chore: update gitignore and remove env.example

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>

.gitignore

@@ -40,6 +40,9 @@ logs/
 .claude/
 .worktrees/
 
+# Demo (excluded from version control)
+demo/
+
 # Lock files (use in development, commit in production)
 # uv.lock - uncomment if you want to commit lock file
 # package-lock.json - uncomment if you want to commit lock file

backend/app/agents/graph.py

@@ -27,6 +27,9 @@ from app.agents.prompts import (
     MASTER_SYSTEM_PROMPT,
     SCHEDULE_PLANNER_SYSTEM_PROMPT,
 )
+from app.agents.orchestration.result_merge import merge_task_results
+from app.agents.orchestration.scheduler import build_subtask_specs, ensure_child_links
+from app.agents.orchestration.subagent_runtime import subtask_spec_to_agent_task
 from app.agents.registry import load_builtin_registry_indexes
 from app.agents.runtime_metrics import (
     coerce_cost_thresholds,
@@ -36,6 +39,14 @@ from app.agents.runtime_metrics import (
 )
 from app.agents.schemas.event import AgentEvent
 from app.agents.schemas.message import AgentMessage
+from app.agents.schemas.orchestration import (
+    ExecutionDecision,
+    MergeReport,
+    RuntimeRequestContext,
+    TaskGraph,
+    VerificationReport,
+    render_runtime_request_context_summary,
+)
 from app.agents.schemas.task import (
     AgentTask,
     CollaborationBudget,
@@ -44,6 +55,7 @@ from app.agents.schemas.task import (
     TaskResult,
 )
 from app.agents.skill_registry import build_skill_context
+from app.agents.skills.retriever import build_shortlisted_skill_context
 from app.agents.state import AgentRole, AgentState
 from app.agents.tools import SUB_COMMANDER_TOOLSETS
 from app.agents.tools.time_reasoning import normalize_tool_time_arguments
@@ -1148,6 +1160,57 @@ def _parse_json_action(content: str, allowed_tools: list[str]) -> dict[str, Any]
     return None
 
 
+def _looks_like_internal_tool_markup(content: str) -> bool:
+    text = (content or "").strip()
+    if not text:
+        return False
+
+    lowered = text.lower()
+    xml_markers = (
+        "<minimax:tool_call",
+        "</minimax:tool_call>",
+        "<invoke name=",
+        "</invoke>",
+        "<parameter name=",
+        "</parameter>",
+    )
+    if any(marker in lowered for marker in xml_markers):
+        return True
+
+    return "分发说明" in text and ("<invoke name=" in lowered or "tool_call" in lowered)
+
+
+def _clean_tool_result_for_user(tool_result: str | None) -> str:
+    text = (tool_result or "").strip()
+    if not text:
+        return ""
+
+    cleaned_lines = [
+        re.sub(r"^\[[^\]]+\]\s*", "", line).strip()
+        for line in text.splitlines()
+        if line.strip()
+    ]
+    return "\n".join(line for line in cleaned_lines if line).strip()
+
+
+def _build_internal_markup_fallback_response(
+    state: AgentState,
+    *,
+    sub_commander: str,
+) -> str | None:
+    tool_result = state.get("last_tool_result")
+    cleaned_tool_result = _clean_tool_result_for_user(tool_result)
+    if not cleaned_tool_result or _tool_result_indicates_failure(cleaned_tool_result):
+        return None
+
+    if sub_commander == "librarian_retrieval":
+        if _is_missing_knowledge_result(tool_result):
+            return "这次检索没有拿到有效证据。我先不展示内部调度过程；如果您愿意，我可以直接基于常识回答，或改为联网搜索后再整理。"
+        return f"我已经完成检索，直接给您可用信息：\n\n{cleaned_tool_result}"
+
+    return cleaned_tool_result
+
+
 def _has_active_structured_continuation(state: AgentState) -> bool:
     pending_action = state.get("pending_action") or {}
     routing_decision = state.get("routing_decision") or {}
@@ -1205,6 +1268,70 @@ def _build_structured_continuity_summary(state: AgentState) -> str | None:
     return "\n".join(lines)
 
 
+def _build_retrospective_context_summary(state: AgentState) -> str | None:
+    retrospectives = list(state.get("recalled_retrospectives") or [])
+    if not retrospectives:
+        return None
+
+    lines = ["【相关历史复盘】"]
+    for item in retrospectives[:2]:
+        if not isinstance(item, dict):
+            continue
+        request_summary = str(item.get("request_summary") or item.get("task_type") or "").strip()
+        execution_mode = str(item.get("execution_mode") or "").strip()
+        success_score = float(item.get("success_score") or 0.0)
+        reusable_patterns = list(item.get("reusable_patterns") or [])
+        avoid_patterns = list(item.get("avoid_patterns") or [])
+
+        summary_parts = [request_summary[:80] or execution_mode or "历史任务"]
+        if execution_mode:
+            summary_parts.append(f"mode={execution_mode}")
+        summary_parts.append(f"score={success_score:.2f}")
+        if reusable_patterns:
+            summary_parts.append(f"可复用={','.join(reusable_patterns[:2])}")
+        elif avoid_patterns:
+            summary_parts.append(f"避坑={','.join(avoid_patterns[:2])}")
+
+        lines.append(f"- {'；'.join(summary_parts)}")
+
+    return "\n".join(lines) if len(lines) > 1 else None
+
+
+def _estimate_request_complexity(user_query: str, selected_roles: list[str]) -> float:
+    text = (user_query or "").strip()
+    base = min(len(text) / 120.0, 1.0)
+    role_boost = min(len(selected_roles) * 0.2, 0.6)
+    return round(min(base + role_boost, 1.0), 2)
+
+
+def _record_execution_decision(
+    state: AgentState,
+    *,
+    user_query: str,
+    mode: Literal["direct", "collaboration", "parallel"],
+    reason: str,
+    selected_roles: list[str] | None = None,
+    parallel_worthiness_score: float | None = None,
+) -> None:
+    runtime_request_context = state.get("runtime_request_context") or {}
+    request_id = str(runtime_request_context.get("request_id") or state.get("conversation_id") or "")
+    roles = list(selected_roles or [])
+    decision = ExecutionDecision(
+        request_id=request_id or f"request-{uuid4().hex[:8]}",
+        mode=mode,
+        reason=reason,
+        complexity_score=_estimate_request_complexity(user_query, roles),
+        parallel_worthiness_score=parallel_worthiness_score,
+        selected_roles=roles,
+    )
+    state["execution_decision"] = decision.model_dump(mode="json")
+    _append_event_trace(
+        state,
+        "agent.execution.decided",
+        payload=state["execution_decision"],
+    )
+
+
 def _build_system_messages(
     state: AgentState, system_prompt: str, role: AgentRole, sub_commander: str
 ) -> list[BaseMessage]:
@@ -1214,6 +1341,19 @@ def _build_system_messages(
     if current_datetime_context:
         messages.append(SystemMessage(content=current_datetime_context))
 
+    runtime_request_context = state.get("runtime_request_context")
+    if isinstance(runtime_request_context, dict) and runtime_request_context:
+        try:
+            runtime_context_model = RuntimeRequestContext.model_validate(runtime_request_context)
+        except Exception:
+            runtime_context_model = None
+        if runtime_context_model is not None:
+            messages.append(
+                SystemMessage(
+                    content=render_runtime_request_context_summary(runtime_context_model)
+                )
+            )
+
     continuity_summary = _build_structured_continuity_summary(state)
     if continuity_summary:
         messages.append(SystemMessage(content=continuity_summary))
@@ -1226,6 +1366,10 @@ def _build_system_messages(
     if collaboration_summary:
         messages.append(SystemMessage(content=collaboration_summary))
 
+    retrospective_summary = _build_retrospective_context_summary(state)
+    if retrospective_summary:
+        messages.append(SystemMessage(content=retrospective_summary))
+
     role_context_map = {
         AgentRole.SCHEDULE_PLANNER: state.get("schedule_context_summary"),
         AgentRole.LIBRARIAN: state.get("knowledge_context"),
@@ -1237,7 +1381,11 @@ def _build_system_messages(
 
     role_skill_key = ROLE_SKILL_CONTEXT.get(role)
     if role_skill_key:
-        skill_context = build_skill_context(role_skill_key)
+        shortlisted_context = build_shortlisted_skill_context(
+            state.get("skill_shortlist"),
+            agent_type=role_skill_key,
+        )
+        skill_context = shortlisted_context or build_skill_context(role_skill_key)
         if skill_context:
             messages.append(SystemMessage(content=skill_context))
 
@@ -1322,6 +1470,29 @@ def _build_collaboration_tasks(user_query: str) -> list[AgentTask]:
     return tasks
 
 
+def _build_collaboration_plan_from_task_graph(
+    state: AgentState,
+    user_query: str,
+) -> tuple[list[AgentTask], list[dict[str, Any]]]:
+    raw_task_graph = state.get("task_graph")
+    if not isinstance(raw_task_graph, dict) or not raw_task_graph.get("nodes"):
+        return _build_collaboration_tasks(user_query), []
+
+    task_graph = TaskGraph.model_validate(raw_task_graph)
+    specs = [
+        spec
+        for spec in build_subtask_specs(task_graph, query_text=user_query)
+        if spec.role != "master"
+    ]
+    child_links = ensure_child_links(specs)
+    tasks: list[AgentTask] = []
+    for spec in specs:
+        task = subtask_spec_to_agent_task(spec)
+        task.child_task_ids = child_links.get(spec.subtask_id, [])
+        tasks.append(task)
+    return tasks, [spec.model_dump(mode="json") for spec in specs]
+
+
 def _build_collaboration_context_summary(state: AgentState) -> str | None:
     if state.get("execution_mode") != "collaboration":
         return None
@@ -2076,7 +2247,36 @@ async def _run_sub_commander(
                 )
                 _record_response_usage(state, response)
                 response_text = _stringify_message_content(response.content)
+                response_text_stripped = response_text.strip()
                 parsed = _parse_json_action(response_text, allowed_tools)
+                if parsed is None and response_text_stripped and _looks_like_internal_tool_markup(
+                    response_text_stripped
+                ):
+                    if int(state.get("retry_count") or 0) >= int(state.get("max_retries") or 0):
+                        state["fallback_parse_error"] = "internal_tool_markup"
+                        state["final_response"] = _build_internal_markup_fallback_response(
+                            state,
+                            sub_commander=sub_commander,
+                        ) or (
+                            "这次内部调度没有正确收束成最终答复。我先不展示内部调用过程；您重试一次，我会直接用自然语言回答。"
+                        )
+                        break
+                    if not _guard_sub_commander_budget(
+                        state, "iteration_count", "max_iterations", "max_iterations_exceeded"
+                    ):
+                        parsed = None
+                        break
+                    state["iteration_count"] = int(state.get("iteration_count") or 0) + 1
+                    state["retry_count"] = int(state.get("retry_count") or 0) + 1
+                    retry_instruction = SystemMessage(
+                        content=(
+                            "上一轮输出了内部调度或工具调用标记，这是协议错误。"
+                            "不要再输出分发说明、XML 标签、<invoke>、<parameter>、JSON 或 tool_call。"
+                            "请直接面向用户给出最终自然语言答复；如果已有工具结果，就基于结果整理；"
+                            "如果工具没有找到证据，可以基于常识直接回答。"
+                        )
+                    )
+                    continue
                 if parsed is None and response_text.strip() and state.get("tool_round_count"):
                     state["fallback_parse_error"] = None
                     state["final_response"] = response_text.strip()
@@ -2382,6 +2582,35 @@ def _build_collaboration_final_response(task_results: list[TaskResult | dict[str
     return "\n".join(lines)
 
 
+def _build_serial_fallback_response(
+    user_query: str,
+    task_results: list[TaskResult | dict[str, Any]],
+    merge_report: MergeReport | dict[str, Any] | None,
+) -> str:
+    normalized_results = [normalize_task_result(item) for item in task_results]
+    completed = [item for item in normalized_results if item.status == "completed"]
+    merge_payload = (
+        merge_report.model_dump(mode="json")
+        if isinstance(merge_report, MergeReport)
+        else dict(merge_report or {})
+    )
+    lines = [
+        "并行/协作结果出现冲突或失败，我已切回保守收敛路径。",
+        f"原始请求：{user_query}",
+    ]
+    if completed:
+        lines.append("当前仍可确认的结果：")
+        for item in completed[:3]:
+            lines.append(f"- [{item.owner_agent_id or 'unknown'}] {item.summary or '已完成'}")
+    if merge_payload.get("conflict_flags"):
+        lines.append("冲突/回退原因：")
+        for flag in list(merge_payload.get("conflict_flags") or [])[:3]:
+            lines.append(f"- {flag}")
+    if not completed:
+        lines.append("目前没有足够稳定的子任务结果，建议改走 direct 或更小范围的 collaboration。")
+    return "\n".join(lines)
+
+
 def _verify_collaboration_results(
     state: AgentState,
     tasks: list[AgentTask],
@@ -2411,6 +2640,14 @@ def _verify_collaboration_results(
         }
         for item in normalized_results
     ]
+    merge_report = merge_task_results([item.model_dump(mode="json") for item in normalized_results])
+    state["merge_report"] = merge_report.model_dump(mode="json")
+    _append_event_trace(
+        state,
+        "agent.merge.completed",
+        payload=state["merge_report"],
+    )
+
     if missing_task_ids:
         summary = f"协作结果不完整，缺少任务结果: {', '.join(missing_task_ids)}"
         verdict = verify_task_result(
@@ -2426,20 +2663,51 @@ def _verify_collaboration_results(
         verdict = verify_task_result(
             status="failed", summary=summary, evidence=verification_evidence
         )
+    elif merge_report.status == "conflicted":
+        verdict = verify_task_result(
+            status="failed",
+            summary=merge_report.summary,
+            evidence=[
+                *verification_evidence,
+                {"type": "merge_conflict", "conflict_flags": merge_report.conflict_flags},
+            ],
+        )
+    elif merge_report.status == "fallback":
+        verdict = verify_task_result(
+            status="failed",
+            summary=merge_report.summary,
+            evidence=[
+                *verification_evidence,
+                {"type": "merge_fallback", "conflict_flags": merge_report.conflict_flags},
+            ],
+        )
     else:
-        summary = f"协作模式已完成 {len(normalized_results)}/{len(tasks)} 个子任务，并为每个子任务回收了结果与 evidence。"
+        summary = (
+            merge_report.summary
+            or f"协作模式已完成 {len(normalized_results)}/{len(tasks)} 个子任务，并为每个子任务回收了结果与 evidence。"
+        )
         verdict = verify_task_result(
             status="passed", summary=summary, evidence=verification_evidence
         )
 
     updated_state = apply_verification_verdict(state, verdict)
     state.update(updated_state)
+    state["verification_report"] = VerificationReport(
+        status=state.get("verification_status") or "skipped",
+        summary=state.get("verification_summary"),
+        evidence=list(state.get("verification_evidence") or []),
+    ).model_dump(mode="json")
+    _append_event_trace(
+        state,
+        "agent.verify.completed",
+        payload=state["verification_report"],
+    )
 
 
 async def _run_collaboration_flow(state: AgentState, user_query: str) -> AgentState:
     _set_phase(state, "phase_2_controlled_collaboration", reason="collaboration_flow_started")
     _record_checkpoint(state, "collaboration.tasks_planning", reason="collaboration_flow_started")
-    tasks = _build_collaboration_tasks(user_query)
+    tasks, scheduled_subtasks = _build_collaboration_plan_from_task_graph(state, user_query)
     if len(tasks) < 2:
         state["execution_mode"] = "direct"
         state["routing_decision"] = {"mode": "direct", "reason": "collaboration_plan_fell_back"}
@@ -2487,6 +2755,7 @@ async def _run_collaboration_flow(state: AgentState, user_query: str) -> AgentSt
         "agent.collaboration.budget.updated",
         payload=budget_snapshot,
     )
+    state["scheduled_subtasks"] = scheduled_subtasks
     state["active_tasks"] = [task.model_dump(mode="json") for task in tasks]
     _record_checkpoint(
         state, "collaboration.tasks_ready", reason="tasks_built", payload={"task_count": len(tasks)}
@@ -2500,6 +2769,21 @@ async def _run_collaboration_flow(state: AgentState, user_query: str) -> AgentSt
 
     _set_phase(state, "phase_3_dynamic_collaboration", reason="collaboration_workers_dispatch")
     for task in tasks:
+        scheduled_subtask = next(
+            (item for item in scheduled_subtasks if item.get("subtask_id") == task.task_id),
+            None,
+        )
+        if scheduled_subtask is not None:
+            _append_event_trace(
+                state,
+                "agent.subtask.started",
+                payload={
+                    "subtask_id": scheduled_subtask.get("subtask_id"),
+                    "role": scheduled_subtask.get("role"),
+                    "dependencies": scheduled_subtask.get("dependencies") or [],
+                },
+                task_id=task.task_id,
+            )
         _record_checkpoint(
             state,
             "collaboration.task_dispatch",
@@ -2583,6 +2867,17 @@ async def _run_collaboration_flow(state: AgentState, user_query: str) -> AgentSt
             child_task_id=(task.child_task_ids or [None])[0],
             message_id=str(state.get("last_message_id") or "") or None,
         )
+        if scheduled_subtask is not None:
+            _append_event_trace(
+                state,
+                "agent.subtask.completed",
+                payload={
+                    "subtask_id": scheduled_subtask.get("subtask_id"),
+                    "status": task_result.status,
+                    "summary": task_result.summary,
+                },
+                task_id=task.task_id,
+            )
         _apply_task_result_to_state(state, task, task_result)
 
         if task_result.status != "completed":
@@ -2618,6 +2913,26 @@ async def _run_collaboration_flow(state: AgentState, user_query: str) -> AgentSt
         },
         severity="error" if state.get("verification_status") == "failed" else "info",
     )
+    merge_report = state.get("merge_report") or {}
+    if state.get("verification_status") == "failed" and merge_report.get("fallback_used"):
+        state["final_response"] = _build_serial_fallback_response(
+            user_query,
+            state.get("task_results") or [],
+            merge_report,
+        )
+        state["routing_decision"] = {
+            "mode": "direct",
+            "reason": "fallback_to_serial_recovery",
+        }
+        _append_event_trace(
+            state,
+            "agent.rollback.triggered",
+            payload={
+                "layer": "collaboration_runtime",
+                "reason": "merge_fallback_used",
+            },
+            severity="warning",
+        )
     _record_checkpoint(
         state,
         "collaboration.completed",
@@ -2679,23 +2994,99 @@ async def master_node(state: AgentState) -> AgentState:
         return state
 
     state["current_agent"] = _normalize_current_agent(state.get("current_agent"))
+    parallel_worthiness = state.get("parallel_worthiness")
+    if not isinstance(parallel_worthiness, dict):
+        runtime_request_context = state.get("runtime_request_context") or {}
+        if isinstance(runtime_request_context, dict):
+            candidate_parallel = runtime_request_context.get("parallel_worthiness")
+            if isinstance(candidate_parallel, dict):
+                parallel_worthiness = candidate_parallel
+    if isinstance(parallel_worthiness, dict) and parallel_worthiness:
+        _append_event_trace(
+            state,
+            "agent.parallel.assessed",
+            payload=parallel_worthiness,
+        )
+    skill_shortlist = list(state.get("skill_shortlist") or [])
+    if skill_shortlist:
+        _append_event_trace(
+            state,
+            "agent.skill.shortlisted",
+            payload={
+                "count": len(skill_shortlist),
+                "skills": [str(item.get("skill_name") or "") for item in skill_shortlist[:4]],
+            },
+        )
+    task_graph = state.get("task_graph")
+    if isinstance(task_graph, dict) and task_graph.get("nodes"):
+        _append_event_trace(
+            state,
+            "agent.task_graph.built",
+            payload={
+                "graph_id": task_graph.get("graph_id"),
+                "node_count": len(task_graph.get("nodes") or []),
+                "entry_node_ids": task_graph.get("entry_node_ids") or [],
+                "max_parallelism": task_graph.get("max_parallelism"),
+            },
+        )
+    elif (
+        isinstance(parallel_worthiness, dict)
+        and parallel_worthiness.get("preferred_mode") in {"collaboration", "parallel"}
+        and not (state.get("feature_flags") or {}).get("ENABLE_PARALLEL_TASK_GRAPH", True)
+    ):
+        _append_event_trace(
+            state,
+            "agent.rollback.triggered",
+            payload={
+                "layer": "parallel_task_graph",
+                "reason": "feature_flag_disabled",
+            },
+        )
 
     structured_continuity_route = _route_from_structured_continuity(state, user_query)
     clarification_route = _route_from_clarification_context(state, user_query)
     if structured_continuity_route is not None:
         state["execution_mode"] = "direct"
         routed_agent = structured_continuity_route
+        _record_execution_decision(
+            state,
+            user_query=user_query,
+            mode="direct",
+            reason="continue_pending_action",
+            selected_roles=[routed_agent.value],
+        )
     elif clarification_route is not None:
         state["execution_mode"] = "direct"
         routed_agent = clarification_route
+        _record_execution_decision(
+            state,
+            user_query=user_query,
+            mode="direct",
+            reason="clarification_follow_up",
+            selected_roles=[routed_agent.value],
+        )
     elif _is_short_confirmation(user_query) and _previous_turn_proposed_schedule_creation(
         state.get("messages", [])
     ):
         state["execution_mode"] = "direct"
         routed_agent = AgentRole.SCHEDULE_PLANNER
+        _record_execution_decision(
+            state,
+            user_query=user_query,
+            mode="direct",
+            reason="schedule_confirmation_follow_up",
+            selected_roles=[routed_agent.value],
+        )
     else:
         request_mode, routing_metadata = _select_request_mode(user_query)
         state["routing_decision"] = routing_metadata
+        _record_execution_decision(
+            state,
+            user_query=user_query,
+            mode=request_mode,
+            reason=str(routing_metadata.get("reason") or request_mode),
+            selected_roles=list(routing_metadata.get("roles") or []),
+        )
         if request_mode == "collaboration":
             collaboration_state = await _run_collaboration_flow(state, user_query)
             if collaboration_state.get(

backend/app/agents/learning/__init__.py

@@ -0,0 +1,19 @@
+from app.agents.learning.jobs import persist_retrospective, schedule_retrospective_job
+from app.agents.learning.pattern_miner import LearningPatternMiner
+from app.agents.learning.retrospector import build_session_retrospective
+from app.agents.learning.session_search import SessionRetrospectiveSearch
+from app.agents.learning.signal_extractor import RetrospectiveSignalExtractor
+from app.agents.learning.skill_candidate_builder import SkillCandidateBuilder
+from app.agents.learning.store import LearningArtifactStore, SessionRetrospectiveStore
+
+__all__ = [
+    "build_session_retrospective",
+    "LearningArtifactStore",
+    "LearningPatternMiner",
+    "persist_retrospective",
+    "RetrospectiveSignalExtractor",
+    "schedule_retrospective_job",
+    "SessionRetrospectiveSearch",
+    "SessionRetrospectiveStore",
+    "SkillCandidateBuilder",
+]

backend/app/agents/learning/audit.py

@@ -0,0 +1,16 @@
+from __future__ import annotations
+
+from app.agents.schemas.learning import LearningDecision, SessionRetrospective
+
+
+def build_learning_audit_entry(retrospective: SessionRetrospective) -> dict[str, object]:
+    decision = retrospective.learning_decision
+    return {
+        "retrospective_id": retrospective.retrospective_id,
+        "decision": decision.decision if isinstance(decision, LearningDecision) else None,
+        "explanation": decision.explanation if isinstance(decision, LearningDecision) else None,
+        "signal_count": len(retrospective.learning_signals),
+        "pattern_count": len(retrospective.pattern_candidates),
+        "skill_candidate_count": len(retrospective.skill_candidates),
+        "outcome": retrospective.outcome,
+    }

