """Repair the chat scroll state after the popup resized the viewport.
+
+    Textual's compositor recomputes ``scroll_y`` for anchored containers
+    bypassing the validator, so when the popup hides and the content fits
+    again, ``scroll_y`` can go negative — the scrollbar then renders as if
+    scrolled to the bottom while the content sits at the top (issue #301
+    family). Runs after refresh so sizes are current.
+    """
+    # force=True: with the content fitting, the scrollbar is hidden and
+    # allow_vertical_scroll is False — an unforced scroll_home would
+    # silently no-op and leave the negative scroll_y in place.
+    if container.is_anchored:
+        if container.max_scroll_y <= 0:
+            container.anchor(False)
+            container.scroll_home(animate=False, immediate=True, force=True)
+    elif container.scroll_y < 0:
+        container.scroll_home(animate=False, immediate=True, force=True)
+    # Resync the scrollbar thumb: watch_scroll_y skips the update while
+    # the scrollbar is hidden (or when the compositor wrote scroll_y via
+    # set_reactive), so a stale position survives until the scrollbar
+    # reappears — rendering as "scrolled to bottom" at the top.
+    scrollbar = getattr(container, "vertical_scrollbar", None)
+    if scrollbar is not None and scrollbar.position != container.scroll_y:
+        scrollbar.position = container.scroll_y
+
+
+def _session_auto_approve_decisions(action_requests: list) -> list[dict]:
+    """TUI session "approve all": an explicit human opt-in, so blanket-approve
+    everything (dangerous set included), matching the Rich CLI and channel."""
+    return [{"type": "approve"} for _ in action_requests]
+
+
 def run_textual_interactive(
     *,
     show_thinking: bool,
@@ -287,6 +492,7 @@ def run_textual_interactive(
     load_agent: Callable[..., Any],
     create_session_workspace: Callable[[str | None], str],
     config: Any | None = None,
+    async_runtime: AsyncRuntime | None = None,
 ) -> None:
     """Run full-screen Textual interactive chat loop."""
     if config is None:
@@ -294,7 +500,15 @@ def run_textual_interactive(
 
         config = get_effective_config()
 
-    runtime_gateways = create_runtime_gateways()
+    # One frontend event sink for the whole TUI session — injected into the
+    # agent's middleware (write side) and the local gateway's streaming path
+    # (read side). The fallback-notice display is bound to the App's
+    # _append_system once the App exists (on_mount); tool-selection needs no
+    # display hook (its widget is mounted from the stream event).
+    from ..stream.sink import SessionEventSink
+
+    event_sink = SessionEventSink()
+    runtime_gateways = create_runtime_gateways(events=event_sink)
     graph_gateway = runtime_gateways.graph_gateway
 
     try:
@@ -311,6 +525,7 @@ def run_textual_interactive(
             CompactingWidget,
             LoadingWidget,
             MCPLoaderWidget,
+            PanelWidget,
             SubAgentWidget,
             SummarizationWidget,
             SystemMessage,
@@ -333,7 +548,8 @@ def run_textual_interactive(
         def supports_interactive(self) -> bool:
             return True
 
-        CSS = """
+        CSS = (
+            """
         Screen {
             layout: vertical;
             background: #16161a;
@@ -385,14 +601,9 @@ def run_textual_interactive(
             padding: 0 2;
             color: #9ca3af;
         }
-        #completions {
-            display: none;
-            height: auto;
-            max-height: 15;
-            background: #1e1f26;
-            padding: 0 1;
-            border-bottom: solid #0284c7;
-        }
+        """
+            + _COMPLETIONS_CSS
+            + """
         #status {
             height: 1;
             min-height: 1;
@@ -401,6 +612,7 @@ def run_textual_interactive(
             padding: 0 1;
         }
         """
+        )
         BINDINGS: ClassVar[list[Binding]] = [
             Binding("ctrl+c", "request_quit", "Quit", show=False, priority=True),
             Binding("ctrl+v", "paste_clipboard", "Paste", show=False),
@@ -438,7 +650,8 @@ def run_textual_interactive(
             self._resumed = resumed
             self._resume_warning = resume_warning
             self._channel_timer: Any = None
-            self._started_channel_types: list[str] = []
+            self._channel_start_results: list[tuple[str, bool, str]] = []
+            self._channel_start_stop = threading.Event()
             self._busy = False
             self._notification_consuming: bool = (
                 False  # prevent overlapping consume coroutines
@@ -449,6 +662,7 @@ def run_textual_interactive(
             ] = []  # queued messages to send after current turn
             self._comp_items: list = []
             self._comp_index: int = -1
+            self._comp_last_height: int = 0
             self._comp_base: str = ""
             self._hitl_auto_approve: bool = False
             self._approval_future: asyncio.Future | None = None
@@ -465,6 +679,7 @@ def run_textual_interactive(
 
             self._channel_runtime = ChannelRuntime()
             self._quit_pending: bool = False
+            self._exiting: bool = False
             self._current_model: str | None = model
             self._current_provider: str | None = provider
             self._status_started_at = datetime.now()
@@ -516,6 +731,7 @@ def run_textual_interactive(
             self._agent_loader.start(
                 workspace_dir=workspace,
                 checkpointer=self._checkpointer,
+                events=self._runtime_gateways.graph_gateway.events,
             )
 
         def _mount_mcp_loader_widget(self) -> None:
@@ -794,11 +1010,13 @@ def run_textual_interactive(
             yield Static("", id="status")
 
         def on_mount(self) -> None:
-            # Register fallback middleware UI callback so messages appear
-            # as SystemMessage widgets in the chat container.
-            from ..middleware.model_fallback import set_ui_emit
-
-            set_ui_emit(lambda text, style: self._append_system(text, style))
+            # Bind the session sink's fallback-notice display so model-fallback
+            # messages appear as SystemMessage widgets in the chat container.
+            # ``event_sink`` is the concrete SessionEventSink created by the
+            # enclosing factory — the same instance the gateway carries.
+            event_sink.set_fallback_display(
+                lambda text, style: self._append_system(text, style)
+            )
 
             self._render_welcome()
             self._render_status()
@@ -850,12 +1068,23 @@ def run_textual_interactive(
                         exc_info=True,
                     )
                     return
-                self._start_channels()
+                await self._start_channels()
 
             ch_task = asyncio.create_task(_deferred_start_channels())
             self._background_tasks.add(ch_task)
             ch_task.add_done_callback(self._background_tasks.discard)
 
+        def on_resize(self, event: Any) -> None:
+            """Re-window the completion popup for the new terminal height."""
+            try:
+                comp_widget = self.query_one("#completions", Static)
+            except Exception:
+                return
+            if comp_widget.display and self._comp_items:
+                # Deferred: this handler can run before the base App
+                # handler updates self.size with the new dimensions.
+                self.call_after_refresh(self._render_completions)
+
         # ── Update check ──────────────────────────────────────
 
         async def _check_for_updates(self) -> None:
@@ -877,28 +1106,46 @@ def run_textual_interactive(
 
         # ── Channel integration ────────────────────────────────
 
-        def _start_channels(self) -> None:
+        async def _start_channels(self) -> None:
             """Auto-start channels if enabled in config."""
             try:
                 from ..config import load_config
 
-                cfg = load_config()
+                cfg = await asyncio.to_thread(load_config)
                 if cfg and cfg.channel_enabled and not _channels_is_running():
-                    _auto_start_channel(
+                    results = await _auto_start_channel_in_worker(
                         self._agent_loader.agent,
                         self._conversation_tid,
                         cfg,
                         send_thinking=self._channel_send_thinking,
                         runtime=self._channel_runtime,
+                        stop_requested=self._channel_start_stop,
                     )
-                    types = [
-                        t.strip() for t in cfg.channel_enabled.split(",") if t.strip()
-                    ]
-                    self._started_channel_types = types
+                    if self._exiting:
+                        return
+                    current_agent = self._agent_loader.agent
+                    if current_agent is not None and _channels_is_running():
+                        self._channel_runtime.bind(
+                            current_agent,
+                            self._conversation_tid,
+                        )
+                    self._channel_start_results = results
                     self._render_welcome()
+            except asyncio.CancelledError:
+                self._channel_start_stop.set()
+                raise
             except Exception as e:
                 _channel_logger.debug(f"Channel auto-start failed: {e}")
-            self._channel_timer = self.set_interval(0.1, self._poll_channel_queue)
+            finally:
+                if (
+                    not self._exiting
+                    and not self._channel_start_stop.is_set()
+                    and self._channel_timer is None
+                ):
+                    self._channel_timer = self.set_interval(
+                        0.1,
+                        self._poll_channel_queue,
+                    )
 
         def _poll_channel_queue(self) -> None:
             """Poll the channel + notification queues (every 100ms)."""
@@ -1075,7 +1322,9 @@ def run_textual_interactive(
             go negative and pushes the welcome banner out of view (issue #301).
             """
             container.anchor(False)
-            container.scroll_home(animate=False, immediate=True)
+            # force=True: with content fitting, the scrollbar is hidden and
+            # allow_vertical_scroll is False — unforced scroll_home no-ops.
+            container.scroll_home(animate=False, immediate=True, force=True)
 
         def _append_system(self, text: str, style: str = "dim") -> None:
             """Mount a SystemMessage widget into #chat."""
@@ -1125,7 +1374,7 @@ def run_textual_interactive(
             Returns the ``ApprovalWidget.Decided`` message, or ``None`` on
             timeout / cancellation.
             """
-            self._approval_future = asyncio.get_event_loop().create_future()
+            self._approval_future = asyncio.get_running_loop().create_future()
             try:
                 return await asyncio.wait_for(self._approval_future, timeout=300)
             except (TimeoutError, asyncio.CancelledError):
@@ -1171,7 +1420,7 @@ def run_textual_interactive(
 
             Returns the selected thread_id, or ``None`` on cancel/timeout.
             """
-            self._picker_future = asyncio.get_event_loop().create_future()
+            self._picker_future = asyncio.get_running_loop().create_future()
             try:
                 return await asyncio.wait_for(self._picker_future, timeout=120)
             except (TimeoutError, asyncio.CancelledError):
@@ -1199,7 +1448,7 @@ def run_textual_interactive(
 
             Returns list of install sources, or None on cancel/timeout.
             """
-            self._browser_future = asyncio.get_event_loop().create_future()
+            self._browser_future = asyncio.get_running_loop().create_future()
             try:
                 return await asyncio.wait_for(self._browser_future, timeout=300)
             except (TimeoutError, asyncio.CancelledError):
@@ -1226,7 +1475,7 @@ def run_textual_interactive(
 
         async def _wait_for_mcp_browse(self, browser_widget) -> list | None:
             """Wait for user to complete MCP server browsing."""
-            self._mcp_browser_future = asyncio.get_event_loop().create_future()
+            self._mcp_browser_future = asyncio.get_running_loop().create_future()
             try:
                 return await asyncio.wait_for(self._mcp_browser_future, timeout=300)
             except (TimeoutError, asyncio.CancelledError):
@@ -1254,7 +1503,7 @@ def run_textual_interactive(
 
             Returns ``(name, provider)`` or ``None`` on cancel/timeout.
             """
-            self._model_picker_future = asyncio.get_event_loop().create_future()
+            self._model_picker_future = asyncio.get_running_loop().create_future()
             try:
                 return await asyncio.wait_for(self._model_picker_future, timeout=120)
             except (TimeoutError, asyncio.CancelledError):
@@ -1316,6 +1565,7 @@ def run_textual_interactive(
             """
             from ..stream.display import (
                 is_stream_cancel_requested,
+                iter_with_stream_cancel,
             )
 
             container = self.query_one("#chat", VerticalScroll)
@@ -1341,6 +1591,7 @@ def run_textual_interactive(
             todo_w: TodoWidget | None = None
             tool_widgets: dict[str, ToolCallWidget] = {}
             subagent_widgets: dict[str, SubAgentWidget] = {}
+            panel_widgets: dict[str, PanelWidget] = {}
 
             @dataclass
             class _ResponseDisplayState:
@@ -1542,16 +1793,26 @@ def run_textual_interactive(
                     summarization_w = None
                 try:
                     _anchor_engaged = False
-                    async for event in graph_gateway.stream_events(
-                        RunRequest(
-                            message=_stream_input,
-                            thread_id=thread_id_override or self._conversation_tid,
-                            metadata=metadata,
-                            target=GraphTarget(
-                                local_graph=agent,
-                                workspace_dir=self._workspace_dir,
-                            ),
-                        )
+                    _active_teams = list(self._channel_runtime.active_teams)
+                    _configurable_extra = (
+                        {"active_teams": _active_teams} if _active_teams else None
+                    )
+                    async for event in iter_with_stream_cancel(
+                        graph_gateway.stream_events(
+                            RunRequest(
+                                message=_stream_input,
+                                thread_id=(
+                                    thread_id_override or self._conversation_tid
+                                ),
+                                metadata=metadata,
+                                target=GraphTarget(
+                                    local_graph=agent,
+                                    workspace_dir=self._workspace_dir,
+                                ),
+                                configurable_extra=_configurable_extra,
+                            )
+                        ),
+                        cancel_scope,
                     ):
                         if is_stream_cancel_requested(cancel_scope):
                             response = await _mark_cancelled_response()
@@ -1846,6 +2107,40 @@ def run_textual_interactive(
                             if sa_w is not None:
                                 sa_w.finalize()
 
+                        elif event_type == "panel_dispatch_start":
+                            eval_id = event.get("eval_id", "") or "_unbatched"
+                            panel_w = panel_widgets.get(eval_id)
+                            if panel_w is None:
+                                panel_w = PanelWidget(eval_id)
+                                # Register before awaiting mount: a cancel
+                                # during the await would otherwise orphan a
+                                # ticking panel outside the cleanup loop.
+                                panel_widgets[eval_id] = panel_w
+                                await container.mount(panel_w)
+                            await panel_w.start_dispatch(
+                                event["id"],
+                                event.get("subagent_type", ""),
+                                event.get("label", "") or event.get("description", ""),
+                            )
+
+                        elif event_type == "panel_dispatch_complete":
+                            eval_id = event.get("eval_id", "") or "_unbatched"
+                            panel_w = panel_widgets.get(eval_id)
+                            if panel_w is not None:
+                                panel_w.complete_dispatch(
+                                    event["id"], int(event.get("duration_ms", 0))
+                                )
+
+                        elif event_type == "panel_dispatch_error":
+                            eval_id = event.get("eval_id", "") or "_unbatched"
+                            panel_w = panel_widgets.get(eval_id)
+                            if panel_w is not None:
+                                panel_w.fail_dispatch(
+                                    event["id"],
+                                    int(event.get("duration_ms", 0)),
+                                    event.get("error", ""),
+                                )
+
                         elif event_type == "ask_user":
                             questions = event.get("questions", [])
                             if questions:
@@ -1888,20 +2183,15 @@ def run_textual_interactive(
 
                         elif event_type == "interrupt":
                             action_reqs = event.get("action_requests", [])
-                            n = len(action_reqs) or 1
+                            interrupt_id = event.get("interrupt_id")
 
-                            # HITL: check session auto-approve first
+                            # HITL: session "approve all" blanket-approves.
                             if self._hitl_auto_approve:
-                                from langgraph.types import (
-                                    Command,  # type: ignore[import-untyped]
-                                )
+                                from ..backends import build_hitl_resume
 
-                                _stream_input = Command(
-                                    resume={
-                                        "decisions": [
-                                            {"type": "approve"} for _ in range(n)
-                                        ]
-                                    }
+                                decisions = _session_auto_approve_decisions(action_reqs)
+                                _stream_input = build_hitl_resume(
+                                    interrupt_id, decisions
                                 )
                                 _hitl_resuming = True
                                 break  # re-enter outer HITL loop
@@ -1921,12 +2211,10 @@ def run_textual_interactive(
                                     response = await _mark_cancelled_response()
                                     break
                                 if decisions is not None:
-                                    from langgraph.types import (
-                                        Command,  # type: ignore[import-untyped]
-                                    )
+                                    from ..backends import build_hitl_resume
 
-                                    _stream_input = Command(
-                                        resume={"decisions": decisions}
+                                    _stream_input = build_hitl_resume(
+                                        interrupt_id, decisions
                                     )
                                     _hitl_resuming = True
                                     break  # re-enter outer HITL loop
@@ -1955,12 +2243,10 @@ def run_textual_interactive(
                             if decided_event and decided_event.decisions is not None:
                                 if decided_event.auto_approve_session:
                                     self._hitl_auto_approve = True
-                                from langgraph.types import (
-                                    Command,  # type: ignore[import-untyped]
-                                )
+                                from ..backends import build_hitl_resume
 
-                                _stream_input = Command(
-                                    resume={"decisions": decided_event.decisions}
+                                _stream_input = build_hitl_resume(
+                                    interrupt_id, decided_event.decisions
                                 )
                                 _hitl_resuming = True
                                 break  # re-enter outer HITL loop with resume
@@ -2075,6 +2361,13 @@ def run_textual_interactive(
                                 sa_w.finalize()
                             except Exception:
                                 pass
+                    # Finalize any still-running panel dispatches so their
+                    # per-row spinner timers stop instead of ticking forever.
+                    for panel_w in panel_widgets.values():
+                        try:
+                            panel_w.finalize_running()
+                        except Exception:
+                            pass
                     # Finalize thinking widget
                     if thinking_w is not None and thinking_w._is_active:
                         try:
@@ -2142,6 +2435,11 @@ def run_textual_interactive(
             cancelled = False
             response = ""
             try:
+                # Foreground turns share the legacy default scope. Reset it at
+                # the turn boundary; scoped channel stop requests remain armed.
+                from ..stream.display import clear_stream_cancel
+
+                clear_stream_cancel()
                 self._busy = True
                 self._turn_started_at = datetime.now()
                 self._status_phase = ResearchPhase.THINKING
@@ -2176,6 +2474,17 @@ def run_textual_interactive(
                 )
             except asyncio.CancelledError:
                 cancelled = True
+                try:
+                    from ..middleware.code_interpreter import (
+                        aclose_code_interpreters,
+                    )
+
+                    await aclose_code_interpreters()
+                except Exception:
+                    _channel_logger.debug(
+                        "code interpreter cleanup after cancellation failed",
+                        exc_info=True,
+                    )
                 self._append_system("\nInterrupted by user", style="dim italic #ffe082")
             finally:
                 self._busy = False
@@ -2309,6 +2618,7 @@ def run_textual_interactive(
                     on_cmd_completed=self._on_channel_cmd_completed,
                     channel_runtime=self._channel_runtime,
                     graph_gateway=self._runtime_gateways.graph_gateway,
+                    async_runtime=async_runtime,
                 )
                 if _slash_handled:
                     # A channel-issued /new or /resume rotates the thread in
@@ -2784,30 +3094,36 @@ def run_textual_interactive(
         def _hide_completions(self) -> None:
             self._comp_items = []
             self._comp_index = -1
-            self.query_one("#completions", Static).display = False
+            self._comp_last_height = 0
+            comp_widget = self.query_one("#completions", Static)
+            was_visible = comp_widget.display
+            comp_widget.display = False
+            # Called on every ordinary input change — only a popup that was
+            # actually visible changed the chat viewport.
+            if was_visible:
+                self.call_after_refresh(self._normalize_chat_after_popup)
+
+        def _completion_max_rows(self) -> int:
+            return _completion_row_budget(int(getattr(self.size, "height", 0) or 0))
 
         def _render_completions(self) -> None:
-            comp_text = Text()
-            last_cat = ""
-            for i, candidate in enumerate(self._comp_items):
-                cmd, desc = candidate.text, candidate.description
-                cat = getattr(candidate, "category", "")
-                if cat and cat != last_cat:
-                    if last_cat:
-                        comp_text.append("\n")
-                    comp_text.append(f" {cat}\n", style="bold #6b7280")
-                    last_cat = cat
-                if i == self._comp_index:
-                    comp_text.append("  \u25b8 ", style="bold")
-                    comp_text.append(f"{cmd:<28}", style="bold")
-                    comp_text.append(desc, style="bold")
-                else:
-                    comp_text.append("    ", style="#888888")
-                    comp_text.append(f"{cmd:<28}", style="#888888")
-                    comp_text.append(desc, style="#888888")
-                if i < len(self._comp_items) - 1:
-                    comp_text.append("\n")
+            comp_text = _render_completion_text(
+                self._comp_items, self._comp_index, self._completion_max_rows()
+            )
             self.query_one("#completions", Static).update(comp_text)
+            # Selection-only navigation keeps the height — skip the (cheap
+            # but per-keystroke) normalize unless the viewport can change.
+            n_lines = len(comp_text.plain.splitlines()) if comp_text.plain else 0
+            if n_lines != self._comp_last_height:
+                self._comp_last_height = n_lines
+                self.call_after_refresh(self._normalize_chat_after_popup)
+
+        def _normalize_chat_after_popup(self) -> None:
+            try:
+                container = self.query_one("#chat", VerticalScroll)
+            except Exception:
+                return
+            _normalize_chat_scroll(container)
 
         # ── Slash commands ─────────────────────────────────────
 
@@ -2847,6 +3163,7 @@ def run_textual_interactive(
                     input_tokens_hint=self._status_last_input_tokens,
                     channel_runtime=self._channel_runtime,
                     graph_gateway=self._runtime_gateways.graph_gateway,
+                    async_runtime=async_runtime,
                 )
 
                 if await cmd_manager.execute(command, ctx):
@@ -2864,8 +3181,9 @@ def run_textual_interactive(
                 self._render_status()
             finally:
                 self._busy = False
-                prompt_widget.disabled = False
-                prompt_widget.focus()
+                if not self._exiting:
+                    prompt_widget.disabled = False
+                    prompt_widget.focus()
 
         async def _render_history(self, thread_id_value: str) -> None:
             """Render conversation history from a saved thread.
@@ -2970,13 +3288,13 @@ def run_textual_interactive(
 
         def _do_exit(self) -> None:
             """Clean up channels, unregister callbacks, and exit."""
-            from ..middleware.model_fallback import set_ui_emit
-
-            set_ui_emit(None)
+            self._exiting = True
+            self._channel_start_stop.set()
+            event_sink.set_fallback_display(None)
             if self._channel_timer is not None:
                 self._channel_timer.stop()
                 self._channel_timer = None
-            self._started_channel_types.clear()
+            self._channel_start_results.clear()
             if _channels_is_running():
                 try:
                     _channels_stop(runtime=self._channel_runtime)
@@ -2992,6 +3310,9 @@ def run_textual_interactive(
                     self._queued_messages.clear()
                     self._render_queue_indicator()
                 if self._run_task is not None and not self._run_task.done():
+                    from ..stream.display import request_stream_cancel
+
+                    request_stream_cancel()
                     self._run_task.cancel()
                 else:
                     # Edge case: busy but no task — force reset
@@ -3105,11 +3426,11 @@ def run_textual_interactive(
         def _render_welcome(self) -> None:
             channels_info: list[tuple[str, bool, str]] | None = None
             try:
-                running = _channels_running_list()
-                started = self._started_channel_types
-                if running or started:
-                    all_types = list(dict.fromkeys(running + started))
-                    channels_info = [(ct, True, "connected (bus)") for ct in all_types]
+                current = get_channel_startup_results()
+                if current:
+                    self._channel_start_results = current
+                if self._channel_start_results:
+                    channels_info = self._channel_start_results
                 else:
                     from ..config import load_config
 
@@ -3359,6 +3680,18 @@ def run_textual_interactive(
             finally:
                 from .resume_hint import print_resume_hint
 
+                try:
+                    from ..middleware.code_interpreter import (
+                        aclose_code_interpreters,
+                    )
+
+                    await aclose_code_interpreters()
+                except Exception:
+                    _channel_logger.debug(
+                        "code interpreter cleanup failed",
+                        exc_info=True,
+                    )
+
                 # Best-effort resume hint — guarded so failures here (e.g.
                 # DB teardown race during abnormal shutdown) cannot shadow
                 # the original run_async traceback.
@@ -3378,12 +3711,4 @@ def run_textual_interactive(
                 except Exception:
                     _channel_logger.debug("print_resume_hint failed", exc_info=True)
 
-    import nest_asyncio  # type: ignore[import-untyped]
-
-    nest_asyncio.apply()
-    try:
-        loop = asyncio.get_event_loop()
-    except RuntimeError:
-        loop = asyncio.new_event_loop()
-        asyncio.set_event_loop(loop)
-    loop.run_until_complete(_amain())
+    asyncio.run(_amain())

EvoScientist/cli/tui_runtime.py

@@ -2,14 +2,20 @@
 
 from __future__ import annotations
 
+import asyncio
 from collections.abc import Callable
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 from ..gateway import GraphGateway
+from ..runtime import AsyncRuntimeError
 from ..stream.console import console
 from .tui_backends import RichStreamingBackend, StreamingTUIBackend
 
+if TYPE_CHECKING:
+    from ..runtime import AsyncRuntime
+
 DEFAULT_UI_BACKEND = "cli"
+STREAM_CANCEL_SETTLE_TIMEOUT = 5.0
 # "webui" launches the browser front-end instead of an in-terminal UI; it is
 # intercepted earlier (cli/commands.py:_main_callback) and never reaches the
 # streaming backends, but is listed here so normalize/resolve preserve it
@@ -18,6 +24,41 @@ SUPPORTED_UI_BACKENDS = ("cli", "tui", "webui")
 _LEGACY_BACKEND_MAP = {"textual": "tui", "rich": "cli"}
 
 
+class StreamCancellationTimeout(RuntimeError):
+    """A blocking renderer did not settle after its turn was cancelled."""
+
+
+def _consume_late_worker_result(worker: asyncio.Task[Any]) -> None:
+    """Retrieve a detached worker result so eventual failure is not unhandled."""
+    try:
+        worker.exception()
+    except asyncio.CancelledError:
+        pass
+
+
+async def settle_cancelled_worker(
+    worker: asyncio.Task[Any],
+    *,
+    on_cancel: Callable[[], Any],
+) -> Any:
+    """Request cooperative cancellation and wait a bounded time for settlement."""
+    on_cancel()
+    done, _ = await asyncio.wait(
+        {worker},
+        timeout=STREAM_CANCEL_SETTLE_TIMEOUT,
+    )
+    if not done:
+        worker.add_done_callback(_consume_late_worker_result)
+        raise StreamCancellationTimeout(
+            "The active turn did not stop within "
+            f"{STREAM_CANCEL_SETTLE_TIMEOUT:g} seconds after cancellation."
+        )
+    try:
+        return worker.result()
+    except Exception:
+        return ""
+
+
 def normalize_ui_backend(value: str | None) -> str:
     """Normalize user-provided backend name with a safe default."""
     if not value:
@@ -77,10 +118,12 @@ def run_streaming(
     on_stream_event: Callable[[str, Any], Any] | None = None,
     status_footer_builder: Callable[[], Any] | None = None,
     metadata: dict | None = None,
+    configurable_extra: dict[str, Any] | None = None,
     hitl_prompt_fn: Callable[[list], list[dict] | None] | None = None,
     ask_user_prompt_fn: Callable[[dict], dict] | None = None,
     cancel_scope: str | None = None,
     gateway: GraphGateway,
+    runtime: AsyncRuntime | None = None,
 ) -> str:
     """Run streaming with the selected backend."""
     backend = get_backend(ui_backend, warn_fallback=True)
@@ -97,11 +140,15 @@ def run_streaming(
             on_stream_event=on_stream_event,
             status_footer_builder=status_footer_builder,
             metadata=metadata,
+            configurable_extra=configurable_extra,
             hitl_prompt_fn=hitl_prompt_fn,
             ask_user_prompt_fn=ask_user_prompt_fn,
             cancel_scope=cancel_scope,
             gateway=gateway,
+            runtime=runtime,
         )
+    except AsyncRuntimeError:
+        raise
     except RuntimeError:
         requested = normalize_ui_backend(ui_backend)
         if requested == "tui":
@@ -120,9 +167,45 @@ def run_streaming(
                 on_stream_event=on_stream_event,
                 status_footer_builder=status_footer_builder,
                 metadata=metadata,
+                configurable_extra=configurable_extra,
                 hitl_prompt_fn=hitl_prompt_fn,
                 ask_user_prompt_fn=ask_user_prompt_fn,
                 cancel_scope=cancel_scope,
                 gateway=gateway,
+                runtime=runtime,
             )
         raise
+
+
+async def run_streaming_async(
+    *,
+    recover_on_cancel: bool = False,
+    **kwargs: Any,
+) -> str:
+    """Run the synchronous Rich renderer without blocking a frontend loop.
+
+    Cancellation requests the matching stream scope and gives the worker a
+    bounded interval to unwind.  Foreground interactive turns may opt into
+    recovering the frontend task after cleanup so Ctrl+C returns to the prompt.
+    """
+    from ..stream.display import request_stream_cancel
+
+    worker = asyncio.create_task(asyncio.to_thread(run_streaming, **kwargs))
+    try:
+        return await asyncio.shield(worker)
+    except asyncio.CancelledError:
+        try:
+            response = await settle_cancelled_worker(
+                worker,
+                on_cancel=lambda: request_stream_cancel(kwargs.get("cancel_scope")),
+            )
+        finally:
+            from ..middleware.code_interpreter import aclose_code_interpreters
+
+            await aclose_code_interpreters()
+        if recover_on_cancel:
+            current = asyncio.current_task()
+            if current is not None and current.uncancel() > 0:
+                raise
+            return response
+        raise

EvoScientist/cli/widgets/__init__.py

@@ -7,6 +7,7 @@ from .compact_summary_widget import CompactSummaryWidget
 from .compacting_widget import CompactingWidget
 from .loading_widget import LoadingWidget
 from .mcp_loader_widget import MCPLoaderWidget
+from .panel_widget import PanelWidget
 from .subagent_widget import SubAgentWidget
 from .summarization_widget import SummarizationWidget
 from .system_message import SystemMessage
@@ -25,6 +26,7 @@ __all__ = [
     "CompactingWidget",
     "LoadingWidget",
     "MCPLoaderWidget",
+    "PanelWidget",
     "SubAgentWidget",
     "SummarizationWidget",
     "SystemMessage",

EvoScientist/cli/widgets/approval_widget.py

@@ -101,6 +101,15 @@ class ApprovalWidget(Widget):
         self._selected = 0
         self._option_widgets: list[Static] = []
 
+    @staticmethod
+    def _extract_command(args: dict) -> str:
+        """Pull the display-worthy target out of a tool's args dict.
+
+        Checks `command`/`path` first, then deepagents 0.7.0's `delete`
+        tool key `file_path` — without it, `delete` shows no target.
+        """
+        return args.get("command", args.get("path", args.get("file_path", "")))
+
     def compose(self) -> ComposeResult:
         self._option_widgets = []
         count = len(self._action_requests)
@@ -115,10 +124,7 @@ class ApprovalWidget(Widget):
         for req in self._action_requests:
             name = req.get("name", "")
             args = req.get("args", {})
-            if isinstance(args, dict):
-                command = args.get("command", args.get("path", ""))
-            else:
-                command = ""
+            command = self._extract_command(args) if isinstance(args, dict) else ""
             if command:
                 cmd_str = str(command)
                 if len(cmd_str) > _COMMAND_TRUNCATE_LENGTH:

EvoScientist/cli/widgets/mcp_browser.py

@@ -1,6 +1,7 @@
 """Inline MCP server browser widget for /install-mcp in TUI.
 
-Two-phase keyboard-driven widget (mirrors SkillBrowserWidget):
+Two-phase keyboard-driven widget built on the shared picker engine
+(``picker_base.TagCheckboxBrowserBase``):
   Phase 1 — tag picker (arrow keys + Enter to select, or Esc for all)
   Phase 2 — server checkbox (arrow keys to navigate, Space to toggle, Enter to confirm)
 
@@ -12,73 +13,20 @@ from __future__ import annotations
 
 from typing import TYPE_CHECKING, Any, ClassVar
 
-from rich.text import Text
-from textual.binding import Binding, BindingType
-from textual.containers import Container
 from textual.message import Message
-from textual.widget import Widget
-from textual.widgets import Static
+
+from .picker_base import TagCheckboxBrowserBase
 
 if TYPE_CHECKING:
-    from textual import events
-    from textual.app import ComposeResult
-
     from ...mcp.registry import MCPServerEntry
 
 
-class MCPBrowserWidget(Widget):
-    """Inline MCP server browser — mounts in chat, keyboard-driven.
+class MCPBrowserWidget(TagCheckboxBrowserBase):
+    """Inline MCP server browser — mounts in chat, keyboard-driven."""
 
-    Phase 1: Tag picker (select a tag filter or "All").
-    Phase 2: Server checkbox (toggle servers, confirm to install).
-    """
-
-    can_focus = True
-    can_focus_children = False
-
-    DEFAULT_CSS = """
-    MCPBrowserWidget {
-        height: auto;
-        max-height: 30;
-        margin: 1 0;
-        padding: 0 1;
-        background: $surface;
-        border: solid $primary;
-    }
-    MCPBrowserWidget .browser-title {
-        height: 1;
-        text-style: bold;
-        color: $primary;
-    }
-    MCPBrowserWidget .browser-rows {
-        height: auto;
-        max-height: 20;
-        overflow-y: auto;
-    }
-    MCPBrowserWidget .browser-row {
-        height: 1;
-        padding: 0 1;
-    }
-    MCPBrowserWidget .browser-row-selected {
-        background: $primary;
-        text-style: bold;
-    }
-    MCPBrowserWidget .browser-help {
-        height: 1;
-        color: $text-muted;
-        text-style: italic;
-    }
-    """
-
-    BINDINGS: ClassVar[list[BindingType]] = [
-        Binding("up", "move_up", "Up", show=False),
-        Binding("k", "move_up", "Up", show=False),
-        Binding("down", "move_down", "Down", show=False),
-        Binding("j", "move_down", "Down", show=False),
-        Binding("enter", "confirm", "Confirm", show=False),
-        Binding("space", "toggle", "Toggle", show=False),
-        Binding("escape", "cancel", "Cancel", show=False),
-    ]
+    _INSTALLED_SUFFIX: ClassVar[str] = "  (configured)"
+    _PHASE2_TITLE: ClassVar[str] = "Select MCP servers to install"
+    _PHASE2_CONFIRM_LABEL: ClassVar[str] = "install"
 
     class Confirmed(Message):
         """Posted when user confirms server selection."""
@@ -90,242 +38,17 @@ class MCPBrowserWidget(Widget):
     class Cancelled(Message):
         """Posted when user cancels."""
 
-    def __init__(
-        self,
-        servers: list[MCPServerEntry],
-        installed_names: set[str],
-        *,
-        pre_filter_tag: str = "",
-        **kwargs: Any,
-    ) -> None:
-        super().__init__(**kwargs)
-        self._servers = servers
-        self._installed_names = installed_names
-        self._pre_filter_tag = pre_filter_tag.lower()
-        self._selected = 0
-        self._row_widgets: list[Static] = []
-        self._title_widget: Static | None = None
-        self._help_widget: Static | None = None
+    def _item_name(self, item: Any) -> str:
+        return item.name
 
-        # Phase 1: tag picker
-        # Phase 2: server checkbox
-        self._phase: int = 1
-        self._tag_items: list[tuple[str, int]] = []
-        self._server_items: list[MCPServerEntry] = []
-        self._checked: set[int] = set()
+    def _item_tags(self, item: Any) -> list[str]:
+        return item.tags
 
-        # Build tag list
-        from collections import Counter
+    def _item_desc(self, item: Any) -> str:
+        return item.description or item.label
 
-        tag_counter: Counter[str] = Counter()
-        for s in self._servers:
-            for t in s.tags:
-                tag_counter[t.lower()] += 1
-        sorted_tags = sorted(tag_counter.items(), key=lambda x: (-x[1], x[0]))
-        self._tag_items = [("all", len(self._servers)), *sorted_tags]
+    def _post_confirmed(self, items: list[Any]) -> None:
+        self.post_message(self.Confirmed(items))
 
-        # If pre-filtered, skip to phase 2
-        if self._pre_filter_tag:
-            self._server_items = [
-                s
-                for s in self._servers
-                if self._pre_filter_tag in [t.lower() for t in s.tags]
-            ]
-            if self._server_items:
-                self._phase = 2
-            else:
-                self._pre_filter_tag = ""
-
-    def compose(self) -> ComposeResult:
-        self._title_widget = Static("", classes="browser-title")
-        yield self._title_widget
-        with Container(classes="browser-rows"):
-            max_rows = max(len(self._tag_items), len(self._servers))
-            for _ in range(max_rows):
-                widget = Static("", classes="browser-row")
-                self._row_widgets.append(widget)
-                yield widget
-        self._help_widget = Static("", classes="browser-help")
-        yield self._help_widget
-
-    def on_mount(self) -> None:
-        self.call_after_refresh(self._update_display)
-        self.call_later(self.focus)
-
-    def _update_display(self) -> None:
-        if self._phase == 1:
-            self._render_tag_picker()
-        else:
-            self._render_server_checkbox()
-
-    def _render_tag_picker(self) -> None:
-        if self._title_widget:
-            self._title_widget.update("Filter by tag:")
-        if self._help_widget:
-            self._help_widget.update(
-                "\u2191/\u2193 navigate \u00b7 Enter select \u00b7 Esc cancel"
-            )
-
-        for i, widget in enumerate(self._row_widgets):
-            if i < len(self._tag_items):
-                tag, count = self._tag_items[i]
-                is_selected = i == self._selected
-                text = Text()
-                cursor = "\u25b8 " if is_selected else "  "
-                text.append(cursor, style="bold cyan" if is_selected else "dim")
-                label = f"{tag} ({count})"
-                text.append(label, style="bold" if is_selected else "")
-                widget.update(text)
-                widget.display = True
-                widget.remove_class("browser-row-selected")
-                if is_selected:
-                    widget.add_class("browser-row-selected")
-                    widget.scroll_visible()
-            else:
-                widget.update("")
-                widget.display = False
-
-    def _row_content_width(self) -> int:
-        try:
-            w = self.size.width
-            if w > 0:
-                return w - 6
-        except Exception:
-            pass
-        try:
-            return self.app.size.width - 10
-        except Exception:
-            return 100
-
-    def _truncate(self, desc: str, name: str, *, suffix: str = "") -> str:
-        overhead = 2 + 2 + len(name) + 3 + len(suffix)
-        max_len = max(20, self._row_content_width() - overhead)
-        if len(desc) <= max_len:
-            return desc
-        return desc[: max_len - 1] + "\u2026"
-
-    def _render_server_checkbox(self) -> None:
-        n_checked = len(
-            [
-                i
-                for i in self._checked
-                if self._server_items[i].name not in self._installed_names
-            ]
-        )
-        if self._title_widget:
-            self._title_widget.update(
-                f"Select MCP servers to install ({n_checked} selected):"
-            )
-        if self._help_widget:
-            self._help_widget.update(
-                "\u2191/\u2193 navigate \u00b7 Space toggle \u00b7 Enter install \u00b7 Esc cancel"
-            )
-
-        for i, widget in enumerate(self._row_widgets):
-            if i < len(self._server_items):
-                entry = self._server_items[i]
-                is_selected = i == self._selected
-                is_installed = entry.name in self._installed_names
-                is_checked = i in self._checked
-
-                text = Text()
-                cursor = "\u25b8 " if is_selected else "  "
-                text.append(cursor, style="bold cyan" if is_selected else "dim")
-
-                desc = entry.description or entry.label
-
-                if is_installed:
-                    suffix = "  (configured)"
-                    desc = self._truncate(desc, entry.name, suffix=suffix)
-                    text.append("\u2713 ", style="green")
-                    text.append(entry.name, style="green dim")
-                    text.append(f" \u2014 {desc}", style="dim")
-                    text.append(suffix, style="dim italic")
-                elif is_checked:
-                    desc = self._truncate(desc, entry.name)
-                    text.append("\u25cf ", style="green bold")
-                    text.append(entry.name, style="bold")
-                    text.append(f" \u2014 {desc}", style="")
-                else:
-                    desc = self._truncate(desc, entry.name)
-                    text.append("\u25cb ", style="dim")
-                    text.append(entry.name, style="bold" if is_selected else "")
-                    text.append(f" \u2014 {desc}", style="dim")
-
-                widget.update(text)
-                widget.display = True
-                widget.remove_class("browser-row-selected")
-                if is_selected:
-                    widget.add_class("browser-row-selected")
-                    widget.scroll_visible()
-            else:
-                widget.update("")
-                widget.display = False
-
-    def _current_items_count(self) -> int:
-        if self._phase == 1:
-            return len(self._tag_items)
-        return len(self._server_items)
-
-    def action_move_up(self) -> None:
-        n = self._current_items_count()
-        if not n:
-            return
-        self._selected = (self._selected - 1) % n
-        self._update_display()
-
-    def action_move_down(self) -> None:
-        n = self._current_items_count()
-        if not n:
-            return
-        self._selected = (self._selected + 1) % n
-        self._update_display()
-
-    def action_toggle(self) -> None:
-        if self._phase != 2:
-            return
-        if not self._server_items:
-            return
-        entry = self._server_items[self._selected]
-        if entry.name in self._installed_names:
-            return
-        if self._selected in self._checked:
-            self._checked.discard(self._selected)
-        else:
-            self._checked.add(self._selected)
-        self._update_display()
-
-    def action_confirm(self) -> None:
-        if self._phase == 1:
-            if not self._tag_items:
-                return
-            tag, _ = self._tag_items[self._selected]
-            if tag == "all":
-                self._server_items = list(self._servers)
-            else:
-                self._server_items = [
-                    s for s in self._servers if tag in [t.lower() for t in s.tags]
-                ]
-            self._phase = 2
-            self._selected = 0
-            self._checked = set()
-            self._update_display()
-        else:
-            entries = [
-                self._server_items[i]
-                for i in sorted(self._checked)
-                if self._server_items[i].name not in self._installed_names
-            ]
-            self.post_message(self.Confirmed(entries))
-
-    def action_cancel(self) -> None:
-        if self._phase == 2 and not self._pre_filter_tag:
-            self._phase = 1
-            self._selected = 0
-            self._checked = set()
-            self._update_display()
-        else:
-            self.post_message(self.Cancelled())
-
-    def on_blur(self, event: events.Blur) -> None:
-        self.call_after_refresh(self.focus)
+    def _post_cancelled(self) -> None:
+        self.post_message(self.Cancelled())

EvoScientist/cli/widgets/model_picker.py

@@ -12,9 +12,10 @@ from rich.text import Text
 from textual.binding import Binding, BindingType
 from textual.containers import Container
 from textual.message import Message
-from textual.widget import Widget
 from textual.widgets import Input, Static
 
+from .picker_base import PickerWidgetBase, first_selectable_index, move_selection
+
 if TYPE_CHECKING:
     from textual import events
     from textual.app import ComposeResult
@@ -81,14 +82,13 @@ def _build_items(
     return items
 
 
-class ModelPickerWidget(Widget):
+class ModelPickerWidget(PickerWidgetBase):
     """Inline model picker -- mounts in chat, keyboard-driven.
 
