ouyangbo/hermes-agent
1
0
Fork 0
Code Issues Pull Requests Actions 9 Packages Projects Releases Wiki Activity

refactor(tui_gateway): tighten W5 modules (host_supervisor, agent_callbacks, model_switch, change_watcher, compute_host_bridge, browser, hosted-room adapters, git_probe, method_ctx, bot_relay) 3037->2680 LOC, zero behavior change

Browse Source
This commit is contained in:
Teknium
2026-09-03 02:28:12 -07:00
parent e1dc1ff22d
commit e7e287cdd8
11 changed files with 493 additions and 850 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tui_gateway/agent_callbacks.py
+92 -145
View File
@@ -1,9 +1,6 @@
"""Agent callback wiring: child-session live mirror, per-session agent callbacks,
personality overlay, background/preview agent kwargs, agent reset.
Bodies are rebound onto server.py's globals at install time (see
method_ctx.bind_module), so they reference server.py globals bare.
"""
"""Agent callback wiring: child-session live mirror, per-session agent callbacks, personality
overlay, background/preview agent kwargs, agent reset. Bodies are rebound onto server.py's
globals at install time (method_ctx.bind_module), so they reference server.py globals bare."""
from __future__ import annotations
@@ -12,18 +9,18 @@ import threading
from .method_ctx import bind_module
# Child-session live mirror: a delegated child's activity reaches the gateway only
# as relayed ``subagent.*`` events on the PARENT sid, so a window opened on the
# child's own session would sit silent until the run persists. Translate them into
# the native stream events emitted on the CHILD sid (write_json routes by sid).
# Child-session live mirror: a delegated child's activity reaches the gateway only as
# relayed ``subagent.*`` events on the PARENT sid; translate them into native stream
# events on the CHILD sid (write_json routes by sid) so its own window is not silent.
_child_mirrors: dict[str, dict] = {}
_child_mirrors_lock = threading.Lock()
# Child session ids with a run in flight (refreshed per relayed event, popped on
# complete) so a lazy watch resume reports running=true during a silent long tool.
# Child sids with a run in flight (refreshed per relayed event, popped on complete) so a
# lazy watch resume reports running=true during a silent long tool.
_active_child_runs: dict[str, float] = {}
# Anything quiet this long lost its completion event (callback raised, parent
# crashed) — don't pin "running".
# Anything quiet this long lost its completion event — don't pin "running".
_CHILD_RUN_STALE_S = 3600.0
_CHILD_DELTA_EVENTS = {"subagent.thinking": "reasoning.delta", "subagent.text": "message.delta",
"subagent.start": "message.delta"}
def _child_run_active(child_key: str) -> bool:
@@ -35,15 +32,13 @@ def _mirror_subagent_to_child(event_type: str, payload: dict) -> None:
child_key = str(payload.get("child_session_id") or "")
if not child_key:
return
# Liveness registry first: accurate with no window open, so one opened mid-run
# immediately knows the child is busy.
# Liveness registry first: accurate with no window open (one opened mid-run knows busy).
if event_type == "subagent.complete":
_active_child_runs.pop(child_key, None)
else:
_active_child_runs[child_key] = time.time()
# Mirror only into a live watch session NOT upgraded to a full agent: an
# upgraded one owns a real native stream and mirroring would interleave two
# turns on one sid. Either way drop state so a reopened window starts fresh.
# Mirror only into a live watch session NOT upgraded to a full agent (an upgraded one owns
# a real native stream). Either way drop state so a reopened window starts fresh.
live = _find_live_session_by_key(child_key)
if live is None or live[1].get("agent") is not None:
with _child_mirrors_lock:
@@ -51,19 +46,16 @@ def _mirror_subagent_to_child(event_type: str, payload: dict) -> None:
return
csid = live[0]
text = str(payload.get("text") or "")