backend/app/agents/learning/bridge.py

@@ -0,0 +1,45 @@
+from __future__ import annotations
+
+from app.agents.schemas.learning import LearningDecision, LearningSignal
+
+
+def route_learning_signal(signal: LearningSignal) -> str:
+    if signal.signal_type == "preference":
+        return "memory"
+    if signal.signal_type in {"workflow", "decomposition", "tool_success"}:
+        return "skill"
+    if signal.signal_type == "correction":
+        return "audit"
+    return "memory"
+
+
+def build_learning_bridge_summary(signals: list[LearningSignal]) -> dict[str, object]:
+    memory_count = 0
+    skill_count = 0
+    audit_count = 0
+
+    for signal in signals:
+        route = route_learning_signal(signal)
+        if route == "memory":
+            memory_count += 1
+        elif route == "skill":
+            skill_count += 1
+        else:
+            audit_count += 1
+
+    return {
+        "memory_signal_count": memory_count,
+        "skill_signal_count": skill_count,
+        "audit_signal_count": audit_count,
+    }
+
+
+def update_learning_decision_with_bridge(
+    decision: LearningDecision,
+    signals: list[LearningSignal],
+) -> LearningDecision:
+    bridge_summary = build_learning_bridge_summary(signals)
+    metadata = dict(decision.metadata or {})
+    metadata["bridge"] = bridge_summary
+    decision.metadata = metadata
+    return decision

backend/app/agents/learning/jobs.py

@@ -0,0 +1,222 @@
+from __future__ import annotations
+
+import asyncio
+import logging
+from typing import Any
+
+from app.config import settings
+from app.database import async_session
+from app.agents.learning.bridge import update_learning_decision_with_bridge
+from app.agents.learning.pattern_miner import LearningPatternMiner
+from app.agents.learning.audit import build_learning_audit_entry
+from app.agents.learning.retrospector import build_session_retrospective
+from app.agents.learning.signal_extractor import RetrospectiveSignalExtractor
+from app.agents.learning.skill_candidate_builder import SkillCandidateBuilder
+from app.agents.learning.store import LearningArtifactStore, SessionRetrospectiveStore
+from app.agents.schemas.learning import LearningDecision, SessionRetrospective
+from app.agents.skills.evaluator import SkillPromotionEvaluator
+
+logger = logging.getLogger(__name__)
+
+
+def _enrich_retrospective(retrospective: SessionRetrospective) -> SessionRetrospective:
+    signals = RetrospectiveSignalExtractor().extract(retrospective)
+    patterns = LearningPatternMiner().mine(signals)
+    skill_candidates = SkillCandidateBuilder().build(patterns)
+
+    decision = LearningDecision(
+        decision="create_candidate" if skill_candidates else ("reinforce_memory" if signals else "defer"),
+        explanation=(
+            "Retrospective produced reusable candidate skills."
+            if skill_candidates
+            else "Retrospective only reinforces memory-like evidence."
+            if signals
+            else "No stable signal was extracted from this retrospective."
+        ),
+        evidence_refs=(skill_candidates[0].evidence_refs if skill_candidates else retrospective.evidence_refs[:3]),
+        metadata={
+            "signal_count": len(signals),
+            "pattern_count": len(patterns),
+            "skill_candidate_count": len(skill_candidates),
+        },
+    )
+
+    retrospective.learning_signals = signals
+    retrospective.pattern_candidates = patterns
+    retrospective.skill_candidates = skill_candidates
+    retrospective.learning_decision = update_learning_decision_with_bridge(decision, signals)
+    return retrospective
+
+
+def _build_learning_artifacts(retrospective: SessionRetrospective) -> list[dict[str, object]]:
+    artifacts: list[dict[str, object]] = []
+    for signal in retrospective.learning_signals:
+        artifacts.append(
+            {
+                "artifact_type": "signal",
+                "artifact_key": signal.signal_type,
+                "summary_text": signal.explanation or signal.signal_type,
+                "payload": signal.model_dump(mode="json"),
+            }
+        )
+    for pattern in retrospective.pattern_candidates:
+        artifacts.append(
+            {
+                "artifact_type": "pattern_candidate",
+                "artifact_key": pattern.pattern_type,
+                "summary_text": pattern.description,
+                "payload": pattern.model_dump(mode="json"),
+            }
+        )
+    for candidate in retrospective.skill_candidates:
+        artifacts.append(
+            {
+                "artifact_type": "skill_candidate",
+                "artifact_key": candidate.name,
+                "summary_text": candidate.summary,
+                "payload": candidate.model_dump(mode="json"),
+            }
+        )
+    if retrospective.learning_decision is not None:
+        artifacts.append(
+            {
+                "artifact_type": "learning_decision",
+                "artifact_key": retrospective.learning_decision.decision,
+                "summary_text": retrospective.learning_decision.explanation,
+                "payload": retrospective.learning_decision.model_dump(mode="json"),
+            }
+        )
+        artifacts.append(
+            {
+                "artifact_type": "learning_audit",
+                "artifact_key": retrospective.retrospective_id or "retrospective",
+                "summary_text": retrospective.learning_decision.explanation,
+                "payload": build_learning_audit_entry(retrospective),
+            }
+        )
+    return artifacts
+
+
+def _build_lifecycle_artifacts(decisions: list) -> list[dict[str, object]]:
+    artifacts: list[dict[str, object]] = []
+    for decision in decisions:
+        artifacts.append(
+            {
+                "artifact_type": "skill_lifecycle_decision",
+                "artifact_key": getattr(decision, "skill_name", None) or "skill",
+                "summary_text": getattr(decision, "reason", ""),
+                "payload": decision.model_dump(mode="json"),
+            }
+        )
+    return artifacts
+
+
+async def persist_retrospective(
+    *,
+    user_id: str,
+    conversation_id: str,
+    request_message_id: str | None,
+    response_message_id: str | None,
+    query_text: str,
+    final_response: str | None,
+    state: dict[str, Any] | None,
+) -> None:
+    retrospective = build_session_retrospective(
+        request_id=response_message_id or request_message_id or conversation_id,
+        session_id=conversation_id,
+        user_query=query_text,
+        state=state,
+        runtime_context={"user_id": user_id},
+    )
+    retrospective = _enrich_retrospective(retrospective)
+
+    async with async_session() as session:
+        saved = await SessionRetrospectiveStore(session).save(retrospective)
+        lifecycle_decisions = []
+        if settings.ENABLE_SKILL_PROMOTION:
+            lifecycle_decisions = await SkillPromotionEvaluator(session).sync_retrospective(
+                user_id=user_id,
+                retrospective=retrospective,
+            )
+        if settings.ENABLE_LEARNING_SIGNALS:
+            await LearningArtifactStore(session).save_batch(
+                user_id=user_id,
+                conversation_id=conversation_id,
+                retrospective_id=saved.id,
+                artifacts=[
+                    *_build_learning_artifacts(retrospective),
+                    *_build_lifecycle_artifacts(lifecycle_decisions),
+                ],
+            )
+
+
+def schedule_retrospective_job(**kwargs) -> asyncio.Task[None] | None:
+    if not settings.ENABLE_RETROSPECTIVE:
+        return None
+    try:
+        task = asyncio.create_task(persist_retrospective(**kwargs))
+    except RuntimeError:
+        return None
+
+    def _handle_completion(done_task: asyncio.Task[None]) -> None:
+        try:
+            done_task.result()
+        except Exception:
+            logger.exception("retrospective_job_failed")
+
+    task.add_done_callback(_handle_completion)
+    return task
+
+
+def schedule_retrospective_learning_event(
+    *,
+    user_id: str,
+    conversation_id: str,
+    retrospective: SessionRetrospective,
+    session_factory=async_session,
+) -> asyncio.Task[None] | None:
+    if not settings.ENABLE_RETROSPECTIVE:
+        return None
+
+    async def _persist_existing() -> None:
+        async with session_factory() as session:
+            enriched = _enrich_retrospective(retrospective)
+            saved = await SessionRetrospectiveStore(session).save(enriched)
+            lifecycle_decisions = []
+            if settings.ENABLE_SKILL_PROMOTION:
+                lifecycle_decisions = await SkillPromotionEvaluator(session).sync_retrospective(
+                    user_id=user_id,
+                    retrospective=enriched,
+                )
+            if settings.ENABLE_LEARNING_SIGNALS:
+                await LearningArtifactStore(session).save_batch(
+                    user_id=user_id,
+                    conversation_id=conversation_id,
+                    retrospective_id=saved.id,
+                    artifacts=[
+                        *_build_learning_artifacts(enriched),
+                        *_build_lifecycle_artifacts(lifecycle_decisions),
+                    ],
+                )
+
+    try:
+        task = asyncio.create_task(_persist_existing())
+    except RuntimeError:
+        return None
+
+    def _handle_completion(done_task: asyncio.Task[None]) -> None:
+        try:
+            done_task.result()
+        except Exception:
+            logger.exception(
+                "retrospective_learning_event_failed",
+                extra={
+                    "details": {
+                        "user_id": user_id,
+                        "conversation_id": conversation_id,
+                    }
+                },
+            )
+
+    task.add_done_callback(_handle_completion)
+    return task

backend/app/agents/learning/pattern_miner.py

@@ -0,0 +1,42 @@
+from __future__ import annotations
+
+from uuid import uuid4
+
+from app.agents.schemas.learning import LearningSignal, PatternCandidate
+
+
+class LearningPatternMiner:
+    def mine(self, signals: list[LearningSignal]) -> list[PatternCandidate]:
+        patterns: list[PatternCandidate] = []
+
+        for signal in signals:
+            if signal.signal_type not in {"workflow", "decomposition", "preference"}:
+                continue
+
+            description = self._build_description(signal)
+            patterns.append(
+                PatternCandidate(
+                    pattern_id=f"pattern-{uuid4().hex[:10]}",
+                    pattern_type=signal.signal_type,
+                    description=description,
+                    confidence=signal.confidence,
+                    evidence_refs=signal.evidence_refs[:4],
+                )
+            )
+
+        return patterns
+
+    @staticmethod
+    def _build_description(signal: LearningSignal) -> str:
+        payload = signal.payload or {}
+        if signal.signal_type == "workflow":
+            task_type = payload.get("task_type") or "general"
+            execution_mode = payload.get("execution_mode") or "direct"
+            return f"Completed {task_type} requests worked under {execution_mode} execution."
+        if signal.signal_type == "decomposition":
+            task_count = payload.get("task_count") or 0
+            return f"Requests with {task_count} concrete task refs benefit from structured decomposition."
+        if signal.signal_type == "preference":
+            preference = payload.get("preference") or "structured response"
+            return f"User preference repeatedly points to {preference}."
+        return signal.explanation or signal.signal_type

backend/app/agents/learning/retrospector.py

@@ -0,0 +1,115 @@
+from __future__ import annotations
+
+from typing import Any
+
+from app.agents.schemas.learning import SessionRetrospective
+
+
+def _classify_task_type(query_text: str) -> str:
+    normalized = (query_text or "").lower()
+    if any(token in normalized for token in ("总结", "分析", "对比", "report", "analyze")):
+        return "analysis"
+    if any(token in normalized for token in ("安排", "提醒", "日程", "todo", "task")):
+        return "planning_or_execution"
+    if any(token in normalized for token in ("文档", "资料", "年报", "search", "查")):
+        return "retrieval"
+    return "general"
+
+
+def build_session_retrospective(
+    *,
+    request_id: str,
+    session_id: str,
+    user_query: str,
+    state: dict[str, Any] | None,
+    runtime_context: dict[str, Any] | None = None,
+) -> SessionRetrospective:
+    state = state or {}
+    if hasattr(runtime_context, "model_dump"):
+        runtime_context = runtime_context.model_dump(mode="json")
+    runtime_context = runtime_context or {}
+    skill_shortlist = state.get("skill_shortlist") or []
+    used_skill_names = [
+        item.get("skill_name")
+        for item in skill_shortlist
+        if isinstance(item, dict) and item.get("skill_name")
+    ]
+
+    task_refs = []
+    for task in (state.get("completed_tasks") or [])[:4]:
+        if isinstance(task, dict):
+            task_refs.append(
+                {
+                    "task_id": task.get("task_id"),
+                    "title": task.get("title"),
+                    "status": task.get("status"),
+                }
+            )
+
+    event_refs = []
+    for event in (state.get("event_trace") or [])[:8]:
+        if isinstance(event, dict):
+            event_refs.append(
+                {
+                    "event_type": event.get("event_type"),
+                    "task_id": event.get("task_id"),
+                    "agent_id": event.get("agent_id"),
+                }
+            )
+
+    verification_evidence = []
+    for evidence in (state.get("verification_evidence") or [])[:6]:
+        if isinstance(evidence, dict):
+            verification_evidence.append(evidence)
+
+    verification_status = state.get("verification_status")
+    execution_mode = state.get("execution_mode")
+    primary_agent = state.get("current_agent") or "master"
+    retrospective_shortlist = state.get("retrospective_shortlist") or []
+
+    summary_parts = [
+        f"本轮请求按 {execution_mode or 'unknown'} 模式处理",
+        f"主要负责 agent 为 {primary_agent}",
+    ]
+    if verification_status:
+        summary_parts.append(f"验证结果为 {verification_status}")
+    if used_skill_names:
+        summary_parts.append(f"命中技能候选 {', '.join(used_skill_names[:3])}")
+    if retrospective_shortlist:
+        summary_parts.append(f"参考了 {len(retrospective_shortlist)} 条历史复盘")
+
+    final_response = state.get("final_response")
+    outcome = "completed" if final_response else "failed"
+    if not final_response and verification_status == "passed":
+        outcome = "completed"
+    if final_response and verification_status == "skipped":
+        outcome = "partial"
+
+    return SessionRetrospective(
+        retrospective_id=request_id,
+        user_id=str(runtime_context.get("user_id") or ""),
+        conversation_id=session_id,
+        response_message_id=request_id,
+        query_text=user_query,
+        final_response=final_response,
+        summary="；".join(summary_parts) + "。",
+        task_type=_classify_task_type(user_query),
+        execution_mode=execution_mode,
+        primary_agent=primary_agent,
+        verification_status=verification_status,
+        verification_summary=state.get("verification_summary"),
+        used_skill_names=used_skill_names,
+        evidence_refs=verification_evidence,
+        task_refs=task_refs,
+        event_refs=event_refs,
+        context_snapshot={
+            "runtime_request_context": runtime_context,
+            "recommended_runtime_mode": runtime_context.get("recommended_runtime_mode"),
+            "parallel_worthiness": state.get("parallel_worthiness"),
+            "retrospective_shortlist_count": len(retrospective_shortlist),
+            "scheduled_subtask_count": len(state.get("scheduled_subtasks") or []),
+            "merge_report": dict(state.get("merge_report") or {}),
+            "verification_report": dict(state.get("verification_report") or {}),
+        },
+        outcome=outcome,
+    )

backend/app/agents/learning/session_search.py

@@ -0,0 +1,95 @@
+from __future__ import annotations
+
+from app.agents.schemas.learning import SessionRetrospective
+from app.agents.skills.matcher import score_text_match
+from app.agents.learning.store import SessionRetrospectiveStore
+from app.config import settings
+
+
+class SessionRetrospectiveSearch:
+    def __init__(self, db):
+        self.db = db
+
+    async def shortlist(
+        self,
+        *,
+        user_id: str,
+        query_text: str,
+        conversation_id: str | None = None,
+        task_type: str | None = None,
+        skill_name: str | None = None,
+        limit: int = 3,
+    ) -> list[SessionRetrospective]:
+        records = await SessionRetrospectiveStore(self.db).list_recent(user_id=user_id, limit=25)
+        scored: list[tuple[float, SessionRetrospective]] = []
+
+        for record in records:
+            if task_type and record.task_type != task_type:
+                continue
+            if skill_name and skill_name not in (record.skill_names or []):
+                continue
+            score, _matched_terms = score_text_match(
+                query_text,
+                record.query_text,
+                record.summary_text,
+                " ".join(record.skill_names or []),
+            )
+            if conversation_id and record.conversation_id == conversation_id:
+                score = min(1.0, score + 0.1)
+            if score <= 0:
+                continue
+
+            payload = dict(record.payload or {})
+            payload["retrospective_id"] = record.id
+            retrospective = SessionRetrospective.model_validate(payload)
+            scored.append((score, retrospective))
+
+        scored.sort(key=lambda item: item[0], reverse=True)
+        return [item for _score, item in scored[:limit]]
+
+
+async def search_recent_retrospectives(
+    db,
+    *,
+    user_id: str,
+    query: str,
+    conversation_id: str | None = None,
+    task_type: str | None = None,
+    skill_name: str | None = None,
+    limit: int = 3,
+) -> list[SessionRetrospective]:
+    if not settings.ENABLE_SESSION_RETROSPECTIVE_SEARCH:
+        return []
+    return await SessionRetrospectiveSearch(db).shortlist(
+        user_id=user_id,
+        query_text=query,
+        conversation_id=conversation_id,
+        task_type=task_type,
+        skill_name=skill_name,
+        limit=limit,
+    )
+
+
+def summarize_retrospective(retrospective: SessionRetrospective) -> dict[str, object]:
+    verification_status = retrospective.verification_status or retrospective.outcome
+    success_score = 1.0 if verification_status == "passed" else 0.6 if verification_status == "skipped" else 0.2
+    reusable_patterns = []
+    if retrospective.used_skill_names:
+        reusable_patterns.append("skill_shortlist_hit")
+    if retrospective.execution_mode:
+        reusable_patterns.append(f"mode:{retrospective.execution_mode}")
+
+    avoid_patterns = []
+    if retrospective.outcome == "failed":
+        avoid_patterns.append("failed_outcome")
+
+    return {
+        "retrospective_id": retrospective.retrospective_id,
+        "task_type": retrospective.task_type,
+        "request_summary": retrospective.query_text[:120],
+        "summary": retrospective.summary,
+        "execution_mode": retrospective.execution_mode,
+        "success_score": round(success_score, 2),
+        "reusable_patterns": reusable_patterns,
+        "avoid_patterns": avoid_patterns,
+    }

backend/app/agents/learning/signal_extractor.py

@@ -0,0 +1,72 @@
+from __future__ import annotations
+
+from app.agents.schemas.learning import LearningSignal, SessionRetrospective
+
+
+class RetrospectiveSignalExtractor:
+    def extract(self, retrospective: SessionRetrospective) -> list[LearningSignal]:
+        signals: list[LearningSignal] = []
+
+        if retrospective.outcome == "completed":
+            signals.append(
+                LearningSignal(
+                    signal_type="workflow",
+                    confidence=0.8,
+                    evidence_refs=retrospective.evidence_refs[:3],
+                    explanation="Completed runs can be mined as workflow hints later.",
+                    payload={
+                        "task_type": retrospective.task_type,
+                        "execution_mode": retrospective.execution_mode,
+                    },
+                )
+            )
+
+        if len(retrospective.task_refs) > 1:
+            context_snapshot = retrospective.context_snapshot or {}
+            merge_report = dict(context_snapshot.get("merge_report") or {})
+            verification_report = dict(context_snapshot.get("verification_report") or {})
+            effectiveness_score = 1.0
+            if merge_report.get("status") == "conflicted":
+                effectiveness_score = 0.45
+            elif merge_report.get("status") == "fallback":
+                effectiveness_score = 0.25
+            elif verification_report.get("status") == "failed":
+                effectiveness_score = 0.3
+            signals.append(
+                LearningSignal(
+                    signal_type="decomposition",
+                    confidence=0.7,
+                    evidence_refs=retrospective.task_refs[:3],
+                    explanation="Multiple completed task refs indicate a decomposition pattern.",
+                    payload={
+                        "task_count": len(retrospective.task_refs),
+                        "scheduled_subtask_count": context_snapshot.get("scheduled_subtask_count", 0),
+                        "effectiveness_score": effectiveness_score,
+                        "merge_status": merge_report.get("status"),
+                    },
+                )
+            )
+
+        if retrospective.used_skill_names:
+            signals.append(
+                LearningSignal(
+                    signal_type="tool_success",
+                    confidence=0.65 if retrospective.outcome == "completed" else 0.35,
+                    evidence_refs=retrospective.evidence_refs[:2],
+                    explanation="Task-scoped skill shortlist was available during this run.",
+                    payload={"skills": retrospective.used_skill_names[:3]},
+                )
+            )
+
+        if retrospective.outcome == "failed":
+            signals.append(
+                LearningSignal(
+                    signal_type="correction",
+                    confidence=0.75,
+                    evidence_refs=retrospective.evidence_refs[:2],
+                    explanation="Failed retrospectives should remain auditable before any promotion.",
+                    payload={"verification_status": retrospective.verification_status},
+                )
+            )
+
+        return signals

backend/app/agents/learning/skill_candidate_builder.py

@@ -0,0 +1,54 @@
+from __future__ import annotations
+
+import hashlib
+
+from app.agents.schemas.learning import PatternCandidate, SkillCandidate
+
+
+class SkillCandidateBuilder:
+    def build(self, patterns: list[PatternCandidate]) -> list[SkillCandidate]:
+        candidates: list[SkillCandidate] = []
+
+        for pattern in patterns:
+            if pattern.confidence < 0.55:
+                continue
+
+            name = self._build_name(pattern)
+            candidates.append(
+                SkillCandidate(
+                    candidate_id=f"candidate-{self._stable_suffix(pattern)}",
+                    name=name,
+                    summary=pattern.description,
+                    candidate_type=self._map_candidate_type(pattern.pattern_type),
+                    source_pattern_ids=[pattern.pattern_id],
+                    confidence=pattern.confidence,
+                    evidence_refs=pattern.evidence_refs[:4],
+                    recommended_status="candidate",
+                )
+            )
+
+        return candidates
+
+    @staticmethod
+    def _build_name(pattern: PatternCandidate) -> str:
+        prefix = {
+            "workflow": "workflow",
+            "decomposition": "decomposition",
+            "preference": "preference",
+        }.get(pattern.pattern_type, "learned")
+        stable_suffix = SkillCandidateBuilder._stable_suffix(pattern)
+        return f"{prefix}-{stable_suffix}"
+
+    @staticmethod
+    def _map_candidate_type(pattern_type: str) -> str:
+        mapping = {
+            "workflow": "workflow_skill",
+            "decomposition": "decomposition_skill",
+            "preference": "preference_skill",
+        }
+        return mapping.get(pattern_type, "workflow_skill")
+
+    @staticmethod
+    def _stable_suffix(pattern: PatternCandidate) -> str:
+        raw = f"{pattern.pattern_type}:{pattern.description}".encode("utf-8")
+        return hashlib.sha1(raw).hexdigest()[:10]

backend/app/agents/learning/store.py

@@ -0,0 +1,129 @@
+from __future__ import annotations
+
+from sqlalchemy import desc, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.agents.schemas.learning import SessionRetrospective
+from app.models.learning import LearningArtifactRecord, SessionRetrospectiveRecord
+
+
+class SessionRetrospectiveStore:
+    def __init__(self, db: AsyncSession):
+        self.db = db
+
+    async def save(self, retrospective: SessionRetrospective) -> SessionRetrospectiveRecord:
+        payload = retrospective.model_dump(mode="json")
+        record = SessionRetrospectiveRecord(
+            user_id=retrospective.user_id,
+            conversation_id=retrospective.conversation_id,
+            request_message_id=retrospective.request_message_id,
+            response_message_id=retrospective.response_message_id,
+            query_text=retrospective.query_text,
+            final_response=retrospective.final_response,
+            summary_text=retrospective.summary,
+            task_type=retrospective.task_type,
+            execution_mode=retrospective.execution_mode,
+            primary_agent=retrospective.primary_agent,
+            verification_status=retrospective.verification_status,
+            verification_summary=retrospective.verification_summary,
+            skill_names=retrospective.used_skill_names,
+            evidence=retrospective.evidence_refs,
+            task_refs=retrospective.task_refs,
+            payload=payload,
+        )
+        self.db.add(record)
+        await self.db.commit()
+        await self.db.refresh(record)
+        return record
+
+    async def list_recent(
+        self,
+        *,
+        user_id: str,
+        limit: int = 20,
+    ) -> list[SessionRetrospectiveRecord]:
+        result = await self.db.execute(
+            select(SessionRetrospectiveRecord)
+            .where(SessionRetrospectiveRecord.user_id == user_id)
+            .order_by(desc(SessionRetrospectiveRecord.recorded_at), desc(SessionRetrospectiveRecord.created_at))
+            .limit(limit)
+        )
+        return list(result.scalars().all())
+
+
+class LearningArtifactStore:
+    def __init__(self, db: AsyncSession):
+        self.db = db
+
+    async def save_batch(
+        self,
+        *,
+        user_id: str,
+        conversation_id: str,
+        retrospective_id: str | None,
+        artifacts: list[dict[str, object]],
+    ) -> list[LearningArtifactRecord]:
+        records: list[LearningArtifactRecord] = []
+        for artifact in artifacts:
+            record = LearningArtifactRecord(
+                user_id=user_id,
+                conversation_id=conversation_id,
+                retrospective_id=retrospective_id,
+                artifact_type=str(artifact.get("artifact_type") or "unknown"),
+                artifact_key=str(artifact.get("artifact_key") or "") or None,
+                summary_text=str(artifact.get("summary_text") or ""),
+                payload=dict(artifact.get("payload") or {}),
+            )
+            self.db.add(record)
+            records.append(record)
+
+        await self.db.commit()
+        for record in records:
+            await self.db.refresh(record)
+        return records
+
+    async def list_recent(
+        self,
+        *,
+        user_id: str,
+        artifact_type: str | None = None,
+        limit: int = 50,
+    ) -> list[LearningArtifactRecord]:
+        query = select(LearningArtifactRecord).where(LearningArtifactRecord.user_id == user_id)
+        if artifact_type:
+            query = query.where(LearningArtifactRecord.artifact_type == artifact_type)
+        result = await self.db.execute(
+            query.order_by(
+                desc(LearningArtifactRecord.recorded_at),
+                desc(LearningArtifactRecord.created_at),
+            ).limit(limit)
+        )
+        return list(result.scalars().all())
+
+    async def aggregate_counts_by_key(
+        self,
+        *,
+        user_id: str,
+        artifact_type: str,
+        limit: int = 100,
+    ) -> dict[str, int]:
+        records = await self.list_recent(user_id=user_id, artifact_type=artifact_type, limit=limit)
+        counts: dict[str, int] = {}
+        for record in records:
+            key = record.artifact_key or "unknown"
+            counts[key] = counts.get(key, 0) + 1
+        return counts
+
+
+def append_retrospective_attachment(
+    attachments: list[dict] | None,
+    retrospective: SessionRetrospective,
+) -> list[dict]:
+    next_attachments = list(attachments or [])
+    next_attachments.append(
+        {
+            "kind": "session_retrospective",
+            "payload": retrospective.model_dump(mode="json"),
+        }
+    )
+    return next_attachments

backend/app/agents/orchestration/__init__.py

@@ -1,7 +1,16 @@
 """高级编排系统 - Phase 10"""
 
