foxclaw 0.1.2 → 0.1.3

package/python/tests/test_parallel_dispatch.py

@@ -20,7 +20,7 @@ from pathlib import Path
 sys.path.insert(0, str(Path(__file__).parent.parent))
 
 import foxclaw.core.web_search as ws
-ws.needs_search = lambda q: False          # 关联网:离线测试不碰网络
+ws.needs_search = lambda q: False
 
 from foxclaw import Foxclaw
 from foxclaw.core.thought_chain import ThoughtNode
@@ -43,7 +43,7 @@ def _slow(system, messages, max_tokens=2048, agent_id=""):
 
 def _build(data_dir=None):
     os_ = Foxclaw(data_dir=data_dir or tempfile.mkdtemp())
-    os_.router.complete = _slow      # 注入固定延迟,让并发/串行差异可测
+    os_.router.complete = _slow
     return os_
 
 
@@ -55,12 +55,10 @@ def os_(tmp_path_factory):
 
 @pytest.mark.xfail(
     reason="pre-existing 失败:HKC_EMBEDDING=stub 下 LLM 全部本地 stub 调用极快,并行线程/await 开销反而把并行拖慢于顺序,断言 par<seq*0.75 在 stub 环境不可靠。test_hkc_index_build_manager 等 P1 测试仍覆盖冷构建核心。",
-    strict=False,   # 若日后通过/在某些环境通过,不强制转 fail,允许绿
+    strict=False,
 )
 def test_fork_join(os_):
     print("\n【1】并行 fork-join vs 顺序")
-    # 深审硬化:register_subagents 已在 setup 期预建 _scheduler/_steering_queue,
-    # 并发任务到达前属性就绪 → 消除懒建 check-then-act race
     check(hasattr(os_.supervisor, "_scheduler") and hasattr(os_.supervisor, "_steering_queue"),
           "主管 setup 期预建 scheduler/steering_queue(消除并发懒建 race)")
     r = asyncio.run(os_.supervisor.run_task(PAR_TASK))
@@ -83,7 +81,7 @@ def test_fork_join(os_):
 def test_fault_isolation(os_):
     print("\n【2】故障隔离:分支超时不拖垮整批")
     r = asyncio.run(os_.supervisor.run_task(
-        PAR_TASK, context={"parallel_timeout": 0.05}))   # 远小于 0.4s 的 stub
+        PAR_TASK, context={"parallel_timeout": 0.05}))
     res = r["result"]
     check(res["dispatched_to"] == "parallel", "仍走并行")
     steps = res["parallel_steps"]
@@ -126,7 +124,6 @@ def test_concurrency_safety(os_):
     check(cnt >= N * M, f"ThoughtChain 并发写不丢({cnt} ≥ {N*M})")
     st = pm.stats()
     check(st["by_state"].get("running", 0) == 0, "ProcessManager 无残留 running(状态机并发一致)")
-    # os_ 为多个子测试共享,故用 ≥(本测试贡献 N*M 个 completed,叠加此前的)
     check(st["by_state"].get("completed", 0) >= N * M, f"≥{N*M} 进程 completed(并发流转一致)")
     check(len(seen) == N * M, f"SystemBus 事件回调零丢失({len(seen)}/{N*M})")
 
@@ -152,7 +149,6 @@ def test_memory_isolation_concurrent(os_):
     for t in ts: t.start()
     for t in ts: t.join()
 
-    # 从 office 视角召回"保密草稿"：只能拿到自己的 + 共享,绝不能泄露 finance/admin 私有
     res = hmr.recall(query="保密草稿", agent_id="office", top_k=10)
     def foreign_private(r):
         tags = r.get("tags", [])
@@ -174,9 +170,6 @@ def test_gatekeeper_no_race(os_):
 
     def fire():
         try:
-            # 金额取 5000:**远离**任何阈值边界。本测试验的是"并发裁决无竞态",
-            # 与阈值定多少无关 —— 原先写 500 恰好卡在 large_purchase_check 边界上
-            # (规则是 > 500,500 不大于 500),调整默认阈值就把它误伤了。
             req = CrossAgentRequest(
                 from_agent="admin", to_agent="finance", request_type="budget_request",
                 payload={"amount": 5000, "category": "travel"}, reason="差旅")
@@ -192,7 +185,6 @@ def test_gatekeeper_no_race(os_):
     for t in ts: t.join()
 