# thinking/text/start (the child's goal, as a one-time header) are plain deltas.
delta = {"subagent.thinking": "reasoning.delta", "subagent.text": "message.delta",
"subagent.start": "message.delta"}
with _child_mirrors_lock:
st = _child_mirrors.setdefault(child_key, {"seq": 0, "open_tool": None, "started": False})
if not st["started"]:
st["started"] = True
_emit("message.start", csid)
if event_type in delta:
# thinking/text/start (the child's goal, as a one-time header) are plain deltas.
if event_type in _CHILD_DELTA_EVENTS:
if text:
if event_type == "subagent.start":
text = f"{text}\n"
_emit(delta[event_type], csid, {"text": text})
_emit(_CHILD_DELTA_EVENTS[event_type], csid,
{"text": f"{text}\n" if event_type == "subagent.start" else text})
return
if event_type not in ("subagent.tool", "subagent.complete"):
return
@@ -85,11 +77,11 @@ def _mirror_subagent_to_child(event_type: str, payload: dict) -> None:
def _agent_cbs(sid: str) -> dict:
def _read_block(event: str, timeout: int):
# read_terminal / read_preview (desktop GUI): blocking bridge like clarify; the
# preview read gets longer since a URL tab extracts text from a live page.
# read_terminal / read_preview (desktop GUI): blocking bridge like clarify; the preview
# read gets longer since a URL tab extracts text from a live page.
return lambda start=None, count=None: _block(
event, sid, {k: v for k, v in (("start", start), ("count", count)) if v is not None}, timeout=timeout
)
event, sid, {k: v for k, v in (("start", start), ("count", count)) if v is not None},
timeout=timeout)
callbacks = {
"tool_start_callback": lambda tc_id, name, args: _on_tool_start(sid, tc_id, name, args),
@@ -98,49 +90,42 @@ def _agent_cbs(sid: str) -> dict:
sid, event_type, name, preview, args, **kwargs),
"tool_gen_callback": lambda name: _tool_progress_enabled(sid) and _emit("tool.generating", sid, {"name": name}),
"thinking_callback": lambda text: _emit("thinking.delta", sid, {"text": text}),
# Affection reaction (ily / <3 / good bot) → hearts; core-detected so TUI and desktop share it.
# Affection reaction (ily / <3 / good bot) → hearts; core-detected so TUI/desktop share it.
"reaction_callback": lambda kind: _emit("reaction", sid, {"kind": kind}),
"reasoning_callback": lambda text: _emit(
"reasoning.delta", sid, {"text": text, **({"verbose": True} if _session_verbose(sid) else {})}
),
"reasoning.delta", sid, {"text": text, **({"verbose": True} if _session_verbose(sid) else {})}),
"status_callback": lambda kind, text=None: _status_update(sid, str(kind), None if text is None else str(text)),
# Credits/notice spine: AgentNotice → notification.show; recovery clear → notification.clear.
# Credits/notice spine: AgentNotice → notification.show; recovery → notification.clear.
"notice_callback": lambda n: _emit(
"notification.show", sid,
{"text": n.text, "level": n.level, "kind": n.kind, "ttl_ms": n.ttl_ms, "key": n.key, "id": n.id},
),
{"text": n.text, "level": n.level, "kind": n.kind, "ttl_ms": n.ttl_ms, "key": n.key, "id": n.id}),
"notice_clear_callback": lambda key: _emit("notification.clear", sid, {"key": key}),
"clarify_callback": lambda q, c, multi_select=False, questions=None: (
_clarify_block(sid, q, c, multi_select=multi_select, questions=questions)),
"read_terminal_callback": _read_block("terminal.read.request", 30),
"read_preview_callback": _read_block("preview.read.request", 45),
# drive_preview / annotate_preview (desktop GUI): renderer drives the preview webview and
# answers with outcome + refreshed element inventory; same budget as the preview read it ends with.
# drive_preview / annotate_preview (desktop GUI): same budget as the preview read it ends with.
"drive_preview_callback": lambda payload: _block("preview.act.request", sid, dict(payload), timeout=45),