+from app.agents.orchestration.budget import build_subtask_budget
+from app.agents.orchestration.result_merge import merge_task_results
+from app.agents.orchestration.scheduler import (
+    ParallelExecutionScheduler,
+    build_subtask_specs,
+    ensure_child_links,
+)
+from app.agents.orchestration.subagent_runtime import subtask_spec_to_agent_task
 from app.agents.team.leader import TeamLeader, TeamTask, TaskStatus
 from app.agents.transport.remote import RemoteTransport, StructuredMessage
+from app.agents.orchestration.task_graph import build_bounded_task_graph, render_task_graph_summary
 from app.agents.background.manager import (
     BackgroundTaskManager,
     BackgroundTask,
@@ -14,7 +23,15 @@ __all__ = [
     "TaskStatus",
     "RemoteTransport",
     "StructuredMessage",
+    "ParallelExecutionScheduler",
+    "build_bounded_task_graph",
+    "build_subtask_budget",
+    "build_subtask_specs",
     "BackgroundTaskManager",
     "BackgroundTask",
+    "ensure_child_links",
     "get_background_task_manager",
+    "merge_task_results",
+    "render_task_graph_summary",
+    "subtask_spec_to_agent_task",
 ]

backend/app/agents/orchestration/budget.py

@@ -0,0 +1,24 @@
+from __future__ import annotations
+
+from app.agents.schemas.task import CollaborationBudget
+
+
+def build_subtask_budget(
+    *,
+    execution_mode: str,
+    max_parallel_tasks: int,
+    max_tool_calls: int = 2,
+    max_iterations: int = 2,
+    metadata: dict | None = None,
+) -> CollaborationBudget:
+    return CollaborationBudget(
+        mode="collaboration" if execution_mode != "direct" else "direct",
+        max_parallel_tasks=max_parallel_tasks,
+        remaining_parallel_tasks=max_parallel_tasks,
+        max_tool_calls=max_tool_calls,
+        remaining_tool_calls=max_tool_calls,
+        max_iterations=max_iterations,
+        remaining_iterations=max_iterations,
+        escalation_threshold=1,
+        metadata=metadata or {},
+    )

backend/app/agents/orchestration/monitor.py

@@ -0,0 +1,31 @@
+from __future__ import annotations
+
+from typing import Any
+
+
+def build_parallel_runtime_metrics(
+    *,
+    task_graph: dict[str, Any] | None,
+    scheduled_subtasks: list[dict[str, Any]] | None,
+    task_results: list[dict[str, Any]] | None,
+    merge_report: dict[str, Any] | None,
+) -> dict[str, Any]:
+    task_graph = task_graph or {}
+    scheduled_subtasks = list(scheduled_subtasks or [])
+    task_results = list(task_results or [])
+    merge_report = merge_report or {}
+
+    completed = sum(1 for item in task_results if item.get("status") == "completed")
+    failed = sum(1 for item in task_results if item.get("status") == "failed")
+    blocked = sum(1 for item in task_results if item.get("status") == "blocked")
+
+    return {
+        "task_graph_node_count": len(task_graph.get("nodes") or []),
+        "scheduled_subtask_count": len(scheduled_subtasks),
+        "completed_subtask_count": completed,
+        "failed_subtask_count": failed,
+        "blocked_subtask_count": blocked,
+        "merge_status": merge_report.get("status"),
+        "merge_conflict_count": len(merge_report.get("conflict_flags") or []),
+        "fallback_used": bool(merge_report.get("fallback_used") or False),
+    }

backend/app/agents/orchestration/result_merge.py

@@ -0,0 +1,69 @@
+from __future__ import annotations
+
+from app.agents.schemas.orchestration import MergeReport
+from app.agents.verifier import normalize_task_result
+
+
+def merge_task_results(task_results: list[dict] | list[object]) -> MergeReport:
+    normalized = [normalize_task_result(item) for item in (task_results or [])]
+    completed = [item for item in normalized if item.status == "completed"]
+    failed_or_blocked = [item for item in normalized if item.status in {"failed", "blocked"}]
+
+    evidence_union: list[dict] = []
+    summaries = []
+    for item in normalized:
+        evidence_union.extend(list(item.evidence or []))
+        if item.summary:
+            summaries.append(item.summary.strip())
+
+    unique_summaries = list(dict.fromkeys(summary for summary in summaries if summary))
+    conflict_flags: list[str] = []
+    status = "merged"
+    fallback_used = False
+
+    if failed_or_blocked:
+        status = "fallback"
+        fallback_used = True
+        conflict_flags.append(
+            "failed_or_blocked_tasks:" + ",".join(item.task_id for item in failed_or_blocked)
+        )
+        resolution_strategy = "serial_recovery"
+        resolved_summary = (
+            completed[-1].summary
+            if completed and completed[-1].summary
+            else None
+        )
+    elif len(unique_summaries) > 1 and len(completed) > 1:
+        status = "conflicted"
+        conflict_flags.append("multiple_distinct_completed_summaries")
+        resolution_strategy = "rank_by_evidence_count"
+        ranked = sorted(
+            completed,
+            key=lambda item: (len(item.evidence or []), bool(item.summary)),
+            reverse=True,
+        )
+        resolved_summary = ranked[0].summary if ranked and ranked[0].summary else None
+    else:
+        resolution_strategy = "evidence_union"
+        resolved_summary = unique_summaries[-1] if unique_summaries else None
+
+    if status == "merged":
+        summary = (
+            unique_summaries[-1]
+            if unique_summaries
+            else f"已收敛 {len(normalized)} 个子任务结果。"
+        )
+    elif status == "conflicted":
+        summary = "并行子任务摘要存在冲突，需要 verifier 或串行收敛。"
+    else:
+        summary = "存在失败或阻塞子任务，需要回退到更保守的收敛路径。"
+
+    return MergeReport(
+        status=status,
+        summary=summary,
+        evidence_union=evidence_union,
+        conflict_flags=conflict_flags,
+        resolution_strategy=resolution_strategy,
+        resolved_summary=resolved_summary,
+        fallback_used=fallback_used,
+    )

backend/app/agents/orchestration/scheduler.py

@@ -0,0 +1,93 @@
+from __future__ import annotations
+
+from collections import defaultdict, deque
+from uuid import uuid4
+
+from app.agents.orchestration.budget import build_subtask_budget
+from app.agents.schemas.orchestration import SubTaskSpec, TaskGraph, TaskNode
+
+
+class ParallelExecutionScheduler:
+    def plan(self, task_graph: TaskGraph, *, query_text: str) -> list[SubTaskSpec]:
+        ordered_nodes = _topological_nodes(task_graph)
+        specs: list[SubTaskSpec] = []
+        for node in ordered_nodes:
+            budget = build_subtask_budget(
+                execution_mode=node.execution_mode,
+                max_parallel_tasks=max(1, task_graph.max_parallelism),
+                metadata={
+                    "task_graph_id": task_graph.graph_id,
+                    "depends_on": node.depends_on,
+                },
+            )
+            specs.append(
+                SubTaskSpec(
+                    subtask_id=node.node_id,
+                    parent_run_id=task_graph.graph_id,
+                    title=node.title,
+                    role=node.role or "master",
+                    goal=node.goal or query_text,
+                    context_slice=_build_context_slice(node, query_text),
+                    allowed_tools=[],
+                    budget_tokens=1200,
+                    budget_tool_calls=budget.max_tool_calls or 2,
+                    expected_output_schema={
+                        "summary": "string",
+                        "evidence": "list",
+                        "status": "completed|failed|blocked",
+                    },
+                    expected_evidence=node.expected_evidence,
+                    dependencies=node.depends_on,
+                )
+            )
+        return specs
+
+
+def build_subtask_specs(task_graph: TaskGraph, *, query_text: str) -> list[SubTaskSpec]:
+    return ParallelExecutionScheduler().plan(task_graph, query_text=query_text)
+
+
+def _build_context_slice(node: TaskNode, query_text: str) -> dict[str, object]:
+    return {
+        "query": query_text,
+        "role": node.role,
+        "title": node.title,
+        "goal": node.goal,
+        "depends_on": node.depends_on,
+    }
+
+
+def _topological_nodes(task_graph: TaskGraph) -> list[TaskNode]:
+    by_id = {node.node_id: node for node in task_graph.nodes}
+    indegree = {node.node_id: 0 for node in task_graph.nodes}
+    edges: dict[str, list[str]] = defaultdict(list)
+
+    for node in task_graph.nodes:
+        for dep in node.depends_on:
+            if dep not in by_id:
+                continue
+            edges[dep].append(node.node_id)
+            indegree[node.node_id] += 1
+
+    ready = deque(node_id for node_id, count in indegree.items() if count == 0)
+    ordered: list[TaskNode] = []
+
+    while ready:
+        node_id = ready.popleft()
+        ordered.append(by_id[node_id])
+        for target in edges.get(node_id, []):
+            indegree[target] -= 1
+            if indegree[target] == 0:
+                ready.append(target)
+
+    if len(ordered) != len(task_graph.nodes):
+        return list(task_graph.nodes)
+    return ordered
+
+
+def ensure_child_links(specs: list[SubTaskSpec]) -> dict[str, list[str]]:
+    graph: dict[str, list[str]] = defaultdict(list)
+    for spec in specs:
+        for dep in spec.dependencies:
+            graph[dep].append(spec.subtask_id)
+    return dict(graph)

backend/app/agents/orchestration/subagent_runtime.py

@@ -0,0 +1,17 @@
+from __future__ import annotations
+
+from app.agents.schemas.orchestration import SubTaskSpec
+from app.agents.schemas.task import AgentTask
+
+
+def subtask_spec_to_agent_task(spec: SubTaskSpec) -> AgentTask:
+    return AgentTask(
+        task_id=spec.subtask_id,
+        title=spec.title,
+        owner_agent_id=spec.role,
+        role=spec.role,
+        goal=spec.goal,
+        parent_task_id=spec.parent_run_id,
+        child_task_ids=[],
+        expected_evidence=spec.expected_evidence,
+    )

backend/app/agents/orchestration/task_graph.py

@@ -0,0 +1,128 @@
+from __future__ import annotations
+
+from uuid import uuid4
+
+from app.agents.schemas.orchestration import ParallelWorthiness, TaskGraph, TaskNode
+
+
+ROLE_KEYWORDS: list[tuple[str, tuple[str, ...]]] = [
+    ("librarian", ("查", "检索", "资料", "文档", "知识库", "年报", "forum", "search")),
+    ("analyst", ("分析", "判断", "风险", "总结", "对比", "洞察", "结论")),
+    ("schedule_planner", ("计划", "安排", "下周", "日程", "提醒", "优先级")),
+    ("executor", ("执行", "创建", "更新", "落库", "提交", "发帖")),
+]
+
+
+def build_bounded_task_graph(
+    *,
+    query_text: str,
+    parallel_worthiness: ParallelWorthiness,
+    max_nodes: int = 4,
+) -> TaskGraph | None:
+    roles = _infer_roles(query_text)
+    if not roles:
+        return None
+
+    independent_roles = roles[: min(max_nodes - 1, max(1, parallel_worthiness.estimated_subtasks))]
+    nodes: list[TaskNode] = []
+
+    for index, role in enumerate(independent_roles, start=1):
+        node_id = f"task-{index}-{uuid4().hex[:6]}"
+        nodes.append(
+            TaskNode(
+                node_id=node_id,
+                title=_build_title(role),
+                role=role,
+                goal=_build_goal(role, query_text),
+                depends_on=[],
+                execution_mode=(
+                    "parallel"
+                    if parallel_worthiness.preferred_mode in {"collaboration", "parallel"}
+                    and len(independent_roles) > 1
+                    else "serial"
+                ),
+                expected_evidence=_build_expected_evidence(role),
+            )
+        )
+
+    if len(nodes) > 1:
+        merge_id = f"merge-{uuid4().hex[:6]}"
+        nodes.append(
+            TaskNode(
+                node_id=merge_id,
+                title="汇总并收敛最终结论",
+                role="master",
+                goal="汇总前置子任务结果，形成统一可验证的输出。",
+                depends_on=[node.node_id for node in nodes],
+                execution_mode="serial",
+                expected_evidence=[{"type": "merge", "detail": "merged summary and conflict notes"}],
+            )
+        )
+
+    return TaskGraph(
+        nodes=nodes,
+        entry_node_ids=[node.node_id for node in nodes if not node.depends_on],
+        max_parallelism=max(1, len(independent_roles)),
+        rationale=_build_rationale(parallel_worthiness, independent_roles),
+    )
+
+
+def render_task_graph_summary(task_graph: TaskGraph | None) -> str | None:
+    if task_graph is None or not task_graph.nodes:
+        return None
+
+    lines = ["- 任务图:"]
+    for node in task_graph.nodes[:4]:
+        deps = f" deps={','.join(node.depends_on)}" if node.depends_on else ""
+        lines.append(f"  - [{node.execution_mode}] {node.title} ({node.role}){deps}")
+    return "\n".join(lines)
+
+
+def _infer_roles(query_text: str) -> list[str]:
+    selected: list[str] = []
+    text = query_text or ""
+    for role, keywords in ROLE_KEYWORDS:
+        if any(keyword in text for keyword in keywords):
+            selected.append(role)
+
+    if not selected:
+        return ["analyst"]
+    return selected
+
+
+def _build_title(role: str) -> str:
+    mapping = {
+        "librarian": "收集事实与外部/内部证据",
+        "analyst": "形成判断与风险分析",
+        "schedule_planner": "整理计划和优先级",
+        "executor": "执行必要操作并回收结果",
+    }
+    return mapping.get(role, "处理子任务")
+
+
+def _build_goal(role: str, query_text: str) -> str:
+    mapping = {
+        "librarian": f"围绕请求收集支持结论的事实和资料：{query_text}",
+        "analyst": f"基于当前请求输出结构化判断：{query_text}",
+        "schedule_planner": f"把当前请求收束为计划、安排或优先级：{query_text}",
+        "executor": f"基于请求执行必要动作并返回结果：{query_text}",
+    }
+    return mapping.get(role, query_text)
+
+
+def _build_expected_evidence(role: str) -> list[dict[str, str]]:
+    mapping = {
+        "librarian": [{"type": "evidence", "detail": "retrieval findings"}],
+        "analyst": [{"type": "analysis", "detail": "structured judgment"}],
+        "schedule_planner": [{"type": "plan", "detail": "explicit schedule or priorities"}],
+        "executor": [{"type": "execution", "detail": "tool output or mutation result"}],
+    }
+    return mapping.get(role, [{"type": "summary", "detail": "task summary"}])
+
+
+def _build_rationale(parallel_worthiness: ParallelWorthiness, roles: list[str]) -> str:
+    return (
+        f"preferred_mode={parallel_worthiness.preferred_mode}; "
+        f"score={parallel_worthiness.score:.2f}; "
+        f"roles={','.join(roles)}"
+    )

backend/app/agents/schemas/__init__.py

@@ -1,5 +1,24 @@
 from app.agents.schemas.event import AgentEvent
+from app.agents.schemas.learning import (
+    LearningDecision,
+    LearningSignal,
+    PatternCandidate,
+    SessionRetrospective,
+    SkillCandidate,
+)
 from app.agents.schemas.message import AgentMessage
+from app.agents.schemas.orchestration import (
+    ExecutionDecision,
+    MergeReport,
+    ParallelWorthiness,
+    RuntimeRequestContext,
+    SubTaskResult,
+    SubTaskSpec,
+    TaskGraph,
+    TaskNode,
+    VerificationReport,
+)
+from app.agents.schemas.skills import SkillActivationRecord, SkillShortlistEntry
 from app.agents.schemas.task import (
     AgentTask,
     CollaborationBudget,
@@ -14,12 +33,28 @@ from app.agents.schemas.task import (
 __all__ = [
     "AgentEvent",
     "AgentMessage",
+    "ExecutionDecision",
     "AgentTask",
     "CollaborationBudget",
     "InterruptRecord",
+    "LearningDecision",
+    "LearningSignal",
+    "MergeReport",
+    "ParallelWorthiness",
+    "PatternCandidate",
     "RecoveryRecord",
+    "RuntimeRequestContext",
+    "SessionRetrospective",
+    "SkillActivationRecord",
+    "SkillCandidate",
+    "SkillShortlistEntry",
+    "SubTaskResult",
+    "SubTaskSpec",
+    "TaskGraph",
+    "TaskNode",
     "TaskLifecycleStatus",
     "TaskResult",
     "TaskResultStatus",
+    "VerificationReport",
     "VerificationStatus",
 ]

backend/app/agents/schemas/event.py

@@ -7,10 +7,21 @@ from pydantic import BaseModel, Field
 
 
 AgentEventType = Literal[
+    "agent.execution.decided",
+    "agent.parallel.assessed",
+    "agent.skill.shortlisted",
+    "agent.task_graph.built",
+    "agent.subtask.started",
+    "agent.subtask.completed",
+    "agent.merge.completed",
     "agent.tool.start",
     "agent.tool.result",
     "agent.verify.started",
     "agent.verify.completed",
+    "agent.retrospective.created",
+    "agent.learning.decision",
+    "agent.skill.lifecycle.changed",
+    "agent.rollback.triggered",
     "agent.created",
     "agent.spawn.blocked",
     "agent.message.sent",

backend/app/agents/schemas/learning.py

@@ -0,0 +1,76 @@
+from __future__ import annotations
+
+from datetime import datetime, timezone
+from typing import Any, Literal
+
+from pydantic import BaseModel, Field
+
+
+LearningSignalType = Literal[
+    "preference",
+    "workflow",
+    "decomposition",
+    "tool_success",
+    "correction",
+]
+
+
+class SessionRetrospective(BaseModel):
+    retrospective_id: str | None = None
+    user_id: str
+    conversation_id: str
+    request_message_id: str | None = None
+    response_message_id: str | None = None
+    query_text: str
+    final_response: str | None = None
+    summary: str
+    task_type: str | None = None
+    execution_mode: str | None = None
+    primary_agent: str | None = None
+    verification_status: str | None = None
+    verification_summary: str | None = None
+    used_skill_names: list[str] = Field(default_factory=list)
+    evidence_refs: list[dict[str, Any]] = Field(default_factory=list)
+    task_refs: list[dict[str, Any]] = Field(default_factory=list)
+    event_refs: list[dict[str, Any]] = Field(default_factory=list)
+    context_snapshot: dict[str, Any] = Field(default_factory=dict)
+    learning_signals: list["LearningSignal"] = Field(default_factory=list)
+    pattern_candidates: list["PatternCandidate"] = Field(default_factory=list)
+    skill_candidates: list["SkillCandidate"] = Field(default_factory=list)
+    learning_decision: "LearningDecision | None" = None
+    outcome: Literal["completed", "partial", "failed"] = "completed"
+    captured_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+
+
+class LearningSignal(BaseModel):
+    signal_type: LearningSignalType
+    confidence: float = 0.0
+    evidence_refs: list[dict[str, Any]] = Field(default_factory=list)
+    explanation: str | None = None
+    payload: dict[str, Any] = Field(default_factory=dict)
+
+
+class PatternCandidate(BaseModel):
+    pattern_id: str
+    pattern_type: str
+    description: str
+    confidence: float = 0.0
+    evidence_refs: list[dict[str, Any]] = Field(default_factory=list)
+
+
+class SkillCandidate(BaseModel):
+    candidate_id: str
+    name: str
+    summary: str
+    candidate_type: Literal["workflow_skill", "preference_skill", "decomposition_skill"] = "workflow_skill"
+    source_pattern_ids: list[str] = Field(default_factory=list)
+    confidence: float = 0.0
+    evidence_refs: list[dict[str, Any]] = Field(default_factory=list)
+    recommended_status: Literal["candidate", "shadow"] = "candidate"
+
+
+class LearningDecision(BaseModel):
+    decision: Literal["reinforce_memory", "create_candidate", "promote_skill", "defer", "reject"]
+    explanation: str
+    evidence_refs: list[dict[str, Any]] = Field(default_factory=list)
+    metadata: dict[str, Any] = Field(default_factory=dict)