     Posts ``Picked(name, provider)`` on Enter, ``Cancelled()`` on Esc.
     Type to filter models.
     """
 
-    can_focus = True
     # Required so the Custom Ollama ``Input`` child can hold focus when the
     # user is typing a model name.
     can_focus_children = True
@@ -188,22 +188,19 @@ class ModelPickerWidget(Widget):
         self._mode: Literal["list", "input"] = "list"
         self._custom_input: Input | None = None
 
+    @staticmethod
+    def _is_model(item: dict) -> bool:
+        return item["type"] == "model"
+
     def _first_model_index(self) -> int:
-        for i, item in enumerate(self._items):
-            if item["type"] == "model":
-                return i
-        return 0
+        return first_selectable_index(self._items, self._is_model)
 
     def _move(self, direction: int) -> None:
         if not self._items:
             return
-        i = (self._selected + direction) % len(self._items)
-        steps = 0
-        while self._items[i]["type"] != "model" and steps < len(self._items):
-            i = (i + direction) % len(self._items)
-            steps += 1
-        if self._items[i]["type"] == "model":
-            self._selected = i
+        new = move_selection(self._items, self._selected, direction, self._is_model)
+        if self._is_model(self._items[new]):
+            self._selected = new
             self._update_rows()
 
     def _rebuild(self) -> None:
@@ -251,10 +248,9 @@ class ModelPickerWidget(Widget):
             classes="picker-help",
         )
 
-    def on_mount(self) -> None:
+    def _refresh_view(self) -> None:
         self._update_rows()
         self._update_filter()
-        self.call_later(self.focus)
 
     def _update_filter(self) -> None:
         if self._filter_widget is not None:
@@ -273,8 +269,8 @@ class ModelPickerWidget(Widget):
         for i, (item, widget) in enumerate(
             zip(self._items, self._row_widgets, strict=False)
         ):
-            widget.remove_class("picker-row-selected")
             if item["type"] == "header":
+                widget.remove_class("picker-row-selected")
                 t = Text()
                 t.append("\u2500\u2500 ", style="bold cyan")
                 t.append(item["label"], style="bold cyan")
@@ -289,9 +285,7 @@ class ModelPickerWidget(Widget):
                     t.append(" *", style="bold green")
                 t.append(f"  ({item['provider']})", style="dim italic")
                 widget.update(t)
-                if is_selected:
-                    widget.add_class("picker-row-selected")
-                    widget.scroll_visible()
+                self.apply_row_highlight(widget, is_selected)
 
     def on_key(self, event: events.Key) -> None:
         # In input mode, the Input child owns printable keys + backspace.
@@ -350,11 +344,9 @@ class ModelPickerWidget(Widget):
             return
         self.post_message(self.Cancelled())
 
-    def on_blur(self, event: events.Blur) -> None:
+    def _should_refocus_on_blur(self) -> bool:
         # When the Input child has focus we must NOT steal it back.
-        if self._mode == "input":
-            return
-        self.call_after_refresh(self.focus)
+        return self._mode != "input"
 
     def on_input_submitted(self, event: Input.Submitted) -> None:
         """Safety net: Enter fired inside the Input widget rather than

EvoScientist/cli/widgets/panel_widget.py

@@ -0,0 +1,257 @@
+"""Panel widget — in-eval ``task()`` fan-out live view.
+
+Groups all sub-agent dispatches from a single ``code_interpreter`` eval into
+one bordered container, one row per dispatch. Each row shows the expert /
+subagent type, the label, a running elapsed timer, and a status dot that
+flips to ``ok``/``err`` on completion.
+
+Sourced from the ``custom`` stream events emitted by ``langchain_quickjs``
+(see ``.stream.emitter.panel_dispatch_start`` etc.). Keyed by ``eval_id``
+so parallel dispatches from the same eval appear stacked; distinct evals
+get distinct panels.
+"""
+
+from __future__ import annotations
+
+import time
+
+from rich.text import Text
+from textual.containers import Vertical
+from textual.widget import Widget
+from textual.widgets import Static
+
+from ..status_bar import SPINNER_FRAMES
+
+_ROW_LABEL_MAX_CHARS = 48
+
+
+class _DispatchRow(Widget):
+    """One row inside a PanelWidget — a single ``task()`` dispatch.
+
+    Subclasses ``Widget`` directly and overrides ``render()`` to build the
+    row's ``Text`` on demand. Earlier revisions subclassed ``Static`` (visual
+    stayed ``None`` past the first paint) and ``Vertical`` with an inner
+    ``Static`` (container-layout race); rendering from ``render()`` is the
+    standard pattern for single-line widgets and dodges both issues by
+    letting Textual manage the visual lifecycle itself.
+    """
+
+    DEFAULT_CSS = """
+    _DispatchRow {
+        height: 1;
+        width: 100%;
+    }
+    """
+
+    def __init__(self, subagent_type: str, label: str) -> None:
+        super().__init__()
+        self._subagent_type = subagent_type
+        self._label = label
+        self._started_at = time.monotonic()
+        self._status: str = "running"  # "running" | "ok" | "err"
+        self._duration_ms: int | None = None
+        self._error: str = ""
+        self._frame = 0
+
+    def render(self) -> Text:
+        line = Text()
+        if self._status == "running":
+            line.append(f"  {SPINNER_FRAMES[self._frame]} ", style="cyan")
+        elif self._status == "ok":
+            line.append("  \u2713 ", style="green")
+        else:
+            line.append("  \u2717 ", style="red")
+        line.append(f"{self._subagent_type} ", style="bold")
+        if self._label:
+            trimmed = self._label
+            if len(trimmed) > _ROW_LABEL_MAX_CHARS:
+                trimmed = trimmed[: _ROW_LABEL_MAX_CHARS - 1] + "\u2026"
+            line.append(f"\u2014 {trimmed} ", style="dim")
+        line.append(self._elapsed_display(), style="dim")
+        if self._status == "err" and self._error:
+            err = self._error.split("\n", 1)[0]
+            if len(err) > 60:
+                err = err[:59] + "\u2026"
+            line.append(f"  {err}", style="red")
+        return line
+
+    def tick(self) -> None:
+        if self._status == "running":
+            self._frame = (self._frame + 1) % len(SPINNER_FRAMES)
+        self.refresh()
+
+    def complete(self, duration_ms: int) -> None:
+        self._status = "ok"
+        self._duration_ms = duration_ms
+        self.refresh()
+
+    def fail(self, duration_ms: int, error: str) -> None:
+        self._status = "err"
+        self._duration_ms = duration_ms
+        self._error = error
+        self.refresh()
+
+    def _elapsed_display(self) -> str:
+        if self._duration_ms is not None:
+            secs = self._duration_ms / 1000.0
+        else:
+            secs = time.monotonic() - self._started_at
+        return f"{secs:5.1f}s"
+
+
+class PanelWidget(Vertical):
+    """Container for one eval's fan-out — bordered box, one row per dispatch."""
+
+    DEFAULT_CSS = """
+    PanelWidget {
+        height: auto;
+        margin: 0 0;
+    }
+    PanelWidget .panel-header {
+        height: auto;
+        color: #22d3ee;
+    }
+    PanelWidget .panel-rows {
+        height: auto;
+        padding: 0 0 0 2;
+    }
+    PanelWidget .panel-footer {
+        height: auto;
+        color: #22d3ee;
+    }
+    PanelWidget.--completed .panel-header {
+        color: #4ade80;
+    }
+    PanelWidget.--completed .panel-footer {
+        color: #4ade80;
+    }
+    """
+
+    def __init__(self, eval_id: str) -> None:
+        super().__init__()
+        self._eval_id = eval_id
+        self._rows: dict[str, _DispatchRow] = {}
+        self._timer_handle = None
+        self._is_active = True
+
+    @property
+    def eval_id(self) -> str:
+        return self._eval_id
+
+    @property
+    def dispatch_count(self) -> int:
+        return len(self._rows)
+
+    def compose(self):
+        yield Static("", classes="panel-header")
+        yield Vertical(classes="panel-rows")
+        yield Static("", classes="panel-footer")
+
+    def on_mount(self) -> None:
+        self._timer_handle = self.set_interval(0.1, self._tick)
+        self._render_header()
+        self._render_footer()
+
+    def _tick(self) -> None:
+        for row in self._rows.values():
+            row.tick()
+        self._render_header()
+
+    async def start_dispatch(
+        self, dispatch_id: str, subagent_type: str, label: str
+    ) -> None:
+        if dispatch_id in self._rows:
+            return
+        # Re-arm if the panel already finalized: a Promise.allSettled retry
+        # of a failed subset dispatches under the same eval_id, so a new
+        # start after _maybe_finalize() has stopped the timer must undo
+        # the three effects of finalize (latch, class, timer) or the new
+        # row's spinner/elapsed stay frozen and future completions never
+        # refresh the header.
+        if not self._is_active:
+            self._is_active = True
+            self.remove_class("--completed")
+            self._timer_handle = self.set_interval(0.1, self._tick)
+            self._render_footer()
+        row = _DispatchRow(subagent_type, label)
+        rows_container = self.query_one(".panel-rows", Vertical)
+        await rows_container.mount(row)
+        self._rows[dispatch_id] = row
+        self._render_header()
+
+    def complete_dispatch(self, dispatch_id: str, duration_ms: int) -> None:
+        row = self._rows.get(dispatch_id)
+        if row is not None:
+            row.complete(duration_ms)
+        self._maybe_finalize()
+
+    def fail_dispatch(self, dispatch_id: str, duration_ms: int, error: str) -> None:
+        row = self._rows.get(dispatch_id)
+        if row is not None:
+            row.fail(duration_ms, error)
+        self._maybe_finalize()
+
+    def finalize_running(self, reason: str = "interrupted") -> None:
+        """Fail all still-running rows with their measured elapsed time.
+
+        Called from the TUI turn-cleanup ``finally`` so the 100ms interval
+        timer stops when a turn is cancelled mid-dispatch — without this,
+        no reference remains to stop the timer once the outer scope exits.
+        """
+        if not self._is_active:
+            return
+        now = time.monotonic()
+        for dispatch_id, row in list(self._rows.items()):
+            if row._status == "running":
+                elapsed_ms = int((now - row._started_at) * 1000)
+                self.fail_dispatch(dispatch_id, elapsed_ms, reason)
+        # Cover the zero-running-rows case: cancel-during-mount can leave
+        # the timer armed with either no rows registered (first dispatch)
+        # or every registered row already terminal (allSettled retry).
+        # The loop skips both, so call _maybe_finalize unconditionally —
+        # it is a no-op once _is_active has flipped.
+        self._maybe_finalize()
+
+    def _maybe_finalize(self) -> None:
+        if not self._is_active:
+            return
+        if all(row._status != "running" for row in self._rows.values()):
+            self._is_active = False
+            if self._timer_handle is not None:
+                self._timer_handle.stop()
+                self._timer_handle = None
+            self.add_class("--completed")
+            self._render_header()
+            self._render_footer()
+
+    def _summary_counts(self) -> tuple[int, int, int]:
+        running = ok = err = 0
+        for row in self._rows.values():
+            if row._status == "running":
+                running += 1
+            elif row._status == "ok":
+                ok += 1
+            else:
+                err += 1
+        return running, ok, err
+
+    def _render_header(self) -> None:
+        header = self.query_one(".panel-header", Static)
+        running, ok, err = self._summary_counts()
+        line = Text()
+        if self._is_active:
+            line.append("\u250c \u25b6 Expert panel ", style="bold cyan")
+            line.append(
+                f"({running} running, {ok} done, {err} failed)", style="dim cyan"
+            )
+        else:
+            line.append("\u2713 Expert panel ", style="bold green")
+            line.append(f"({ok} done, {err} failed)", style="dim green")
+        header.update(line)
+
+    def _render_footer(self) -> None:
+        footer = self.query_one(".panel-footer", Static)
+        if self._is_active:
+            footer.update(Text("\u2514 running...", style="dim cyan"))
+        else:
+            footer.update(Text(""))

EvoScientist/cli/widgets/picker_base.py

@@ -0,0 +1,427 @@
+"""Shared engine for the TUI's keyboard-driven picker/browser widgets.
+
+Every inline picker (model picker, thread picker, skill/MCP browsers)
+follows the same pattern: a flat item list where some rows are
+selectable, a wrapping highlight cursor, Enter/Esc terminal messages,
+and focus trapped inside the widget until a decision is made.  This
+module owns that machinery so the widgets only provide their data
+model and row rendering.
+
+Subclassing contract: Textual dispatches same-named handlers at EVERY
+level of the MRO, so subclasses must NOT define ``on_mount``/``on_blur``
+— they implement the ``_refresh_view()`` hook (and override
+``_should_refocus_on_blur()`` if focus may legitimately leave, e.g. a
+child ``Input``).  Message classes (``Picked``/``Confirmed``/
+``Cancelled``) stay defined in each widget: their handler names
+(``on_<widget>_<message>``) derive from the defining class.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Callable
+from typing import TYPE_CHECKING, Any, ClassVar
+
+from rich.text import Text
+from textual.binding import Binding, BindingType
+from textual.containers import Container
+from textual.widget import Widget
+from textual.widgets import Static
+
+if TYPE_CHECKING:
+    from textual import events
+    from textual.app import ComposeResult
+
+# ---------------------------------------------------------------------------
+# Pure selection helpers
+# ---------------------------------------------------------------------------
+
+
+def first_selectable_index(
+    items: list[Any], is_selectable: Callable[[Any], bool]
+) -> int:
+    """Index of the first selectable item, or 0 when none qualifies."""
+    for i, item in enumerate(items):
+        if is_selectable(item):
+            return i
+    return 0
+
+
+def move_selection(
+    items: list[Any],
+    current: int,
+    direction: int,
+    is_selectable: Callable[[Any], bool],
+) -> int:
+    """Next selectable index from *current*, wrapping around the list.
+
+    Non-selectable rows (headers, separators) are skipped; when no
+    selectable row exists the *current* index is returned unchanged.
+    """
+    if not items:
+        return current
+    i = (current + direction) % len(items)
+    steps = 0
+    while not is_selectable(items[i]) and steps < len(items):
+        i = (i + direction) % len(items)
+        steps += 1
+    return i if is_selectable(items[i]) else current
+
+
+# ---------------------------------------------------------------------------
+# Widget base
+# ---------------------------------------------------------------------------
+
+
+class PickerWidgetBase(Widget):
+    """Focus-trapped inline picker: mount-focus, blur-refocus, row
+    highlight bookkeeping and description truncation."""
+
+    can_focus = True
+    can_focus_children = False
+
+    def _refresh_view(self) -> None:
+        """Render the current state into the row widgets."""
+        raise NotImplementedError
+
+    def _should_refocus_on_blur(self) -> bool:
+        """Whether blur should snap focus back (focus trap)."""
+        return True
+
+    def on_mount(self) -> None:
+        # Deferred so self.size is populated for width-aware rendering.
+        self.call_after_refresh(self._refresh_view)
+        self.call_later(self.focus)
+
+    def on_blur(self, event: events.Blur) -> None:
+        if self._should_refocus_on_blur():
+            self.call_after_refresh(self.focus)
+
+    @staticmethod
+    def apply_row_highlight(
+        widget: Static, selected: bool, css_class: str = "picker-row-selected"
+    ) -> None:
+        """Toggle the selected-row CSS class and keep the row in view."""
+        widget.remove_class(css_class)
+        if selected:
+            widget.add_class(css_class)
+            widget.scroll_visible()
+
+    def _row_content_width(self) -> int:
+        """Usable character width for a row's text content (accounts for
+        widget border/padding; falls back to terminal width pre-layout)."""
+        try:
+            w = self.size.width
+            if w > 0:
+                # border (2) + widget padding (2) + row padding (2)
+                return w - 6
+        except Exception:
+            pass
+        try:
+            return self.app.size.width - 10
+        except Exception:
+            return 100
+
+    def _truncate(self, desc: str, name: str, *, suffix: str = "") -> str:
+        """Truncate a description to fit the row, adding ellipsis."""
+        # cursor(2) + indicator(2) + name + " — "(3) + suffix
+        overhead = 2 + 2 + len(name) + 3 + len(suffix)
+        max_len = max(20, self._row_content_width() - overhead)
+        if len(desc) <= max_len:
+            return desc
+        return desc[: max_len - 1] + "…"
+
+
+# ---------------------------------------------------------------------------
+# Two-phase tag-filter → checkbox browser
+# ---------------------------------------------------------------------------
+
+
+class TagCheckboxBrowserBase(PickerWidgetBase):
+    """Two-phase multi-select browser shared by the skill and MCP browsers.
+
+    Phase 1 — tag picker (Enter selects a tag filter, "all" included).
+    Phase 2 — checkbox list (Space toggles, Enter confirms, Esc goes back
+    to phase 1 unless the widget was constructed pre-filtered).
+
+    Subclasses provide the data adapters (``_item_name`` / ``_item_tags``
+    / ``_item_desc``), the phase-2 texts, and ``_post_confirmed()``.
+    """
+
+    DEFAULT_CSS = """
+    TagCheckboxBrowserBase {
+        height: auto;
+        max-height: 30;
+        margin: 1 0;
+        padding: 0 1;
+        background: $surface;
+        border: solid $primary;
+    }
+    TagCheckboxBrowserBase .browser-title {
+        height: 1;
+        text-style: bold;
+        color: $primary;
+    }
+    TagCheckboxBrowserBase .browser-rows {
+        height: auto;
+        max-height: 20;
+        overflow-y: auto;
+    }
+    TagCheckboxBrowserBase .browser-row {
+        height: 1;
+        padding: 0 1;
+    }
+    TagCheckboxBrowserBase .browser-row-selected {
+        background: $primary;
+        text-style: bold;
+    }
+    TagCheckboxBrowserBase .browser-help {
+        height: 1;
+        color: $text-muted;
+        text-style: italic;
+    }
+    """
+
+    BINDINGS: ClassVar[list[BindingType]] = [
+        Binding("up", "move_up", "Up", show=False),
+        Binding("k", "move_up", "Up", show=False),
+        Binding("down", "move_down", "Down", show=False),
+        Binding("j", "move_down", "Down", show=False),
+        Binding("enter", "confirm", "Confirm", show=False),
+        Binding("space", "toggle", "Toggle", show=False),
+        Binding("escape", "cancel", "Cancel", show=False),
+    ]
+
+    # -- subclass adapters --------------------------------------------
+
+    _INSTALLED_SUFFIX: ClassVar[str] = "  (installed)"
+    _PHASE2_TITLE: ClassVar[str] = "Select items to install"
+    _PHASE2_CONFIRM_LABEL: ClassVar[str] = "install"
+
+    def _item_name(self, item: Any) -> str:
+        raise NotImplementedError
+
+    def _item_tags(self, item: Any) -> list[str]:
+        raise NotImplementedError
+
+    def _item_desc(self, item: Any) -> str:
+        raise NotImplementedError
+
+    def _post_confirmed(self, items: list[Any]) -> None:
+        """Post the widget-specific ``Confirmed`` message."""
+        raise NotImplementedError
+
+    def _post_cancelled(self) -> None:
+        """Post the widget-specific ``Cancelled`` message."""
+        raise NotImplementedError
+
+    # -- state ---------------------------------------------------------
+
+    def __init__(
+        self,
+        items: list[Any],
+        installed_names: set[str],
+        *,
+        pre_filter_tag: str = "",
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(**kwargs)
+        self._all_items = items
+        self._installed_names = installed_names
+        self._pre_filter_tag = pre_filter_tag.lower()
+        self._selected = 0
+        self._row_widgets: list[Static] = []
+        self._title_widget: Static | None = None
+        self._help_widget: Static | None = None
+
+        self._phase: int = 1
+        self._filtered_items: list[Any] = []
+        self._checked: set[int] = set()
+
+        # Build tag list (sorted by count desc, then alphabetically)
+        from collections import Counter
+
+        tag_counter: Counter[str] = Counter()
+        for item in self._all_items:
+            for t in self._item_tags(item):
+                tag_counter[t.lower()] += 1
+        sorted_tags = sorted(tag_counter.items(), key=lambda x: (-x[1], x[0]))
+        self._tag_items: list[tuple[str, int]] = [
+            ("all", len(self._all_items)),
+            *sorted_tags,
+        ]
+
+        # If pre-filtered, skip to phase 2
+        if self._pre_filter_tag:
+            self._filtered_items = self._items_with_tag(self._pre_filter_tag)
+            if self._filtered_items:
+                self._phase = 2
+            else:
+                self._pre_filter_tag = ""
+
+    def _items_with_tag(self, tag: str) -> list[Any]:
+        if tag == "all":
+            return list(self._all_items)
+        return [
+            item
+            for item in self._all_items
+            if tag in [t.lower() for t in self._item_tags(item)]
+        ]
+
+    # -- layout ---------------------------------------------------------
+
+    def compose(self) -> ComposeResult:
+        self._title_widget = Static("", classes="browser-title")
+        yield self._title_widget
+        with Container(classes="browser-rows"):
+            max_rows = max(len(self._tag_items), len(self._all_items))
+            for _ in range(max_rows):
+                widget = Static("", classes="browser-row")
+                self._row_widgets.append(widget)
+                yield widget
+        self._help_widget = Static("", classes="browser-help")
+        yield self._help_widget
+
+    # -- rendering ------------------------------------------------------
+
+    def _refresh_view(self) -> None:
+        if self._phase == 1:
+            self._render_tag_picker()
+        else:
+            self._render_checkbox_list()
+
+    def _render_tag_picker(self) -> None:
+        if self._title_widget:
+            self._title_widget.update("Filter by tag:")
+        if self._help_widget:
+            self._help_widget.update("↑/↓ navigate · Enter select · Esc cancel")
+
+        for i, widget in enumerate(self._row_widgets):
+            if i < len(self._tag_items):
+                tag, count = self._tag_items[i]
+                is_selected = i == self._selected
+                text = Text()
+                cursor = "▸ " if is_selected else "  "
+                text.append(cursor, style="bold cyan" if is_selected else "dim")
+                text.append(f"{tag} ({count})", style="bold" if is_selected else "")
+                widget.update(text)
+                widget.display = True
+                self.apply_row_highlight(widget, is_selected, "browser-row-selected")
+            else:
+                widget.update("")
+                widget.display = False
+
+    def _render_checkbox_list(self) -> None:
+        n_checked = len(
+            [
+                i
+                for i in self._checked
+                if self._item_name(self._filtered_items[i]) not in self._installed_names
+            ]
+        )
+        if self._title_widget:
+            self._title_widget.update(f"{self._PHASE2_TITLE} ({n_checked} selected):")
+        if self._help_widget:
+            self._help_widget.update(
+                "↑/↓ navigate · Space toggle · "
+                f"Enter {self._PHASE2_CONFIRM_LABEL} · Esc cancel"
+            )
+
+        for i, widget in enumerate(self._row_widgets):
+            if i < len(self._filtered_items):
+                item = self._filtered_items[i]
+                name = self._item_name(item)
+                is_selected = i == self._selected
+                is_installed = name in self._installed_names
+                is_checked = i in self._checked
+
+                text = Text()
+                cursor = "▸ " if is_selected else "  "
+                text.append(cursor, style="bold cyan" if is_selected else "dim")
+
+                if is_installed:
+                    suffix = self._INSTALLED_SUFFIX
+                    desc = self._truncate(self._item_desc(item), name, suffix=suffix)
+                    text.append("✓ ", style="green")
+                    text.append(name, style="green dim")
+                    text.append(f" — {desc}", style="dim")
+                    text.append(suffix, style="dim italic")
+                elif is_checked:
+                    desc = self._truncate(self._item_desc(item), name)
+                    text.append("● ", style="green bold")
+                    text.append(name, style="bold")
+                    text.append(f" — {desc}", style="")
+                else:
+                    desc = self._truncate(self._item_desc(item), name)
+                    text.append("○ ", style="dim")
+                    text.append(name, style="bold" if is_selected else "")
+                    text.append(f" — {desc}", style="dim")
+
+                widget.update(text)
+                widget.display = True
+                self.apply_row_highlight(widget, is_selected, "browser-row-selected")
+            else:
+                widget.update("")
+                widget.display = False
+
+    # -- actions ----------------------------------------------------------
+
+    def _current_items_count(self) -> int:
+        if self._phase == 1:
+            return len(self._tag_items)
+        return len(self._filtered_items)
+
+    def action_move_up(self) -> None:
+        n = self._current_items_count()
+        if not n:
+            return
+        self._selected = (self._selected - 1) % n
+        self._refresh_view()
+
+    def action_move_down(self) -> None:
+        n = self._current_items_count()
+        if not n:
+            return
+        self._selected = (self._selected + 1) % n
+        self._refresh_view()
+
+    def action_toggle(self) -> None:
+        """Toggle checkbox selection (phase 2 only)."""
+        if self._phase != 2 or not self._filtered_items:
+            return
+        if self._item_name(self._filtered_items[self._selected]) in (
+            self._installed_names
+        ):
+            return  # Can't toggle already-installed items
+        if self._selected in self._checked:
+            self._checked.discard(self._selected)
+        else:
+            self._checked.add(self._selected)
+        self._refresh_view()
+
+    def action_confirm(self) -> None:
+        if self._phase == 1:
+            if not self._tag_items:
+                return
+            tag, _ = self._tag_items[self._selected]
+            self._filtered_items = self._items_with_tag(tag)
+            self._phase = 2
+            self._selected = 0
+            self._checked = set()
+            self._refresh_view()
+        else:
+            items = [
+                self._filtered_items[i]
+                for i in sorted(self._checked)
+                if self._item_name(self._filtered_items[i]) not in self._installed_names
+            ]
+            self._post_confirmed(items)
+
+    def action_cancel(self) -> None:
+        if self._phase == 2 and not self._pre_filter_tag:
+            # Go back to tag picker
+            self._phase = 1
+            self._selected = 0
+            self._checked = set()
+            self._refresh_view()
+        else:
+            self._post_cancelled()

EvoScientist/cli/widgets/skill_browser.py

@@ -1,6 +1,7 @@
 """Inline skill browser widget for /evoskills in TUI.
 
-Two-phase keyboard-driven widget:
+Two-phase keyboard-driven widget built on the shared picker engine
+(``picker_base.TagCheckboxBrowserBase``):
   Phase 1 — tag picker (arrow keys + Enter to select, or Esc for all)
   Phase 2 — skill checkbox (arrow keys to navigate, Space to toggle, Enter to confirm)
 
@@ -10,73 +11,19 @@ or ``SkillBrowserWidget.Cancelled`` on Esc.
 
 from __future__ import annotations
 
-from typing import TYPE_CHECKING, Any, ClassVar
+from typing import Any, ClassVar
 
-from rich.text import Text
-from textual.binding import Binding, BindingType
-from textual.containers import Container
 from textual.message import Message
-from textual.widget import Widget
-from textual.widgets import Static
 
-if TYPE_CHECKING:
-    from textual import events
-    from textual.app import ComposeResult
+from .picker_base import TagCheckboxBrowserBase
 
 
-class SkillBrowserWidget(Widget):
-    """Inline skill browser — mounts in chat, keyboard-driven.
+class SkillBrowserWidget(TagCheckboxBrowserBase):
+    """Inline skill browser — mounts in chat, keyboard-driven."""
 
-    Phase 1: Tag picker (select a tag filter or "All").
-    Phase 2: Skill checkbox (toggle skills, confirm to install).
-    """
-
-    can_focus = True
-    can_focus_children = False
-
-    DEFAULT_CSS = """
-    SkillBrowserWidget {
-        height: auto;
-        max-height: 30;
-        margin: 1 0;
-        padding: 0 1;
-        background: $surface;
-        border: solid $primary;
-    }
-    SkillBrowserWidget .browser-title {
-        height: 1;
-        text-style: bold;
-        color: $primary;
-    }
-    SkillBrowserWidget .browser-rows {
-        height: auto;
-        max-height: 20;
-        overflow-y: auto;
-    }
-    SkillBrowserWidget .browser-row {
-        height: 1;
-        padding: 0 1;
-    }
-    SkillBrowserWidget .browser-row-selected {
-        background: $primary;
-        text-style: bold;
-    }
-    SkillBrowserWidget .browser-help {
-        height: 1;
-        color: $text-muted;
-        text-style: italic;
-    }
-    """
-
-    BINDINGS: ClassVar[list[BindingType]] = [
-        Binding("up", "move_up", "Up", show=False),
-        Binding("k", "move_up", "Up", show=False),
-        Binding("down", "move_down", "Down", show=False),
-        Binding("j", "move_down", "Down", show=False),
-        Binding("enter", "confirm", "Confirm", show=False),
-        Binding("space", "toggle", "Toggle", show=False),
-        Binding("escape", "cancel", "Cancel", show=False),
-    ]
+    _INSTALLED_SUFFIX: ClassVar[str] = "  (installed)"
+    _PHASE2_TITLE: ClassVar[str] = "Select skills to install"
+    _PHASE2_CONFIRM_LABEL: ClassVar[str] = "install"
 
     class Confirmed(Message):
         """Posted when user confirms skill selection."""
@@ -88,261 +35,17 @@ class SkillBrowserWidget(Widget):
     class Cancelled(Message):
         """Posted when user cancels."""
 
-    def __init__(
-        self,
-        index: list[dict],
-        installed_names: set[str],
-        *,
-        pre_filter_tag: str = "",
-        **kwargs: Any,
-    ) -> None:
-        super().__init__(**kwargs)
-        self._index = index
-        self._installed_names = installed_names
-        self._pre_filter_tag = pre_filter_tag.lower()
-        self._selected = 0
-        self._row_widgets: list[Static] = []
-        self._title_widget: Static | None = None
-        self._help_widget: Static | None = None
+    def _item_name(self, item: Any) -> str:
+        return item["name"]
 
-        # Phase 1: tag picker
-        # Phase 2: skill checkbox
-        self._phase: int = 1
-        self._tag_items: list[tuple[str, int]] = []  # (tag, count)
-        self._skill_items: list[dict] = []  # filtered skills
-        self._checked: set[int] = set()  # indices of checked skills
+    def _item_tags(self, item: Any) -> list[str]:
+        return item.get("tags", [])
 
-        # Build tag list (sorted by count desc, then alphabetically)
-        from collections import Counter
+    def _item_desc(self, item: Any) -> str:
+        return item["description"]
 
-        tag_counter: Counter[str] = Counter()
-        for s in self._index:
-            for t in s.get("tags", []):
-                tag_counter[t.lower()] += 1
-        sorted_tags = sorted(tag_counter.items(), key=lambda x: (-x[1], x[0]))
-        self._tag_items = [("all", len(self._index)), *sorted_tags]
+    def _post_confirmed(self, items: list[Any]) -> None:
+        self.post_message(self.Confirmed([s["install_source"] for s in items]))
 
-        # If pre-filtered, skip to phase 2
-        if self._pre_filter_tag:
-            self._skill_items = [
-                s
-                for s in self._index
-                if self._pre_filter_tag in [t.lower() for t in s.get("tags", [])]
-            ]
-            if self._skill_items:
-                self._phase = 2
-            else:
-                # No matches — show tag picker anyway
-                self._pre_filter_tag = ""
-
-    def compose(self) -> ComposeResult:
-        self._title_widget = Static("", classes="browser-title")
-        yield self._title_widget
-        with Container(classes="browser-rows"):
-            # Pre-allocate enough rows for the larger of tag list or skill list
-            max_rows = max(len(self._tag_items), len(self._index))
-            for _ in range(max_rows):
-                widget = Static("", classes="browser-row")
-                self._row_widgets.append(widget)
-                yield widget
-        self._help_widget = Static("", classes="browser-help")
-        yield self._help_widget
-
-    def on_mount(self) -> None:
-        # Defer rendering until after layout so self.size is populated
-        self.call_after_refresh(self._update_display)
-        self.call_later(self.focus)
-
-    def _update_display(self) -> None:
-        if self._phase == 1:
-            self._render_tag_picker()
-        else:
-            self._render_skill_checkbox()
-
-    def _render_tag_picker(self) -> None:
-        if self._title_widget:
-            self._title_widget.update("Filter by tag:")
-        if self._help_widget:
-            self._help_widget.update("↑/↓ navigate · Enter select · Esc cancel")
-
-        for i, widget in enumerate(self._row_widgets):
-            if i < len(self._tag_items):
-                tag, count = self._tag_items[i]
-                is_selected = i == self._selected
-                text = Text()
-                cursor = "▸ " if is_selected else "  "
-                text.append(cursor, style="bold cyan" if is_selected else "dim")
-                label = f"{tag} ({count})"
-                text.append(label, style="bold" if is_selected else "")
-                widget.update(text)
-                widget.display = True
-                widget.remove_class("browser-row-selected")
-                if is_selected:
-                    widget.add_class("browser-row-selected")
-                    widget.scroll_visible()
-            else:
-                widget.update("")
-                widget.display = False
-
-    def _row_content_width(self) -> int:
-        """Get the usable character width for a row's text content.
-
-        Accounts for widget border, widget padding, and row padding.
-        Falls back to terminal width if the widget hasn't been laid out yet.
-        """
-        try:
-            w = self.size.width
-            if w > 0:
-                # border (2) + widget padding-left/right (2) + row padding-left/right (2)
-                return w - 6
-        except Exception:
-            pass
-        # Fallback: use terminal width minus reasonable chrome
-        try:
-            return self.app.size.width - 10
-        except Exception:
-            return 100
-
-    def _truncate(self, desc: str, name: str, *, suffix: str = "") -> str:
-        """Truncate a description to fit the row, adding ellipsis if needed."""
-        # cursor(2) + indicator(2) + name + " — "(3) + suffix
-        overhead = 2 + 2 + len(name) + 3 + len(suffix)
-        max_len = max(20, self._row_content_width() - overhead)
-        if len(desc) <= max_len:
-            return desc
-        return desc[: max_len - 1] + "…"
-
-    def _render_skill_checkbox(self) -> None:
-        n_checked = len(
-            [
-                i
-                for i in self._checked
-                if self._skill_items[i]["name"] not in self._installed_names
-            ]
-        )
-        if self._title_widget:
-            self._title_widget.update(
-                f"Select skills to install ({n_checked} selected):"
-            )
-        if self._help_widget:
-            self._help_widget.update(
-                "↑/↓ navigate · Space toggle · Enter install · Esc cancel"
-            )
-
-        for i, widget in enumerate(self._row_widgets):
-            if i < len(self._skill_items):
-                skill = self._skill_items[i]
-                is_selected = i == self._selected
-                is_installed = skill["name"] in self._installed_names
-                is_checked = i in self._checked
-
-                text = Text()
-                cursor = "▸ " if is_selected else "  "
-                text.append(cursor, style="bold cyan" if is_selected else "dim")
-
-                if is_installed:
-                    suffix = "  (installed)"
-                    desc = self._truncate(
-                        desc=skill["description"],
-                        name=skill["name"],
-                        suffix=suffix,
-                    )
-                    text.append("✓ ", style="green")
-                    text.append(skill["name"], style="green dim")
-                    text.append(f" — {desc}", style="dim")
-                    text.append(suffix, style="dim italic")
-                elif is_checked:
-                    desc = self._truncate(skill["description"], skill["name"])
-                    text.append("● ", style="green bold")
-                    text.append(skill["name"], style="bold")
-                    text.append(f" — {desc}", style="")
-                else:
-                    desc = self._truncate(skill["description"], skill["name"])
-                    text.append("○ ", style="dim")
-                    text.append(skill["name"], style="bold" if is_selected else "")
-                    text.append(f" — {desc}", style="dim")
-
-                widget.update(text)
-                widget.display = True
-                widget.remove_class("browser-row-selected")
-                if is_selected:
-                    widget.add_class("browser-row-selected")
-                    widget.scroll_visible()
-            else:
-                widget.update("")
-                widget.display = False
-
-    def _current_items_count(self) -> int:
-        if self._phase == 1:
-            return len(self._tag_items)
-        return len(self._skill_items)
-
-    def action_move_up(self) -> None:
-        n = self._current_items_count()
-        if not n:
-            return
-        self._selected = (self._selected - 1) % n
-        self._update_display()
-
-    def action_move_down(self) -> None:
-        n = self._current_items_count()
-        if not n:
-            return
-        self._selected = (self._selected + 1) % n
-        self._update_display()
-
-    def action_toggle(self) -> None:
-        """Toggle skill selection (phase 2 only)."""
-        if self._phase != 2:
-            return
-        if not self._skill_items:
-            return
-        skill = self._skill_items[self._selected]
-        if skill["name"] in self._installed_names:
-            return  # Can't toggle installed skills
-        if self._selected in self._checked:
-            self._checked.discard(self._selected)
-        else:
-            self._checked.add(self._selected)
-        self._update_display()
-
-    def action_confirm(self) -> None:
-        if self._phase == 1:
-            # Transition to phase 2
-            if not self._tag_items:
-                return
-            tag, _ = self._tag_items[self._selected]
-            if tag == "all":
-                self._skill_items = list(self._index)
-            else:
-                self._skill_items = [
-                    s
-                    for s in self._index
-                    if tag in [t.lower() for t in s.get("tags", [])]
-                ]
-            self._phase = 2
-            self._selected = 0
-            self._checked = set()
-            self._update_display()
-        else:
-            # Confirm selection
-            sources = [
-                self._skill_items[i]["install_source"]
-                for i in sorted(self._checked)
-                if self._skill_items[i]["name"] not in self._installed_names
-            ]
-            self.post_message(self.Confirmed(sources))
-
-    def action_cancel(self) -> None:
-        if self._phase == 2 and not self._pre_filter_tag:
-            # Go back to tag picker
-            self._phase = 1
-            self._selected = 0
-            self._checked = set()
-            self._update_display()
-        else:
-            self.post_message(self.Cancelled())
-
-    def on_blur(self, event: events.Blur) -> None:
-        """Re-focus to keep focus trapped until decision is made."""
-        self.call_after_refresh(self.focus)
+    def _post_cancelled(self) -> None:
+        self.post_message(self.Cancelled())

EvoScientist/cli/widgets/thread_selector.py

@@ -23,11 +23,11 @@ from rich.text import Text
 from textual.binding import Binding, BindingType
 from textual.containers import Container
 from textual.message import Message
-from textual.widget import Widget
 from textual.widgets import Static
 
+from .picker_base import PickerWidgetBase, first_selectable_index, move_selection
+
 if TYPE_CHECKING:
-    from textual import events
     from textual.app import ComposeResult
 
 
@@ -237,16 +237,13 @@ def build_row_text(
 # ---------------------------------------------------------------------------
 
 
-class ThreadPickerWidget(Widget):
+class ThreadPickerWidget(PickerWidgetBase):
     """Inline thread picker — mounts in chat, keyboard-driven.
 