     check(not errs, f"并发守门人裁决无异常（否则: {errs[:2]}）")
-    # 大额确定性触发 large_purchase_check → 全部一致(不写死具体阈值:那是可调项)
     check(set(results) == {"needs_confirmation"},
           f"{N} 次并发裁决结果一致（无竞态）: {set(results)}")
     check(len(orch.request_log) - before == N, f"请求日志无丢失（+{N}）")

package/python/tests/test_pipeline.py

@@ -32,7 +32,7 @@ class FakeAgent:
     def __init__(self, nodes=None, fail_on=None):
         self.thought_chain = _FakeChain(nodes or [])
         self.exec_count = 0
-        self._fail_on = fail_on          # 第 N 次 _execute 抛错(模拟重试期间 BLOCK/出错)
+        self._fail_on = fail_on
     def _recall_query(self, task): return f"q:{task}"
     def _recall_memory(self, query, top_k=5):
         self.last_recall = (query, top_k)
@@ -57,7 +57,6 @@ def test_default_pipeline():
     assert ctx.perceived.get("task_len") == 1, "perceive 记录输入特征"
     assert ctx.plan.get("multi_step") is False and ctx.route.get("target") == "self"
     assert ctx.verified.get("passed") is True, "verify 汇总校验结果到 ctx.verified"
-    # 默认 reflect:接受,不重试(behavior-preserving) → _execute 只调 1 次
     assert ctx.reflect["decision"] == "accept" and ctx.reflect["retries"] == 0
     assert a.exec_count == 1, "默认不重试,_execute 只调一次"
     print(f"  ✓ recall→plan→route→act→verify→reflect 按序;plan={ctx.plan['flow']} "
@@ -67,12 +66,10 @@ def test_default_pipeline():
 def test_plan_route_stage():
     print("\n【1c】plan/route stage:P1-2 引擎分析结构(informational,不改执行)")
     from foxclaw.pipeline import stage_plan, stage_route
-    # 多步任务(顺序连接词)
     ctx = PipelineContext(agent=FakeAgent(), task="先分析支出 然后 给出建议")
     stage_plan(ctx); stage_route(ctx)
     assert ctx.plan["multi_step"] is True and len(ctx.plan["steps"]) == 2, ctx.plan
     assert ctx.plan["flow"] == "sequential" and ctx.route["target"] == "self"
-    # 单步任务
     ctx2 = PipelineContext(agent=FakeAgent(), task="分析我的开支")
     stage_plan(ctx2)
     assert ctx2.plan["multi_step"] is False and ctx2.plan["steps"] == ["分析我的开支"]
@@ -94,7 +91,7 @@ def test_insertable_stage():
     print("\n【2】可插桩:插入自定义 stage 按位置执行")
     hits = []
     def stage_probe(ctx):
-        hits.append(("probe", len(ctx.memories)))   # 在 recall 之后、act 之前
+        hits.append(("probe", len(ctx.memories)))
     stage_probe.stage_name = "probe"
     a = FakeAgent()
     ctx = PipelineContext(agent=a, task="X")
@@ -133,14 +130,13 @@ def test_stage_events():
     print("\n【6】逐 stage 事件:每个 stage emit agent.cognition.<stage> 到脊柱")
     spine = _CapSpine()
     a = FakeAgent()
-    a._caps = _Caps(spine)                     # 注入能力中枢(解析出脊柱)
+    a._caps = _Caps(spine)
     ctx = PipelineContext(agent=a, task="E", session_id="s3")
     asyncio.run(CognitivePipeline(DEFAULT_STAGES).run(ctx))
     expected = ["agent.cognition." + n for n in
                 ["perceive", "recall", "plan", "route", "act", "verify", "reflect"]]
     assert spine.events == expected, f"逐 stage 事件,实得 {spine.events}"
-    # 无中枢时优雅跳过(不崩)
-    a2 = FakeAgent()   # 无 _caps
+    a2 = FakeAgent()
     asyncio.run(CognitivePipeline(DEFAULT_STAGES).run(PipelineContext(agent=a2, task="X")))
     print("  ✓ 7 个 agent.cognition.* 事件按序 emit;无中枢时优雅跳过")
 
@@ -149,33 +145,25 @@ def test_route_skill_optin_inert():
     print("\n【7】A·Step5:route 默认自动选 skill(仅只读+高阈值);可关;有副作用的不自动选")
     from foxclaw.pipeline import stage_route
     from foxclaw.skills import build_default_registry
-    reg = build_default_registry()          # 含 calculator(触发词「计算」等)
+    reg = build_default_registry()
 