backend/app/agents/schemas/orchestration.py

@@ -0,0 +1,211 @@
+from __future__ import annotations
+
+import re
+from datetime import datetime, timezone
+from typing import Any, Literal
+from uuid import uuid4
+
+from pydantic import BaseModel, Field
+
+from app.agents.schemas.skills import SkillShortlistEntry
+
+
+ExecutionMode = Literal["direct", "collaboration", "parallel", "delegated"]
+ParallelPreference = Literal["direct", "collaboration", "parallel"]
+
+
+class ParallelWorthiness(BaseModel):
+    should_parallelize: bool = False
+    score: float = 0.0
+    estimated_subtasks: int = 1
+    preferred_mode: ParallelPreference = "direct"
+    reasons: list[str] = Field(default_factory=list)
+    risk_flags: list[str] = Field(default_factory=list)
+
+
+class TaskNode(BaseModel):
+    node_id: str
+    title: str
+    role: str | None = None
+    goal: str | None = None
+    depends_on: list[str] = Field(default_factory=list)
+    execution_mode: Literal["serial", "parallel"] = "serial"
+    expected_evidence: list[dict[str, Any]] = Field(default_factory=list)
+
+
+class TaskGraph(BaseModel):
+    graph_id: str = Field(default_factory=lambda: str(uuid4()))
+    nodes: list[TaskNode] = Field(default_factory=list)
+    entry_node_ids: list[str] = Field(default_factory=list)
+    max_parallelism: int = 1
+    rationale: str | None = None
+
+
+class SubTaskSpec(BaseModel):
+    subtask_id: str
+    parent_run_id: str
+    title: str
+    role: str
+    goal: str
+    context_slice: dict[str, Any] = Field(default_factory=dict)
+    allowed_tools: list[str] = Field(default_factory=list)
+    budget_tokens: int = 1200
+    budget_tool_calls: int = 2
+    expected_output_schema: dict[str, Any] = Field(default_factory=dict)
+    expected_evidence: list[dict[str, Any]] = Field(default_factory=list)
+    dependencies: list[str] = Field(default_factory=list)
+
+
+class SubTaskResult(BaseModel):
+    subtask_id: str
+    status: Literal["completed", "failed", "blocked"]
+    summary: str | None = None
+    evidence: list[dict[str, Any]] = Field(default_factory=list)
+    output: dict[str, Any] = Field(default_factory=dict)
+
+
+class MergeReport(BaseModel):
+    merge_id: str = Field(default_factory=lambda: str(uuid4()))
+    status: Literal["merged", "conflicted", "fallback"]
+    summary: str | None = None
+    evidence_union: list[dict[str, Any]] = Field(default_factory=list)
+    conflict_flags: list[str] = Field(default_factory=list)
+    resolution_strategy: str | None = None
+    resolved_summary: str | None = None
+    fallback_used: bool = False
+
+
+class VerificationReport(BaseModel):
+    status: Literal["passed", "failed", "skipped"]
+    summary: str | None = None
+    evidence: list[dict[str, Any]] = Field(default_factory=list)
+
+
+class ExecutionDecision(BaseModel):
+    request_id: str = Field(default_factory=lambda: str(uuid4()))
+    mode: ExecutionMode = "direct"
+    reason: str
+    complexity_score: float = 0.0
+    parallel_worthiness_score: float | None = None
+    selected_roles: list[str] = Field(default_factory=list)
+
+
+class RuntimeRequestContext(BaseModel):
+    request_id: str = Field(default_factory=lambda: str(uuid4()))
+    session_id: str | None = None
+    user_id: str
+    conversation_id: str | None = None
+    query_text: str | None = None
+    raw_user_query: str | None = None
+    recalled_memories: list[str] = Field(default_factory=list)
+    retrospective_shortlist: list[dict[str, Any]] = Field(default_factory=list)
+    recalled_retrospectives: list[dict[str, Any]] = Field(default_factory=list)
+    skill_shortlist: list[SkillShortlistEntry] = Field(default_factory=list)
+    shortlisted_skills: list[str] = Field(default_factory=list)
+    parallel_worthiness: ParallelWorthiness = Field(default_factory=ParallelWorthiness)
+    task_graph: TaskGraph | None = None
+    recommended_runtime_mode: Literal["direct", "collaboration"] = "direct"
+    execution_mode: Literal["direct", "collaboration"] | None = None
+    current_agent_role: str | None = None
+    conversation_state_ref: str | None = None
+    assembly_metrics: dict[str, float] = Field(default_factory=dict)
+    assembled_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+
+
+def assess_parallel_worthiness(
+    query_text: str,
+    *,
+    retrospective_count: int = 0,
+    skill_count: int = 0,
+) -> ParallelWorthiness:
+    normalized = (query_text or "").strip().lower()
+    reasons: list[str] = []
+    score = 0.0
+
+    multi_step_markers = ("然后", "接着", "同时", "并且", "最后", "汇总", "对比", "分析", "research")
+    artifact_markers = ("文档", "代码", "文件", "数据库", "论坛", "知识库", "计划")
+
+    if any(marker in normalized for marker in multi_step_markers):
+        score += 0.35
+        reasons.append("multi_step_request")
+
+    if sum(1 for marker in artifact_markers if marker in normalized) >= 2:
+        score += 0.25
+        reasons.append("multi_source_context")
+
+    if len(re.findall(r"[，,、；;]", query_text or "")) >= 2:
+        score += 0.15
+        reasons.append("compound_instruction")
+
+    if retrospective_count > 0:
+        score += 0.1
+        reasons.append("historical_support")
+
+    if skill_count > 0:
+        score += 0.1
+        reasons.append("skill_candidates_available")
+
+    score = min(score, 1.0)
+    should_parallelize = score >= 0.55
+    preferred_mode: ParallelPreference = "parallel" if should_parallelize else "direct"
+    if not should_parallelize and score >= 0.3:
+        preferred_mode = "collaboration"
+
+    estimated_subtasks = 1
+    if preferred_mode == "parallel":
+        estimated_subtasks = 3 if score >= 0.8 else 2
+    elif preferred_mode == "collaboration":
+        estimated_subtasks = 2
+
+    return ParallelWorthiness(
+        should_parallelize=should_parallelize,
+        score=round(score, 3),
+        estimated_subtasks=estimated_subtasks,
+        preferred_mode=preferred_mode,
+        reasons=reasons,
+    )
+
+
+def render_runtime_request_context_summary(context: RuntimeRequestContext) -> str:
+    lines = ["【Runtime Request Context】"]
+    lines.append(f"- 推荐运行模式: {context.recommended_runtime_mode}")
+    lines.append(
+        f"- 并行潜力: score={context.parallel_worthiness.score:.2f}, "
+        f"preferred={context.parallel_worthiness.preferred_mode}, "
+        f"estimated_subtasks={context.parallel_worthiness.estimated_subtasks}"
+    )
+
+    if context.parallel_worthiness.reasons:
+        lines.append(f"- 并行判断依据: {', '.join(context.parallel_worthiness.reasons)}")
+    if context.assembly_metrics:
+        total_ms = context.assembly_metrics.get("total_ms")
+        if total_ms is not None:
+            lines.append(f"- 上下文装配耗时: {total_ms:.1f} ms")
+
+    if context.task_graph and context.task_graph.nodes:
+        lines.append(
+            f"- 任务图: nodes={len(context.task_graph.nodes)}, max_parallelism={context.task_graph.max_parallelism}"
+        )
+        for node in context.task_graph.nodes[:4]:
+            deps = f", deps={len(node.depends_on)}" if node.depends_on else ""
+            lines.append(f"  - [{node.execution_mode}] {node.title} ({node.role}{deps})")
+
+    if context.retrospective_shortlist:
+        lines.append("- 历史复盘命中:")
+        for item in context.retrospective_shortlist[:3]:
+            summary = (item.get("summary") or item.get("summary_text") or "").strip()
+            task_type = item.get("task_type") or "unknown"
+            lines.append(f"  - [{task_type}] {summary[:160]}")
+
+    if context.skill_shortlist:
+        lines.append("- 技能候选:")
+        for item in context.skill_shortlist[:3]:
+            lines.append(
+                f"  - {item.skill_name} ({item.injection_mode}, score={item.score:.2f})"
+                + (f": {item.rationale}" if item.rationale else "")
+            )
+
+    if context.recalled_memories:
+        lines.append("- 记忆上下文已装配，可在回答中按需引用。")
+
+    return "\n".join(lines)

backend/app/agents/schemas/skills.py

@@ -0,0 +1,38 @@
+from __future__ import annotations
+
+from datetime import datetime, timezone
+from typing import Any, Literal
+
+from pydantic import BaseModel, Field
+
+
+SkillStatus = Literal["candidate", "shadow", "active", "deprecated", "retired"]
+SkillInjectionMode = Literal["metadata_only", "summary", "full"]
+
+
+class SkillShortlistEntry(BaseModel):
+    skill_name: str
+    source: str = "runtime"
+    source_id: str | None = None
+    status: SkillStatus = "active"
+    scope: list[str] = Field(default_factory=list)
+    effectiveness: float | None = None
+    score: float = 0.0
+    rationale: str | None = None
+    summary: str | None = None
+    matched_terms: list[str] = Field(default_factory=list)
+    injection_mode: SkillInjectionMode = "metadata_only"
+    metadata: dict[str, Any] = Field(default_factory=dict)
+
+
+class SkillActivationRecord(BaseModel):
+    skill_name: str
+    source: str = "runtime"
+    source_id: str | None = None
+    status: SkillStatus = "active"
+    injection_mode: SkillInjectionMode = "metadata_only"
+    matched_terms: list[str] = Field(default_factory=list)
+    rationale: str | None = None
+    activated_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+    outcome: str | None = None
+    metadata: dict[str, Any] = Field(default_factory=dict)

backend/app/agents/skills/__init__.py

@@ -1,16 +1 @@
-"""Skills 注册表 - Phase 9"""
-
-from app.agents.skills.registry import SkillRegistry, get_skill_registry
-from app.agents.skills.metadata import SkillMetadata
-from app.agents.skills.loaders.local_loader import LocalSkillLoader
-from app.agents.skills.loaders.plugin_loader import PluginSkillLoader
-from app.agents.skills.mcp_builder import MCPSkillBuilder
-
-__all__ = [
-    "SkillRegistry",
-    "SkillMetadata",
-    "LocalSkillLoader",
-    "PluginSkillLoader",
-    "MCPSkillBuilder",
-    "get_skill_registry",
-]
+"""Skill package."""

backend/app/agents/skills/effectiveness.py

@@ -0,0 +1,14 @@
+from __future__ import annotations
+
+from app.models.skill import Skill
+
+
+def summarize_skill_effectiveness(skill: Skill) -> dict[str, object]:
+    return {
+        "name": skill.name,
+        "status": skill.status,
+        "effectiveness": skill.effectiveness,
+        "activation_count": skill.activation_count,
+        "candidate_count": getattr(skill, "candidate_count", 0),
+        "last_activated_at": skill.last_activated_at.isoformat() if skill.last_activated_at else None,
+    }

backend/app/agents/skills/evaluator.py

@@ -0,0 +1,58 @@
+from __future__ import annotations
+
+from datetime import UTC, datetime, timedelta
+
+from app.agents.schemas.learning import SessionRetrospective, SkillCandidate
+from app.agents.skills.models import SkillLifecycleDecision
+from app.services.skill_service import SkillService
+
+
+class SkillPromotionEvaluator:
+    def __init__(self, db):
+        self.db = db
+        self.skill_service = SkillService(db)
+
+    async def sync_retrospective(
+        self,
+        *,
+        user_id: str,
+        retrospective: SessionRetrospective,
+    ) -> list[SkillLifecycleDecision]:
+        decisions: list[SkillLifecycleDecision] = []
+
+        for candidate in retrospective.skill_candidates:
+            decisions.append(
+                await self.skill_service.upsert_learned_candidate(
+                    user_id=user_id,
+                    candidate=candidate,
+                    primary_agent=retrospective.primary_agent,
+                    evidence_refs=candidate.evidence_refs,
+                )
+            )
+
+        outcome_score = self._derive_outcome_score(retrospective)
+        for skill_name in retrospective.used_skill_names:
+            decision = await self.skill_service.record_activation_feedback(
+                user_id=user_id,
+                skill_name=skill_name,
+                outcome_score=outcome_score,
+                evidence_refs=retrospective.evidence_refs,
+            )
+            if decision is not None:
+                decisions.append(decision)
+
+        return decisions
+
+    @staticmethod
+    def _derive_outcome_score(retrospective: SessionRetrospective) -> float:
+        if retrospective.verification_status == "passed":
+            return 0.9
+        if retrospective.verification_status == "skipped":
+            return 0.55
+        if retrospective.verification_status == "failed":
+            return 0.15
+        return 0.7 if retrospective.outcome == "completed" else 0.2
+
+
+def next_review_after(days: int = 7) -> datetime:
+    return datetime.now(UTC) + timedelta(days=days)

backend/app/agents/skills/matcher.py

@@ -0,0 +1,32 @@
+from __future__ import annotations
+
+import re
+
+
+def extract_match_terms(text: str | None) -> list[str]:
+    source = (text or "").lower()
+    terms = [token for token in re.findall(r"[a-z0-9_]+", source) if len(token) >= 3]
+
+    for chunk in re.findall(r"[\u4e00-\u9fff]+", text or ""):
+        if len(chunk) >= 2:
+            terms.append(chunk)
+            if len(chunk) > 4:
+                for index in range(len(chunk) - 1):
+                    terms.append(chunk[index : index + 2])
+
+    return list(dict.fromkeys(terms))
+
+
+def score_text_match(query_text: str, *corpus_parts: str | None) -> tuple[float, list[str]]:
+    query_terms = extract_match_terms(query_text)
+    if not query_terms:
+        return 0.0, []
+
+    corpus = " ".join(part for part in corpus_parts if part).lower()
+    matched_terms = [term for term in query_terms if term and term in corpus]
+    if not matched_terms:
+        return 0.0, []
+
+    coverage = len(matched_terms) / max(len(query_terms), 1)
+    density = min(len(matched_terms), 4) / 4
+    return round(min(1.0, coverage * 0.7 + density * 0.3), 3), matched_terms

backend/app/agents/skills/metadata.py

@@ -20,6 +20,10 @@ class SkillMetadata:
     source_id: str = ""  # 来源 ID
     enabled: bool = True  # 是否启用
     tools: list[str] = field(default_factory=list)  # 关联的工具
+    status: str = "active"  # candidate/shadow/active/deprecated/retired
+    scope: list[str] = field(default_factory=list)
+    effectiveness: float | None = None
+    review_after: str | None = None
 
     def to_dict(self) -> dict[str, Any]:
         return {
@@ -35,6 +39,10 @@ class SkillMetadata:
             "source_id": self.source_id,
             "enabled": self.enabled,
             "tools": self.tools,
+            "status": self.status,
+            "scope": self.scope,
+            "effectiveness": self.effectiveness,
+            "review_after": self.review_after,
         }
 
     @classmethod

backend/app/agents/skills/models.py

@@ -0,0 +1,29 @@
+from __future__ import annotations
+
+from datetime import datetime
+from typing import Literal
+
+from pydantic import BaseModel, Field
+
+
+SkillLifecycleAction = Literal[
+    "created_candidate",
+    "promoted_to_shadow",
+    "promoted_to_active",
+    "degraded_to_deprecated",
+    "retired",
+    "reactivated",
+    "feedback_recorded",
+    "no_change",
+]
+
+
+class SkillLifecycleDecision(BaseModel):
+    skill_name: str
+    action: SkillLifecycleAction
+    previous_status: str | None = None
+    new_status: str
+    reason: str
+    evidence_refs: list[dict[str, object]] = Field(default_factory=list)
+    confidence: float | None = None
+    review_after: datetime | None = None

backend/app/agents/skills/policy.py

@@ -0,0 +1,27 @@
+from __future__ import annotations
+
+from app.agents.schemas.skills import SkillInjectionMode, SkillShortlistEntry
+
+MAX_SUMMARY_CHARS = 120
+
+
+def choose_injection_mode(score: float, summary_available: bool) -> SkillInjectionMode:
+    if score >= 0.75 and summary_available:
+        return "summary"
+    return "metadata_only"
+
+
+def render_skill_shortlist_context(entries: list[SkillShortlistEntry]) -> str:
+    if not entries:
+        return ""
+
+    lines = ["[Task-Scoped Skills]"]
+    for entry in entries[:3]:
+        detail = entry.summary or "Relevant to the current request."
+        detail = detail[:MAX_SUMMARY_CHARS]
+        lines.append(f"- {entry.skill_name} | mode={entry.injection_mode} | score={entry.score:.2f}")
+        lines.append(f"  {detail}")
+        if entry.matched_terms:
+            lines.append(f"  matched_terms={', '.join(entry.matched_terms[:6])}")
+
+    return "\n".join(lines)

backend/app/agents/skills/retriever.py

@@ -0,0 +1,153 @@
+from __future__ import annotations
+
+from collections import OrderedDict
+
+from app.agents.schemas.skills import SkillShortlistEntry
+from app.agents.skills.matcher import score_text_match
+from app.agents.skills.policy import choose_injection_mode, render_skill_shortlist_context
+from app.agents.skills.registry import get_skill_registry
+from app.services.skill_service import SkillService
+
+
+class RuntimeSkillRetriever:
+    def __init__(self, db):
+        self.db = db
+
+    async def shortlist(
+        self,
+        *,
+        user_id: str,
+        query_text: str,
+        memory_context: str | None = None,
+        retrospectives: list[dict] | None = None,
+        include_learned: bool = True,
+        limit: int = 3,
+    ) -> list[SkillShortlistEntry]:
+        deduped: "OrderedDict[str, SkillShortlistEntry]" = OrderedDict()
+        retrospective_text = "\n".join(
+            (item.get("summary") or item.get("summary_text") or "")
+            for item in (retrospectives or [])
+            if isinstance(item, dict)
+        )
+
+        service = SkillService(self.db)
+        for skill in await service.list_runtime_candidates(user_id, include_learned=include_learned):
+            score, matched_terms = score_text_match(
+                query_text,
+                skill.name,
+                skill.description,
+                skill.instructions,
+                retrospective_text,
+                memory_context,
+            )
+            if score <= 0:
+                continue
+            entry = SkillShortlistEntry(
+                skill_name=skill.name,
+                source="database",
+                source_id=skill.id,
+                scope=[skill.agent_type, skill.visibility],
+                status=skill.status,
+                effectiveness=skill.effectiveness,
+                score=score,
+                matched_terms=matched_terms,
+                rationale=(
+                    "Shadow skill matched current request; keep metadata-only injection."
+                    if skill.status == "shadow"
+                    else "Matched against DB skill metadata and instructions."
+                ),
+                summary=skill.description or (skill.instructions[:160] if skill.instructions else None),
+                injection_mode=(
+                    "metadata_only"
+                    if skill.status == "shadow"
+                    else choose_injection_mode(score, bool(skill.description or skill.instructions))
+                ),
+            )
+            self._upsert(deduped, entry)
+
+        registry = get_skill_registry()
+        if not registry.list_all():
+            try:
+                registry.load_all()
+            except Exception:
+                pass
+
+        for skill in registry.list_all():
+            score, matched_terms = score_text_match(
+                query_text,
+                skill.name,
+                skill.description,
+                " ".join(skill.tags),
+                " ".join(skill.triggers),
+                skill.content[:400],
+                retrospective_text,
+                memory_context,
+            )
+            if score <= 0:
+                continue
+            entry = SkillShortlistEntry(
+                skill_name=skill.name,
+                source=skill.source,
+                source_id=skill.source_id or skill.id,
+                scope=skill.scope or list(skill.tags),
+                status=skill.status,
+                effectiveness=skill.effectiveness,
+                score=score,
+                matched_terms=matched_terms,
+                rationale="Matched against local or external skill metadata.",
+                summary=skill.description or skill.content[:160],
+                injection_mode=choose_injection_mode(
+                    score,
+                    bool(skill.description or skill.content),
+                ),
+            )
+            self._upsert(deduped, entry)
+
+        return sorted(deduped.values(), key=lambda item: item.score, reverse=True)[:limit]
+
+    @staticmethod
+    def _upsert(
+        deduped: "OrderedDict[str, SkillShortlistEntry]",
+        entry: SkillShortlistEntry,
+    ) -> None:
+        existing = deduped.get(entry.skill_name)
+        if existing is None or existing.score < entry.score:
+            deduped[entry.skill_name] = entry
+
+
+def build_shortlisted_skill_context(
+    shortlist: list[dict] | list[SkillShortlistEntry] | None,
+    *,
+    agent_type: str | None = None,
+) -> str:
+    if not shortlist:
+        return ""
+
+    entries: list[SkillShortlistEntry] = []
+    for item in shortlist:
+        entry = item if isinstance(item, SkillShortlistEntry) else SkillShortlistEntry.model_validate(item)
+        if agent_type and entry.scope and agent_type not in entry.scope:
+            continue
+        entries.append(entry)
+
+    return render_skill_shortlist_context(entries)
+
+
+async def shortlist_skills_for_request(
+    db,
+    *,
+    user_id: str,
+    user_query: str,
+    memory_context: str | None = None,
+    retrospectives: list[dict] | None = None,
+    include_learned: bool = True,
+    limit: int = 3,
+) -> list[SkillShortlistEntry]:
+    return await RuntimeSkillRetriever(db).shortlist(
+        user_id=user_id,
+        query_text=user_query,
+        memory_context=memory_context,
+        retrospectives=retrospectives,
+        include_learned=include_learned,
+        limit=limit,
+    )

backend/app/agents/state.py

@@ -138,6 +138,18 @@ class AgentState(TypedDict):
     memory_context: str | None
     current_datetime_context: str | None
     current_datetime_reference: dict[str, str] | None
+    runtime_request_context: dict[str, Any] | None
+    task_graph: dict[str, Any] | None
+    scheduled_subtasks: list[dict[str, Any]]
+    recalled_retrospectives: list[dict[str, Any]]
+    retrospective_shortlist: list[dict[str, Any]]
+    skill_shortlist: list[dict[str, Any]]
+    skill_activation_records: list[dict[str, Any]]
+    execution_decision: dict[str, Any] | None
+    merge_report: dict[str, Any] | None
+    verification_report: dict[str, Any] | None
+    feature_flags: dict[str, bool]
+    observability_report: dict[str, Any] | None
 
     turn_context: dict[str, Any] | None
     routing_decision: dict[str, Any] | None
@@ -254,6 +266,18 @@ def initial_state(user_id: str, conversation_id: str) -> AgentState:
         memory_context=None,
         current_datetime_context=None,
         current_datetime_reference=None,
+        runtime_request_context=None,
+        task_graph=None,
+        scheduled_subtasks=[],
+        recalled_retrospectives=[],
+        retrospective_shortlist=[],
+        skill_shortlist=[],
+        skill_activation_records=[],
+        execution_decision=None,
+        merge_report=None,
+        verification_report=None,
+        feature_flags={},
+        observability_report=None,
         turn_context=None,
         routing_decision=None,
         continuity_state=None,

backend/app/config.py

@@ -104,6 +104,15 @@ class Settings(BaseSettings):
     WEB_SEARCH_DEFAULT_LIMIT: int = 5
     WEB_SEARCH_TIMEOUT_SECONDS: int = 10
 
+    # === Hermes 风格升级开关 ===
+    ENABLE_RETROSPECTIVE: bool = True
+    ENABLE_SESSION_RETROSPECTIVE_SEARCH: bool = True
+    ENABLE_RUNTIME_SKILL_SHORTLIST: bool = True
+    ENABLE_LEARNING_SIGNALS: bool = True
+    ENABLE_SKILL_PROMOTION: bool = True
+    ENABLE_LEARNED_SKILL_LOADING: bool = True
+    ENABLE_PARALLEL_TASK_GRAPH: bool = True
+
 
 settings = Settings()
 settings.DATABASE_URL = settings.DATABASE_URL.replace("./data", _resolve_path("./data"), 1)

backend/app/database.py

@@ -39,10 +39,12 @@ async def init_db():
         await ensure_message_columns(conn)
         await ensure_conversation_columns(conn)
         await ensure_document_columns(conn)
+        await ensure_memory_columns(conn)
         await ensure_user_columns(conn)
         await ensure_forum_columns(conn)
         await ensure_agent_columns(conn)
         await ensure_skill_columns(conn)
+        await ensure_learning_artifact_tables(conn)
 
 
 async def ensure_log_columns(conn):
@@ -115,6 +117,28 @@ async def ensure_document_columns(conn):
             await conn.execute(text(ddl))
 
 
+async def ensure_memory_columns(conn):
+    rows = await _get_table_info(conn, 'user_memories')
+    if not rows:
+        return
+
+    columns = {row[1] for row in rows}
+    required_columns = {
+        'frequency_count': "ALTER TABLE user_memories ADD COLUMN frequency_count INTEGER DEFAULT 0",
+        'emotion_tags': "ALTER TABLE user_memories ADD COLUMN emotion_tags JSON",
+        'importance_score': "ALTER TABLE user_memories ADD COLUMN importance_score FLOAT DEFAULT 0.5",
+        'importance_level': "ALTER TABLE user_memories ADD COLUMN importance_level VARCHAR(20) DEFAULT 'medium'",
+        'associated_topics': "ALTER TABLE user_memories ADD COLUMN associated_topics JSON",
+        'decay_score': "ALTER TABLE user_memories ADD COLUMN decay_score FLOAT DEFAULT 1.0",
+        'is_archived': "ALTER TABLE user_memories ADD COLUMN is_archived BOOLEAN DEFAULT 0",
+        'last_accessed_at': "ALTER TABLE user_memories ADD COLUMN last_accessed_at DATETIME",
+        'archive_at': "ALTER TABLE user_memories ADD COLUMN archive_at DATETIME",
+    }
+    for column, ddl in required_columns.items():
+        if column not in columns:
+            await conn.execute(text(ddl))
+
+
 async def ensure_user_columns(conn):
     rows = await _get_table_info(conn, 'users')
     if not rows:
@@ -181,6 +205,14 @@ async def ensure_skill_columns(conn):
         'output_format': "ALTER TABLE skills ADD COLUMN output_format TEXT",
         'is_builtin': "ALTER TABLE skills ADD COLUMN is_builtin BOOLEAN DEFAULT 0 NOT NULL",
         'team_id': "ALTER TABLE skills ADD COLUMN team_id VARCHAR(36)",
+        'status': "ALTER TABLE skills ADD COLUMN status VARCHAR(20) DEFAULT 'active' NOT NULL",
+        'scope': "ALTER TABLE skills ADD COLUMN scope JSON DEFAULT '[]' NOT NULL",
+        'effectiveness': "ALTER TABLE skills ADD COLUMN effectiveness FLOAT DEFAULT 0.0 NOT NULL",
+        'review_after': "ALTER TABLE skills ADD COLUMN review_after DATETIME",
+        'candidate_count': "ALTER TABLE skills ADD COLUMN candidate_count INTEGER DEFAULT 0 NOT NULL",
+        'candidate_source_hashes': "ALTER TABLE skills ADD COLUMN candidate_source_hashes JSON DEFAULT '[]' NOT NULL",
+        'activation_count': "ALTER TABLE skills ADD COLUMN activation_count INTEGER DEFAULT 0 NOT NULL",
+        'last_activated_at': "ALTER TABLE skills ADD COLUMN last_activated_at DATETIME",
     }
     for column, ddl in required_columns.items():
         if column not in columns:
@@ -205,6 +237,48 @@ async def ensure_skill_columns(conn):
         )
 