     Posts ``Picked(thread_id)`` on Enter, ``Cancelled()`` on Esc.
     Threads are displayed in a two-level workspace hierarchy.
     """
 
-    can_focus = True
-    can_focus_children = False
-
     DEFAULT_CSS = """
     ThreadPickerWidget {
         height: auto;
@@ -323,22 +320,19 @@ class ThreadPickerWidget(Widget):
         self._selected = self._first_thread_index()
         self._row_widgets: list[Static] = []
 
+    @staticmethod
+    def _is_thread(item: dict) -> bool:
+        return item["type"] == "thread"
+
     def _first_thread_index(self) -> int:
-        for i, item in enumerate(self._items):
-            if item["type"] == "thread":
-                return i
-        return 0
+        return first_selectable_index(self._items, self._is_thread)
 
     def _move(self, direction: int) -> None:
         if not self._items:
             return
-        i = (self._selected + direction) % len(self._items)
-        steps = 0
-        while self._items[i]["type"] != "thread" and steps < len(self._items):
-            i = (i + direction) % len(self._items)
-            steps += 1
-        if self._items[i]["type"] == "thread":
-            self._selected = i
+        new = move_selection(self._items, self._selected, direction, self._is_thread)
+        if self._is_thread(self._items[new]):
+            self._selected = new
             self._update_rows()
 
     def compose(self) -> ComposeResult:
@@ -358,18 +352,18 @@ class ThreadPickerWidget(Widget):
             classes="picker-help",
         )
 
-    def on_mount(self) -> None:
+    def _refresh_view(self) -> None:
         self._update_rows()
-        self.call_later(self.focus)
 
     def _update_rows(self) -> None:
         for i, (item, widget) in enumerate(
             zip(self._items, self._row_widgets, strict=False)
         ):
-            widget.remove_class("picker-row-selected")
             if item["type"] == "header":
+                widget.remove_class("picker-row-selected")
                 widget.update(build_header_text(item["label"]))
             elif item["type"] == "subheader":
+                widget.remove_class("picker-row-selected")
                 widget.update(build_subheader_text(item["label"]))
             else:
                 thread = item["thread"]
@@ -381,9 +375,7 @@ class ThreadPickerWidget(Widget):
                     indented=item.get("indented", False),
                 )
                 widget.update(text)
-                if is_selected:
-                    widget.add_class("picker-row-selected")
-                    widget.scroll_visible()
+                self.apply_row_highlight(widget, is_selected)
 
     def action_move_up(self) -> None:
         self._move(-1)
@@ -403,6 +395,3 @@ class ThreadPickerWidget(Widget):
 
     def action_cancel(self) -> None:
         self.post_message(self.Cancelled())
-
-    def on_blur(self, event: events.Blur) -> None:
-        self.call_after_refresh(self.focus)

EvoScientist/commands/base.py

@@ -6,6 +6,7 @@ from typing import TYPE_CHECKING, Any, ClassVar, Protocol, runtime_checkable
 
 if TYPE_CHECKING:
     from ..gateway import GraphGateway
+    from ..runtime import AsyncRuntime
 
 
 @dataclass
@@ -65,20 +66,47 @@ class CommandUI(Protocol):
 
 @dataclass
 class ChannelRuntime:
-    """Mutable handle to the agent + thread bound to running channels."""
+    """Mutable handle to the agent + thread bound to running channels.
+
+    Also holds session-scoped bindings mutated by slash commands — the
+    ``active_teams`` list backs the ``/expert`` command, feeding into
+    ``RunRequest.configurable_extra`` at stream call time.
+    """
 
     agent: Any = None
     thread_id: str | None = None
+    active_teams: list[str] = field(default_factory=list)
 
     def bind(self, agent: Any, thread_id: str) -> None:
         self.agent = agent
         self.thread_id = thread_id
 
     def clear(self) -> None:
+        # ``active_teams`` is session-scoped and reset explicitly by ``/new``
+        # (session.py) and ``/expert clear`` — not tied to channel lifecycle.
+        # Clearing here on channel shutdown would silently dismiss the user's
+        # invited experts, which they never asked for.
         self.agent = None
         self.thread_id = None
 
 
+def active_teams_configurable_extra(
+    runtime: ChannelRuntime | None,
+) -> dict[str, Any] | None:
+    """Build ``RunRequest.configurable_extra`` from a channel runtime.
+
+    Returns ``{"active_teams": [...]}`` when the runtime has invited
+    experts, or ``None`` when there is no runtime or no active invites —
+    lets stream call sites forward the field unconditionally without
+    each duplicating the "read runtime slot, build dict, drop when
+    empty" three-liner.
+    """
+    if runtime is None:
+        return None
+    invited = list(runtime.active_teams)
+    return {"active_teams": invited} if invited else None
+
+
 @dataclass
 class CommandContext:
     """Context passed to commands during execution."""
@@ -91,6 +119,7 @@ class CommandContext:
     config: Any = None
     channel_runtime: ChannelRuntime | None = None
     graph_gateway: GraphGateway | None = None
+    async_runtime: AsyncRuntime | None = None
     command_error: str | None = None
     # Real LLM input token count from last usage_metadata (includes system
     # prompt + tool schemas).  Used by /compact for accurate display.

EvoScientist/commands/channel_ui.py

@@ -12,6 +12,8 @@ if TYPE_CHECKING:
 
 _logger = logging.getLogger(__name__)
 
+_COMMAND_OUTPUT_FAILURE_NOTICE = "Command output could not be delivered."
+
 
 class ChannelCommandUI(CommandUI):
     """CommandUI implementation for messaging channels with output buffering."""
@@ -37,6 +39,10 @@ class ChannelCommandUI(CommandUI):
         self.handle_session_resume_callback = handle_session_resume_callback
         self.graph_gateway = graph_gateway
         self._system_buffer: list[str] = []
+        # Whether any output was delivered (or scheduled for delivery) to the
+        # channel. The slash dispatcher consults this to decide between a
+        # bare completion ack and staying silent.
+        self.sent_to_channel: bool = False
 
     def _queue_system(
         self,
@@ -107,6 +113,7 @@ class ChannelCommandUI(CommandUI):
             content=grouped_text,
             reply_to=self.msg.message_id,
             metadata=self.msg.metadata,
+            failure_notice=_COMMAND_OUTPUT_FAILURE_NOTICE,
         )
 
         if self.msg.bus_ref:
@@ -114,6 +121,7 @@ class ChannelCommandUI(CommandUI):
         else:
             coro = self.msg.channel_ref.send(outbound)
 
+        self.sent_to_channel = True
         asyncio.run_coroutine_threadsafe(coro, loop)
 
     def mount_renderable(self, renderable: Any) -> None:
@@ -147,6 +155,7 @@ class ChannelCommandUI(CommandUI):
             content=f"```\n{text}\n```",
             reply_to=self.msg.message_id,
             metadata=self.msg.metadata,
+            failure_notice=_COMMAND_OUTPUT_FAILURE_NOTICE,
         )
 
         if self.msg.bus_ref:
@@ -154,6 +163,7 @@ class ChannelCommandUI(CommandUI):
         else:
             coro = self.msg.channel_ref.send(outbound)
 
+        self.sent_to_channel = True
         asyncio.run_coroutine_threadsafe(coro, loop)
 
     async def wait_for_thread_pick(

EvoScientist/commands/implementation/__init__.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 from . import (
     autoskills,
     channel,
+    experts,
     general,
     mcp,
     model,
@@ -15,6 +16,7 @@ from . import (
 __all__ = [
     "autoskills",
     "channel",
+    "experts",
     "general",
     "mcp",
     "model",

EvoScientist/commands/implementation/experts.py

@@ -0,0 +1,257 @@
+"""Slash commands for TUI expert-skill selection.
+
+``/experts`` — list installed expert skills.
+``/expert <name>`` — toggle an expert into the current session's
+``active_teams`` list; the next turn's ``configurable.active_teams`` picks
+this up and ``ActiveTeamMiddleware`` biases the main-agent's delegation
+toward the invited expert(s).
+``/expert clear`` — reset the list.
+
+User-facing verbs match the WebUI gallery: **invite** to add an expert,
+**dismiss** to remove one. Internal state field stays ``active_teams``
+for wire compatibility.
+
+Backing store is ``ChannelRuntime.active_teams`` (see
+``EvoScientist/commands/base.py``). WebUI users get the same effect via
+its gallery + langgraph-sdk ``config.configurable``; these commands are
+the TUI-side equivalent.
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, ClassVar
+
+from rich.table import Table
+
+from ..base import Argument, Command, CommandContext, SubCommand
+from ..manager import manager
+
+if TYPE_CHECKING:
+    from ...tools.skills_manager import SkillInfo
+
+_dispatchable_experts_cache: list[SkillInfo] | None = None
+
+
+def invalidate_experts_cache() -> None:
+    """Reset the /expert dispatchable-experts cache.
+
+    Called after ``install_skill`` / ``uninstall_skill`` mutations so a
+    freshly installed expert shows up in the /expert popup on the next
+    keystroke.
+    """
+    global _dispatchable_experts_cache
+    _dispatchable_experts_cache = None
+
+
+def _subscribe_cache_invalidation() -> None:
+    """Register with ``skills_manager`` so every install/uninstall path
+    (slash commands, agent ``skill_manager`` @tool, onboarding) busts
+    the /expert popup — no caller has to remember.
+    """
+    try:
+        from ...tools.skills_manager import register_skills_changed_callback
+
+        register_skills_changed_callback(invalidate_experts_cache)
+    except Exception:
+        # ``skills_manager`` not importable in some early-init contexts;
+        # cache staleness is a UX inconvenience, not a correctness bug.
+        pass
+
+
+_subscribe_cache_invalidation()
+
+
+def _dispatchable_experts() -> list[SkillInfo]:
+    """Cached list of experts that /expert can safely invite.
+
+    Filters ``list_expert_skills`` down to those that pass the same
+    empty-body + name-collision guards ``build_expert_subagent_specs``
+    and ``_fold_expert_subagents`` apply at agent-construction time, so
+    the /expert popup and invite-accept path only ever surface names
+    that will actually reach ``ActiveTeamMiddleware``'s cue.
+    """
+    global _dispatchable_experts_cache
+    if _dispatchable_experts_cache is None:
+        try:
+            from ...subagents.expert_container import list_dispatchable_experts
+
+            _dispatchable_experts_cache = list_dispatchable_experts(include_system=True)
+        except Exception:
+            return []
+    return _dispatchable_experts_cache
+
+
+class ExpertsCommand(Command):
+    """List installed expert skills."""
+
+    name: ClassVar[str] = "/experts"
+    description: ClassVar[str] = "List installed expert skills"
+    category: ClassVar[str] = "Experts"
+
+    async def execute(self, ctx: CommandContext, args: list[str]) -> None:
+        from ...tools.skills_manager import list_expert_skills
+
+        experts = list_expert_skills(include_system=True)
+        active = _current_active_teams(ctx)
+
+        if not experts:
+            ctx.ui.append_system("No expert skills installed.", style="dim")
+            ctx.ui.append_system(
+                "Install with: /install-skill <path-or-url>", style="dim"
+            )
+            return
+
+        table = Table(title=f"Expert Skills ({len(experts)})", show_header=True)
+        table.add_column("Name", style="cyan")
+        table.add_column("Role", style="dim")
+        table.add_column("Active", style="green")
+        for skill in experts:
+            marker = "*" if skill.name in active else ""
+            table.add_row(
+                skill.name,
+                skill.role or skill.description,
+                marker,
+            )
+        ctx.ui.mount_renderable(table)
+
+        if active:
+            ctx.ui.append_system(
+                f"Active: {', '.join(active)}. Toggle with `/expert <name>`, "
+                "clear with `/expert clear`.",
+                style="dim",
+            )
+        else:
+            ctx.ui.append_system(
+                "No experts invited. `/expert <name>` to invite one.",
+                style="dim",
+            )
+
+
+class ExpertCommand(Command):
+    """Invite, dismiss, or clear expert skills for the current thread."""
+
+    name: ClassVar[str] = "/expert"
+    description: ClassVar[str] = "Invite or dismiss an expert skill"
+    category: ClassVar[str] = "Experts"
+    arguments: ClassVar[list[Argument]] = [
+        Argument(
+            name="name_or_clear",
+            type=str,
+            description="Expert skill name to toggle, or 'clear' to reset",
+            required=True,
+        )
+    ]
+    subcommands: ClassVar[list[SubCommand]] = [
+        SubCommand("clear", "Dismiss all invited experts"),
+    ]
+
+    def _get_expert_candidates(self) -> list[tuple[str, str]]:
+        return [(s.name, s.role or s.description) for s in _dispatchable_experts()]
+
+    def get_completions(self, tokens: list[str]) -> list[tuple[str, str]]:
+        """Complete expert names + the ``clear`` subcommand."""
+        # /expert takes a single positional arg; anything past it (including a
+        # trailing space that turns tokens into ["name", ""]) has nothing to offer.
+        if len(tokens) > 1:
+            return []
+        prefix = tokens[0].lower() if tokens else ""
+        candidates = [
+            *self._get_expert_candidates(),
+            ("clear", "Dismiss all invited experts"),
+        ]
+        matches = [
+            (name, desc) for name, desc in candidates if name.lower().startswith(prefix)
+        ]
+        # Exact match — argument already complete, hide the popup.
+        if len(matches) == 1 and matches[0][0].lower() == prefix:
+            return []
+        return matches
+
+    async def execute(self, ctx: CommandContext, args: list[str]) -> None:
+        runtime = ctx.channel_runtime
+        if runtime is None:
+            ctx.ui.append_system(
+                "/expert requires a session runtime; not available in this context.",
+                style="yellow",
+            )
+            return
+
+        if not args:
+            ctx.ui.append_system(
+                "Usage: /expert <name>   toggle an expert into the invited list",
+                style="yellow",
+            )
+            ctx.ui.append_system(
+                "       /expert clear    dismiss all invited experts",
+                style="dim",
+            )
+            return
+
+        target = args[0].strip()
+        if target.lower() == "clear":
+            if not runtime.active_teams:
+                ctx.ui.append_system("No experts invited.", style="dim")
+                return
+            dismissed = list(runtime.active_teams)
+            runtime.active_teams = []
+            ctx.ui.append_system(
+                f"Dismissed experts: {', '.join(dismissed)}", style="dim"
+            )
+            return
+
+        # Completion matches case-insensitively; honour the same here by
+        # resolving a case-variant to the on-disk name before membership.
+        by_lower = {s.name.lower(): s.name for s in _dispatchable_experts()}
+        canonical = by_lower.get(target.lower())
+        if canonical is None:
+            from ...tools.skills_manager import list_expert_skills
+
+            installed = {
+                s.name.lower() for s in list_expert_skills(include_system=True)
+            }
+            if target.lower() not in installed:
+                ctx.ui.append_system(
+                    f"No expert skill named '{target}'. `/experts` lists "
+                    "installed ones.",
+                    style="red",
+                )
+            else:
+                ctx.ui.append_system(
+                    f"Expert '{target}' can't be dispatched (empty actor "
+                    "definition or name collision with a built-in sub-agent).",
+                    style="red",
+                )
+            return
+
+        if canonical in runtime.active_teams:
+            runtime.active_teams = [n for n in runtime.active_teams if n != canonical]
+            ctx.ui.append_system(f"Dismissed expert: {canonical}", style="dim")
+        else:
+            runtime.active_teams = [*runtime.active_teams, canonical]
+            ctx.ui.append_system(f"Invited expert: {canonical}", style="green")
+            # An expert installed mid-session: the background reach
+            # (``start_async_task``) resolves it on first dispatch, but the
+            # in-turn ``task`` reach is frozen into the running agent, so it
+            # needs a rebuilt agent. An expert installed before this session
+            # started is already inside that frozen set — its in-turn reach
+            # works without a rebuild — so the hint scopes the /new boundary
+            # to newly installed experts instead of stating it
+            # unconditionally.
+            ctx.ui.append_system(
+                "Newly installed experts: background dispatch is available "
+                "immediately; in-turn task dispatch needs /new.",
+                style="dim",
+            )
+        if runtime.active_teams:
+            ctx.ui.append_system(
+                f"Active: {', '.join(runtime.active_teams)}", style="dim"
+            )
+
+
+def _current_active_teams(ctx: CommandContext) -> list[str]:
+    runtime = ctx.channel_runtime
+    return list(runtime.active_teams) if runtime is not None else []
+
+
+manager.register(ExpertsCommand())
+manager.register(ExpertCommand())

EvoScientist/commands/implementation/mcp_install.py

@@ -37,7 +37,7 @@ class InstallMCPCommand(Command):
         try:
             import asyncio
 
-            servers = await asyncio.get_event_loop().run_in_executor(
+            servers = await asyncio.get_running_loop().run_in_executor(
                 None, fetch_marketplace_index
             )
         except Exception as e:

EvoScientist/commands/implementation/model.py

@@ -130,6 +130,7 @@ class ModelCommand(Command):
         *,
         save: bool = False,
     ) -> None:
+        import asyncio
         import copy
 
         from ...cli.agent import _load_agent
@@ -139,6 +140,7 @@ class ModelCommand(Command):
             set_active_config,
             set_chat_model_instance,
         )
+        from ...runtime import AsyncRuntime
 
         cfg = _ensure_config()
 
@@ -151,13 +153,26 @@ class ModelCommand(Command):
         temp_cfg.model = model_name
         temp_cfg.provider = provider
 
+        # Re-thread the session's frontend event sink so the rebuilt agent's
+        # middleware keeps driving the tool-selection widget / fallback notices
+        # after a /model switch (the sink lives on the gateway, not the agent).
+        events = ctx.graph_gateway.events
+
         try:
             new_chat_model = _build_chat_model(temp_cfg)
-            new_agent = _load_agent(
-                workspace_dir=ctx.workspace_dir,
-                checkpointer=ctx.checkpointer,
-                config=temp_cfg,
-                chat_model=new_chat_model,
+            load_kwargs = {
+                "workspace_dir": ctx.workspace_dir,
+                "checkpointer": ctx.checkpointer,
+                "config": temp_cfg,
+                "chat_model": new_chat_model,
+                "events": events,
+            }
+            async_runtime = getattr(ctx, "async_runtime", None)
+            if isinstance(async_runtime, AsyncRuntime):
+                load_kwargs["runtime"] = async_runtime
+            new_agent = await asyncio.to_thread(
+                _load_agent,
+                **load_kwargs,
             )
         except Exception as e:
             ctx.ui.append_system(f"Failed to switch model: {e}", style="red")

EvoScientist/commands/implementation/schedule.py

@@ -10,13 +10,21 @@ from ..base import Command, CommandContext, SubCommand
 from ..manager import manager
 
 
+def _clean(text: str) -> str:
+    """Trim a shlex-joined argument and drop a stray wrapping quote pair."""
+    return text.strip().strip('"').strip("'")
+
+
 class ScheduleCommand(Command):
     """Manage scheduled (cron) tasks."""
 
     name = "/schedule"
     description = "Manage scheduled (cron) tasks"
     subcommands: ClassVar[list[SubCommand]] = [
-        SubCommand("add", 'Add: /schedule add <m h dom mon dow> "<prompt>"'),
+        SubCommand(
+            "add",
+            'Add: /schedule add <m h dom mon dow> "<prompt>" [--rubric "<checklist>"]',
+        ),
         SubCommand("list", "List scheduled tasks"),
         SubCommand("remove", "Remove a schedule by id"),
         SubCommand("run", "Run a schedule's prompt once now (test)"),
@@ -78,10 +86,19 @@ class ScheduleCommand(Command):
             schedule, prompt_tokens = " ".join(rest[:5]), rest[5:]
         else:
             ctx.ui.append_system(
-                'Usage: /schedule add "<m h dom mon dow>" "<prompt>"', style="yellow"
+                'Usage: /schedule add "<m h dom mon dow>" "<prompt>" '
+                '[--rubric "<checklist>"]',
+                style="yellow",
             )
             return
-        prompt = " ".join(prompt_tokens).strip().strip('"').strip("'")
+        # Optional trailing acceptance checklist; everything after --rubric is it.
+        rubric = None
+        if "--rubric" in prompt_tokens:
+            # Last occurrence wins so an unquoted prompt may mention the flag.
+            split_at = len(prompt_tokens) - 1 - prompt_tokens[::-1].index("--rubric")
+            rubric = _clean(" ".join(prompt_tokens[split_at + 1 :])) or None
+            prompt_tokens = prompt_tokens[:split_at]
+        prompt = _clean(" ".join(prompt_tokens))
         if not prompt:
             ctx.ui.append_system("A task prompt is required.", style="yellow")
             return
@@ -90,7 +107,11 @@ class ScheduleCommand(Command):
         name = re.sub(r"[^a-z0-9]+", "-", raw).strip("-")[:32] or "task"
         try:
             rec = await asyncio.to_thread(
-                crons.create_schedule, name=name, schedule=schedule, prompt=prompt
+                crons.create_schedule,
+                name=name,
+                schedule=schedule,
+                prompt=prompt,
+                rubric=rubric,
             )
         except Exception as exc:
             ctx.ui.append_system(f"Error: {exc}", style="red")
@@ -119,6 +140,7 @@ class ScheduleCommand(Command):
         table.add_column("Schedule", style="green")
         table.add_column("Enabled", style="yellow")
         table.add_column("Next run (UTC)", style="white")
+        table.add_column("Rubric", style="blue")
         for r in rows:
             meta = r.get("metadata") or {}
             table.add_row(
@@ -127,6 +149,7 @@ class ScheduleCommand(Command):
                 str(r.get("schedule", "")),
                 "yes" if r.get("enabled", True) else "no",
                 str(r.get("next_run_date", "")),
+                "yes" if meta.get("rubric") else "",
             )
         ctx.ui.mount_renderable(table)
 
@@ -191,7 +214,8 @@ class ScheduleCommand(Command):
         match = await self._resolve_or_report(ctx, crons, prefix)
         if match is None:
             return
-        prompt = (match.get("metadata") or {}).get("prompt", "")
+        meta = match.get("metadata") or {}
+        prompt = meta.get("prompt", "")
         if not str(prompt).strip():
             ctx.ui.append_system(
                 f"Schedule {prefix} has no stored prompt — cannot run it.",
@@ -199,7 +223,9 @@ class ScheduleCommand(Command):
             )
             return
         try:
-            rec = await asyncio.to_thread(crons.run_now, prompt)
+            rec = await asyncio.to_thread(
+                crons.run_now, prompt, rubric=meta.get("rubric") or None
+            )
         except Exception as exc:
             ctx.ui.append_system(f"Error: {exc}", style="red")
             return

EvoScientist/commands/implementation/session.py

@@ -182,8 +182,21 @@ class ResumeCommand(Command):
         if restored_workspace:
             ctx.workspace_dir = restored_workspace
 
+        switched_thread = resolved != ctx.thread_id
         ctx.thread_id = resolved
 
+        # Invitations are session-scoped (see ChannelRuntime.active_teams);
+        # resuming a different thread is a session switch, so release them —
+        # uniform with /new. Resuming the current thread keeps them.
+        runtime = ctx.channel_runtime
+        if switched_thread and runtime is not None and runtime.active_teams:
+            dismissed = list(runtime.active_teams)
+            runtime.active_teams = []
+            ctx.ui.append_system(
+                f"Dismissed experts on session switch: {', '.join(dismissed)}",
+                style="dim",
+            )
+
         # Signal session change to UI
         if hasattr(ctx.ui, "handle_session_resume"):
             await ctx.ui.handle_session_resume(resolved, restored_workspace)
@@ -214,7 +227,23 @@ class NewCommand(Command):
     category = "Session"
 
     async def execute(self, ctx: CommandContext, args: list[str]) -> None:
+        # ``/new`` means fresh state — release any invited experts. Uniform
+        # with the explicit ``/expert clear`` path; avoids
+        # the "why is idea-brainstorm still active in my new thread?"
+        # surprise. Users who want to reuse an invite in the next thread can
+        # re-invite explicitly. Cleared only after the new session actually
+        # exists, so a failed start leaves the current session intact.
+        runtime = ctx.channel_runtime
+        dismissed: list[str] = []
+        if runtime is not None and runtime.active_teams:
+            dismissed = list(runtime.active_teams)
         await ctx.ui.start_new_session()
+        if dismissed:
+            runtime.active_teams = []
+            ctx.ui.append_system(
+                f"Dismissed experts on new session: {', '.join(dismissed)}",
+                style="dim",
+            )
 
 
 class ClearCommand(Command):

EvoScientist/config/onboard/channels.py

@@ -10,6 +10,7 @@ from __future__ import annotations
 import questionary
 from questionary import Choice
 
+from ...runtime import AsyncRuntime
 from ..settings import EvoScientistConfig
 from .helpers import (
     _setup_imessage,
@@ -21,7 +22,11 @@ from .style import (
 )
 
 
-def _step_channels(config: EvoScientistConfig) -> dict[str, object]:
+def _step_channels(
+    config: EvoScientistConfig,
+    *,
+    runtime: AsyncRuntime | None = None,
+) -> dict[str, object]:
     """Step: Select channels to enable on startup.
 
     Presents a multi-select list of supported channels.
@@ -35,6 +40,12 @@ def _step_channels(config: EvoScientistConfig) -> dict[str, object]:
         Dict mapping config field names to their new values.
         Empty dict when the user skips or selects nothing.
     """
+    # Direct/programmatic callers still get a single owned runtime for the
+    # whole step. CLI callers pass their application-scoped runtime instead.
+    if runtime is None:
+        with AsyncRuntime(thread_name="evosci-onboard-runtime") as owned_runtime:
+            return _step_channels(config, runtime=owned_runtime)
+
     # Currently enabled channels
     _currently_enabled = {
         t.strip()
@@ -592,11 +603,9 @@ def _step_channels(config: EvoScientistConfig) -> dict[str, object]:
                         f" to {_accounts_path}.[/dim]"
                     )
                     try:
-                        import asyncio
-
                         from ...channels.wechat.personal import qr_login
 
-                        creds = asyncio.run(qr_login())
+                        creds = runtime.run_sync(qr_login)
                     except Exception as exc:
                         console.print(f"  [red]✗ Scan failed: {exc}[/red]")
                         creds = None
@@ -783,7 +792,7 @@ def _step_channels(config: EvoScientistConfig) -> dict[str, object]:
             updates[senders_field] = senders.strip()
 
         # Probe validation
-        _probe_channel(ch_name, config, updates)
+        _probe_channel(ch_name, config, updates, runtime=runtime)
 
         enabled_channels.append(ch_name)
 
@@ -820,12 +829,13 @@ def _probe_channel(
     ch_name: str,
     config: EvoScientistConfig,
     updates: dict[str, object],
+    *,
+    runtime: AsyncRuntime,
 ) -> None:
     """Run the probe for a channel type and print the result.
 
     Non-fatal: prints a warning on failure but does not prevent enabling.
     """
-    import asyncio
 
     def _val(key: str, fallback: str = "") -> str:
         """Get a value from updates first, then config, then fallback."""
@@ -928,17 +938,7 @@ def _probe_channel(
             return True, "No probe available"
 
     try:
-        try:
-            loop = asyncio.get_event_loop()
-            if loop.is_running():
-                import nest_asyncio  # type: ignore[import-untyped]
-
-                nest_asyncio.apply()
-        except RuntimeError:
-            loop = asyncio.new_event_loop()
-            asyncio.set_event_loop(loop)
-
-        ok, detail = loop.run_until_complete(_run())
+        ok, detail = runtime.run_sync(_run)
         if ok:
             console.print(f"  [green]✓ {detail}[/green]")
         else:

EvoScientist/config/onboard/constants.py

@@ -21,6 +21,7 @@ VALID_PROVIDERS: frozenset[str] = frozenset(
         "zhipu",
         "zhipu-code",
         "volcengine",
+        "volcengine-code",
         "dashscope",
         "dashscope-code",
         "deepseek",
@@ -30,6 +31,9 @@ VALID_PROVIDERS: frozenset[str] = frozenset(
         "nvidia",
         "siliconflow",
         "openrouter",
+        "atlascloud",
+        "requesty",
+        "novita",
         "custom-openai",
         "custom-anthropic",
     }

EvoScientist/config/onboard/helpers.py

@@ -15,6 +15,7 @@ from ..settings import EvoScientistConfig
 from .style import QMARK, WIZARD_STYLE, console
 from .validators import (
     validate_anthropic_key,
+    validate_atlascloud_key,
     validate_dashscope_code_key,
     validate_dashscope_key,
     validate_deepseek_key,
@@ -22,9 +23,11 @@ from .validators import (
     validate_kimi_key,
     validate_minimax_key,
     validate_moonshot_key,
+    validate_novita_key,
     validate_nvidia_key,
     validate_openai_key,
     validate_openrouter_key,
+    validate_requesty_key,
     validate_siliconflow_key,
     validate_volcengine_key,
     validate_zhipu_key,
@@ -70,6 +73,21 @@ def _provider_key_info(config: EvoScientistConfig, provider: str):
             config.openrouter_api_key or os.environ.get("OPENROUTER_API_KEY", ""),
             validate_openrouter_key,
         ),
+        "atlascloud": (
+            "Atlas Cloud",
+            config.atlascloud_api_key or os.environ.get("ATLASCLOUD_API_KEY", ""),
+            validate_atlascloud_key,
+        ),
+        "requesty": (
+            "Requesty",
+            config.requesty_api_key or os.environ.get("REQUESTY_API_KEY", ""),
+            validate_requesty_key,
+        ),
+        "novita": (
+            "Novita",
+            config.novita_api_key or os.environ.get("NOVITA_API_KEY", ""),
+            validate_novita_key,
+        ),
         "deepseek": (
             "DeepSeek",
             config.deepseek_api_key or os.environ.get("DEEPSEEK_API_KEY", ""),
@@ -90,6 +108,11 @@ def _provider_key_info(config: EvoScientistConfig, provider: str):
             config.volcengine_api_key or os.environ.get("VOLCENGINE_API_KEY", ""),
             validate_volcengine_key,
         ),
+        "volcengine-code": (
+            "Volcengine Coding Plan",
+            config.volcengine_api_key or os.environ.get("VOLCENGINE_API_KEY", ""),
+            validate_volcengine_key,
+        ),
         "dashscope": (
             "DashScope",
             config.dashscope_api_key or os.environ.get("DASHSCOPE_API_KEY", ""),

EvoScientist/config/onboard/steps.py

@@ -156,8 +156,14 @@ def _step_langgraph_dev_port(config: EvoScientistConfig) -> int:
             f"EvoSci config set langgraph_dev_port <other-port>[/yellow]"
         )
     else:
+        # Render the address the configured bind actually produces rather than
+        # a hard-coded loopback URL — the two diverge once langgraph_dev_host
+        # is pinned to a specific interface.
+        from ...langgraph_dev.manager import _base_url
+
+        host = getattr(config, "langgraph_dev_host", "")
         console.print(
-            f"  [green]✓ EvoScientist will run on http://127.0.0.1:{port}[/green]"
+            f"  [green]✓ EvoScientist will run on {_base_url(port, host)}[/green]"
         )
     return port
 
@@ -220,7 +226,15 @@ def _step_webui_port(config: EvoScientistConfig) -> int:
         raise KeyboardInterrupt()
 
     port = int(raw) if raw else current_port
-    console.print(f"  [green]✓ WebUI will open at http://localhost:{port}[/green]")
+    # Same reasoning as the langgraph-dev step: render the configured bind, not
+    # a hard-coded localhost. A wildcard bind still shows loopback here — that
+    # is the address this machine's own browser opens.
+    from ...langgraph_dev.manager import _format_hostport
+
+    host = getattr(config, "webui_host", "")
+    console.print(
+        f"  [green]✓ WebUI will open at http://{_format_hostport(host, port)}[/green]"
+    )
     console.print(
         "  [yellow]⚠️  Note: the WebUI won't show your CLI/TUI chat history "
         "yet.[/yellow]"
@@ -264,6 +278,10 @@ def _step_provider(
             title="Volcengine (火山引擎 — Doubao models)",
             value="volcengine",
         ),
+        Choice(
+            title="Volcengine Coding Plan (火山引擎代码计划 — coding models)",
+            value="volcengine-code",
+        ),
         Choice(
             title="DashScope (阿里云 — Qwen models)",
             value="dashscope",
@@ -296,6 +314,18 @@ def _step_provider(
             title="OpenRouter (aggregator — Grok, Gemini, Qwen, etc.)",
             value="openrouter",
         ),
+        Choice(
+            title="Atlas Cloud (aggregator — DeepSeek, Qwen, etc.)",
+            value="atlascloud",
+        ),
+        Choice(
+            title="Requesty (aggregator — OpenAI, Anthropic, Gemini, xAI, etc.)",
+            value="requesty",
+        ),
+        Choice(
+            title="Novita (aggregator — DeepSeek, Qwen, GLM, etc.)",
+            value="novita",
+        ),
         Choice(
             title="OpenAI-compatible (third-party OpenAI endpoint)",
             value="custom-openai",
@@ -981,6 +1011,11 @@ _RECOMMENDED_SKILLS = [
         "label": "HuggingFace Skills  (dataset creation, model training & evaluation, third party by HuggingFace)",
         "source": "huggingface/skills@skills",
     },
+    # ── Third-party (NVIDIA BioNeMo) ──
+    {
+        "label": "BioNeMo Skills  (31 protein folding, docking, generative chemistry & genomics skills, third party by NVIDIA)",
+        "source": "NVIDIA-BioNeMo/bionemo-agent-toolkit@plugins/bionemo-agent-toolkit/skills",
+    },
 ]
 
 

EvoScientist/config/onboard/validators.py

@@ -321,6 +321,159 @@ def validate_openrouter_key(api_key: str) -> tuple[bool, str]:
         return False, f"Error: {e}"
 
 
+def validate_atlascloud_key(api_key: str) -> tuple[bool, str]:
+    """Validate an Atlas Cloud key with a nonexistent sentinel model.
+
+    The probe deliberately targets a nonexistent sentinel model. A 404 means
+    authentication passed and model resolution failed; 200 also confirms
+    authentication if the sentinel unexpectedly resolves. A 401/403 means the
+    key was rejected. Other statuses remain inconclusive until verified.
+    """
+    if not api_key:
+        return True, "Skipped (no key provided)"
+
+    try:
+        import httpx
+
+        resp = httpx.post(
+            "https://api.atlascloud.ai/v1/chat/completions",
+            headers={
+                "Authorization": f"Bearer {api_key}",
+                "Content-Type": "application/json",
+            },
+            json={
+                "model": "atlascloud/auth-preflight",
+                "messages": [{"role": "user", "content": "ping"}],
+                "max_tokens": 1,
+            },
+            timeout=10,
+        )
+        if resp.status_code in (200, 404):
+            return True, "Valid"
+        # Atlas checks account balance before model resolution: a valid key
+        # on an uncredited account gets 402 from the sentinel probe.
+        if resp.status_code == 402:
+            return True, "Valid (insufficient balance — top up to use)"
+        if resp.status_code in (401, 403):
+            return False, "Invalid API key"
+        return False, f"Validation inconclusive (HTTP {resp.status_code})"
+    except Exception as e:
+        classified = _classify_validation_error(e)
+        if classified is not None:
+            return classified
+        return False, f"Error: {e}"
+
+
+def validate_requesty_key(api_key: str) -> tuple[bool, str]:
+    """Validate a Requesty API key against the router's auth layer.
+
+    Unlike OpenRouter, Requesty's ``/v1/models`` endpoint returns HTTP 200
+    (the public model catalog) even for a missing or invalid key, so it
+    cannot be used to check a key. We instead issue a minimal
+    ``/v1/chat/completions`` request, but deliberately target a nonexistent
+    sentinel model: the router checks auth *before* resolving the model, so
+    the response distinguishes the two failures without depending on any
+    real model staying available upstream.
+
+    - valid key → 404 ("Model and/or policy not supported"), i.e. auth passed
+      (or 200 in the unlikely event the sentinel ever resolves);
+    - invalid/missing key → 401/403 ("Invalid authorization token");
+    - 429 (rate-limit) / 5xx (router incident) leave validity unknown, so a
+      transient outage doesn't reject a good key.
+
+    Returns:
+        Tuple of (is_valid, message).
+    """
+    if not api_key:
+        return True, "Skipped (no key provided)"
+
+    try:
+        import httpx
+
+        resp = httpx.post(
+            "https://router.requesty.ai/v1/chat/completions",
+            headers={
+                "Authorization": f"Bearer {api_key}",
+                "Content-Type": "application/json",
+            },
+            json={
+                # Deliberately nonexistent sentinel: auth is resolved before
+                # the model, so a valid key gets a 404 (model-not-found)
+                # rather than depending on a specific model being available.
+                "model": "requesty/auth-preflight",
+                "messages": [{"role": "user", "content": "ping"}],
+                "max_tokens": 1,
+            },
+            timeout=10,
+        )
+        # 200 (accepted) or 404 (auth passed, model not found) → key is good.
+        if resp.status_code in (200, 404):
+            return True, "Valid"
+        # Only 401/403 mean the key is actually rejected. 429 (rate-limit)
+        # and 5xx (router incident) leave the key validity unknown — surface
+        # the real status so the user doesn't go re-roll a good key during
+        # an outage.
+        if resp.status_code in (401, 403):
+            return False, "Invalid API key"
+        return False, f"Validation inconclusive (HTTP {resp.status_code})"
+    except Exception as e:
+        classified = _classify_validation_error(e)
+        if classified is not None:
+            return classified
+        return False, f"Error: {e}"
+
+
+def validate_novita_key(api_key: str) -> tuple[bool, str]:
+    """Validate a Novita API key against the router's auth layer.
+
+    Like Requesty and Atlas Cloud, Novita's ``/v1/models`` endpoint returns
+    HTTP 200 (the public model catalog) even for a missing or invalid key, so
+    it cannot be used to check a key (verified against the live endpoint). We
+    instead issue a minimal ``/v1/chat/completions`` request with a
+    deliberately nonexistent sentinel model: auth is resolved before the
+    model, so a valid key doesn't depend on any real model staying available
+    upstream.
+
+    - invalid/missing key → 401/403 (confirmed against the live endpoint);
+    - valid key → 200 or 404 (model-not-found, auth passed), mirroring the
+      Requesty/Atlas Cloud sentinel pattern;
+    - 429 (rate-limit) / 5xx (service incident) leave validity unknown, so a
+      transient outage doesn't reject a good key.
+
+    Returns:
+        Tuple of (is_valid, message).
+    """
+    if not api_key:
+        return True, "Skipped (no key provided)"
+
+    try:
+        import httpx
+
+        resp = httpx.post(
+            "https://api.novita.ai/openai/v1/chat/completions",
+            headers={
+                "Authorization": f"Bearer {api_key}",
+                "Content-Type": "application/json",
+            },
+            json={
+                "model": "novita/auth-preflight",
+                "messages": [{"role": "user", "content": "ping"}],
+                "max_tokens": 1,
+            },
+            timeout=10,
+        )
+        if resp.status_code in (200, 404):
+            return True, "Valid"
+        if resp.status_code in (401, 403):
+            return False, "Invalid API key"
+        return False, f"Validation inconclusive (HTTP {resp.status_code})"
+    except Exception as e:
+        classified = _classify_validation_error(e)
+        if classified is not None:
+            return classified
+        return False, f"Error: {e}"
+
+
 def validate_deepseek_key(api_key: str) -> tuple[bool, str]:
     """Validate a DeepSeek API key by making a test request.
 
@@ -373,6 +526,9 @@ def validate_zhipu_key(api_key: str) -> tuple[bool, str]:
 def validate_volcengine_key(api_key: str) -> tuple[bool, str]:
     """Validate a Volcengine API key by making a test request.
 
+    Uses the general endpoint for validation; volcengine and volcengine-code
+    share the same API key and only differ in their runtime base URL.
+
     Returns:
         Tuple of (is_valid, message).
     """

EvoScientist/config/onboard/wizard.py

@@ -9,6 +9,7 @@ import questionary
 from rich.panel import Panel
 from rich.text import Text
 
+from ...runtime import AsyncRuntime
 from ..settings import (
     EvoScientistConfig,
     get_config_path,
@@ -117,10 +118,14 @@ _PROVIDER_KEY_ATTR = {
     "google-genai": "google_api_key",
     "siliconflow": "siliconflow_api_key",
     "openrouter": "openrouter_api_key",
+    "atlascloud": "atlascloud_api_key",
+    "requesty": "requesty_api_key",
+    "novita": "novita_api_key",
     "deepseek": "deepseek_api_key",
     "zhipu": "zhipu_api_key",
     "zhipu-code": "zhipu_api_key",
     "volcengine": "volcengine_api_key",
+    "volcengine-code": "volcengine_api_key",
     "dashscope": "dashscope_api_key",
     "dashscope-code": "dashscope_api_key",
     "moonshot": "moonshot_api_key",
@@ -475,6 +480,7 @@ def run_onboard(
     skip_validation: bool = False,
     prompter=None,
     only_sections: set[str] | frozenset[str] | None = None,
+    runtime: AsyncRuntime | None = None,
 ) -> bool:
     """Run the interactive onboarding wizard.
 
@@ -487,6 +493,9 @@ def run_onboard(
         only_sections: If given, restrict the wizard to exactly these section
             ids — the Keep/Modify/Reset prompt is skipped. Used by ``EvoSci
             configure <section>`` to re-run a single phase.
+        runtime: Optional application-scoped async runtime used by channel
+            login and credential probes. Direct callers may omit it; the
+            channel step then owns a runtime for the duration of that step.
 
     Returns:
         True if configuration was saved, False if cancelled.
@@ -883,7 +892,7 @@ def run_onboard(
                 _step_tinytex()
 
             if "channels" in sections_to_run:
-                for key, value in _step_channels(config).items():
+                for key, value in _step_channels(config, runtime=runtime).items():
                     setattr(config, key, value)
                 _autosave(config)
 

EvoScientist/config/settings.py

@@ -1,8 +1,9 @@
 """Configuration management for EvoScientist.
 