-    # (a) 【契约变更 2026-07-16】原为"默认关"(behavior-preserving 放量期)。
-    #     现默认**开**:此前 route/dispatch 双关导致工具中心/MCP/技能在正常对话里根本够不着
-    #     (「帮我算 3+5」只会 LLM 硬答)。现默认自动选中**只读**技能。
     a = FakeAgent(); a.skill_registry = reg
     ctx = PipelineContext(agent=a, task="帮我计算 (12+8)*5")
     stage_route(ctx)
     assert ctx.route["target"] == "skill" and ctx.route["skill_id"] == "calculator", \
         f"默认应自动选中只读的 calculator: {ctx.route}"
 
-    # (a2) 可关(显式 False / 环境变量)→ 回落 self,老行为随时可拿回
     ctx_off = PipelineContext(agent=a, task="帮我计算 (12+8)*5",
                               context={"_route_skill": False})
     stage_route(ctx_off)
     assert ctx_off.route["target"] == "self", ctx_off.route
 
-    # (b) 开启(显式阈值 0.5,避免耦合具体分值):计算类 → router 选中 calculator,target=skill
     ctx2 = PipelineContext(agent=a, task="帮我计算 (12+8)*5",
                            context={"_route_skill": {"threshold": 0.5}})
     stage_route(ctx2)
     assert ctx2.route["target"] == "skill" and ctx2.route["skill_id"] == "calculator", ctx2.route
     assert ctx2.route["confidence"] >= 0.5, ctx2.route
 
-    # (c) 【契约变更】原断言"act 不消费 route"(Step1 铺线期)。Step2 起 act 会真派发;
-    #     此处验证:选中 skill 但**无 skill_adapter**(FakeAgent 没挂)→ _act_via_skill 回落
-    #     _execute,fail-safe 不崩。
     a3 = FakeAgent(); a3.skill_registry = reg
     ctx3 = PipelineContext(agent=a3, task="帮我计算 (12+8)*5",
                            context={"_route_skill": {"threshold": 0.5}})
@@ -184,13 +172,11 @@ def test_route_skill_optin_inert():
     assert ctx3.result == "done:帮我计算 (12+8)*5#1" and a3.exec_count == 1, \
         "无 skill_adapter → 回落 _execute(fail-safe)"
 
-    # (d) 无 skill_registry 的 Agent(finance/admin/office)→ 恒 self(fail-safe 回落)
     a4 = FakeAgent()
     ctx4 = PipelineContext(agent=a4, task="计算 1+1", context={"_route_skill": True})
     stage_route(ctx4)
     assert ctx4.route["target"] == "self", "无 skill_registry → 回落 self"
 
-    # (e) 高阈值(>1)→ 实际不选,回落 self(阈值可调证明)
     ctx5 = PipelineContext(agent=a, task="帮我计算 1+1",
                            context={"_route_skill": {"threshold": 1.01}})
     stage_route(ctx5)
@@ -209,7 +195,6 @@ def test_act_skill_dispatch_step2():
         a = FakeAgent(); a.skill_registry = reg; a.skill_adapter = adapter
         return a
 
-    # (a) 开启 dispatch + 计算类 → 真正执行 calculator,结果=100,且领域 _execute(LLM) 不被调
     a = mk_agent()
     ctx = PipelineContext(agent=a, task="帮我计算 (12+8)*5",
                           context={"_route_skill": {"threshold": 0.5}, "_skill_dispatch": True})
@@ -219,9 +204,6 @@ def test_act_skill_dispatch_step2():
     assert ctx.result["output"]["result"] == 100, ctx.result
     assert a.exec_count == 0, "skill 执行了 → 领域 _execute(LLM) 不再被调(首个真改执行)"
 
-    # (b) 【安全铁律】开启 dispatch + 发邮件(destructive/send)未确认 → needs_confirmation,零副作用、不回落 LLM
-    #     注:默认的只读闸会挡住 gmail_send(有副作用不自动选),故此处**显式** read_only=False
-    #     表示"我就是要测有副作用的路由被 PolicyGate 拦住"。
     a2 = mk_agent()
     ctx2 = PipelineContext(agent=a2, task="帮我发送邮件给客户",
                            context={"_route_skill": {"threshold": 0.5, "read_only": False},
@@ -231,37 +213,31 @@ def test_act_skill_dispatch_step2():
     assert ctx2.result["status"] == "needs_confirmation" and ctx2.result["pending"] is True, ctx2.result
     assert a2.exec_count == 0, "未确认的危险 skill → 不执行、不回落 LLM(绝不 fail-open)"
 