 
+async def ensure_learning_artifact_tables(conn):
+    await conn.execute(
+        text(
+            """
+            CREATE TABLE IF NOT EXISTS learning_artifacts (
+                id VARCHAR(36) PRIMARY KEY,
+                created_at DATETIME NOT NULL,
+                updated_at DATETIME NOT NULL,
+                user_id VARCHAR(36) NOT NULL,
+                conversation_id VARCHAR(36) NOT NULL,
+                retrospective_id VARCHAR(36),
+                artifact_type VARCHAR(32) NOT NULL,
+                artifact_key VARCHAR(128),
+                summary_text TEXT NOT NULL,
+                payload JSON NOT NULL,
+                recorded_at DATETIME NOT NULL
+            )
+            """
+        )
+    )
+    await conn.execute(
+        text(
+            "CREATE INDEX IF NOT EXISTS ix_learning_artifacts_user_id ON learning_artifacts (user_id)"
+        )
+    )
+    await conn.execute(
+        text(
+            "CREATE INDEX IF NOT EXISTS ix_learning_artifacts_conversation_id ON learning_artifacts (conversation_id)"
+        )
+    )
+    await conn.execute(
+        text(
+            "CREATE INDEX IF NOT EXISTS ix_learning_artifacts_retrospective_id ON learning_artifacts (retrospective_id)"
+        )
+    )
+    await conn.execute(
+        text(
+            "CREATE INDEX IF NOT EXISTS ix_learning_artifacts_artifact_type ON learning_artifacts (artifact_type)"
+        )
+    )
+
+
 async def _backfill_usernames(conn):
     result = await conn.execute(text("SELECT id, email, username FROM users ORDER BY created_at, id"))
     users = result.fetchall()

backend/app/models/__init__.py

@@ -7,6 +7,7 @@ from app.models.forum import ForumPost, ForumReply
 from app.models.agent import Agent, AgentMessage
 from app.models.conversation import Conversation, Message
 from app.models.knowledge_graph import KGNode, KGEdge
+from app.models.learning import LearningArtifactRecord, SessionRetrospectiveRecord
 from app.models.memory import MemorySummary, UserMemory
 from app.models.brain import (
     BrainEvent,
@@ -20,6 +21,7 @@ from app.models.brain import (
 from app.models.todo import DailyTodo, TodoSource
 from app.models.reminder import Reminder, ReminderStatus
 from app.models.goal import Goal, GoalStatus
+from app.models.skill import Skill
 from app.models.log import Log, LogType, LogLevel
 
 __all__ = [
@@ -38,6 +40,8 @@ __all__ = [
     "Message",
     "KGNode",
     "KGEdge",
+    "LearningArtifactRecord",
+    "SessionRetrospectiveRecord",
     "MemorySummary",
     "UserMemory",
     "BrainEvent",
@@ -53,6 +57,7 @@ __all__ = [
     "ReminderStatus",
     "Goal",
     "GoalStatus",
+    "Skill",
     "Log",
     "LogType",
     "LogLevel",

backend/app/models/learning.py

@@ -0,0 +1,38 @@
+from sqlalchemy import Column, DateTime, ForeignKey, JSON, String, Text
+
+from app.models.base import BaseModel, utc_now
+
+
+class SessionRetrospectiveRecord(BaseModel):
+    __tablename__ = "session_retrospectives"
+
+    user_id = Column(String(36), ForeignKey("users.id"), nullable=False, index=True)
+    conversation_id = Column(String(36), ForeignKey("conversations.id"), nullable=False, index=True)
+    request_message_id = Column(String(36), ForeignKey("messages.id"), nullable=True, index=True)
+    response_message_id = Column(String(36), ForeignKey("messages.id"), nullable=True, index=True)
+    query_text = Column(Text, nullable=False)
+    final_response = Column(Text, nullable=True)
+    summary_text = Column(Text, nullable=False)
+    task_type = Column(String(64), nullable=True, index=True)
+    execution_mode = Column(String(32), nullable=True, index=True)
+    primary_agent = Column(String(64), nullable=True)
+    verification_status = Column(String(32), nullable=True)
+    verification_summary = Column(Text, nullable=True)
+    skill_names = Column(JSON, default=list, nullable=False)
+    evidence = Column(JSON, default=list, nullable=False)
+    task_refs = Column(JSON, default=list, nullable=False)
+    payload = Column(JSON, default=dict, nullable=False)
+    recorded_at = Column(DateTime, default=utc_now, nullable=False)
+
+
+class LearningArtifactRecord(BaseModel):
+    __tablename__ = "learning_artifacts"
+
+    user_id = Column(String(36), ForeignKey("users.id"), nullable=False, index=True)
+    conversation_id = Column(String(36), ForeignKey("conversations.id"), nullable=False, index=True)
+    retrospective_id = Column(String(36), ForeignKey("session_retrospectives.id"), nullable=True, index=True)
+    artifact_type = Column(String(32), nullable=False, index=True)
+    artifact_key = Column(String(128), nullable=True, index=True)
+    summary_text = Column(Text, nullable=False)
+    payload = Column(JSON, default=dict, nullable=False)
+    recorded_at = Column(DateTime, default=utc_now, nullable=False)

backend/app/models/skill.py

@@ -1,4 +1,4 @@
-from sqlalchemy import Column, String, Text, Boolean, JSON, ForeignKey
+from sqlalchemy import Column, String, Text, Boolean, JSON, ForeignKey, Float, Integer, DateTime
 from sqlalchemy.orm import relationship
 from app.models.base import BaseModel
 
@@ -17,6 +17,14 @@ class Skill(BaseModel):
     is_builtin = Column(Boolean, default=False, nullable=False)
     team_id = Column(String(36), ForeignKey("users.id"), nullable=True)
     is_active = Column(Boolean, default=True)
+    status = Column(String(20), default="active", nullable=False, index=True)  # candidate/shadow/active/deprecated/retired
+    scope = Column(JSON, default=list, nullable=False)
+    effectiveness = Column(Float, default=0.0, nullable=False)
+    review_after = Column(DateTime, nullable=True)
+    candidate_count = Column(Integer, default=0, nullable=False)
+    candidate_source_hashes = Column(JSON, default=list, nullable=False)
+    activation_count = Column(Integer, default=0, nullable=False)
+    last_activated_at = Column(DateTime, nullable=True)
     owner_id = Column(String(36), ForeignKey("users.id"), nullable=False)
 
     owner = relationship("User", foreign_keys=[owner_id])

backend/app/routers/agent.py

@@ -6,6 +6,7 @@ from sqlalchemy import select
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from app.database import get_db
+from app.agents.learning.store import LearningArtifactStore, SessionRetrospectiveStore
 from app.agents.registry import load_builtin_registry_indexes
 from app.agents.runtime_metrics import coerce_cost_thresholds, estimate_token_cost, is_cost_budget_warning
 from app.models.agent import Agent
@@ -37,6 +38,7 @@ from app.schemas.agent import (
     AgentVisibilityVerifierOut,
 )
 from app.services.agent_service import _extract_continuity_snapshot
+from app.services.runtime_observability import build_runtime_observability_report
 
 router = APIRouter(prefix="/api/agents", tags=["Agent"])
 
@@ -662,6 +664,59 @@ async def get_visibility_tools(
     return _build_tool_governance(state, conversation_id=conversation_id)
 
 
+@router.get("/visibility/debug")
+async def get_visibility_debug(
+    conversation_id: str,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    state = await _get_visibility_state(conversation_id, current_user=current_user, db=db)
+    observability = build_runtime_observability_report(
+        state=state,
+        feature_flags=dict(state.get("feature_flags") or {}),
+    )
+    retrospective_store = SessionRetrospectiveStore(db)
+    artifact_store = LearningArtifactStore(db)
+    recent_retrospectives = await retrospective_store.list_recent(
+        user_id=current_user.id,
+        limit=5,
+    )
+    recent_artifacts = await artifact_store.list_recent(
+        user_id=current_user.id,
+        limit=10,
+    )
+
+    return {
+        "conversation_id": conversation_id,
+        "observability": observability,
+        "skill_shortlist": list(state.get("skill_shortlist") or []),
+        "retrospective_shortlist": list(state.get("retrospective_shortlist") or []),
+        "merge_report": state.get("merge_report"),
+        "verification_report": state.get("verification_report"),
+        "recent_retrospectives": [
+            {
+                "id": item.id,
+                "task_type": item.task_type,
+                "summary": item.summary_text,
+                "execution_mode": item.execution_mode,
+                "verification_status": item.verification_status,
+                "recorded_at": item.recorded_at.isoformat() if item.recorded_at else None,
+            }
+            for item in recent_retrospectives
+        ],
+        "recent_learning_artifacts": [
+            {
+                "id": item.id,
+                "artifact_type": item.artifact_type,
+                "artifact_key": item.artifact_key,
+                "summary": item.summary_text,
+                "recorded_at": item.recorded_at.isoformat() if item.recorded_at else None,
+            }
+            for item in recent_artifacts
+        ],
+    }
+
+
 @router.post("", response_model=AgentOut, status_code=201)
 async def create_agent(
     data: AgentCreate,

backend/app/routers/conversation.py

@@ -145,6 +145,9 @@ async def chat_stream(
             except ValueError as exc:
                 yield f"event: error\ndata: {json.dumps({'error': str(exc)}, ensure_ascii=False)}\n\n"
                 return
+            except Exception as exc:
+                yield f"event: error\ndata: {json.dumps({'error': str(exc)}, ensure_ascii=False)}\n\n"
+                return
 
             yield f"event: metadata\ndata: {json.dumps({'conversation_id': conv_id, 'message_id': msg_id})}\n\n"
 

backend/app/routers/skill.py

@@ -29,6 +29,10 @@ async def create_skill(
         visibility=data.visibility,
         team_id=data.team_id,
         is_active=data.is_active,
+        status=data.status,
+        scope=data.scope,
+        effectiveness=data.effectiveness,
+        review_after=data.review_after,
         owner_id=current_user.id,
     )
     db.add(skill)
@@ -103,6 +107,14 @@ async def update_skill(
         skill.team_id = data.team_id
     if data.is_active is not None:
         skill.is_active = data.is_active
+    if data.status is not None:
+        skill.status = data.status
+    if data.scope is not None:
+        skill.scope = data.scope
+    if data.effectiveness is not None:
+        skill.effectiveness = data.effectiveness
+    if data.review_after is not None:
+        skill.review_after = data.review_after
 
     await db.commit()
     await db.refresh(skill)

backend/app/routers/system.py

@@ -12,4 +12,4 @@ async def get_system_status():
 @router.get('/config')
 async def get_system_config():
     """Get public system configuration."""
-    return SystemService().get_config()
+    return await SystemService().get_config()

backend/app/schemas/skill.py

@@ -14,6 +14,10 @@ class SkillCreate(BaseModel):
     visibility: str = "private"
     team_id: Optional[str] = None
     is_active: bool = True
+    status: str = "active"
+    scope: list[str] = []
+    effectiveness: Optional[float] = None
+    review_after: Optional[datetime] = None
 
 
 class SkillUpdate(BaseModel):
@@ -28,6 +32,10 @@ class SkillUpdate(BaseModel):
     visibility: Optional[str] = None
     team_id: Optional[str] = None
     is_active: Optional[bool] = None
+    status: Optional[str] = None
+    scope: Optional[list[str]] = None
+    effectiveness: Optional[float] = None
+    review_after: Optional[datetime] = None
 
 
 class SkillOut(BaseModel):
@@ -43,6 +51,12 @@ class SkillOut(BaseModel):
     is_builtin: bool
     team_id: Optional[str]
     is_active: bool
+    status: str
+    scope: list[str]
+    effectiveness: Optional[float]
+    review_after: Optional[datetime]
+    activation_count: int
+    last_activated_at: Optional[datetime]
     owner_id: str
     created_at: datetime
     updated_at: datetime

backend/app/services/agent_service.py

@@ -7,12 +7,13 @@ import json
 import uuid
 import logging
 from datetime import UTC, datetime
+from time import perf_counter
 from typing import Any, AsyncGenerator
 import asyncio
 from openai import BadRequestError
 from sqlalchemy.ext.asyncio import AsyncSession
 from sqlalchemy import select
-from langchain_core.messages import HumanMessage, AIMessage
+from langchain_core.messages import HumanMessage, AIMessage, SystemMessage
 
 from app.database import async_session
 from app.logging_utils import summarize_llm_config
@@ -21,10 +22,24 @@ from app.models.conversation import Conversation, Message
 from app.models.user import User
 from app.agents.graph import get_agent_graph
 from app.agents.context import set_current_user, clear_current_user
+from app.agents.learning.jobs import schedule_retrospective_job
+from app.agents.learning.retrospector import build_session_retrospective
+from app.agents.learning.session_search import SessionRetrospectiveSearch, summarize_retrospective
+from app.agents.orchestration.task_graph import build_bounded_task_graph
+from app.agents.learning.store import append_retrospective_attachment
+from app.agents.schemas.orchestration import (
+    RuntimeRequestContext,
+    assess_parallel_worthiness,
+    render_runtime_request_context_summary,
+)
+from app.agents.schemas.skills import SkillActivationRecord
 from app.agents.skills.registry import get_skill_registry
+from app.agents.skills.retriever import shortlist_skills_for_request
 from app.services import memory_service
 from app.services.brain_service import BrainService
 from app.services.llm_service import create_llm_from_config, resolve_provider_capabilities
+from app.services.rollback_controller import RollbackController
+from app.services.runtime_observability import build_runtime_observability_report
 from app.agents.tools.time_reasoning import extract_reference_datetime
 from app.agents.state import initial_state
 
@@ -36,6 +51,7 @@ MEMORY_SECTION_HEADERS = (
     "【之前对话摘要】",
     "【知识大脑】",
 )
+MEMORY_INLINE_HEADERS = {"[关于你的记忆]"}
 
 
 def _split_memory_context_sections(memory_context: str | None) -> dict[str, str]:
@@ -81,6 +97,41 @@ def _derive_role_memory_contexts(memory_context: str | None) -> dict[str, str |
     }
 
 
+def _extract_memory_highlights(memory_context: str | None, *, limit: int = 5) -> list[str]:
+    text = (memory_context or "").strip()
+    if not text:
+        return []
+
+    highlights: list[str] = []
+    for raw_line in text.splitlines():
+        line = raw_line.strip()
+        if not line or line in MEMORY_SECTION_HEADERS or line in MEMORY_INLINE_HEADERS:
+            continue
+        if line.startswith("-"):
+            normalized = line[1:].strip()
+        else:
+            normalized = line
+        if normalized:
+            highlights.append(normalized)
+        if len(highlights) >= limit:
+            break
+    return highlights
+
+
+def _summarize_retrospective(retrospective: Any) -> str:
+    summary = str(getattr(retrospective, "summary", "") or "").strip()
+    task_type = str(getattr(retrospective, "task_type", "") or "").strip()
+    execution_mode = str(getattr(retrospective, "execution_mode", "") or "").strip()
+    outcome = str(getattr(retrospective, "outcome", "") or "").strip()
+
+    parts = [summary[:80] or task_type or "历史复盘"]
+    if execution_mode:
+        parts.append(f"mode={execution_mode}")
+    if outcome:
+        parts.append(f"outcome={outcome}")
+    return "；".join(parts)
+
+
 def _is_streaming_rejection_error(error: Exception, user_llm_config: dict | None) -> bool:
     capabilities = resolve_provider_capabilities(user_llm_config)
     error_text = str(error).lower()
@@ -461,18 +512,27 @@ class AgentService:
     async def _build_agent_state(
         self,
         *,
+        request_id: str,
         user_id: str,
         conversation: Conversation,
+        raw_user_query: str,
         full_message: str,
         memory_context: str | None,
         current_datetime_context: str,
         current_datetime_reference: dict[str, str],
         user_llm_config: dict | None,
+        runtime_request_context: RuntimeRequestContext,
+        recalled_retrospectives: list[dict[str, Any]],
+        skill_shortlist: list[dict[str, Any]],
     ) -> dict[str, Any]:
         state = initial_state(user_id, conversation.id)
+        runtime_summary = render_runtime_request_context_summary(runtime_request_context)
         state.update(
             {
-                "messages": [HumanMessage(content=full_message)],
+                "messages": [
+                    SystemMessage(content=runtime_summary),
+                    HumanMessage(content=full_message),
+                ],
                 "memory_context": memory_context,
                 "current_datetime_context": current_datetime_context,
                 "current_datetime_reference": current_datetime_reference,
@@ -482,9 +542,119 @@ class AgentService:
         previous_snapshot = await self._load_continuity_snapshot(conversation)
         if previous_snapshot:
             state.update(previous_snapshot)
-            state["messages"] = [HumanMessage(content=full_message)]
+            state["messages"] = [
+                SystemMessage(content=runtime_summary),
+                HumanMessage(content=full_message),
+            ]
+        state.update(
+            {
+                "runtime_request_context": runtime_request_context.model_dump(mode="json"),
+                "task_graph": (
+                    runtime_request_context.task_graph.model_dump(mode="json")
+                    if runtime_request_context.task_graph is not None
+                    else None
+                ),
+                "feature_flags": RollbackController().snapshot_flags(),
+                "recalled_retrospectives": recalled_retrospectives,
+                "retrospective_shortlist": recalled_retrospectives,
+                "skill_shortlist": skill_shortlist,
+                "skill_activation_records": [
+                    SkillActivationRecord(
+                        skill_name=item.get("skill_name"),
+                        source=item.get("source", "runtime"),
+                        source_id=item.get("source_id"),
+                        status=item.get("status", "active"),
+                        injection_mode=item.get("injection_mode", "metadata_only"),
+                        matched_terms=item.get("matched_terms", []),
+                        rationale=item.get("rationale"),
+                    ).model_dump(mode="json")
+                    for item in skill_shortlist
+                    if item.get("skill_name")
+                ],
+                "parallel_worthiness": runtime_request_context.parallel_worthiness.model_dump(
+                    mode="json"
+                ),
+            }
+        )
         return state
 
+    async def _build_runtime_request_context(
+        self,
+        *,
+        request_id: str,
+        user_id: str,
+        conversation: Conversation,
+        user_query: str,
+        memory_context: str | None,
+    ) -> tuple[RuntimeRequestContext, list[dict[str, Any]], list[dict[str, Any]]]:
+        started_at = perf_counter()
+        retrospectives_started = perf_counter()
+        recent_retrospectives = await SessionRetrospectiveSearch(self.db).shortlist(
+            user_id=user_id,
+            query_text=user_query,
+            conversation_id=conversation.id,
+            limit=3,
+        )
+        retrospective_ms = (perf_counter() - retrospectives_started) * 1000
+        feature_flags = RollbackController().snapshot_flags()
+        shortlist_started = perf_counter()
+        skill_shortlist = await shortlist_skills_for_request(
+            self.db,
+            user_id=user_id,
+            user_query=user_query,
+            memory_context=memory_context,
+            retrospectives=[item.model_dump(mode="json") for item in recent_retrospectives],
+            include_learned=feature_flags["ENABLE_LEARNED_SKILL_LOADING"],
+            limit=4,
+        )
+        skill_shortlist_ms = (perf_counter() - shortlist_started) * 1000
+        parallel_worthiness = assess_parallel_worthiness(
+            user_query,
+            retrospective_count=len(recent_retrospectives),
+            skill_count=len(skill_shortlist),
+        )
+        recommended_runtime_mode = (
+            "collaboration" if parallel_worthiness.preferred_mode != "direct" else "direct"
+        )
+        task_graph = (
+            build_bounded_task_graph(
+                query_text=user_query,
+                parallel_worthiness=parallel_worthiness,
+            )
+            if feature_flags["ENABLE_PARALLEL_TASK_GRAPH"]
+            else None
+        )
+        runtime_request_context = RuntimeRequestContext(
+            request_id=request_id,
+            session_id=conversation.id,
+            conversation_id=conversation.id,
+            user_id=user_id,
+            query_text=user_query,
+            raw_user_query=user_query,
+            recalled_memories=_extract_memory_highlights(memory_context),
+            recalled_retrospectives=[
+                summarize_retrospective(retrospective) for retrospective in recent_retrospectives
+            ],
+            shortlisted_skills=[entry.skill_name for entry in skill_shortlist],
+            skill_shortlist=skill_shortlist,
+            current_agent_role="master",
+            execution_mode=recommended_runtime_mode,
+            conversation_state_ref=conversation.id,
+            parallel_worthiness=parallel_worthiness,
+            task_graph=task_graph,
+            recommended_runtime_mode=recommended_runtime_mode,
+            assembly_metrics={
+                "retrospective_ms": round(retrospective_ms, 3),
+                "skill_shortlist_ms": round(skill_shortlist_ms, 3),
+                "total_ms": round((perf_counter() - started_at) * 1000, 3),
+            },
+        )
+        return (
+            runtime_request_context,
+            [item.model_dump(mode="json") for item in recent_retrospectives],
+            [item.model_dump(mode="json") for item in skill_shortlist],
+        )
+
     async def chat(
         self,
         user_id: str,
@@ -610,21 +780,38 @@ class AgentService:
         async def run_agent():
             collected = ""
             state: dict[str, Any] | None = None
+            runtime_request_context: RuntimeRequestContext | None = None
             set_current_user(user_id)
             try:
                 graph = get_agent_graph()
                 current_datetime_context, current_datetime_reference = (
                     self._build_current_datetime_context()
                 )
-
-                state = await self._build_agent_state(
+                (
+                    runtime_request_context,
+                    recalled_retrospectives,
+                    skill_shortlist,
+                ) = await self._build_runtime_request_context(
+                    request_id=assistant_msg.id,
                     user_id=user_id,
                     conversation=conv,
+                    user_query=message,
+                    memory_context=memory_ctx,
+                )
+
+                state = await self._build_agent_state(
+                    request_id=assistant_msg.id,
+                    user_id=user_id,
+                    conversation=conv,
+                    raw_user_query=message,
                     full_message=full_message,
                     memory_context=memory_ctx,
                     current_datetime_context=current_datetime_context,
                     current_datetime_reference=current_datetime_reference,
                     user_llm_config=user_llm_config,
+                    runtime_request_context=runtime_request_context,
+                    recalled_retrospectives=recalled_retrospectives,
+                    skill_shortlist=skill_shortlist,
                 )
                 state.update(_derive_role_memory_contexts(memory_ctx))
 