-Handles loading, saving, and merging configuration from multiple sources
-with the following priority (highest to lowest):
-    CLI arguments > Environment variables > Config file > Defaults
+Handles loading, saving, and merging configuration from multiple sources.
+See :func:`get_effective_config` for the authoritative priority chain —
+``EVOSCIENTIST_*`` shell values and third-party keys are treated
+asymmetrically with respect to workspace ``.env`` handling.
 """
 
 from __future__ import annotations
@@ -16,14 +17,18 @@ from pathlib import Path
 from typing import Any, Literal, get_type_hints
 
 import yaml
-from dotenv import find_dotenv, load_dotenv
+from dotenv import dotenv_values, find_dotenv
 
 # Tools that run shell commands and need manual HITL approval (subject to
 # shell_allow_list). Single source of truth for every interrupt consumer
-# (stream/display.py, channels/consumer.py) — keep aligned with the agent's
+# (stream/display.py, channels/interaction.py) — keep aligned with the agent's
 # `interrupt_on` set in EvoScientist.py.
 HITL_SHELL_TOOLS = ("execute", "run_in_background")
 
+# Armed non-shell destructive tools must always prompt — no allow-list carve-outs
+# (their args carry paths, not commands). Keep aligned with HITL_INTERRUPT_ON.
+HITL_ALWAYS_PROMPT_TOOLS = ("delete", "schedule_task")
+
 
 class MemoryObservationTarget(StrEnum):
     """Runtime locations that can receive `record_observation`."""
@@ -178,6 +183,9 @@ class EvoScientistConfig:
     minimax_base_url: str = ""
     siliconflow_api_key: str = ""
     openrouter_api_key: str = ""
+    atlascloud_api_key: str = ""
+    requesty_api_key: str = ""
+    novita_api_key: str = ""
     deepseek_api_key: str = ""
     zhipu_api_key: str = ""
     volcengine_api_key: str = ""
@@ -218,11 +226,24 @@ class EvoScientistConfig:
     # the Ai4Sci-Web Gateway's recoverable runtime URL.
     langgraph_dev_port: int = 3076
 
+    # Network interface the langgraph dev subprocess binds to. Loopback by
+    # default — this is the unauthenticated agent API (the agent can run
+    # shell), so "0.0.0.0" is opt-in and every launcher prints a PUBLIC BIND
+    # banner while exposed. Internal callers *connect* via manager._probe_host,
+    # so widening never redirects their traffic off-box.
+    langgraph_dev_host: str = "127.0.0.1"
+
     # Port for the WebUI front-end (Next.js server from @evoscientist/webui),
     # used only when ui_backend == "webui". The backend keeps
     # its own port (langgraph_dev_port); this is just the browser server.
     webui_port: int = 4716
 
+    # Network interface the WebUI front-end binds to. Loopback by default,
+    # matching langgraph_dev_host: this server is not a passive app shell —
+    # its API reads, writes and uploads workspace files and installs skills,
+    # all unauthenticated. Set "0.0.0.0" (with langgraph_dev_host) for LAN.
+    webui_host: str = "127.0.0.1"
+
     # --- Scheduled tasks (cron) ---
     # Master switch for scheduled tasks (/schedule, NL tools, scheduler context). Defaults
     # True so the feature is available out-of-the-box; set False to disable.
@@ -248,6 +269,15 @@ class EvoScientistConfig:
     # slowdown.
     langgraph_dev_jobs_per_worker: int = 10
 
+    # Keep the auto-started langgraph dev subprocess running after the CLI
+    # exits. The next `EvoSci` start in the same workspace reuses it instantly
+    # instead of paying the cold boot (~15s). Starting in a DIFFERENT workspace
+    # raises WorkspaceMismatchError with the leftover server's pid — stop it
+    # manually (the server is pinned to one workspace per process). Known
+    # limitation: changing langgraph_dev_port/host while a keepalive server
+    # runs orphans its records — run `EvoSci server stop` before switching.
+    langgraph_dev_keepalive: bool = False
+
     # Max LangGraph super-steps (LLM call / tool call / sub-agent delegation
     # each count as 1) before raising GraphRecursionError. Resets on every
     # ``agent.invoke()`` — i.e., this is per-turn, NOT per-conversation. For
@@ -294,6 +324,14 @@ class EvoScientistConfig:
         DEFAULT_MEMORY_SKILL_SYNTHESIS_CADENCE
     )
     memory_skill_synthesis_time: str = DEFAULT_MEMORY_SKILL_SYNTHESIS_TIME
+    # Max number of parsed observation files kept in the process-wide parse
+    # cache. Each entry holds one parsed document keyed on the file path; at
+    # the end of a call the LRU trims down to max(cap, entries touched by
+    # the call), so an active store larger than the cap temporarily exceeds
+    # it instead of thrashing. 2048 is generous for the single-workspace
+    # deploy model; raise for a long-running server that cycles through many
+    # large workspaces.
+    memory_observation_cache_max_files: int = 2048
 
     # Workspace Settings
     default_mode: Literal["daemon", "run"] = "daemon"
@@ -310,7 +348,9 @@ class EvoScientistConfig:
     openrouter_anthropic_prompt_cache: bool = True
     # OpenRouter app attribution (issue #339). Sent only for the openrouter
     # provider; identifies EvoScientist in OpenRouter's app rankings/analytics.
-    # Override (e.g. a private fork) via these fields or their env vars.
+    # Override (e.g. a private fork) via these fields or their env vars. A custom
+    # title only takes effect together with a custom referer: OpenRouter keys app
+    # pages by referer, so a lone title would rename the shared EvoScientist page.
     # Defaults live in the module constants above (also imported by llm/models.py).
     openrouter_http_referer: str = OPENROUTER_DEFAULT_HTTP_REFERER
     openrouter_app_title: str = OPENROUTER_DEFAULT_APP_TITLE
@@ -492,14 +532,36 @@ class EvoScientistConfig:
             )
             self.sandbox_execute_timeout = 300
 
-        # Dangerous mode implies auto_approve regardless of source (CLI, env,
-        # config file). Mirrors how auto_mode implies auto_approve — done here so
-        # the coupling holds even when dangerous_mode is set via `config set`.
-        if self.dangerous_mode:
+        # A non-positive cache cap would evict every file entry immediately,
+        # defeating the cache entirely.
+        cap = self.memory_observation_cache_max_files
+        if not isinstance(cap, int) or isinstance(cap, bool) or cap < 1:
+            logging.getLogger(__name__).warning(
+                "Invalid memory_observation_cache_max_files %r; falling back to 2048.",
+                cap,
+            )
+            self.memory_observation_cache_max_files = 2048
+
+        # auto_mode and dangerous_mode both imply auto_approve regardless of
+        # source (CLI, env, config file, direct construction) — done here so the
+        # "unattended → zero prompts" contract holds even when either is set via
+        # `config set` or a config file rather than a CLI flag.
+        if self.auto_mode or self.dangerous_mode:
             self.auto_approve = True
 
         _normalize_str_enum_fields(self)
 
+        # Bind hosts reach socket.bind() / the langgraph CLI verbatim, where a
+        # stray-whitespace or empty value surfaces as an opaque gaierror at
+        # startup. Normalize to the field's own default instead.
+        for _host_field, _host_default in (
+            ("langgraph_dev_host", "127.0.0.1"),
+            ("webui_host", "127.0.0.1"),
+        ):
+            _host = getattr(self, _host_field, _host_default)
+            _host = _host.strip() if isinstance(_host, str) else ""
+            setattr(self, _host_field, _host or _host_default)
+
         synthesis_time = _normalize_hhmm(self.memory_skill_synthesis_time)
         if synthesis_time is None:
             logging.getLogger(__name__).warning(
@@ -783,6 +845,9 @@ _ENV_MAPPINGS = {
     "minimax_base_url": "MINIMAX_BASE_URL",
     "siliconflow_api_key": "SILICONFLOW_API_KEY",
     "openrouter_api_key": "OPENROUTER_API_KEY",
+    "atlascloud_api_key": "ATLASCLOUD_API_KEY",
+    "requesty_api_key": "REQUESTY_API_KEY",
+    "novita_api_key": "NOVITA_API_KEY",
     "deepseek_api_key": "DEEPSEEK_API_KEY",
     "zhipu_api_key": "ZHIPU_API_KEY",
     "volcengine_api_key": "VOLCENGINE_API_KEY",
@@ -815,7 +880,9 @@ _ENV_MAPPINGS = {
     "checkpoint_keep_per_thread": "EVOSCIENTIST_CHECKPOINT_KEEP_PER_THREAD",
     "enable_async_subagents": "EVOSCIENTIST_ENABLE_ASYNC_SUBAGENTS",
     "langgraph_dev_port": "EVOSCIENTIST_LANGGRAPH_DEV_PORT",
+    "langgraph_dev_host": "EVOSCIENTIST_LANGGRAPH_DEV_HOST",
     "webui_port": "EVOSCIENTIST_WEBUI_PORT",
+    "webui_host": "EVOSCIENTIST_WEBUI_HOST",
     "enable_scheduler": "EVOSCIENTIST_ENABLE_SCHEDULER",
     "scheduler_default_timezone": "EVOSCIENTIST_SCHEDULER_DEFAULT_TIMEZONE",
     "code_interpreter_timeout": "EVOSCIENTIST_CODE_INTERPRETER_TIMEOUT",
@@ -823,6 +890,7 @@ _ENV_MAPPINGS = {
     "sandbox_execute_timeout": "EVOSCIENTIST_SANDBOX_EXECUTE_TIMEOUT",
     "langgraph_dev_file_persistence": "EVOSCIENTIST_LANGGRAPH_DEV_FILE_PERSISTENCE",
     "langgraph_dev_jobs_per_worker": "EVOSCIENTIST_LANGGRAPH_DEV_JOBS_PER_WORKER",
+    "langgraph_dev_keepalive": "EVOSCIENTIST_LANGGRAPH_DEV_KEEPALIVE",
     "recursion_limit": "EVOSCIENTIST_RECURSION_LIMIT",
     "repetitive_tool_call_threshold": (
         "EVOSCIENTIST_REPETITIVE_TOOL_CALL_THRESHOLD"
@@ -836,6 +904,7 @@ _ENV_MAPPINGS = {
     "memory_skill_synthesis_mode": "EVOSCIENTIST_MEMORY_SKILL_SYNTHESIS_MODE",
     "memory_skill_synthesis_cadence": "EVOSCIENTIST_MEMORY_SKILL_SYNTHESIS_CADENCE",
     "memory_skill_synthesis_time": "EVOSCIENTIST_MEMORY_SKILL_SYNTHESIS_TIME",
+    "memory_observation_cache_max_files": "EVOSCIENTIST_MAX_CACHED_FILES",
 }
 
 
@@ -845,10 +914,33 @@ def get_effective_config(
     """Get effective configuration by merging all sources.
 
     Priority (highest to lowest):
-        1. CLI arguments (cli_overrides)
-        2. Environment variables
-        3. Config file
-        4. Defaults
+        1. CLI arguments (``cli_overrides``)
+        2. Parent-process environment variables for any ``EVOSCIENTIST_*`` key
+        3. ``.env`` file at (or above) the current working directory
+        4. Parent-process environment variables for everything else
+           (third-party API keys / base URLs, plus arbitrary unmapped keys)
+        5. Config file (``~/.config/evoscientist/config.yaml``)
+        6. Dataclass defaults
+
+    Rows 2 and 4 differ because ``.env`` values need different treatment
+    for our own namespaced config knobs vs third-party credentials.
+    Third-party keys (``ANTHROPIC_API_KEY``, ``OPENAI_API_KEY``, ...)
+    follow the industry convention that ``.env`` is the per-project
+    credential store; extending shell-wins to them would silently flip
+    a workspace key back to a global ``.bashrc`` key. Our own
+    ``EVOSCIENTIST_*`` keys are the opposite: an explicit CLI/parent-
+    process value (e.g. the bind port that ``EvoSci deploy --port X``
+    hands to the langgraph dev subprocess) must not be shadowed by a
+    workspace ``.env``. We implement this by reading ``.env`` into a
+    dict via ``dotenv_values`` (no ``os.environ`` mutation), then
+    writing third-party keys unconditionally and ``EVOSCIENTIST_*`` keys
+    only when the shell doesn't already have a non-empty value.
+
+    Tradeoff: ``OPENAI_API_KEY=xxx evoscientist ...`` inline overrides
+    still lose to a workspace ``.env`` containing ``OPENAI_API_KEY``,
+    because the merge writes third-party keys from ``.env``
+    unconditionally. Users who need to override a ``.env``-defined
+    credential inline must edit or unset the ``.env`` entry.
 
     Args:
         cli_overrides: Dictionary of CLI argument overrides.
@@ -856,7 +948,34 @@ def get_effective_config(
     Returns:
         EvoScientistConfig with merged values.
     """
-    load_dotenv(find_dotenv(usecwd=True), override=True)
+    # Merge workspace ``.env`` into ``os.environ`` without going through
+    # ``load_dotenv``. The previous snapshot → ``load_dotenv`` → restore
+    # sequence was a read-modify-write on ``os.environ`` that could race with
+    # concurrent ``get_effective_config`` calls in the langgraph dev subprocess
+    # (per-request threads in ``langgraph_dev/http.py``, ``sessions.py``
+    # checkpoint writes, memory workers): one thread's mid-flight ``.env``
+    # value could be re-captured by another as "parent env" and then restored
+    # last, promoting the ``.env`` value into the snapshot permanently.
+    #
+    # ``dotenv_values`` returns a dict without touching ``os.environ``, so the
+    # merge below is a pure write sequence and idempotent under interleaving.
+    # Third-party keys keep ``.env``-wins (industry convention).
+    # ``EVOSCIENTIST_*`` keys are our own namespaced config knobs where
+    # CLI/parent-process intent should stay authoritative — write from ``.env``
+    # only when the shell doesn't already have a non-empty value. Treating an
+    # empty shell value as "unset" matches the ``if env_value:`` truthy check
+    # in the ``_ENV_MAPPINGS`` loop below; without this, an empty parent export
+    # would silently regress vs main by falling through to file/defaults.
+    dotenv_path = find_dotenv(usecwd=True)
+    dotenv_map = dotenv_values(dotenv_path) if dotenv_path else {}
+    for env_key, env_value in dotenv_map.items():
+        if env_value is None:
+            continue  # bare ``FOO`` without ``=`` — nothing to write
+        if env_key.startswith("EVOSCIENTIST_"):
+            if not os.environ.get(env_key):
+                os.environ[env_key] = env_value
+        else:
+            os.environ[env_key] = env_value
 
     # Start with file config (includes defaults for missing values)
     config = load_config()
@@ -913,6 +1032,12 @@ def apply_config_to_env(config: EvoScientistConfig) -> None:
         os.environ["SILICONFLOW_API_KEY"] = config.siliconflow_api_key
     if config.openrouter_api_key and not os.environ.get("OPENROUTER_API_KEY"):
         os.environ["OPENROUTER_API_KEY"] = config.openrouter_api_key
+    if config.atlascloud_api_key and not os.environ.get("ATLASCLOUD_API_KEY"):
+        os.environ["ATLASCLOUD_API_KEY"] = config.atlascloud_api_key
+    if config.requesty_api_key and not os.environ.get("REQUESTY_API_KEY"):
+        os.environ["REQUESTY_API_KEY"] = config.requesty_api_key
+    if config.novita_api_key and not os.environ.get("NOVITA_API_KEY"):
+        os.environ["NOVITA_API_KEY"] = config.novita_api_key
     if config.deepseek_api_key and not os.environ.get("DEEPSEEK_API_KEY"):
         os.environ["DEEPSEEK_API_KEY"] = config.deepseek_api_key
     if config.zhipu_api_key and not os.environ.get("ZHIPU_API_KEY"):

EvoScientist/cron/schedule.py

@@ -12,7 +12,7 @@ multiple clients at one hand-started server they will share the same cron store.
 
 from __future__ import annotations
 
-from typing import TYPE_CHECKING
+from typing import TYPE_CHECKING, Any
 
 if TYPE_CHECKING:
     from langgraph_sdk.schema import Cron, Run
@@ -28,6 +28,33 @@ SCHEDULER_GRAPH_ID = "scheduler"
 SCHEDULED_RUN_KIND = "scheduled_task"
 
 
+def _normalize_rubric(rubric: str | None) -> str | None:
+    text = (rubric or "").strip()
+    return text or None
+
+
+def _scheduled_input(prompt: str, rubric: str | None) -> dict[str, Any]:
+    """Run input for the scheduler graph; ``rubric`` rides along only when set.
+
+    The key is read by ``RubricMiddleware`` mounted on the scheduler graph — an
+    absent key means no grading pass at all, so unset stays byte-identical to
+    the pre-rubric payload.
+    """
+    payload: dict[str, Any] = messages_input(prompt)
+    if rubric:
+        payload["rubric"] = rubric
+    return payload
+
+
+def _scheduled_metadata(
+    *, name: str, prompt: str, rubric: str | None
+) -> dict[str, str]:
+    metadata = {"run_kind": SCHEDULED_RUN_KIND, "name": name, "prompt": prompt}
+    if rubric:
+        metadata["rubric"] = rubric
+    return metadata
+
+
 def _scheduler_url() -> str:
     return configured_langgraph_dev_url()
 
@@ -48,16 +75,26 @@ def is_available() -> bool:
 
 
 def create_schedule(
-    *, name: str, schedule: str, prompt: str, timezone: str | None = None
+    *,
+    name: str,
+    schedule: str,
+    prompt: str,
+    timezone: str | None = None,
+    rubric: str | None = None,
 ) -> Cron:
-    """Create a recurring scheduled task on the scheduler graph."""
+    """Create a recurring scheduled task on the scheduler graph.
+
+    ``rubric`` is an optional acceptance checklist graded after each run; blank
+    means the run is never graded.
+    """
+    rubric = _normalize_rubric(rubric)
     # Crons are stored in the langgraph-dev process's .langgraph_api store, not
     # tagged by workspace. Isolation is process-level (see module docstring).
     return _client().crons.create(
         assistant_id=SCHEDULER_GRAPH_ID,
         schedule=schedule,
-        input=messages_input(prompt),
-        metadata={"run_kind": SCHEDULED_RUN_KIND, "name": name, "prompt": prompt},
+        input=_scheduled_input(prompt, rubric),
+        metadata=_scheduled_metadata(name=name, prompt=prompt, rubric=rubric),
         timezone=timezone or _default_timezone(),
     )
 
@@ -87,20 +124,17 @@ def set_enabled(cron_id: str, enabled: bool) -> Cron:
     return _client().crons.update(cron_id, enabled=enabled)
 
 
-def run_now(prompt: str) -> Run:
+def run_now(prompt: str, *, rubric: str | None = None) -> Run:
     """Fire a one-off scheduler run immediately (for ``/schedule run``).
 
     Output goes wherever the task's prompt specifies; there is no push notification.
     """
+    rubric = _normalize_rubric(rubric)
     client = _client()
     thread = client.threads.create(graph_id=SCHEDULER_GRAPH_ID)
     return client.runs.create(
         thread_id=str(thread["thread_id"]),
         assistant_id=SCHEDULER_GRAPH_ID,
-        input=messages_input(prompt),
-        metadata={
-            "run_kind": SCHEDULED_RUN_KIND,
-            "name": "manual-run",
-            "prompt": prompt,
-        },
+        input=_scheduled_input(prompt, rubric),
+        metadata=_scheduled_metadata(name="manual-run", prompt=prompt, rubric=rubric),
     )

EvoScientist/deploy/server.py

@@ -46,6 +46,13 @@ def deploy(
         "--port",
         help="Port for langgraph dev (default: config.langgraph_dev_port = 3076)",
     ),
+    host: str | None = typer.Option(
+        None,
+        "--host",
+        help="Interface to bind (default: config.langgraph_dev_host = "
+        "127.0.0.1, i.e. this machine only — pass 0.0.0.0 to reach it from "
+        "other machines, but note the server has no auth)",
+    ),
     tunnel: bool = typer.Option(
         False,
         "--tunnel",
@@ -66,9 +73,15 @@ def deploy(
     """
     from ..config import apply_config_to_env, get_effective_config
     from ..langgraph_dev.manager import (
+        _DEFAULT_HOST,
         _DEFAULT_PORT,
         RUNTIME,
+        _base_url,
+        _is_loopback_host,
         _is_port_occupied,
+        _pid_serves_port,
+        _read_workspace_sidecar,
+        _server_config_fingerprint,
         is_langgraph_dev_running,
         read_tunnel_url,
         start_langgraph_dev,
@@ -114,20 +127,46 @@ def deploy(
         )
         raise typer.Exit(1)
 
+    # A blank ``--host`` means "not passed" (matching serve), so it can never
+    # discard the configured bind. Both branches strip: whitespace reaching
+    # socket.bind() surfaces as an opaque gaierror, and duck-typed configs
+    # handed to this function never ran ``__post_init__`` normalization.
+    cli_host = host.strip() if host is not None else ""
+    effective_host = (
+        cli_host
+        or str(getattr(config, "langgraph_dev_host", _DEFAULT_HOST) or "").strip()
+        or _DEFAULT_HOST
+    )
+
     # 4. Pre-flight port check — refuse to start if a non-EvoSci process is
     # holding the port. If an existing EvoSci langgraph dev is already up,
     # also refuse (deploy is the "primary server" — running multiple on the
     # same port is a configuration error).
-    if _is_port_occupied(effective_port):
-        if is_langgraph_dev_running(port=effective_port):
+    if _is_port_occupied(effective_port, effective_host):
+        if is_langgraph_dev_running(port=effective_port, host=effective_host):
             console.print(
                 f"[red]Port {effective_port} is already serving a langgraph dev "
                 f"instance.[/red]"
             )
-            console.print(
-                "[dim]Stop the existing EvoSci/serve session first, or use "
-                "[bold]--port[/bold] to deploy on a different port.[/dim]"
-            )
+            sidecar = _read_workspace_sidecar()
+            if sidecar is not None and _pid_serves_port(
+                sidecar.get("pid"), effective_port
+            ):
+                # Surface what we know about the occupant — with keepalive it
+                # may be an ownerless leftover rather than a live session.
+                # Only when the recorded pid verifiably serves THIS port, so a
+                # stale or other-port record is never blamed.
+                console.print(
+                    f"[dim]It serves workspace {sidecar.get('workspace')} "
+                    f"(pid {sidecar.get('pid')}). Stop it with "
+                    f"[bold]EvoSci server stop[/bold], or use "
+                    f"[bold]--port[/bold] to deploy on a different port.[/dim]"
+                )
+            else:
+                console.print(
+                    "[dim]Stop the existing EvoSci/serve session first, or use "
+                    "[bold]--port[/bold] to deploy on a different port.[/dim]"
+                )
         else:
             console.print(
                 f"[red]Port {effective_port} is occupied by another process.[/red]"
@@ -144,6 +183,7 @@ def deploy(
         Panel(
             Text.from_markup(
                 f"[bold]Workspace:[/bold] {_shorten(ws)}\n"
+                f"[bold]Host:[/bold]      {effective_host}\n"
                 f"[bold]Port:[/bold]      {effective_port}\n"
                 f"[bold]Auth:[/bold]      {_auth_label}"
             ),
@@ -162,6 +202,13 @@ def deploy(
             f"[bold red]{DANGEROUS_BANNER_MESSAGE}[/bold red]"
         )
 
+    if not _is_loopback_host(effective_host):
+        console.print(
+            "[bold white on red] ⚠ PUBLIC BIND [/bold white on red] "
+            f"[bold red]Listening on {effective_host} — no auth, and the agent "
+            f"can run shell. Trusted networks only.[/bold red]"
+        )
+
     if tunnel:
         console.print(
             "[bold white on red] ⚠ PUBLIC TUNNEL [/bold white on red] "
@@ -197,10 +244,12 @@ def deploy(
             proc = start_langgraph_dev(
                 workspace_dir=Path(ws),
                 port=effective_port,
+                host=effective_host,
                 file_persistence=file_persistence,
                 jobs_per_worker=jobs_per_worker,
                 deploy_mode=True,
                 tunnel=tunnel,
+                config_fingerprint=_server_config_fingerprint(config),
             )
         atexit.register(stop_langgraph_dev, proc)
     except Exception as exc:
@@ -236,7 +285,7 @@ def deploy(
         Panel(
             Text.from_markup(
                 f"[bold]Endpoint:[/bold]     "
-                f"http://localhost:{effective_port}\n"
+                f"{_base_url(effective_port, effective_host)}\n"
                 f"{public_line}"
                 f"[bold]Assistant ID:[/bold] EvoScientist\n"
                 f"[bold]Connect via:[/bold]  any LangChain SDK / "

EvoScientist/deploy/webui.py

@@ -42,6 +42,7 @@ from ..stream.console import console
 # Front-end npm package + spec. ``@latest`` → always the newest published UI.
 _WEBUI_PACKAGE = "@evoscientist/webui@latest"
 _DEFAULT_WEBUI_PORT = 4716
+_DEFAULT_WEBUI_HOST = "127.0.0.1"
 
 
 def run_webui(config: Any, workspace_dir: str | None = None) -> None:
@@ -58,10 +59,15 @@ def run_webui(config: Any, workspace_dir: str | None = None) -> None:
     """
     from ..config import apply_config_to_env
     from ..langgraph_dev.manager import (
+        _DEFAULT_HOST,
         _DEFAULT_PORT,
         RUNTIME,
+        _base_url,
+        _format_hostport,
+        _is_loopback_host,
         _is_port_occupied,
         _read_workspace_sidecar,
+        _server_config_fingerprint,
         is_langgraph_dev_running,
         start_langgraph_dev,
         stop_langgraph_dev,
@@ -84,6 +90,14 @@ def run_webui(config: Any, workspace_dir: str | None = None) -> None:
     # webui_port = the local Next.js server the browser actually opens.
     backend_port = int(getattr(config, "langgraph_dev_port", _DEFAULT_PORT))
     webui_port = int(getattr(config, "webui_port", _DEFAULT_WEBUI_PORT))
+    # ...and their bind interfaces, both loopback by default — the front-end
+    # carries workspace/skill APIs of its own (see config.webui_host).
+    backend_host = (
+        str(getattr(config, "langgraph_dev_host", _DEFAULT_HOST) or _DEFAULT_HOST)
+    ).strip() or _DEFAULT_HOST
+    webui_host = (
+        str(getattr(config, "webui_host", _DEFAULT_WEBUI_HOST) or _DEFAULT_WEBUI_HOST)
+    ).strip() or _DEFAULT_WEBUI_HOST
     for label, p in (("langgraph dev", backend_port), ("WebUI", webui_port)):
         if not (1 <= p <= 65535):
             console.print(
@@ -128,8 +142,8 @@ def run_webui(config: Any, workspace_dir: str | None = None) -> None:
     # else start a fresh deploy-mode one (full MCP + async). Refuse a foreign
     # occupant — that's a configuration error, not something to silently share.
     started_proc = None
-    if _is_port_occupied(backend_port):
-        if is_langgraph_dev_running(port=backend_port):
+    if _is_port_occupied(backend_port, backend_host):
+        if is_langgraph_dev_running(port=backend_port, host=backend_host):
             # Reuse an existing EvoSci server only when it serves THIS workspace
             # — mirror the sidecar guard in ensure_langgraph_dev so WebUI started
             # from workspace B never silently binds to a server pinned to
@@ -150,6 +164,31 @@ def run_webui(config: Any, workspace_dir: str | None = None) -> None:
                     f"[/dim]"
                 )
                 raise typer.Exit(1)
+            if sidecar is not None and sidecar.get("deploy_mode") is False:
+                # A stripped (CLI-started) server has no MCP and no async
+                # sub-agents — silently reusing it would degrade the WebUI
+                # with no visible cause. Refuse; never auto-kill.
+                console.print(
+                    f"[red]Port {backend_port} is serving a stripped "
+                    f"(CLI-mode) langgraph dev — the WebUI needs the full "
+                    f"deploy-mode server (MCP + async sub-agents).[/red]"
+                )
+                console.print(
+                    "[dim]Stop it with [bold]EvoSci server stop[/bold], then "
+                    "re-run [bold]EvoSci[/bold].[/dim]"
+                )
+                raise typer.Exit(1)
+            if sidecar is not None:
+                recorded_fp = sidecar.get("config_fingerprint")
+                if isinstance(
+                    recorded_fp, str
+                ) and recorded_fp != _server_config_fingerprint(config):
+                    console.print(
+                        "[yellow]⚠ Config changed since this server was "
+                        "launched — it still serves the old settings. Apply "
+                        "them with [bold]EvoSci server stop[/bold], then "
+                        "re-run EvoSci.[/yellow]"
+                    )
             console.print(
                 f"[green]✓[/green] Reusing langgraph dev already serving "
                 f"port {backend_port}"
@@ -174,17 +213,28 @@ def run_webui(config: Any, workspace_dir: str | None = None) -> None:
                 started_proc = start_langgraph_dev(
                     workspace_dir=Path(ws),
                     port=backend_port,
+                    host=backend_host,
                     file_persistence=file_persistence,
                     jobs_per_worker=jobs_per_worker,
                     deploy_mode=True,
+                    config_fingerprint=_server_config_fingerprint(config),
                 )
-            atexit.register(stop_langgraph_dev, started_proc)
+            if getattr(config, "langgraph_dev_keepalive", False):
+                # Keepalive: the deploy-mode backend outlives this session so
+                # the next same-workspace launch reuses it instantly. The npx
+                # front-end below still stops on exit as usual.
+                console.print(
+                    "[dim]keepalive: backend server stays up after exit — "
+                    "stop it with [bold]EvoSci server stop[/bold].[/dim]"
+                )
+            else:
+                atexit.register(stop_langgraph_dev, started_proc)
         except Exception as exc:
             console.print(f"[red]langgraph dev startup failed:[/red] {exc}")
             raise typer.Exit(1) from exc
         console.print("[green]✓[/green] langgraph dev ready")
 
-    if _is_port_occupied(webui_port):
+    if _is_port_occupied(webui_port, webui_host):
         console.print(
             f"[yellow]⚠ Port {webui_port} is already in use; the WebUI server "
             f"may fail to start. Change it with "
@@ -197,20 +247,37 @@ def run_webui(config: Any, workspace_dir: str | None = None) -> None:
     # inherited so it all shows in THIS terminal. EVOSCIENTIST_LANGGRAPH_DEV_PORT
     # lets the UI's config prefill point at our backend automatically. Secrets
     # are scrubbed — the browser UI never needs LLM provider API keys.
+    #
+    # HOSTNAME is the front-end's only bind knob: the package has no --host
+    # flag; its launcher forwards `HOSTNAME || "127.0.0.1"` to the Next server.
     webui_env = _scrubbed_env(
         {
             "EVOSCIENTIST_LANGGRAPH_DEV_PORT": str(backend_port),
             "PORT": str(webui_port),
+            "HOSTNAME": webui_host,
         }
     )
+    # The UI reaches the backend from the BROWSER; when only the front-end is
+    # exposed, remote pages load but every request fails — say so.
+    remote_backend_hint = ""
+    if not _is_loopback_host(webui_host) and _is_loopback_host(backend_host):
+        remote_backend_hint = (
+            f"\n[yellow]Note:[/yellow] the UI connects to the backend from the "
+            f"browser. Remote visitors cannot reach a loopback backend — run "
+            f"[bold]EvoSci config set langgraph_dev_host 0.0.0.0[/bold] and "
+            f"point the UI at [bold]http://<this-machine-ip>:{backend_port}"
+            f"[/bold].\n"
+        )
     console.print(
         Panel(
             Text.from_markup(
-                f"[bold]Backend:[/bold]  http://localhost:{backend_port}  "
+                f"[bold]Backend:[/bold]  {_base_url(backend_port, backend_host)}  "
                 f"[dim](langgraph dev — Assistant: EvoScientist)[/dim]\n"
-                f"[bold]WebUI:[/bold]    http://localhost:{webui_port}  "
+                f"[bold]WebUI:[/bold]    "
+                f"http://{_format_hostport(webui_host, webui_port)}  "
                 f"[dim](opens in your browser)[/dim]\n"
-                f"[bold]Logs:[/bold]     {_shorten(str(RUNTIME.log_file))}\n\n"
+                f"[bold]Logs:[/bold]     {_shorten(str(RUNTIME.log_file))}\n"
+                f"{remote_backend_hint}\n"
                 f"[dim]Fetching {_WEBUI_PACKAGE} via npx (first run may take a "
                 f"moment)…  Press Ctrl+C to stop.[/dim]"
             ),
@@ -218,6 +285,19 @@ def run_webui(config: Any, workspace_dir: str | None = None) -> None:
             border_style="green",
         )
     )
+    if not _is_loopback_host(backend_host):
+        console.print(
+            "[bold white on red] ⚠ PUBLIC BIND [/bold white on red] "
+            f"[bold red]Backend listening on {backend_host} — no auth, and the "
+            f"agent can run shell. Trusted networks only.[/bold red]"
+        )
+    if not _is_loopback_host(webui_host):
+        console.print(
+            "[bold white on red] ⚠ PUBLIC BIND [/bold white on red] "
+            f"[bold red]WebUI listening on {webui_host} — its API reads, writes "
+            f"and uploads workspace files and installs skills, with no auth. "
+            f"Trusted networks only.[/bold red]"
+        )
 
     popen_kwargs: dict[str, Any] = {"env": webui_env}
     if os.name == "posix":

EvoScientist/gateway/__init__.py

@@ -4,19 +4,16 @@ The gateway package is the migration seam between UI surfaces and graph
 execution. CLI, TUI, channels, and future frontends should depend on this
 package for thread/run operations instead of reaching directly into
 ``sessions.py``, ``stream.events``, or the LangGraph SDK.
+
+Backend implementations are attached lazily via :mod:`lazy_loader` (SPEC-1 /
+PEP 562): importing the shared :mod:`.types` protocols must not cascade into
+``sessions``/langgraph/langgraph_sdk, which every CLI invocation would pay.
 """
 
-from . import background_runs
-from .local import LocalGraphGateway, LocalThreadStore
-from .runtime import (
-    RuntimeGatewayBackend,
-    RuntimeGateways,
-    create_runtime_gateways,
-)
-from .server import (
-    LangGraphServerGateway,
-    LangGraphServerThreadStore,
-)
+from typing import TYPE_CHECKING
+
+import lazy_loader as _lazy
+
 from .types import (
     DEFAULT_GRAPH_ID,
     GraphEvent,
@@ -29,6 +26,44 @@ from .types import (
     ThreadStore,
 )
 
+if TYPE_CHECKING:
+    # Static counterparts of the lazy attach below — type checkers don't
+    # infer names served through __getattr__.
+    from . import background_runs
+    from .local import LocalGraphGateway, LocalThreadStore
+    from .runtime import (
+        RuntimeGatewayBackend,
+        RuntimeGateways,
+        create_runtime_gateways,
+    )
+    from .server import (
+        LangGraphServerGateway,
+        LangGraphServerThreadStore,
+    )
+
+__getattr__, _attach_dir, _ = _lazy.attach(
+    __name__,
+    submodules=["background_runs"],
+    submod_attrs={
+        "local": ["LocalGraphGateway", "LocalThreadStore"],
+        "runtime": [
+            "RuntimeGatewayBackend",
+            "RuntimeGateways",
+            "create_runtime_gateways",
+        ],
+        "server": [
+            "LangGraphServerGateway",
+            "LangGraphServerThreadStore",
+        ],
+    },
+)
+
+
+def __dir__() -> list[str]:
+    # attach() only knows the lazy names; include the eager type exports too.
+    return sorted(set(_attach_dir()) | set(__all__))
+
+
 __all__ = [
     "DEFAULT_GRAPH_ID",
     "GraphEvent",

EvoScientist/gateway/background_runs.py

@@ -19,7 +19,7 @@ import asyncio
 import logging
 import threading
 import time
-from collections.abc import Callable, Mapping
+from collections.abc import Callable, Mapping, Sequence
 from dataclasses import dataclass
 from typing import TYPE_CHECKING, Protocol, TypedDict
 
@@ -75,6 +75,12 @@ class _SyncRunsClient(Protocol):
 
     def get(self, thread_id: str, run_id: str) -> Run: ...
 
+    def list(
+        self, thread_id: str, *, limit: int, offset: int, status: str
+    ) -> list[Run]: ...
+
+    def cancel_many(self, *, thread_id: str, run_ids: Sequence[str]) -> object: ...
+
 
 class SyncLangGraphClient(Protocol):
     """Sync subset of the LangGraph SDK used by background runs."""
@@ -107,6 +113,14 @@ class _AsyncRunsClient(Protocol):
 
     async def get(self, thread_id: str, run_id: str) -> Run: ...
 
+    async def list(
+        self, thread_id: str, *, limit: int, offset: int, status: str
+    ) -> list[Run]: ...
+
+    async def cancel_many(
+        self, *, thread_id: str, run_ids: Sequence[str]
+    ) -> object: ...
+
 
 class AsyncLangGraphClient(Protocol):
     """Async subset of the LangGraph SDK used by background runs."""
@@ -247,12 +261,97 @@ async def _aget_run_status(
     return run["status"]
 
 
+# Page size for enumerating a thread's runs before deletion. The SDK's
+# ``runs.list`` defaults to limit=10, which would silently skip runs on
+# threads with a longer history.
+_RUN_CANCEL_PAGE_SIZE = 100
+
+# Statuses worth cancelling; listed server-side so terminal history is
+# never paged through.
+_CANCELABLE_RUN_STATUSES = ("pending", "running")
+
+
+def _cancel_thread_runs(
+    client: SyncLangGraphClient,
+    thread_id: str,
+    *,
+    name: str,
+) -> None:
+    """Best-effort interrupt of the thread's pending/running runs.
+
+    The server's ``threads.delete`` cascade-removes queued runs from the
+    registry, but it does not interrupt a run that is already executing —
+    cancelling first sends the interrupt control message so in-flight work
+    actually stops (issue #358). It also protects cleanup paths that
+    mutate the registry without going through ``threads.delete``. The bulk
+    cancel is skipped when nothing is cancellable (the server 404s on an
+    empty cancel set), which keeps the common terminal-only path to two
+    cheap filtered GETs.
+    """
+    try:
+        run_ids: list[str] = []
+        for status in _CANCELABLE_RUN_STATUSES:
+            offset = 0
+            while True:
+                page = client.runs.list(
+                    thread_id,
+                    limit=_RUN_CANCEL_PAGE_SIZE,
+                    offset=offset,
+                    status=status,
+                )
+                run_ids.extend(run["run_id"] for run in page)
+                if len(page) < _RUN_CANCEL_PAGE_SIZE:
+                    break
+                offset += _RUN_CANCEL_PAGE_SIZE
+        if run_ids:
+            client.runs.cancel_many(
+                thread_id=thread_id, run_ids=list(dict.fromkeys(run_ids))
+            )
+    except Exception:
+        logger.warning(
+            "Failed to cancel %s runs on thread %s", name, thread_id, exc_info=True
+        )
+
+
+async def _acancel_thread_runs(
+    client: AsyncLangGraphClient,
+    thread_id: str,
+    *,
+    name: str,
+) -> None:
+    """Async variant of :func:`_cancel_thread_runs`."""
+    try:
+        run_ids: list[str] = []
+        for status in _CANCELABLE_RUN_STATUSES:
+            offset = 0
+            while True:
+                page = await client.runs.list(
+                    thread_id,
+                    limit=_RUN_CANCEL_PAGE_SIZE,
+                    offset=offset,
+                    status=status,
+                )
+                run_ids.extend(run["run_id"] for run in page)
+                if len(page) < _RUN_CANCEL_PAGE_SIZE:
+                    break
+                offset += _RUN_CANCEL_PAGE_SIZE
+        if run_ids:
+            await client.runs.cancel_many(
+                thread_id=thread_id, run_ids=list(dict.fromkeys(run_ids))
+            )
+    except Exception:
+        logger.warning(
+            "Failed to cancel %s runs on thread %s", name, thread_id, exc_info=True
+        )
+
+
 def _delete_thread(
     client: SyncLangGraphClient,
     thread_id: str,
     *,
     name: str,
 ) -> None:
+    _cancel_thread_runs(client, thread_id, name=name)
     try:
         client.threads.delete(thread_id)
     except Exception:
@@ -265,6 +364,7 @@ async def _adelete_thread(
     *,
     name: str,
 ) -> None:
+    await _acancel_thread_runs(client, thread_id, name=name)
     try:
         await client.threads.delete(thread_id)
     except Exception:

EvoScientist/gateway/local.py

@@ -19,6 +19,8 @@ from .types import (
 if TYPE_CHECKING:
     from langgraph.graph.state import CompiledStateGraph
 
+    from ..middleware.events import SessionEvents
+
 
 @dataclass(frozen=True, slots=True)
 class LocalThreadStore:
@@ -61,9 +63,16 @@ class LocalThreadStore:
 
 @dataclass(slots=True)
 class LocalGraphGateway:
-    """Gateway backed by the current in-process graph and session helpers."""
+    """Gateway backed by the current in-process graph and session helpers.
+
+    ``events`` is the frontend/session event sink for this runtime — normally
+    the same instance injected into the agent's middleware. If it is ``None``,
+    ``stream_agent_events`` creates a per-run session sink and binds it for
+    default main-agent middleware via ``RunScopedEventSink``.
+    """
 
     thread_store: ThreadStore = field(default_factory=LocalThreadStore)
+    events: SessionEvents | None = None
 
     async def create_thread(
         self,
@@ -155,6 +164,8 @@ class LocalGraphGateway:
             request.thread_id,
             metadata=request.metadata,
             media=request.media,
+            events=self.events,
+            configurable_extra=request.configurable_extra,
         )
         try:
             async for event in inner:

EvoScientist/gateway/runtime.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 
 from dataclasses import dataclass
-from typing import Literal
+from typing import TYPE_CHECKING, Literal
 
 from langgraph_sdk import get_client
 from langgraph_sdk.client import LangGraphClient
@@ -14,6 +14,9 @@ from .server import (
 )
 from .types import GraphGateway, ThreadStore
 
+if TYPE_CHECKING:
+    from ..middleware.events import SessionEvents
+
 RuntimeGatewayBackend = Literal["local", "langgraph_server"]
 
 
@@ -32,8 +35,14 @@ def create_runtime_gateways(
     graph_id: str = DEFAULT_GRAPH_ID,
     headers: dict[str, str] | None = None,
     langgraph_client: LangGraphClient | None = None,
+    events: SessionEvents | None = None,
 ) -> RuntimeGateways:
-    """Create gateway handles for CLI/TUI/serve execution."""
+    """Create gateway handles for CLI/TUI/serve execution.
+
+    ``events`` is the frontend event sink; it is attached to the local gateway
+    so the streaming path shares the same sink instance the frontend injects
+    into the agent's middleware. Server backends ignore it (headless).
+    """
     if backend == "langgraph_server":
         if base_url is None and langgraph_client is None:
             raise ValueError("base_url is required for langgraph_server gateways")
@@ -59,5 +68,5 @@ def create_runtime_gateways(
 
     return RuntimeGateways(
         thread_store=local_thread_store,
-        graph_gateway=LocalGraphGateway(thread_store=local_thread_store),
+        graph_gateway=LocalGraphGateway(thread_store=local_thread_store, events=events),
     )

EvoScientist/gateway/server.py

@@ -7,7 +7,10 @@ import uuid
 from collections.abc import AsyncIterator, Mapping
 from dataclasses import dataclass, field
 from datetime import UTC, datetime
-from typing import Any
+from typing import TYPE_CHECKING, Any
+
+if TYPE_CHECKING:
+    from ..middleware.events import SessionEvents
 
 from langchain_core.messages import BaseMessage, convert_to_messages, messages_from_dict
 from langgraph.types import Command
@@ -25,6 +28,7 @@ from ..stream.events import (
 )
 from ..stream.summarization import _find_summarization_event_payload
 from ..stream.v3_payloads import _as_raw_map, _event_namespace
+from .background_runs import _acancel_thread_runs
 from .types import (
     DEFAULT_GRAPH_ID,
     GraphEvent,
@@ -321,6 +325,10 @@ class LangGraphServerThreadStore(ThreadStore):
         return True
 
     async def delete_thread(self, thread_id: str) -> bool:
+        # Interrupt live runs first: the server's cascade delete clears
+        # queued runs from the registry but does not stop a run that is
+        # already executing (issue #358).
+        await _acancel_thread_runs(self.client, thread_id, name="thread delete")
         try:
             await self.client.threads.delete(thread_id)
         except NotFoundError:
@@ -448,6 +456,7 @@ class LangGraphServerGateway:
     thread_store: LangGraphServerThreadStore
     graph_id: str = DEFAULT_GRAPH_ID
     interrupt_wait_seconds: float = 5.0
+    events: SessionEvents | None = None
 
     def _target_graph_id(self, target: GraphTarget | None = None) -> str:
         return target.graph_id if target is not None else self.graph_id

EvoScientist/gateway/types.py

@@ -6,13 +6,15 @@ from collections.abc import AsyncIterator
 from dataclasses import dataclass
 from typing import TYPE_CHECKING, Any, Protocol, TypeAlias
 
-from langgraph.types import Command
-
 if TYPE_CHECKING:
     from langgraph.graph.state import CompiledStateGraph
+    from langgraph.types import Command
+
+    from ..middleware.events import SessionEvents
 
 GraphEvent: TypeAlias = dict[str, Any]
-GraphRunInput: TypeAlias = str | Command
+# String alias keeps this module langgraph-free at import time (~950 modules).
+GraphRunInput: TypeAlias = "str | Command"
 GraphStateValues: TypeAlias = dict[str, Any]
 DEFAULT_GRAPH_ID = "EvoScientist"
 
@@ -39,6 +41,13 @@ class RunRequest:
     metadata: dict[str, Any] | None = None
     media: list[str] | None = None
     target: GraphTarget | None = None
+    configurable_extra: dict[str, Any] | None = None
+    """Extra keys to merge into the LangGraph ``configurable`` dict alongside
+    ``thread_id`` — e.g. ``{"active_teams": [...]}`` from the TUI
+    ``/expert`` command. WebUI callers achieve the same effect via
+    ``langgraph_sdk``'s ``config.configurable`` on their own; this field is
+    the local-gateway equivalent so CLI / TUI / headless serve can bias
+    the run identically."""
 