# read_window_below (desktop GUI): main process enumerates native windows.
"read_window_below_callback": lambda: _block("window.read.request", sid, {}, timeout=30),
# setup_mcp (desktop GUI): consent card + install/enable/OAuth. Long timeout on purpose (typing
# an API key, browser OAuth); like clarify, timeout returns "unanswered" and a late answer is tolerated.
# setup_mcp (desktop GUI): consent card + install/enable/OAuth; long timeout on purpose
# (typing an API key, browser OAuth) and, like clarify, a late answer is tolerated.
"setup_mcp_callback": lambda server, action, reason: _block(
"mcp.setup.request", sid, {"server": server, "action": action, "reason": reason}, timeout=600
),
"mcp.setup.request", sid, {"server": server, "action": action, "reason": reason}, timeout=600),
# tour (desktop GUI): renderer drives driver.js and answers tour.respond.
"tour_callback": lambda payload: _tour_request(sid, payload)}
# Interim assistant commentary (text alongside tool calls). Gated on
# display.interim_assistant_messages (default true); _run_prompt_submit overwrites
# it per turn and clears it in its finally so a stale closure can't fire.
# Interim assistant commentary (text alongside tool calls), gated on display.interim_assistant_
# messages; _run_prompt_submit overwrites it per turn and clears it so a stale closure can't fire.
if _load_interim_assistant_messages():
callbacks["interim_assistant_callback"] = lambda text, *, already_streamed=False: _emit(
"message.interim", sid, {"text": str(text), "already_streamed": bool(already_streamed)})
return callbacks
def _apply_project_workspace(task_id: str, path: str, _name: str = "") -> None:
"""Intentional workspace move from the project_* tools: re-anchor the live
session's cwd and push session.info so the desktop follows. This is the ONLY
auto-cwd path — driven by an explicit tool call, never a terminal `cd`."""
"""Intentional workspace move from the project_* tools: re-anchor the live session's cwd
and push session.info. The ONLY auto-cwd path — an explicit tool call, never a `cd`."""
if not path:
return
# task_id is the durable session_key; _sessions (and desktop event routing) key by sid.
@@ -149,24 +134,19 @@ def _apply_project_workspace(task_id: str, path: str, _name: str = "") -> None:
sid, session = (key, _sessions[key]) if key in _sessions else next(
((s, c) for s, c in _sessions.items()
if c.get("session_key") == key or getattr(c.get("agent"), "session_id", None) == key),
("", None),
)
if session is None:
return
("", None))
resolved = os.path.abspath(os.path.expanduser(str(path)))
if not os.path.isdir(resolved):
if session is None or not os.path.isdir(resolved):
return
session["cwd"] = resolved
session["explicit_cwd"] = True
session["cwd_from_settle"] = False # explicit switch supersedes a settle-adopted cwd
# explicit switch supersedes a settle-adopted cwd
session.update(cwd=resolved, explicit_cwd=True, cwd_from_settle=False)
_register_session_cwd(session)
_persist_session_cwd_and_schedule_git_meta(session, resolved)
try:
agent = session.get("agent")
info = _session_info(agent, session) if agent is not None else {
"cwd": resolved, "branch": _git_branch_for_cwd(resolved),
"project": _project_info_for_cwd(resolved), "lazy": True,
}
"project": _project_info_for_cwd(resolved), "lazy": True}
_emit("session.info", sid, info)
except Exception:
logger.debug("failed to emit session.info after project workspace move", exc_info=True)
@@ -176,19 +156,17 @@ def _wire_callbacks(sid: str):
from tools.terminal_tool import set_sudo_password_callback
from tools.skills_tool import set_secret_capture_callback
from tools.project_tools import set_project_workspace_callback
set_sudo_password_callback(lambda: _block("sudo.request", sid, {}, timeout=120))
set_project_workspace_callback(_apply_project_workspace)
def secret_cb(env_var, prompt, metadata=None):
pl = {"prompt": prompt, "env_var": env_var}
if metadata:
pl["metadata"] = metadata
pl = {"prompt": prompt, "env_var": env_var, **({"metadata": metadata} if metadata else {})}
val = _block("secret.request", sid, pl)
if not val:
return {"success": True, "stored_as": env_var, "validated": False, "skipped": True, "message": "skipped"}
from hermes_cli.config import save_env_value_secure
return {**save_env_value_secure(env_var, val), "skipped": False, "message": "ok"}
set_sudo_password_callback(lambda: _block("sudo.request", sid, {}, timeout=120))
set_project_workspace_callback(_apply_project_workspace)
set_secret_capture_callback(secret_cb)
@@ -199,12 +177,10 @@ def _available_personalities(cfg: dict | None = None) -> dict:
def _validate_personality(value: str, cfg: dict | None = None) -> tuple[str, str]:
"""Resolve a requested personality to (name, prompt) or raise ValueError. Same
contract as hermes_cli.personality.resolve_personality, but goes through the
module-level _available_personalities so tests keep a single patch point."""