-    # (c) 显式关 dispatch(只 route)→ 即便 target=skill,act 仍走 _execute(两道门相互独立)
-    #     注:dispatch 现在**默认开**,故此处显式 False 才能验证"关掉"这条路径。
     a3 = mk_agent()
     ctx3 = PipelineContext(agent=a3, task="帮我计算 (12+8)*5",
                            context={"_route_skill": {"threshold": 0.5}, "_skill_dispatch": False})
     asyncio.run(CognitivePipeline(DEFAULT_STAGES).run(ctx3))
     assert ctx3.route["target"] == "skill" and a3.exec_count == 1, "dispatch 关 → 回落 _execute"
 
-    # (d) 开 dispatch 但 Agent 无 skill_adapter → 回落 _execute(fail-safe)
-    a4 = FakeAgent(); a4.skill_registry = reg     # 有 registry 无 adapter
+    a4 = FakeAgent(); a4.skill_registry = reg
     ctx4 = PipelineContext(agent=a4, task="帮我计算 1+1",
                            context={"_route_skill": {"threshold": 0.5}, "_skill_dispatch": True})
     asyncio.run(CognitivePipeline(DEFAULT_STAGES).run(ctx4))
     assert a4.exec_count == 1, "无 adapter → 回落 _execute"
 
-    # (b2) 【新安全闸·A·Step5】默认(只读闸开)下,发邮件这类有副作用的技能**根本不会被自动选中**
-    #      —— 自动执行不得越过"会改变世界"的线;要用仍可显式 use_skill(那时用户是知情的)。
     a2b = mk_agent()
-    ctx2b = PipelineContext(agent=a2b, task="帮我发送邮件给客户")   # 全默认
+    ctx2b = PipelineContext(agent=a2b, task="帮我发送邮件给客户")
     asyncio.run(CognitivePipeline(DEFAULT_STAGES).run(ctx2b))
     assert ctx2b.route["target"] == "self", f"有副作用的技能不该被自动选中: {ctx2b.route}"
     assert a2b.exec_count == 1, "未自动动手 → 由 LLM 作答"
 
-    # (e) 【fail-closed 加固】invoke 意外抛错 → 结构化 error,不崩、不回落 LLM
     class _BoomAdapter:
         def invoke(self, skill, args, confirmed=False, mode="strict"):
             raise RuntimeError("dispatch 内部炸了")
     a5 = FakeAgent(); a5.skill_registry = reg; a5.skill_adapter = _BoomAdapter()
     ctx5 = PipelineContext(agent=a5, task="帮我计算 1+1",
                            context={"_route_skill": {"threshold": 0.5}, "_skill_dispatch": True})
-    asyncio.run(CognitivePipeline(DEFAULT_STAGES).run(ctx5))   # 不抛异常
+    asyncio.run(CognitivePipeline(DEFAULT_STAGES).run(ctx5))
     assert ctx5.result["status"] == "error" and "异常" in ctx5.result["message"], ctx5.result
     assert a5.exec_count == 0, "dispatch 异常 → fail-closed 结构化 error,绝不回落 LLM"
     print("  ✓ 开启:calculator 真执行(=100,不调LLM);发邮件未确认→needs_confirmation零副作用;"
@@ -303,35 +279,28 @@ def test_run_task_confirm_suspend_step3():
         ag = _Mini(hmr=_FakeHMR(), thought_chain=ThoughtChain(db_path=f"{d}/tc.db"),
                    data_dir=d, shadow_rounds=5)
         ag.skill_registry = reg; ag.skill_adapter = adapter
-        # read_only=False:本测试专门验证**写入类** skill 的 confirm 挂起路径,
-        # 而默认的只读闸不会自动选中它 —— 故显式声明(调用方自担风险的显式选择)。
         base_ctx = {"_route_skill": {"threshold": 0.4, "read_only": False},
                     "_skill_dispatch": True}
 
