foxclaw 0.1.0 → 0.1.3

package/python/tests/run_all.py

@@ -0,0 +1,200 @@
+"""
+Foxclaw — 全量回归测试运行器
+一次跑完 tests/ 下所有测试,汇总通过/失败,确认整个系统端到端健康。
+用法:python tests/run_all.py
+"""
+import os
+import shutil
+import subprocess
+import sys
+import tempfile
+import time
+from pathlib import Path
+
+for _s in (sys.stdout, sys.stderr):
+    try:
+        _s.reconfigure(encoding="utf-8", line_buffering=True)
+    except Exception:
+        pass
+
+TESTS = [
+    ("test_e2e.py",          "认知运行时端到端"),
+    ("test_api.py",          "API 端点"),
+    ("test_integration.py",  "A · 前后端实时链路"),
+    ("test_dispatch.py",     "D · 主管调度 + 共享记忆"),
+    ("test_finance.py",      "B · 财务内核"),
+    ("test_learning.py",     "B · 学习内核"),
+    ("test_itinerary.py",    "B · 行程内核"),
+    ("test_orchestration.py","主管编排 + 跨 Agent"),
+    ("test_skills.py",       "助手 Skill 接入层"),
+    ("test_skills_fetch.py", "技能安装(ClawHub/GitHub/URL/本地包)"),
+    ("test_kernel.py",       "内核新模块回归"),
+    ("test_governance.py",   "运行时治理(幂等/恢复/队列/压缩/循环/钩子)"),
+    ("test_compaction_toolpair.py", "M5·压缩器工具消息配对(原子组切分+畸形输入安全降级)"),
+    ("test_tool_params.py",  "M4·工具入参校验(执行侧 schema/Stage1/解析失败可见)"),
+    ("test_shared_write.py", "P0-3 · 共享层写入策略(audit/warn/enforce)"),
+    ("test_audit_sink.py",   "P0-3 · 审计落盘 sink(环形缓冲外的持久样本)"),
+    ("test_governance_audit.py", "批次3 · 治理审计链(P1-2/P1-3/P2-0/P3-4 decision_id)"),
+    ("test_shadow_learning.py", "批次4 · P0-1 影子学习(shadow→promote/discard+回滚+边界)"),
+    ("test_channels.py",     "消息渠道(四适配器+闭环+幂等去重)"),
+    ("test_evals_smoke.py",  "M1 · 治理评测冒烟(拦截率/误报/收敛)"),
+    ("test_parallel_dispatch.py", "M2 · 并行子 Agent(fork-join/隔离/并发安全)"),
+    ("test_governance_middleware.py", "M3 · 治理中间件(框架无关/装饰器/审计)"),
+    ("test_memory_recall.py", "M4 · 记忆隔离召回完整性(根治)"),
+    ("test_toplevel_concurrency.py", "顶层任务并发隔离(contextvars)"),
+    ("test_api_auth.py",     "API 可选鉴权(X-API-Key)"),
+    ("test_supervisor_triage.py", "主管分诊与自答策略(关键词直路 / 自答 / stub 兜底)"),
+    ("test_supervisor_pipeline.py", "批次5 · P1-1 主管 pipeline 化(八 stage/统一 emit/流式/恢复记账)"),
+    ("test_ahs_p2.py",       "批次6 · P2层(compose pending/run_async/idempotency日志/recovery计划)+P3-2"),
+    ("test_p3_p5_hardening.py", "批次7 · P3收尾(isinstance/常量时间/exact/blocked短路/system归因)+P5-1/P5-3"),
+    ("test_browser.py",      "浏览器子系统(HIR-Browser:核心逻辑/线程包装/降级)"),
+    ("test_hkc.py",          "知识核心(HKC融入:装配/检索/H1资源修复/惰性teardown)"),
+    ("test_hkc_manifest_reuse.py",        "P1·Stage A — manifest v2 复用(roundtrip/8步校验/bm25 0-corpus)"),
+    ("test_hkc_search_event_sync.py",     "P1·Stage B — 事件→索引 C/U/D 五分支同步(text_hash 跳过/立即落盘)"),
+    ("test_hkc_index_build_manager.py",   "P1·Stage C — 后台冷构建(状态机/检查点/锁/batch/取消/重启续传)"),
+    ("test_search.py",       "认知搜索(HSR融入:编排/证据融合/迭代上限/端口降级)"),
+    ("test_embedding_singleton.py", "P0·EmbeddingService单例(同名模型进程内只加载一次)"),
+    ("test_event_spine.py",  "P0·统一事件脊柱(EventSpine/SystemBus门面/harness转发/统一投影)"),
+    ("test_agent_lifecycle.py", "P1·任务生命周期收敛(5 run_task→1 模板 + _execute 钩子)"),
+    ("test_planner.py",      "P1-2·规划原语统一(主管/AHS/search 三处委托一个引擎)"),
+    ("test_capabilities.py", "P1-3·能力中枢(Composition Root/打断两处对象环)"),
+    ("test_truth_flow.py",   "P2·真值流全路径触发(直连任务也进真值链 + _no_emit 防双发)"),