"""(name, prompt) for a requested personality or ValueError; like resolve_personality but
via the module-level _available_personalities so tests keep a single patch point."""
from hermes_cli.personality import normalize_personality_name, render_personality_prompt
name = normalize_personality_name(value)
if not name:
if not (name := normalize_personality_name(value)):
return "", ""
personalities = _available_personalities(cfg)
if name not in personalities:
@@ -221,16 +197,13 @@ def _prompt_text(value) -> str:
def _apply_personality_to_session(
sid: str, session: dict, new_prompt: str, personality: str = "") -> tuple[bool, dict | None]:
"""Apply a personality change to a live session without resetting history: the
ephemeral system prompt is updated in place (appended at API-call time, so
prompt-cache hits survive) plus a pivot marker so the model stops pattern-matching
its earlier tone. Returns (history_reset=False, info)."""
"""Apply a personality change without resetting history: the ephemeral system prompt is
updated in place (appended at API-call time, so prompt-cache hits survive) plus a pivot
marker so the model stops pattern-matching its earlier tone. Returns (False, info)."""
if not session:
return False, None
session["personality"] = personality
agent = session.get("agent")
if not agent:
if not (agent := session.get("agent")):
return False, None
agent.ephemeral_system_prompt = new_prompt or None
marker = (
@@ -239,12 +212,10 @@ def _apply_personality_to_session(
f"accordingly: {new_prompt}]"
if new_prompt else
"[System: The user has cleared the personality overlay. "
"From this point forward, respond in your normal default style.]"
)
# Like the model-switch marker: role=user so strict providers accept it
# mid-conversation, but `display_kind` keeps it out of the
# `truncate_before_user_ordinal` addressing space (untagged, every rewind would
# land one turn early and `replace_messages` hard-delete the difference).
"From this point forward, respond in your normal default style.]")
# Like the model-switch marker: role=user so strict providers accept it mid-conversation,
# but `display_kind` keeps it out of the `truncate_before_user_ordinal` addressing space
# (untagged, every rewind would land one turn early and hard-delete the difference).
with session["history_lock"]:
session["history"].append({"role": "user", "content": marker, "display_kind": "personality_switch"})
session["history_version"] = int(session.get("history_version", 0)) + 1
@@ -256,8 +227,7 @@ def _apply_personality_to_session(
def _cfg_max_turns(cfg: dict, default: int) -> int:
from hermes_cli.config import resolve_turn_limit as _resolve_turn_limit
# Env override wins; resolve_turn_limit makes "none"/"unlimited"/0 first-class spellings.
env_val = os.environ.get("HERMES_TUI_MAX_TURNS")
if env_val:
if env_val := os.environ.get("HERMES_TUI_MAX_TURNS"):
return _resolve_turn_limit(env_val, default=default)
raw = (cfg.get("agent") or {}).get("max_turns")
if raw is None:
@@ -267,12 +237,7 @@ def _cfg_max_turns(cfg: dict, default: int) -> int:
def _parse_tui_skills_env() -> list[str]:
raw = os.environ.get("HERMES_TUI_SKILLS", "")
skills: list[str] = []
for part in raw.replace("\n", ",").split(","):
item = part.strip()
if item and item not in skills:
skills.append(item)
return skills
return list(dict.fromkeys(p.strip() for p in raw.replace("\n", ",").split(",") if p.strip()))
def _load_fallback_model():
@@ -282,40 +247,33 @@ def _load_fallback_model():
return get_fallback_chain(_load_cfg())
def _agent_fallback_model(agent):
"""Return an agent's fallback chain without rehydrating deliberately empty chains."""