-        # 未确认:写 skill → run_task 顶层 status=needs_confirmation + pending_action,零副作用
         r1 = asyncio.run(ag.run_task("帮我记一下 买牛奶", dict(base_ctx)))
         assert r1.get("status") == "needs_confirmation", r1
         assert r1.get("pending_action", {}).get("skill_id") == "save_note", r1
         assert saved == [], "未确认 → handler 不执行(零副作用)"
 
-        # 【深审加固】确认绑定错 skill_id → 仍 needs_confirmation、不执行(防确认漂移)
         r_bad = asyncio.run(ag.run_task("帮我记一下 买牛奶",
                                         {**base_ctx, "_confirmed": "gmail_send"}))
         assert r_bad.get("status") == "needs_confirmation", r_bad
         assert saved == [], "确认绑定到别的 skill → 不执行本 skill(绑定优先)"
 
-        # 确认恢复:绑定到正确 skill_id 重跑 → skill 真执行,无 needs_confirmation
         r2 = asyncio.run(ag.run_task("帮我记一下 买牛奶",
                                      {**base_ctx, "_confirmed": "save_note"}))
         assert r2.get("status") != "needs_confirmation", r2
         assert len(saved) == 1, f"确认后 handler 执行(有副作用),saved={saved}"
 
-        # 向后兼容:_confirmed=True 概括确认仍生效
         saved.clear()
         r2b = asyncio.run(ag.run_task("帮我记一下 买酸奶", {**base_ctx, "_confirmed": True}))
         assert r2b.get("status") != "needs_confirmation" and len(saved) == 1, (r2b, saved)
 
-        # 纯 LLM 路径(dispatch 关)→ 无 status 字段(行为逐字不变)
         r3 = asyncio.run(ag.run_task("随便聊聊"))
         assert "status" not in r3 and r3["result"] == "llm:随便聊聊", r3
         print(f"  ✓ 未确认→顶层 needs_confirmation 零副作用;_confirmed 重跑→执行(saved={saved});"
@@ -343,7 +312,7 @@ def test_run_task_confirm_suspend_step3():
 def test_reflect_retry_optin():
     print("\n【4】reflect opt-in:漂移超阈值 → 有上限重试")
     from foxclaw.pipeline import stage_verify   # noqa
-    a = FakeAgent(nodes=[_Node("WARN", 0.8)])   # verify 得 max_delta=0.8 > 阈值 0.5
+    a = FakeAgent(nodes=[_Node("WARN", 0.8)])
     ctx = PipelineContext(agent=a, task="V",
                           context={"_reflect": {"max_retries": 2, "drift_threshold": 0.5}})
     asyncio.run(CognitivePipeline(DEFAULT_STAGES).run(ctx))
@@ -358,8 +327,8 @@ class _DynChain:
     def __init__(self, agent): self.agent = agent
     def get_chain(self, session_id, agent_id=None):
         if self.agent.exec_count <= 1:
-            return [_Node("WARN", 0.8)]                      # 初次:高漂移
-        return [_Node("WARN", 0.8), _Node("PASS", 0.2)]     # 重试后:旧高(0.8) + 新低(0.2)
+            return [_Node("WARN", 0.8)]
+        return [_Node("WARN", 0.8), _Node("PASS", 0.2)]
 
 
 def test_reflect_early_exit_on_improve():
@@ -369,10 +338,8 @@ def test_reflect_early_exit_on_improve():
     ctx = PipelineContext(agent=a, task="D",
                           context={"_reflect": {"max_retries": 3, "drift_threshold": 0.5}})
     asyncio.run(CognitivePipeline(DEFAULT_STAGES).run(ctx))
-    # 初次 exec=1 → last_delta=0.8>0.5 → 重试1 exec=2 → last_delta=0.2<=0.5 → 早退(非到上限3)
     assert a.exec_count == 2, f"改善后应只重试 1 次(而非到上限 3),实得 {a.exec_count}"
     assert ctx.reflect["decision"] == "retry_accepted" and ctx.reflect["retries"] == 1, ctx.reflect
-    # 关键:此时 max_delta 仍是 0.8(旧节点),若用它会错误地重试到上限;last_delta=0.2 才对
     assert ctx.verified["max_delta"] == 0.8 and ctx.verified["last_delta"] == 0.2
     print(f"  ✓ 早退:exec={a.exec_count} reflect={ctx.reflect['decision']} "
           f"(max_delta={ctx.verified['max_delta']} 仍高,但 last_delta={ctx.verified['last_delta']} 达标)")
@@ -380,7 +347,7 @@ def test_reflect_early_exit_on_improve():
 