 
 @dataclass(frozen=True, slots=True)
@@ -94,6 +103,8 @@ class ThreadStore(Protocol):
 class GraphGateway(Protocol):
     """One authority for graph runs and thread lifecycle operations."""
 
+    events: SessionEvents | None
+
     async def create_thread(
         self,
         target: GraphTarget | None = None,

EvoScientist/langgraph_dev/graphs.py

@@ -29,10 +29,19 @@ from EvoScientist.memory.agents import (
 )
 from EvoScientist.memory.types import MemorySourceType
 from EvoScientist.subagents._factory import build_async_subagent_graph
+from EvoScientist.subagents.expert_container_async import (
+    build_expert_container_async_graph,
+)
 
 writing_agent = build_async_subagent_graph("writing-agent")
 data_analysis_agent = build_async_subagent_graph("data-analysis-agent")
 scheduler = build_async_subagent_graph("scheduler")
+# Generic async container for expert-skill dispatch. One graph, parameterised
+# per invocation by the ``skill_name`` payload the main agent passes through
+# ``EvoAsyncSubAgentMiddleware.start_async_task``. Any installed expert skill
+# dispatches through this graph; the loader middleware resolves the skill
+# body at model-call time.
+expert_container_async = build_expert_container_async_graph()
 evomemory_subagent_worker = build_memory_worker_graph(MemorySourceType.SUBAGENT)
 evomemory_turn_worker = build_memory_worker_graph(MemorySourceType.TURN)
 evomemory_observation_linker = build_observation_linker_graph()

EvoScientist/langgraph_dev/http.py

@@ -552,6 +552,63 @@ async def cancel_recoverable_run(request: Request) -> JSONResponse:
     return JSONResponse(receipt)
 
 
+async def get_teams(_request: Request) -> JSONResponse:
+    """Return installed expert skills as ``{teams: [...]}`` for the WebUI gallery.
+
+    A "team" in the WebUI vocabulary is an installed expert skill — a skill
+    directory carrying a sibling ``EXPERT.md`` (or, on the deprecated path,
+    ``type: expert`` SKILL.md frontmatter). The response is a curated,
+    gallery-safe projection: name + description, plus optional ``byline`` /
+    ``capability_tags`` / ``avatar_hint`` when the skill populates them.
+
+    Cards for experts on the current contract carry name + description only:
+    the decoration fields were actor metadata in SKILL.md frontmatter, which
+    that contract removes rather than relocates (``EXPERT.md`` has no
+    frontmatter to hold them). The omit-when-unpopulated projection below is
+    what makes those cards degrade rather than break; restoring richer cards
+    means sourcing decoration from index metadata, not re-adding frontmatter
+    fields.
+
+    Backend implementation details (SKILL.md body / system prompt, role
+    line, tool list, source tier, filesystem path,
+    tags) are intentionally NOT projected. The gallery only needs
+    identity + descriptor fields to render the card; anything richer
+    belongs in a dedicated info endpoint.
+
+    Sourced from ``list_expert_skills(include_system=True)`` so
+    first-party experts shipped as builtin skills surface alongside
+    workspace/global installs.
+
+    Offloaded to a thread because the skill loader does synchronous
+    filesystem walking + yaml parsing, which langgraph-dev's
+    ``blockbuster`` middleware refuses on the async event loop.
+
+    Response shape (each entry): ``{name, description, byline?,
+    capability_tags?, avatar_hint?}`` — the WebUI gallery consumes these.
+    """
+    from EvoScientist.tools.skills_manager import list_expert_skills
+
+    experts = await asyncio.to_thread(list_expert_skills, True)
+    teams = []
+    for info in experts:
+        entry = {
+            "name": info.name,
+            "description": info.description,
+        }
+        # Optional gallery fields — omit when unpopulated so the WebUI
+        # card degrades gracefully (SkillInfo defaults `byline` /
+        # `avatar_hint` to "" and `capability_tags` to [], which we
+        # treat as "not declared").
+        if info.byline:
+            entry["byline"] = info.byline
+        if info.capability_tags:
+            entry["capability_tags"] = list(info.capability_tags)
+        if info.avatar_hint:
+            entry["avatar_hint"] = info.avatar_hint
+        teams.append(entry)
+    return JSONResponse({"teams": teams})
+
+
 app = Starlette(
     routes=[
         Route("/api/models", get_models, methods=["GET"]),
@@ -575,5 +632,6 @@ app = Starlette(
         Route("/internal/workspace-scopes/by-thread/{thread_id}", delete_workspace_scope, methods=["DELETE"]),
         Route("/internal/workspace-scopes/{scope_id}/runs/reserve", reserve_workspace_run, methods=["POST"]),
         Route("/internal/workspace-scopes/{scope_id}/runs/{run_request_id}", bind_workspace_run, methods=["PATCH"]),
+        Route("/api/teams", get_teams, methods=["GET"]),
     ]
 )

EvoScientist/langgraph_dev/langgraph.json

@@ -5,6 +5,7 @@
         "writing-agent": "EvoScientist.langgraph_dev.graphs:writing_agent",
         "data-analysis-agent": "EvoScientist.langgraph_dev.graphs:data_analysis_agent",
         "scheduler": "EvoScientist.langgraph_dev.graphs:scheduler",
+        "expert-container-async": "EvoScientist.langgraph_dev.graphs:expert_container_async",
         "evomemory-subagent-worker": "EvoScientist.langgraph_dev.graphs:evomemory_subagent_worker",
         "evomemory-turn-worker": "EvoScientist.langgraph_dev.graphs:evomemory_turn_worker",
         "evomemory-observation-linker": "EvoScientist.langgraph_dev.graphs:evomemory_observation_linker",

EvoScientist/langgraph_dev/manager.py

@@ -12,6 +12,7 @@ Mirrors the lifecycle pattern used by ``ccproxy_manager.py``.
 from __future__ import annotations
 
 import atexit
+import hashlib
 import json
 import logging
 import os
@@ -21,6 +22,7 @@ import subprocess
 import threading
 import time
 from dataclasses import dataclass
+from dataclasses import fields as dataclass_fields
 from pathlib import Path
 
 import httpx
@@ -110,15 +112,56 @@ def needs_langgraph_dev(config: EvoScientistConfig) -> bool:
 _LOCK = threading.RLock()
 
 
+# Set by ``ensure_langgraph_dev`` when it reuses a keepalive server whose
+# recorded launch-time config fingerprint differs from the current effective
+# config. The CLI reads it after startup to surface a "restart to apply"
+# hint — the server itself is never restarted automatically.
+CONFIG_DRIFT_SINCE_LAUNCH = False
+
+
 # Default port shared with the Ai4Sci-Web recoverable runtime.
 # Overridable per-call via ``start_langgraph_dev(port=...)`` /
 # ``ensure_langgraph_dev`` (which reads ``config.langgraph_dev_port``) and the
 # corresponding url= field on AsyncSubAgent specs.
 _DEFAULT_PORT = 3076
 
+# Default bind interface — loopback, matching ``config.langgraph_dev_host``.
+# SECURITY: this is the unauthenticated agent API; launchers print a PUBLIC
+# BIND banner while it is exposed.
+_DEFAULT_HOST = "127.0.0.1"
 
-def _base_url(port: int = _DEFAULT_PORT) -> str:
-    return f"http://localhost:{port}"
+# Wildcard bind addresses: the server listens on every interface, but you
+# cannot meaningfully *connect* to them (0.0.0.0 is routed to loopback on
+# Linux and outright rejected on Windows), so clients target loopback instead.
+_WILDCARD_HOSTS = frozenset({"0.0.0.0", "::", ""})
+
+
+def _probe_host(host: str = _DEFAULT_HOST) -> str:
+    """Map a bind address to one a client can actually connect to.
+
+    A wildcard bind includes loopback, so clients use ``127.0.0.1``; a
+    specific interface is returned as-is — loopback would not reach it.
+    """
+    return "127.0.0.1" if host in _WILDCARD_HOSTS else host
+
+
+def _is_loopback_host(host: str) -> bool:
+    """Return True if binding ``host`` keeps the server unreachable off-box.
+
+    Drives the PUBLIC BIND warning, so it is conservative: anything not
+    provably loopback counts as exposed.
+    """
+    return host.strip().lower() in {"127.0.0.1", "::1", "localhost"}
+
+
+def _format_hostport(host: str, port: int) -> str:
+    """Render ``host:port`` for a URL, bracketing IPv6 literals per RFC 3986."""
+    probe = _probe_host(host)
+    return f"[{probe}]:{port}" if ":" in probe else f"{probe}:{port}"
+
+
+def _base_url(port: int = _DEFAULT_PORT, host: str = _DEFAULT_HOST) -> str:
+    return f"http://{_format_hostport(host, port)}"
 
 
 # Default rollover threshold for ``RUNTIME.log_file`` — once the log
@@ -177,9 +220,17 @@ class WorkspaceMismatchError(RuntimeError):
     """
 
 
-def _write_workspace_sidecar(workspace_dir: Path, pid: int) -> None:
+def _write_workspace_sidecar(
+    workspace_dir: Path,
+    pid: int,
+    config_fingerprint: str | None = None,
+    deploy_mode: bool | None = None,
+) -> None:
     """Record the workspace + pid of the langgraph dev we just started.
 
+    ``config_fingerprint`` (optional) captures the launch-time config subset
+    the server consumed; keepalive reuse compares it to detect drift.
+
     Atomic write via temp-file + ``os.replace``: without this, a concurrent
     reader could observe a partially-written file, fail JSON parse, and
     silently downgrade to the "no sidecar" fallback path — which skips the
@@ -194,9 +245,12 @@ def _write_workspace_sidecar(workspace_dir: Path, pid: int) -> None:
     try:
         RUNTIME.pid_dir.mkdir(parents=True, exist_ok=True)
         tmp = RUNTIME.workspace_sidecar.with_suffix(".json.tmp")
-        tmp.write_text(
-            json.dumps({"workspace": str(workspace_dir), "pid": pid}), encoding="utf-8"
-        )
+        payload: dict = {"workspace": str(workspace_dir), "pid": pid}
+        if config_fingerprint is not None:
+            payload["config_fingerprint"] = config_fingerprint
+        if deploy_mode is not None:
+            payload["deploy_mode"] = deploy_mode
+        tmp.write_text(json.dumps(payload), encoding="utf-8")
         os.replace(tmp, RUNTIME.workspace_sidecar)
     except OSError as exc:
         logger.warning(
@@ -331,32 +385,37 @@ def is_langgraph_dev_running(
     base_url: str | None = None,
     *,
     port: int = _DEFAULT_PORT,
+    host: str = _DEFAULT_HOST,
 ) -> bool:
     """Check whether a langgraph dev API is already serving at ``base_url``.
 
-    ``base_url`` overrides ``port`` when given.
+    ``base_url`` overrides ``port``/``host`` when given.
     """
-    url = base_url or _base_url(port)
+    url = base_url or _base_url(port, host)
     try:
         return httpx.get(f"{url}/ok", timeout=1.0).status_code == 200
     except (httpx.TransportError, OSError):
         return False
 
 
-def _is_port_occupied(port: int) -> bool:
-    """Return True if anything is listening on ``port`` (TCP, IPv4)."""
+def _is_port_occupied(port: int, host: str = _DEFAULT_HOST) -> bool:
+    """Return True if anything is listening on ``host:port`` (TCP)."""
     import socket as _socket
 
-    s = _socket.socket(_socket.AF_INET, _socket.SOCK_STREAM)
+    probe = _probe_host(host)
+    family = _socket.AF_INET6 if ":" in probe else _socket.AF_INET
+    s = _socket.socket(family, _socket.SOCK_STREAM)
     try:
         s.settimeout(0.5)
         # connect_ex returns 0 on success (something accepted), nonzero otherwise
-        return s.connect_ex(("127.0.0.1", port)) == 0
+        return s.connect_ex((probe, port)) == 0
     finally:
         s.close()
 
 
-def _wait_for_port_release(port: int, timeout: float = 10.0) -> bool:
+def _wait_for_port_release(
+    port: int, timeout: float = 10.0, host: str = _DEFAULT_HOST
+) -> bool:
     """Poll until ``port`` is released or ``timeout`` elapses.
 
     Used after ``stop_langgraph_dev`` / ``_kill_owned_stale_process`` to
@@ -364,13 +423,13 @@ def _wait_for_port_release(port: int, timeout: float = 10.0) -> bool:
     True if the port is free, False on timeout.
     """
     deadline = time.monotonic() + timeout
-    while _is_port_occupied(port) and time.monotonic() < deadline:
+    while _is_port_occupied(port, host) and time.monotonic() < deadline:
         time.sleep(0.5)
-    return not _is_port_occupied(port)
+    return not _is_port_occupied(port, host)
 
 
-def _can_bind_port(port: int) -> bool:
-    """Return True if a fresh ``bind()`` to ``port`` succeeds right now.
+def _can_bind_port(port: int, host: str = _DEFAULT_HOST) -> bool:
+    """Return True if a fresh ``bind()`` to ``host:port`` succeeds right now.
 
     More reliable than ``_is_port_occupied`` when the previous listener has
     just exited: ``connect_ex`` can already report "free" while ``bind()``
@@ -378,12 +437,17 @@ def _can_bind_port(port: int) -> bool:
     (TIME_WAIT for accepted connections, SO_REUSEADDR rules, etc.). This
     actually attempts the bind that langgraph dev would attempt, then
     closes immediately.
+
+    Binds the *literal* ``host`` — not ``_probe_host(host)`` — because this
+    must replicate the server's own bind: a loopback probe can succeed while
+    the real wildcard bind still fails on another interface's conflict.
     """
     import socket as _socket
 
-    s = _socket.socket(_socket.AF_INET, _socket.SOCK_STREAM)
+    family = _socket.AF_INET6 if ":" in host else _socket.AF_INET
+    s = _socket.socket(family, _socket.SOCK_STREAM)
     try:
-        s.bind(("127.0.0.1", port))
+        s.bind((host, port))
         return True
     except OSError:
         return False
@@ -394,7 +458,9 @@ def _can_bind_port(port: int) -> bool:
             pass
 
 
-def _wait_for_port_bindable(port: int, timeout: float = 60.0) -> bool:
+def _wait_for_port_bindable(
+    port: int, timeout: float = 60.0, host: str = _DEFAULT_HOST
+) -> bool:
     """Poll until a real ``bind()`` to ``port`` can succeed, or timeout.
 
     Use this immediately before ``subprocess.Popen("langgraph dev")`` —
@@ -408,7 +474,7 @@ def _wait_for_port_bindable(port: int, timeout: float = 60.0) -> bool:
     """
     deadline = time.monotonic() + timeout
     while time.monotonic() < deadline:
-        if _can_bind_port(port):
+        if _can_bind_port(port, host):
             return True
         time.sleep(0.5)
     return False
@@ -523,6 +589,186 @@ def _kill_owned_stale_process(port: int) -> bool:
     return True
 
 
+# Config fields that provably never reach the langgraph dev subprocess:
+# the channel stack + STT run in the CLI process, display/workspace/frontend
+# knobs shape the CLI itself, and keepalive is a lifecycle flag. Everything
+# NOT listed here counts toward the drift fingerprint, so a newly added
+# config field defaults to "affects the server" — the failure mode is a
+# spurious restart hint, never silent staleness.
+# Packaged sub-agent specs — consumed at graph build; module constant so
+# tests can redirect it.
+_SUBAGENTS_DIR = Path(__file__).resolve().parent.parent / "subagents"
+
+_FINGERPRINT_EXCLUDED_PREFIXES = (
+    "channel_",
+    "imessage_",
+    "telegram_",
+    "discord_",
+    "slack_",
+    "feishu_",
+    "wechat_",
+    "dingtalk_",
+    "email_",
+    "qq_",
+    "signal_",
+    "stt_",
+)
+_FINGERPRINT_EXCLUDED_FIELDS = frozenset(
+    {
+        "require_mention",
+        "text_chunk_limit",
+        "allowed_channels",
+        "dm_policy",
+        "shared_webhook_port",
+        "show_thinking",
+        "ui_backend",
+        "log_level",
+        "default_mode",
+        "default_workdir",
+        "webui_port",
+        "webui_host",
+        "langgraph_dev_keepalive",
+        "shell_allow_list",
+    }
+)
+
+
+def _server_config_fingerprint(config: EvoScientistConfig) -> str:
+    """Hash of everything the langgraph dev subprocess consumes at launch.
+
+    Deployed graphs read config once at import (``subagents/_factory.py``,
+    ``EvoScientist.py``), so a keepalive server keeps serving those values
+    until restarted. Iterates the full ``EvoScientistConfig`` field list
+    minus the explicit exclusion set above — a new config field counts
+    toward drift by default — and folds in ``mcp.yaml`` plus the packaged
+    ``subagents/*.yaml``, which are consumed at graph build too. Secrets
+    only feed a truncated one-way digest; nothing recoverable is stored.
+    getattr with defaults: deploy/WebUI (and their tests) routinely hand
+    this module duck-typed config objects missing dataclass fields.
+    """
+    parts = []
+    for field in dataclass_fields(EvoScientistConfig):
+        name = field.name
+        if name in _FINGERPRINT_EXCLUDED_FIELDS or name.startswith(
+            _FINGERPRINT_EXCLUDED_PREFIXES
+        ):
+            continue
+        parts.append((name, str(getattr(config, name, None))))
+    digest = hashlib.sha256(repr(parts).encode("utf-8"))
+    try:
+        from EvoScientist.config.settings import get_config_dir
+
+        mcp_yaml = get_config_dir() / "mcp.yaml"
+        if mcp_yaml.exists():
+            digest.update(mcp_yaml.read_bytes())
+    except OSError:
+        pass
+    try:
+        for yaml_path in sorted(_SUBAGENTS_DIR.glob("*.yaml")):
+            digest.update(yaml_path.name.encode("utf-8"))
+            digest.update(yaml_path.read_bytes())
+    except OSError:
+        pass
+    return digest.hexdigest()[:16]
+
+
+def stop_recorded_server() -> int | None:
+    """Explicitly stop the langgraph dev recorded in our PID file.
+
+    Backs the user-facing ``EvoSci server stop`` command — the deliberate
+    counterpart to ``langgraph_dev_keepalive``: an opt-in server that
+    outlives its CLI needs a first-class way to stop it. Ownership = our
+    PID file + a live process whose cmdline still contains ``langgraph``
+    (same loose anti-PID-recycling match as ``_kill_owned_stale_process``,
+    with PID-file ownership as the primary guard). Holds the cross-process
+    file lock so a concurrent start can't have its fresh PID/sidecar records
+    wiped by this stop's cleanup. Kills the whole process tree, then removes
+    the PID file + sidecar. Returns the stopped pid, or ``None`` when nothing
+    was stopped (stale/corrupt files, if any, are still cleaned up).
+    """
+    try:
+        with FileLock(str(RUNTIME.lock_file), timeout=_FILE_LOCK_TIMEOUT):
+            return _stop_recorded_server_locked()
+    except FileLockTimeout:
+        logger.warning(
+            "Timed out waiting for the langgraph dev lock — another EvoSci "
+            "process is mid lifecycle change; not stopping anything."
+        )
+        return None
+
+
+def _stop_recorded_server_locked() -> int | None:
+    with _LOCK:
+        if _PROCESS is not None and _PROCESS.poll() is None:
+            pid = _PROCESS.pid
+            stop_langgraph_dev()
+            return pid
+    if not RUNTIME.pid_file.exists():
+        return None
+    try:
+        owned_pid = int(RUNTIME.pid_file.read_text(encoding="utf-8").strip())
+    except ValueError:
+        stop_langgraph_dev()  # corrupt PID file — clean it up as promised
+        return None
+    except OSError:
+        return None
+    try:
+        proc = psutil.Process(owned_pid)
+        cmdline = proc.cmdline()
+    except (psutil.NoSuchProcess, psutil.AccessDenied):
+        stop_langgraph_dev()  # dead/inaccessible — clean the stale files
+        return None
+    if not any("langgraph" in arg for arg in cmdline):
+        stop_langgraph_dev()  # pid recycled by a foreign process — files only
+        return None
+    try:
+        children = proc.children(recursive=True)
+        for child in children:
+            try:
+                child.terminate()
+            except (psutil.NoSuchProcess, psutil.AccessDenied):
+                pass
+        proc.terminate()
+        try:
+            proc.wait(timeout=5)
+        except psutil.TimeoutExpired:
+            for child in children:
+                try:
+                    child.kill()
+                except (psutil.NoSuchProcess, psutil.AccessDenied):
+                    pass
+            proc.kill()
+        # The parent exiting promptly doesn't prove its workers did — sweep
+        # the pre-kill snapshot for survivors.
+        for child in children:
+            try:
+                if child.is_running():
+                    child.kill()
+            except (psutil.NoSuchProcess, psutil.AccessDenied):
+                pass
+    except (psutil.NoSuchProcess, psutil.AccessDenied):
+        pass
+    stop_langgraph_dev()
+    return owned_pid
+
+
+def _pid_serves_port(pid: object, port: int) -> bool:
+    """Best-effort check that ``pid`` is a langgraph dev serving ``port``.
+
+    Used to attribute an occupied port to the sidecar's recorded server
+    before printing its details — avoids blaming a stale record. Relies on
+    ``--port`` always being in ``start_langgraph_dev``'s argv, not on
+    port→PID mapping (root-only on macOS via psutil).
+    """
+    if not isinstance(pid, int) or isinstance(pid, bool) or pid <= 0:
+        return False
+    try:
+        cmdline = psutil.Process(pid).cmdline()
+    except (psutil.NoSuchProcess, psutil.AccessDenied):
+        return False
+    return any("langgraph" in arg for arg in cmdline) and str(port) in cmdline
+
+
 def _packaged_langgraph_config() -> Path:
     """Return path to the package-shipped ``langgraph.json``.
 
@@ -544,10 +790,12 @@ def start_langgraph_dev(
     workspace_dir: Path | None = None,
     *,
     port: int = _DEFAULT_PORT,
+    host: str = _DEFAULT_HOST,
     file_persistence: bool = True,
     jobs_per_worker: int = 10,
     deploy_mode: bool = False,
     tunnel: bool = False,
+    config_fingerprint: str | None = None,
 ) -> subprocess.Popen:
     """Start langgraph dev as a background subprocess.
 
@@ -557,6 +805,9 @@ def start_langgraph_dev(
             (``CustomSandboxBackend`` derives its workspace root from cwd via
             ``paths.WORKSPACE_ROOT``). Defaults to ``Path.cwd()``.
         port: TCP port to bind. Defaults to 3076.
+        host: Network interface to bind. Defaults to loopback. SECURITY:
+            widening this exposes an unauthenticated API whose agent can run
+            shell commands — only pass ``0.0.0.0`` on trusted networks.
         file_persistence: When True (default), langgraph dev writes its full
             ``.langgraph_api/`` cache so async-task / Store / scheduler state
             survives subprocess restarts. Set False to suppress periodic
@@ -609,7 +860,9 @@ def start_langgraph_dev(
     # only verifies PID-file ownership, so absence of a match conflates "stale
     # TIME_WAIT" with "foreign process". Falling through to the bind poll
     # disambiguates by behavior — TIME_WAIT clears, foreign listeners don't.
-    if not is_langgraph_dev_running(port=port) and _is_port_occupied(port):
+    if not is_langgraph_dev_running(port=port, host=host) and _is_port_occupied(
+        port, host
+    ):
         if _kill_owned_stale_process(port):
             logger.warning(
                 "Cleaned up stale langgraph dev (pid from %s) on port %d",
@@ -620,7 +873,7 @@ def start_langgraph_dev(
             # several seconds before fully releasing it. Poll until the port
             # is genuinely free so the upcoming bind() doesn't race a
             # half-released socket and crash with "Port already in use".
-            _wait_for_port_release(port)
+            _wait_for_port_release(port, host=host)
         else:
             # No owned stale PID — could be foreign or kernel-only TIME_WAIT
             # from a previous subprocess. Defer to the bind poll below.
@@ -638,9 +891,9 @@ def start_langgraph_dev(
     # "Port already in use" even though our pre-checks passed. By probing
     # the same operation langgraph dev will do, we either wait it out or
     # fail clearly with an actionable message. 60s covers macOS TIME_WAIT.
-    if not _wait_for_port_bindable(port):
+    if not _wait_for_port_bindable(port, host=host):
         raise RuntimeError(
-            f"Port {port} cannot be bound after waiting 60s (kernel TIME_WAIT "
+            f"{host}:{port} cannot be bound after waiting 60s (kernel TIME_WAIT "
             f"or another process holds it). Free the port with `lsof -ti:{port}`, "
             f"or change ports with: `EvoSci config set langgraph_dev_port <other-port>`"
         )
@@ -713,6 +966,25 @@ def start_langgraph_dev(
     sub_env.pop("EVOSCIENTIST_DEPLOY_MODE", None)
     sub_env["EVOSCIENTIST_DEPLOY_MODE"] = "full" if deploy_mode else "stripped"
 
+    # Propagate the effective bind port into the subprocess's config resolution
+    # via the standard ``EVOSCIENTIST_LANGGRAPH_DEV_PORT`` override (see
+    # ``EvoScientist/config/settings.py``). Without this, ``EvoSci deploy
+    # --port X`` binds to X but the deployed main agent still reads
+    # ``cfg.langgraph_dev_port`` from disk and dispatches self-loop async
+    # tasks (start_async_task → http://localhost:{cfg.port}) to whatever
+    # the config file says — which desyncs from the bind port whenever
+    # ``--port`` differs from the persisted ``langgraph_dev_port``, and
+    # every async subagent launch fails with "All connection attempts failed".
+    # ``get_effective_config`` treats ``EVOSCIENTIST_*`` shell values as
+    # authoritative over any workspace ``.env`` (see its docstring), so a
+    # ``.env`` in the subprocess cwd cannot shadow the caller-resolved port.
+    sub_env["EVOSCIENTIST_LANGGRAPH_DEV_PORT"] = str(port)
+    # Same reasoning for the bind interface: the deployed agent resolves its
+    # self-dispatch URL from ``cfg.langgraph_dev_host``, so a host resolved by
+    # this caller (``EvoSci deploy --host X``) must reach the subprocess too,
+    # or async sub-agent launches would target whatever the config file says.
+    sub_env["EVOSCIENTIST_LANGGRAPH_DEV_HOST"] = host
+
     try:
         logger.info("Starting langgraph dev with CLI: %s", exe)
         proc = subprocess.Popen(
@@ -721,6 +993,8 @@ def start_langgraph_dev(
                 "dev",
                 "--config",
                 str(config_file),
+                "--host",
+                host,
                 "--port",
                 str(port),
                 "--n-jobs-per-worker",
@@ -743,7 +1017,12 @@ def start_langgraph_dev(
         except Exception:
             pass
     RUNTIME.pid_file.write_text(str(proc.pid), encoding="utf-8")
-    _write_workspace_sidecar(workspace_dir=workspace_dir, pid=proc.pid)
+    _write_workspace_sidecar(
+        workspace_dir=workspace_dir,
+        pid=proc.pid,
+        config_fingerprint=config_fingerprint,
+        deploy_mode=deploy_mode,
+    )
     global _PROCESS_WORKSPACE
     _PROCESS = proc
     _PROCESS_WORKSPACE = workspace_dir
@@ -773,9 +1052,9 @@ def start_langgraph_dev(
                 f"langgraph dev exited immediately with code {proc.returncode}.\n"
                 f"Log tail:\n{tail}"
             )
-        if is_langgraph_dev_running(port=port):
+        if is_langgraph_dev_running(port=port, host=host):
             logger.info(
-                "langgraph dev started on %s (pid=%d)", _base_url(port), proc.pid
+                "langgraph dev started on %s (pid=%d)", _base_url(port, host), proc.pid
             )
             return proc
         time.sleep(0.5)
@@ -920,7 +1199,8 @@ def ensure_langgraph_dev(
     still chat with sync sub-agents; only async sub-agent calls and EvoMemory
     background workers will fail.
     """
-    global _ASYNC_SUBAGENTS_AVAILABLE
+    global _ASYNC_SUBAGENTS_AVAILABLE, CONFIG_DRIFT_SINCE_LAUNCH
+    CONFIG_DRIFT_SINCE_LAUNCH = False
 
     if not needs_langgraph_dev(config):
         _ASYNC_SUBAGENTS_AVAILABLE = False
@@ -959,8 +1239,10 @@ def _ensure_langgraph_dev_locked(
     workspace_dir: Path | str | None,
 ) -> subprocess.Popen | None:
     """Locked critical section of ``ensure_langgraph_dev`` — must hold ``_LOCK``."""
-    global _ASYNC_SUBAGENTS_AVAILABLE
+    global _ASYNC_SUBAGENTS_AVAILABLE, CONFIG_DRIFT_SINCE_LAUNCH
+    config_fp = _server_config_fingerprint(config)
     port = int(getattr(config, "langgraph_dev_port", _DEFAULT_PORT))
+    host = str(getattr(config, "langgraph_dev_host", _DEFAULT_HOST) or _DEFAULT_HOST)
     file_persistence = bool(getattr(config, "langgraph_dev_file_persistence", True))
     jobs_per_worker = int(getattr(config, "langgraph_dev_jobs_per_worker", 10))
 
@@ -993,10 +1275,10 @@ def _ensure_langgraph_dev_locked(
         # and abort with a hard "non-langgraph process" error — turning a
         # clean owned restart into a permanent async-disable. Wait inline for
         # the kernel to release the port before continuing.
-        _wait_for_port_release(port)
+        _wait_for_port_release(port, host=host)
         _ASYNC_SUBAGENTS_AVAILABLE = False  # cleared until restart succeeds
 
-    if is_langgraph_dev_running(port=port):
+    if is_langgraph_dev_running(port=port, host=host):
         # If WE own the running process AND it's still alive, workspace was
         # already verified above via _PROCESS_WORKSPACE comparison. Otherwise
         # — we never owned it (EvoSci deploy in another terminal, or a
@@ -1012,17 +1294,38 @@ def _ensure_langgraph_dev_locked(
             if sidecar is not None:
                 recorded = Path(sidecar["workspace"]).resolve()
                 if recorded != ws_path.resolve():
+                    hint = ""
+                    if getattr(config, "langgraph_dev_keepalive", False):
+                        # Only under keepalive can the server be an ownerless
+                        # leftover; without the flag the mismatch means a live
+                        # session, where a stop suggestion would be misleading.
+                        # Point at `EvoSci server stop` (not a raw kill): it
+                        # verifies ownership and cleans the PID/sidecar files,
+                        # so no stale records are left behind.
+                        hint = (
+                            " If it is a leftover keepalive server, stop it"
+                            " with: EvoSci server stop."
+                        )
                     raise WorkspaceMismatchError(
                         f"An EvoSci langgraph dev is already running on "
-                        f"{_base_url(port)} for workspace {recorded}, but the "
+                        f"{_base_url(port, host)} for workspace {recorded}, but the "
                         f"current process requested workspace {ws_path}. "
                         f"Stop the other EvoSci session (deploy / TUI / serve) "
-                        f"or rerun with --workdir {recorded}."
+                        f"or rerun with --workdir {recorded}." + hint
+                    )
+                recorded_fp = sidecar.get("config_fingerprint")
+                if isinstance(recorded_fp, str) and recorded_fp != config_fp:
+                    CONFIG_DRIFT_SINCE_LAUNCH = True
+                    logger.warning(
+                        "Config changed since the running langgraph dev was "
+                        "launched — async sub-agents still use the old "
+                        "settings until the server is restarted "
+                        "(EvoSci server stop)."
                     )
                 logger.info(
                     "Reusing externally-managed langgraph dev on %s; sidecar "
                     "confirms matching workspace %s.",
-                    _base_url(port),
+                    _base_url(port, host),
                     recorded,
                 )
             else:
@@ -1034,11 +1337,13 @@ def _ensure_langgraph_dev_locked(
                     "workspace sidecar, cannot verify it matches the requested "
                     "%s. Async sub-agents may operate on a different workspace's "
                     "files.",
-                    _base_url(port),
+                    _base_url(port, host),
                     ws_path,
                 )
         else:
-            logger.info("langgraph dev already running on %s, reusing", _base_url(port))
+            logger.info(
+                "langgraph dev already running on %s, reusing", _base_url(port, host)
+            )
         _ASYNC_SUBAGENTS_AVAILABLE = True
         return None
 
@@ -1046,8 +1351,10 @@ def _ensure_langgraph_dev_locked(
         proc = start_langgraph_dev(
             workspace_dir=ws_path,
             port=port,
+            host=host,
             file_persistence=file_persistence,
             jobs_per_worker=jobs_per_worker,
+            config_fingerprint=config_fp,
         )
     except (FileNotFoundError, RuntimeError) as exc:
         # Startup failed — keep async subagents disabled so the main agent
@@ -1064,5 +1371,10 @@ def _ensure_langgraph_dev_locked(
         return None
 
     _ASYNC_SUBAGENTS_AVAILABLE = True
-    atexit.register(stop_langgraph_dev, proc)
+    if getattr(config, "langgraph_dev_keepalive", False):
+        # Keepalive: leave the server (plus PID file + sidecar) behind on CLI
+        # exit so the next start in this workspace reuses it instantly.
+        logger.info("langgraph_dev_keepalive enabled — server will outlive this CLI.")
+    else:
+        atexit.register(stop_langgraph_dev, proc)
     return proc

EvoScientist/langgraph_dev/sdk.py

@@ -6,20 +6,47 @@ import os
 from collections.abc import Mapping
 
 DEFAULT_LANGGRAPH_DEV_PORT = 3076
+# Mirrors ``config.langgraph_dev_host`` / ``manager._DEFAULT_HOST``. The value
+# only matters as a stand-in for the *bind* host — ``_format_hostport`` runs it
+# through ``_probe_host``, so both this and "0.0.0.0" yield the same client URL.
+DEFAULT_LANGGRAPH_DEV_HOST = "127.0.0.1"
 LANGGRAPH_DEV_AUTH_HEADERS = {"x-auth-scheme": "langsmith"}
 
 
-def langgraph_dev_url(config: object | None = None, *, port: int | None = None) -> str:
-    """Return the local langgraph-dev base URL for a config or explicit port."""
+def langgraph_dev_url(
+    config: object | None = None,
+    *,
+    port: int | None = None,
+    host: str | None = None,
+) -> str:
+    """Return the local langgraph-dev base URL for a config or explicit port/host.
+
+    An explicit ``LANGGRAPH_SERVER_URL`` (container / prod deploy) wins when no
+    port or host override is supplied. Otherwise the configured bind interface is
+    mapped through ``manager._probe_host``: a wildcard bind (``0.0.0.0``) still
+    resolves to loopback here, while a specific interface is honored so
+    self-dispatch keeps working when the server is pinned to one address.
+    """
     runtime_url = os.environ.get("LANGGRAPH_SERVER_URL", "").strip().rstrip("/")
-    if port is None and runtime_url:
+    if port is None and host is None and runtime_url:
         return runtime_url
+
+    from .manager import _format_hostport
+
     selected_port = (
         int(port)
         if port is not None
         else int(getattr(config, "langgraph_dev_port", DEFAULT_LANGGRAPH_DEV_PORT))
     )
-    return f"http://localhost:{selected_port}"
+    selected_host = (
+        host
+        if host is not None
+        else str(
+            getattr(config, "langgraph_dev_host", DEFAULT_LANGGRAPH_DEV_HOST)
+            or DEFAULT_LANGGRAPH_DEV_HOST
+        )
+    )
+    return f"http://{_format_hostport(selected_host, selected_port)}"
 
 
 def configured_langgraph_dev_url() -> str:

EvoScientist/llm/__init__.py

@@ -18,6 +18,7 @@ __getattr__, __dir__, __all__ = _lazy.attach(
         "context_window",
         "models",
         "patches",
+        "registry",
         "contracts",
         "config_admin",
         "configuration",
@@ -35,9 +36,12 @@ __getattr__, __dir__, __all__ = _lazy.attach(
             "resolve_context_window",
         ],
         "models": [
+            "get_chat_model",
+        ],
+        # Registry data resolves without the langchain/provider-SDK stack.
+        "registry": [
             "DEFAULT_MODEL",
             "MODELS",
-            "get_chat_model",
             "get_model_info",
             "get_models_for_provider",
             "list_models",