@@ -749,7 +936,7 @@ class AgentService:
                     if collected:
                         assistant_msg.content = collected
                         continuity_snapshot = _build_continuity_snapshot(state or {})
-                        assistant_msg.attachments = (
+                        attachments = (
                             [
                                 {
                                     "kind": "agent_continuity_state",
@@ -757,8 +944,26 @@ class AgentService:
                                 }
                             ]
                             if continuity_snapshot
-                            else None
+                            else []
                         )
+                        if state is not None and runtime_request_context is not None:
+                            retrospective = build_session_retrospective(
+                                request_id=assistant_msg.id,
+                                session_id=conversation_id,
+                                user_query=message,
+                                state=state,
+                                runtime_context=runtime_request_context,
+                            )
+                            attachments = append_retrospective_attachment(attachments, retrospective)
+                            attachments.append(
+                                {
+                                    "kind": "runtime_observability",
+                                    "payload": build_runtime_observability_report(
+                                        state=state,
+                                        feature_flags=state.get("feature_flags") or {},
+                                    ),
+                                }
+                            )
                         conv.agent_state = (
                             {
                                 "kind": "agent_continuity_state",
@@ -771,8 +976,18 @@ class AgentService:
                             user_id,
                             **_build_assistant_event_payload(collected),
                         )
+                        assistant_msg.attachments = attachments or None
                         await self.db.commit()
                         await self.db.refresh(assistant_msg)
+                        schedule_retrospective_job(
+                            user_id=user_id,
+                            conversation_id=conversation_id,
+                            request_message_id=user_msg.id,
+                            response_message_id=assistant_msg.id,
+                            query_text=message,
+                            final_response=collected,
+                            state=state,
+                        )
                 except Exception:
                     logger.exception("save_assistant_message_failed")
                 asyncio.create_task(self._try_auto_summarize_background(user_id, conversation_id))
@@ -863,14 +1078,30 @@ class AgentService:
             current_datetime_context, current_datetime_reference = (
                 self._build_current_datetime_context()
             )
-            state = await self._build_agent_state(
+            (
+                runtime_request_context,
+                recalled_retrospectives,
+                skill_shortlist,
+            ) = await self._build_runtime_request_context(
+                request_id=assistant_msg.id,
                 user_id=user_id,
                 conversation=conv,
+                user_query=message,
+                memory_context=memory_ctx,
+            )
+            state = await self._build_agent_state(
+                request_id=assistant_msg.id,
+                user_id=user_id,
+                conversation=conv,
+                raw_user_query=message,
                 full_message=message,
                 memory_context=memory_ctx,
                 current_datetime_context=current_datetime_context,
                 current_datetime_reference=current_datetime_reference,
                 user_llm_config=user_llm_config,
+                runtime_request_context=runtime_request_context,
+                recalled_retrospectives=recalled_retrospectives,
+                skill_shortlist=skill_shortlist,
             )
             state.update(_derive_role_memory_contexts(memory_ctx))
             result_state = await graph.ainvoke(state)
@@ -900,7 +1131,7 @@ class AgentService:
         continuity_snapshot = (
             _build_continuity_snapshot(result_state) if "result_state" in locals() else None
         )
-        assistant_msg.attachments = (
+        attachments = (
             [
                 {
                     "kind": "agent_continuity_state",
@@ -908,8 +1139,26 @@ class AgentService:
                 }
             ]
             if continuity_snapshot
-            else None
+            else []
         )
+        if "result_state" in locals() and "runtime_request_context" in locals():
+            retrospective = build_session_retrospective(
+                request_id=assistant_msg.id,
+                session_id=conversation_id,
+                user_query=message,
+                state=result_state,
+                runtime_context=runtime_request_context,
+            )
+            attachments = append_retrospective_attachment(attachments, retrospective)
+            attachments.append(
+                {
+                    "kind": "runtime_observability",
+                    "payload": build_runtime_observability_report(
+                        state=result_state,
+                        feature_flags=result_state.get("feature_flags") or {},
+                    ),
+                }
+            )
         conv.agent_state = (
             {
                 "kind": "agent_continuity_state",
@@ -918,7 +1167,17 @@ class AgentService:
             if continuity_snapshot
             else None
         )
+        assistant_msg.attachments = attachments or None
         await self.db.commit()
         await self.db.refresh(assistant_msg)
+        schedule_retrospective_job(
+            user_id=user_id,
+            conversation_id=conversation_id,
+            request_message_id=user_msg.id,
+            response_message_id=assistant_msg.id,
+            query_text=message,
+            final_response=response_content,
+            state=result_state if "result_state" in locals() else None,
+        )
 
         return conversation_id, assistant_msg.id, response_content, model_name_used

backend/app/services/rollback_controller.py

@@ -0,0 +1,25 @@
+from __future__ import annotations
+
+from app.config import settings
+
+
+FEATURE_FLAG_NAMES = (
+    "ENABLE_RETROSPECTIVE",
+    "ENABLE_SESSION_RETROSPECTIVE_SEARCH",
+    "ENABLE_RUNTIME_SKILL_SHORTLIST",
+    "ENABLE_LEARNING_SIGNALS",
+    "ENABLE_SKILL_PROMOTION",
+    "ENABLE_LEARNED_SKILL_LOADING",
+    "ENABLE_PARALLEL_TASK_GRAPH",
+)
+
+
+class RollbackController:
+    def snapshot_flags(self) -> dict[str, bool]:
+        return {
+            flag_name: bool(getattr(settings, flag_name, False))
+            for flag_name in FEATURE_FLAG_NAMES
+        }
+
+    def is_enabled(self, flag_name: str) -> bool:
+        return bool(getattr(settings, flag_name, False))

backend/app/services/runtime_observability.py

@@ -0,0 +1,32 @@
+from __future__ import annotations
+
+from typing import Any
+
+from app.agents.orchestration.monitor import build_parallel_runtime_metrics
+
+
+def build_runtime_observability_report(
+    *,
+    state: dict[str, Any],
+    feature_flags: dict[str, bool] | None = None,
+) -> dict[str, Any]:
+    task_graph = state.get("task_graph") if isinstance(state.get("task_graph"), dict) else None
+    scheduled_subtasks = (
+        state.get("scheduled_subtasks") if isinstance(state.get("scheduled_subtasks"), list) else []
+    )
+    task_results = state.get("task_results") if isinstance(state.get("task_results"), list) else []
+    merge_report = state.get("merge_report") if isinstance(state.get("merge_report"), dict) else None
+
+    return {
+        "execution_mode": state.get("execution_mode"),
+        "verification_status": state.get("verification_status"),
+        "skill_shortlist_count": len(state.get("skill_shortlist") or []),
+        "retrospective_shortlist_count": len(state.get("retrospective_shortlist") or []),
+        "feature_flags": feature_flags or {},
+        "parallel_metrics": build_parallel_runtime_metrics(
+            task_graph=task_graph,
+            scheduled_subtasks=scheduled_subtasks,
+            task_results=task_results,
+            merge_report=merge_report,
+        ),
+    }

backend/app/services/skill_service.py

@@ -3,9 +3,13 @@ Skill Service - 技能管理服务层
 负责技能的创建、查询、更新、删除等操作
 """
 
+import hashlib
+from datetime import UTC, datetime, timedelta
 from typing import Optional
 from sqlalchemy.ext.asyncio import AsyncSession
 from sqlalchemy import select, and_, or_
+from app.agents.schemas.learning import SkillCandidate
+from app.agents.skills.models import SkillLifecycleDecision
 from app.models.skill import Skill
 from app.models.user import User
 
@@ -28,6 +32,10 @@ class SkillService:
             visibility=data.get("visibility", "private"),
             team_id=data.get("team_id"),
             is_active=data.get("is_active", True),
+            status=data.get("status", "active"),
+            scope=data.get("scope", []),
+            effectiveness=data.get("effectiveness", 0.0),
+            review_after=data.get("review_after"),
         )
         self.db.add(skill)
         await self.db.commit()
@@ -41,6 +49,17 @@ class SkillService:
         )
         return result.scalar_one_or_none()
 
+    async def get_by_name_for_user(self, user_id: str, name: str) -> Optional[Skill]:
+        access_scope = or_(
+            Skill.owner_id == user_id,
+            Skill.visibility == "market",
+            Skill.team_id == user_id,
+        )
+        result = await self.db.execute(
+            select(Skill).where(and_(Skill.name == name, access_scope))
+        )
+        return result.scalar_one_or_none()
+
     async def list_for_user(
         self,
         user_id: str,
@@ -56,7 +75,7 @@ class SkillService:
             Skill.team_id == user_id,
         )
 
-        filters = [access_scope, Skill.is_active == True]
+        filters = [access_scope, Skill.is_active == True, Skill.status != "retired"]
 
         if agent_type:
             filters.append(Skill.agent_type == agent_type)
@@ -83,7 +102,7 @@ class SkillService:
         update_fields = [
             "name", "description", "instructions", "agent_type",
             "tools", "required_context", "output_format", "visibility",
-            "team_id", "is_active"
+            "team_id", "is_active", "status", "scope", "effectiveness", "review_after"
         ]
 
         for field in update_fields:
@@ -117,6 +136,7 @@ class SkillService:
                 and_(
                     Skill.agent_type == agent_type,
                     Skill.is_active == True,
+                    Skill.status == "active",
                     or_(
                         Skill.visibility == "market",
                         Skill.visibility == "private"
@@ -125,3 +145,234 @@ class SkillService:
             )
         )
         return list(result.scalars().all())
+
+    async def list_runtime_candidates(
+        self,
+        user_id: str,
+        *,
+        agent_type: Optional[str] = None,
+        include_shadow: bool = True,
+        include_learned: bool = True,
+    ) -> list[Skill]:
+        allowed_statuses = ["active", "shadow"] if include_shadow else ["active"]
+        access_scope = or_(
+            Skill.owner_id == user_id,
+            Skill.visibility == "market",
+            Skill.team_id == user_id,
+        )
+
+        filters = [
+            access_scope,
+            Skill.is_active == True,
+            Skill.status.in_(allowed_statuses),
+        ]
+        if not include_learned:
+            filters.append(Skill.is_builtin == True)
+        if agent_type:
+            filters.append(Skill.agent_type == agent_type)
+
+        result = await self.db.execute(select(Skill).where(and_(*filters)))
+        return list(result.scalars().all())
+
+    async def upsert_learned_candidate(
+        self,
+        *,
+        user_id: str,
+        candidate: SkillCandidate,
+        primary_agent: str | None,
+        evidence_refs: list[dict] | None = None,
+    ) -> SkillLifecycleDecision:
+        source_hash = self._build_candidate_source_hash(candidate)
+        skill = await self.get_by_name_for_user(user_id, candidate.name)
+        if skill is None:
+            review_after = datetime.now(UTC) + timedelta(days=7)
+            skill = Skill(
+                owner_id=user_id,
+                name=candidate.name,
+                description=candidate.summary,
+                instructions=candidate.summary,
+                agent_type=primary_agent or "master",
+                tools=[],
+                required_context=[],
+                output_format=None,
+                visibility="private",
+                is_active=True,
+                status="candidate",
+                scope=[primary_agent or "master", "learned", candidate.candidate_type],
+                effectiveness=candidate.confidence,
+                review_after=review_after,
+                candidate_count=1,
+                candidate_source_hashes=[source_hash],
+            )
+            self.db.add(skill)
+            await self.db.commit()
+            await self.db.refresh(skill)
+            return SkillLifecycleDecision(
+                skill_name=skill.name,
+                action="created_candidate",
+                previous_status=None,
+                new_status="candidate",
+                reason="First learned candidate created from retrospective evidence.",
+                evidence_refs=evidence_refs or [],
+                confidence=candidate.confidence,
+                review_after=review_after,
+            )
+
+        previous_status = skill.status
+        known_hashes = list(skill.candidate_source_hashes or [])
+        is_duplicate_candidate = source_hash in known_hashes
+        if not is_duplicate_candidate:
+            skill.candidate_count = int(skill.candidate_count or 0) + 1
+            known_hashes.append(source_hash)
+        skill.candidate_source_hashes = known_hashes
+        current_effectiveness = float(skill.effectiveness or 0.0)
+        skill.effectiveness = round(max(current_effectiveness, float(candidate.confidence or 0.0)), 3)
+        skill.review_after = datetime.now(UTC) + timedelta(days=7)
+        if primary_agent and primary_agent not in (skill.scope or []):
+            skill.scope = [*(skill.scope or []), primary_agent]
+
+        action = "no_change"
+        reason = "Candidate evidence refreshed."
+        if is_duplicate_candidate:
+            reason = "Duplicate candidate evidence ignored for promotion counting."
+        if (
+            not is_duplicate_candidate
+            and skill.status == "candidate"
+            and skill.candidate_count >= 2
+            and skill.effectiveness >= 0.6
+        ):
+            skill.status = "shadow"
+            action = "promoted_to_shadow"
+            reason = "Repeated candidate evidence promoted the learned skill to shadow."
+
+        await self.db.commit()
+        await self.db.refresh(skill)
+        return SkillLifecycleDecision(
+            skill_name=skill.name,
+            action=action,
+            previous_status=previous_status,
+            new_status=skill.status,
+            reason=reason,
+            evidence_refs=evidence_refs or [],
+            confidence=skill.effectiveness,
+            review_after=skill.review_after,
+        )
+
+    async def record_activation_feedback(
+        self,
+        *,
+        user_id: str,
+        skill_name: str,
+        outcome_score: float,
+        evidence_refs: list[dict] | None = None,
+    ) -> SkillLifecycleDecision | None:
+        skill = await self.get_by_name_for_user(user_id, skill_name)
+        if skill is None or skill.status not in {"shadow", "active", "deprecated"}:
+            return None
+
+        previous_status = skill.status
+        previous_activation_count = int(skill.activation_count or 0)
+        skill.activation_count = previous_activation_count + 1
+        skill.last_activated_at = datetime.now(UTC)
+
+        previous_effectiveness = float(skill.effectiveness or 0.0)
+        if previous_activation_count <= 0:
+            skill.effectiveness = round(outcome_score, 3)
+        else:
+            skill.effectiveness = round(
+                ((previous_effectiveness * previous_activation_count) + outcome_score)
+                / skill.activation_count,
+                3,
+            )
+
+        action = "feedback_recorded"
+        reason = "Activation outcome recorded."
+        if skill.status == "shadow" and skill.activation_count >= 2 and skill.effectiveness >= 0.7:
+            skill.status = "active"
+            action = "promoted_to_active"
+            reason = "Shadow skill proved effective enough to become active."
+        elif skill.status == "active" and skill.activation_count >= 3 and skill.effectiveness < 0.35:
+            skill.status = "deprecated"
+            action = "degraded_to_deprecated"
+            reason = "Active skill underperformed repeatedly and was deprecated."
+        elif skill.status == "deprecated" and skill.activation_count >= 4 and skill.effectiveness < 0.2:
+            skill.status = "retired"
+            action = "retired"
+            reason = "Deprecated skill stayed ineffective and was retired."
+        elif skill.status == "deprecated" and skill.effectiveness >= 0.65 and outcome_score >= 0.8:
+            skill.status = "active"
+            action = "reactivated"
+            reason = "Deprecated skill recovered with strong positive feedback."
+
+        skill.review_after = datetime.now(UTC) + timedelta(days=7)
+        await self.db.commit()
+        await self.db.refresh(skill)
+        return SkillLifecycleDecision(
+            skill_name=skill.name,
+            action=action,
+            previous_status=previous_status,
+            new_status=skill.status,
+            reason=reason,
+            evidence_refs=evidence_refs or [],
+            confidence=skill.effectiveness,
+            review_after=skill.review_after,
+        )
+
+    async def run_decay_review(
+        self,
+        *,
+        user_id: str,
+        as_of: datetime | None = None,
+    ) -> list[SkillLifecycleDecision]:
+        review_time = as_of or datetime.now(UTC)
+        result = await self.db.execute(
+            select(Skill).where(
+                and_(
+                    Skill.owner_id == user_id,
+                    Skill.is_active == True,
+                    Skill.status.in_(["shadow", "active", "deprecated"]),
+                    Skill.review_after.is_not(None),
+                    Skill.review_after <= review_time,
+                )
+            )
+        )
+        skills = list(result.scalars().all())
+        decisions: list[SkillLifecycleDecision] = []
+        for skill in skills:
+            previous_status = skill.status
+            action = "no_change"
+            reason = "Review completed without status change."
+
+            if skill.status == "shadow" and float(skill.effectiveness or 0.0) < 0.45:
+                skill.status = "deprecated"
+                action = "degraded_to_deprecated"
+                reason = "Shadow skill review found low effectiveness."
+            elif skill.status == "deprecated" and float(skill.effectiveness or 0.0) < 0.2:
+                skill.status = "retired"
+                action = "retired"
+                reason = "Deprecated skill remained weak through review."
+
+            skill.review_after = review_time + timedelta(days=7)
+            decisions.append(
+                SkillLifecycleDecision(
+                    skill_name=skill.name,
+                    action=action,
+                    previous_status=previous_status,
+                    new_status=skill.status,
+                    reason=reason,
+                    confidence=skill.effectiveness,
+                    review_after=skill.review_after,
+                )
+            )
+
+        await self.db.commit()
+        return decisions
+
+    @staticmethod
+    def _build_candidate_source_hash(candidate: SkillCandidate) -> str:
+        raw = (
+            f"{candidate.name}|{candidate.summary}|"
+            f"{','.join(candidate.source_pattern_ids)}|"
+            f"{len(candidate.evidence_refs)}"
+        ).encode("utf-8")
+        return hashlib.sha1(raw).hexdigest()

backend/app/services/system_service.py

@@ -4,6 +4,9 @@ import os
 import platform
 import socket
 import subprocess
+import time
+
+import httpx
 
 try:
     import psutil
@@ -15,6 +18,10 @@ class SystemService:
     _last_net_bytes_sent: int | None = None
     _last_net_bytes_recv: int | None = None
     _last_net_sample_at: float | None = None
+    _weather_cache: dict | None = None
+    _weather_cached_at: float | None = None
+    _weather_cached_location: str | None = None
+    _weather_cache_ttl_seconds: float = 10 * 60  # 10 minutes
 
     def __init__(self):
         # Import settings here to avoid circular imports
@@ -134,8 +141,95 @@ class SystemService:
             'timestamp': datetime.now(UTC).isoformat(),
         }
 
-    def get_config(self) -> dict:
+    async def _fetch_weather(self, location: str) -> dict:
+        try:
+            timeout = httpx.Timeout(10.0, connect=5.0)
+            async with httpx.AsyncClient(timeout=timeout) as client:
+                response = await client.get(f'https://wttr.in/{location}', params={'format': 'j1'})
+                response.raise_for_status()
+            payload = response.json()
+            current = (payload.get('current_condition') or [{}])[0]
+            weather_code = current.get('weatherCode')
+            temp = current.get('temp_C')
+            parsed_code = int(weather_code) if weather_code is not None and str(weather_code).isdigit() else None
+            if parsed_code is None or temp in (None, ''):
+                return {'weather_code': None, 'weather_summary': 'Weather unavailable'}
+
+            label = self._weather_code_label(parsed_code)
+            return {
+                'weather_code': parsed_code,
+                'weather_summary': f'{label} {temp}°C',
+            }
+        except (httpx.HTTPError, ValueError, TypeError):
+            return {'weather_code': None, 'weather_summary': 'Weather unavailable'}
+
+    @staticmethod
+    def _weather_code_label(code: int | None) -> str:
+        if code == 0:
+            return 'Clear'
+        if code in {1, 2}:
+            return 'Partly Cloudy'
+        if code == 3:
+            return 'Overcast'
+        if code in {45, 48}:
+            return 'Fog'
+        if code in {51, 53, 55, 56, 57}:
+            return 'Drizzle'
+        if code in {61, 63, 65, 66, 67, 80, 81, 82}:
+            return 'Rain'
+        if code in {71, 73, 75, 77, 85, 86}:
+            return 'Snow'
+        if code in {95, 96, 99}:
+            return 'Thunderstorm'
+        return 'Weather'
+
+    async def get_config(self) -> dict:
         """Get public system configuration."""
+        location = self._settings.LOCATION
+        now = time.time()
+        cached_weather = self.__class__._weather_cache
+        cached_at = self.__class__._weather_cached_at
+        cached_location = self.__class__._weather_cached_location
+
+        cache_is_valid = (
+            cached_weather is not None
+            and cached_at is not None
+            and cached_location == location
+            and (now - cached_at) < self.__class__._weather_cache_ttl_seconds
+        )
+
+        if cache_is_valid:
+            return {
+                'location': location,
+                **cached_weather,
+                'weather_cached': True,
+                'weather_cached_at': cached_at,
+            }
+
+        weather = await self._fetch_weather(location)
+
+        # If fetch failed but we have *any* last known weather for same location, return it to avoid UI flicker.
+        if (
+            (weather.get('weather_code') is None)
+            and cached_weather is not None
+            and cached_location == location
+        ):
+            return {
+                'location': location,
+                **cached_weather,
+                'weather_cached': True,
+                'weather_cached_at': cached_at,
+                'weather_stale': True,
+            }
+
+        # Update cache on successful/meaningful payload (or keep "unavailable" if never succeeded).
+        self.__class__._weather_cache = weather
+        self.__class__._weather_cached_at = now
+        self.__class__._weather_cached_location = location
+
         return {
-            'location': self._settings.LOCATION,
+            'location': location,
+            **weather,
+            'weather_cached': False,
+            'weather_cached_at': now,
         }

backend/tests/backend/app/agents/test_graph.py

@@ -314,6 +314,22 @@ class FailIfCalledLLM:
         raise AssertionError('LLM should not be called for simple greetings')
 
 
+class InternalMarkupRecoveryLLM:
+    def __init__(self, responses: list[str]):
+        self.responses = responses
+        self.calls = 0
+        self._jarvis_provider_capabilities = SimpleNamespace(
+            provider='minimax',
+            supports_native_tools=False,
+            preferred_tool_strategy='json_fallback',
+        )
+
+    async def ainvoke(self, messages):
+        self.calls += 1
+        index = min(self.calls - 1, len(self.responses) - 1)
+        return AIMessage(content=self.responses[index])
+
+
 def test_initial_state_sets_structured_continuity_defaults():
     state = initial_state('u1', 'c1')
 
@@ -2047,6 +2063,75 @@ async def test_run_sub_commander_uses_web_search_in_json_fallback(monkeypatch):
     assert result['final_response'] == '我查了外部网页，下面是最新结果摘要。'
 