+    ("test_orchestration_decoupled.py", "P3·编排层去领域硬编码(通用聚合/路由,领域归属Agent)"),
+    ("test_pipeline.py",     "v2②·认知管线(perceive/recall/plan/route/act/verify/reflect)"),
+    ("test_scheduler.py",    "v2①·微内核调度器(并发配额/背压/生命周期/记账)+Lifecycle统一"),
+    ("test_policy_gate.py",  "v2①Step4·统一准入门(Harness+skill-guard→一套Verdict+审计)"),
+    ("test_truth_crystallizer.py", "v2④Step1·真值走HKC/KEE(默认关双写+后台结晶+订阅接线)"),
+    ("test_reasoner.py",     "v2③Step1·领域认知化(观察→假设→证据→裁决;财务守门人形式化)"),
+    ("test_domain_reasoners.py", "v2③Step4·认知化推广(itinerary冲突/learning复习;规则权威+LLM只增强)"),
+    ("test_tool_center.py",  "工具中心①·技能台真实化(安装/启停/删除/测试/持久化/SSRF审核门/治理)"),
+    ("test_mcp_connectors.py", "工具中心②·MCP 连接器(接入/工具成技能/审核/调用/移除/持久化)"),
+    ("test_markdown_skills.py", "工具中心③·SKILL.md 指令技能(上传/解析/审核/注入/持久化)"),
+    ("test_markdown_skills_openclaw.py", "工具中心③·OpenClaw 兼容(zip包/描述触发/编辑触发词·进程隔离防原生库崩溃)"),
+    ("test_mcp_server.py",   "内置 MCP 服务器(Foxclaw 作提供者:只读知识+记忆暴露成 MCP 工具)"),
+    ("test_mcp_no_memory.py", "MCP ask_foxclaw 不污染记忆(外部任务不写 HMR,正常任务照写)"),
+    ("test_streaming.py",    "LLM 流式输出(router.stream 分块 + /api/chat/stream SSE 逐字)"),
+    ("test_task_stream.py",  "D1 · /api/task/stream 任务流式化(逐 stage 进度 + act token + 结果)"),
+    ("test_config_layering.py", "E1 · 分层配置(LayeredConfig 三层 + policy 层 deny 不可覆盖)"),
+    ("test_semantic.py",     "真语义检索 + KEE 语义化(共享ST单例/余弦/矛盾方向;无缓存优雅跳过)"),
+    ("test_metrics.py",      "ops 监控(请求/LLM 延迟分位+错误率 · 深度健康 · Prometheus)"),
+    ("test_routing.py",      "任务感知模型路由(fast/smart 分档 + 跨 provider fallback 链)"),
+    ("test_file_ingest.py",  "文件摄入打通交互层(docx/xlsx 加载器 + KDE 分派 + 上传格式门)"),
+    ("test_multimodal.py",   "多模态对话(图片随问题发视觉模型 · 按 provider 组装 · 护栏 · stub 降级)"),
+    ("test_attachments.py",  "多模态渲染·聊天附件(上传/下载/正文提取/护栏/LLM 注入/会话落盘)"),
+    ("test_gateway.py",      "统一 API 网关(per-client 令牌桶+配额+空闲淘汰+429/Retry-After+启用哲学)"),
+    ("test_kee_semantic.py", "KEE 身份/去重(别名确定性匹配·类型隔离·Fact 完整 statement 守 exact)"),
+    ("test_domain_import.py", "OS 领域导入超越 csv(finance/learning/itinerary 上传 csv/xlsx/json)"),
+    ("test_tracing.py",      "请求链路追踪(trace_id 贯穿 请求→Agent→LLM·跨线程 span·有界环存)"),
+    ("test_emb_cache.py",    "检索/embedding 共享向量缓存(批量只编未命中·跨层共享·LRU 有界·并发)"),
+    ("test_multi_chain.py",  "智能体层·多步跨 Agent 链式协作(后一步看得到前序结果·上下文封顶)"),
+    ("test_web_trigger.py",  "联网触发判定(个人数据不误搜·外部话题照搜·纯时间词不触发)"),
+    ("test_evidence_semantic.py", "证据融合语义化(否定词边界·幻影矛盾·语义支持边·降级)"),
+    ("test_domain_evals.py", "领域推理评测(漏报/误报率+误报归因到规则+规则实例隔离)"),
+    ("test_negotiation.py",  "智能体层·跨域协商闭环(被拒→让步→再裁决·有界·拒绝不降级)"),
+    ("test_tool_use.py",        "P2·function-calling 工具化(多轮工具循环/注册表/降级)"),
+    ("test_injection_guard.py", "提示词注入纵深防御(检出/wrap_untrusted/内核前缀/隔离)"),
+    ("test_sessions.py",        "会话存储(增删改查/标题提炼/路径穿越/并发追加)"),
+    ("test_stats_and_events.py","记忆统计 + 活动事件(hmr.stats 计数修复/chat.completed)"),
+    ("test_llm_planner.py",            "P0-1·LLM 驱动规划器(llm_plan/replan·opt-in+fail-safe+sv_replan)"),
+    ("test_anthropic_tools.py",        "P0-2·Anthropic tool_use 原生接入(格式转换/解析/回灌/降级)"),
+    ("test_taint.py",                  "数据污点追踪+动态审批门槛(污点分级/传播/TAINT_CONFIRM)"),
+    ("test_mcp_guardian_discovery.py", "MCP 发现进守门人(凭证不出主进程·代发现·序列化)"),