if hasattr(agent, "_fallback_chain"):
return agent._fallback_chain or []
return agent._fallback_model if hasattr(agent, "_fallback_model") else _load_fallback_model()
def _background_agent_kwargs(agent, task_id: str) -> dict:
cfg = _load_cfg()
def g(name, default=None):
return getattr(agent, name, default)
kwargs = {k: g(k) or None for k in (
"base_url", "api_key", "provider", "api_mode", "acp_command", "acp_args",
"ephemeral_system_prompt")}
kwargs.update({k: g(k) for k in (
"providers_allowed", "providers_ignored", "providers_order", "provider_sort",
"provider_data_collection", "openrouter_min_coding_score")})
kwargs.update(
model=g("model") or _resolve_model(),
max_iterations=_cfg_max_turns(cfg, 25),
# Detached tasks declare platform="tui" (no UI sid for renderer-routed
# events), so resolve toolsets against it — never GUI schema they can't use.
enabled_toolsets=g("enabled_toolsets") or _load_enabled_toolsets("tui"),
quiet_mode=True, verbose_logging=False,
provider_require_parameters=g("provider_require_parameters", False),
session_id=task_id,
reasoning_config=g("reasoning_config") or _load_reasoning_config(str(g("model", "") or "")),
service_tier=g("service_tier") or _load_service_tier(),
request_overrides=dict(g("request_overrides", {}) or {}),
platform="tui", session_db=_get_db(), fallback_model=_agent_fallback_model(agent),
)
return kwargs
# Don't rehydrate a deliberately empty fallback chain.
if hasattr(agent, "_fallback_chain"):
fallback = agent._fallback_chain or []
else:
fallback = (agent._fallback_model if hasattr(agent, "_fallback_model")
else _load_fallback_model())
# Detached tasks declare platform="tui" (no UI sid for renderer-routed events), so resolve
# toolsets against it — never GUI schema they can't use.
return {
**{k: g(k) or None for k in ("base_url", "api_key", "provider", "api_mode", "acp_command",
"acp_args", "ephemeral_system_prompt")},
**{k: g(k) for k in ("providers_allowed", "providers_ignored", "providers_order", "provider_sort",
"provider_data_collection", "openrouter_min_coding_score")},
"model": g("model") or _resolve_model(), "max_iterations": _cfg_max_turns(cfg, 25),
"enabled_toolsets": g("enabled_toolsets") or _load_enabled_toolsets("tui"),
"quiet_mode": True, "verbose_logging": False,
"provider_require_parameters": g("provider_require_parameters", False), "session_id": task_id,
"reasoning_config": g("reasoning_config") or _load_reasoning_config(str(g("model", "") or "")),
"service_tier": g("service_tier") or _load_service_tier(),
"request_overrides": dict(g("request_overrides", {}) or {}),
"platform": "tui", "session_db": _get_db(), "fallback_model": fallback}
def _ephemeral_preview_agent_kwargs(agent, task_id: str) -> dict:
@@ -323,13 +281,10 @@ def _ephemeral_preview_agent_kwargs(agent, task_id: str) -> dict:
"enabled_toolsets": ["terminal", "file"], "session_db": None, "skip_memory": True}
_PREVIEW_HISTORY_ROLES = ("user", "assistant", "tool", "system")
def _preview_restart_history(session: dict, max_messages: int = 24, max_tool_chars: int = 1200) -> list[dict]:
"""Distill recent parent history for the ephemeral preview-restart agent (else it
guesses app/server/cwd/port from the bare URL). Keeps the last ``max_messages``
(always back to the last user turn); tool results truncated to ``max_tool_chars``."""