 
+async def test_run_sub_commander_recovers_from_internal_tool_markup_after_tool_round(monkeypatch):
+    fake_llm = InternalMarkupRecoveryLLM([
+        '{"mode":"tool_call","tool_calls":[{"name":"web_search","arguments":{"query":"武汉 介绍","top_k":2}}]}',
+        '我来让知识管理员为你整理武汉的详细介绍。\n\n分发说明：这个问题需要调用知识库信息，由 librarian（知识管理员）处理最合适。\n<minimax:tool_call>\n<invoke name="librarian">\n<parameter name="info_type">city_introduction</parameter>\n<parameter name="parameters">{"city":"武汉","word_count":2000,"language":"zh-CN"}</parameter>\n</invoke>\n</minimax:tool_call>',
+        '武汉是湖北省省会，位于长江与汉江交汇处，是中部重要的交通、科教和工业中心。',
+    ])
+    fake_tool = FakeTool('web_search', 'found 2 web results')
+
+    monkeypatch.setattr('app.agents.graph._get_llm_for_state', lambda state: fake_llm)
+    monkeypatch.setitem(
+        __import__('app.agents.graph', fromlist=['SUB_COMMANDER_TOOLSETS']).SUB_COMMANDER_TOOLSETS,
+        'librarian_retrieval',
+        [fake_tool],
+    )
+
+    state = _base_state('请介绍一下武汉', {'provider': 'openai', 'model': 'MiniMax-M2.7-highspeed', 'base_url': 'https://api.minimaxi.com/v1'})
+    state['current_agent'] = AgentRole.LIBRARIAN
+    state['max_retries'] = 1
+
+    result = await _run_sub_commander(
+        state,
+        AgentRole.LIBRARIAN,
+        'manager prompt',
+        '请介绍一下武汉',
+        use_tools=True,
+        summary_target='knowledge_context',
+    )
+
+    assert fake_llm.calls == 3
+    assert fake_tool.invocations == [{'query': '武汉 介绍', 'top_k': 2}]
+    assert result['fallback_parse_error'] is None
+    assert '<invoke name=' not in result['final_response']
+    assert '分发说明' not in result['final_response']
+    assert result['final_response'] == '武汉是湖北省省会，位于长江与汉江交汇处，是中部重要的交通、科教和工业中心。'
+
+
+async def test_run_sub_commander_falls_back_to_tool_summary_when_internal_markup_persists(monkeypatch):
+    fake_llm = InternalMarkupRecoveryLLM([
+        '{"mode":"tool_call","tool_calls":[{"name":"web_search","arguments":{"query":"武汉 介绍","top_k":2}}]}',
+        '分发说明：交给 librarian。\n<minimax:tool_call><invoke name="librarian"></invoke></minimax:tool_call>',
+    ])
+    fake_tool = FakeTool('web_search', 'found 2 web results')
+
+    monkeypatch.setattr('app.agents.graph._get_llm_for_state', lambda state: fake_llm)
+    monkeypatch.setitem(
+        __import__('app.agents.graph', fromlist=['SUB_COMMANDER_TOOLSETS']).SUB_COMMANDER_TOOLSETS,
+        'librarian_retrieval',
+        [fake_tool],
+    )
+
+    state = _base_state('请介绍一下武汉', {'provider': 'openai', 'model': 'MiniMax-M2.7-highspeed', 'base_url': 'https://api.minimaxi.com/v1'})
+    state['current_agent'] = AgentRole.LIBRARIAN
+    state['max_retries'] = 0
+
+    result = await _run_sub_commander(
+        state,
+        AgentRole.LIBRARIAN,
+        'manager prompt',
+        '请介绍一下武汉',
+        use_tools=True,
+        summary_target='knowledge_context',
+    )
+
+    assert fake_llm.calls == 2
+    assert result['fallback_parse_error'] == 'internal_tool_markup'
+    assert result['final_response'] == '我已经完成检索，直接给您可用信息：\n\nfound 2 web results'
+    assert '<invoke name=' not in result['final_response']
+
+
 async def test_run_sub_commander_supports_multiple_json_fallback_tool_rounds(monkeypatch):
     fake_llm = TripleResponseFallbackLLM([
         '{"mode":"tool_call","tool_calls":[{"name":"web_search","arguments":{"query":"Jarvis 最新模型更新","top_k":2}}]}',

backend/tests/backend/app/agents/test_result_merge_runtime.py

@@ -0,0 +1,138 @@
+import app.agents.graph as graph_module
+from app.agents.orchestration.result_merge import merge_task_results
+from app.agents.schemas.task import AgentTask
+from app.agents.state import AgentRole, initial_state
+
+
+def test_merge_task_results_marks_conflict_for_distinct_completed_summaries():
+    report = merge_task_results(
+        [
+            {
+                "task_id": "task-1",
+                "status": "completed",
+                "summary": "结论 A",
+                "evidence": [{"type": "source"}],
+                "owner_agent_id": "librarian",
+            },
+            {
+                "task_id": "task-2",
+                "status": "completed",
+                "summary": "结论 B",
+                "evidence": [{"type": "analysis"}, {"type": "analysis"}],
+                "owner_agent_id": "analyst",
+            },
+        ]
+    )
+
+    assert report.status == "conflicted"
+    assert "multiple_distinct_completed_summaries" in report.conflict_flags
+    assert report.resolution_strategy == "rank_by_evidence_count"
+    assert report.resolved_summary == "结论 B"
+
+
+def test_verify_collaboration_results_persists_merge_and_verification_reports():
+    state = initial_state("u1", "c1")
+    tasks = [
+        AgentTask(
+            task_id="task-1",
+            title="收集证据",
+            role=AgentRole.LIBRARIAN.value,
+            owner_agent_id=AgentRole.LIBRARIAN.value,
+            goal="检索资料",
+            expected_evidence=[{"type": "evidence"}],
+        ),
+        AgentTask(
+            task_id="task-2",
+            title="给出分析",
+            role=AgentRole.ANALYST.value,
+            owner_agent_id=AgentRole.ANALYST.value,
+            goal="分析风险",
+            expected_evidence=[{"type": "analysis"}],
+        ),
+    ]
+
+    graph_module._verify_collaboration_results(
+        state,
+        tasks,
+        task_results=[
+            {
+                "task_id": "task-1",
+                "status": "completed",
+                "summary": "证据显示风险中等",
+                "evidence": [{"type": "evidence"}],
+                "owner_agent_id": AgentRole.LIBRARIAN.value,
+            },
+            {
+                "task_id": "task-2",
+                "status": "completed",
+                "summary": "证据显示风险中等",
+                "evidence": [{"type": "analysis"}],
+                "owner_agent_id": AgentRole.ANALYST.value,
+            },
+        ],
+    )
+
+    assert state["merge_report"] is not None
+    assert state["merge_report"]["status"] == "merged"
+    assert state["verification_report"] is not None
+    assert state["verification_report"]["status"] == "passed"
+    event_types = [item["event_type"] for item in state["event_trace"]]
+    assert "agent.merge.completed" in event_types
+    assert "agent.verify.completed" in event_types
+
+
+def test_serial_fallback_response_is_used_when_merge_report_requires_fallback():
+    state = initial_state("u1", "c1")
+    tasks = [
+        AgentTask(
+            task_id="task-1",
+            title="收集证据",
+            role=AgentRole.LIBRARIAN.value,
+            owner_agent_id=AgentRole.LIBRARIAN.value,
+            goal="检索资料",
+            expected_evidence=[{"type": "evidence"}],
+        ),
+        AgentTask(
+            task_id="task-2",
+            title="给出分析",
+            role=AgentRole.ANALYST.value,
+            owner_agent_id=AgentRole.ANALYST.value,
+            goal="分析风险",
+            expected_evidence=[{"type": "analysis"}],
+        ),
+    ]
+    state["task_results"] = [
+        {
+            "task_id": "task-1",
+            "status": "completed",
+            "summary": "已确认可用证据",
+            "evidence": [{"type": "evidence"}],
+            "owner_agent_id": AgentRole.LIBRARIAN.value,
+        },
+        {
+            "task_id": "task-2",
+            "status": "failed",
+            "summary": "分析失败",
+            "evidence": [{"type": "analysis"}],
+            "owner_agent_id": AgentRole.ANALYST.value,
+        },
+    ]
+    state["final_response"] = "原始协作汇总"
+
+    graph_module._verify_collaboration_results(state, tasks, state["task_results"])
+    if state["verification_status"] == "failed" and state["merge_report"]["fallback_used"]:
+        state["final_response"] = graph_module._build_serial_fallback_response(
+            "先查资料再分析",
+            state["task_results"],
+            state["merge_report"],
+        )
+        graph_module._append_event_trace(
+            state,
+            "agent.rollback.triggered",
+            payload={"layer": "collaboration_runtime", "reason": "merge_fallback_used"},
+            severity="warning",
+        )
+
+    assert "切回保守收敛路径" in state["final_response"]
+    event_types = [item["event_type"] for item in state["event_trace"]]
+    assert "agent.rollback.triggered" in event_types

backend/tests/backend/app/agents/test_runtime_context.py

@@ -0,0 +1,170 @@
+from langchain_core.messages import HumanMessage
+
+import app.agents.graph as graph_module
+from app.agents.graph import master_node
+from app.agents.learning.retrospector import build_session_retrospective
+from app.agents.schemas.orchestration import (
+    RuntimeRequestContext,
+    assess_parallel_worthiness,
+    render_runtime_request_context_summary,
+)
+from app.agents.schemas.skills import SkillShortlistEntry
+from app.agents.state import initial_state
+
+
+def test_runtime_request_context_summary_renders_parallel_and_shortlists():
+    worthiness = assess_parallel_worthiness(
+        "先分析需求，再查资料，同时整理成计划",
+        retrospective_count=1,
+        skill_count=2,
+    )
+    context = RuntimeRequestContext(
+        user_id="u1",
+        session_id="c1",
+        query_text="先分析需求，再查资料，同时整理成计划",
+        recalled_memories=["最近偏好结构化输出"],
+        retrospective_shortlist=[
+            {
+                "task_type": "analysis",
+                "summary": "上次先检索再分析，结果更稳。",
+            }
+        ],
+        shortlisted_skills=["weekly-planning"],
+        skill_shortlist=[
+            SkillShortlistEntry(
+                skill_name="weekly-planning",
+                score=0.82,
+                rationale="命中计划关键词",
+                injection_mode="summary",
+            )
+        ],
+        parallel_worthiness=worthiness,
+        recommended_runtime_mode="collaboration",
+    )
+
+    summary = render_runtime_request_context_summary(context)
+
+    assert "Runtime Request Context" in summary
+    assert "collaboration" in summary
+    assert "weekly-planning" in summary
+
+
+def test_build_session_retrospective_captures_skill_and_history_context():
+    retrospective = build_session_retrospective(
+        request_id="resp-1",
+        session_id="conv-1",
+        user_query="帮我分析并安排下周任务",
+        state={
+            "execution_mode": "collaboration",
+            "current_agent": "analyst",
+            "verification_status": "passed",
+            "verification_summary": "ok",
+            "final_response": "已经给出建议",
+            "skill_shortlist": [{"skill_name": "weekly-planning"}],
+            "event_trace": [{"event_type": "agent.execution.decided", "agent_id": "master"}],
+            "verification_evidence": [{"type": "verification"}],
+            "completed_tasks": [{"task_id": "t1", "title": "收集信息", "status": "completed"}],
+            "retrospective_shortlist": [{"summary": "上次周计划拆解有效"}],
+            "parallel_worthiness": {"score": 0.6},
+        },
+        runtime_context={
+            "user_id": "u1",
+            "recommended_runtime_mode": "collaboration",
+        },
+    )
+
+    assert retrospective.user_id == "u1"
+    assert retrospective.execution_mode == "collaboration"
+    assert retrospective.used_skill_names == ["weekly-planning"]
+    assert retrospective.context_snapshot["retrospective_shortlist_count"] == 1
+    assert retrospective.outcome == "completed"
+
+
+async def test_master_node_records_execution_decision_and_skill_shortlist_event():
+    state = initial_state("u1", "c1")
+    state["messages"] = [HumanMessage(content="帮我查一下资料并分析重点")]
+    state["skill_shortlist"] = [
+        {
+            "skill_name": "research-synthesis",
+            "score": 0.73,
+            "injection_mode": "summary",
+        }
+    ]
+    state["runtime_request_context"] = {
+        "request_id": "req-1",
+        "recommended_runtime_mode": "direct",
+        "parallel_worthiness": {
+            "preferred_mode": "direct",
+            "score": 0.2,
+            "estimated_subtasks": 1,
+        },
+    }
+    state["task_graph"] = {
+        "graph_id": "graph-1",
+        "nodes": [{"node_id": "task-1", "title": "收集证据", "role": "librarian"}],
+        "entry_node_ids": ["task-1"],
+        "max_parallelism": 1,
+    }
+
+    result = await master_node(state)
+
+    assert result["execution_decision"] is not None
+    event_types = [item["event_type"] for item in result["event_trace"]]
+    assert "agent.parallel.assessed" in event_types
+    assert "agent.skill.shortlisted" in event_types
+    assert "agent.task_graph.built" in event_types
+    assert "agent.execution.decided" in event_types
+
+
+async def test_master_node_records_rollback_event_when_parallel_task_graph_flag_is_disabled():
+    async def fake_collaboration_flow(state, _user_query):
+        state["execution_mode"] = "collaboration"
+        state["final_response"] = "collaboration skipped in test"
+        return state
+
+    graph_module._run_collaboration_flow = fake_collaboration_flow
+    state = initial_state("u1", "c1")
+    state["messages"] = [HumanMessage(content="先查资料再分析风险再安排计划")]
+    state["feature_flags"] = {"ENABLE_PARALLEL_TASK_GRAPH": False}
+    state["parallel_worthiness"] = {
+        "preferred_mode": "parallel",
+        "score": 0.8,
+        "estimated_subtasks": 3,
+    }
+    state["runtime_request_context"] = {
+        "request_id": "req-2",
+        "user_id": "u1",
+        "session_id": "c1",
+        "recommended_runtime_mode": "collaboration",
+    }
+
+    result = await master_node(state)
+
+    event_types = [item["event_type"] for item in result["event_trace"]]
+    assert "agent.rollback.triggered" in event_types
+
+
+async def test_master_node_direct_mode_baseline_still_returns_simple_response():
+    state = initial_state("u1", "c1")
+    state["messages"] = [HumanMessage(content="你好")]
+
+    result = await master_node(state)
+
+    assert result["execution_mode"] == "direct"
+    assert result["final_response"] is not None
+
+
+async def test_master_node_collaboration_mode_baseline_still_respects_complex_request(monkeypatch):
+    async def fake_collaboration_flow(state, _user_query):
+        state["execution_mode"] = "collaboration"
+        state["final_response"] = "collaboration baseline ok"
+        return state
+
+    graph_module._run_collaboration_flow = fake_collaboration_flow
+    state = initial_state("u1", "c1")
+    state["messages"] = [HumanMessage(content="先查资料，再分析风险，再安排计划")]
+
+    result = await master_node(state)
+
+    assert result["execution_mode"] == "collaboration"
+    assert result["final_response"] == "collaboration baseline ok"

backend/tests/backend/app/agents/test_scheduler_runtime.py

@@ -0,0 +1,66 @@
+from langchain_core.messages import HumanMessage
+
+import app.agents.graph as graph_module
+from app.agents.orchestration.scheduler import build_subtask_specs
+from app.agents.orchestration.task_graph import build_bounded_task_graph
+from app.agents.schemas.orchestration import TaskGraph, assess_parallel_worthiness
+from app.agents.state import initial_state
+
+
+def test_build_subtask_specs_keeps_dependencies_and_contract_fields():
+    worthiness = assess_parallel_worthiness(
+        "先查资料、再分析风险、再安排下周计划",
+        retrospective_count=2,
+        skill_count=1,
+    )
+    task_graph = build_bounded_task_graph(
+        query_text="先查资料、再分析风险、再安排下周计划",
+        parallel_worthiness=worthiness,
+    )
+
+    specs = build_subtask_specs(task_graph, query_text="先查资料、再分析风险、再安排下周计划")
+
+    assert specs
+    assert all(spec.parent_run_id == task_graph.graph_id for spec in specs)
+    assert all(isinstance(spec.context_slice, dict) for spec in specs)
+    assert all(spec.expected_output_schema for spec in specs)
+    assert any(spec.dependencies for spec in specs)
+
+
+async def test_run_collaboration_flow_uses_task_graph_plan_and_records_subtask_events(monkeypatch):
+    async def fake_run_sub_commander(
+        state,
+        assigned_role,
+        _system_prompt,
+        task_goal,
+        **_kwargs,
+    ):
+        state["final_response"] = f"{assigned_role.value} handled: {task_goal}"
+        return state
+
+    monkeypatch.setattr(graph_module, "_run_sub_commander", fake_run_sub_commander)
+
+    state = initial_state("u1", "c1")
+    state["messages"] = [HumanMessage(content="先查资料、再分析风险、再安排下周计划")]
+    state["current_datetime_context"] = "CURRENT_TIME: 2026-03-28T12:00:00+08:00"
+    state["task_graph"] = TaskGraph.model_validate(
+        build_bounded_task_graph(
+            query_text="先查资料、再分析风险、再安排下周计划",
+            parallel_worthiness=assess_parallel_worthiness(
+                "先查资料、再分析风险、再安排下周计划",
+                retrospective_count=2,
+                skill_count=1,
+            ),
+        ).model_dump(mode="json")
+    ).model_dump(mode="json")
+
+    result = await graph_module._run_collaboration_flow(
+        state,
+        "先查资料、再分析风险、再安排下周计划",
+    )
+
+    assert result["scheduled_subtasks"]
+    event_types = [item["event_type"] for item in result["event_trace"]]
+    assert "agent.subtask.started" in event_types
+    assert "agent.subtask.completed" in event_types
+    assert result["task_results"]

backend/tests/backend/app/agents/test_task_graph_runtime.py

@@ -0,0 +1,59 @@
+from app.agents.orchestration.task_graph import build_bounded_task_graph
+from app.agents.schemas.orchestration import RuntimeRequestContext, assess_parallel_worthiness, render_runtime_request_context_summary
+
+
+def test_build_bounded_task_graph_creates_independent_nodes_and_merge_node():
+    worthiness = assess_parallel_worthiness(
+        "先查资料、再分析风险、再安排下周计划",
+        retrospective_count=2,
+        skill_count=1,
+    )
+
+    graph = build_bounded_task_graph(
+        query_text="先查资料、再分析风险、再安排下周计划",
+        parallel_worthiness=worthiness,
+    )
+
+    assert graph is not None
+    assert len(graph.nodes) >= 2
+    assert graph.entry_node_ids
+    assert any(node.execution_mode == "parallel" for node in graph.nodes[:-1])
+    assert graph.nodes[-1].role == "master"
+
+
+def test_runtime_request_context_summary_renders_task_graph():
+    worthiness = assess_parallel_worthiness(
+        "先查资料、再分析风险、再安排下周计划",
+        retrospective_count=1,
+        skill_count=1,
+    )
+    task_graph = build_bounded_task_graph(
+        query_text="先查资料、再分析风险、再安排下周计划",
+        parallel_worthiness=worthiness,
+    )
+    context = RuntimeRequestContext(
+        user_id="u1",
+        session_id="c1",
+        query_text="先查资料、再分析风险、再安排下周计划",
+        parallel_worthiness=worthiness,
+        task_graph=task_graph,
+        recommended_runtime_mode="collaboration",
+    )
+
+    summary = render_runtime_request_context_summary(context)
+
+    assert "任务图" in summary
+    assert "max_parallelism" in summary
+
+
+def test_runtime_request_context_summary_renders_assembly_metrics():
+    context = RuntimeRequestContext(
+        user_id="u1",
+        session_id="c1",
+        query_text="帮我分析一下资料",
+        assembly_metrics={"total_ms": 12.3},
+    )
+
+    summary = render_runtime_request_context_summary(context)
+
+    assert "上下文装配耗时" in summary

backend/tests/backend/app/agents/test_visibility_api.py

@@ -503,7 +503,9 @@ async def test_visibility_tools_returns_governance_metadata_and_usage_counts(vis
     payload = response.json()
     assert payload['total_tools'] >= 1
     assert payload['used_tools'] >= 1
-    search_tool = next(item for item in payload['items'] if item['tool_name'] == 'search_web')
+    search_tool = next(
+        item for item in payload['items'] if item['tool_name'] in {'search_web', 'web_search'}
+    )
     assert search_tool['permission_class'] == 'external'
     assert search_tool['side_effect_scope'] == 'network'
     assert search_tool['usage_count'] == 1
@@ -516,6 +518,26 @@ async def test_visibility_tools_returns_governance_metadata_and_usage_counts(vis
     ]
 
 
+@pytest.mark.asyncio
+async def test_visibility_debug_returns_observability_and_learning_views(visibility_env):
+    app, ids = visibility_env
+    transport = ASGITransport(app=app)
+    async with AsyncClient(transport=transport, base_url='http://testserver') as client:
+        response = await client.get(
+            '/api/agents/visibility/debug',
+            params={'conversation_id': ids['conversation_id']},
+        )
+
+    assert response.status_code == 200
+    payload = response.json()
+    assert payload['conversation_id'] == ids['conversation_id']
+    assert 'observability' in payload
+    assert 'skill_shortlist' in payload
+    assert 'retrospective_shortlist' in payload
+    assert 'recent_retrospectives' in payload
+    assert 'recent_learning_artifacts' in payload
+
+
 @pytest.mark.asyncio
 async def test_visibility_events_reject_invalid_datetime(visibility_env):
     app, ids = visibility_env

backend/tests/backend/app/test_conversation_router.py

@@ -73,3 +73,25 @@ async def test_list_conversations_succeeds_when_agent_state_column_was_missing(c
     assert len(payload) == 1
     assert payload[0]['title'] == 'Existing conversation'
     assert payload[0]['message_count'] == 3
+
+
+@pytest.mark.asyncio
+async def test_chat_stream_emits_error_event_when_agent_service_fails_before_stream_starts(
+    conversation_env,
+    monkeypatch,
+):
+    async def fail_chat(*args, **kwargs):
+        raise RuntimeError('stream boot failed')
+
+    monkeypatch.setattr('app.routers.conversation.AgentService.chat', fail_chat)
+
+    transport = ASGITransport(app=conversation_env)
+    async with AsyncClient(transport=transport, base_url='http://testserver') as client:
+        response = await client.post(
+            '/api/conversations/chat/stream',
+            json={'message': 'hello'},
+        )
+
+    assert response.status_code == 200
+    assert 'event: error' in response.text
+    assert 'stream boot failed' in response.text

backend/tests/backend/app/test_database_schema_bootstrap.py

@@ -0,0 +1,115 @@
+import pytest
+from sqlalchemy import text
+from sqlalchemy.ext.asyncio import create_async_engine
+
+from app.database import ensure_learning_artifact_tables, ensure_memory_columns, ensure_skill_columns
+
+
+@pytest.mark.anyio
+async def test_ensure_memory_columns_adds_importance_tracking_fields_for_existing_user_memories_table(tmp_path):
+    db_path = tmp_path / 'test_user_memories.db'
+    engine = create_async_engine(f"sqlite+aiosqlite:///{db_path}", future=True)
+
+    async with engine.begin() as conn:
+        await conn.execute(text(
+            '''
+            CREATE TABLE user_memories (
+                id VARCHAR(36) PRIMARY KEY,
+                user_id VARCHAR(36) NOT NULL,
+                memory_type VARCHAR(50) NOT NULL,
+                content TEXT NOT NULL,
+                importance INTEGER,
+                is_recalled BOOLEAN,
+                recall_count INTEGER,
+                source_conversation_id VARCHAR(36),
+                extracted_at DATETIME,
+                last_recalled_at DATETIME,
+                created_at DATETIME,
+                updated_at DATETIME
+            )
+            '''
+        ))
+        result = await conn.execute(text("PRAGMA table_info(user_memories)"))
+        columns_before = {row[1] for row in result.fetchall()}
+        assert 'frequency_count' not in columns_before
+        assert 'importance_score' not in columns_before
+        assert 'decay_score' not in columns_before
+
+        await ensure_memory_columns(conn)
+
+        result = await conn.execute(text("PRAGMA table_info(user_memories)"))
+        columns_after = {row[1] for row in result.fetchall()}
+        assert 'frequency_count' in columns_after
+        assert 'emotion_tags' in columns_after
+        assert 'importance_score' in columns_after
+        assert 'importance_level' in columns_after
+        assert 'associated_topics' in columns_after
+        assert 'decay_score' in columns_after
+        assert 'is_archived' in columns_after
+        assert 'last_accessed_at' in columns_after
+        assert 'archive_at' in columns_after
+
+    await engine.dispose()
+
+
+@pytest.mark.anyio
+async def test_ensure_skill_columns_adds_lifecycle_fields_for_existing_skills_table(tmp_path):
+    db_path = tmp_path / 'test_skills.db'
+    engine = create_async_engine(f"sqlite+aiosqlite:///{db_path}", future=True)
+
+    async with engine.begin() as conn:
+        await conn.execute(text(
+            '''
+            CREATE TABLE skills (
+                id VARCHAR(36) PRIMARY KEY,
+                name VARCHAR(100) NOT NULL,
+                description TEXT,
+                instructions TEXT NOT NULL,
+                agent_type VARCHAR(50) NOT NULL,
+                visibility VARCHAR(20),
+                is_active BOOLEAN,
+                owner_id VARCHAR(36),
+                created_at DATETIME,
+                updated_at DATETIME
+            )
+            '''
+        ))
+        result = await conn.execute(text("PRAGMA table_info(skills)"))
+        columns_before = {row[1] for row in result.fetchall()}
+        assert 'status' not in columns_before
+        assert 'effectiveness' not in columns_before
+
+        await ensure_skill_columns(conn)
+
+        result = await conn.execute(text("PRAGMA table_info(skills)"))
+        columns_after = {row[1] for row in result.fetchall()}
+        assert 'status' in columns_after
+        assert 'scope' in columns_after
+        assert 'effectiveness' in columns_after
+        assert 'review_after' in columns_after
+        assert 'activation_count' in columns_after
+        assert 'last_activated_at' in columns_after
+
+    await engine.dispose()
+
+
+@pytest.mark.anyio
+async def test_ensure_learning_artifact_tables_creates_table_and_indexes(tmp_path):
+    db_path = tmp_path / 'test_learning_artifacts.db'
+    engine = create_async_engine(f"sqlite+aiosqlite:///{db_path}", future=True)
+
+    async with engine.begin() as conn:
+        await ensure_learning_artifact_tables(conn)
+        result = await conn.execute(text("PRAGMA table_info(learning_artifacts)"))
+        columns = {row[1] for row in result.fetchall()}
+        assert 'artifact_type' in columns
+        assert 'artifact_key' in columns
+        assert 'summary_text' in columns
+        assert 'payload' in columns
+
+        indexes = await conn.execute(text("PRAGMA index_list(learning_artifacts)"))
+        index_names = {row[1] for row in indexes.fetchall()}
+        assert 'ix_learning_artifacts_user_id' in index_names
+        assert 'ix_learning_artifacts_artifact_type' in index_names
+
+    await engine.dispose()

