release: v0.3.0 (#462)

2026-09-11 21:45:42 +01:00
parent 0410b40f57
commit 418abca4f4
9 changed files with 556 additions and 509 deletions
@@ -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">v0.2.10</text>
+        font-size="13" font-weight="700" fill="#ffffff">v0.3.0</text>
 </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">v0.2.10</text>
+        font-size="13" font-weight="700" fill="#ffffff">v0.3.0</text>
 </svg>
@@ -61,12 +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.
    """
@@ -78,11 +78,11 @@ class EvoCodeInterpreterMiddleware(CodeInterpreterMiddleware):
        """Evict active REPLs on their worker loops before event-loop shutdown."""
        registry = self._registry
        with registry._lock:
-            thread_ids = tuple(registry._slots)
+            slot_ids = tuple(registry._slots)
-        for thread_id in thread_ids:
+        for slot_id in slot_ids:
            with contextlib.suppress(Exception):
-                await registry.aevict(thread_id)
+                await registry.aevict(slot_id)
-        self._ptc_tools_by_thread.clear()
+        self._ptc_tools_by_slot.clear()
 _live_interpreters = weakref.WeakSet()
@@ -10,7 +10,7 @@
 <a href="https://pypi.org/project/EvoScientist/"><picture>
  <source media="(prefers-color-scheme: light)" srcset="https://raw.githubusercontent.com/EvoScientist/EvoScientist/main/.github/assets/badge-pypi-light.svg">
  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/EvoScientist/EvoScientist/main/.github/assets/badge-pypi-dark.svg">
-  <img alt="PyPI v0.2.10" src="https://raw.githubusercontent.com/EvoScientist/EvoScientist/main/.github/assets/badge-pypi-light.svg" height="28">
+  <img alt="PyPI v0.3.0" src="https://raw.githubusercontent.com/EvoScientist/EvoScientist/main/.github/assets/badge-pypi-light.svg" height="28">
 </picture></a><a href="https://EvoScientist.github.io/"><picture>
  <source media="(prefers-color-scheme: light)" srcset="https://raw.githubusercontent.com/EvoScientist/EvoScientist/main/.github/assets/badge-website-light.svg">
  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/EvoScientist/EvoScientist/main/.github/assets/badge-website-dark.svg">
@@ -151,6 +151,7 @@ Moving beyond traditional human-in-the-loop systems, EvoScientist adopts a human
 <details>
 <summary>📦 Release Highlights — version changelog</summary>
 - **[11 Sep 2026]** **[v0.3.0](https://github.com/EvoScientist/EvoScientist/releases/tag/v0.3.0)** — Async sub-agents follow the caller's model; mid-session expert install without `/new`; MiniMax multi-turn fix; channel retry and OpenRouter attribution fixes.
 - **[05 Sep 2026]** **[v0.2.10](https://github.com/EvoScientist/EvoScientist/releases/tag/v0.2.10)** — New models: Claude Fable 5.1, GPT-6 Astra, Gemini 3.8 Flash, and Meta Muse Spark 1.3; proactivity phase 1: a first-contact profile bootstrap that eases cold start; optional acceptance checklists for scheduled tasks; tool selection only kicks in above 42 tools; deepagents 0.7.13.
 - **[29 Aug 2026]** **[v0.2.9](https://github.com/EvoScientist/EvoScientist/releases/tag/v0.2.9)** — New models: GLM-5.3-Flash (Zhipu + OpenRouter), Qwen3.8-Flash (DashScope + OpenRouter), and Tencent HY4 preview (OpenRouter), all with 1M context; deepagents 0.7.11.
 - **[21 Aug 2026]** **[v0.2.8](https://github.com/EvoScientist/EvoScientist/releases/tag/v0.2.8)** — Deploy-mode graph rebuilds drop from ~15 s to under a second; resume commands no longer crash and empty session history; Novita AI as a new LLM provider; NVIDIA BioNeMo Agent Toolkit in onboarding's recommended skill packs; bounded routed reasoning with empty truncations surfaced as errors; a subscription OAuth recipe in the docs; deepagents 0.7.8.