EvoScientist/llm/context_window.py

@@ -17,11 +17,15 @@ _KNOWN_MODEL_CONTEXT_WINDOWS: dict[str, int] = {
     # Qwen 3.6 open-source variants — exceptions to the ``qwen3.6`` family.
     "qwen3.6-27b": 262_000,
     "qwen3.6-35b-a3b": 262_000,
+    # Qwen 3.8 closed-source tiers — Max flagship and Flash (1M).
+    "qwen3.8-max": 1_000_000,
+    "qwen3.8-flash": 1_000_000,
     # Qwen 3.7 closed-source tiers — Max flagship and Plus (1M).
     "qwen3.7-max": 1_000_000,
     "qwen3.7-plus": 1_000_000,
-    # xAI Grok — per-model windows (build-0.1: 256K, 4.5: 500K).
+    # xAI Grok — per-model windows (build-0.1: 256K, 4.5/4.6: 500K).
     "grok-build-0.1": 256_000,
+    "grok-4.6": 500_000,
     "grok-4.5": 500_000,
     # Claude Haiku 4.5 — exception to the ``claude-`` family (200K, not 1M).
     "claude-haiku-4-5": 200_000,
@@ -29,10 +33,15 @@ _KNOWN_MODEL_CONTEXT_WINDOWS: dict[str, int] = {
     # Covers OpenRouter ``minimax/minimax-m3`` (via split('/')[-1]) and direct
     # ``MiniMax-M3`` (via lowercased exact match).
     "minimax-m3": 1_000_000,
-    # Zhipu GLM-5.2 — 1M context, an exception to the ``glm-5`` family (203K).
-    # Matches OpenRouter ``z-ai/glm-5.2`` via split('/')[-1].
+    # Zhipu GLM-5.3/5.2 — 1M context, exceptions to the ``glm-5`` family (203K).
+    # Matches OpenRouter ``z-ai/glm-5.x`` via split('/')[-1].
+    "glm-5.3": 1_000_000,
+    "glm-5.3-flash": 1_000_000,
     "glm-5.2": 1_000_000,
-    # Tencent Hunyuan HY3 — 262K context (OpenRouter ``tencent/hy3``).
+    # Volcengine Coding Plan's OpenAI-compatible alias for GLM-5.2.
+    "glm-5-2": 1_000_000,
+    # Tencent Hunyuan — HY4 preview 1M, HY3 262K (OpenRouter ``tencent/hy*``).
+    "hy4-preview": 1_048_576,
     "hy3": 262_000,
 }
 
@@ -42,12 +51,17 @@ _KNOWN_MODEL_CONTEXT_WINDOWS: dict[str, int] = {
 _KNOWN_MODEL_FAMILIES: list[tuple[str, int]] = [
     # All Claude — 1M via the ``context-1m-2025-08-07`` beta header.
     ("claude-", 1_000_000),
+    # OpenAI GPT-6 family — astra, astra-pro, future variants
+    ("gpt-6", 1_050_000),
     # OpenAI GPT-5.6 family — sol, terra, luna variants
     ("gpt-5.6", 1_050_000),
     # OpenAI GPT-5.5 family — base, pro, future variants
     ("gpt-5.5", 1_050_000),
     # Google Gemini 3.x family — flash, flash-lite, pro (1.05M). Excludes 2.5.
     ("gemini-3", 1_050_000),
+    # Moonshot Kimi K3 — 1M context; covers bare ``kimi-k3`` (native Moonshot),
+    # OpenRouter ``moonshotai/kimi-k3``, and dated slugs like ``kimi-k3-20260715``.
+    ("kimi-k3", 1_048_576),
     # Moonshot Kimi K2 family — k2.5, k2.6, k2-thinking, k2-thinking-turbo
     ("kimi-k2", 262_000),
     # Zhipu GLM-5 family — base, 5.1, 5-turbo, 5v-turbo, etc.
@@ -56,6 +70,8 @@ _KNOWN_MODEL_FAMILIES: list[tuple[str, int]] = [
     ("deepseek-v4", 1_050_000),
     # Xiaomi MiMo v2.5 family — base, pro, future variants
     ("mimo-v2.5", 1_050_000),
+    # Meta Muse Spark family — 1.1/1.2/1.3 (OpenRouter ``meta/muse-spark-*``, 1M).
+    ("muse-spark", 1_048_576),
     # Qwen 3.6 closed-source family — flash, plus, max-preview, etc.
     # Open-source ``-<size>b`` variants are 262K — listed in the dict above.
     ("qwen3.6", 1_000_000),

EvoScientist/llm/deepseek.py

@@ -0,0 +1,80 @@
+"""DeepSeek chat model integration."""
+
+from __future__ import annotations
+
+import logging
+from collections.abc import Mapping
+from typing import Any
+
+from langchain_core.language_models import LanguageModelInput
+from langchain_core.messages import AIMessage, BaseMessage
+from langchain_deepseek import ChatDeepSeek
+
+from .openai_compat import OpenAICompatContentMixin
+
+logger = logging.getLogger(__name__)
+
+DEEPSEEK_THINKING_DISABLED = {"type": "disabled"}
+
+
+def is_deepseek_thinking_disabled(
+    extra_body: Mapping[str, object] | None,
+) -> bool:
+    """Return whether a request body explicitly disables DeepSeek thinking."""
+    if not extra_body:
+        return False
+    thinking = extra_body.get("thinking")
+    return isinstance(thinking, Mapping) and thinking.get("type") == "disabled"
+
+
+def _inject_reasoning_content(
+    messages: list[BaseMessage],
+    payload: dict[str, object],
+) -> dict[str, object]:
+    """Copy captured DeepSeek reasoning into serialized assistant messages."""
+    reasoning = [
+        message.additional_kwargs.get("reasoning_content")
+        for message in messages
+        if isinstance(message, AIMessage)
+    ]
+    serialized = payload.get("messages")
+    if not isinstance(serialized, list):
+        return payload
+
+    ai_index = 0
+    for message in serialized:
+        if not isinstance(message, dict) or message.get("role") != "assistant":
+            continue
+        value = reasoning[ai_index] if ai_index < len(reasoning) else None
+        if value:
+            message["reasoning_content"] = value
+        elif "reasoning_content" not in message:
+            message["reasoning_content"] = ""
+        ai_index += 1
+    return payload
+
+
+class EvoChatDeepSeek(OpenAICompatContentMixin, ChatDeepSeek):
+    """ChatDeepSeek with EvoScientist's media and history compatibility."""
+
+    def _get_request_payload(
+        self,
+        input_: LanguageModelInput,
+        *,
+        stop: list[str] | None = None,
+        **kwargs: Any,
+    ) -> dict[str, Any]:
+        payload = super()._get_request_payload(input_, stop=stop, **kwargs)
+        if is_deepseek_thinking_disabled(self.extra_body):
+            return payload
+
+        try:
+            messages = self._convert_input(input_).to_messages()
+        except Exception:
+            logger.warning(
+                "DeepSeek reasoning passback: input conversion failed",
+                exc_info=True,
+            )
+            return payload
+
+        return _inject_reasoning_content(messages, payload)

EvoScientist/llm/errors.py

@@ -353,6 +353,7 @@ def _redact_api_keys(message: str) -> str:
 _HOST_TO_PROVIDER: dict[str, str] = {
     "api.openai.com": "openai",
     "api.anthropic.com": "anthropic",
+    "api.atlascloud.ai": "atlascloud",
     "api.deepseek.com": "deepseek",
     "api.moonshot.cn": "moonshot",
     "api.siliconflow.cn": "siliconflow",
@@ -363,6 +364,7 @@ _HOST_TO_PROVIDER: dict[str, str] = {
     "api.minimaxi.com": "minimax",
     "api.kimi.com": "kimi",  # kimi-coding shares this host
     "openrouter.ai": "openrouter",
+    "api.novita.ai": "novita",
 }
 
 
@@ -377,16 +379,22 @@ def _provider_from_model(model: Any) -> str | None:
     (``ErrorNormalizationMiddleware``) then passes the exception
     through unchanged.
     """
-    cls_module = type(model).__module__ or ""
-    if cls_module.startswith("langchain_openrouter"):
+    cls_modules = {cls.__module__ for cls in type(model).__mro__}
+
+    def _uses_sdk(module_prefix: str) -> bool:
+        return any(module.startswith(module_prefix) for module in cls_modules)
+
+    if _uses_sdk("langchain_openrouter"):
         return "openrouter"
-    if cls_module.startswith("langchain_google_genai"):
+    if _uses_sdk("langchain_google_genai"):
         return "google_genai"
-    if cls_module.startswith("langchain_openai"):
+    if _uses_sdk("langchain_deepseek"):
+        return "deepseek"
+    if _uses_sdk("langchain_openai"):
         return _lookup_host_or_compat(
             getattr(model, "openai_api_base", None), module_tag="openai"
         )
-    if cls_module.startswith("langchain_anthropic"):
+    if _uses_sdk("langchain_anthropic"):
         return _lookup_host_or_compat(
             getattr(model, "anthropic_api_url", None), module_tag="anthropic"
         )

EvoScientist/llm/models.py

@@ -1,10 +1,11 @@
 """LLM model configuration based on LangChain init_chat_model.
 
 This module provides a unified interface for creating chat model instances
-with support for multiple providers (Anthropic, OpenAI, Google GenAI, MiniMax
-(Anthropic-compatible), NVIDIA, SiliconFlow, OpenRouter, ZhipuAI, Volcengine,
-DashScope, DashScope-Code, DeepSeek, Ollama, and custom OpenAI/Anthropic-compatible
-endpoints) and convenient short names for common models.
+with support for multiple providers (Anthropic, OpenAI, Google GenAI, Atlas
+Cloud, MiniMax (Anthropic-compatible), NVIDIA, SiliconFlow, OpenRouter, Requesty,
+Novita, ZhipuAI, Volcengine, DashScope, DashScope-Code, DeepSeek, Ollama, and
+custom OpenAI/Anthropic-compatible endpoints) and convenient short names for
+common models.
 """
 
 from __future__ import annotations
@@ -15,6 +16,7 @@ import subprocess
 import warnings
 from functools import lru_cache
 from typing import Any
+from urllib.parse import urlparse
 
 from langchain.chat_models import init_chat_model
 
@@ -24,26 +26,31 @@ from ..config.settings import (
     OPENROUTER_DEFAULT_HTTP_REFERER,
 )
 from .context_window import apply_known_context_window
+from .deepseek import EvoChatDeepSeek
 from .patches import (
     _is_ccproxy_codex,
+    _patch_anthropic_strip_foreign_reasoning,
+    _patch_anthropic_structured_output,
     _patch_ccproxy_system_to_developer,
-    _patch_deepseek_reasoning_passback,
     _patch_openai_compat_content,
     _patch_openrouter_strip_responses_reasoning,
+    _patch_openrouter_structured_output,
+)
+from .registry import (
+    _ANTHROPIC_ROUTED_PROVIDERS,
+    _MODEL_ENTRIES,
+    _OPENAI_ROUTED_PROVIDERS,
+    _OPENROUTER_JSON_SCHEMA_STRUCTURED_OUTPUT_MODELS,  # noqa: F401 — re-exported
+    _THINKING_CAPABLE_PROVIDERS,
+    DEFAULT_MODEL,
+    MODELS,
+    _is_mandatory_thinking_kimi,
+    get_model_info,  # noqa: F401 — re-exported for existing import sites
+    get_models_for_provider,  # noqa: F401 — re-exported for existing import sites
+    list_model_picker_entries,  # noqa: F401 — re-exported for existing import sites
+    list_models,  # noqa: F401 — re-exported for existing import sites
+    list_models_by_provider,  # noqa: F401 — re-exported for existing import sites
 )
-
-_MINIMAX_ANTHROPIC_BASE_URL = "https://api.minimaxi.com/anthropic"
-_SILICONFLOW_BASE_URL = "https://api.siliconflow.cn/v1"
-
-_ZHIPU_BASE_URL = "https://open.bigmodel.cn/api/paas/v4"
-_ZHIPU_CODE_BASE_URL = "https://open.bigmodel.cn/api/coding/paas/v4"
-_VOLCENGINE_BASE_URL = "https://ark.cn-beijing.volces.com/api/v3"
-_DASHSCOPE_BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1"
-_DASHSCOPE_CODE_BASE_URL = "https://coding.dashscope.aliyuncs.com/v1"
-
-_DEEPSEEK_BASE_URL = "https://api.deepseek.com"
-_MOONSHOT_BASE_URL = "https://api.moonshot.cn/v1"
-_KIMI_CODING_BASE_URL = "https://api.kimi.com/coding/"
 
 # Minimum Codex CLI version advertised when no explicit override is set. Newer
 # installed versions are advertised automatically.
@@ -84,33 +91,79 @@ def _resolve_codex_client_version() -> str:
     return _CODEX_CLIENT_VERSION_FALLBACK
 
 
-# Providers routed through the OpenAI provider with a custom base_url.
-# Maps provider name → (base_url or None, env var for API key).
-_OPENAI_ROUTED_PROVIDERS: dict[str, tuple[str | None, str]] = {
-    "deepseek": (_DEEPSEEK_BASE_URL, "DEEPSEEK_API_KEY"),
-    "moonshot": (_MOONSHOT_BASE_URL, "MOONSHOT_API_KEY"),
-    "siliconflow": (_SILICONFLOW_BASE_URL, "SILICONFLOW_API_KEY"),
-    "zhipu": (_ZHIPU_BASE_URL, "ZHIPU_API_KEY"),
-    "zhipu-code": (_ZHIPU_CODE_BASE_URL, "ZHIPU_API_KEY"),
-    "volcengine": (_VOLCENGINE_BASE_URL, "VOLCENGINE_API_KEY"),
-    "dashscope": (_DASHSCOPE_BASE_URL, "DASHSCOPE_API_KEY"),
-    "dashscope-code": (_DASHSCOPE_CODE_BASE_URL, "DASHSCOPE_API_KEY"),
-    "custom-openai": (
-        None,
-        "CUSTOM_OPENAI_API_KEY",
-    ),  # base_url from CUSTOM_OPENAI_BASE_URL env
-}
 
-# Providers routed through the Anthropic provider with a custom base_url.
-# Maps provider name → (base_url or None, env var for API key).
-_ANTHROPIC_ROUTED_PROVIDERS: dict[str, tuple[str | None, str]] = {
-    "minimax": (_MINIMAX_ANTHROPIC_BASE_URL, "MINIMAX_API_KEY"),
-    "kimi-coding": (_KIMI_CODING_BASE_URL, "KIMI_API_KEY"),
-    "custom-anthropic": (None, "CUSTOM_ANTHROPIC_API_KEY"),
-}
 
-# Anthropic-routed providers that support extended thinking.
-_THINKING_CAPABLE_PROVIDERS: set[str] = {"minimax"}
+def _resolve_reasoning_effort(default: str) -> str:
+    """Return the configured reasoning effort or a provider-specific default."""
+    return os.environ.get("EVOSCIENTIST_REASONING_EFFORT", "").strip() or default
+
+
+# Qwen 3.8 Max canonical levels and documented OpenAI alias mappings:
+# https://docs.qwencloud.com/api-reference/chat/openai-chat#reasoning-effort
+_DASHSCOPE_QWEN38_REASONING_EFFORTS = frozenset(
+    {"none", "minimal", "low", "medium", "high", "xhigh", "max"}
+)
+
+
+def _validate_dashscope_reasoning_effort(
+    provider: str,
+    model_id: str,
+    effort: str,
+) -> None:
+    """Reject reasoning levels unsupported by DashScope Qwen 3.8 Max."""
+    if effort not in _DASHSCOPE_QWEN38_REASONING_EFFORTS:
+        choices = ", ".join(sorted(_DASHSCOPE_QWEN38_REASONING_EFFORTS))
+        raise ValueError(
+            f"Unsupported EVOSCIENTIST_REASONING_EFFORT={effort!r} for "
+            f"{provider} model {model_id!r}. Supported values: {choices}."
+        )
+
+
+def _apply_openai_compat_reasoning_config(
+    provider: str,
+    model_id: str,
+    kwargs: dict[str, Any],
+) -> None:
+    """Apply reasoning controls supported by OpenAI-compatible providers.
+
+    Routed providers deliberately skip the native-OpenAI branch in
+    :func:`_apply_auto_config`, because most compatible endpoints reject
+    OpenAI-only ``reasoning`` payloads.  A small subset does support the
+    standard ``reasoning_effort`` field, though:
+
+    * DashScope Qwen 3.8 Max supports ``low`` / ``medium`` / ``xhigh`` and
+      maps the OpenAI aliases (including ``none``).  Its server default is
+      extremely large, so use the standard ``medium`` level unless the user
+      selected another level.
+    * ``custom-openai`` is user-owned.  Forward an *explicit* setting only;
+      with no setting, preserve compatibility with endpoints that reject the
+      field (including many non-reasoning OpenAI-compatible APIs).
+
+    Explicit caller kwargs always win.
+    """
+    configured = os.environ.get("EVOSCIENTIST_REASONING_EFFORT", "").strip()
+    short_model_id = model_id.rsplit("/", 1)[-1]
+
+    if provider == "dashscope" and short_model_id.startswith("qwen3.8-max"):
+        if "reasoning_effort" not in kwargs:
+            effort = configured or "medium"
+            _validate_dashscope_reasoning_effort(provider, model_id, effort)
+            kwargs["reasoning_effort"] = effort
+        return
+
+    if provider == "custom-openai" and configured:
+        kwargs.setdefault("reasoning_effort", configured)
+
+
+def _is_deepseek_endpoint(base_url: str | None) -> bool:
+    """Return whether an OpenAI-compatible endpoint is DeepSeek's API."""
+    if not base_url:
+        return False
+    try:
+        return urlparse(base_url).hostname == "api.deepseek.com"
+    except ValueError:
+        return False
+
 
 _TRUTHY_ENV_VALUES = {"1", "true", "yes", "on"}
 _FALSEY_ENV_VALUES = {"0", "false", "no", "off"}
@@ -128,192 +181,6 @@ _OPENROUTER_MAX_CATEGORIES_PER_REQUEST = 2
 # LangChain move them into model_kwargs and can later leak them into SDK calls.
 _UNSUPPORTED_CHAT_MODEL_KWARGS = frozenset({"sanitize_openai_sdk_headers"})
 
-# Model registry: list of (short_name, model_id, provider)
-# Allows same short_name across different providers.
-_MODEL_ENTRIES: list[tuple[str, str, str]] = [
-    # Custom Anthropic (third-party Claude-compatible endpoints, current-gen defaults)
-    # Listed BEFORE native anthropic so MODELS dict defaults to native provider
-    ("claude-sonnet-4-6", "claude-sonnet-4-6", "custom-anthropic"),
-    ("claude-haiku-4-5", "claude-haiku-4-5", "custom-anthropic"),
-    # Custom OpenAI (third-party OpenAI-compatible endpoints, 3 defaults)
-    # Listed BEFORE native openai so MODELS dict defaults to native provider
-    ("gpt-5.5-pro", "gpt-5.5-pro", "custom-openai"),
-    ("gpt-5.5", "gpt-5.5", "custom-openai"),
-    ("gpt-5.4", "gpt-5.4", "custom-openai"),
-    ("gpt-5.3-codex", "gpt-5.3-codex", "custom-openai"),
-    ("gpt-5-mini", "gpt-5-mini", "custom-openai"),
-    # Anthropic (current generation)
-    ("claude-fable-5", "claude-fable-5", "anthropic"),
-    ("claude-opus-4-8", "claude-opus-4-8", "anthropic"),
-    ("claude-sonnet-5", "claude-sonnet-5", "anthropic"),
-    ("claude-sonnet-4-6", "claude-sonnet-4-6", "anthropic"),
-    ("claude-haiku-4-5", "claude-haiku-4-5", "anthropic"),
-    # OpenAI
-    ("gpt-5.6-sol", "gpt-5.6-sol", "openai"),
-    ("gpt-5.6-terra", "gpt-5.6-terra", "openai"),
-    ("gpt-5.6-luna", "gpt-5.6-luna", "openai"),
-    ("gpt-5.5-pro", "gpt-5.5-pro", "openai"),
-    ("gpt-5.5", "gpt-5.5", "openai"),
-    ("gpt-5.4", "gpt-5.4", "openai"),
-    ("gpt-5.4-mini", "gpt-5.4-mini", "openai"),
-    ("gpt-5.4-nano", "gpt-5.4-nano", "openai"),
-    ("gpt-5.3-codex", "gpt-5.3-codex", "openai"),
-    ("gpt-5.2-codex", "gpt-5.2-codex", "openai"),
-    ("gpt-5.2", "gpt-5.2", "openai"),
-    ("gpt-5.1", "gpt-5.1", "openai"),
-    ("gpt-5", "gpt-5", "openai"),
-    ("gpt-5-mini", "gpt-5-mini", "openai"),
-    ("gpt-5-nano", "gpt-5-nano", "openai"),
-    # Google GenAI
-    ("gemini-3.5-flash", "gemini-3.5-flash", "google-genai"),
-    ("gemini-3.1-pro", "gemini-3.1-pro-preview", "google-genai"),
-    (
-        "gemini-3.1-pro-customtools",
-        "gemini-3.1-pro-preview-customtools",
-        "google-genai",
-    ),
-    ("gemini-3.1-flash-lite", "gemini-3.1-flash-lite-preview", "google-genai"),
-    ("gemini-3-flash", "gemini-3-flash-preview", "google-genai"),
-    ("gemini-2.5-flash", "gemini-2.5-flash", "google-genai"),
-    ("gemini-2.5-flash-lite", "gemini-2.5-flash-lite", "google-genai"),
-    ("gemini-2.5-pro", "gemini-2.5-pro", "google-genai"),
-    # MiniMax (direct API — Anthropic-compatible; default: api.minimaxi.com, global: api.minimax.io)
-    ("minimax-m3", "MiniMax-M3", "minimax"),
-    ("minimax-m2.7", "MiniMax-M2.7", "minimax"),
-    ("minimax-m2.7-highspeed", "MiniMax-M2.7-highspeed", "minimax"),
-    ("minimax-m2.5", "MiniMax-M2.5", "minimax"),
-    ("minimax-m2.5-highspeed", "MiniMax-M2.5-highspeed", "minimax"),
-    # NVIDIA
-    ("nemotron-super", "nvidia/nemotron-3-super-120b-a12b", "nvidia"),
-    ("nemotron-nano", "nvidia/nemotron-3-nano-30b-a3b", "nvidia"),
-    ("glm-5.2", "z-ai/glm-5.2", "nvidia"),
-    ("glm4.7", "z-ai/glm4.7", "nvidia"),
-    ("deepseek-v3.2", "deepseek-ai/deepseek-v3.2", "nvidia"),
-    ("deepseek-v3.1", "deepseek-ai/deepseek-v3.1-terminus", "nvidia"),
-    ("kimi-k2.5", "moonshotai/kimi-k2.5", "nvidia"),
-    ("kimi-k2-thinking", "moonshotai/kimi-k2-thinking", "nvidia"),
-    ("minimax-m2.5", "minimaxai/minimax-m2.5", "nvidia"),
-    ("minimax-m2.1", "minimaxai/minimax-m2.1", "nvidia"),
-    ("qwen3.5-397b", "qwen/qwen3.5-397b-a17b", "nvidia"),
-    ("step-3.5-flash", "stepfun-ai/step-3.5-flash", "nvidia"),
-    # SiliconFlow
-    ("minimax-m2.5", "Pro/MiniMaxAI/MiniMax-M2.5", "siliconflow"),
-    ("glm-5.2", "Pro/zai-org/GLM-5.2", "siliconflow"),
-    ("glm-5", "Pro/zai-org/GLM-5", "siliconflow"),
-    ("kimi-k2.5", "Pro/moonshotai/Kimi-K2.5", "siliconflow"),
-    ("glm-4.7", "Pro/zai-org/GLM-4.7", "siliconflow"),
-    # OpenRouter
-    ("claude-fable-5", "anthropic/claude-fable-5", "openrouter"),
-    ("claude-opus-4.8", "anthropic/claude-opus-4.8", "openrouter"),
-    ("claude-opus-4.8-fast", "anthropic/claude-opus-4.8-fast", "openrouter"),
-    ("claude-sonnet-5", "anthropic/claude-sonnet-5", "openrouter"),
-    ("claude-sonnet-4.6", "anthropic/claude-sonnet-4.6", "openrouter"),
-    ("gpt-5.6-sol", "openai/gpt-5.6-sol", "openrouter"),
-    ("gpt-5.6-terra", "openai/gpt-5.6-terra", "openrouter"),
-    ("gpt-5.6-luna", "openai/gpt-5.6-luna", "openrouter"),
-    ("gpt-5.5-pro", "openai/gpt-5.5-pro", "openrouter"),
-    ("gpt-5.5", "openai/gpt-5.5", "openrouter"),
-    ("gpt-5.4", "openai/gpt-5.4", "openrouter"),
-    ("gpt-5.3-codex", "openai/gpt-5.3-codex", "openrouter"),
-    ("gemini-3.5-flash", "google/gemini-3.5-flash", "openrouter"),
-    ("gemini-3.1-pro", "google/gemini-3.1-pro-preview", "openrouter"),
-    ("gemini-3-flash", "google/gemini-3-flash-preview", "openrouter"),
-    ("kimi-k2.6", "moonshotai/kimi-k2.6", "openrouter"),
-    ("glm-5.2", "z-ai/glm-5.2", "openrouter"),
-    ("glm-5v-turbo", "z-ai/glm-5v-turbo", "openrouter"),
-    ("minimax-m3", "minimax/minimax-m3", "openrouter"),
-    ("mimo-v2.5-pro", "xiaomi/mimo-v2.5-pro", "openrouter"),
-    ("mimo-v2.5", "xiaomi/mimo-v2.5", "openrouter"),
-    ("grok-build-0.1", "x-ai/grok-build-0.1", "openrouter"),
-    ("grok-4.5", "x-ai/grok-4.5", "openrouter"),
-    ("hy3", "tencent/hy3", "openrouter"),
-    ("qwen3.7-max", "qwen/qwen3.7-max", "openrouter"),
-    ("qwen3.7-plus", "qwen/qwen3.7-plus", "openrouter"),
-    ("qwen3.6-flash", "qwen/qwen3.6-flash", "openrouter"),
-    ("qwen3.5-122b", "qwen/qwen3.5-122b-a10b", "openrouter"),
-    ("deepseek-v4-pro", "deepseek/deepseek-v4-pro", "openrouter"),
-    ("deepseek-v4-flash", "deepseek/deepseek-v4-flash", "openrouter"),
-    # Zhipu CodePlan (智谱代码计划 — coding-only endpoint)
-    ("glm-5.2", "glm-5.2", "zhipu-code"),
-    ("glm-5.1", "glm-5.1", "zhipu-code"),
-    ("glm-5", "glm-5", "zhipu-code"),
-    ("glm-5-turbo", "glm-5-turbo", "zhipu-code"),
-    ("glm-5v-turbo", "glm-5v-turbo", "zhipu-code"),
-    ("glm-4.7", "glm-4.7", "zhipu-code"),
-    # Zhipu (智谱 — general endpoint, default for simple lookups)
-    ("glm-5.2", "glm-5.2", "zhipu"),
-    ("glm-5.1", "glm-5.1", "zhipu"),
-    ("glm-5", "glm-5", "zhipu"),
-    ("glm-5-turbo", "glm-5-turbo", "zhipu"),
-    ("glm-5v-turbo", "glm-5v-turbo", "zhipu"),
-    ("glm-4.7", "glm-4.7", "zhipu"),
-    # Volcengine (火山引擎 — Doubao models)
-    ("doubao-seed-2.0-pro", "doubao-seed-2-0-pro-260215", "volcengine"),
-    ("doubao-seed-2.0-lite", "doubao-seed-2-0-lite-260215", "volcengine"),
-    ("doubao-seed-2.0-mini", "doubao-seed-2-0-mini-260215", "volcengine"),
-    ("doubao-seed-2.0-code", "doubao-seed-2-0-code-preview-260215", "volcengine"),
-    ("doubao-seed-1.6", "doubao-seed-1.6", "volcengine"),
-    ("doubao-1.5-pro", "doubao-1.5-pro-256k", "volcengine"),
-    ("doubao-1.5-thinking-pro", "doubao-1.5-thinking-pro", "volcengine"),
-    # DashScope Coding Plan (阿里云代码计划 — subscription sk-sp-* endpoint)
-    ("qwen3.7-max", "qwen3.7-max", "dashscope-code"),
-    ("qwen3.7-plus", "qwen3.7-plus", "dashscope-code"),
-    ("qwen3.6-max", "qwen3.6-max-preview", "dashscope-code"),
-    ("qwen3.6-plus", "qwen3.6-plus", "dashscope-code"),
-    ("qwen3.6-flash", "qwen3.6-flash", "dashscope-code"),
-    ("qwen3-coder", "qwen3-coder-plus", "dashscope-code"),
-    ("qwen3-coder-next", "qwen3-coder-next", "dashscope-code"),
-    ("qwen3-max", "qwen3-max", "dashscope-code"),
-    ("qwen3.5-plus", "qwen3.5-plus", "dashscope-code"),
-    # DashScope (阿里云 — Qwen models, default for simple lookups)
-    ("qwen3.7-max", "qwen3.7-max", "dashscope"),
-    ("qwen3.7-plus", "qwen3.7-plus", "dashscope"),
-    ("qwen3.6-max", "qwen3.6-max-preview", "dashscope"),
-    ("qwen3.6-plus", "qwen3.6-plus", "dashscope"),
-    ("qwen3.6-flash", "qwen3.6-flash", "dashscope"),
-    ("qwen3-coder", "qwen3-coder-plus", "dashscope"),
-    ("qwen3-235b", "qwen3-235b-a22b", "dashscope"),
-    ("qwen-max", "qwen-max", "dashscope"),
-    ("qwq-plus", "qwq-plus", "dashscope"),
-    # DeepSeek
-    ("deepseek-v4-pro", "deepseek-v4-pro", "deepseek"),
-    ("deepseek-v4-flash", "deepseek-v4-flash", "deepseek"),
-    # Legacy aliases (deprecated 2026-07-24; route to v4-flash thinking/non-thinking)
-    ("deepseek-r1", "deepseek-reasoner", "deepseek"),
-    ("deepseek-v3", "deepseek-chat", "deepseek"),
-    # Moonshot (OpenAI-compatible)
-    ("kimi-k2.6", "kimi-k2.6", "moonshot"),
-    ("kimi-k2.5", "kimi-k2.5", "moonshot"),
-    ("kimi-k2-thinking", "kimi-k2-thinking", "moonshot"),
-    ("kimi-k2-thinking-turbo", "kimi-k2-thinking-turbo", "moonshot"),
-    ("moonshot-v1-auto", "moonshot-v1-auto", "moonshot"),
-    ("moonshot-v1-128k", "moonshot-v1-128k", "moonshot"),
-    ("moonshot-v1-32k", "moonshot-v1-32k", "moonshot"),
-    ("moonshot-v1-8k", "moonshot-v1-8k", "moonshot"),
-    # Kimi Coding Plan (Anthropic-compatible)
-    ("kimi-for-coding", "kimi-for-coding", "kimi-coding"),
-]
-
-# Public dict for simple lookups (last entry wins for duplicate names).
-# Use get_models_for_provider() for provider-aware lookups.
-MODELS: dict[str, tuple[str, str]] = {
-    name: (model_id, provider) for name, model_id, provider in _MODEL_ENTRIES
-}
-
-DEFAULT_MODEL = "claude-sonnet-4-6"
-
-
-def get_models_for_provider(provider: str) -> list[tuple[str, str]]:
-    """Get all models for a specific provider.
-
-    Args:
-        provider: Provider name (e.g., 'anthropic', 'openrouter').
-
-    Returns:
-        List of (short_name, model_id) tuples for the provider.
-    """
-    return [(name, model_id) for name, model_id, p in _MODEL_ENTRIES if p == provider]
-
 
 def _env_flag_enabled(name: str) -> bool:
     return os.environ.get(name, "").strip().lower() in _TRUTHY_ENV_VALUES
@@ -335,26 +202,34 @@ def _drop_unsupported_chat_model_kwargs(kwargs: dict[str, Any]) -> None:
 
 _IMPLICIT_CACHE_PROVIDERS = frozenset({"zhipu", "zhipu-code", "siliconflow", "nvidia"})
 
+# OpenAI-compatible routers that forward an Anthropic-style ``cache_control``
+# declaration through to Claude models addressed as ``anthropic/...``.
+_EXPLICIT_CACHE_PROVIDERS = frozenset({"openrouter", "requesty"})
+
 
 def _cache_strategy(provider: str, model_id: str) -> str:
     """Return the provider's prompt-cache mechanism.
 
-    ``explicit`` — needs Anthropic-style ``cache_control`` markers (OpenRouter
-    Claude routes only, matching the pre-existing condition).
+    ``explicit`` — needs Anthropic-style ``cache_control`` markers (Claude routes
+    on the OpenAI-compatible routers).
     ``implicit`` — provider prefixes-cache automatically; no markers, but the
     prompt prefix must stay byte-stable for hits (see memory injection order).
     ``none`` — no cache model to declare.
     """
-    if provider == "openrouter" and model_id.startswith(("anthropic/", "~anthropic/")):
+    if provider in _EXPLICIT_CACHE_PROVIDERS and model_id.startswith(
+        ("anthropic/", "~anthropic/")
+    ):
         return "explicit"
     if provider in _IMPLICIT_CACHE_PROVIDERS:
         return "implicit"
     return "none"
 
 
-def _supports_openrouter_anthropic_prompt_cache(provider: str, model_id: str) -> bool:
-    """Return whether EvoScientist should declare OpenRouter Claude caching."""
-    return _cache_strategy(provider, model_id) == "explicit"
+def _supports_openrouter_anthropic_prompt_cache(
+    provider: str | None, model_id: str
+) -> bool:
+    """Return whether EvoScientist should declare Claude caching for a router."""
+    return _cache_strategy(provider or "", model_id) == "explicit"
 
 
 def _has_cache_control_override(kwargs: dict[str, Any]) -> bool:
@@ -377,16 +252,25 @@ def _has_cache_control_override(kwargs: dict[str, Any]) -> bool:
 
 
 def _apply_openrouter_anthropic_prompt_cache(
-    provider: str,
+    provider: str | None,
     model_id: str,
     kwargs: dict[str, Any],
 ) -> None:
-    """Declare OpenRouter Claude prompt caching unless explicitly disabled.
+    """Declare router Claude prompt caching unless explicitly disabled.
 
-    OpenRouter already handles implicit caching for most providers, but Claude
-    prompt caching needs Anthropic-style cache-control declaration.
+    OpenRouter and Requesty both handle implicit caching for most providers,
+    but Claude prompt caching needs an Anthropic-style cache-control
+    declaration. Each router honours its own opt-out env flag
+    (``EVOSCIENTIST_OPENROUTER_ANTHROPIC_PROMPT_CACHE`` /
+    ``EVOSCIENTIST_REQUESTY_ANTHROPIC_PROMPT_CACHE``).
     """
-    if _env_flag_disabled("EVOSCIENTIST_OPENROUTER_ANTHROPIC_PROMPT_CACHE"):
+    if provider is None:
+        return
+    disable_flag = {
+        "openrouter": "EVOSCIENTIST_OPENROUTER_ANTHROPIC_PROMPT_CACHE",
+        "requesty": "EVOSCIENTIST_REQUESTY_ANTHROPIC_PROMPT_CACHE",
+    }.get(provider)
+    if disable_flag is not None and _env_flag_disabled(disable_flag):
         return
     if not _supports_openrouter_anthropic_prompt_cache(provider, model_id):
         return
@@ -395,6 +279,22 @@ def _apply_openrouter_anthropic_prompt_cache(
     kwargs.setdefault("model_kwargs", {})["cache_control"] = {"type": "ephemeral"}
 
 
+def _enable_openrouter_429_retry(chat_model: Any) -> None:
+    """Add 429 to the OpenRouter SDK's retryable status codes (default ["5XX"]).
+
+    Upstream rate limits ("temporarily rate-limited upstream", whose
+    Retry-After the SDK backoff already honors) otherwise fail the run outright.
+    """
+    sdk_config = getattr(getattr(chat_model, "client", None), "sdk_configuration", None)
+    retry_config: Any = getattr(sdk_config, "retry_config", None)
+    # Skip the UNSET sentinel (max_retries=0) and explicit caller overrides.
+    if not hasattr(retry_config, "status_codes_override"):
+        return
+    if retry_config.status_codes_override:
+        return
+    retry_config.status_codes_override = ["429", "5XX"]
+
+
 def _apply_auto_config(
     provider: str,
     model_id: str,
@@ -427,8 +327,15 @@ def _apply_auto_config(
         else:
             _is_proxy = False
         if _is_proxy or (is_third_party and not _supports_thinking):
-            pass
-        elif "fable" in model_id or model_id.endswith(("4-6", "4-7", "4-8")):
+            # Mandatory-thinking Kimi models (K3 / Kimi For Coding) must declare
+            # thinking so with_structured_output avoids forced tool_choice (400).
+            # max_tokens must exceed budget_tokens (default resolves to 4096).
+            if is_third_party and _is_mandatory_thinking_kimi(model_id):
+                kwargs["thinking"] = {"type": "enabled", "budget_tokens": 10000}
+                kwargs.setdefault("max_tokens", 16000)
+        elif "fable" in model_id or model_id.endswith(
+            ("opus-5", "sonnet-5", "4-6", "4-7", "4-8")
+        ):
             kwargs["thinking"] = {"type": "adaptive", "display": "summarized"}
             kwargs.setdefault("effort", "max")
         else:
@@ -567,6 +474,11 @@ def get_chat_model(
         if api_key:
             kwargs.setdefault("api_key", api_key)
 
+    elif provider == "deepseek":
+        api_key = os.environ.get("DEEPSEEK_API_KEY", "")
+        if api_key:
+            kwargs["api_key"] = api_key
+
     # OpenAI-routed providers → route through OpenAI provider with base_url
     elif provider in _OPENAI_ROUTED_PROVIDERS:
         _original_provider = provider
@@ -587,14 +499,17 @@ def get_chat_model(
         api_key = os.environ.get(api_key_env, "")
         if api_key:
             kwargs.setdefault("api_key", api_key)
+        _apply_openai_compat_reasoning_config(provider, model_id, kwargs)
         # SiliconFlow: disable thinking — LangChain drops reasoning_content
         # from history, causing error 20015 on multi-turn requests.
         if provider == "siliconflow":
             kwargs.setdefault("extra_body", {})["enable_thinking"] = False
-        # Moonshot: disable thinking for all models to prevent LangChain from dropping
-        # reasoning_content, which causes multi-turn conversation errors (error 20015).
-        # Even native thinking models like kimi-k2-thinking operate in non-thinking mode.
-        if provider == "moonshot":
+        # Moonshot: disable thinking for pre-K3 models to prevent LangChain from
+        # dropping reasoning_content, which causes multi-turn conversation errors
+        # (error 20015). Even native thinking models like kimi-k2-thinking operate
+        # in non-thinking mode. kimi-k3+ is exempt: always-thinking, and
+        # Moonshot's K3 guide forbids the K2.x `thinking` parameter for it.
+        if provider == "moonshot" and not model_id.startswith("kimi-k3"):
             kwargs.setdefault("extra_body", {})["thinking"] = {"type": "disabled"}
         provider = "openai"
 
@@ -610,7 +525,12 @@ def get_chat_model(
         # passback (OpenRouter's `/responses` beta is stateless, store=false —
         # "Item with id 'rs_...' not found"); the patch strips them on passback,
         # so enabling `summary` is safe. See langchain-ai/langchain#37777.
-        kwargs.setdefault("reasoning", {"effort": "medium", "summary": "auto"})
+        # Note: mandatory-reasoning endpoints (kimi-k3, grok-4.5, …) reject
+        # effort "none" with HTTP 400 — that error is surfaced to the user
+        # as-is; pick a real effort (low+) for those models.
+        # Ai4Sci: default medium, overridable via EVOSCIENTIST_REASONING_EFFORT.
+        effort = _resolve_reasoning_effort("medium")
+        kwargs.setdefault("reasoning", {"effort": effort, "summary": "auto"})
         # App attribution (issue #339): identify EvoScientist to OpenRouter so
         # usage is credited to the project (app rankings, model app tabs,
         # analytics) rather than langchain-openrouter's LangChain-branded
@@ -628,6 +548,11 @@ def get_chat_model(
             os.environ.get("EVOSCIENTIST_OPENROUTER_APP_TITLE", "").strip()
             or OPENROUTER_DEFAULT_APP_TITLE,
         )
+        # OpenRouter keys app pages by HTTP-Referer and X-Title only renames that
+        # page, so a custom title on the default referer would rename the shared
+        # EvoScientist page for everyone. Honor it only with a custom referer.
+        if kwargs["app_url"] == OPENROUTER_DEFAULT_HTTP_REFERER:
+            kwargs["app_title"] = OPENROUTER_DEFAULT_APP_TITLE
         # app_categories must be a list[str] (langchain-openrouter joins it into
         # the X-OpenRouter-Categories header); split the comma-separated config
         # value and drop blanks so a stray comma/space can't emit an empty one.
@@ -658,6 +583,7 @@ def get_chat_model(
         if _app_categories:
             kwargs.setdefault("app_categories", _app_categories)
         _patch_openrouter_strip_responses_reasoning()
+        _patch_openrouter_structured_output()
 
     # Anthropic-routed providers → route through Anthropic provider with base_url
     elif provider in _ANTHROPIC_ROUTED_PROVIDERS:
@@ -693,25 +619,72 @@ def get_chat_model(
 