backend/tests/backend/app/test_skill_router.py

@@ -54,6 +54,9 @@ async def skill_env(tmp_path, monkeypatch):
                 required_context=[],
                 visibility='private',
                 is_active=True,
+                status='active',
+                scope=['schedule_planner'],
+                effectiveness=0.88,
                 owner_id=user.id,
             ),
             Skill(
@@ -65,6 +68,9 @@ async def skill_env(tmp_path, monkeypatch):
                 required_context=[],
                 visibility='private',
                 is_active=True,
+                status='shadow',
+                scope=['executor'],
+                effectiveness=0.41,
                 owner_id=user.id,
             ),
             Skill(
@@ -76,6 +82,8 @@ async def skill_env(tmp_path, monkeypatch):
                 required_context=[],
                 visibility='private',
                 is_active=True,
+                status='active',
+                scope=['schedule_planner'],
                 owner_id=other_user.id,
             ),
         ])
@@ -188,3 +196,9 @@ async def test_list_skills_without_agent_type_returns_current_user_skills(skill_
     assert all(isinstance(item['updated_at'], str) for item in payload)
     assert all('is_builtin' in item for item in payload)
     assert all(item['is_builtin'] is False for item in payload)
+    assert all('status' in item for item in payload)
+    assert all('scope' in item for item in payload)
+    assert any(item['status'] == 'shadow' for item in payload)
+    executor = next(item for item in payload if item['name'] == 'Executor skill')
+    assert executor['scope'] == ['executor']
+    assert executor['effectiveness'] == 0.41

backend/tests/backend/app/test_system_router.py

@@ -0,0 +1,130 @@
+import httpx
+import pytest
+from httpx import ASGITransport, AsyncClient
+
+from app.main import app
+
+
+@pytest.mark.asyncio
+async def test_system_config_returns_location_and_weather(monkeypatch):
+    async def fake_get_config(self):
+        return {
+            'location': 'wuhan',
+            'weather_code': 3,
+            'weather_summary': 'Overcast 22°C',
+        }
+
+    monkeypatch.setattr('app.routers.system.SystemService.get_config', fake_get_config)
+    transport = ASGITransport(app=app)
+
+    async with AsyncClient(transport=transport, base_url='http://testserver') as client:
+        response = await client.get('/api/system/config')
+
+    assert response.status_code == 200
+    assert response.json() == {
+        'location': 'wuhan',
+        'weather_code': 3,
+        'weather_summary': 'Overcast 22°C',
+    }
+
+
+@pytest.mark.asyncio
+async def test_system_config_gracefully_returns_unavailable_weather(monkeypatch):
+    async def fake_get_config(self):
+        return {
+            'location': 'wuhan',
+            'weather_code': None,
+            'weather_summary': 'Weather unavailable',
+        }
+
+    monkeypatch.setattr('app.routers.system.SystemService.get_config', fake_get_config)
+    transport = ASGITransport(app=app)
+
+    async with AsyncClient(transport=transport, base_url='http://testserver') as client:
+        response = await client.get('/api/system/config')
+
+    assert response.status_code == 200
+    assert response.json() == {
+        'location': 'wuhan',
+        'weather_code': None,
+        'weather_summary': 'Weather unavailable',
+    }
+
+
+class FakeWeatherResponse:
+    def __init__(self, payload: dict, status_code: int = 200):
+        self._payload = payload
+        self.status_code = status_code
+
+    def raise_for_status(self):
+        if self.status_code >= 400:
+            raise httpx.HTTPStatusError(
+                'request failed',
+                request=httpx.Request('GET', 'https://wttr.in/wuhan?format=j1'),
+                response=httpx.Response(self.status_code, request=httpx.Request('GET', 'https://wttr.in/wuhan?format=j1')),
+            )
+
+    def json(self):
+        return self._payload
+
+
+class FakeAsyncClient:
+    def __init__(self, *, response=None, error=None, **kwargs):
+        self._response = response
+        self._error = error
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, exc_type, exc, tb):
+        return False
+
+    async def get(self, url, *, params=None):
+        if self._error is not None:
+            raise self._error
+        return self._response
+
+
+@pytest.mark.asyncio
+async def test_system_service_get_config_fetches_weather(monkeypatch):
+    monkeypatch.setattr(
+        'app.services.system_service.httpx.AsyncClient',
+        lambda **kwargs: FakeAsyncClient(
+            response=FakeWeatherResponse({'current_condition': [{'weatherCode': '61', 'temp_C': '18'}]}),
+            **kwargs,
+        ),
+    )
+
+    from app.services.system_service import SystemService
+
+    service = SystemService()
+    monkeypatch.setattr(service._settings, 'LOCATION', 'wuhan')
+
+    payload = await service.get_config()
+
+    assert payload == {
+        'location': 'wuhan',
+        'weather_code': 61,
+        'weather_summary': 'Rain 18°C',
+    }
+
+
+@pytest.mark.asyncio
+async def test_system_service_get_config_handles_weather_failure(monkeypatch):
+    monkeypatch.setattr(
+        'app.services.system_service.httpx.AsyncClient',
+        lambda **kwargs: FakeAsyncClient(error=httpx.TimeoutException('timed out'), **kwargs),
+    )
+
+    from app.services.system_service import SystemService
+
+    service = SystemService()
+    monkeypatch.setattr(service._settings, 'LOCATION', 'wuhan')
+
+    payload = await service.get_config()
+
+    assert payload == {
+        'location': 'wuhan',
+        'weather_code': None,
+        'weather_summary': 'Weather unavailable',
+    }

frontend/.env.example

@@ -1 +0,0 @@
-VITE_API_URL=http://127.0.0.1:3337

frontend/src/api/conversation.ts

@@ -101,7 +101,8 @@ export const conversationApi = {
     handlers: ChatStreamHandlers = {},
   ) {
     const token = localStorage.getItem('access_token')
-    const response = await fetch(`${import.meta.env.VITE_API_URL}/api/conversations/chat/stream`, {
+    const baseURL = import.meta.env.DEV ? '' : import.meta.env.VITE_API_URL
+    const response = await fetch(`${baseURL}/api/conversations/chat/stream`, {
       method: 'POST',
       headers: {
         'Content-Type': 'application/json',

frontend/src/pages/chat/composables/useClientTime.ts

@@ -13,17 +13,53 @@ export function useClientTime() {
   const weatherSummary = ref('Weather unavailable')
   const weatherCode = ref<number | null>(null)
   let clientTimeTimer: ReturnType<typeof setInterval> | null = null
+  const WEATHER_CACHE_KEY = 'jarvis:clientWeather'
 
-  // Load location from backend config
-  async function loadLocation() {
+  function loadWeatherCache() {
+    if (typeof window === 'undefined') return
+    try {
+      const raw = window.localStorage.getItem(WEATHER_CACHE_KEY)
+      if (!raw) return
+      const parsed = JSON.parse(raw) as { city?: string; weather_summary?: string; weather_code?: number | null; cached_at?: number }
+      if (typeof parsed.city === 'string' && parsed.city.trim()) city.value = parsed.city
+      if (typeof parsed.weather_summary === 'string' && parsed.weather_summary.trim()) weatherSummary.value = parsed.weather_summary
+      if (typeof parsed.weather_code === 'number') weatherCode.value = parsed.weather_code
+    } catch {
+      // ignore cache parse errors
+    }
+  }
+
+  function saveWeatherCache(payload: { city: string; weather_summary: string; weather_code: number | null }) {
+    if (typeof window === 'undefined') return
+    try {
+      window.localStorage.setItem(WEATHER_CACHE_KEY, JSON.stringify({ ...payload, cached_at: Date.now() }))
+    } catch {
+      // ignore storage errors
+    }
+  }
+
+  // Read cache synchronously during setup to avoid initial render flicker (icon showing as X/na).
+  loadWeatherCache()
+
+  async function loadSystemConfig() {
     try {
       const response = await fetch('/api/system/config')
-      if (response.ok) {
-        const config = await response.json()
-        city.value = config.location || 'Location'
-      }
+      if (!response.ok) throw new Error('system config request failed')
+      const config = await response.json()
+      city.value = typeof config.location === 'string' && config.location.trim() ? config.location : 'Location'
+      weatherCode.value = typeof config.weather_code === 'number' ? config.weather_code : null
+      weatherSummary.value = typeof config.weather_summary === 'string' && config.weather_summary.trim()
+        ? config.weather_summary
+        : 'Weather unavailable'
+      saveWeatherCache({ city: city.value, weather_summary: weatherSummary.value, weather_code: weatherCode.value })
     } catch {
-      city.value = import.meta.env.VITE_LOCATION || 'Location'
+      // If we already have cached weather on screen, keep it to avoid UI flicker (icon showing as "X"/na).
+      const hasExistingWeather = weatherCode.value !== null || (weatherSummary.value && weatherSummary.value !== 'Weather unavailable')
+      if (!hasExistingWeather) {
+        city.value = 'Location'
+        weatherCode.value = null
+        weatherSummary.value = 'Weather unavailable'
+      }
     }
   }
 
@@ -40,6 +76,27 @@ export function useClientTime() {
   }
 
   function weatherCodeLabel(code: number | null | undefined) {
+    // Backend may return wttr.in condition codes (e.g. 113/116/119...), normalize to user-friendly labels.
+    if (typeof code === 'number' && code > 99) {
+      if (code === 113) return 'Clear'
+      if (code === 116) return 'Partly Cloudy'
+      if (code === 119 || code === 122) return 'Overcast'
+      if (code === 143) return 'Fog'
+      if ([200, 386, 389].includes(code)) return 'Thunderstorm'
+      if (
+        [
+          176, 263, 266, 281, 293, 296, 299, 302, 305, 308,
+          311, 314, 317, 350, 353, 356, 359, 362,
+        ].includes(code)
+      ) return 'Rain'
+      if (
+        [
+          179, 182, 185, 227, 230, 323, 326, 329, 332, 335,
+          338, 368, 371, 374, 377, 392, 395,
+        ].includes(code)
+      ) return 'Snow'
+      return 'Weather'
+    }
     if (code === 0) return 'Clear'
     if (code === 1 || code === 2) return 'Partly Cloudy'
     if (code === 3) return 'Overcast'
@@ -48,11 +105,34 @@ export function useClientTime() {
     if ([61, 63, 65, 66, 67, 80, 81, 82].includes(code ?? -1)) return 'Rain'
     if ([71, 73, 75, 77, 85, 86].includes(code ?? -1)) return 'Snow'
     if ([95, 96, 99].includes(code ?? -1)) return 'Thunderstorm'
-    return 'Weather'
+    return 'Weather unavailable'
   }
 
   const weatherIcon = computed(() => {
     const code = weatherCode.value
+    if (code === null || code === undefined) return ''
+    // Support wttr.in weather codes (commonly 113/116/119/122/143/...).
+    if (typeof code === 'number' && code > 99) {
+      if (code === 113) return 'wi-day-sunny'
+      if (code === 116) return 'wi-day-cloudy'
+      if (code === 119) return 'wi-cloudy'
+      if (code === 122) return 'wi-cloudy'
+      if (code === 143) return 'wi-fog'
+      if ([200, 386, 389].includes(code)) return 'wi-thunderstorm'
+      if (
+        [
+          176, 263, 266, 281, 293, 296, 299, 302, 305, 308,
+          311, 314, 317, 350, 353, 356, 359, 362,
+        ].includes(code)
+      ) return 'wi-rain'
+      if (
+        [
+          179, 182, 185, 227, 230, 323, 326, 329, 332, 335,
+          338, 368, 371, 374, 377, 392, 395,
+        ].includes(code)
+      ) return 'wi-snow'
+      return ''
+    }
     if (code === 0) return 'wi-day-sunny'
     if (code === 1) return 'wi-day-cloudy'
     if (code === 2) return 'wi-day-cloudy-gusts'
@@ -65,56 +145,13 @@ export function useClientTime() {
     if ([66, 67, 81, 82].includes(code ?? -1)) return 'wi-rain'
     if ([71, 73, 75, 77, 85, 86].includes(code ?? -1)) return 'wi-snow'
     if ([95, 96, 99].includes(code ?? -1)) return 'wi-thunderstorm'
-    return 'wi-day-sunny'
+    return ''
   })
 
-  async function loadWeather(latitude: number, longitude: number) {
-    try {
-      // Fetch weather data
-      const weatherResp = await fetch(
-        `https://api.open-meteo.com/v1/forecast?latitude=${latitude}&longitude=${longitude}&current=temperature_2m,weather_code&timezone=auto`,
-      )
-      if (!weatherResp.ok) throw new Error('weather request failed')
-      const weatherData = await weatherResp.json()
-      const current = weatherData.current ?? {}
-      weatherCode.value = typeof current.weather_code === 'number' ? current.weather_code : null
-      const temp = typeof current.temperature_2m === 'number' ? `${Math.round(current.temperature_2m)}°C` : '--'
-      weatherSummary.value = `${weatherCodeLabel(current.weather_code)} ${temp}`
-
-      // Only fetch city name if not already set by config
-      if (city.value === 'Location') {
-        try {
-          const geoResp = await fetch(
-            `https://nominatim.openstreetmap.org/reverse?format=json&lat=${latitude}&lon=${longitude}&zoom=10`,
-          )
-          if (geoResp.ok) {
-            const geoData = await geoResp.json()
-            city.value = geoData.address?.city ?? geoData.address?.town ?? geoData.address?.county ?? geoData.display_name?.split(',')[0] ?? 'Location'
-          }
-        } catch {
-          city.value = 'Location'
-        }
-      }
-    } catch {
-      weatherCode.value = null
-      weatherSummary.value = 'Weather unavailable'
-    }
-  }
-
-  onMounted(() => {
+  onMounted(async () => {
     updateClientTime()
     clientTimeTimer = setInterval(updateClientTime, 1000)
-    void loadLocation()
-    if (!navigator.geolocation) {
-      weatherCode.value = null
-      weatherSummary.value = 'Weather unavailable'
-      return
-    }
-    navigator.geolocation.getCurrentPosition(
-      (position) => { void loadWeather(position.coords.latitude, position.coords.longitude) },
-      () => { weatherCode.value = null; weatherSummary.value = 'Weather unavailable' },
-      { enableHighAccuracy: false, timeout: 8000, maximumAge: 300000 },
-    )
+    await loadSystemConfig()
   })
 
   onUnmounted(() => {
@@ -122,7 +159,15 @@ export function useClientTime() {
   })
 
   return {
-    clientTime, city, weatherSummary, weatherCode, weatherIcon,
-    updateClientTime, formatClientDate, formatClientClock, weatherCodeLabel, loadWeather
+    clientTime,
+    city,
+    weatherSummary,
+    weatherCode,
+    weatherIcon,
+    updateClientTime,
+    formatClientDate,
+    formatClientClock,
+    weatherCodeLabel,
+    loadSystemConfig,
   }
 }

frontend/src/pages/chat/composables/useSidebarPlan.ts

@@ -1,4 +1,4 @@
-import { computed, onMounted, ref, watch, toRef } from 'vue'
+import { computed, onMounted, ref, watch, type Ref } from 'vue'
 import { CornerDownLeft, Database, Sparkles, Sun, ListTodo } from 'lucide-vue-next'
 import { scheduleCenterApi, type ScheduleCenterDateResponse, type ScheduleCenterDaySummary } from '@/api/scheduleCenter'
 import type { Conversation } from '@/api/conversation'
@@ -18,20 +18,24 @@ export const sidebarCollapsedModules = [
   { id: 'review', label: '复盘', icon: CornerDownLeft },
 ]
 
-function formatDateKey(date: Date) {
-  const year = date.getUTCFullYear()
-  const month = String(date.getUTCMonth() + 1).padStart(2, '0')
-  const day = String(date.getUTCDate()).padStart(2, '0')
+export function formatDateKey(date: Date) {
+  const year = date.getFullYear()
+  const month = String(date.getMonth() + 1).padStart(2, '0')
+  const day = String(date.getDate()).padStart(2, '0')
   return `${year}-${month}-${day}`
 }
 
 function formatMonthKey(date: Date) {
-  const year = date.getUTCFullYear()
-  const month = String(date.getUTCMonth() + 1).padStart(2, '0')
+  const year = date.getFullYear()
+  const month = String(date.getMonth() + 1).padStart(2, '0')
   return `${year}-${month}`
 }
 
-export function useSidebarPlan(clientTimeRef: { value: Date }, loadDailyDigestFn: () => void, conversationsRef: Conversation[] = []) {
+export function useSidebarPlan(
+  clientTimeRef: { value: Date },
+  loadDailyDigestFn: () => void,
+  conversationsRef: Ref<Conversation[]> | Conversation[] = [],
+) {
   const todayPlanDetail = ref<ScheduleCenterDateResponse | null>(null)
   const monthPlanDays = ref<ScheduleCenterDaySummary[]>([])
   const selectedDate = ref<string | null>(null)
@@ -41,9 +45,7 @@ export function useSidebarPlan(clientTimeRef: { value: Date }, loadDailyDigestFn
     const map = new Map<string, boolean>()
     const conversations = Array.isArray(conversationsRef) ? conversationsRef : (conversationsRef.value ?? [])
     conversations.forEach((conv) => {
-      const date = new Date(conv.updated_at)
-      const dateKey = formatDateKey(date)
-      map.set(dateKey, true)
+      map.set(formatDateKey(new Date(conv.updated_at)), true)
     })
     return map
   })
@@ -52,22 +54,22 @@ export function useSidebarPlan(clientTimeRef: { value: Date }, loadDailyDigestFn
   const monthPlanSummaryMap = computed(() => new Map(monthPlanDays.value.map((item) => [item.date, item])))
 
   const calendarCells = computed(() => {
-    const year = clientTimeRef.value.getUTCFullYear()
-    const month = clientTimeRef.value.getUTCMonth()
-    const daysInMonth = new Date(Date.UTC(year, month + 1, 0)).getUTCDate()
-    const firstDayOffset = (new Date(Date.UTC(year, month, 1)).getUTCDay() + 6) % 7
-    const today = clientTimeRef.value.getUTCDate()
+    const year = clientTimeRef.value.getFullYear()
+    const month = clientTimeRef.value.getMonth()
+    const daysInMonth = new Date(year, month + 1, 0).getDate()
+    const firstDayOffset = (new Date(year, month, 1).getDay() + 6) % 7
+    const todayKey = todayDateKey.value
     const cells: Array<{ key: string; value: number | null; active: boolean; busy: boolean; selected: boolean; hasConversation: boolean }> = []
     for (let index = 0; index < firstDayOffset; index += 1) {
       cells.push({ key: `blank-start-${index}`, value: null, active: false, busy: false, selected: false, hasConversation: false })
     }
     for (let day = 1; day <= daysInMonth; day += 1) {
-      const monthDate = new Date(Date.UTC(year, month, day))
+      const monthDate = new Date(year, month, day)
       const dateKey = formatDateKey(monthDate)
       const summary = monthPlanSummaryMap.value.get(dateKey)
       const busy = Boolean(summary && (summary.todo_total + summary.task_due_total + summary.goal_total + summary.reminder_total) > 0)
       const hasConv = conversationDateMap.value.get(dateKey) || false
-      cells.push({ key: dateKey, value: day, active: day === today, busy, selected: dateKey === selectedDate.value, hasConversation: hasConv })
+      cells.push({ key: dateKey, value: day, active: dateKey === todayKey, busy, selected: dateKey === selectedDate.value, hasConversation: hasConv })
     }
     while (cells.length % 7 !== 0) {
       cells.push({ key: `blank-end-${cells.length}`, value: null, active: false, busy: false, selected: false, hasConversation: false })
@@ -75,8 +77,8 @@ export function useSidebarPlan(clientTimeRef: { value: Date }, loadDailyDigestFn
     return cells
   })
 
-  const calendarYear = computed(() => clientTimeRef.value.getUTCFullYear())
-  const calendarMonth = computed(() => clientTimeRef.value.getUTCMonth() + 1)
+  const calendarYear = computed(() => clientTimeRef.value.getFullYear())
+  const calendarMonth = computed(() => clientTimeRef.value.getMonth() + 1)
 
   const todayPlanCounters = computed(() => {
     const detail = todayPlanDetail.value

frontend/src/pages/chat/index.vue

@@ -1,5 +1,5 @@
 <script setup lang="ts">
-import { ref } from 'vue'
+import { ref, toRef } from 'vue'
 import {
   ChevronRight,
   Send,
@@ -25,7 +25,7 @@ import { useChatView } from '@/pages/chat/composables/useChatView'
 import { useKnowledgeView } from '@/pages/knowledge/composables/useKnowledgeView'
 import { useDailyDigest } from '@/pages/chat/composables/useDailyDigest'
 import { useClientTime, formatNetworkRate } from '@/pages/chat/composables/useClientTime'
-import { useSidebarPlan } from '@/pages/chat/composables/useSidebarPlan'
+import { useSidebarPlan, formatDateKey } from '@/pages/chat/composables/useSidebarPlan'
 
 // --- Chat view (core messaging logic) ---
 const {
@@ -79,7 +79,7 @@ const {
   sidebarFocusItems, sidebarReviewAchievements, sidebarReviewReflections,
   sidebarFeedItems, topbarFeedItems, sidebarCollapsedModules,
   selectedDate, selectCalendarDate
-} = useSidebarPlan(clientTime, loadDailyDigest, store.conversations)
+} = useSidebarPlan(clientTime, loadDailyDigest, toRef(store, 'conversations'))
 
 // --- Local UI state ---
 const sidebarCollapsed = ref(false)
@@ -118,27 +118,16 @@ function handleOpenPreview(doc: any) { previewDoc.value = doc }
 function handleCalendarDateSelect(dateKey: string) {
   selectCalendarDate(dateKey)
 
-  // Reload conversations to get latest data
   loadConversations().then(() => {
-    // Find conversation that matches the selected date (by updated_at)
-    // Use UTC to avoid timezone issues
-    const [year, month, day] = dateKey.split('-').map(Number)
-    const targetDateStart = new Date(Date.UTC(year, month - 1, day))
-    const targetDateEnd = new Date(Date.UTC(year, month - 1, day + 1))
-
-    // Find conversation that falls on the selected date
-    const conversations = store.conversations.value ?? store.conversations
-    const conversation = conversations.find((conv) => {
-      const convDate = new Date(conv.updated_at)
-      return convDate >= targetDateStart && convDate < targetDateEnd
-    })
+    const conversations = store.conversations
+    const conversation = conversations.find((conv: { id: string; updated_at: string }) => formatDateKey(new Date(conv.updated_at)) === dateKey)
 
     if (conversation) {
       selectConversation(conversation.id)
-    } else {
-      // No conversation for this date, create a new one
-      newConversation()
+      return
     }
+
+    newConversation()
   }).catch((err) => {
     console.error('[Calendar] Error loading conversations:', err)
   })

frontend/vite.config.ts

@@ -1,10 +1,44 @@
 import { defineConfig, loadEnv } from 'vite'
 import vue from '@vitejs/plugin-vue'
 import path from 'path'
+import fs from 'fs'
 
 export default defineConfig(({ mode }) => {
   const env = loadEnv(mode, __dirname, '')
 
+  function parseDotenvFile(filePath: string) {
+    try {
+      if (!fs.existsSync(filePath)) return {}
+      const text = fs.readFileSync(filePath, 'utf-8')
+      const result: Record<string, string> = {}
+      for (const rawLine of text.split(/\r?\n/)) {
+        const line = rawLine.trim()
+        if (!line || line.startsWith('#')) continue
+        const eq = line.indexOf('=')
+        if (eq <= 0) continue
+        const key = line.slice(0, eq).trim()
+        let value = line.slice(eq + 1).trim()
+        if ((value.startsWith('"') && value.endsWith('"')) || (value.startsWith("'") && value.endsWith("'"))) {
+          value = value.slice(1, -1)
+        }
+        result[key] = value
+      }
+      return result
+    } catch {
+      return {}
+    }
+  }
+
+  // Vite only loads env files under `frontend/` by default.
+  // Many Jarvis setups keep HOST/PORT in repo root `.env`, so we read it explicitly as a fallback.
+  const rootEnvPath = path.resolve(__dirname, '../.env')
+  const rootEnv = parseDotenvFile(rootEnvPath)
+  const rootHost = (rootEnv.HOST || '127.0.0.1').trim()
+  const rootPort = (rootEnv.PORT || '').trim()
+  const rootApi = (rootEnv.VITE_API_URL || (rootPort ? `http://${rootHost}:${rootPort}` : '')).trim()
+
+  const apiTarget = env.VITE_API_URL || rootApi || 'http://localhost:8000'
+
   return {
   plugins: [vue()],
   resolve: {
@@ -15,7 +49,7 @@ export default defineConfig(({ mode }) => {
   server: {
     proxy: {
       '/api': {
-        target: env.VITE_API_URL,
+        target: apiTarget,
         changeOrigin: true,
       },
     },