+    ("test_verify_evidence.py",        "三个薄点·失败策略表/执行后核验/执行证据(重试只限只读)"),
+    ("test_evidence_sse.py",           "执行凭据 SSE 命名事件(与正文分离·纯对话不发)"),
+    ("test_profile.py",                "称呼(user profile)存取:写入/替换/清空/不误删偏好"),
+    ("test_update_check.py",           "新版本检查:默认关/版本比较/24h 缓存/失败静默"),
+]
+
+
+def main():
+    here = Path(__file__).parent
+    _argv = sys.argv[1:]
+    global TESTS
+    if "--list" in _argv:
+        print(f"共 {len(TESTS)} 项:")
+        for _f, _d in TESTS:
+            print(f"  {_f:<40} {_d}")
+        sys.exit(0)
+    _only = None
+    if "--only" in _argv:
+        _i = _argv.index("--only")
+        _only = _argv[_i + 1] if _i + 1 < len(_argv) else None
+        if not _only:
+            print("--only 需要一个子串参数,例:python tests/run_all.py --only governance")
+            sys.exit(2)
+        _before = len(TESTS)
+        TESTS = [(f, d) for f, d in TESTS if _only in f or _only in d]
+        print(f"[filter] '{_only}' 匹配 {len(TESTS)}/{_before} 项\n")
+    env = dict(os.environ, FOXCLAW_SHADOW_ROUNDS="50",
+               HF_HUB_OFFLINE="1", TRANSFORMERS_OFFLINE="1",
+               PYTHONUTF8="1", PYTHONIOENCODING="utf-8",
+               FOXCLAW_SKIP_DOTENV="1",
+               DEEPSEEK_API_KEY="", ANTHROPIC_API_KEY="", OPENAI_API_KEY="",
+               GOOGLE_API_KEY="", DASHSCOPE_API_KEY="", FOXCLAW_API_KEY="",
+               FOXCLAW_MCP_SERVER="0",
+               FOXCLAW_PROVIDER="", FOXCLAW_MODEL="",
+               HKC_EMBEDDING="stub")
+    print("=" * 64)
+    print("Foxclaw · 全量回归测试")
+    print("=" * 64)
+    results = []
+    try:
+        import pytest  # noqa: F401
+        have_pytest = True
+    except Exception:
+        have_pytest = False
+    if not have_pytest:
+        print("⚠ 未安装 pytest —— 依赖它的测试将标记为「跳过」。"
+              "完整回归请先 `pip install -e \".[dev]\"` 或 `pip install pytest`。\n")
+    for fn, desc in TESTS:
+        if not have_pytest:
+            results.append((desc, "跳过", 0.0))
+            print(f"  ○ {desc:<28} [跳过:需 pytest]")
+            continue
+        path = here / fn
+        if not path.exists():
+            results.append((desc, "缺失", 0)); continue
+        t0 = time.time()
+        _bt = Path(tempfile.mkdtemp(prefix="foxclaw-pt-"))
+        try:
+            _cmd = [sys.executable, "-m", "pytest", str(path), "-q",
+                    f"--basetemp={_bt}"]
+            r = subprocess.run(_cmd,
+                               capture_output=True, text=True, timeout=300, env=env,
+                               encoding="utf-8", errors="replace")
+            ok = r.returncode == 0
+            if "Traceback" in (r.stdout + r.stderr):
+                ok = False
+            _ACCESS_VIOLATION = (3221225477, -1073741819)
+            if not ok and r.returncode in _ACCESS_VIOLATION and \
+                    "Traceback" not in (r.stdout + r.stderr):
+                r = subprocess.run(_cmd,
+                                   capture_output=True, text=True, timeout=300, env=env,
+                                   encoding="utf-8", errors="replace")
+                ok = r.returncode == 0 and "Traceback" not in (r.stdout + r.stderr)
+                if ok:
+                    print(f"  ↻ {desc} 首次 0xC0000005 原生崩溃,重试通过(已知环境性偶发)")