"""Distill recent parent history for the ephemeral preview-restart agent (else it guesses
app/cwd/port from the bare URL): last ``max_messages`` back to the last user turn, tool
results truncated to ``max_tool_chars``."""
try:
with session["history_lock"]:
history = list(session.get("history") or [])
@@ -337,14 +292,13 @@ def _preview_restart_history(session: dict, max_messages: int = 24, max_tool_cha
history = list(session.get("history") or [])
if not history:
return []
last_user = next((i for i in range(len(history) - 1, -1, -1) if history[i].get("role") == "user"), None)
start = max(0, len(history) - max_messages)
for idx in range(len(history) - 1, -1, -1):
if history[idx].get("role") == "user":
start = min(start, idx)
break
if last_user is not None:
start = min(start, last_user)
trimmed: list[dict] = []
for msg in history[start:]:
if not isinstance(msg, dict) or msg.get("role") not in _PREVIEW_HISTORY_ROLES:
if not isinstance(msg, dict) or msg.get("role") not in ("user", "assistant", "tool", "system"):
continue
copy = {k: v for k, v in msg.items() if k != "reasoning"}
content = copy.get("content")
@@ -358,7 +312,7 @@ def _preview_tool_result_preview(name: str, result: str) -> str:
try:
data = json.loads(result)
except Exception:
return ""
data = None
if not isinstance(data, dict):
return ""
if name == "terminal":
@@ -375,8 +329,7 @@ def _preview_restart_callbacks(parent: str, task_id: str) -> dict:
started_at: dict[str, float] = {}
def progress(message: str, level: str = "info") -> None:
text = str(message or "").strip()
if text:
if text := str(message or "").strip():
_emit("preview.restart.progress", parent, {"task_id": task_id, "level": level, "text": text})
def tool_start(tool_call_id: str, name: str, args: dict) -> None:
@@ -391,27 +344,22 @@ def _preview_restart_callbacks(parent: str, task_id: str) -> dict:
progress(summary + (f"\n{output}" if output else ""))
def tool_progress(event_type: str, name: str | None = None, preview: str | None = None, **_kwargs) -> None:
if preview:
progress(str(preview))
elif name:
progress(f"{event_type.replace('.', ' ')}: {name}")
if preview or name:
progress(str(preview) if preview else f"{event_type.replace('.', ' ')}: {name}")
return {
"tool_start_callback": tool_start, "tool_complete_callback": tool_complete,
"tool_progress_callback": tool_progress,
"tool_gen_callback": lambda name: progress(f"Preparing {name}"),
"status_callback": lambda kind, text=None: progress(text if text is not None else kind),
}
"status_callback": lambda kind, text=None: progress(text if text is not None else kind)}
def _reset_session_agent(sid: str, session: dict) -> dict:
tokens = _set_session_context(session["session_key"])
try:
# /new is a full conversation boundary: session-scoped runtime overrides
# (/model, /reasoning, /fast) do NOT carry forward — the fresh agent
# re-derives them from config.yaml, and the pins are cleared so a rebuild
# can't resurrect them. Global process state is never touched (see the
# cross-session-contamination note in _apply_model_switch).
# /new is a full conversation boundary: session-scoped runtime overrides (/model,
# /reasoning, /fast) do NOT carry forward and the pins are cleared so a rebuild can't
# resurrect them. Global process state is never touched (see _apply_model_switch).
for k in ("model_override", "create_reasoning_override", "create_service_tier_override", "one_turn_model_restore"):
session.pop(k, None)
new_agent = _make_agent(
@@ -425,8 +373,7 @@ def _reset_session_agent(sid: str, session: dict) -> dict:
queued_prompt=None,
_queued_prompt_generation=int(session.get("_queued_prompt_generation", 0)) + 1,
edit_snapshots={}, image_counter=0, running=False, show_reasoning=_load_show_reasoning(),
tool_progress_mode=_load_tool_progress_mode(), tool_started_at={},
)
tool_progress_mode=_load_tool_progress_mode(), tool_started_at={})
session.pop("queued_prompts", None)
with session["history_lock"]:
session["history"] = []