 def test_reflect_failsafe_rollback():
     print("\n【5】reflect fail-safe:重试自身出错 → 回滚到重试前结果")
-    a = FakeAgent(nodes=[_Node("WARN", 0.8)], fail_on=2)   # 第 2 次 _execute(首个重试)抛错
+    a = FakeAgent(nodes=[_Node("WARN", 0.8)], fail_on=2)
     ctx = PipelineContext(agent=a, task="X",
                           context={"_reflect": {"max_retries": 2, "drift_threshold": 0.5}})
     asyncio.run(CognitivePipeline(DEFAULT_STAGES).run(ctx))
@@ -401,10 +368,10 @@ def test_concurrent_ctx_isolation():
 
     class SlowAgent(FakeAgent):
         async def _execute(self, task, context, session_id, memories):
-            await asyncio.sleep(0.005)          # 让出控制权 → 制造交错 → 暴露串状态
+            await asyncio.sleep(0.005)
             return f"done:{task}", {"task_echo": task}
 
-    a = SlowAgent()                             # 同一 agent 实例,被 24 个任务并发打
+    a = SlowAgent()
 
     async def run_one(task):
         ctx = PipelineContext(agent=a, task=task, session_id=f"s_{task}")
@@ -434,40 +401,35 @@ def test_cooperative_interrupt():
     import threading
     print("\n【7】P3 协作式中断:stop_event 在 stage 边界 / 重试迭代前优雅停")
 
-    # (a) 无 stop_event → 跑满 7 stage、interrupted=False(behavior-preserving)
     a = FakeAgent()
     ctx = PipelineContext(agent=a, task="T")
     asyncio.run(CognitivePipeline(DEFAULT_STAGES).run(ctx))
     assert ctx.interrupted is False and len(ctx.trace) == 7, (ctx.interrupted, ctx.trace)
 
-    # (b) 起步即置位 → 0 个 stage 执行、interrupted=True、无结果
     ev = threading.Event(); ev.set()
     ctx2 = PipelineContext(agent=FakeAgent(), task="T", stop_event=ev)
     asyncio.run(CognitivePipeline(DEFAULT_STAGES).run(ctx2))
     assert ctx2.interrupted is True and ctx2.trace == [] and ctx2.result is None, \
         (ctx2.interrupted, ctx2.trace, ctx2.result)
 
-    # (c) 中途置位(自定义 stage 跑完就 set)→ 停在下一边界,后续 stage 全不跑
     ev2 = threading.Event()
     def _tripwire(c): ev2.set()
     _tripwire.stage_name = "tripwire"
-    stages = DEFAULT_STAGES[:2] + [_tripwire] + DEFAULT_STAGES[2:]   # perceive,recall,tripwire,plan...
+    stages = DEFAULT_STAGES[:2] + [_tripwire] + DEFAULT_STAGES[2:]
     ctx3 = PipelineContext(agent=FakeAgent(), task="T", stop_event=ev2)
     asyncio.run(CognitivePipeline(stages).run(ctx3))
     assert ctx3.trace == ["perceive", "recall", "tripwire"], ctx3.trace
     assert ctx3.interrupted is True and ctx3.result is None, (ctx3.interrupted, ctx3.result)
     print(f"  ✓ 边界中断:无 stop 跑满 7;起步停 0 stage;中途停 trace={ctx3.trace}(后续未跑)")
 