+            results.append((desc, "通过" if ok else "失败", time.time() - t0))
+            if not ok:
+                tail = (r.stdout + r.stderr).strip().splitlines()[-15:]
+                print(f"  ✗ {desc} 失败:")
+                for line in tail: print(f"      {line}")
+                (here / f"_fail_{fn}.txt").write_text(
+                    r.stdout + r.stderr, encoding="utf-8")
+        except subprocess.TimeoutExpired:
+            results.append((desc, "超时", 300))
+        finally:
+            shutil.rmtree(_bt, ignore_errors=True)
+        print(f"  {'✓' if results[-1][1]=='通过' else '✗'} {desc:<28} [{results[-1][1]}] {results[-1][2]:.0f}s")
+
+    print("\n" + "=" * 64)
+    passed = sum(1 for _, s, _ in results if s == "通过")
+    skipped = sum(1 for _, s, _ in results if s == "跳过")
+    total = len(results) - skipped
+    suffix = f"（另有 {skipped} 项因缺 pytest 跳过）" if skipped else ""
+    print(f"结果:{passed}/{total} 通过{suffix}")
+    print("=" * 64)
+    return 0 if passed == total else 1
+
+if __name__ == "__main__":
+    sys.exit(main())

package/python/tests/test_agent_lifecycle.py

@@ -0,0 +1,326 @@
+"""
+统一任务生命周期(P1:5 个 run_task 收敛为 1)守卫测试 —— 全离线。
+
+覆盖:
+  1. 结构收敛:4 个领域 Agent 复用 base 的模板 run_task(不各自重写),各自实现 _execute;
+     SupervisorAgent 作为路由器合法地保留自己的 run_task。
+  2. 统一异常兜底:_execute 抛 HarnessBlockedError → 模板统一返回结构化错误(一处定义,全 Agent 生效)。
+  3. 领域钩子生效:各 Agent 的 entities/recall 前缀/召回条数按覆盖生效(行为逐字保留)。
+"""
+import os, sys, asyncio, tempfile, shutil, types
+from pathlib import Path
+os.environ.setdefault("HF_HUB_OFFLINE", "1")
+os.environ.setdefault("TRANSFORMERS_OFFLINE", "1")
+os.environ["FOXCLAW_EMBEDDING_MODEL"] = "foxclaw-nonexistent-fast"
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+from foxclaw.agents.base import CognitiveAgent, HarnessBlockedError
+from foxclaw.agents.agents import (
+    FinanceAgent, OfficeAgent, SupervisorAgent)
+
+
+def test_structural_convergence():
+    print("\n【1】结构收敛:全部 Agent 复用模板 run_task(主管经 P1-1 pipeline 化收敛)")
+    for A in (FinanceAgent, OfficeAgent):
+        assert A.run_task is CognitiveAgent.run_task, f"{A.__name__} 不应重写 run_task"
+        assert A._execute is not CognitiveAgent._execute, f"{A.__name__} 应实现 _execute"
+    assert SupervisorAgent.run_task is CognitiveAgent.run_task, \
+        "P1-1:主管不再重写 run_task(路由分发在 pipeline stage 里)"
+    assert SupervisorAgent._cognitive_pipeline is not CognitiveAgent._cognitive_pipeline, \
+        "主管应覆盖 _cognitive_pipeline 使用 SUPERVISOR_STAGES"
+    print("  ✓ 5 个 Agent 共用唯一 run_task 模板;主管差异收敛为 pipeline 覆盖(P1-1)")
+
+
+def test_domain_hooks():
+    print("\n【2】领域钩子:entities / recall 前缀 / 召回条数 逐字保留")
+    assert FinanceAgent._task_entities(None, "x") == ["财务", "收支", "储蓄"]
+    assert FinanceAgent._recall_query(None, "买车") == "财务 买车"
+    assert OfficeAgent._recall_query(None, "开会") == "学习 知识 日程 行程 开会"
+    assert OfficeAgent._task_entities(None, "x") == ["学习", "知识", "文档", "日程", "行程", "会议"]
+    assert OfficeAgent._recall_topk == 8, "办公学习 Agent 覆盖为广泛召回 8"
+    assert CognitiveAgent._recall_topk == 5, "基类默认仍是 5"
+    print("  ✓ finance/admin/office 的 entities、recall 前缀、top_k 均按覆盖生效")
+
+