@@ -15,7 +15,7 @@
 <a href="https://pypi.org/project/EvoScientist/"><picture>
  <source media="(prefers-color-scheme: light)" srcset="https://raw.githubusercontent.com/EvoScientist/EvoScientist/main/.github/assets/badge-pypi-light.svg">
  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/EvoScientist/EvoScientist/main/.github/assets/badge-pypi-dark.svg">
-  <img alt="PyPI v0.2.10" src="https://raw.githubusercontent.com/EvoScientist/EvoScientist/main/.github/assets/badge-pypi-light.svg" height="28">
+  <img alt="PyPI v0.3.0" src="https://raw.githubusercontent.com/EvoScientist/EvoScientist/main/.github/assets/badge-pypi-light.svg" height="28">
 </picture></a><a href="https://EvoScientist.github.io/"><picture>
  <source media="(prefers-color-scheme: light)" srcset="https://raw.githubusercontent.com/EvoScientist/EvoScientist/main/.github/assets/badge-website-light.svg">
  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/EvoScientist/EvoScientist/main/.github/assets/badge-website-dark.svg">
@@ -160,6 +160,7 @@ EvoScientist 超越了传统的人在回路（Human-in-the-Loop）模式，采
 <details>
 <summary>📦 版本更新摘要（changelog）</summary>
 - **[2026 年 9 月 11 日]** **[v0.3.0](https://github.com/EvoScientist/EvoScientist/releases/tag/v0.3.0)** — 异步子智能体跟随调用方模型；expert 安装即用，无需 `/new`；修复 MiniMax 多轮崩溃、渠道重试与 OpenRouter 归因问题。
 - **[2026 年 9 月 5 日]** **[v0.2.10](https://github.com/EvoScientist/EvoScientist/releases/tag/v0.2.10)** — 新增模型：Claude Fable 5.1、GPT-6 Astra、Gemini 3.8 Flash、Meta Muse Spark 1.3；主动性第一阶段：首次对话建立用户档案，缓解冷启动；定时任务可附加验收清单；工具选择器仅在工具超过 42 个时启用；升级 deepagents 0.7.13。
 - **[2026 年 8 月 29 日]** **[v0.2.9](https://github.com/EvoScientist/EvoScientist/releases/tag/v0.2.9)** — 新增模型：GLM-5.3-Flash（智谱与 OpenRouter）、Qwen3.8-Flash（DashScope 与 OpenRouter）、腾讯 HY4 preview（OpenRouter），均为 1M 上下文；升级 deepagents 0.7.11。
 - **[2026 年 8 月 21 日]** **[v0.2.8](https://github.com/EvoScientist/EvoScientist/releases/tag/v0.2.8)** — deploy 模式图重建从约 15 秒降至 1 秒内；resume 命令不再崩溃清空会话历史；新增 Novita AI provider；onboard 推荐技能包新增 NVIDIA BioNeMo Agent Toolkit；路由 reasoning 参数收敛、空截断显式报错；文档新增订阅 OAuth 配置指南；升级 deepagents 0.7.8。
@@ -1,6 +1,6 @@
 [project]
 name = "EvoScientist"
-version = "0.2.10"
+version = "0.3.0"
 description = "EvoScientist: Towards Self-Evolving AI Scientists for End-to-End Scientific Discovery"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -17,6 +17,7 @@ classifiers = [
 ]
 dependencies = [
    "deepagents[quickjs]~=0.7.6",
    "langchain-quickjs>=0.3.7",
    "langchain>=1.3",
    "langchain-anthropic>=1.5",
    "langchain-openai>=1.2",
@@ -104,11 +104,11 @@ def test_after_agent_evicts_slot_on_untouched_turn():
    Upstream ``after_agent`` in ``langchain_quickjs/middleware.py`` performs
    two things: snapshot the REPL AND evict the slot (``finally:
-    self._registry.evict(thread_id)``). ``before_agent`` restores the REPL
+    self._registry.evict(slot_id)``). ``before_agent`` restores the REPL
    on any turn that follows a touched one via ``self._registry.get`` —
    which is get-or-create. So if ``after_agent`` returns early without
    evicting, one ``ThreadWorker`` + QuickJS Runtime leaks per persistent
-    ``thread_id`` that ever went touched → quiet.
+    slot that ever went touched → quiet.
    Fix: don't override ``after_agent`` / ``aafter_agent`` at all — inherit
    upstream's unconditional snapshot+evict behavior. This test creates a
@@ -116,17 +116,18 @@ def test_after_agent_evicts_slot_on_untouched_turn():
    untouched-state input, and asserts the slot was evicted.
    """
    mw = create_code_interpreter_middleware()
-    tid = mw._fallback_thread_id
+    slot_id = mw._slot_update_for_runtime()["_quickjs_slot_id"]
    # Simulate the slot creation that ``before_agent`` performs when it sees
    # a prior turn's snapshot payload in state.
-    mw._registry.get(tid)
+    mw._registry.get(slot_id)
    assert len(mw._registry._slots) == 1
    # Untouched-turn state: no ``code_interpreter`` tool call between the
    # last ``HumanMessage`` and end. Under the earlier buggy gate this
    # returned ``{}`` without evicting — leaking the slot created above.
    untouched_state = {
        "_quickjs_slot_id": slot_id,
        "_quickjs_snapshot_payload": b"payload-from-prior-turn",
        "messages": [
            HumanMessage(content="thanks"),