-    # (d) reflect 重试循环中断:agent 在**首次重试的 _execute** 时置位 → 下一轮迭代前停,
-    #     证明重试在用尽 max_retries(=5)之前被截断。这是本管线最接近"agentic loop 每轮检查中断"处。
     ev3 = threading.Event()
     class _StopOnRetryAgent(FakeAgent):
         def __init__(self, e):
-            super().__init__(nodes=[_Node("WARN", 0.8)])   # 恒高漂移 → 不中断会一直重试到上限
+            super().__init__(nodes=[_Node("WARN", 0.8)])
             self._e = e
         async def _execute(self, task, context, session_id, memories):
             self.exec_count += 1
-            if self.exec_count == 2:            # 第 1 次重试的 LLM 调用时请求中断
+            if self.exec_count == 2:
                 self._e.set()
             return f"done:{task}#{self.exec_count}", {"call": self.exec_count}
     a4 = _StopOnRetryAgent(ev3)
@@ -485,18 +447,15 @@ def test_cooperative_interrupt():
 def test_on_stage_callback():
     """【8】D1:`pipeline.run(on_stage)` 逐 stage 回调(opt-in);无回调=行为不变;回调异常 fail-safe。"""
     print("\n【8】D1 on_stage 回调:逐 stage 通知 + opt-in + fail-safe")
-    # (a) 有回调:每 stage 完成后回调一次,顺序与 trace 一致
     seen = []
     ctx = PipelineContext(agent=FakeAgent(), task="T")
     asyncio.run(CognitivePipeline(DEFAULT_STAGES).run(
         ctx, on_stage=lambda name, c: seen.append(name)))
     assert seen == ctx.trace == ["perceive", "recall", "plan", "route", "act", "verify", "reflect"], \
         (seen, ctx.trace)
-    # (b) 无回调:行为逐字不变(默认路径)
     ctx2 = PipelineContext(agent=FakeAgent(), task="T")
     asyncio.run(CognitivePipeline(DEFAULT_STAGES).run(ctx2))
     assert len(ctx2.trace) == 7 and ctx2.result == "done:T#1"
-    # (c) 回调抛异常:被吞,不影响管线跑完(fail-safe:流式反馈绝不拖垮任务)
     def _boom(name, c):
         raise RuntimeError("回调故障")
     ctx3 = PipelineContext(agent=FakeAgent(), task="T")

package/python/tests/test_plan_visible.py

@@ -44,12 +44,12 @@ def test_stage_plan_records_plan_node() -> None:
 
 
 def test_record_plan_node_never_breaks_execution() -> None:
-    _record_plan_node(_ctx("x", agent=None))          # 无 _think → 静默跳过
+    _record_plan_node(_ctx("x", agent=None))
 
     def _boom(**kwargs):
         raise RuntimeError("harness 挂了")
 
-    _record_plan_node(_ctx("x", agent=types.SimpleNamespace(_think=_boom)))  # 不得抛出
+    _record_plan_node(_ctx("x", agent=types.SimpleNamespace(_think=_boom)))
 
 
 def test_plan_payload_accepts_both_shapes() -> None:
@@ -59,7 +59,7 @@ def test_plan_payload_accepts_both_shapes() -> None:
                                      "multi_step": True}
     sup = types.SimpleNamespace(plan_obj=types.SimpleNamespace(steps=["x", "y"], flow="parallel"))
     assert _plan_payload(sup)["steps"] == ["x", "y"]
-    assert _plan_payload(sup)["multi_step"] is True      # 未声明 → 按步数推断
+    assert _plan_payload(sup)["multi_step"] is True
     one = types.SimpleNamespace(plan_obj=types.SimpleNamespace(steps=["x"]))
     assert _plan_payload(one)["multi_step"] is False
     assert _plan_payload(types.SimpleNamespace(plan=None)) is None
@@ -72,7 +72,7 @@ def test_stream_emits_plan_event_after_plan_stage() -> None:
 
     async def _core(task, context=None, stop_event=None, on_stage=None):
         on_stage("recall", ctx)
-        on_stage("plan", ctx)          # 管线在 plan stage 完成后回调
+        on_stage("plan", ctx)
         return {"result": "done"}
 
     fake = types.SimpleNamespace(_run_task_core=_core)

package/python/tests/test_planner.py

@@ -38,33 +38,26 @@ def test_trigger_route_engine():
         C("web", ["搜索", "查"], "联网 搜索 网页"),
         C("email", ["发邮件", "email"], "发送 邮件"),
     ]
-    # 否定:"不要计算" → calculator 触发被否定 → 不选它
     r = P.trigger_route("不要计算一下", cands)
     assert r.primary != "calculator", f"否定应排除 calculator,实得 {r.primary}"
-    # 命中:"帮我搜索北京" → web
     r2 = P.trigger_route("帮我搜索北京", cands)
     assert r2.primary == "web" and r2.confidence >= 0.15
-    # 兜底:完全不相关 → 用最后一个候选兜底,低置信
     r3 = P.trigger_route("xyzzy 无关内容", cands)
     assert r3.primary == cands[-1].id and r3.confidence < 0.15
-    # 确定性:空候选
     assert P.trigger_route("x", []).primary is None
     print(f"  ✓ 否定排除 / 命中选中 / 兜底(={r3.primary}) / 空候选")
 