+def test_unified_block_handling():
+    print("\n【3】统一异常兜底:_execute 抛 HarnessBlockedError → 结构化错误")
+    d = tempfile.mkdtemp(prefix="life_")
+    try:
+        from foxclaw import Foxclaw
+        os_ = Foxclaw(data_dir=d, shadow_rounds=50)
+
+        async def boom(self, task, context, session_id, memories):
+            raise HarnessBlockedError(types.SimpleNamespace(reason="模拟越界"), None)
+
+        os_.finance._execute = types.MethodType(boom, os_.finance)
+        r = asyncio.run(os_.finance.run_task("危险任务"))
+        assert r.get("error") == "harness_blocked", f"应兜底为 harness_blocked,实得 {r}"
+        assert r.get("reason") == "模拟越界" and "session_id" in r
+        print(f"  ✓ 模板统一兜底: {r}")
+        asyncio.run(os_.aclose())
+    finally:
+        shutil.rmtree(d, ignore_errors=True)
+
+
+def test_finance_verify_before_recommend():
+    """【4·P5】记忆验证优先:财务事实以**当前账本(live analyze)**为准,召回记忆仅作历史上下文。
+    锁定 finance._execute 的 verify-before-recommend 设计——即便召回了**过时**记忆,权威事实源
+    仍是当前交易(`fin.analyze()`),记忆被显式降级为"历史财务记录"。防未来重构把过时记忆当事实源
+    (呼应范式"记忆是快照,非事实")。评估结论见方案 P5:该设计已存在,本测试是把它锁死的回归护栏。"""
+    print("\n【4】P5 记忆验证优先:财务事实以当前账本为准、记忆仅历史上下文")
+    d = tempfile.mkdtemp(prefix="p5_")
+    os_ = None
+    try:
+        from foxclaw import Foxclaw
+        os_ = Foxclaw(data_dir=d, shadow_rounds=50)
+        fin = os_.finance
+        os_.finance_runtime.import_csv(
+            str(Path(__file__).parent.parent / "examples" / "sample_transactions.csv"))
+        captured = {}
+        async def _capture(self, task, messages):
+            captured["prompt"] = messages[-1]["content"]
+            return "（模拟回答）"
+        fin._call_llm_web = types.MethodType(_capture, fin)
+        stale = [{"content": "历史:上月预算结余 8000 元", "id": "m1"}]
+        asyncio.run(fin._execute("这个月还能花多少?", {}, "s1", stale))
+        p = captured["prompt"]
+        assert "基于真实交易,请据此回答" in p, "缺 live analyze() 权威事实源(verify-before-recommend 退化)"
+        assert "历史财务记录" in p and "8000" in p, "召回记忆应作历史上下文出现,而非当前事实"
+        assert p.index("历史财务记录") < p.index("基于真实交易"), \
+            "结构应为:历史记忆在前、当前账本权威在后(措辞上 live『请据此回答』压过历史记忆)"
+        print("  ✓ live analyze() 作权威源『请据此回答』;过时记忆降级为『历史财务记录』(P5 已锁定)")
+    finally:
+        if os_ is not None:
+            asyncio.run(os_.aclose())
+        shutil.rmtree(d, ignore_errors=True)
+
+
+def test_run_task_stream():
+    """【5·D1】run_task 流式化:同步 generator 逐 stage 进度 + act 逐 token + 最终结果,且与非流式
+    run_task **行为一致**(共享 `_run_task_core`)。关键验证:`_act_stream` contextvar 真的跨
+    worker 线程 / asyncio 边界传到 `_call_llm`(否则一个 delta token 都不会有)。"""
+    print("\n【5】D1 run_task 流式化:stage 进度 + act token + 结果")
+    d = tempfile.mkdtemp(prefix="d1_")
+    os_ = None
+    try:
+        from foxclaw import Foxclaw
+        os_ = Foxclaw(data_dir=d, shadow_rounds=50)
+        ag = os_.office
+        ag.router.provider = "stub"
+
+        events = list(ag.run_task_stream("帮我总结一下今天"))
+        stages = [e["stage"] for e in events if "stage" in e]
+        deltas = [e["delta"] for e in events if "delta" in e]
+        results = [e["result"] for e in events if "result" in e]
+        assert stages == ["perceive", "recall", "plan", "route", "act", "verify", "reflect"], stages