     _drop_unsupported_chat_model_kwargs(kwargs)
     _apply_auto_config(provider, model_id, _is_third_party, kwargs, _original_provider)
-    _apply_openrouter_anthropic_prompt_cache(provider, model_id, kwargs)
+    # OpenAI-routed routers (e.g. Requesty) reassign ``provider`` to "openai"
+    # above, so use the original provider name to detect router-level caching.
+    _cache_provider = _original_provider or provider
+    _apply_openrouter_anthropic_prompt_cache(_cache_provider, model_id, kwargs)
 
+    _uses_native_deepseek = provider == "deepseek" or (
+        provider == "openai"
+        and _original_provider == "custom-openai"
+        and _is_deepseek_endpoint(kwargs.get("base_url"))
+    )
+
+    # User-level override for the OpenAI Responses API vs Chat Completions.
+    # When "false", force Chat Completions and drop reasoning (which triggers
+    # the Responses API path in langchain-openai). Only applies to OpenAI.
+    if _uses_native_deepseek:
+        if kwargs.get("use_responses_api") is True:
+            raise ValueError(
+                "DeepSeek does not support the OpenAI Responses API. "
+                "Remove use_responses_api=True."
+            )
+        kwargs.pop("use_responses_api", None)
+    elif provider == "openai":
+        _responses_api_setting = (
+            os.environ.get("EVOSCIENTIST_USE_RESPONSES_API", "").strip().lower()
+        )
+        if _responses_api_setting == "false":
+            kwargs["use_responses_api"] = False
+            kwargs.pop("reasoning", None)
+        elif _responses_api_setting == "true":
+            kwargs["use_responses_api"] = True
+
+    if _is_openai_proxy and kwargs.get("use_responses_api") is True:
+        reasoning = kwargs.setdefault("reasoning", {})
+        if isinstance(reasoning, dict):
+            reasoning = dict(reasoning)
+            reasoning.setdefault("context", "all_turns")
+            kwargs["reasoning"] = reasoning
+
+    # Ai4Sci: an ambient ANTHROPIC_AUTH_TOKEN would silently override the
+    # explicit api_key resolved for this provider, so hide it for this call.
     anthropic_auth_token = None
     if provider == "anthropic" and kwargs.get("api_key"):
         anthropic_auth_token = os.environ.pop("ANTHROPIC_AUTH_TOKEN", None)
     try:
-        chat_model = init_chat_model(model=model_id, model_provider=provider, **kwargs)
+        if _uses_native_deepseek:
+            chat_model = EvoChatDeepSeek(model=model_id, **kwargs)
+        else:
+            chat_model = init_chat_model(model=model_id, model_provider=provider, **kwargs)
     finally:
         if anthropic_auth_token is not None:
             os.environ["ANTHROPIC_AUTH_TOKEN"] = anthropic_auth_token
 
     # Flatten list content to strings for strict OpenAI-compatible providers
-    # (DeepSeek, SiliconFlow, OpenRouter, custom-openai, etc.) and
+    # (SiliconFlow, OpenRouter, custom-openai, etc.) and
     # native OpenAI through a proxy, to avoid "sequence expected string" errors.
     # Moonshot and Kimi Coding support standard format, no patch needed.
+    # Mandatory-thinking Kimi models on Anthropic-routed endpoints are exempt:
+    # flatten drops thinking blocks, which Kimi requires on tool-call turns.
     _no_patch_providers = {"moonshot", "kimi-coding"}
     if (
-        _is_third_party or _is_openai_proxy
-    ) and _original_provider not in _no_patch_providers:
+        (_is_third_party or _is_openai_proxy)
+        and _original_provider not in _no_patch_providers
+        and not _uses_native_deepseek
+        and not (provider == "anthropic" and _is_mandatory_thinking_kimi(model_id))
+        and _original_provider not in _ANTHROPIC_ROUTED_PROVIDERS
+    ):
         # Anthropic-routed providers accept media in tool results natively;
         # only OpenAI-compatible providers need tool-media hoisting.
         _hoist = _original_provider not in _ANTHROPIC_ROUTED_PROVIDERS
@@ -729,77 +702,17 @@ def get_chat_model(
             ),
         )
 
-    # DeepSeek thinking mode requires reasoning_content passback in multi-turn
-    # + tool_use scenarios.
-    if _original_provider == "deepseek":
-        _patch_deepseek_reasoning_passback(chat_model)
 
     if _is_openai_proxy:
         _patch_ccproxy_system_to_developer(chat_model)
 
+    if provider == "openrouter":
+        _enable_openrouter_429_retry(chat_model)
+
+    if provider == "anthropic":
+        _patch_anthropic_strip_foreign_reasoning()
+        _patch_anthropic_structured_output()
+
     apply_known_context_window(chat_model)
 
     return chat_model
-
-
-def list_models() -> list[str]:
-    """List all available model short names.
-
-    Returns:
-        List of unique model short names that can be passed to get_chat_model().
-    """
-    seen = set()
-    result = []
-    for name, _, _ in _MODEL_ENTRIES:
-        if name not in seen:
-            seen.add(name)
-            result.append(name)
-    return result
-
-
-def list_models_by_provider() -> list[tuple[str, str, str]]:
-    """List all unique (short_name, model_id, provider) entries.
-
-    Returns:
-        De-duplicated list of model entries preserving registry order.
-    """
-    seen: set[tuple[str, str]] = set()
-    result: list[tuple[str, str, str]] = []
-    for name, model_id, provider in _MODEL_ENTRIES:
-        key = (name, provider)
-        if key not in seen:
-            seen.add(key)
-            result.append((name, model_id, provider))
-    return result
-
-
-async def list_model_picker_entries(
-    ollama_base_url: str | None,
-    *,
-    include_custom_ollama: bool,
-) -> list[tuple[str, str, str]]:
-    """Return model picker entries, optionally including local Ollama models."""
-    entries = list_models_by_provider()
-    if ollama_base_url:
-        from .ollama_discovery import discover_ollama_models
-
-        for detected_name in await discover_ollama_models(
-            ollama_base_url,
-            timeout=1.5,
-        ):
-            entries.append((detected_name, detected_name, "ollama"))
-        if include_custom_ollama:
-            entries.append(("Custom Ollama model...", "__custom_ollama__", "ollama"))
-    return entries
-
-
-def get_model_info(model: str) -> tuple[str, str] | None:
-    """Get the (model_id, provider) tuple for a short name.
-
-    Args:
-        model: Short model name.
-
-    Returns:
-        Tuple of (model_id, provider) or None if not found.
-    """
-    return MODELS.get(model)

EvoScientist/llm/openai_compat.py

@@ -0,0 +1,94 @@
+"""Reusable behavior for OpenAI-compatible chat model integrations."""
+
+from __future__ import annotations
+
+from collections.abc import AsyncIterator, Iterator, Mapping
+from typing import Any
+
+from langchain_core.callbacks import (
+    AsyncCallbackManagerForLLMRun,
+    CallbackManagerForLLMRun,
+)
+from langchain_core.messages import BaseMessage
+from langchain_core.outputs import ChatGenerationChunk, ChatResult
+
+from .patches import _OpenAICompatContent
+
+
+class OpenAICompatContentMixin:
+    """Normalize message content before calling an OpenAI-compatible model."""
+
+    def _content_compat(self) -> _OpenAICompatContent:
+        compat = self.__dict__.get("_evosci_content_compat")
+        if not isinstance(compat, _OpenAICompatContent):
+            profile = getattr(self, "profile", None)
+            compat = _OpenAICompatContent(
+                profile if isinstance(profile, Mapping) else None,
+                hoist_tool_media=True,
+            )
+            self.__dict__["_evosci_content_compat"] = compat
+        return compat
+
+    def _generate(
+        self,
+        messages: list[BaseMessage],
+        stop: list[str] | None = None,
+        run_manager: CallbackManagerForLLMRun | None = None,
+        **kwargs: Any,
+    ) -> ChatResult:
+        return self._content_compat().invoke(
+            super()._generate,  # type: ignore[attr-defined]
+            messages,
+            stop=stop,
+            run_manager=run_manager,
+            **kwargs,
+        )
+
+    async def _agenerate(
+        self,
+        messages: list[BaseMessage],
+        stop: list[str] | None = None,
+        run_manager: AsyncCallbackManagerForLLMRun | None = None,
+        **kwargs: Any,
+    ) -> ChatResult:
+        return await self._content_compat().ainvoke(
+            super()._agenerate,  # type: ignore[attr-defined]
+            messages,
+            stop=stop,
+            run_manager=run_manager,
+            **kwargs,
+        )
+
+    def _stream(
+        self,
+        messages: list[BaseMessage],
+        stop: list[str] | None = None,
+        run_manager: CallbackManagerForLLMRun | None = None,
+        **kwargs: Any,
+    ) -> Iterator[ChatGenerationChunk]:
+        yield from self._content_compat().stream(
+            super()._stream,  # type: ignore[attr-defined]
+            messages,
+            stop=stop,
+            run_manager=run_manager,
+            **kwargs,
+        )
+
+    async def _astream(
+        self,
+        messages: list[BaseMessage],
+        stop: list[str] | None = None,
+        run_manager: AsyncCallbackManagerForLLMRun | None = None,
+        *,
+        stream_usage: bool | None = None,
+        **kwargs: Any,
+    ) -> AsyncIterator[ChatGenerationChunk]:
+        async for chunk in self._content_compat().astream(
+            super()._astream,  # type: ignore[attr-defined]
+            messages,
+            stop=stop,
+            run_manager=run_manager,
+            stream_usage=stream_usage,
+            **kwargs,
+        ):
+            yield chunk

EvoScientist/llm/registry.py

@@ -0,0 +1,370 @@
+"""Model registry data — short names, model ids, providers, routing tables.
+
+Pure data with no langchain/provider-SDK imports: the onboard wizard, the
+``/model`` pickers, and provider validation read this registry without paying
+for the chat-model construction stack in :mod:`.models` (~2000 modules).
+"""
+
+from __future__ import annotations
+
+_MINIMAX_ANTHROPIC_BASE_URL = "https://api.minimaxi.com/anthropic"
+_SILICONFLOW_BASE_URL = "https://api.siliconflow.cn/v1"
+
+_ZHIPU_BASE_URL = "https://open.bigmodel.cn/api/paas/v4"
+_ZHIPU_CODE_BASE_URL = "https://open.bigmodel.cn/api/coding/paas/v4"
+_VOLCENGINE_BASE_URL = "https://ark.cn-beijing.volces.com/api/v3"
+_VOLCENGINE_CODE_BASE_URL = "https://ark.cn-beijing.volces.com/api/coding/v3"
+_DASHSCOPE_BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1"
+_DASHSCOPE_CODE_BASE_URL = "https://coding.dashscope.aliyuncs.com/v1"
+
+_ATLASCLOUD_BASE_URL = "https://api.atlascloud.ai/v1"
+_MOONSHOT_BASE_URL = "https://api.moonshot.cn/v1"
+_KIMI_CODING_BASE_URL = "https://api.kimi.com/coding/"
+_REQUESTY_BASE_URL = "https://router.requesty.ai/v1"
+_NOVITA_BASE_URL = "https://api.novita.ai/openai/v1"
+
+# Providers routed through the OpenAI provider with a custom base_url.
+# Maps provider name → (base_url or None, env var for API key).
+_OPENAI_ROUTED_PROVIDERS: dict[str, tuple[str | None, str]] = {
+    "atlascloud": (_ATLASCLOUD_BASE_URL, "ATLASCLOUD_API_KEY"),
+    "moonshot": (_MOONSHOT_BASE_URL, "MOONSHOT_API_KEY"),
+    "siliconflow": (_SILICONFLOW_BASE_URL, "SILICONFLOW_API_KEY"),
+    "zhipu": (_ZHIPU_BASE_URL, "ZHIPU_API_KEY"),
+    "zhipu-code": (_ZHIPU_CODE_BASE_URL, "ZHIPU_API_KEY"),
+    "volcengine": (_VOLCENGINE_BASE_URL, "VOLCENGINE_API_KEY"),
+    "volcengine-code": (_VOLCENGINE_CODE_BASE_URL, "VOLCENGINE_API_KEY"),
+    "dashscope": (_DASHSCOPE_BASE_URL, "DASHSCOPE_API_KEY"),
+    "dashscope-code": (_DASHSCOPE_CODE_BASE_URL, "DASHSCOPE_API_KEY"),
+    "requesty": (_REQUESTY_BASE_URL, "REQUESTY_API_KEY"),
+    "novita": (_NOVITA_BASE_URL, "NOVITA_API_KEY"),
+    "custom-openai": (
+        None,
+        "CUSTOM_OPENAI_API_KEY",
+    ),  # base_url from CUSTOM_OPENAI_BASE_URL env
+}
+
+# Providers routed through the Anthropic provider with a custom base_url.
+# Maps provider name → (base_url or None, env var for API key).
+_ANTHROPIC_ROUTED_PROVIDERS: dict[str, tuple[str | None, str]] = {
+    "minimax": (_MINIMAX_ANTHROPIC_BASE_URL, "MINIMAX_API_KEY"),
+    "kimi-coding": (_KIMI_CODING_BASE_URL, "KIMI_API_KEY"),
+    "custom-anthropic": (None, "CUSTOM_ANTHROPIC_API_KEY"),
+}
+
+# Anthropic-routed providers that support extended thinking.
+_THINKING_CAPABLE_PROVIDERS: set[str] = {"minimax"}
+
+# Moonshot rejects a forced tool choice while thinking is enabled, and kimi-k3
+# cannot disable thinking — structured output must use json_schema there.
+# Moonshot-specific: do NOT widen to other mandatory-reasoning models.
+_OPENROUTER_JSON_SCHEMA_STRUCTURED_OUTPUT_MODELS = frozenset(
+    {"moonshotai/kimi-k3", "moonshotai/kimi-k3-20260715"}
+)
+
+
+def _is_mandatory_thinking_kimi(model_id: str) -> bool:
+    """True for Kimi models whose thinking cannot be disabled (K3 family)."""
+    short_id = model_id.split("/")[-1]
+    return short_id.startswith("kimi-k3") or short_id == "kimi-for-coding"
+
+
+# Model registry: list of (short_name, model_id, provider)
+# Allows same short_name across different providers.
+_MODEL_ENTRIES: list[tuple[str, str, str]] = [
+    # Custom Anthropic (third-party Claude-compatible endpoints, current-gen defaults)
+    # Listed BEFORE native anthropic so MODELS dict defaults to native provider
+    ("claude-sonnet-4-6", "claude-sonnet-4-6", "custom-anthropic"),
+    ("claude-haiku-4-5", "claude-haiku-4-5", "custom-anthropic"),
+    # Custom OpenAI (third-party OpenAI-compatible endpoints, 3 defaults)
+    # Listed BEFORE native openai so MODELS dict defaults to native provider
+    ("gpt-5.5-pro", "gpt-5.5-pro", "custom-openai"),
+    ("gpt-5.5", "gpt-5.5", "custom-openai"),
+    ("gpt-5.4", "gpt-5.4", "custom-openai"),
+    ("gpt-5.3-codex", "gpt-5.3-codex", "custom-openai"),
+    ("gpt-5-mini", "gpt-5-mini", "custom-openai"),
+    # Atlas Cloud (OpenAI-compatible)
+    ("qwen3.5-27b", "qwen/qwen3.5-27b", "atlascloud"),
+    # Anthropic (current generation)
+    ("claude-fable-5-1", "claude-fable-5-1", "anthropic"),
+    ("claude-fable-5", "claude-fable-5", "anthropic"),
+    ("claude-opus-5", "claude-opus-5", "anthropic"),
+    ("claude-opus-4-8", "claude-opus-4-8", "anthropic"),
+    ("claude-sonnet-5", "claude-sonnet-5", "anthropic"),
+    ("claude-sonnet-4-6", "claude-sonnet-4-6", "anthropic"),
+    ("claude-haiku-4-5", "claude-haiku-4-5", "anthropic"),
+    # OpenAI
+    ("gpt-6-astra", "gpt-6-astra", "openai"),
+    ("gpt-5.6-sol", "gpt-5.6-sol", "openai"),
+    ("gpt-5.6-terra", "gpt-5.6-terra", "openai"),
+    ("gpt-5.6-luna", "gpt-5.6-luna", "openai"),
+    ("gpt-5.5-pro", "gpt-5.5-pro", "openai"),
+    ("gpt-5.5", "gpt-5.5", "openai"),
+    ("gpt-5.4", "gpt-5.4", "openai"),
+    ("gpt-5.4-mini", "gpt-5.4-mini", "openai"),
+    ("gpt-5.4-nano", "gpt-5.4-nano", "openai"),
+    ("gpt-5.3-codex", "gpt-5.3-codex", "openai"),
+    ("gpt-5.2-codex", "gpt-5.2-codex", "openai"),
+    ("gpt-5.2", "gpt-5.2", "openai"),
+    ("gpt-5.1", "gpt-5.1", "openai"),
+    ("gpt-5", "gpt-5", "openai"),
+    ("gpt-5-mini", "gpt-5-mini", "openai"),
+    ("gpt-5-nano", "gpt-5-nano", "openai"),
+    # Google GenAI
+    ("gemini-3.8-flash", "gemini-3.8-flash", "google-genai"),
+    ("gemini-3.7-flash", "gemini-3.7-flash", "google-genai"),
+    ("gemini-3.6-flash", "gemini-3.6-flash", "google-genai"),
+    ("gemini-3.5-flash", "gemini-3.5-flash", "google-genai"),
+    ("gemini-3.5-flash-lite", "gemini-3.5-flash-lite", "google-genai"),
+    ("gemini-3.1-pro", "gemini-3.1-pro-preview", "google-genai"),
+    (
+        "gemini-3.1-pro-customtools",
+        "gemini-3.1-pro-preview-customtools",
+        "google-genai",
+    ),
+    ("gemini-3.1-flash-lite", "gemini-3.1-flash-lite-preview", "google-genai"),
+    ("gemini-3-flash", "gemini-3-flash-preview", "google-genai"),
+    ("gemini-2.5-flash", "gemini-2.5-flash", "google-genai"),
+    ("gemini-2.5-flash-lite", "gemini-2.5-flash-lite", "google-genai"),
+    ("gemini-2.5-pro", "gemini-2.5-pro", "google-genai"),
+    # MiniMax (direct API — Anthropic-compatible; default: api.minimaxi.com, global: api.minimax.io)
+    ("minimax-m3", "MiniMax-M3", "minimax"),
+    ("minimax-m2.7", "MiniMax-M2.7", "minimax"),
+    ("minimax-m2.7-highspeed", "MiniMax-M2.7-highspeed", "minimax"),
+    ("minimax-m2.5", "MiniMax-M2.5", "minimax"),
+    ("minimax-m2.5-highspeed", "MiniMax-M2.5-highspeed", "minimax"),
+    # NVIDIA
+    ("nemotron-super", "nvidia/nemotron-3-super-120b-a12b", "nvidia"),
+    ("nemotron-nano", "nvidia/nemotron-3-nano-30b-a3b", "nvidia"),
+    ("glm-5.2", "z-ai/glm-5.2", "nvidia"),
+    ("glm4.7", "z-ai/glm4.7", "nvidia"),
+    ("deepseek-v3.2", "deepseek-ai/deepseek-v3.2", "nvidia"),
+    ("deepseek-v3.1", "deepseek-ai/deepseek-v3.1-terminus", "nvidia"),
+    ("kimi-k2.5", "moonshotai/kimi-k2.5", "nvidia"),
+    ("kimi-k2-thinking", "moonshotai/kimi-k2-thinking", "nvidia"),
+    ("minimax-m2.5", "minimaxai/minimax-m2.5", "nvidia"),
+    ("minimax-m2.1", "minimaxai/minimax-m2.1", "nvidia"),
+    ("qwen3.5-397b", "qwen/qwen3.5-397b-a17b", "nvidia"),
+    ("step-3.5-flash", "stepfun-ai/step-3.5-flash", "nvidia"),
+    # SiliconFlow
+    ("minimax-m2.5", "Pro/MiniMaxAI/MiniMax-M2.5", "siliconflow"),
+    ("glm-5.2", "Pro/zai-org/GLM-5.2", "siliconflow"),
+    ("glm-5", "Pro/zai-org/GLM-5", "siliconflow"),
+    ("kimi-k2.5", "Pro/moonshotai/Kimi-K2.5", "siliconflow"),
+    ("glm-4.7", "Pro/zai-org/GLM-4.7", "siliconflow"),
+    # Requesty (aggregator — OpenAI-compatible router, provider/model IDs).
+    # Listed before OpenRouter so that for model names shared with OpenRouter
+    # or a native provider, Requesty does not override them (the dict below is
+    # last-entry-wins); Requesty is selected explicitly via get_models_for_provider.
+    ("claude-sonnet-4.6", "anthropic/claude-sonnet-4-6", "requesty"),
+    ("claude-opus-4.8", "anthropic/claude-opus-4-8", "requesty"),
+    ("gemini-3.5-flash", "google/gemini-3.5-flash", "requesty"),
+    ("grok-4.3", "xai/grok-4.3", "requesty"),
+    ("grok-build-0.1", "xai/grok-build-0.1", "requesty"),
+    # Novita (aggregator — OpenAI-compatible, Kimi/GLM/DeepSeek, etc.)
+    ("kimi-k3", "moonshotai/kimi-k3", "novita"),
+    ("glm-5.2", "zai-org/glm-5.2", "novita"),
+    ("deepseek-v4-flash", "deepseek/deepseek-v4-flash-0731", "novita"),
+    # OpenRouter
+    ("claude-fable-5.1", "anthropic/claude-fable-5.1", "openrouter"),
+    ("claude-fable-5", "anthropic/claude-fable-5", "openrouter"),
+    ("claude-opus-5", "anthropic/claude-opus-5", "openrouter"),
+    ("claude-opus-5-fast", "anthropic/claude-opus-5-fast", "openrouter"),
+    ("claude-opus-4.8", "anthropic/claude-opus-4.8", "openrouter"),
+    ("claude-opus-4.8-fast", "anthropic/claude-opus-4.8-fast", "openrouter"),
+    ("claude-sonnet-5", "anthropic/claude-sonnet-5", "openrouter"),
+    ("claude-sonnet-4.6", "anthropic/claude-sonnet-4.6", "openrouter"),
+    ("gpt-6-astra", "openai/gpt-6-astra", "openrouter"),
+    ("gpt-5.6-sol", "openai/gpt-5.6-sol", "openrouter"),
+    ("gpt-5.6-terra", "openai/gpt-5.6-terra", "openrouter"),
+    ("gpt-5.6-luna", "openai/gpt-5.6-luna", "openrouter"),
+    ("gpt-5.5-pro", "openai/gpt-5.5-pro", "openrouter"),
+    ("gpt-5.5", "openai/gpt-5.5", "openrouter"),
+    ("gpt-5.4", "openai/gpt-5.4", "openrouter"),
+    ("gpt-5.3-codex", "openai/gpt-5.3-codex", "openrouter"),
+    ("gemini-3.8-flash", "google/gemini-3.8-flash", "openrouter"),
+    ("gemini-3.7-flash", "google/gemini-3.7-flash", "openrouter"),
+    ("gemini-3.6-flash", "google/gemini-3.6-flash", "openrouter"),
+    ("gemini-3.5-flash", "google/gemini-3.5-flash", "openrouter"),
+    ("gemini-3.5-flash-lite", "google/gemini-3.5-flash-lite", "openrouter"),
+    ("gemini-3.1-pro", "google/gemini-3.1-pro-preview", "openrouter"),
+    ("gemini-3-flash", "google/gemini-3-flash-preview", "openrouter"),
+    ("kimi-k3", "moonshotai/kimi-k3", "openrouter"),
+    ("kimi-k2.6", "moonshotai/kimi-k2.6", "openrouter"),
+    ("glm-5.3", "z-ai/glm-5.3", "openrouter"),
+    ("glm-5.3-flash", "z-ai/glm-5.3-flash", "openrouter"),
+    ("glm-5.2", "z-ai/glm-5.2", "openrouter"),
+    ("glm-5v-turbo", "z-ai/glm-5v-turbo", "openrouter"),
+    ("minimax-m3", "minimax/minimax-m3", "openrouter"),
+    ("mimo-v2.5-pro", "xiaomi/mimo-v2.5-pro", "openrouter"),
+    ("mimo-v2.5", "xiaomi/mimo-v2.5", "openrouter"),
+    ("grok-build-0.1", "x-ai/grok-build-0.1", "openrouter"),
+    ("grok-4.6", "x-ai/grok-4.6", "openrouter"),
+    ("grok-4.5", "x-ai/grok-4.5", "openrouter"),
+    ("hy4-preview", "tencent/hy4-preview", "openrouter"),
+    ("hy3", "tencent/hy3", "openrouter"),
+    ("qwen3.8-max", "qwen/qwen3.8-max", "openrouter"),
+    ("qwen3.8-flash", "qwen/qwen3.8-flash", "openrouter"),
+    ("qwen3.7-max", "qwen/qwen3.7-max", "openrouter"),
+    ("qwen3.7-plus", "qwen/qwen3.7-plus", "openrouter"),
+    ("qwen3.6-flash", "qwen/qwen3.6-flash", "openrouter"),
+    ("qwen3.5-122b", "qwen/qwen3.5-122b-a10b", "openrouter"),
+    ("deepseek-v4-pro-0813", "deepseek/deepseek-v4-pro-0813", "openrouter"),
+    ("deepseek-v4-pro", "deepseek/deepseek-v4-pro", "openrouter"),
+    ("deepseek-v4-flash", "deepseek/deepseek-v4-flash", "openrouter"),
+    ("muse-spark-1.3", "meta/muse-spark-1.3", "openrouter"),
+    # Volcengine Coding Plan (火山引擎代码计划 — coding-only endpoint)
+    # Listed before Zhipu so simple GLM lookups keep their existing default.
+    ("glm-5.2", "glm-5-2", "volcengine-code"),
+    ("kimi-k2.5", "kimi-k2-5", "volcengine-code"),
+    # Zhipu CodePlan (智谱代码计划 — coding-only endpoint)
+    ("glm-5.3", "glm-5.3", "zhipu-code"),
+    ("glm-5.3-flash", "glm-5.3-flash", "zhipu-code"),
+    ("glm-5.2", "glm-5.2", "zhipu-code"),
+    ("glm-5.1", "glm-5.1", "zhipu-code"),
+    ("glm-5", "glm-5", "zhipu-code"),
+    ("glm-5-turbo", "glm-5-turbo", "zhipu-code"),
+    ("glm-5v-turbo", "glm-5v-turbo", "zhipu-code"),
+    ("glm-4.7", "glm-4.7", "zhipu-code"),
+    # Zhipu (智谱 — general endpoint, default for simple lookups)
+    ("glm-5.3", "glm-5.3", "zhipu"),
+    ("glm-5.3-flash", "glm-5.3-flash", "zhipu"),
+    ("glm-5.2", "glm-5.2", "zhipu"),
+    ("glm-5.1", "glm-5.1", "zhipu"),
+    ("glm-5", "glm-5", "zhipu"),
+    ("glm-5-turbo", "glm-5-turbo", "zhipu"),
+    ("glm-5v-turbo", "glm-5v-turbo", "zhipu"),
+    ("glm-4.7", "glm-4.7", "zhipu"),
+    # Volcengine (火山引擎 — Doubao models)
+    ("doubao-seed-2.0-pro", "doubao-seed-2-0-pro-260215", "volcengine"),
+    ("doubao-seed-2.0-lite", "doubao-seed-2-0-lite-260215", "volcengine"),
+    ("doubao-seed-2.0-mini", "doubao-seed-2-0-mini-260215", "volcengine"),
+    ("doubao-seed-2.0-code", "doubao-seed-2-0-code-preview-260215", "volcengine"),
+    ("doubao-seed-1.6", "doubao-seed-1.6", "volcengine"),
+    ("doubao-1.5-pro", "doubao-1.5-pro-256k", "volcengine"),
+    ("doubao-1.5-thinking-pro", "doubao-1.5-thinking-pro", "volcengine"),
+    # DashScope Coding Plan (阿里云代码计划 — subscription sk-sp-* endpoint)
+    ("qwen3.8-max", "qwen3.8-max", "dashscope-code"),
+    ("qwen3.8-flash", "qwen3.8-flash", "dashscope-code"),
+    ("qwen3.7-max", "qwen3.7-max", "dashscope-code"),
+    ("qwen3.7-plus", "qwen3.7-plus", "dashscope-code"),
+    ("qwen3.6-max", "qwen3.6-max-preview", "dashscope-code"),
+    ("qwen3.6-plus", "qwen3.6-plus", "dashscope-code"),
+    ("qwen3.6-flash", "qwen3.6-flash", "dashscope-code"),
+    ("qwen3-coder", "qwen3-coder-plus", "dashscope-code"),
+    ("qwen3-coder-next", "qwen3-coder-next", "dashscope-code"),
+    ("qwen3-max", "qwen3-max", "dashscope-code"),
+    ("qwen3.5-plus", "qwen3.5-plus", "dashscope-code"),
+    # DashScope (阿里云 — Qwen models, default for simple lookups)
+    ("qwen3.8-max", "qwen3.8-max", "dashscope"),
+    ("qwen3.8-flash", "qwen3.8-flash", "dashscope"),
+    ("qwen3.7-max", "qwen3.7-max", "dashscope"),
+    ("qwen3.7-plus", "qwen3.7-plus", "dashscope"),
+    ("qwen3.6-max", "qwen3.6-max-preview", "dashscope"),
+    ("qwen3.6-plus", "qwen3.6-plus", "dashscope"),
+    ("qwen3.6-flash", "qwen3.6-flash", "dashscope"),
+    ("qwen3-coder", "qwen3-coder-plus", "dashscope"),
+    ("qwen3-235b", "qwen3-235b-a22b", "dashscope"),
+    ("qwen-max", "qwen-max", "dashscope"),
+    ("qwq-plus", "qwq-plus", "dashscope"),
+    # DeepSeek
+    ("deepseek-v4-pro", "deepseek-v4-pro", "deepseek"),
+    ("deepseek-v4-flash", "deepseek-v4-flash", "deepseek"),
+    # Legacy aliases (deprecated 2026-07-24; route to v4-flash thinking/non-thinking)
+    ("deepseek-r1", "deepseek-reasoner", "deepseek"),
+    ("deepseek-v3", "deepseek-chat", "deepseek"),
+    # Moonshot (OpenAI-compatible)
+    ("kimi-k3", "kimi-k3", "moonshot"),
+    ("kimi-k2.6", "kimi-k2.6", "moonshot"),
+    ("kimi-k2.5", "kimi-k2.5", "moonshot"),
+    ("kimi-k2-thinking", "kimi-k2-thinking", "moonshot"),
+    ("kimi-k2-thinking-turbo", "kimi-k2-thinking-turbo", "moonshot"),
+    ("moonshot-v1-auto", "moonshot-v1-auto", "moonshot"),
+    ("moonshot-v1-128k", "moonshot-v1-128k", "moonshot"),
+    ("moonshot-v1-32k", "moonshot-v1-32k", "moonshot"),
+    ("moonshot-v1-8k", "moonshot-v1-8k", "moonshot"),
+    # Kimi Coding Plan (Anthropic-compatible)
+    ("kimi-for-coding", "kimi-for-coding", "kimi-coding"),
+]
+
+# Public dict for simple lookups (last entry wins for duplicate names).
+# Use get_models_for_provider() for provider-aware lookups.
+MODELS: dict[str, tuple[str, str]] = {
+    name: (model_id, provider) for name, model_id, provider in _MODEL_ENTRIES
+}
+
+DEFAULT_MODEL = "claude-sonnet-4-6"
+
+
+def get_models_for_provider(provider: str) -> list[tuple[str, str]]:
+    """Get all models for a specific provider.
+
+    Args:
+        provider: Provider name (e.g., 'anthropic', 'openrouter').
+
+    Returns:
+        List of (short_name, model_id) tuples for the provider.
+    """
+    return [(name, model_id) for name, model_id, p in _MODEL_ENTRIES if p == provider]
+
+
+def list_models() -> list[str]:
+    """List all available model short names.
+
+    Returns:
+        List of unique model short names that can be passed to get_chat_model().
+    """
+    seen = set()
+    result = []
+    for name, _, _ in _MODEL_ENTRIES:
+        if name not in seen:
+            seen.add(name)
+            result.append(name)
+    return result
+
+
+def list_models_by_provider() -> list[tuple[str, str, str]]:
+    """List all unique (short_name, model_id, provider) entries.
+
+    Returns:
+        De-duplicated list of model entries preserving registry order.
+    """
+    seen: set[tuple[str, str]] = set()
+    result: list[tuple[str, str, str]] = []
+    for name, model_id, provider in _MODEL_ENTRIES:
+        key = (name, provider)
+        if key not in seen:
+            seen.add(key)
+            result.append((name, model_id, provider))
+    return result
+
+
+async def list_model_picker_entries(
+    ollama_base_url: str | None,
+    *,
+    include_custom_ollama: bool,
+) -> list[tuple[str, str, str]]:
+    """Return model picker entries, optionally including local Ollama models."""
+    entries = list_models_by_provider()
+    if ollama_base_url:
+        from .ollama_discovery import discover_ollama_models
+
+        for detected_name in await discover_ollama_models(
+            ollama_base_url,
+            timeout=1.5,
+        ):
+            entries.append((detected_name, detected_name, "ollama"))
+        if include_custom_ollama:
+            entries.append(("Custom Ollama model...", "__custom_ollama__", "ollama"))
+    return entries
+
+
+def get_model_info(model: str) -> tuple[str, str] | None:
+    """Get the (model_id, provider) tuple for a short name.
+
+    Args:
+        model: Short model name.
+
+    Returns:
+        Tuple of (model_id, provider) or None if not found.
+    """
+    return MODELS.get(model)

EvoScientist/mcp/README.md

@@ -319,6 +319,15 @@ stdio server fails to start — install Node.js and `npx`, or replace `npx` with
 
 </details>
 
+<details>
+<summary><strong>Windows: stdio server fails with <code>[Errno 9] Bad file descriptor</code></strong></summary>
+
+On Windows, the TUI redirects `sys.stderr` to an in-memory capture whose `fileno()` is not a real OS handle. The MCP SDK forwards that `stderr` to the stdio server subprocess, and `subprocess.Popen` rejects the invalid handle with `OSError: [Errno 9] Bad file descriptor` — so only stdio servers fail to load (HTTP/SSE servers are unaffected).
+
+EvoScientist wraps the SDK's stdio client so that, whenever the configured `stderr` has no usable file descriptor, it falls back to the original console handle (`sys.__stderr__`, or `os.devnull` in GUI hosts). If you still see this error, run from a real console (not `pythonw.exe`) and check the server's own startup output.
+
+</details>
+
 <details>
 <summary><strong><code>--env-ref</code> or <code>${VAR}</code> not resolving</strong></summary>
 

EvoScientist/mcp/client.py

@@ -14,11 +14,15 @@ import re
 import shutil
 import sys
 from collections.abc import Callable
+from contextlib import asynccontextmanager
+from functools import wraps
 from pathlib import Path
 from typing import Any
 
 import yaml
 
+from ..runtime import AsyncRuntime, AsyncRuntimeError
+
 logger = logging.getLogger(__name__)
 
 
@@ -96,6 +100,178 @@ def _patch_mcp_windows_command_resolver() -> None:
 _patch_mcp_windows_command_resolver()
 
 
+# =============================================================================
+# Windows MCP SDK patch — give stdio subprocesses a real stderr file descriptor
+# =============================================================================
+#
+# ``mcp.client.stdio.stdio_client`` forwards the parent process's ``stderr``
+# (``errlog``, defaulting to ``sys.stderr``) to the MCP server subprocess via
+# ``anyio.open_process`` → ``subprocess.Popen(stderr=...)``. ``Popen`` resolves
+# that to an OS handle by calling ``errlog.fileno()``.
+#
+# Under the Textual TUI (and any non-console host), ``sys.stderr`` is
+# redirected to ``textual.app._PrintCapture``, whose ``fileno()`` returns
+# ``-1``. ``subprocess.Popen`` dutifully converts fd ``-1`` into an invalid
+# Windows handle and the child inherits a broken stderr pipe — every stdio
+# MCP server then fails to spawn with ``OSError: [Errno 9] Bad file
+# descriptor``. HTTP/SSE servers are unaffected (no subprocess), which is why
+# only the stdio server (e.g. arxiv) shows up as failed in the loader. See
+# issue #418; the same class of bug is tracked upstream as
+# modelcontextprotocol/python-sdk#1103.
+#
+# We can't pass ``errlog`` through ``langchain-mcp-adapters`` (it calls
+# ``stdio_client(server_params)`` with no ``errlog``), and the SDK's default
+# is bound once at import time — which may itself already capture a
+# redirected ``sys.stderr``. So we wrap ``stdio_client`` to swap in a safe
+# ``errlog`` at call time whenever the configured one has no usable fileno.
+# The substitute is ``sys.__stderr__`` (the real console handle) when
+# available, otherwise a discarded ``os.devnull`` handle.
+
+
+def _stdio_errlog_is_usable(errlog: object) -> bool:
+    """Return ``True`` if *errlog* can back a subprocess ``stderr`` pipe.
+
+    A usable errlog exposes a ``fileno()`` that resolves to a live OS file
+    descriptor. Textual's ``_PrintCapture`` and similar redirected streams
+    return ``-1`` (or raise), so they are rejected here. A closed stream may
+    still report its former (positive) fd, so we additionally ``os.fstat``
+    the descriptor to confirm it is still open.
+    """
+    fileno = getattr(errlog, "fileno", None)
+    if not callable(fileno):
+        return False
+    try:
+        fd = fileno()
+    except Exception:
+        return False
+    if not isinstance(fd, int) or fd < 0:
+        return False
+    try:
+        os.fstat(fd)
+    except (OSError, OverflowError):
+        return False
+    return True
+
+
+def _safe_stdio_errlog() -> tuple[Any, bool]:
+    """Return a ``(stream, opened_by_us)`` pair for a usable stderr.
+
+    Prefers ``sys.__stderr__`` (the original console handle, so the server's
+    diagnostic output still lands where the user expects — note
+    ``sys.__stderr__`` is the process's original handle and stays usable even
+    while the Textual TUI controls the screen, since Textual only redirects
+    ``sys.stderr``). Falls back to an ``os.devnull`` handle when even
+    ``__stderr__`` is unavailable (e.g. in a GUI/pythonw host with no console).
+
+    The second element is ``True`` when *we* allocated the stream (the
+    ``os.devnull`` case) and therefore own its lifecycle; it is ``False`` for
+    ``sys.__stderr__``, which is process-owned and must never be closed here.
+    Callers use that flag to decide whether to close the stream after the
+    stdio session exits.
+    """
+    dunder = getattr(sys, "__stderr__", None)
+    if dunder is not None and _stdio_errlog_is_usable(dunder):
+        return dunder, False
+    # Last resort: discard the server's stderr so the spawn still succeeds.
+    return open(os.devnull, "w", encoding="utf-8", errors="replace"), True
+
+
+def _patch_mcp_stdio_errlog_safe() -> None:
+    """Wrap the SDK's ``stdio_client`` to guarantee a usable ``errlog``.
+
+    Idempotent. A no-op when the MCP SDK is absent (optional dependency).
+    When the caller already supplied a usable ``errlog`` it is forwarded
+    unchanged; only the unsafe default (redirected ``sys.stderr``) is
+    replaced. This keeps the patch transparent for embedders that pass their
+    own ``errlog`` explicitly.
+
+    The wrapper is installed on both ``mcp.client.stdio.stdio_client`` and
+    ``langchain_mcp_adapters.sessions.stdio_client``: the adapter binds the
+    name via a ``from … import`` at its module load, so updating only the
+    SDK module would leave an already-imported adapter pointing at the
+    unwrapped function.
+    """
+    try:
+        import mcp.client.stdio as _stdio_mod
+    except ImportError:
+        return  # MCP SDK not installed — nothing to patch.
+
+    original = getattr(_stdio_mod, "stdio_client", None)
+    if original is None:
+        # The SDK renamed/removed stdio_client — nothing to wrap. Log so a
+        # future SDK refactor doesn't silently drop this guard.
+        logger.warning(
+            "MCP SDK layout changed: mcp.client.stdio.stdio_client is missing; "
+            "the Windows stdio errlog safety patch was NOT applied. MCP stdio "
+            "tool loading may fail with [Errno 9] under a redirected stderr."
+        )
+        return
+
+    if getattr(original, "_evosci_errlog_safe", False):
+        return  # Already patched.
+
+    @wraps(original)
+    def _stdio_client_safe(server: Any, errlog: Any = ..., *args: Any, **kwargs: Any):
+        # When the caller didn't supply a usable errlog we allocate a fallback
+        # stream (sys.__stderr__ or os.devnull). The SDK never closes a
+        # caller-provided errlog, so a devnull fallback would leak its fd on
+        # every MCP reload. We allocate the fallback inside the async context
+        # manager below so it is closed on exit — and, if the CM is discarded
+        # before being entered, Python finalises the async generator and runs
+        # the same ``finally``. ``errlog`` is forwarded by keyword so a future
+        # SDK that inserts a positional parameter before it can't mis-bind it.
+        caller_errlog = errlog
+        needs_fallback = errlog is ... or not _stdio_errlog_is_usable(errlog)
+
+        @asynccontextmanager
+        async def _close_owned_errlog():
+            if needs_fallback:
+                # ``opened_by_us`` is True only for the os.devnull case;
+                # sys.__stderr__ is process-owned and must not be closed.
+                errlog, opened_by_us = _safe_stdio_errlog()
+            else:
+                errlog, opened_by_us = caller_errlog, False
+            try:
+                # Construct inside the try so a failure here still reaches the
+                # finally and closes a wrapper-owned fallback stream.
+                # Forward by keyword: robust against future SDK signature changes
+                # that insert a positional parameter before ``errlog``.
+                cm = original(server, *args, errlog=errlog, **kwargs)
+                async with cm as streams:
+                    yield streams
+            finally:
+                if opened_by_us:
+                    close = getattr(errlog, "close", None)
+                    if callable(close):
+                        try:
+                            close()
+                        except Exception:
+                            logger.debug(
+                                "Failed to close fallback stdio errlog", exc_info=True
+                            )
+
+        return _close_owned_errlog()
+
+    _stdio_client_safe._evosci_errlog_safe = True  # type: ignore[attr-defined]
+    _stdio_mod.stdio_client = _stdio_client_safe
+
+    # langchain-mcp-adapters binds stdio_client via a ``from`` import at its
+    # module load, so a pre-imported adapter keeps the unwrapped reference.
+    # Re-bind it too (best-effort; ignore if the layout differs).
+    try:
+        import langchain_mcp_adapters.sessions as _adapter_sessions
+
+        if getattr(_adapter_sessions, "stdio_client", None) is original:
+            _adapter_sessions.stdio_client = _stdio_client_safe
+    except ImportError:
+        pass  # Adapter not installed — nothing extra to rebind.
+
+    logger.debug("Applied MCP stdio errlog safety patch")
+
+
+_patch_mcp_stdio_errlog_safe()
+
+
 # =============================================================================
 # Constants
 # =============================================================================
@@ -853,6 +1029,7 @@ def load_mcp_tools(
     config: dict[str, Any] | None = None,
     *,
     on_progress: ProgressCallback | None = None,
+    runtime: AsyncRuntime | None = None,
 ) -> dict[str, list]:
     """Load MCP tools and return them grouped by target agent.
 