 
 def test_three_sites_delegate():
     print("\n【3】三处委托同一引擎(单一真源)")
-    # AHS
     from foxclaw.skills import build_default_registry
     from foxclaw.skills.pipeline import route, plan
     reg = build_default_registry()
     assert route("不要计算", reg).primary != "calculator", "AHS route 否定生效"
     assert len(plan("先算 5+5 然后再算 9*9", reg).tasks) == 2, "AHS plan 拆解"
-    # search
     from foxclaw.search.planner import QueryPlanner
     sub = QueryPlanner()._decompose("查A，查B、查C")
     assert len(sub) == 3, "search 拆解走共享原语"
-    # supervisor(用引擎的 keyword_hits;此处直接验证原语等价)
     assert P.keyword_hits("买基金理财", ["理财", "基金", "收支"]) == 2
     print("  ✓ AHS / search / 主管 三处均由 foxclaw.planning 单一真源驱动")
 

package/python/tests/test_policy_gate.py

@@ -17,7 +17,7 @@ sys.path.insert(0, str(Path(__file__).parent.parent))
 from foxclaw.kernel import PolicyGate
 
 
-class _HR:                          # 假 HarnessResult
+class _HR:
     def __init__(self, blocked=False, warned=False, reason=""):
         self._b, self._w, self.reason = blocked, warned, reason
     @property
@@ -31,7 +31,7 @@ class _Harness:
     def check(self, node): self.calls += 1; return self.hr
 
 
-class _PD:                          # 假 PermissionDecision
+class _PD:
     def __init__(self, decision, reason=""):
         self.decision, self.stage, self.reason, self.trust_downgrade = decision, 2, reason, False
 
@@ -81,7 +81,6 @@ def test_integration_skill_path():
         os_ = Foxclaw(data_dir=d, shadow_rounds=50)
         seen = []
         os_.event_spine.project(lambda e: seen.append(e.topic))
-        # 发邮件是副作用 skill → 需确认 → 经 PolicyGate 发 policy.decision.confirm
         os_.supervisor.use_skill("帮我发邮件给客户",
                                 args={"to": "a@b.com", "subject": "x", "body": "y"})
         assert any(t.startswith("policy.decision") for t in seen), \

package/python/tests/test_profile.py

@@ -0,0 +1,67 @@
+"""称呼(user profile)存取:round-trip、改一次不留旧条目、可清空、不误删别的偏好。
+
+用 stub 宿主直接测 mixin,不初始化 Agent / API(那条路径由 test_api 覆盖)。
+"""
+
+from __future__ import annotations
+
+from foxclaw.agents.mixins.preference import PreferenceMixin
+
+
+class _Stub(PreferenceMixin):
+    AGENT_ID = "stub"
+
+    def __init__(self):
+        self.items: list[dict] = []
+        self._seq = 0
+
+    def _store_memory(self, content, mem_type="concept", **kwargs):
+        self._seq += 1
+        self.items.insert(0, {"id": f"m{self._seq}", "content": content})
+
+    def list_preferences(self, limit=50):
+        return self.items[:limit]
+
+    def delete_preference(self, pref_id):
+        before = len(self.items)
+        self.items = [p for p in self.items if p["id"] != pref_id]
+        return len(self.items) < before
+
+
+def test_nickname_roundtrip():
+    s = _Stub()
+    assert s.get_nickname() == ""
+    assert s.set_nickname("  snow  ") == "snow"
+    assert s.get_nickname() == "snow"
+
+
+def test_set_nickname_replaces_old():
+    """改称呼不能留下两条互相矛盾的偏好。"""
+    s = _Stub()
+    s.set_nickname("snow")
+    s.set_nickname("snow2")
+    assert [s._nick_of(p["content"]) for p in s.items] == ["snow2"]
+    assert s.get_nickname() == "snow2"
+
+
+def test_clear_nickname():
+    s = _Stub()
+    s.set_nickname("snow")
+    assert s.set_nickname("") == ""
+    assert s.get_nickname() == ""
+    assert s.items == []
+
+
+def test_other_preferences_untouched():
+    s = _Stub()
+    s.remember_preference("回答尽量简短")
+    s.set_nickname("snow")
+    contents = [p["content"] for p in s.items]
+    assert "用户偏好:回答尽量简短" in contents
+    assert "用户偏好:称呼我为 snow" in contents
+
+
+def test_nickname_truncated():
+    s = _Stub()
+    long = "x" * 60
+    assert s.set_nickname(long) == long[:24]