+        assert len(deltas) > 0, "act 应逐 token 出字;为空说明 contextvar 没传到 _call_llm"
+        assert len(results) == 1, f"应恰有一个最终 result,实得 {len(results)}"
+        res = results[0]
+        assert res.get("agent") == ag.AGENT_ID and "session_id" in res and "result" in res
+        assert "".join(deltas) == res["result"], "流式 token 拼接应等于最终答案"
+
+        plain = asyncio.run(ag.run_task("帮我总结一下今天"))
+        assert "result" in plain and "delta" not in plain and plain.get("agent") == ag.AGENT_ID
+
+        import threading as _th
+        ev = _th.Event(); ev.set()
+        ev_events = list(ag.run_task_stream("这个任务会被中断", stop_event=ev))
+        ev_stages = [e for e in ev_events if "stage" in e]
+        ev_results = [e["result"] for e in ev_events if "result" in e]
+        assert ev_stages == [], f"预置 stop → 0 个 stage 事件,实得 {ev_stages}"
+        assert len(ev_results) == 1 and ev_results[0].get("interrupted") is True, ev_results
+
+        def _boom(*a, **k):
+            raise RuntimeError("流式炸了")
+        ag.router.stream = _boom
+        from foxclaw.agents.base import _act_stream as _AS
+        async def _one_call():
+            tok = _AS.set({"on_token": lambda c: None, "stop_event": None})
+            try:
+                return await ag._call_llm([{"role": "user", "content": "x"}])
+            finally:
+                _AS.reset(tok)
+        out = asyncio.run(_one_call())
+        assert isinstance(out, str) and out, "流式失败应 fail-safe 降级 complete(返回非空串),而非崩"
+
+        print(f"  ✓ stage={len(stages)} delta={len(deltas)} result=1;token 拼接==答案;"
+              f"非流式不受影响;预置 stop 起步即停;流式异常 fail-safe 降级 complete")
+    finally:
+        if os_ is not None:
+            asyncio.run(os_.aclose())
+        shutil.rmtree(d, ignore_errors=True)
+
+
+def test_interrupted_task_not_sunk():
+    """【6·B3 深审 G2】被 stop_event 中断的任务 = 未完成,绝不沉淀为已完成真值(blocked=True)。
+    深审发现:_emit_execution 的 blocked 原只看 skill_status,中断任务(skill_status=None)会
+    blocked=False → 部分/空结果被 TruthPipeline 当完成写进 HMR 共享记忆(违反 P5/Principle07)。
+    修复后 blocked 联动 ctx.interrupted。这条路径由 B3 的 /api/task/stream(断连 set stop_event)
+    首次在生产可达,故 happy-path 测不到、深审才暴露。"""
+    print("\n【6】B3 深审 G2:中断任务不沉淀真值(blocked=True)")
+    d = tempfile.mkdtemp(prefix="g2_")
+    os_ = None
+    try:
+        import threading as _th
+        from foxclaw import Foxclaw
+        os_ = Foxclaw(data_dir=d, shadow_rounds=50)
+        ag = os_.supervisor
+        ag.router.provider = "stub"
+        captured = {}
+        orig = ag._emit_execution
+        def _spy(task, result, blocked, context):
+            captured["blocked"] = blocked
+            return orig(task, result, blocked=blocked, context=context)
+        ag._emit_execution = _spy
+        ev = _th.Event(); ev.set()
+        r = asyncio.run(ag.run_task("会被中断的任务", stop_event=ev))
+        assert r.get("interrupted") is True, r
+        assert captured.get("blocked") is True, "中断任务必须 blocked=True(不把部分/空结果当完成真值沉淀)"
+        captured.clear()
+        r2 = asyncio.run(ag.run_task("正常任务"))
+        assert "interrupted" not in r2 and captured.get("blocked") is False, \
+            "正常完成任务 blocked=False(行为逐字不变,照常沉淀真值)"
+        print("  ✓ 中断→blocked=True(不沉淀);正常→blocked=False(照常沉淀,行为不变)")
+    finally:
+        if os_ is not None:
+            asyncio.run(os_.aclose())
+        shutil.rmtree(d, ignore_errors=True)
+
+
+def test_reasoning_lifecycle_hooks():
+    """【7·E2】认知管线生命周期钩子:pre_reasoning(agent_start)可注入推理上下文(前插进 LLM);
+    task_complete(agent_end)只读观测;默认无钩子=行为逐字不变。这两个语义名是此前从未 fire 的
+    孤儿 agent_start/agent_end 的别名——E2 顺带复活了它们并给认知管线 run_task 装上钩子。"""
+    print("\n【7】E2 生命周期钩子:pre_reasoning 注入 + task_complete 观测")
+    d = tempfile.mkdtemp(prefix="e2_")
+    os_ = None
+    try:
+        from foxclaw import Foxclaw
+        os_ = Foxclaw(data_dir=d, shadow_rounds=50)
+        ag = os_.office
+        ag.router.provider = "stub"
+        seen = {}
+        def _cap_complete(system, messages, max_tokens, agent_id):
+            seen["messages"] = messages
+            return "（stub 回答）"
+        ag.router.complete = _cap_complete
+
+        asyncio.run(ag.run_task("基线任务"))
+        base_msgs = seen.get("messages", [])
+        assert base_msgs and "【E2注入】" not in base_msgs[0].get("content", ""), \
+            f"无 pre_reasoning 钩子时不应注入,实得 {base_msgs[:1]}"
+
+        completed = []
+        ag.hooks.register("pre_reasoning", lambda c: {"additional_context": "【E2注入】优先考虑省钱"})
+        ag.hooks.register("task_complete", lambda c: completed.append(c.get("result")))
+        asyncio.run(ag.run_task("帮我看看要不要买"))
+        msgs = seen.get("messages", [])
+        joined = " ".join(m.get("content", "") for m in msgs)
+        assert "【E2注入】优先考虑省钱" in joined, f"pre_reasoning 注入应进入 LLM messages,实得 {msgs}"
+        roles = [m.get("role") for m in msgs]
+        assert not any(roles[i] == roles[i + 1] == "user" for i in range(len(roles) - 1)), \
+            f"注入不应产生连续两条 user 消息(违反模型无关铁律),roles={roles}"
+        assert len(completed) == 1, f"task_complete 应触发一次并收到 result,实得 {len(completed)}"
+
+        import types
+        async def _boom(self, task, context, session_id, memories):
+            raise HarnessBlockedError(types.SimpleNamespace(reason="E2越界"), None)
+        ag._execute = types.MethodType(_boom, ag)
+        completed.clear()
+        rb = asyncio.run(ag.run_task("越界任务"))
+        assert rb.get("error") == "harness_blocked", rb
+        assert len(completed) == 1, "task_complete 应在 HarnessBlock 结束路径也触发(观测被拦截任务)"
+        print("  ✓ 无钩子不注入(behavior-preserving);pre_reasoning 注入前插;task_complete 正常+拦截路径都观测")
+    finally:
+        if os_ is not None:
+            asyncio.run(os_.aclose())
+        shutil.rmtree(d, ignore_errors=True)
+
+
+def test_preference_priority_recall():
+    """【8·偏好优先召回】明确偏好标 tag=preference(shared 持久)+ **独立于语义 query** 优先注入。
+    核心对照:偏好与当前任务**语义完全不相关**时,纯语义 recall 召不到,但偏好召回仍注入——
+    修复"越用越懂你"断在召回链的洞。默认开;FOXCLAW_PREF_CAPTURE=0 关采集。"""
+    print("\n【8】偏好优先召回:语义不相关也注入")
+    d = tempfile.mkdtemp(prefix="pref_")
+    os_ = None
+    try:
+        from foxclaw import Foxclaw
+        os_ = Foxclaw(data_dir=d, shadow_rounds=50)
+        ag = os_.office
+        ag.router.provider = "stub"
+        seen = {}
+        def _cap(system, messages, max_tokens, agent_id):
+            seen["messages"] = messages; return "（stub）"
+        ag.router.complete = _cap
+
+        asyncio.run(ag.run_task("我喜欢简洁直接的回答，不要长篇大论"))
+        prefs = ag.hmr.recall_by_tag(agent_id="supervisor", tag="preference", limit=8)
+        assert any("简洁" in p.get("content", "") for p in prefs), f"偏好应带 preference tag 存下,实得 {prefs}"
+
+        seen.clear()
+        asyncio.run(ag.run_task("解释一下光合作用的原理"))
+        joined = " ".join(m.get("content", "") for m in seen.get("messages", []))
+        assert "用户偏好" in joined and "简洁" in joined, \
+            f"语义不相关任务也应注入偏好(越用越懂你),实得 {seen.get('messages')}"
+
+        before = len(ag.hmr.recall_by_tag(agent_id="assistant", tag="preference"))
+        asyncio.run(ag.run_task("光合作用发生在植物的哪个部位"))
+        after = len(ag.hmr.recall_by_tag(agent_id="assistant", tag="preference"))
+        assert after == before, "无偏好信号的任务不应新增偏好"
+
+        os.environ["FOXCLAW_PREF_CAPTURE"] = "0"
+        try:
+            asyncio.run(ag.run_task("我喜欢用表格展示数据"))
+            got = ag.hmr.recall_by_tag(agent_id="assistant", tag="preference", limit=20)
+            assert not any("表格" in p.get("content", "") for p in got), "关采集后不应存偏好"
+        finally:
+            os.environ.pop("FOXCLAW_PREF_CAPTURE", None)
+
+        seen_stream = []
+        def _cap_stream(system, messages, *a, **k):
+            seen_stream.append(messages); yield "（stub）"
+        ag.router.stream = _cap_stream
+        list(ag.stream_reply("随便讲点什么"))
+        js = " ".join(m.get("content", "") for m in (seen_stream[0] if seen_stream else []))
+        assert "用户偏好" in js and "简洁" in js, f"对话路径(stream_reply)也应注入偏好,实得 {seen_stream}"
+
+        for _ in range(3):
+            ag._store_memory("用户偏好:我喜欢简洁直接的回答，不要长篇大论",
+                             mem_type="preference", tags=["preference"], shared=True)
+        got = ag.hmr.recall_by_tag(agent_id="assistant", tag="preference", limit=8)
+        same = [p for p in got if "简洁直接的回答" in p.get("content", "")]
+        assert len(same) == 1, f"重复偏好召回应去重为 1 条,实得 {len(same)}"
+
+        ag._store_memory("用户偏好:请用英文回复", mem_type="preference", tags=["preference"], shared=True)
+        target = [p for p in ag.list_preferences() if "英文" in p.get("content", "")]
+        assert target and target[0].get("id"), f"唯一偏好应可列出带 id,实得 {target}"
+        assert ag.delete_preference(target[0]["id"]), "删偏好应成功"
+        gone = [p for p in ag.list_preferences() if "英文" in p.get("content", "")]
+        assert not gone, "删除后该偏好不应再出现(F3:误存可清理)"
+        print("  ✓ 偏好存 tag;两路径注入;去重;不误存;可关;可 list/删除(F3)")
+    finally:
+        if os_ is not None:
+            asyncio.run(os_.aclose())
+        shutil.rmtree(d, ignore_errors=True)
+
+
+def main():
+    print("=" * 60)
+    print("统一任务生命周期 · run_task 收敛(P1)")
+    print("=" * 60)
+    test_structural_convergence()
+    test_domain_hooks()
+    test_unified_block_handling()
+    test_finance_verify_before_recommend()
+    test_run_task_stream()
+    test_interrupted_task_not_sunk()
+    test_reasoning_lifecycle_hooks()
+    test_preference_priority_recall()
+    print("\n" + "=" * 60)
+    print("✓ P1 验证完成:5 run_task→1 模板 + 领域 _execute 钩子 + 统一兜底 + P5 记忆验证优先")
+    print("=" * 60)
+
+
+if __name__ == "__main__":
+    main()