@@ -869,6 +1046,10 @@ def load_mcp_tools(
             warnings when the caller has already loaded the config.
         on_progress: Optional callback invoked per server with
             ``(event, server_name, detail)``.  See :data:`ProgressCallback`.
+        runtime: Runtime that owns MCP discovery work. When omitted, this
+            function creates one scoped to this call. The returned adapters
+            open a fresh MCP session for each tool call and do not retain the
+            discovery loop.
 
     Returns:
         Dict mapping agent name -> list of LangChain ``BaseTool`` objects.
@@ -880,19 +1061,28 @@ def load_mcp_tools(
     if not config:
         return {}
 
-    try:
-        loop = asyncio.get_running_loop()
-    except RuntimeError:
-        loop = None
+    if runtime is None:
+        with AsyncRuntime(thread_name="evosci-mcp-runtime") as owned_runtime:
+            return load_mcp_tools(
+                config,
+                on_progress=on_progress,
+                runtime=owned_runtime,
+            )
 
     try:
-        if loop and loop.is_running():
-            # Inside an already-running event loop (e.g. Jupyter) —
-            # nest_asyncio patches the loop so asyncio.run() works.
-            import nest_asyncio
-
-            nest_asyncio.apply()
-        server_tools = asyncio.run(_load_tools(config, on_progress=on_progress))
+        server_tools = runtime.run_sync(
+            lambda: _load_tools(config, on_progress=on_progress)
+        )
+    except AsyncRuntimeError as exc:
+        # A bridge lifecycle/call-site error is not an MCP availability
+        # failure.  In particular, hiding a running-loop violation here makes
+        # callers cache an empty tool set for the rest of the process.
+        if "cannot block a running event loop" in str(exc):
+            raise AsyncRuntimeError(
+                "load_mcp_tools() cannot run inside an async context; use "
+                "`await aload_mcp_tools(config, on_progress=...)` instead"
+            ) from exc
+        raise
     except Exception as exc:
         logger.warning("MCP tool loading failed: %s", exc)
         return {}

EvoScientist/memory/agents/_factory.py

@@ -18,6 +18,7 @@ from ... import paths as _paths
 MEMORY_AGENT_RECURSION_LIMIT = 100
 MEMORY_MAINTENANCE_EXCLUDED_TOOLS = frozenset(
     {
+        "delete",
         "edit_file",
         "execute",
         "task",

EvoScientist/memory/agents/autoskills.py

@@ -21,7 +21,7 @@ from ._factory import (
     resolve_memory_agent_paths,
 )
 
-_AUTOSKILLS_EXCLUDED_TOOLS = frozenset({"task", "write_todos"})
+_AUTOSKILLS_EXCLUDED_TOOLS = frozenset({"delete", "task", "write_todos"})
 
 
 def _autoskills_system_prompt() -> str:

EvoScientist/memory/agents/memory_worker.py

@@ -34,7 +34,7 @@ from ._factory import (
 logger = logging.getLogger(__name__)
 
 _MEMORY_WORKER_EXCLUDED_TOOLS = frozenset(
-    {"execute", "task", "write_file", "write_todos"}
+    {"delete", "execute", "task", "write_file", "write_todos"}
 )
 
 

EvoScientist/memory/observations/store.py

@@ -11,6 +11,8 @@ import hashlib
 import json
 import os
 import tempfile
+import threading
+from collections import OrderedDict
 from dataclasses import replace
 from datetime import UTC, date, datetime
 from pathlib import Path
@@ -275,19 +277,11 @@ def _observation_files(
 ) -> list[Path]:
     """Return candidate observation files for the current project context."""
     root = Path(memory_dir).expanduser()
-    memory_paths: list[str] = []
-    if scope in {None, MemoryScope.GLOBAL}:
-        memory_paths.append(f"{OBSERVATION_DIR}/global")
-    if scope in {None, MemoryScope.PROJECT}:
-        memory_paths.append(f"{OBSERVATION_DIR}/projects/{project_id}")
-
     paths: list[Path] = []
-    for memory_path in memory_paths:
-        directory = root / memory_path.lstrip("/")
-        try:
-            paths.extend(sorted(directory.glob("*.md")))
-        except OSError:
-            continue
+    if scope in {None, MemoryScope.GLOBAL}:
+        paths.extend(_global_files(root))
+    if scope in {None, MemoryScope.PROJECT}:
+        paths.extend(_project_files(root, project_id))
     return paths
 
 
@@ -355,6 +349,7 @@ def _resolve_document_links(
     parsed: list[tuple[ObservationSearchDocument, list[RelatedObservationEntry]]],
     *,
     root: Path,
+    touched: set[str],
 ) -> list[ObservationSearchDocument]:
     documents_by_id = {document.observation_id: document for document, _ in parsed}
     missing_related_ids = {
@@ -367,7 +362,7 @@ def _resolve_document_links(
         for path in _all_observation_files(root):
             if not missing_related_ids:
                 break
-            parsed_document = _parse_observation_search_document(root=root, path=path)
+            parsed_document = _parse_with_cache(root, path, touched)
             if parsed_document is None:
                 continue
             document, _entries = parsed_document
@@ -388,6 +383,124 @@ def _resolve_document_links(
     ]
 
 
+# ── Parsed-document cache ─────────────────────────────────────────────
+#
+# One layer: a per-file parse cache keyed on the observation file path.
+# Each value carries the file's ``(st_mtime_ns, st_size)`` signature, so a
+# rewrite invalidates exactly that entry and replaces it in place.  Scope
+# filtering is just which directories get globbed, and link resolution
+# reruns on every call over the memoized parses — its fallback walk parses
+# through the same cache, so fallback visits are dict hits instead of
+# whole-tree reparses.
+#
+# Deletion needs no invalidation: a deleted file is never globbed, so its
+# cached entry can never be served again; the LRU bounds the leftover
+# memory.  Eviction runs only at the end of a call, down to
+# ``max(cap, entries touched by the call)``, so a call never evicts its own
+# working set and a store larger than the cap temporarily exceeds it
+# instead of thrashing.
+#
+# The cap is ``config.memory_observation_cache_max_files`` (default 2048,
+# env var ``EVOSCIENTIST_MAX_CACHED_FILES``), read lazily from config on
+# first call.  With the working-set rule it bounds retained memory for
+# inactive workspaces rather than correctness.
+#
+_FileParseValue = tuple[
+    tuple[int, int],
+    tuple[ObservationSearchDocument, list[RelatedObservationEntry]],
+]
+
+_file_parse_cache: OrderedDict[str, _FileParseValue] = OrderedDict()
+
+# Serializes cache transactions (lookup+recency, insert+recency, trim) so
+# concurrent calls cannot evict a key between another call's lookup and its
+# recency update. Parsing stays outside the lock; only dict mutations hold it.
+_cache_lock = threading.Lock()
+
+_cached_max_files: int | None = None
+
+
+def _max_cached_files() -> int:
+    """Return the configured cache cap.
+
+    Read once from ``get_effective_config()`` on first call and cached at
+    module level; a runtime config change requires a process restart to take
+    effect.
+    """
+    global _cached_max_files
+    if _cached_max_files is None:
+        from ...config import get_effective_config
+
+        _cached_max_files = get_effective_config().memory_observation_cache_max_files
+    return _cached_max_files
+
+
+def _parse_with_cache(
+    root: Path,
+    path: Path,
+    touched: set[str],
+) -> tuple[ObservationSearchDocument, list[RelatedObservationEntry]] | None:
+    """Return the parsed document for *path*, memoized per file.
+
+    The cached entry is validated against the file's current
+    ``(st_mtime_ns, st_size)`` signature, so a rewrite replaces it in
+    place.  Parse failures are not cached; the file is retried on the next
+    call.  No copy on read or write: callers only iterate the documents or
+    build new ones via ``replace``.
+    """
+    key = str(path)
+    try:
+        st = path.stat()
+    except OSError:
+        return None
+    signature = (st.st_mtime_ns, st.st_size)
+    with _cache_lock:
+        cached = _file_parse_cache.get(key)
+        if cached is not None and cached[0] == signature:
+            _file_parse_cache.move_to_end(key)
+            touched.add(key)
+            return cached[1]
+    parsed_document = _parse_observation_search_document(root=root, path=path)
+    if parsed_document is None:
+        return None
+    with _cache_lock:
+        _file_parse_cache[key] = (signature, parsed_document)
+        _file_parse_cache.move_to_end(key)
+        touched.add(key)
+    return parsed_document
+
+
+def _trim_parse_cache(touched: set[str]) -> None:
+    """Trim the cache at the end of a call down to ``max(cap, len(touched))``.
+
+    Every entry touched by the call was moved or appended to the
+    most-recent end, so trimming from the least-recent end down to at least
+    ``len(touched)`` entries never evicts the call's own working set.
+    """
+    target = max(_max_cached_files(), len(touched))
+    with _cache_lock:
+        while len(_file_parse_cache) > target:
+            _file_parse_cache.popitem(last=False)
+
+
+def _global_files(root: Path) -> list[Path]:
+    """Glob the global observation directory."""
+    directory = root / OBSERVATION_DIR.lstrip("/") / "global"
+    try:
+        return sorted(directory.glob("*.md"))
+    except OSError:
+        return []
+
+
+def _project_files(root: Path, project_id: str) -> list[Path]:
+    """Glob a project's observation directory."""
+    directory = root / OBSERVATION_DIR.lstrip("/") / "projects" / project_id
+    try:
+        return sorted(directory.glob("*.md"))
+    except OSError:
+        return []
+
+
 def list_observation_documents(
     *,
     memory_dir: str | Path,
@@ -395,26 +508,32 @@ def list_observation_documents(
     scope: MemoryScope | None = None,
     memory_type: MemoryType | None = None,
 ) -> list[ObservationSearchDocument]:
-    """Read candidate observations for the current filters."""
+    """Read candidate observations for the current filters.
+
+    Observation files are parsed once per file and shared across every
+    project_id and scope; each cached entry is validated against the file's
+    current ``(st_mtime_ns, st_size)`` signature.  A deleted file simply
+    drops out of the glob.  Link resolution reruns on each call over the
+    memoized parses.
+    """
     root = Path(memory_dir).expanduser()
+    paths = _observation_files(memory_dir=root, project_id=project_id, scope=scope)
+
+    touched: set[str] = set()
     parsed: list[tuple[ObservationSearchDocument, list[RelatedObservationEntry]]] = []
-    for path in _observation_files(
-        memory_dir=root,
-        project_id=project_id,
-        scope=scope,
-    ):
-        parsed_document = _parse_observation_search_document(root=root, path=path)
+    for path in paths:
+        parsed_document = _parse_with_cache(root, path, touched)
         if parsed_document is not None:
             parsed.append(parsed_document)
 
-    # Resolve links before filtering by memory_type so a procedural hit can still
-    # surface a linked semantic observation, and vice versa.
-    documents = _resolve_document_links(parsed, root=root)
+    documents = _resolve_document_links(parsed, root=root, touched=touched)
+    _trim_parse_cache(touched)
+
     if memory_type is not None:
         return [
             document for document in documents if document.memory_type == memory_type
         ]
-    return documents
+    return list(documents)
 
 
 def search_observation_files(

EvoScientist/middleware/__init__.py

@@ -4,6 +4,7 @@ Re-exports middleware classes and factory functions so that existing
 ``from EvoScientist.middleware import X`` imports continue to work.
 """
 
+from .active_team import ActiveTeamMiddleware, create_active_team_middleware
 from .ask_user import (
     AskUserMiddleware,
     AskUserRequest,
@@ -53,11 +54,13 @@ from .skill_context import (
 )
 from .subagent_timeout import SubagentTimeoutMiddleware
 from .tool_error_handler import ToolErrorHandlerMiddleware
+from .tool_history_repair import ToolHistoryRepairMiddleware
 from .tool_protocol_guard import ToolProtocolGuardMiddleware
 from .tool_selector import create_tool_selector_middleware
 from .utils import disable_thinking
 
 __all__ = [
+    "ActiveTeamMiddleware",
     "DEFAULT_MAX_CONSECUTIVE_TOOL_ERRORS",
     "DEFAULT_MAX_DESCRIPTION_BYTES",
     "DEFAULT_MAX_SKILLS",
@@ -85,9 +88,11 @@ __all__ = [
     "SchedulerMiddleware",
     "SubagentTimeoutMiddleware",
     "ToolErrorHandlerMiddleware",
+    "ToolHistoryRepairMiddleware",
     "ToolProtocolGuardMiddleware",
     "collapse_repetitive_tool_rounds",
     "compute_context_editing_trigger",
+    "create_active_team_middleware",
     "create_code_interpreter_middleware",
     "create_context_editing_middleware",
     "create_memory_lifecycle_middleware",

EvoScientist/middleware/active_team.py

@@ -0,0 +1,166 @@
+"""ActiveTeamMiddleware: the expert prompt for the main agent.
+
+Injects the ``## Experts`` concept into the system message on every
+main-agent turn, so the expert mechanism is always visible — mirroring how
+the skill system's guidance is always present. When the user has invited
+experts (``configurable.active_teams``), an ``<active_expert>`` block naming
+the reachable ones is appended on top.
+
+An expert is a fractal of a skill, so this block is ordered (via the
+middleware stack in ``_get_default_middleware``) to land right after
+``## Skills System``. Gating the whole block on invitation is the trap this
+design avoids: the expert mechanism must not disappear when nothing is
+invited, and the invited-expert list must not read as a standalone "always
+dispatch an expert" directive.
+
+Backend-stateless team binding: WebUI sends ``active_teams`` on every
+``stream.submit()`` for as long as the invited expert is active; this
+middleware reads it fresh per turn via ``langgraph.config.get_config()`` —
+the ``configurable`` primitive, not a server-side thread-state store
+(CLAUDE.md #5). The wire key stays ``active_teams`` (plural, legacy from the
+earlier "teams" framing); the semantic content is a list of expert names.
+
+Not included in the async-subagent middleware stack: an expert running as
+its own graph would otherwise inject the expert prompt into its own system
+message, where its persona is already baked in. See
+``EvoScientist.py::_get_default_middleware``.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Awaitable, Callable
+
+from langchain.agents.middleware.types import (
+    AgentMiddleware,
+    ModelRequest,
+    ModelResponse,
+)
+
+# The expert concept — injected on every main-agent turn so the mechanism is
+# always visible, mirroring how ``## Skills System`` is always present. Moved
+# here from ``DELEGATION_STRATEGY`` (an expert is a fractal of a skill, so its
+# guidance belongs next to the skill system's). The invited-expert list is
+# appended below only when the user has invited experts this session.
+EXPERTS_CONCEPT = """## Experts
+An expert is an installed skill that also ships an actor definition — a persona and a result-envelope contract. Every installed expert is reachable both ways, and the choice is yours per task, not fixed per expert:
+
+- `task({subagent_type: '<expert>', description: ...})` — runs in-turn and returns into the current turn. Use when the answer is short and the user is waiting on it.
+- `start_async_task(subagent_type: '<expert>', description: ...)` — runs in the background, returns a task ID immediately. Use when the work is long-running or its deliverable is a file. **Name a concrete output path in the description** (e.g. "write to `./artifacts/<expert>/<slug>.md`") — the expert honours the path you give it. On `status: 'success'`, `check_async_task` returns a `result` envelope with `output_path`, a one-paragraph `summary`, and an expert-defined `metadata` block; render `summary` and `metadata` to the user directly rather than re-reading the artifact to build a synopsis.
+
+Prefer the background form when unsure — expert work is usually multi-step, and it keeps the conversation responsive.
+
+You need not dispatch at all. An expert's `SKILL.md` is ordinary knowledge on the `/skills/` mount: read it and do the work yourself when the task is small, or when the full conversation context matters more than a fresh sub-agent would. If an `<active_expert>` block appears below, the user invited that expert specifically — prefer it for requests in its scope."""
+
+# Appended to ``EXPERTS_CONCEPT`` only when the user has invited reachable
+# experts. One ``<active_expert>`` tag handles one or many names.
+_INVITE_TEMPLATE = (
+    "\n\n<active_expert>\n"
+    "The user has invited {experts} to this thread. Prefer the right one for "
+    "requests within its scope; do not consult an expert if the request is "
+    "clearly outside its scope. They stay available for the whole session "
+    "until the user dismisses them.\n"
+    "</active_expert>"
+)
+
+
+def _read_active_teams() -> list[str]:
+    """Read ``configurable.active_teams`` from the current RunnableConfig.
+
+    Returns an empty list when the config is absent, malformed, or the
+    call happens outside a runnable context.
+    """
+    try:
+        from langgraph.config import get_config
+
+        cfg = get_config()
+    except Exception:
+        # Outside a runnable context (most common in tests) or
+        # langgraph not importable — nothing to inject.
+        return []
+    if not isinstance(cfg, dict):
+        return []
+    configurable = cfg.get("configurable") or {}
+    if not isinstance(configurable, dict):
+        return []
+    raw = configurable.get("active_teams")
+    if not isinstance(raw, list):
+        return []
+    return [t for t in raw if isinstance(t, str) and t]
+
+
+def _dispatchable_names() -> set[str]:
+    """Return the names of experts the orchestrator can currently reach.
+
+    Fresh filesystem read every call so a ``skill_manager install <expert>``
+    is visible on the next turn without an agent rebuild. Cheap at current
+    scale (a handful of skills, cached bodies).
+
+    Sourced from ``list_dispatchable_experts``, which drops empty-body
+    experts and names colliding with reserved sub-agents. Keeps the cue
+    honest: naming an expert the model cannot reach is worse than saying
+    nothing.
+
+    On import failure returns an empty set — the middleware then emits no
+    cue, matching the outside-runnable-context no-op path.
+    """
+    try:
+        from ..subagents.expert_container import list_dispatchable_experts
+    except Exception:
+        return set()
+    try:
+        return {s.name for s in list_dispatchable_experts()}
+    except Exception:
+        return set()
+
+
+class ActiveTeamMiddleware(AgentMiddleware):
+    """Bias delegation toward the user's active expert(s) on every turn."""
+
+    name = "active_team"
+
+    def _invite_block(self, experts: list[str]) -> str:
+        """Render the ``<active_expert>`` block over the dispatchable subset.
+
+        Invited experts that aren't currently dispatchable (uninstalled,
+        empty actor definition, name collision) are dropped — naming an
+        expert the model cannot reach is worse than saying nothing. Returns
+        the empty string when nothing survives the filter.
+        """
+        reachable = _dispatchable_names()
+        experts = [e for e in experts if e in reachable]
+        if not experts:
+            return ""
+        names = ", ".join(f"`{e}`" for e in experts)
+        return _INVITE_TEMPLATE.format(experts=names)
+
+    def modify_request(self, request: ModelRequest) -> ModelRequest:
+        """Append the expert concept (always) plus the invited-expert block
+        (when the user has invited reachable experts) to the system message.
+        """
+        block = EXPERTS_CONCEPT
+        invited = _read_active_teams()
+        if invited:
+            block += self._invite_block(invited)
+        from .utils import append_to_system_message
+
+        new_system = append_to_system_message(request.system_message, block)
+        return request.override(system_message=new_system)
+
+    def wrap_model_call(
+        self,
+        request: ModelRequest,
+        handler: Callable[[ModelRequest], ModelResponse],
+    ) -> ModelResponse:
+        return handler(self.modify_request(request))
+
+    async def awrap_model_call(
+        self,
+        request: ModelRequest,
+        handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
+    ) -> ModelResponse:
+        return await handler(self.modify_request(request))
+
+
+def create_active_team_middleware() -> ActiveTeamMiddleware:
+    """Build ActiveTeamMiddleware."""
+    return ActiveTeamMiddleware()

EvoScientist/middleware/async_watcher.py

@@ -22,13 +22,16 @@ from __future__ import annotations
 
 import logging
 from collections.abc import Awaitable, Callable
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 from langchain.agents.middleware import AgentMiddleware
 from langchain.agents.middleware.types import ToolCallRequest
 from langchain_core.messages import ToolMessage
 from langgraph.types import Command
 
+if TYPE_CHECKING:
+    from .notifier import NotifierPort
+
 logger = logging.getLogger(__name__)
 
 _LAUNCH_TOOL_NAMES = ("start_async_task", "update_async_task")
@@ -42,42 +45,30 @@ class AsyncWatcherMiddleware(AgentMiddleware):
         async_agents: Mapping of subagent name → ``AsyncSubAgent`` TypedDict
             (must contain at least ``url`` and ``graph_id``). Used to construct
             a ``_ClientCache`` for resolving the LangGraph client per agent.
+        notifier: Injected :class:`~EvoScientist.middleware.notifier.NotifierPort`
+            used to pre-cancel stale watchers and spawn new ones. The composition
+            root supplies ``EvoScientist.cli.async_notifier``.
     """
 
-    def __init__(self, async_agents: dict[str, Any]) -> None:
+    def __init__(self, async_agents: dict[str, Any], notifier: NotifierPort) -> None:
         from deepagents.middleware.async_subagents import _ClientCache
 
         super().__init__()
         self._clients = _ClientCache(async_agents)
+        self._notifier = notifier
 
     async def awrap_tool_call(
         self,
         request: ToolCallRequest,
         handler: Callable[[ToolCallRequest], Awaitable[ToolMessage | Command]],
     ) -> ToolMessage | Command:
-        from EvoScientist.cli import async_notifier
-
         name = request.tool_call.get("name")
         args = request.tool_call.get("args") or {}
 
         # Pre-cancel the existing watcher BEFORE the new run interrupts the old
-        # one. ``update_async_task`` creates a new run on the same thread_id with
-        # ``multitask_strategy="interrupt"``, which closes the old run's stream
-        # cleanly — without pre-cancellation the old watcher would observe a
-        # clean exit and enqueue a stale "success" notification before the new
-        # spawn can replace it.
+        # one (see NotifierPort.pre_cancel_watcher for the full rationale).
         if name == "update_async_task" and (tid := args.get("task_id")):
-            try:
-                old = async_notifier._watcher_by_thread.get(tid)
-                if old is not None and not old.done():
-                    old.cancel()
-            except Exception:
-                logger.warning(
-                    "Pre-cancel of stale watcher for task %s failed; a stale "
-                    "success notification may be enqueued",
-                    tid,
-                    exc_info=True,
-                )
+            self._notifier.pre_cancel_watcher(tid)
 
         result = await handler(request)
 
@@ -96,7 +87,7 @@ class AsyncWatcherMiddleware(AgentMiddleware):
             for task_id, task in tasks_update.items():
                 try:
                     client = self._clients.get_async(task["agent_name"])
-                    async_notifier.spawn_watcher(
+                    self._notifier.spawn_watcher(
                         client,
                         task_id,
                         task["run_id"],

EvoScientist/middleware/background.py

@@ -11,7 +11,7 @@ sub-agents are *tasks*, future cron is *schedules*).
 
 from __future__ import annotations
 
-from datetime import UTC, datetime
+from typing import TYPE_CHECKING
 
 from langchain.agents.middleware import AgentMiddleware
 from langchain.tools import ToolRuntime
@@ -20,6 +20,9 @@ from langchain_core.tools import tool
 from .. import background, paths
 from ..backends import prepare_sandbox_command
 
+if TYPE_CHECKING:
+    from .notifier import NotifierPort
+
 
 def _origin_thread_id(runtime: ToolRuntime | None) -> str | None:
     """Best-effort current CLI thread_id, used to route the completion notification."""
@@ -29,11 +32,15 @@ def _origin_thread_id(runtime: ToolRuntime | None) -> str | None:
         return None
 
 
-def _notify_done(proc: background.BgProcess, origin_thread_id: str | None) -> None:
-    """Watcher ``on_exit`` hook: enqueue a completion notification (reuses async_notifier).
+def _notify_done(
+    proc: background.BgProcess,
+    origin_thread_id: str | None,
+    notifier: NotifierPort,
+) -> None:
+    """Watcher ``on_exit`` hook: enqueue a completion notification via the port.
 
-    Skipped for user-stopped processes (the user already knows). The notifier is imported
-    lazily to keep this module free of a load-time dependency on the CLI layer.
+    Skipped for user-stopped processes (the user already knows). The notifier
+    port owns the notification type, so this module never imports the CLI layer.
     """
     if proc.stopped:
         return
@@ -44,69 +51,78 @@ def _notify_done(proc: background.BgProcess, origin_thread_id: str | None) -> No
         status = "interrupted"  # terminated by a signal
     else:
         status = "error"
-    from ..cli import async_notifier
-
-    async_notifier._enqueue(
-        async_notifier.AsyncTaskNotification(
-            task_id=proc.process_id,
-            agent_name=proc.name,
-            status=status,
-            received_at=datetime.now(UTC).strftime("%Y-%m-%dT%H:%M:%SZ"),
-            prompt=proc.command,
-            kind="bg-process",
-            origin_cli_thread_id=origin_thread_id,
-        )
+    notifier.enqueue_bg_process_notification(
+        task_id=proc.process_id,
+        agent_name=proc.name,
+        status=status,
+        prompt=proc.command,
+        origin_cli_thread_id=origin_thread_id,
     )
 
 
-@tool(parse_docstring=True)
-def run_in_background(
-    command: str, name: str | None = None, runtime: ToolRuntime = None
-) -> str:
-    """Launch a long-running shell command in the background and return immediately.
+def _make_run_in_background(
+    notifier: NotifierPort, dangerous: bool, guard_dangerous: bool = False
+):
+    """Build the ``run_in_background`` tool bound to an injected notifier + policy.
 
-    Use for unbounded or very long tasks (model training, large downloads, servers)
-    that should not block the conversation. Output streams to a log file; poll it with
-    check_process and stop it with stop_process. For a bounded command that just needs
-    more time, prefer execute(..., timeout=N) instead of backgrounding.
-
-    Args:
-        command: The shell command to run in the background.
-        name: Optional short label to recognize the process later.
+    ``dangerous`` is captured from ``cfg.dangerous_mode`` at assembly (the agent
+    is rebuilt when config changes, so the captured value never goes stale), and
+    the notifier is the injected port used for the completion notification.
+    ``guard_dangerous`` mirrors ``execute``'s backstop: with no interactive
+    approval reachable (``auto_approve``), refuse the narrow dangerous set
+    instead of running it unattended.
     """
-    cwd = str(paths.resolve_virtual_path("/"))
-    # Honor dangerous mode so background commands match `execute`'s policy
-    # (real-filesystem access, no virtual-path rewriting). Read the env flag that
-    # apply_config_to_env round-trips at startup (and the subprocess inherits) —
-    # cheaper than reloading the full config from disk on every launch, and uses
-    # the same truthy parsing as every other bool env flag.
-    from ..llm.models import _env_flag_enabled
 
-    dangerous = _env_flag_enabled("EVOSCIENTIST_DANGEROUS_MODE")
-    # Same path-rewriting + validation as execute (shared helper) so virtual paths
-    # resolve to the workspace and the command can't bypass the sandbox checks.
-    command, error = prepare_sandbox_command(
-        command, cwd, virtual_mode=not dangerous, dangerous=dangerous
-    )
-    if error:
-        return error
-    tid = _origin_thread_id(runtime)
-    process_id = background.launch(
-        command, cwd, name, origin_thread_id=tid, on_exit=lambda p: _notify_done(p, tid)
-    )
-    label = f" (name={name!r})" if name else ""
-    # In dangerous mode `/` is the real root, so advertise the real log path;
-    # in virtual mode `/.bg_processes/...` correctly maps to the workspace.
-    log_path = (
-        f"{cwd}/.bg_processes/{process_id}.log"
-        if dangerous
-        else f"/.bg_processes/{process_id}.log"
-    )
-    return (
-        f"Started background process {process_id}{label}. "
-        f"Output -> {log_path}. "
-        f"Poll with check_process('{process_id}'), stop with stop_process('{process_id}')."
-    )
+    @tool(parse_docstring=True)
+    def run_in_background(
+        command: str, name: str | None = None, runtime: ToolRuntime = None
+    ) -> str:
+        """Launch a long-running shell command in the background and return immediately.
+
+        Use for unbounded or very long tasks (model training, large downloads, servers)
+        that should not block the conversation. Output streams to a log file; poll it with
+        check_process and stop it with stop_process. For a bounded command that just needs
+        more time, prefer execute(..., timeout=N) instead of backgrounding.
+
+        Args:
+            command: The shell command to run in the background.
+            name: Optional short label to recognize the process later.
+        """
+        cwd = str(paths.resolve_virtual_path("/"))
+        # Same path-rewriting + validation as execute (shared helper) so virtual paths
+        # resolve to the workspace and the command can't bypass the sandbox checks.
+        command, error = prepare_sandbox_command(
+            command,
+            cwd,
+            virtual_mode=not dangerous,
+            dangerous=dangerous,
+            guard_dangerous=guard_dangerous,
+        )
+        if error:
+            return error
+        tid = _origin_thread_id(runtime)
+        process_id = background.launch(
+            command,
+            cwd,
+            name,
+            origin_thread_id=tid,
+            on_exit=lambda p: _notify_done(p, tid, notifier),
+        )
+        label = f" (name={name!r})" if name else ""
+        # In dangerous mode `/` is the real root, so advertise the real log path;
+        # in virtual mode `/.bg_processes/...` correctly maps to the workspace.
+        log_path = (
+            f"{cwd}/.bg_processes/{process_id}.log"
+            if dangerous
+            else f"/.bg_processes/{process_id}.log"
+        )
+        return (
+            f"Started background process {process_id}{label}. "
+            f"Output -> {log_path}. "
+            f"Poll with check_process('{process_id}'), stop with stop_process('{process_id}')."
+        )
+
+    return run_in_background
 
 
 @tool(parse_docstring=True)
@@ -146,6 +162,17 @@ class BackgroundExecutionMiddleware(AgentMiddleware):
     Attached to the main agent only (async sub-agents must not spawn local processes).
     """
 
-    def __init__(self) -> None:
+    def __init__(
+        self,
+        notifier: NotifierPort,
+        *,
+        dangerous: bool = False,
+        guard_dangerous: bool = False,
+    ) -> None:
         super().__init__()
-        self.tools = [run_in_background, check_process, stop_process, list_processes]
+        self.tools = [
+            _make_run_in_background(notifier, dangerous, guard_dangerous),
+            check_process,
+            stop_process,
+            list_processes,
+        ]

EvoScientist/middleware/code_interpreter.py

@@ -29,6 +29,11 @@ Usage::
 
 from __future__ import annotations
 
+import asyncio
+import contextlib
+import logging
+import weakref
+
 from langchain.agents.middleware.types import ModelRequest
 from langchain_quickjs import CodeInterpreterMiddleware
 
@@ -37,12 +42,17 @@ from langchain_quickjs import CodeInterpreterMiddleware
 # values; tests / ad-hoc callers can omit and get sensible defaults.
 _DEFAULT_TIMEOUT_SECONDS: float = 60.0
 _DEFAULT_MAX_RESULT_CHARS: int = 10000
+_CLOSE_TIMEOUT_SECONDS: float = 10.0
+
+logger = logging.getLogger(__name__)
 
 _MEMORY_FIRST_INTERPRETER_PROMPT = (
     "\n\nWhen memory tools (search_observations, read_memory) are available, use "
     "them before `code_interpreter` for workspace inspection or implementation work."
 )
 
+_live_interpreters: weakref.WeakSet[EvoCodeInterpreterMiddleware]
+
 
 class EvoCodeInterpreterMiddleware(CodeInterpreterMiddleware):
     """Code interpreter middleware with EvoScientist's memory preflight hint.
@@ -51,12 +61,12 @@ class EvoCodeInterpreterMiddleware(CodeInterpreterMiddleware):
     earlier "conditional snapshot" gate that skipped ``after_agent`` on turns
     where ``code_interpreter`` wasn't called saved ~50 ms/turn of
     ``create_snapshot()`` work, but also skipped the slot eviction upstream
-    performs in the same hook (``finally: self._registry.evict(thread_id)``
+    performs in the same hook (``finally: self._registry.evict(slot_id)``
     in ``langchain_quickjs.middleware.CodeInterpreterMiddleware.after_agent``).
     ``before_agent`` restores the REPL on every turn that follows a touched
-    one via ``self._registry.get(thread_id)`` (get-or-create), so skipping
+    one via ``self._registry.get(slot_id)`` (get-or-create), so skipping
     eviction leaked one ``ThreadWorker`` + QuickJS Runtime per persistent
-    ``thread_id`` that ever went touched → quiet. The regression test
+    slot that ever went touched → quiet. The regression test
     ``test_after_agent_evicts_slot_on_untouched_turn`` guards against
     reintroducing the gate.
     """
@@ -64,6 +74,49 @@ class EvoCodeInterpreterMiddleware(CodeInterpreterMiddleware):
     def _prepare_for_call(self, request: ModelRequest) -> str:
         return super()._prepare_for_call(request) + _MEMORY_FIRST_INTERPRETER_PROMPT
 
+    async def aclose(self) -> None:
+        """Evict active REPLs on their worker loops before event-loop shutdown."""
+        registry = self._registry
+        with registry._lock:
+            slot_ids = tuple(registry._slots)
+        for slot_id in slot_ids:
+            with contextlib.suppress(Exception):
+                await registry.aevict(slot_id)
+        self._ptc_tools_by_slot.clear()
+
+
+_live_interpreters = weakref.WeakSet()
+
+
+async def aclose_code_interpreters(
+    *,
+    timeout: float = _CLOSE_TIMEOUT_SECONDS,
+) -> None:
+    """Close live QuickJS middleware without blocking application shutdown."""
+    middlewares = tuple(_live_interpreters)
+    if not middlewares:
+        return
+
+    close_tasks = [middleware.aclose() for middleware in middlewares]
+    try:
+        results = await asyncio.wait_for(
+            asyncio.gather(*close_tasks, return_exceptions=True),
+            timeout=timeout,
+        )
+    except TimeoutError:
+        logger.warning(
+            "code interpreter cleanup did not finish within %g seconds",
+            timeout,
+        )
+        return
+
+    for result in results:
+        if isinstance(result, BaseException):
+            logger.debug(
+                "code interpreter cleanup failed",
+                exc_info=(type(result), result, result.__traceback__),
+            )
+
 
 # Read-only, batchable tools that benefit from being callable inside JS.
 # Multi-agent orchestration is the killer use case: ``Promise.all`` over
@@ -109,9 +162,11 @@ def create_code_interpreter_middleware(
         Configured ``CodeInterpreterMiddleware`` ready to append to an agent's
         middleware stack.
     """
-    return EvoCodeInterpreterMiddleware(
+    middleware = EvoCodeInterpreterMiddleware(
         ptc=_DEFAULT_PTC_ALLOWLIST,
         timeout=timeout,
         max_result_chars=max_result_chars,
         tool_name="code_interpreter",
     )
+    _live_interpreters.add(middleware)
+    return middleware

EvoScientist/middleware/error_normalization.py

@@ -46,6 +46,110 @@ if TYPE_CHECKING:
     from ..llm.errors import ProviderStreamError
 
 
+class ModelOutputTruncatedError(RuntimeError):
+    """The provider exhausted its output budget before producing an answer."""
+
+
+_TRUNCATED_FINISH_REASONS = frozenset(
+    {
+        "length",
+        "max_tokens",
+        "max_output_tokens",
+        "max_completion_tokens",
+        "incomplete",
+    }
+)
+
+
+def _has_answer_content(content: object) -> bool:
+    """Return whether message content contains something beyond reasoning."""
+    if isinstance(content, str):
+        return bool(content.strip())
+    if not isinstance(content, list):
+        return content is not None
+
+    reasoning_types = {
+        "thinking",
+        "redacted_thinking",
+        "reasoning",
+        "reasoning_content",
+    }
+    text_types = {"text", "output_text"}
+    for block in content:
+        if isinstance(block, str):
+            if block.strip():
+                return True
+            continue
+        if not isinstance(block, dict):
+            return True
+        block_type = str(block.get("type", "")).lower()
+        if block_type in reasoning_types:
+            continue
+        if block_type in text_types:
+            text = block.get("text")
+            if isinstance(text, str):
+                if text.strip():
+                    return True
+            elif text:
+                return True
+            continue
+        # Any non-reasoning block is meaningful output (text, image, refusal,
+        # server tool result, etc.), even when its provider-specific payload
+        # does not use a ``text`` key.
+        return True
+    return False
+
+
+def _truncated_empty_message(response: ModelResponse):
+    """Return the empty truncated AI message in *response*, if present."""
+    from langchain_core.messages import AIMessage
+
+    if getattr(response, "structured_response", None) is not None:
+        return None
+    messages = getattr(response, "result", None) or []
+    message = next(
+        (item for item in reversed(messages) if isinstance(item, AIMessage)), None
+    )
+    if message is None:
+        return None
+    if _has_answer_content(message.content):
+        return None
+    if message.tool_calls or getattr(message, "invalid_tool_calls", None):
+        return None
+
+    metadata = message.response_metadata or {}
+    reasons = {
+        str(metadata.get(key, "")).strip().lower()
+        for key in ("finish_reason", "stop_reason", "status")
+    }
+    incomplete_details = metadata.get("incomplete_details")
+    if isinstance(incomplete_details, dict):
+        reasons.add(str(incomplete_details.get("reason", "")).strip().lower())
+    if reasons.isdisjoint(_TRUNCATED_FINISH_REASONS):
+        return None
+    return message
+
+
+def _check_truncated_output(response: ModelResponse) -> ModelResponse:
+    """Raise a visible error instead of silently accepting an empty answer."""
+    message = _truncated_empty_message(response)
+    if message is None:
+        return response
+    metadata = message.response_metadata or {}
+    reason = (
+        metadata.get("finish_reason")
+        or metadata.get("stop_reason")
+        or metadata.get("status")
+        or "output limit"
+    )
+    raise ModelOutputTruncatedError(
+        "The model exhausted its output budget during reasoning and returned "
+        f"no answer (finish reason: {reason}). Lower reasoning_effort, disable "
+        "reasoning with none when supported, or increase the provider "
+        "output-token limit."
+    )
+
+
 def _should_pass_through(exc: BaseException) -> bool:
     """True if *exc* is a LangGraph-level signal that must propagate
     untouched — either a control-flow signal or a structural error
@@ -90,6 +194,7 @@ _PROVIDER_EXC_MODULE_PREFIXES: tuple[str, ...] = (
     "google.api_core",
     "openrouter",
     "langchain_openai",
+    "langchain_deepseek",
     "langchain_anthropic",
     "langchain_google_genai",
     "langchain_openrouter",
@@ -237,7 +342,7 @@ class ErrorNormalizationMiddleware(AgentMiddleware):
         handler: Callable[[ModelRequest], ModelResponse],
     ) -> ModelResponse:
         try:
-            return handler(request)
+            return _check_truncated_output(handler(request))
         except Exception as exc:
             normalized = _normalize(request, exc)
             if normalized is None:
@@ -250,7 +355,7 @@ class ErrorNormalizationMiddleware(AgentMiddleware):
         handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
     ) -> ModelResponse:
         try:
-            return await handler(request)
+            return _check_truncated_output(await handler(request))
         except Exception as exc:
             normalized = _normalize(request, exc)
             if normalized is None: