fix(backend): update conversation and schedule center schemas

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/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/prompts.py

@@ -309,14 +309,14 @@ ANALYST_INSIGHTS_PROMPT = f"""{JARVIS_PERSONA_PROMPT}
 
 你是 analyst 体系下的洞察建议官，负责从任务、论坛和知识线索里提炼趋势、风险与建议。
 
-## 允许使用的工具:
+## 你的允许使用的工具:
 - get_tasks
 - get_forum_posts
 - search_knowledge
 - hybrid_search
 - web_search
 
-## 要求:
+## 你的要求:
 - 先给结论与判断
 - 再说明依据与建议
 - 当需要外部/最新信息时，可使用 `web_search`
@@ -324,6 +324,38 @@ ANALYST_INSIGHTS_PROMPT = f"""{JARVIS_PERSONA_PROMPT}
 """
 
 
+CODE_COMMANDER_SYSTEM_PROMPT = f"""{JARVIS_PERSONA_PROMPT}
+
+你是代码指挥官，负责协调 AI 写代码助手。
+
+## 你的职责:
+1. 接收用户选择的 AI 提供商（Claude/Gemini/Codex/OpenCode）
+2. 接收用户的写代码需求
+3. 进行安全分级判定
+4. 路由到合适的执行器
+
+## 安全分级规则:
+- 低风险：demo、示例、贪食蛇游戏等独立项目
+- 高风险：修改现有项目、涉及 Jarvis 项目、路径操作等
+
+## 执行模式:
+- 直接执行：低风险任务，直接运行
+- 沙盒执行：高风险任务，在临时目录隔离执行
+
+## 你的输出:
+- 简洁汇报执行结果
+- 如果需要用户交互（如确认 "y"），明确提示
+"""
+
+
+SANDBOX_EXECUTION_PROMPT = """将在隔离的临时目录中执行任务。
+任务完成后，工作目录会被保留供下载。"""
+
+
+DIRECT_EXECUTION_PROMPT = """将直接执行任务。
+如果需要交互，请等待用户输入。"""
+
+
 COORDINATOR_SYSTEM_PROMPT = f"""{JARVIS_PERSONA_PROMPT}
 
 你是 Jarvis 的协作协调官，负责把复杂请求收束成最小受控协作，而不是放任系统进入自由 swarm。
@@ -382,6 +414,7 @@ TOP_LEVEL_SYSTEM_PROMPTS_BY_KEY = {
     "executor": EXECUTOR_SYSTEM_PROMPT,
     "librarian": LIBRARIAN_SYSTEM_PROMPT,
     "analyst": ANALYST_SYSTEM_PROMPT,
+    "code_commander": CODE_COMMANDER_SYSTEM_PROMPT,
 }
 
 

backend/app/agents/registry/builtins.py

@@ -29,6 +29,7 @@ TOP_LEVEL_AGENT_DEFAULT_SUB_COMMANDERS: dict[str, tuple[str, ...]] = {
         "analyst_progress",
         "analyst_insights",
     ),
+    AgentRole.CODE_COMMANDER.value: (),
 }
 
 TOP_LEVEL_AGENT_DISPLAY_NAMES: dict[str, str] = {
@@ -37,6 +38,7 @@ TOP_LEVEL_AGENT_DISPLAY_NAMES: dict[str, str] = {
     AgentRole.EXECUTOR.value: "Executor",
     AgentRole.LIBRARIAN.value: "Librarian",
     AgentRole.ANALYST.value: "Analyst",
+    AgentRole.CODE_COMMANDER.value: "Code Commander",
 }
 
 TOP_LEVEL_AGENT_ROUTING_HINTS: dict[str, tuple[str, ...]] = {
@@ -55,6 +57,9 @@ TOP_LEVEL_AGENT_ROUTING_HINTS: dict[str, tuple[str, ...]] = {
     AgentRole.ANALYST.value: (
         "Handle reporting and insight requests using analyst sub-commanders.",
     ),
+    AgentRole.CODE_COMMANDER.value: (
+        "Handle code writing and execution tasks using AI CLI adapters.",
+    ),
 }
 
 TOP_LEVEL_AGENT_ALLOWED_SPAWN_ROLES: dict[str, tuple[str, ...]] = {
@@ -63,11 +68,13 @@ TOP_LEVEL_AGENT_ALLOWED_SPAWN_ROLES: dict[str, tuple[str, ...]] = {
         AgentRole.EXECUTOR.value,
         AgentRole.LIBRARIAN.value,
         AgentRole.ANALYST.value,
+        AgentRole.CODE_COMMANDER.value,
     ),
     AgentRole.SCHEDULE_PLANNER.value: (AgentRole.SCHEDULE_PLANNER.value,),
     AgentRole.EXECUTOR.value: (AgentRole.EXECUTOR.value,),
     AgentRole.LIBRARIAN.value: (AgentRole.LIBRARIAN.value,),
     AgentRole.ANALYST.value: (AgentRole.ANALYST.value,),
+    AgentRole.CODE_COMMANDER.value: (),
 }
 
 SUB_COMMANDER_PARENT_AGENT_IDS: dict[str, str] = {
@@ -99,11 +106,7 @@ BUILTIN_AGENT_MANIFESTS: tuple[AgentManifest, ...] = tuple(
 
 
 _capability_tool_names = tuple(
-    dict.fromkeys(
+    dict.fromkeys(tool.name for tools in SUB_COMMANDER_TOOLSETS.values() for tool in tools)
-        tool.name
-        for tools in SUB_COMMANDER_TOOLSETS.values()
-        for tool in tools
-    )
 )
 
 _CAPABILITY_METADATA_BY_TOOL_NAME: dict[str, dict[str, object]] = {
@@ -260,9 +263,7 @@ BUILTIN_SUB_COMMANDER_MANIFESTS: tuple[SubCommanderManifest, ...] = tuple(
         sub_commander_id=sub_commander_id,
         parent_agent_id=SUB_COMMANDER_PARENT_AGENT_IDS[sub_commander_id],
         prompt_text=SUB_COMMANDER_PROMPTS_BY_KEY[sub_commander_id],
-        capability_ids=list(
+        capability_ids=list(dict.fromkeys(tool.name for tool in tools)),
-            dict.fromkeys(tool.name for tool in tools)
-        ),
     )
     for sub_commander_id, tools in SUB_COMMANDER_TOOLSETS.items()
 )

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/schemas/task.py

@@ -1,7 +1,9 @@
 from __future__ import annotations
 
 from datetime import datetime, timezone
+from enum import Enum
 from typing import Any, Literal
+from uuid import uuid4
 
 from pydantic import BaseModel, Field
 
@@ -13,6 +15,18 @@ InterruptStatus = Literal["requested", "acknowledged", "resolved"]
 BudgetMode = Literal["direct", "collaboration"]
 
 
+class CodeProviderType(str, Enum):
+    CLAUDE = "claude"
+    GEMINI = "gemini"
+    CODEX = "codex"
+    OPENCODE = "opencode"
+
+
+class RiskLevelType(str, Enum):
+    LOW = "low"
+    HIGH = "high"
+
+
 class InterruptRecord(BaseModel):
     interrupt_id: str
     reason: str
@@ -83,3 +97,37 @@ class TaskResult(BaseModel):
     budget_snapshot: CollaborationBudget | dict[str, Any] | None = None
     next_action: str | None = None
     output_data: dict[str, Any] | None = None
+
+
+class CodeTaskType(str, Enum):
+    DEMO = "demo"
+    PROJECT = "project"
+    MODIFICATION = "modification"
+
+
+class CodeTask(BaseModel):
+    """代码任务请求模型"""
+
+    task_id: str = Field(default_factory=lambda: str(uuid4()))
+    task_type: CodeTaskType
+    ai_provider: CodeProviderType
+    sandbox_mode: bool = False
+    workspace_path: str | None = None
+    user_prompt: str
+    parent_task_id: str | None = None
+    thread_id: str | None = None
+    message_id: str | None = None
+    created_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+
+
+class CodeExecutionResultSchema(BaseModel):
+    """代码执行结果模型 (API 响应用)"""
+
+    success: bool
+    message: str
+    files_created: list[str] = Field(default_factory=list)
+    output: str = ""
+    error: str | None = None
+    exit_code: int = 0
+    execution_time: float | None = None
+    sandbox_session_id: str | None = None

backend/app/agents/skills/__init__.py

@@ -1,16 +1 @@
-"""Skills 注册表 - Phase 9"""
+"""Skill package."""
-
-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",
-]

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

@@ -4,7 +4,14 @@ from typing import Annotated, Any, Literal, TypedDict
 
 from app.agents.schemas.event import AgentEvent
 from app.agents.schemas.message import AgentMessage
-from app.agents.schemas.task import AgentTask, CollaborationBudget, InterruptRecord, RecoveryRecord, TaskResult, VerificationStatus
+from app.agents.schemas.task import (
+    AgentTask,
+    CollaborationBudget,
+    InterruptRecord,
+    RecoveryRecord,
+    TaskResult,
+    VerificationStatus,
+)
 from langchain_core.messages import AIMessage, BaseMessage, HumanMessage
 from langgraph.graph.message import add_messages
 
@@ -23,6 +30,7 @@ class AgentRole(str, Enum):
     EXECUTOR = "executor"
     LIBRARIAN = "librarian"
     ANALYST = "analyst"
+    CODE_COMMANDER = "code_commander"
 
 
 @dataclass
@@ -130,6 +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
@@ -141,6 +161,14 @@ class AgentState(TypedDict):
     user_llm_config: dict[str, Any] | None
     provider_capabilities: dict[str, Any] | None
 
+    # Code Commander state
+    code_task_type: Literal["demo", "project", "modification"] | None
+    code_ai_provider: Literal["claude", "gemini", "codex", "opencode"] | None
+    code_sandbox_mode: bool | None
+    code_workspace_path: str | None
+    code_execution_session_id: str | None
+    code_execution_result: dict[str, Any] | None
+
 
 def initial_state(user_id: str, conversation_id: str) -> AgentState:
     return AgentState(
@@ -238,6 +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/agents/tools/__init__.py

@@ -138,3 +138,12 @@ SUB_COMMANDER_TOOLSETS = {
     "analyst_progress": ANALYST_PROGRESS_TOOLS,
     "analyst_insights": ANALYST_INSIGHT_TOOLS,
 }
+
+# Code Commander toolset (tools implemented in later phases)
+CODE_COMMANDER_TOOLSET_NAMES = [
+    "execute_code_task",
+    "get_execution_status",
+    "send_interactive_input",
+    "download_workspace",
+    "cleanup_workspace",
+]

backend/app/agents/tools/ai_adapter.py

@@ -0,0 +1,196 @@
+"""
+AI CLI Adapter - 统一接口适配不同 AI CLI (Claude/Gemini/Codex/OpenCode)
+"""
+
+from abc import ABC, abstractmethod
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Literal
+
+
+@dataclass
+class CodeExecutionResult:
+    """代码执行结果"""
+
+    success: bool
+    message: str
+    files_created: list[str] = field(default_factory=list)
+    output: str = ""
+    error: str | None = None
+    exit_code: int = 0
+
+
+class AICLIAdapter(ABC):
+    """AI CLI 适配器抽象基类"""
+
+    @property
+    @abstractmethod
+    def cli_name(self) -> str:
+        """CLI 命令名称，如 'claude', 'gemini'"""
+        pass
+
+    @property
+    @abstractmethod
+    def requires_workspace(self) -> bool:
+        """是否需要工作目录"""
+        pass
+
+    @property
+    def provider(self) -> Literal["claude", "gemini", "codex", "opencode"]:
+        """AI 提供商标识"""
+        return self.cli_name
+
+    @abstractmethod
+    def build_command(self, prompt: str, workspace: Path | None) -> list[str]:
+        """构建 CLI 命令"""
+        pass
+
+    @abstractmethod
+    def parse_output(self, output: str) -> CodeExecutionResult:
+        """解析 CLI 输出"""
+        pass
+
+    @abstractmethod
+    def is_installed(self) -> bool:
+        """检查 CLI 是否已安装"""
+        pass
+
+
+class ClaudeAdapter(AICLIAdapter):
+    """Claude CLI 适配器"""
+
+    cli_name = "claude"
+    requires_workspace = True
+
+    def build_command(self, prompt: str, workspace: Path | None) -> list[str]:
+        cmd = ["claude", "-p", prompt]
+        if workspace:
+            cmd.extend(["--output-format", "stream-json"])
+        cmd.append("--dangerously-skip-permissions")
+        return cmd
+
+    def parse_output(self, output: str) -> CodeExecutionResult:
+        # Claude CLI 输出可能是纯文本或 JSON
+        # 简化处理：直接返回输出
+        if not output.strip():
+            return CodeExecutionResult(
+                success=False,
+                message="No output from Claude CLI",
+                output=output,
+            )
+        return CodeExecutionResult(
+            success=True,
+            message="Execution completed",
+            output=output,
+        )
+
+    def is_installed(self) -> bool:
+        import shutil
+
+        return shutil.which("claude") is not None
+
+
+class GeminiAdapter(AICLIAdapter):
+    """Gemini CLI 适配器"""
+
+    cli_name = "gemini"
+    requires_workspace = False
+
+    def build_command(self, prompt: str, workspace: Path | None) -> list[str]:
+        cmd = ["gemini", "-p", prompt]
+        return cmd
+
+    def parse_output(self, output: str) -> CodeExecutionResult:
+        if not output.strip():
+            return CodeExecutionResult(
+                success=False,
+                message="No output from Gemini CLI",
+                output=output,
+            )
+        return CodeExecutionResult(
+            success=True,
+            message="Execution completed",
+            output=output,
+        )
+
+    def is_installed(self) -> bool:
+        import shutil
+
+        return shutil.which("gemini") is not None
+
+
+class CodexAdapter(AICLIAdapter):
+    """Codex CLI 适配器"""
+
+    cli_name = "codex"
+    requires_workspace = True
+
+    def build_command(self, prompt: str, workspace: Path | None) -> list[str]:
+        cmd = ["codex", "-p", prompt]
+        return cmd
+
+    def parse_output(self, output: str) -> CodeExecutionResult:
+        if not output.strip():
+            return CodeExecutionResult(
+                success=False,
+                message="No output from Codex CLI",
+                output=output,
+            )
+        return CodeExecutionResult(
+            success=True,
+            message="Execution completed",
+            output=output,
+        )
+
+    def is_installed(self) -> bool:
+        import shutil
+
+        return shutil.which("codex") is not None
+
+
+class OpenCodeAdapter(AICLIAdapter):
+    """OpenCode CLI 适配器"""
+
+    cli_name = "opencode"
+    requires_workspace = True
+
+    def build_command(self, prompt: str, workspace: Path | None) -> list[str]:
+        cmd = ["opencode", "-p", prompt]
+        return cmd
+
+    def parse_output(self, output: str) -> CodeExecutionResult:
+        if not output.strip():
+            return CodeExecutionResult(
+                success=False,
+                message="No output from OpenCode CLI",
+                output=output,
+            )
+        return CodeExecutionResult(
+            success=True,
+            message="Execution completed",
+            output=output,
+        )
+
+    def is_installed(self) -> bool:
+        import shutil
+
+        return shutil.which("opencode") is not None
+
+
+# 提供商注册表
+ADAPTER_REGISTRY: dict[str, AICLIAdapter] = {
+    "claude": ClaudeAdapter(),
+    "gemini": GeminiAdapter(),
+    "codex": CodexAdapter(),
+    "opencode": OpenCodeAdapter(),
+}
+
+
+def get_adapter(provider: str) -> AICLIAdapter:
+    """获取指定提供商的适配器"""
+    adapter = ADAPTER_REGISTRY.get(provider.lower())
+    if adapter is None:
+        raise ValueError(
+            f"Unknown AI provider: {provider}. Available: {list(ADAPTER_REGISTRY.keys())}"
+        )
+    return adapter

backend/app/agents/tools/collaboration.py

@@ -0,0 +1,217 @@
+"""
+Agent Collaboration Protocol
+
+Inter-agent tool collaboration messaging system.
+"""
+
+import uuid
+from datetime import datetime
+from enum import Enum
+from typing import Any, Callable, Dict, Optional
+
+from pydantic import BaseModel, Field
+
+
+class MessageType(str, Enum):
+    """Collaboration message types"""
+
+    REQUEST = "request"  # Request collaboration
+    RESPONSE = "response"  # Response result
+    PROGRESS = "progress"  # Progress update
+    CANCEL = "cancel"  # Cancel request
+
+
+class CollaborationMessage(BaseModel):
+    """Collaboration message model"""
+
+    id: str = Field(default_factory=lambda: str(uuid.uuid4()))
+    type: MessageType
+    from_agent: str
+    to_agent: str
+    content: Dict[str, Any]
+    metadata: Dict[str, Any] = Field(default_factory=dict)
+    timestamp: datetime = Field(default_factory=datetime.utcnow)
+
+    def is_request(self) -> bool:
+        return self.type == MessageType.REQUEST
+
+    def is_response(self) -> bool:
+        return self.type == MessageType.RESPONSE
+
+
+class CollaborationProtocol:
+    """Agent collaboration protocol for inter-agent tool requests"""
+
+    def __init__(self):
+        self._pending_requests: Dict[str, CollaborationMessage] = {}
+        self._handlers: Dict[str, Callable] = {}
+        self._response_futures: Dict[str, asyncio.Future] = {}
+
+    def register_handler(self, tool_name: str, handler: Callable) -> None:
+        """Register a tool handler for collaboration
+
+        Args:
+            tool_name: Name of the tool
+            handler: Async callable to handle the tool execution
+        """
+        self._handlers[tool_name] = handler
+
+    async def request_collaboration(
+        self,
+        from_agent: str,
+        to_agent: str,
+        tool_name: str,
+        parameters: Dict[str, Any],
+        timeout_ms: int = 30000,
+    ) -> Dict[str, Any]:
+        """Request collaboration from another agent
+
+        Args:
+            from_agent: Source agent name
+            to_agent: Target agent name
+            tool_name: Tool to execute
+            parameters: Tool parameters
+            timeout_ms: Timeout in milliseconds
+
+        Returns:
+            Execution result dict with status and result/error
+        """
+        import asyncio
+
+        request_id = str(uuid.uuid4())
+
+        message = CollaborationMessage(
+            id=request_id,
+            type=MessageType.REQUEST,
+            from_agent=from_agent,
+            to_agent=to_agent,
+            content={
+                "tool": tool_name,
+                "parameters": parameters,
+            },
+            metadata={"timeout": timeout_ms},
+        )
+
+        self._pending_requests[request_id] = message
+
+        # Create future for response
+        future = asyncio.get_event_loop().create_future()
+        self._response_futures[request_id] = future
+
+        # Send the message
+        await self._send_message(message)
+
+        # Wait for response with timeout
+        try:
+            result = await asyncio.wait_for(future, timeout=timeout_ms / 1000)
+            return result
+        except asyncio.TimeoutError:
+            return {
+                "status": "error",
+                "error": "Collaboration request timed out",
+            }
+        finally:
+            self._pending_requests.pop(request_id, None)
+            self._response_futures.pop(request_id, None)
+
+    async def handle_request(self, message: CollaborationMessage) -> CollaborationMessage:
+        """Handle an incoming collaboration request
+
+        Args:
+            message: The collaboration message
+
+        Returns:
+            Response message with result or error
+        """
+        import uuid
+
+        tool_name = message.content.get("tool")
+        parameters = message.content.get("parameters", {})
+
+        handler = self._handlers.get(tool_name)
+        if not handler:
+            return CollaborationMessage(
+                id=str(uuid.uuid4()),
+                type=MessageType.RESPONSE,
+                from_agent=message.to_agent,
+                to_agent=message.from_agent,
+                content={
+                    "status": "error",
+                    "error": f"Unknown tool: {tool_name}",
+                },
+                metadata={},
+            )
+
+        try:
+            result = await handler(**parameters)
+            return CollaborationMessage(
+                id=str(uuid.uuid4()),
+                type=MessageType.RESPONSE,
+                from_agent=message.to_agent,
+                to_agent=message.from_agent,
+                content={"status": "success", "result": result},
+                metadata={},
+            )
+        except Exception as e:
+            return CollaborationMessage(
+                id=str(uuid.uuid4()),
+                type=MessageType.RESPONSE,
+                from_agent=message.to_agent,
+                to_agent=message.from_agent,
+                content={"status": "error", "error": str(e)},
+                metadata={},
+            )
+
+    async def handle_response(self, message: CollaborationMessage) -> None:
+        """Handle an incoming response message
+
+        Args:
+            message: The response message
+        """
+        request_id = None
+        for req_id, pending in self._pending_requests.items():
+            if pending.id == message.id:
+                request_id = req_id
+                break
+
+        if request_id and request_id in self._response_futures:
+            future = self._response_futures[request_id]
+            if not future.done():
+                future.set_result(message.content)
+
+    async def _send_message(self, message: CollaborationMessage) -> None:
+        """Send a collaboration message
+
+        This is a placeholder for actual transport implementation.
+        In production, this would use WebSocket, message queue, or shared storage.
+
+        Args:
+            message: The message to send
+        """
+        # TODO: Implement actual message transport
+        # Options: WebSocket, Redis pub/sub, shared database
+        pass
+
+    def get_pending_requests(self) -> list:
+        """Get list of pending requests"""
+        return [
+            {
+                "id": msg.id,
+                "from": msg.from_agent,
+                "to": msg.to_agent,
+                "tool": msg.content.get("tool"),
+            }
+            for msg in self._pending_requests.values()
+        ]
+
+
+# Global collaboration protocol instance
+_collaboration_protocol: Optional[CollaborationProtocol] = None
+
+
+def get_collaboration_protocol() -> CollaborationProtocol:
+    """Get the global collaboration protocol instance"""
+    global _collaboration_protocol
+    if _collaboration_protocol is None:
+        _collaboration_protocol = CollaborationProtocol()
+    return _collaboration_protocol

backend/app/agents/tools/direct_executor.py

@@ -0,0 +1,112 @@
+"""
+Direct Executor - 直接执行器
+用于低风险任务，直接执行不隔离
+"""
+
+import asyncio
+import os
+import shutil
+import tempfile
+from pathlib import Path
+from typing import AsyncGenerator
+
+from app.agents.tools.ai_adapter import AICLIAdapter
+
+
+class ExecutionResult:
+    """执行结果"""
+
+    def __init__(
+        self,
+        success: bool,
+        exit_code: int,
+        stdout: str,
+        stderr: str,
+    ):
+        self.success = success
+        self.exit_code = exit_code
+        self.stdout = stdout
+        self.stderr = stderr
+
+
+class DirectExecutor:
+    """直接执行器（用于低风险任务）"""
+
+    def __init__(self, adapter: AICLIAdapter, timeout: int = 60):
+        self.adapter = adapter
+        self.timeout = timeout
+
+    async def execute(
+        self,
+        prompt: str,
+    ) -> AsyncGenerator[str, None]:
+        """
+        直接执行，不需要沙盒
+
+        Args:
+            prompt: 任务描述
+
+        Yields:
+            str: 实时输出
+        """
+        # 1. 检查 CLI 是否安装
+        if not self.adapter.is_installed():
+            yield f"[ERROR] {self.adapter.cli_name} is not installed\n"
+            yield f"[ERROR] Please install {self.adapter.cli_name} first\n"
+            return
+
+        # 2. 构建命令
+        cmd = self.adapter.build_command(prompt, None)
+
+        # 3. 异步执行，实时 yield 输出
+        process = await asyncio.create_subprocess_exec(
+            *cmd,
+            stdout=asyncio.subprocess.PIPE,
+            stderr=asyncio.subprocess.PIPE,
+            env={**os.environ, "TERM": "xterm-256color"},
+        )
+
+        # 4. 实时读取输出
+        stdout_lines = []
+        stderr_lines = []
+
+        while True:
+            try:
+                line_bytes = await asyncio.wait_for(
+                    process.stdout.readline(),
+                    timeout=self.timeout,
+                )
+                if not line_bytes:
+                    break
+                line = line_bytes.decode("utf-8", errors="replace")
+                stdout_lines.append(line)
+                yield line
+            except asyncio.TimeoutError:
+                process.kill()
+                yield f"\n[ERROR] Execution timed out after {self.timeout}s\n"
+                break
+
+        # 5. 读取 stderr
+        stderr_bytes = await process.communicate()
+        if stderr_bytes[1]:
+            stderr = stderr_bytes[1].decode("utf-8", errors="replace")
+            stderr_lines.append(stderr)
+            yield f"\n[STDERR]\n{stderr}\n"
+
+        # 6. 完成标记
+        yield f"\n[EXIT_CODE] {process.returncode or 0}\n"
+        yield f"\n[COMPLETE] success={process.returncode == 0}\n"
+
+    async def execute_sync(self, prompt: str) -> ExecutionResult:
+        """同步执行并返回完整结果"""
+        output_parts = []
+        async for line in self.execute(prompt):
+            output_parts.append(line)
+
+        output = "".join(output_parts)
+        return ExecutionResult(
+            success="[COMPLETE] success=True" in output,
+            exit_code=0,
+            stdout=output,
+            stderr="",
+        )

backend/app/agents/tools/interactive_input.py

@@ -0,0 +1,58 @@
+"""
+InteractiveInputHandler - 交互输入处理
+"""
+
+import asyncio
+from typing import Any
+
+from app.agents.tools.terminal_engine import PTYManager
+
+
+class InteractiveInputHandler:
+    """交互输入处理器"""
+
+    def __init__(self, pty_manager: PTYManager):
+        self.pty_manager = pty_manager
+        self._pending_inputs: dict[str, asyncio.Event] = {}
+        self._input_cache: dict[str, str] = {}
+
+    async def wait_for_input(self, session_id: str, prompt: str) -> str:
+        """等待用户输入（如 "y" 确认）"""
+        event = asyncio.Event()
+        self._pending_inputs[session_id] = event
+
+        # 发送提示
+        from app.routers.terminal import manager
+
+        try:
+            await manager.send(session_id, f"\n{prompt}\n")
+        except Exception:
+            pass
+
+        # 等待输入完成
+        try:
+            await asyncio.wait_for(event.wait(), timeout=60.0)
+        except asyncio.TimeoutError:
+            del self._pending_inputs[session_id]
+            return self._input_cache.get(session_id, "")
+
+        del self._pending_inputs[session_id]
+
+        return self._input_cache.get(session_id, "")
+
+    async def send_input(self, session_id: str, data: str):
+        """用户发送输入"""
+        self._input_cache[session_id] = data
+        if session_id in self._pending_inputs:
+            self._pending_inputs[session_id].set()
+
+        # 同时写入 PTY
+        await self.pty_manager.write(session_id, data + "\n")
+
+    def clear_input(self, session_id: str):
+        """清除输入缓存"""
+        if session_id in self._input_cache:
+            del self._input_cache[session_id]
+        if session_id in self._pending_inputs:
+            self._pending_inputs[session_id].set()  # 取消等待
+            del self._pending_inputs[session_id]

backend/app/agents/tools/sandbox_executor.py

@@ -0,0 +1,173 @@
+"""
+Sandbox Executor - 沙盒执行器
+在高风险任务在隔离的临时目录中执行
+"""
+
+import asyncio
+import os
+import shutil
+import tempfile
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import AsyncGenerator
+
+from app.agents.tools.ai_adapter import AICLIAdapter, CodeExecutionResult
+
+
+@dataclass
+class SandboxEnvironment:
+    """沙盒环境"""
+
+    workspace_path: Path
+    session_id: str
+
+    @staticmethod
+    async def create(prefix: str = "jarvis_code_") -> "SandboxEnvironment":
+        """创建新的沙盒环境"""
+        temp_dir = tempfile.mkdtemp(prefix=prefix)
+        session_id = Path(temp_dir).name
+        return SandboxEnvironment(
+            workspace_path=Path(temp_dir),
+            session_id=session_id,
+        )
+
+    async def cleanup(self) -> None:
+        """清理沙盒环境"""
+        if self.workspace_path.exists():
+            shutil.rmtree(self.workspace_path)
+
+    def list_created_files(self) -> list[str]:
+        """列出沙盒中创建的文件"""
+        if not self.workspace_path.exists():
+            return []
+        files = []
+        for root, dirs, filenames in os.walk(self.workspace_path):
+            for filename in filenames:
+                full_path = os.path.join(root, filename)
+                rel_path = os.path.relpath(full_path, self.workspace_path)
+                files.append(rel_path)
+        return files
+
+
+@dataclass
+class ExecutionResult:
+    """执行结果"""
+
+    success: bool
+    exit_code: int
+    stdout: str
+    stderr: str
+    files_created: list[str] = field(default_factory=list)
+
+
+class SandboxExecutor:
+    """沙盒执行器"""
+
+    def __init__(self, adapter: AICLIAdapter, timeout: int = 300):
+        self.adapter = adapter
+        self.timeout = timeout
+        self._sessions: dict[str, SandboxEnvironment] = {}
+
+    async def execute(
+        self,
+        prompt: str,
+        session_id: str | None = None,
+    ) -> AsyncGenerator[str, None]:
+        """
+        执行代码任务，yield 实时输出
+
+        Args:
+            prompt: 任务描述
+            session_id: 会话 ID（可选，用于复用沙盒）
+
+        Yields:
+            str: 实时输出
+        """
+        # 1. 创建或复用沙盒环境
+        if session_id and session_id in self._sessions:
+            env = self._sessions[session_id]
+        else:
+            env = await SandboxEnvironment.create()
+            self._sessions[env.session_id] = env
+            session_id = env.session_id
+
+        # 2. 构建命令
+        cmd = self.adapter.build_command(prompt, env.workspace_path)
+
+        # 3. 异步执行，实时 yield 输出
+        process = await asyncio.create_subprocess_exec(
+            *cmd,
+            stdout=asyncio.subprocess.PIPE,
+            stderr=asyncio.subprocess.PIPE,
+            cwd=str(env.workspace_path),
+            env={**os.environ, "TERM": "xterm-256color"},
+        )
+
+        # 4. 实时读取输出
+        stdout_lines = []
+        stderr_lines = []
+
+        while True:
+            try:
+                line_bytes = await asyncio.wait_for(
+                    process.stdout.readline(),
+                    timeout=self.timeout,
+                )
+                if not line_bytes:
+                    break
+                line = line_bytes.decode("utf-8", errors="replace")
+                stdout_lines.append(line)
+                yield line
+            except asyncio.TimeoutError:
+                process.kill()
+                yield f"\n[ERROR] Execution timed out after {self.timeout}s\n"
+                break
+
+        # 5. 读取 stderr
+        stderr_bytes = await process.communicate()
+        if stderr_bytes[1]:
+            stderr = stderr_bytes[1].decode("utf-8", errors="replace")
+            stderr_lines.append(stderr)
+            yield f"\n[STDERR]\n{stderr}\n"
+
+        # 6. 返回结果（通过 yield 完成标记）
+        result = ExecutionResult(
+            success=process.returncode == 0,
+            exit_code=process.returncode or 0,
+            stdout="".join(stdout_lines),
+            stderr="".join(stderr_lines),
+            files_created=env.list_created_files(),
+        )
+        yield f"\n[EXIT_CODE] {result.exit_code}\n"
+        yield f"\n[COMPLETE] success={result.success}\n"
+
+    async def get_result(
+        self,
+        prompt: str,
+        session_id: str | None = None,
+    ) -> ExecutionResult:
+        """同步执行并返回完整结果"""
+        output_parts = []
+        async for line in self.execute(prompt, session_id):
+            output_parts.append(line)
+        # 解析最后的结果
+        # 实际使用中可能需要更复杂的结果收集
+        return ExecutionResult(
+            success="[COMPLETE] success=True" in "".join(output_parts),
+            exit_code=0,
+            stdout="".join(output_parts),
+            stderr="",
+            files_created=[],
+        )
+
+    async def cleanup_session(self, session_id: str) -> bool:
+        """清理指定会话"""
+        if session_id in self._sessions:
+            await self._sessions[session_id].cleanup()
+            del self._sessions[session_id]
+            return True
+        return False
+
+    def get_session(self, session_id: str) -> SandboxEnvironment | None:
+        """获取会话环境"""
+        return self._sessions.get(session_id)

backend/app/agents/tools/security_classifier.py

@@ -0,0 +1,129 @@
+"""
+Security Classifier - 安全分级判定
+低风险任务直接执行，高风险任务沙盒执行
+"""
+
+from enum import Enum
+
+
+class RiskLevel(Enum):
+    LOW = "low"  # 直接执行
+    HIGH = "high"  # 沙盒执行
+
+
+class SecurityClassifier:
+    """安全分级器"""
+
+    HIGH_RISK_KEYWORDS = [
+        # 路径/项目操作
+        "修改",
+        "编辑",
+        "删除",
+        "移动",
+        "重命名",
+        "Jarvis",
+        "backend",
+        "frontend",
+        "git",
+        "config",
+        ".env",
+        "生产环境",
+        # 文件操作
+        "写入",
+        "创建文件在",
+        "移动到",
+        "提交",
+        "push",
+        "pull",
+        "merge",
+        # 系统操作
+        "sudo",
+        "rm ",
+        "chmod",
+        "chown",
+    ]
+
+    LOW_RISK_KEYWORDS = [
+        # demo/示例类
+        "demo",
+        "示例",
+        "贪食蛇",
+        "俄罗斯方块",
+        "小游戏",
+        "独立项目",
+        "新项目",
+        "创建一个",
+        "写一个",
+        "帮我写一个",
+        # 明确无害的请求
+        "生成代码",
+        "代码示例",
+        "练习项目",
+    ]
+
+    def classify(self, task_description: str, target_path: str | None = None) -> RiskLevel:
+        """
+        判断任务风险等级
+
+        Args:
+            task_description: 任务描述
+            target_path: 目标路径（如果有）
+
+        Returns:
+            RiskLevel: LOW 或 HIGH
+        """
+        # 1. 检查高风险关键词
+        task_lower = task_description.lower()
+        if any(kw.lower() in task_lower for kw in self.HIGH_RISK_KEYWORDS):
+            return RiskLevel.HIGH
+
+        # 2. 检查目标路径
+        if target_path and self._is_project_path(target_path):
+            return RiskLevel.HIGH
+
+        # 3. 检查低风险关键词
+        if any(kw.lower() in task_lower for kw in self.LOW_RISK_KEYWORDS):
+            return RiskLevel.LOW
+
+        # 4. 默认高风险（保守策略）
+        return RiskLevel.HIGH
+
+    def _is_project_path(self, path: str) -> bool:
+        """
+        检查路径是否指向项目目录
+
+        Args:
+            path: 文件路径
+
+        Returns:
+            bool: 如果是项目路径返回 True
+        """
+        path_lower = path.lower()
+        project_indicators = [
+            "jarvis",
+            "backend/app",
+            "frontend/src",
+            ".git",
+            "package.json",
+            "pyproject.toml",
+            "requirements.txt",
+        ]
+        return any(indicator in path_lower for indicator in project_indicators)
+
+    def get_risk_factors(
+        self, task_description: str, target_path: str | None = None
+    ) -> dict[str, bool]:
+        """
+        获取详细的风险因素分析
+
+        Returns:
+            dict: 各风险因素及其状态
+        """
+        task_lower = task_description.lower()
+        return {
+            "has_high_risk_keywords": any(
+                kw.lower() in task_lower for kw in self.HIGH_RISK_KEYWORDS
+            ),
+            "has_low_risk_keywords": any(kw.lower() in task_lower for kw in self.LOW_RISK_KEYWORDS),
+            "is_project_path": bool(target_path and self._is_project_path(target_path)),
+        }

backend/app/agents/tools/stream_output.py

@@ -0,0 +1,86 @@
+"""
+StreamOutput - 流式输出封装
+"""
+
+import json
+from datetime import datetime, timezone
+from typing import Any, AsyncGenerator, Callable
+
+from dataclasses import dataclass
+
+
+@dataclass
+class StreamEvent:
+    """流式事件"""
+
+    type: str  # "output" | "error" | "status" | "complete"
+    session_id: str
+    data: str
+    timestamp: str
+
+
+class StreamOutput:
+    """流式输出处理器"""
+
+    def __init__(
+        self,
+        session_id: str,
+        websocket_sender: Callable[[str, str], Any] | None = None,
+    ):
+        self.session_id = session_id
+        self.websocket_sender = websocket_sender
+        self._listeners: list[Callable[[StreamEvent], Any]] = []
+
+    def add_listener(self, listener: Callable[[StreamEvent], Any]):
+        """添加事件监听器"""
+        self._listeners.append(listener)
+
+    def remove_listener(self, listener: Callable[[StreamEvent], Any]):
+        """移除事件监听器"""
+        if listener in self._listeners:
+            self._listeners.remove(listener)
+
+    async def push(self, event_type: str, data: str):
+        """推送事件到 WebSocket 和监听器"""
+        event = StreamEvent(
+            type=event_type,
+            session_id=self.session_id,
+            data=data,
+            timestamp=datetime.now(timezone.utc).isoformat(),
+        )
+
+        # 发送到 WebSocket
+        if self.websocket_sender:
+            try:
+                await self.websocket_sender(self.session_id, json.dumps(event.__dict__))
+            except Exception:
+                pass
+
+        # 通知监听器
+        for listener in self._listeners:
+            try:
+                result = listener(event)
+                if hasattr(result, "__await__"):
+                    await result
+            except Exception:
+                pass
+
+    async def stream_execution(
+        self,
+        executor,
+        prompt: str,
+    ) -> AsyncGenerator[str, None]:
+        """包装执行器，实现流式输出"""
+        async for line in executor.execute(prompt):
+            await self.push("output", line)
+            yield line
+
+        await self.push("complete", "")
+
+    async def push_status(self, status: str):
+        """推送状态消息"""
+        await self.push("status", status)
+
+    async def push_error(self, error: str):
+        """推送错误消息"""
+        await self.push("error", error)

backend/app/agents/tools/terminal_engine.py

@@ -0,0 +1,160 @@
+"""
+PTY Terminal Engine - 跨平台 PTY 终端管理
+"""
+
+import asyncio
+import os
+from dataclasses import dataclass, field
+from typing import AsyncGenerator
+
+from uuid import uuid4
+
+
+@dataclass
+class PTYSession:
+    """PTY 会话"""
+
+    session_id: str
+    process: asyncio.subprocess.Process
+    workspace_path: str
+
+
+class PTYManager:
+    """PTY 会话管理器"""
+
+    def __init__(self):
+        self._sessions: dict[str, PTYSession] = {}
+        self._output_queues: dict[str, asyncio.Queue] = {}
+
+    async def spawn(
+        self,
+        cli: str,
+        args: list[str],
+        cwd: str,
+        session_id: str | None = None,
+        env: dict | None = None,
+    ) -> str:
+        """启动 PTY 会话"""
+        if session_id is None:
+            session_id = f"pty_{uuid4().hex[:8]}"
+
+        # 构建环境变量
+        process_env = {**os.environ, "TERM": "xterm-256color"}
+        if env:
+            process_env.update(env)
+
+        # 创建 PTY 进程
+        process = await asyncio.create_subprocess_exec(
+            cli,
+            *args,
+            stdout=asyncio.subprocess.PIPE,
+            stderr=asyncio.subprocess.PIPE,
+            cwd=cwd,
+            env=process_env,
+        )
+
+        session = PTYSession(
+            session_id=session_id,
+            process=process,
+            workspace_path=cwd,
+        )
+        self._sessions[session_id] = session
+        self._output_queues[session_id] = asyncio.Queue()
+
+        # 启动输出读取协程
+        asyncio.create_task(self._read_output(session_id))
+
+        return session_id
+
+    async def _read_output(self, session_id: str):
+        """读取 PTY 输出并放入队列"""
+        session = self._sessions.get(session_id)
+        if not session:
+            return
+
+        queue = self._output_queues[session_id]
+
+        try:
+            while True:
+                line = await session.process.stdout.readline()
+                if not line:
+                    break
+                decoded_line = line.decode(errors="replace")
+                await queue.put(decoded_line)
+
+                # 广播到 WebSocket
+                await self._broadcast(session_id, decoded_line)
+
+            # 读取 stderr
+            stderr_line = await session.process.stderr.readline()
+            if stderr_line:
+                decoded_err = stderr_line.decode(errors="replace")
+                await queue.put(decoded_err)
+                await self._broadcast(session_id, decoded_err)
+
+        except Exception:
+            pass
+        finally:
+            await queue.put(None)  # 结束标记
+
+    async def write(self, session_id: str, data: str):
+        """写入 PTY（用户输入）"""
+        session = self._sessions.get(session_id)
+        if session and session.process.stdin:
+            session.process.stdin.write(data)
+            await session.process.stdin.drain()
+
+    async def read(self, session_id: str) -> AsyncGenerator[str, None]:
+        """读取 PTY 输出"""
+        queue = self._output_queues.get(session_id)
+        if not queue:
+            return
+
+        while True:
+            line = await queue.get()
+            if line is None:
+                break
+            yield line
+
+    async def resize(self, session_id: str, rows: int, cols: int):
+        """调整终端大小"""
+        # TODO: 实现 resize (需要平台特定实现)
+        pass
+
+    async def kill(self, session_id: str):
+        """终止 PTY 会话"""
+        if session_id in self._sessions:
+            session = self._sessions[session_id]
+            try:
+                session.process.terminate()
+                await asyncio.wait_for(session.process.wait(), timeout=3.0)
+            except asyncio.TimeoutError:
+                session.process.kill()
+                await session.process.wait()
+            except Exception:
+                pass
+            finally:
+                del self._sessions[session_id]
+                if session_id in self._output_queues:
+                    del self._output_queues[session_id]
+
+    async def _broadcast(self, session_id: str, data: str):
+        """广播输出到 WebSocket"""
+        from app.routers.terminal import manager
+
+        try:
+            await manager.send(session_id, data)
+        except Exception:
+            pass
+
+    def get_session(self, session_id: str) -> PTYSession | None:
+        """获取会话"""
+        return self._sessions.get(session_id)
+
+    def list_sessions(self) -> list[str]:
+        """列出所有会话 ID"""
+        return list(self._sessions.keys())
+
+
+# 全局单例
+pty_manager = PTYManager()

backend/app/config.py

@@ -61,6 +61,9 @@ class Settings(BaseSettings):
     DAILY_PLAN_TIME: str = "00:00"
     FORUM_SCAN_INTERVAL_MINUTES: int = 30
 
+    # === 位置配置 ===
+    LOCATION: str = "Location"
+
     # === CORS ===
     CORS_ORIGINS: list[str] = ["http://localhost:5173", "http://localhost:3000"]
 
@@ -101,6 +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

@@ -35,14 +35,206 @@ async def get_db() -> AsyncGenerator[AsyncSession, None]:
 async def init_db():
     async with engine.begin() as conn:
         await conn.run_sync(Base.metadata.create_all)
+        await ensure_task_columns(conn)
         await ensure_log_columns(conn)
         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_task_columns(conn):
+    rows = await _get_table_info(conn, 'tasks')
+    if not rows:
+        return
+
+    columns = {row[1] for row in rows}
+    required_columns = {
+        'source': "ALTER TABLE tasks ADD COLUMN source VARCHAR(32) DEFAULT 'manual' NOT NULL",
+        'conversation_id': "ALTER TABLE tasks ADD COLUMN conversation_id VARCHAR(36)",
+        'quadrant': "ALTER TABLE tasks ADD COLUMN quadrant VARCHAR(64)",
+        'assignee_type': "ALTER TABLE tasks ADD COLUMN assignee_type VARCHAR(32)",
+        'assignee_id': "ALTER TABLE tasks ADD COLUMN assignee_id VARCHAR(255)",
+        'dispatch_status': "ALTER TABLE tasks ADD COLUMN dispatch_status VARCHAR(32) DEFAULT 'idle' NOT NULL",
+        'dispatch_run_id': "ALTER TABLE tasks ADD COLUMN dispatch_run_id VARCHAR(64)",
+        'result_summary': "ALTER TABLE tasks ADD COLUMN result_summary TEXT",
+        'started_at': "ALTER TABLE tasks ADD COLUMN started_at DATETIME",
+        'last_synced_at': "ALTER TABLE tasks ADD COLUMN last_synced_at DATETIME",
+    }
+    for column, ddl in required_columns.items():
+        if column not in columns:
+            await conn.execute(text(ddl))
+
+    indexes = {
+        'ix_tasks_due_date': "CREATE INDEX IF NOT EXISTS ix_tasks_due_date ON tasks (due_date)",
+        'ix_tasks_source': "CREATE INDEX IF NOT EXISTS ix_tasks_source ON tasks (source)",
+        'ix_tasks_conversation_id': "CREATE INDEX IF NOT EXISTS ix_tasks_conversation_id ON tasks (conversation_id)",
+        'ix_tasks_quadrant': "CREATE INDEX IF NOT EXISTS ix_tasks_quadrant ON tasks (quadrant)",
+        'ix_tasks_assignee_type': "CREATE INDEX IF NOT EXISTS ix_tasks_assignee_type ON tasks (assignee_type)",
+        'ix_tasks_assignee_id': "CREATE INDEX IF NOT EXISTS ix_tasks_assignee_id ON tasks (assignee_id)",
+        'ix_tasks_dispatch_status': "CREATE INDEX IF NOT EXISTS ix_tasks_dispatch_status ON tasks (dispatch_status)",
+        'ix_tasks_dispatch_run_id': "CREATE INDEX IF NOT EXISTS ix_tasks_dispatch_run_id ON tasks (dispatch_run_id)",
+    }
+    for ddl in indexes.values():
+        await conn.execute(text(ddl))
+
+    history_rows = await _get_table_info(conn, 'task_histories')
+    if history_rows:
+        history_columns = {row[1] for row in history_rows}
+        if 'subtask_id' not in history_columns:
+            await conn.execute(text("ALTER TABLE task_histories ADD COLUMN subtask_id VARCHAR(36)"))
+        await conn.execute(text("CREATE INDEX IF NOT EXISTS ix_task_histories_subtask_id ON task_histories (subtask_id)"))
+
+    await conn.execute(
+        text(
+            """
+            CREATE TABLE IF NOT EXISTS task_subtasks (
+                id VARCHAR(36) PRIMARY KEY,
+                created_at DATETIME NOT NULL,
+                updated_at DATETIME NOT NULL,
+                task_id VARCHAR(36) NOT NULL,
+                title VARCHAR(500) NOT NULL,
+                description TEXT,
+                status VARCHAR(32) NOT NULL DEFAULT 'todo',
+                order_index INTEGER NOT NULL DEFAULT 0,
+                assignee_type VARCHAR(32),
+                assignee_id VARCHAR(255),
+                dispatch_status VARCHAR(32) NOT NULL DEFAULT 'idle',
+                dispatch_run_id VARCHAR(64),
+                completed_at DATETIME,
+                FOREIGN KEY(task_id) REFERENCES tasks (id)
+            )
+            """
+        )
+    )
+    subtask_rows = await _get_table_info(conn, 'task_subtasks')
+    subtask_columns = {row[1] for row in subtask_rows}
+    if 'result_summary' not in subtask_columns:
+        await conn.execute(text("ALTER TABLE task_subtasks ADD COLUMN result_summary TEXT"))
+    subtask_indexes = {
+        'ix_task_subtasks_task_id': "CREATE INDEX IF NOT EXISTS ix_task_subtasks_task_id ON task_subtasks (task_id)",
+        'ix_task_subtasks_status': "CREATE INDEX IF NOT EXISTS ix_task_subtasks_status ON task_subtasks (status)",
+        'ix_task_subtasks_order_index': "CREATE INDEX IF NOT EXISTS ix_task_subtasks_order_index ON task_subtasks (order_index)",
+        'ix_task_subtasks_assignee_type': "CREATE INDEX IF NOT EXISTS ix_task_subtasks_assignee_type ON task_subtasks (assignee_type)",
+        'ix_task_subtasks_assignee_id': "CREATE INDEX IF NOT EXISTS ix_task_subtasks_assignee_id ON task_subtasks (assignee_id)",
+        'ix_task_subtasks_dispatch_status': "CREATE INDEX IF NOT EXISTS ix_task_subtasks_dispatch_status ON task_subtasks (dispatch_status)",
+        'ix_task_subtasks_dispatch_run_id': "CREATE INDEX IF NOT EXISTS ix_task_subtasks_dispatch_run_id ON task_subtasks (dispatch_run_id)",
+    }
+    for ddl in subtask_indexes.values():
+        await conn.execute(text(ddl))
+
+    # Normalize legacy/invalid enum-like values to prevent ORM Enum decoding failures.
+    await conn.execute(
+        text(
+            """
+            UPDATE tasks
+            SET source = 'manual'
+            WHERE source IS NULL
+               OR TRIM(source) = ''
+               OR source NOT IN ('manual','chat','schedule_center','today_status','commander')
+            """
+        )
+    )
+    await conn.execute(
+        text(
+            """
+            UPDATE tasks
+            SET status = 'todo'
+            WHERE status IS NULL
+               OR TRIM(status) = ''
+               OR status NOT IN ('todo','in_progress','done','cancelled')
+            """
+        )
+    )
+    await conn.execute(
+        text(
+            """
+            UPDATE tasks
+            SET priority = 'medium'
+            WHERE priority IS NULL
+               OR TRIM(priority) = ''
+               OR priority NOT IN ('low','medium','high','urgent')
+            """
+        )
+    )
+    await conn.execute(
+        text(
+            """
+            UPDATE tasks
+            SET quadrant = NULL
+            WHERE quadrant IS NOT NULL
+              AND (TRIM(quadrant) = '' OR quadrant NOT IN (
+                'urgent-important',
+                'not-urgent-important',
+                'urgent-not-important',
+                'not-urgent-not-important'
+              ))
+            """
+        )
+    )
+    await conn.execute(
+        text(
+            """
+            UPDATE tasks
+            SET assignee_type = NULL
+            WHERE assignee_type IS NOT NULL
+              AND (TRIM(assignee_type) = '' OR assignee_type NOT IN (
+                'user','commander','agent','planner','executor','knowledge','analyst','coder','researcher'
+              ))
+            """
+        )
+    )
+    await conn.execute(
+        text(
+            """
+            UPDATE tasks
+            SET dispatch_status = 'idle'
+            WHERE dispatch_status IS NULL
+               OR TRIM(dispatch_status) = ''
+               OR dispatch_status NOT IN ('idle','queued','running','completed','failed')
+            """
+        )
+    )
+
+    await conn.execute(
+        text(
+            """
+            UPDATE task_subtasks
+            SET status = 'todo'
+            WHERE status IS NULL
+               OR TRIM(status) = ''
+               OR status NOT IN ('todo','in_progress','done','cancelled')
+            """
+        )
+    )
+    await conn.execute(
+        text(
+            """
+            UPDATE task_subtasks
+            SET assignee_type = NULL
+            WHERE assignee_type IS NOT NULL
+              AND (TRIM(assignee_type) = '' OR assignee_type NOT IN (
+                'user','commander','agent','planner','executor','knowledge','analyst','coder','researcher'
+              ))
+            """
+        )
+    )
+    await conn.execute(
+        text(
+            """
+            UPDATE task_subtasks
+            SET dispatch_status = 'idle'
+            WHERE dispatch_status IS NULL
+               OR TRIM(dispatch_status) = ''
+               OR dispatch_status NOT IN ('idle','queued','running','completed','failed')
+            """
+        )
+    )
 
 
 async def ensure_log_columns(conn):
@@ -115,6 +307,28 @@ 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 +395,14 @@ 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 +427,48 @@ 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/main.py

@@ -28,6 +28,10 @@ from app.routers import (
     marketplace_router,
     agent_skills_router,
     agent_sessions_router,
+    terminal_router,
+    tools_router,
+    remote_mount_router,
+    office_router,
 )
 from app.routers.scheduler import router as scheduler_router
 from app.services.scheduler_service import start_scheduler, stop_scheduler, get_scheduler_status
@@ -127,6 +131,10 @@ app.include_router(plugins_router)
 app.include_router(marketplace_router)
 app.include_router(agent_skills_router)
 app.include_router(agent_sessions_router)
+app.include_router(terminal_router)
+app.include_router(tools_router)
+app.include_router(remote_mount_router)
+app.include_router(office_router)
 
 
 @app.get("/api/health")

backend/app/models/__init__.py

@@ -2,11 +2,22 @@ from app.models.base import Base
 from app.models.user import User
 from app.models.folder import Folder
 from app.models.document import Document, DocumentChunk
-from app.models.task import Task, TaskHistory
+from app.models.task import (
+    Task,
+    TaskAssigneeType,
+    TaskDispatchStatus,
+    TaskHistory,
+    TaskPriority,
+    TaskQuadrant,
+    TaskSource,
+    TaskStatus,
+    TaskSubTask,
+)
 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,
@@ -20,7 +31,9 @@ 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
+from app.models.remote_mount import RemoteMount, RemoteSyncItem
 
 __all__ = [
     "Base",
@@ -29,7 +42,14 @@ __all__ = [
     "Document",
     "DocumentChunk",
     "Task",
+    "TaskSubTask",
     "TaskHistory",
+    "TaskStatus",
+    "TaskPriority",
+    "TaskSource",
+    "TaskQuadrant",
+    "TaskAssigneeType",
+    "TaskDispatchStatus",
     "ForumPost",
     "ForumReply",
     "Agent",
@@ -38,6 +58,8 @@ __all__ = [
     "Message",
     "KGNode",
     "KGEdge",
+    "LearningArtifactRecord",
+    "SessionRetrospectiveRecord",
     "MemorySummary",
     "UserMemory",
     "BrainEvent",
@@ -53,7 +75,10 @@ __all__ = [
     "ReminderStatus",
     "Goal",
     "GoalStatus",
+    "Skill",
     "Log",
     "LogType",
     "LogLevel",
+    "RemoteMount",
+    "RemoteSyncItem",
 ]

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/memory.py

@@ -1,4 +1,15 @@
-from sqlalchemy import Column, String, Text, Integer, ForeignKey, Boolean, DateTime, Enum as SQLEnum
+from sqlalchemy import (
+    Column,
+    String,
+    Text,
+    Integer,
+    Float,
+    ForeignKey,
+    Boolean,
+    DateTime,
+    Enum as SQLEnum,
+    JSON,
+)
 from app.models.base import BaseModel, utc_now
 
 
@@ -7,6 +18,7 @@ class MemorySummary(BaseModel):
     对话摘要 — 中期记忆
     当一段对话超过阈值轮数时，自动生成摘要存入此表
     """
+
     __tablename__ = "memory_summaries"
 
     user_id = Column(String(36), ForeignKey("users.id"), nullable=False, index=True)
@@ -21,14 +33,28 @@ class UserMemory(BaseModel):
     用户画像记忆 — 长期记忆
     从对话中提取的用户事实、偏好、目标
     """
+
     __tablename__ = "user_memories"
 
     user_id = Column(String(36), ForeignKey("users.id"), nullable=False, index=True)
     memory_type = Column(String(50), nullable=False)  # fact | preference | goal | habit | other
     content = Column(Text, nullable=False)  # 记忆内容
-    importance = Column(Integer, default=5)                # 重要程度 1-10
+    importance = Column(Integer, default=5)  # 重要程度 1-10 (legacy, replaced by importance_score)
     is_recalled = Column(Boolean, default=False)  # 是否在当前对话中被召回
     recall_count = Column(Integer, default=0)  # 被召回次数
     source_conversation_id = Column(String(36), nullable=True)  # 来源对话
     extracted_at = Column(DateTime, default=utc_now, nullable=False)
     last_recalled_at = Column(DateTime, nullable=True)
+    # M.1: 重要性评分系统
+    frequency_count = Column(
+        Integer, default=0
+    )  # 被召回次数 (duplicate of recall_count, for scoring clarity)
+    emotion_tags = Column(JSON, nullable=True)  # List of emotion keywords
+    importance_score = Column(Float, default=0.5)  # 重要性分数 0.0-1.0
+    importance_level = Column(String(20), default="medium")  # high | medium | low
+    associated_topics = Column(JSON, nullable=True)  # List of topic strings
+    # M.2: 遗忘曲线系统
+    decay_score = Column(Float, default=1.0)  # 0.0-1.0, higher=more remembered
+    is_archived = Column(Boolean, default=False)  # 是否已归档到冷存储
+    last_accessed_at = Column(DateTime, nullable=True)  # 上次访问时间（用于遗忘计算）
+    archive_at = Column(DateTime, nullable=True)  # 归档时间

backend/app/models/remote_mount.py

@@ -0,0 +1,34 @@
+from sqlalchemy import Boolean, Column, ForeignKey, String, Text, UniqueConstraint
+
+from app.models.base import BaseModel
+
+
+class RemoteMount(BaseModel):
+    __tablename__ = "remote_mounts"
+    __table_args__ = (
+        UniqueConstraint("user_id", "name", name="uq_remote_mount_user_name"),
+    )
+
+    user_id = Column(String(36), ForeignKey("users.id"), nullable=False, index=True)
+    name = Column(String(255), nullable=False)
+    mount_type = Column(String(32), nullable=False, default="webdav")
+    base_url = Column(String(1000), nullable=False)
+    username = Column(String(255), nullable=True)
+    password_encrypted = Column(Text, nullable=True)
+    root_path = Column(String(1000), nullable=False, default="/")
+    is_active = Column(Boolean, nullable=False, default=True)
+    last_sync_at = Column(String(64), nullable=True)
+
+
+class RemoteSyncItem(BaseModel):
+    __tablename__ = "remote_sync_items"
+
+    mount_id = Column(String(36), ForeignKey("remote_mounts.id"), nullable=False, index=True)
+    remote_path = Column(String(2000), nullable=False)
+    remote_etag = Column(String(512), nullable=True)
+    remote_modified_at = Column(String(128), nullable=True)
+    local_folder_id = Column(String(36), ForeignKey("folders.id"), nullable=True)
+    local_document_id = Column(String(36), ForeignKey("documents.id"), nullable=True)
+    sync_status = Column(String(32), nullable=False, default="synced")
+    last_error = Column(Text, nullable=True)
+    last_synced_at = Column(String(64), nullable=True)

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/models/task.py

@@ -1,7 +1,8 @@
-from sqlalchemy import Column, String, Text, Integer, ForeignKey, DateTime, Enum
-from sqlalchemy.orm import relationship
-from datetime import datetime
 from enum import Enum as PyEnum
+
+from sqlalchemy import Column, DateTime, Enum, ForeignKey, Integer, String, Text
+from sqlalchemy.orm import relationship
+
 from app.models.base import BaseModel
 
 
@@ -19,26 +20,144 @@ class TaskPriority(str, PyEnum):
     URGENT = "urgent"
 
 
+class TaskSource(str, PyEnum):
+    MANUAL = "manual"
+    CHAT = "chat"
+    SCHEDULE_CENTER = "schedule_center"
+    TODAY_STATUS = "today_status"
+    COMMANDER = "commander"
+
+
+class TaskQuadrant(str, PyEnum):
+    URGENT_IMPORTANT = "urgent-important"
+    NOT_URGENT_IMPORTANT = "not-urgent-important"
+    URGENT_NOT_IMPORTANT = "urgent-not-important"
+    NOT_URGENT_NOT_IMPORTANT = "not-urgent-not-important"
+
+
+class TaskAssigneeType(str, PyEnum):
+    USER = "user"
+    COMMANDER = "commander"
+    AGENT = "agent"
+    PLANNER = "planner"
+    EXECUTOR = "executor"
+    KNOWLEDGE = "knowledge"
+    ANALYST = "analyst"
+    CODER = "coder"
+    RESEARCHER = "researcher"
+
+
+class TaskDispatchStatus(str, PyEnum):
+    IDLE = "idle"
+    QUEUED = "queued"
+    RUNNING = "running"
+    COMPLETED = "completed"
+    FAILED = "failed"
+
+
+DispatchStatus = TaskDispatchStatus
+
+
+DispatchStatus = TaskDispatchStatus
+
+
+class TaskHistoryAction(str, PyEnum):
+    CREATED = "created"
+    CREATED_FROM_CHAT = "created_from_chat"
+    UPDATED = "updated"
+    STATUS_CHANGED = "status_changed"
+    ASSIGNED = "assigned"
+    DELETED = "deleted"
+    SUBTASK_CREATED = "subtask_created"
+    SUBTASK_UPDATED = "subtask_updated"
+    SUBTASK_DELETED = "subtask_deleted"
+    SUBTASK_REORDERED = "subtask_reordered"
+    DISPATCHED_TO_COMMANDER = "dispatched_to_commander"
+    DISPATCH_STATUS_CHANGED = "dispatch_status_changed"
+
+
+def enum_values(enum_cls: type[PyEnum]) -> list[str]:
+    return [item.value for item in enum_cls]
+
+
+TASK_STATUS_ENUM = Enum(TaskStatus, values_callable=enum_values)
+TASK_PRIORITY_ENUM = Enum(TaskPriority, values_callable=enum_values)
+TASK_SOURCE_ENUM = Enum(TaskSource, values_callable=enum_values)
+TASK_QUADRANT_ENUM = Enum(TaskQuadrant, values_callable=enum_values)
+TASK_ASSIGNEE_TYPE_ENUM = Enum(TaskAssigneeType, values_callable=enum_values)
+TASK_DISPATCH_STATUS_ENUM = Enum(TaskDispatchStatus, values_callable=enum_values)
+
+
 class Task(BaseModel):
     __tablename__ = "tasks"
 
     user_id = Column(String(36), ForeignKey("users.id"), nullable=False, index=True)
     title = Column(String(500), nullable=False)
     description = Column(Text, nullable=True)
-    status = Column(Enum(TaskStatus), default=TaskStatus.TODO, nullable=False, index=True)
+    status = Column(TASK_STATUS_ENUM, default=TaskStatus.TODO, nullable=False, index=True)
-    priority = Column(Enum(TaskPriority), default=TaskPriority.MEDIUM, nullable=False)
+    priority = Column(TASK_PRIORITY_ENUM, default=TaskPriority.MEDIUM, nullable=False)
-    due_date = Column(DateTime, nullable=True)
+    due_date = Column(DateTime, nullable=True, index=True)
     completed_at = Column(DateTime, nullable=True)
-    tags = Column(String(1000), nullable=True)  # JSON 数组
+    tags = Column(String(1000), nullable=True)  # JSON array
+    source = Column(TASK_SOURCE_ENUM, default=TaskSource.MANUAL, nullable=False, index=True)
+    conversation_id = Column(String(36), nullable=True, index=True)
+    quadrant = Column(TASK_QUADRANT_ENUM, nullable=True, index=True)
+    assignee_type = Column(TASK_ASSIGNEE_TYPE_ENUM, nullable=True, index=True)
+    assignee_id = Column(String(255), nullable=True, index=True)
+    dispatch_status = Column(
+        TASK_DISPATCH_STATUS_ENUM,
+        default=TaskDispatchStatus.IDLE,
+        nullable=False,
+        index=True,
+    )
+    dispatch_run_id = Column(String(64), nullable=True, index=True)
+    result_summary = Column(Text, nullable=True)
+    started_at = Column(DateTime, nullable=True)
+    last_synced_at = Column(DateTime, nullable=True)
 
-    history = relationship("TaskHistory", back_populates="task", cascade="all, delete-orphan")
+    subtasks = relationship(
+        "TaskSubTask",
+        back_populates="task",
+        cascade="all, delete-orphan",
+        order_by="TaskSubTask.order_index.asc()",
+    )
+    history = relationship(
+        "TaskHistory",
+        back_populates="task",
+        cascade="all, delete-orphan",
+        order_by="TaskHistory.created_at.desc()",
+    )
+
+
+class TaskSubTask(BaseModel):
+    __tablename__ = "task_subtasks"
+
+    task_id = Column(String(36), ForeignKey("tasks.id"), nullable=False, index=True)
+    title = Column(String(500), nullable=False)
+    description = Column(Text, nullable=True)
+    status = Column(TASK_STATUS_ENUM, default=TaskStatus.TODO, nullable=False, index=True)
+    order_index = Column(Integer, default=0, nullable=False, index=True)
+    assignee_type = Column(TASK_ASSIGNEE_TYPE_ENUM, nullable=True, index=True)
+    assignee_id = Column(String(255), nullable=True, index=True)
+    dispatch_status = Column(
+        TASK_DISPATCH_STATUS_ENUM,
+        default=TaskDispatchStatus.IDLE,
+        nullable=False,
+        index=True,
+    )
+    dispatch_run_id = Column(String(64), nullable=True, index=True)
+    result_summary = Column(Text, nullable=True)
+    completed_at = Column(DateTime, nullable=True)
+
+    task = relationship("Task", back_populates="subtasks")
 
 
 class TaskHistory(BaseModel):
     __tablename__ = "task_histories"
 
     task_id = Column(String(36), ForeignKey("tasks.id"), nullable=False, index=True)
-    action = Column(String(100), nullable=False)  # created, status_changed, updated, deleted
+    subtask_id = Column(String(36), ForeignKey("task_subtasks.id"), nullable=True, index=True)
+    action = Column(String(100), nullable=False)
     old_value = Column(Text, nullable=True)
     new_value = Column(Text, nullable=True)
 

backend/app/routers/__init__.py

@@ -20,3 +20,7 @@ from app.routers.plugins import router as plugins_router
 from app.routers.plugins import _marketplace_router as marketplace_router
 from app.routers.agent_skills import router as agent_skills_router
 from app.routers.agent_sessions import router as agent_sessions_router
+from app.routers.terminal import router as terminal_router
+from app.routers.tools import router as tools_router
+from app.routers.remote_mount import router as remote_mount_router
+from app.routers.office import router as office_router

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

@@ -100,6 +100,7 @@ async def chat(
             conversation_id=data.conversation_id,
             file_ids=data.file_ids,
             model_name=data.model_name,
+            runtime=data.runtime,
         )
     except ValueError as exc:
         raise HTTPException(status_code=400, detail=str(exc))
@@ -115,7 +116,7 @@ async def chat(
         conversation_id=conv_id,
         message_id=msg_id,
         content=content,
-        agent_name="jarvis",
+        agent_name=data.runtime or "jarvis",
         model_name=model_name,
     )
 
@@ -141,10 +142,14 @@ async def chat_stream(
                     conversation_id=data.conversation_id,
                     file_ids=data.file_ids,
                     model_name=data.model_name,
+                    runtime=data.runtime,
                 )
             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/folder.py

@@ -1,17 +1,20 @@
 from fastapi import APIRouter, Depends, HTTPException, status
+from sqlalchemy import and_, select
 from sqlalchemy.ext.asyncio import AsyncSession
-from sqlalchemy import select, and_
 from typing import List
+import shutil
+
 from app.database import get_db
 from app.models.folder import Folder
 from app.models.user import User
-from app.schemas.folder import FolderCreate, FolderUpdate, FolderOut, FolderTreeOut
 from app.routers.auth import get_current_user
+from app.schemas.folder import FolderCreate, FolderOut, FolderTreeOut, FolderUpdate
+from app.services.document_service import DocumentService
 
 router = APIRouter(prefix="/api/folders", tags=["文件夹"])
 
+
 def build_folder_tree(folders: list[Folder], parent_id: str = None) -> List[FolderTreeOut]:
-    """递归构建文件夹树"""
     tree = []
     for folder in folders:
         if folder.parent_id == parent_id:
@@ -20,30 +23,29 @@ def build_folder_tree(folders: list[Folder], parent_id: str = None) -> List[Fold
                 id=folder.id,
                 name=folder.name,
                 parent_id=folder.parent_id,
-                children=children
+                children=children,
             ))
     return tree
 
+
 @router.get("", response_model=List[FolderTreeOut])
 async def get_folders(
     db: AsyncSession = Depends(get_db),
-    current_user: User = Depends(get_current_user)
+    current_user: User = Depends(get_current_user),
 ):
-    """获取用户的完整文件夹树"""
     result = await db.execute(
         select(Folder).where(Folder.user_id == current_user.id)
     )
     folders = result.scalars().all()
     return build_folder_tree(list(folders))
 
+
 @router.post("", response_model=FolderOut, status_code=status.HTTP_201_CREATED)
 async def create_folder(
     folder_data: FolderCreate,
     db: AsyncSession = Depends(get_db),
-    current_user: User = Depends(get_current_user)
+    current_user: User = Depends(get_current_user),
 ):
-    """创建文件夹"""
-    # 验证父文件夹存在且属于当前用户
     if folder_data.parent_id:
         result = await db.execute(
             select(Folder).where(
@@ -53,13 +55,12 @@ async def create_folder(
         if not result.scalar_one_or_none():
             raise HTTPException(status_code=404, detail="父文件夹不存在")
 
-    # 检查同名文件夹
     result = await db.execute(
         select(Folder).where(
             and_(
                 Folder.user_id == current_user.id,
                 Folder.parent_id == folder_data.parent_id,
-                Folder.name == folder_data.name
+                Folder.name == folder_data.name,
             )
         )
     )
@@ -69,21 +70,24 @@ async def create_folder(
     folder = Folder(
         user_id=current_user.id,
         name=folder_data.name,
-        parent_id=folder_data.parent_id
+        parent_id=folder_data.parent_id,
     )
     db.add(folder)
     await db.commit()
     await db.refresh(folder)
+
+    document_service = DocumentService(db, current_user.id)
+    await document_service.ensure_folder_directory(current_user.id, folder.id)
     return folder
 
+
 @router.put("/{folder_id}", response_model=FolderOut)
 async def rename_folder(
     folder_id: str,
     folder_data: FolderUpdate,
     db: AsyncSession = Depends(get_db),
-    current_user: User = Depends(get_current_user)
+    current_user: User = Depends(get_current_user),
 ):
-    """重命名文件夹"""
     result = await db.execute(
         select(Folder).where(
             and_(Folder.id == folder_id, Folder.user_id == current_user.id)
@@ -93,18 +97,22 @@ async def rename_folder(
     if not folder:
         raise HTTPException(status_code=404, detail="文件夹不存在")
 
+    old_name = folder.name
     folder.name = folder_data.name
+
+    document_service = DocumentService(db, current_user.id)
+    await document_service.rename_folder_directory(current_user.id, folder.id, old_name, folder_data.name)
     await db.commit()
     await db.refresh(folder)
     return folder
 
+
 @router.delete("/{folder_id}", status_code=status.HTTP_204_NO_CONTENT)
 async def delete_folder(
     folder_id: str,
     db: AsyncSession = Depends(get_db),
-    current_user: User = Depends(get_current_user)
+    current_user: User = Depends(get_current_user),
 ):
-    """删除文件夹（级联删除文档）"""
     from app.models.document import Document
     from app.services.knowledge_service import KnowledgeService
 
@@ -117,15 +125,16 @@ async def delete_folder(
     if not folder:
         raise HTTPException(status_code=404, detail="文件夹不存在")
 
+    document_service = DocumentService(db, current_user.id)
+    folder_path = await document_service._get_storage_directory(current_user.id, folder_id)
+
     async def delete_recursive(fid: str):
-        # 删除子文件夹（先递归）
         children = await db.execute(
             select(Folder).where(Folder.parent_id == fid)
         )
         for child in children.scalars():
             await delete_recursive(child.id)
 
-        # 删除文档
         docs = await db.execute(
             select(Document).where(Document.folder_id == fid)
         )
@@ -134,10 +143,12 @@ async def delete_folder(
             await knowledge_service.delete_from_vectorstore(current_user.id, doc.id)
             await db.delete(doc)
 
-        # 删除文件夹本身
         folder_to_delete = await db.get(Folder, fid)
         if folder_to_delete:
             await db.delete(folder_to_delete)
 
     await delete_recursive(folder_id)
     await db.commit()
+
+    if folder_path.exists():
+        shutil.rmtree(folder_path, ignore_errors=True)

backend/app/routers/office.py

@@ -0,0 +1,179 @@
+"""Office Status API - Star Office style visualization for Jarvis agents."""
+
+from datetime import datetime, timedelta
+from typing import Literal
+from fastapi import APIRouter, HTTPException
+from pydantic import BaseModel
+
+router = APIRouter(prefix="/api/office", tags=["office"])
+
+# ============================================================================
+# State Definitions (mapped to spaceship areas)
+# ============================================================================
+# idle → Rest Bay (breakroom)
+# writing/researching/executing → Command Console (writing)
+# syncing → Server Room (syncing)
+# error → Repair Bay (error)
+
+SHIP_AREAS = {
+    "breakroom": {"x": 200, "y": 300},  # Rest Bay - bottom left
+    "writing": {"x": 640, "y": 200},  # Command Console - center top
+    "server": {"x": 640, "y": 400},  # Server Room - center bottom
+    "error": {"x": 1000, "y": 300},  # Repair Bay - right side
+}
+
+STATES = {
+    "idle": {"name": "待命", "area": "breakroom"},
+    "writing": {"name": "执行中", "area": "writing"},
+    "researching": {"name": "研究中", "area": "writing"},
+    "executing": {"name": "执行中", "area": "writing"},
+    "syncing": {"name": "同步中", "area": "server"},
+    "error": {"name": "故障中", "area": "error"},
+}
+
+
+# ============================================================================
+# Data Models
+# ============================================================================
+class AgentState(BaseModel):
+    agent_id: str
+    name: str
+    state: Literal["idle", "writing", "researching", "executing", "syncing", "error"]
+    detail: str | None = None
+    area: str | None = None
+    is_main: bool = False
+    auth_status: str = "approved"  # approved, pending, rejected, offline
+
+
+class SetStateRequest(BaseModel):
+    state: str
+    detail: str | None = None
+
+
+class OfficeStatus(BaseModel):
+    state: str
+    detail: str | None = None
+    agent_name: str
+    timestamp: str
+
+
+class OfficeMemo(BaseModel):
+    success: bool
+    date: str
+    memo: str
+
+
+# ============================================================================
+# In-Memory State (in production, this would come from Jarvis's agent state)
+# ============================================================================
+_current_state: dict = {
+    "agent_id": "jarvis-main",
+    "name": "JARVIS",
+    "state": "idle",
+    "detail": "战舰启动中...",
+    "area": "breakroom",
+    "is_main": True,
+    "auth_status": "approved",
+}
+
+
+def normalize_state(state: str | None) -> str:
+    """Normalize various state names to our canonical states."""
+    if not state:
+        return "idle"
+    state = state.lower().strip()
+    if state in ("working", "run", "running"):
+        return "writing"
+    if state in ("sync", "syncing"):
+        return "syncing"
+    if state in ("research", "researching"):
+        return "researching"
+    if state in ("execute", "executing"):
+        return "executing"
+    if state == "error":
+        return "error"
+    return "idle"
+
+
+def get_state_info(state: str) -> dict:
+    """Get state info including area mapping."""
+    return STATES.get(state, STATES["idle"])
+
+
+# ============================================================================
+# API Endpoints
+# ============================================================================
+@router.get("/status", response_model=OfficeStatus)
+async def get_status():
+    """Get current agent status."""
+    state_info = get_state_info(_current_state["state"])
+    return OfficeStatus(
+        state=_current_state["state"],
+        detail=_current_state.get("detail"),
+        agent_name=_current_state["name"],
+        timestamp=datetime.now().isoformat(),
+    )
+
+
+@router.get("/yesterday-memo", response_model=OfficeMemo)
+async def get_yesterday_memo():
+    """Return a lightweight public memo for the Star Office viewer."""
+    target_date = (datetime.now() - timedelta(days=1)).date().isoformat()
+    detail = (_current_state.get("detail") or "No detailed log was recorded.").strip()
+    memo = (
+        "Yesterday summary\n"
+        f"- Last known state: {_current_state['state']}\n"
+        f"- Detail: {detail}\n"
+        "- Next step: open the command surface and continue from the current work thread."
+    )
+    return OfficeMemo(success=True, date=target_date, memo=memo)
+
+
+@router.post("/set_state")
+async def set_state(req: SetStateRequest):
+    """Set the current agent state."""
+    normalized = normalize_state(req.state)
+    state_info = get_state_info(normalized)
+
+    _current_state["state"] = normalized
+    _current_state["detail"] = req.detail or ""
+    _current_state["area"] = state_info["area"]
+
+    return {
+        "success": True,
+        "state": normalized,
+        "area": state_info["area"],
+        "detail": _current_state["detail"],
+    }
+
+
+@router.get("/agents")
+async def get_agents():
+    """Get all agents in the office (for multi-agent support)."""
+    # For now, return just the main agent
+    # In full implementation, this would query Jarvis's agent registry
+    state_info = get_state_info(_current_state["state"])
+    return [
+        {
+            "agentId": _current_state["agent_id"],
+            "name": _current_state["name"],
+            "state": _current_state["state"],
+            "detail": _current_state.get("detail", ""),
+            "area": state_info["area"],
+            "isMain": _current_state.get("is_main", True),
+            "authStatus": _current_state.get("auth_status", "approved"),
+            "updated_at": datetime.now().isoformat(),
+        }
+    ]
+
+
+@router.get("/areas")
+async def get_areas():
+    """Get all spaceship areas with coordinates."""
+    return SHIP_AREAS
+
+
+@router.get("/health")
+async def health():
+    """Health check."""
+    return {"status": "ok", "service": "office"}

backend/app/routers/remote_mount.py

@@ -0,0 +1,130 @@
+from fastapi import APIRouter, Depends, HTTPException, status
+from sqlalchemy import and_, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.database import get_db
+from app.models.remote_mount import RemoteMount
+from app.models.user import User
+from app.routers.auth import get_current_user
+from app.schemas.remote_mount import (
+    RemoteMountCreate,
+    RemoteMountOut,
+    RemoteMountTreeOut,
+    RemoteNodeOut,
+    RemoteSyncRequest,
+    RemoteSyncResultOut,
+)
+from app.services.remote_sync_service import RemoteSyncService
+from app.services.secret_service import encrypt_secret
+from app.services.webdav_service import WebDavNode, WebDavService
+
+router = APIRouter(prefix="/api/remote-mounts", tags=["远程挂载"])
+
+
+def _to_node_out(node: WebDavNode) -> RemoteNodeOut:
+    return RemoteNodeOut(
+        path=node.path,
+        name=node.name,
+        is_dir=node.is_dir,
+        size=node.size,
+        modified_at=node.modified_at,
+        etag=node.etag,
+        children=[_to_node_out(child) for child in node.children],
+    )
+
+
+@router.get("", response_model=list[RemoteMountOut])
+async def list_remote_mounts(
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    result = await db.execute(
+        select(RemoteMount).where(RemoteMount.user_id == current_user.id).order_by(RemoteMount.created_at.desc())
+    )
+    return list(result.scalars().all())
+
+
+@router.post("", response_model=RemoteMountOut, status_code=status.HTTP_201_CREATED)
+async def create_remote_mount(
+    payload: RemoteMountCreate,
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    existing = await db.execute(
+        select(RemoteMount).where(and_(RemoteMount.user_id == current_user.id, RemoteMount.name == payload.name))
+    )
+    if existing.scalar_one_or_none():
+        raise HTTPException(status_code=400, detail="同名远程挂载已存在")
+
+    mount = RemoteMount(
+        user_id=current_user.id,
+        name=payload.name,
+        mount_type="webdav",
+        base_url=str(payload.base_url),
+        username=payload.username,
+        password_encrypted=encrypt_secret(payload.password),
+        root_path=payload.root_path,
+        is_active=True,
+    )
+    try:
+        await WebDavService(mount).list_directory(payload.root_path)
+    except Exception as exc:  # noqa: BLE001
+        raise HTTPException(status_code=400, detail=f"WebDAV 连接失败: {exc}") from exc
+
+    db.add(mount)
+    await db.commit()
+    await db.refresh(mount)
+    return mount
+
+
+async def _get_user_mount(db: AsyncSession, user_id: str, mount_id: str) -> RemoteMount:
+    result = await db.execute(
+        select(RemoteMount).where(and_(RemoteMount.id == mount_id, RemoteMount.user_id == user_id))
+    )
+    mount = result.scalar_one_or_none()
+    if mount is None:
+        raise HTTPException(status_code=404, detail="远程挂载不存在")
+    return mount
+
+
+@router.get("/{mount_id}/tree", response_model=RemoteMountTreeOut)
+async def get_remote_tree(
+    mount_id: str,
+    path: str | None = None,
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    mount = await _get_user_mount(db, current_user.id, mount_id)
+    try:
+        nodes = await WebDavService(mount).list_tree(path or mount.root_path)
+    except Exception as exc:  # noqa: BLE001
+        raise HTTPException(status_code=400, detail=f"远程目录读取失败: {exc}") from exc
+
+    return RemoteMountTreeOut(
+        mount_id=mount.id,
+        root_path=path or mount.root_path,
+        nodes=[_to_node_out(node) for node in nodes],
+    )
+
+
+@router.post("/{mount_id}/sync", response_model=RemoteSyncResultOut)
+async def sync_remote_mount(
+    mount_id: str,
+    payload: RemoteSyncRequest,
+    db: AsyncSession = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    mount = await _get_user_mount(db, current_user.id, mount_id)
+    try:
+        result = await RemoteSyncService(db, current_user.id).sync_remote_path(
+            mount,
+            payload.remote_path,
+            payload.local_folder_id,
+            payload.mode,
+        )
+    except ValueError as exc:
+        raise HTTPException(status_code=400, detail=str(exc)) from exc
+    except Exception as exc:  # noqa: BLE001
+        raise HTTPException(status_code=500, detail=f"远程同步失败: {exc}") from exc
+
+    return RemoteSyncResultOut(**result)

backend/app/routers/schedule_center.py

@@ -1,25 +1,62 @@
 from calendar import monthrange
 from datetime import UTC, date, datetime
 
-from fastapi import APIRouter, Depends, HTTPException, Query
+from fastapi import APIRouter, Depends, Query
 from sqlalchemy import select
 from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.orm import selectinload
 
 from app.database import get_db
 from app.models.goal import Goal
 from app.models.reminder import Reminder
-from app.models.task import Task, TaskPriority
+from app.models.task import Task, TaskDispatchStatus, TaskPriority, TaskQuadrant, TaskStatus
 from app.models.todo import DailyTodo
 from app.models.user import User
 from app.routers.auth import get_current_user
 from app.schemas.schedule_center import (
+    ScheduleCenterCommanderSummaryOut,
     ScheduleCenterDateOut,
     ScheduleCenterDaySummary,
+    ScheduleCenterFocusTaskOut,
     ScheduleCenterMonthOut,
+    ScheduleCenterQuadrantOut,
+    ScheduleCenterQuadrantTaskOut,
 )
+from app.schemas.task import build_task_out
 
 router = APIRouter(prefix="/api/schedule-center", tags=["调度中心"])
 
+QUADRANT_META: dict[TaskQuadrant, dict[str, str]] = {
+    TaskQuadrant.URGENT_IMPORTANT: {
+        "title": "重要且紧急",
+        "subtitle": "CRITICAL",
+        "color": "#ff4757",
+        "glow_color": "rgba(255, 71, 87, 0.4)",
+        "icon": "◈",
+    },
+    TaskQuadrant.NOT_URGENT_IMPORTANT: {
+        "title": "重要不紧急",
+        "subtitle": "PLANNED",
+        "color": "#ffd93d",
+        "glow_color": "rgba(255, 217, 61, 0.4)",
+        "icon": "◇",
+    },
+    TaskQuadrant.URGENT_NOT_IMPORTANT: {
+        "title": "紧急不重要",
+        "subtitle": "DELEGATE",
+        "color": "#00d4ff",
+        "glow_color": "rgba(0, 212, 255, 0.4)",
+        "icon": "◉",
+    },
+    TaskQuadrant.NOT_URGENT_NOT_IMPORTANT: {
+        "title": "不重要不紧急",
+        "subtitle": "ELIMINATE",
+        "color": "#6bcf7f",
+        "glow_color": "rgba(107, 207, 127, 0.4)",
+        "icon": "○",
+    },
+}
+
 
 def _build_summary(
     target_date: str,
@@ -39,6 +76,146 @@ def _build_summary(
     )
 
 
+def _coerce_enum(value, enum_cls, default=None):
+    if value is None:
+        return default
+    if isinstance(value, enum_cls):
+        return value
+    if isinstance(value, str):
+        raw = value.strip()
+        if not raw:
+            return default
+        for item in enum_cls:
+            if raw == item.value or raw.lower() == item.value:
+                return item
+            if raw.upper() == item.name:
+                return item
+    return default
+
+
+def _derive_quadrant(task: Task) -> TaskQuadrant:
+    quadrant = _coerce_enum(task.quadrant, TaskQuadrant, None)
+    if quadrant is not None:
+        return quadrant
+
+    priority = _coerce_enum(task.priority, TaskPriority, TaskPriority.MEDIUM)
+    status = _coerce_enum(task.status, TaskStatus, TaskStatus.TODO)
+
+    if priority in {TaskPriority.HIGH, TaskPriority.URGENT}:
+        return TaskQuadrant.URGENT_IMPORTANT
+    if status == TaskStatus.IN_PROGRESS:
+        return TaskQuadrant.NOT_URGENT_IMPORTANT
+    if priority == TaskPriority.MEDIUM:
+        return TaskQuadrant.URGENT_NOT_IMPORTANT
+    return TaskQuadrant.NOT_URGENT_NOT_IMPORTANT
+
+
+def _enum_value(value) -> str | None:
+    if value is None:
+        return None
+    if hasattr(value, "value"):
+        return str(value.value)
+    if isinstance(value, str):
+        raw = value.strip()
+        return raw or None
+    return str(value)
+
+
+def _build_focus_tasks(tasks: list[Task]) -> list[ScheduleCenterFocusTaskOut]:
+    priority_rank = {
+        TaskPriority.URGENT: 0,
+        TaskPriority.HIGH: 1,
+        TaskPriority.MEDIUM: 2,
+        TaskPriority.LOW: 3,
+    }
+    status_rank = {
+        TaskStatus.IN_PROGRESS: 0,
+        TaskStatus.TODO: 1,
+        TaskStatus.DONE: 2,
+        TaskStatus.CANCELLED: 3,
+    }
+    ordered = sorted(
+        tasks,
+        key=lambda item: (
+            status_rank.get(_coerce_enum(item.status, TaskStatus, TaskStatus.TODO), 99),
+            priority_rank.get(_coerce_enum(item.priority, TaskPriority, TaskPriority.MEDIUM), 99),
+            item.created_at,
+        ),
+    )
+    return [
+        ScheduleCenterFocusTaskOut(
+            id=item.id,
+            title=item.title,
+            status=_coerce_enum(item.status, TaskStatus, TaskStatus.TODO),
+            priority=_coerce_enum(item.priority, TaskPriority, TaskPriority.MEDIUM),
+            quadrant=_derive_quadrant(item),
+            assignee_type=_enum_value(item.assignee_type),
+            assignee_id=item.assignee_id,
+            dispatch_status=_coerce_enum(item.dispatch_status, TaskDispatchStatus, TaskDispatchStatus.IDLE),
+            due_date=item.due_date,
+        )
+        for item in ordered[:6]
+    ]
+
+
+def _build_quadrants(tasks: list[Task]) -> list[ScheduleCenterQuadrantOut]:
+    buckets: dict[TaskQuadrant, list[ScheduleCenterQuadrantTaskOut]] = {
+        quadrant: [] for quadrant in QUADRANT_META
+    }
+    for task in tasks:
+        quadrant = _derive_quadrant(task)
+        buckets[quadrant].append(
+            ScheduleCenterQuadrantTaskOut(
+                id=task.id,
+                title=task.title,
+                status=_coerce_enum(task.status, TaskStatus, TaskStatus.TODO),
+                priority=_coerce_enum(task.priority, TaskPriority, TaskPriority.MEDIUM),
+                dispatch_status=_coerce_enum(task.dispatch_status, TaskDispatchStatus, TaskDispatchStatus.IDLE),
+                assignee_type=_enum_value(task.assignee_type),
+                assignee_id=task.assignee_id,
+            )
+        )
+
+    return [
+        ScheduleCenterQuadrantOut(
+            id=quadrant,
+            title=meta["title"],
+            subtitle=meta["subtitle"],
+            color=meta["color"],
+            glow_color=meta["glow_color"],
+            icon=meta["icon"],
+            tasks=buckets[quadrant],
+        )
+        for quadrant, meta in QUADRANT_META.items()
+    ]
+
+
+def _build_commander_summary(tasks: list[Task]) -> ScheduleCenterCommanderSummaryOut:
+    counts = ScheduleCenterCommanderSummaryOut()
+    for task in tasks:
+        states = [task.dispatch_status, *(subtask.dispatch_status for subtask in task.subtasks)]
+        for state in states:
+            normalized = _coerce_enum(state, TaskDispatchStatus, TaskDispatchStatus.IDLE)
+            if normalized == TaskDispatchStatus.IDLE:
+                continue
+            counts.total += 1
+            if normalized == TaskDispatchStatus.QUEUED:
+                counts.queued += 1
+            elif normalized == TaskDispatchStatus.RUNNING:
+                counts.running += 1
+            elif normalized == TaskDispatchStatus.COMPLETED:
+                counts.completed += 1
+            elif normalized == TaskDispatchStatus.FAILED:
+                counts.failed += 1
+    if counts.running > 0:
+        counts.overall_status = "running"
+    elif counts.queued > 0:
+        counts.overall_status = "queued"
+    elif counts.failed > 0 and counts.completed == 0:
+        counts.overall_status = "failed"
+    return counts
+
+
 @router.get("/month", response_model=ScheduleCenterMonthOut)
 async def get_month_schedule(
     year: int = Query(..., ge=2000, le=2100),
@@ -53,27 +230,43 @@ async def get_month_schedule(
     start_dt = datetime.combine(month_start, datetime.min.time())
     end_dt = datetime.combine(month_start.replace(day=days_in_month), datetime.max.time())
 
-    todos = (await db.execute(
+    todos = (
-        select(DailyTodo).where(DailyTodo.user_id == current_user.id, DailyTodo.todo_date >= start_key, DailyTodo.todo_date <= end_key)
+        await db.execute(
-    )).scalars().all()
+            select(DailyTodo).where(
-    tasks = (await db.execute(
+                DailyTodo.user_id == current_user.id,
+                DailyTodo.todo_date >= start_key,
+                DailyTodo.todo_date <= end_key,
+            )
+        )
+    ).scalars().all()
+    tasks = (
+        await db.execute(
             select(Task).where(
                 Task.user_id == current_user.id,
                 Task.due_date.is_not(None),
                 Task.due_date >= start_dt,
                 Task.due_date <= end_dt,
             )
-    )).scalars().all()
+        )
-    reminders = (await db.execute(
+    ).scalars().all()
+    reminders = (
+        await db.execute(
             select(Reminder).where(
                 Reminder.user_id == current_user.id,
                 Reminder.reminder_at >= start_dt,
                 Reminder.reminder_at <= end_dt,
             )
-    )).scalars().all()
+        )
-    goals = (await db.execute(
+    ).scalars().all()
-        select(Goal).where(Goal.user_id == current_user.id, Goal.goal_date >= start_key, Goal.goal_date <= end_key)
+    goals = (
-    )).scalars().all()
+        await db.execute(
+            select(Goal).where(
+                Goal.user_id == current_user.id,
+                Goal.goal_date >= start_key,
+                Goal.goal_date <= end_key,
+            )
+        )
+    ).scalars().all()
 
     todo_map: dict[str, list[DailyTodo]] = {}
     for item in todos:
@@ -96,18 +289,20 @@ async def get_month_schedule(
     days = []
     for day in range(1, days_in_month + 1):
         date_key = month_start.replace(day=day).isoformat()
-        days.append(_build_summary(
+        days.append(
+            _build_summary(
                 date_key,
                 todo_map.get(date_key, []),
                 task_map.get(date_key, []),
                 reminder_map.get(date_key, []),
                 goal_map.get(date_key, []),
-        ))
+            )
+        )
 
     return ScheduleCenterMonthOut(month=f"{year:04d}-{month:02d}", days=days)
 
 
-@router.get("/date", response_model=ScheduleCenterDateOut)
+@router.get("/date", response_model=ScheduleCenterDateOut, response_model_exclude_none=True)
 async def get_date_schedule(
     date_str: date = Query(...),
     current_user: User = Depends(get_current_user),
@@ -118,22 +313,28 @@ async def get_date_schedule(
     end_dt = datetime.combine(target_date, datetime.max.time())
     date_key = target_date.isoformat()
 
-    todos = (await db.execute(
+    todos = (
+        await db.execute(
             select(DailyTodo)
             .where(DailyTodo.user_id == current_user.id, DailyTodo.todo_date == date_key)
             .order_by(DailyTodo.created_at.desc())
-    )).scalars().all()
+        )
-    tasks = (await db.execute(
+    ).scalars().all()
+    tasks = (
+        await db.execute(
             select(Task)
+            .options(selectinload(Task.subtasks), selectinload(Task.history))
             .where(
                 Task.user_id == current_user.id,
                 Task.due_date.is_not(None),
                 Task.due_date >= start_dt,
                 Task.due_date <= end_dt,
             )
-        .order_by(Task.created_at.desc())
+            .order_by(Task.priority.desc(), Task.created_at.desc())
-    )).scalars().all()
+        )
-    reminders = (await db.execute(
+    ).scalars().unique().all()
+    reminders = (
+        await db.execute(
             select(Reminder)
             .where(
                 Reminder.user_id == current_user.id,
@@ -141,20 +342,26 @@ async def get_date_schedule(
                 Reminder.reminder_at <= end_dt,
             )
             .order_by(Reminder.reminder_at.asc(), Reminder.created_at.asc())
-    )).scalars().all()
+        )
-    goals = (await db.execute(
+    ).scalars().all()
+    goals = (
+        await db.execute(
             select(Goal)
             .where(Goal.user_id == current_user.id, Goal.goal_date == date_key)
             .order_by(Goal.created_at.desc())
-    )).scalars().all()
+        )
+    ).scalars().all()
 
     summary = _build_summary(date_key, todos, tasks, reminders, goals)
     return ScheduleCenterDateOut(
         date=date_key,
         todos=todos,
-        tasks=tasks,
+        tasks=[build_task_out(task) for task in tasks],
         reminders=reminders,
         goals=goals,
         summary=summary,
+        focus_tasks=_build_focus_tasks(tasks),
+        quadrants=_build_quadrants(tasks),
+        commander_summary=_build_commander_summary(tasks),
         generated_at=datetime.now(UTC),
     )

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

@@ -7,3 +7,9 @@ router = APIRouter(prefix='/api/system', tags=['system'])
 @router.get('/status')
 async def get_system_status():
     return SystemService().get_status()
+
+
+@router.get('/config')
+async def get_system_config():
+    """Get public system configuration."""
+    return await SystemService().get_config()

backend/app/routers/task.py

@@ -1,15 +1,116 @@
+import json
 from datetime import UTC, date, datetime
 
 from fastapi import APIRouter, Depends, HTTPException, Query
 from sqlalchemy import desc, select
 from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.orm import selectinload
+
 from app.database import get_db
-from app.models.task import Task, TaskStatus
+from app.models.task import (
+    Task,
+    TaskAssigneeType,
+    TaskDispatchStatus,
+    TaskQuadrant,
+    TaskSource,
+    TaskStatus,
+    TaskSubTask,
+)
 from app.models.user import User
 from app.routers.auth import get_current_user
-from app.schemas.task import TaskCreate, TaskUpdate, TaskOut
+from app.schemas.task import (
+    TaskCreate,
+    TaskDetailOut,
+    TaskDispatchRequest,
+    TaskDispatchResponse,
+    TaskHistoryOut,
+    TaskOut,
+    TaskSubTaskCreate,
+    TaskSubTaskOut,
+    TaskSubTaskReorderRequest,
+    TaskSubTaskUpdate,
+    TaskUpdate,
+    build_task_detail_out,
+)
+from app.services.task_dispatch import append_task_history, load_task_with_details, queue_task_dispatch
 
-router = APIRouter(prefix="/api/tasks", tags=["看板"])
+router = APIRouter(prefix="/api/tasks", tags=["Tasks"])
+
+
+def _encode_tags(tags: list[str] | None) -> str | None:
+    if not tags:
+        return None
+    return json.dumps(tags, ensure_ascii=False)
+
+
+def _decode_tags(value: str | None) -> list[str]:
+    if not value:
+        return []
+    try:
+        payload = json.loads(value)
+    except json.JSONDecodeError:
+        return [value]
+    if isinstance(payload, list):
+        return [str(item) for item in payload]
+    return [str(payload)]
+
+
+def _subtask_to_out(subtask: TaskSubTask) -> TaskSubTaskOut:
+    return TaskSubTaskOut.model_validate(subtask)
+
+
+def _history_to_out(history) -> TaskHistoryOut:
+    return TaskHistoryOut.model_validate(history)
+
+
+def _task_to_out(task: Task) -> TaskOut:
+    return TaskOut(
+        id=task.id,
+        title=task.title,
+        description=task.description,
+        status=task.status,
+        priority=task.priority,
+        due_date=task.due_date,
+        completed_at=task.completed_at,
+        tags=_decode_tags(task.tags),
+        source=task.source or TaskSource.MANUAL,
+        conversation_id=task.conversation_id,
+        quadrant=task.quadrant,
+        assignee_type=task.assignee_type,
+        assignee_id=task.assignee_id,
+        dispatch_status=task.dispatch_status or TaskDispatchStatus.IDLE,
+        dispatch_run_id=task.dispatch_run_id,
+        result_summary=task.result_summary,
+        started_at=task.started_at,
+        last_synced_at=task.last_synced_at,
+        created_at=task.created_at,
+        updated_at=task.updated_at,
+    )
+
+
+def _task_detail_to_out(task: Task) -> TaskDetailOut:
+    return build_task_detail_out(task)
+
+
+async def _get_task_or_404(db: AsyncSession, *, task_id: str, user_id: str) -> Task:
+    task = await load_task_with_details(db, task_id=task_id, user_id=user_id)
+    if task is None:
+        raise HTTPException(status_code=404, detail="Task not found")
+    return task
+
+
+def _sync_task_completion(task: Task) -> None:
+    if task.status == TaskStatus.DONE:
+        task.completed_at = task.completed_at or datetime.now(UTC)
+    elif task.status != TaskStatus.CANCELLED:
+        task.completed_at = None
+
+
+def _sync_subtask_completion(subtask: TaskSubTask) -> None:
+    if subtask.status == TaskStatus.DONE:
+        subtask.completed_at = subtask.completed_at or datetime.now(UTC)
+    elif subtask.status != TaskStatus.CANCELLED:
+        subtask.completed_at = None
 
 
 @router.get("", response_model=list[TaskOut])
@@ -18,12 +119,28 @@ async def list_tasks(
     due_date: date | None = Query(default=None),
     date_from: date | None = Query(default=None),
     date_to: date | None = Query(default=None),
+    quadrant: TaskQuadrant | None = None,
+    assignee_type: TaskAssigneeType | None = None,
+    dispatch_status: TaskDispatchStatus | None = None,
+    conversation_id: str | None = None,
     current_user: User = Depends(get_current_user),
     db: AsyncSession = Depends(get_db),
 ):
-    query = select(Task).where(Task.user_id == current_user.id)
+    query = (
+        select(Task)
+        .options(selectinload(Task.subtasks), selectinload(Task.history))
+        .where(Task.user_id == current_user.id)
+    )
     if status:
         query = query.where(Task.status == status)
+    if quadrant:
+        query = query.where(Task.quadrant == quadrant)
+    if assignee_type:
+        query = query.where(Task.assignee_type == assignee_type)
+    if dispatch_status:
+        query = query.where(Task.dispatch_status == dispatch_status)
+    if conversation_id:
+        query = query.where(Task.conversation_id == conversation_id)
     if due_date:
         start = datetime.combine(due_date, datetime.min.time())
         end = datetime.combine(due_date, datetime.max.time())
@@ -32,65 +149,109 @@ async def list_tasks(
         start = datetime.combine(date_from, datetime.min.time()) if date_from else None
         end = datetime.combine(date_to, datetime.max.time()) if date_to else None
         if start and end and start > end:
-            raise HTTPException(status_code=400, detail="开始日期不能晚于结束日期")
+            raise HTTPException(status_code=400, detail="date_from cannot be later than date_to")
         if start is not None:
             query = query.where(Task.due_date.is_not(None), Task.due_date >= start)
         if end is not None:
             query = query.where(Task.due_date.is_not(None), Task.due_date <= end)
-    query = query.order_by(desc(Task.created_at))
+
+    query = query.order_by(desc(Task.updated_at), desc(Task.created_at))
     result = await db.execute(query)
-    return result.scalars().all()
+    tasks = result.scalars().unique().all()
+    return [_task_to_out(task) for task in tasks]
 
 
-@router.post("", response_model=TaskOut, status_code=201)
+@router.post("", response_model=TaskDetailOut, status_code=201)
 async def create_task(
     data: TaskCreate,
     current_user: User = Depends(get_current_user),
     db: AsyncSession = Depends(get_db),
 ):
-    import json
     task = Task(
         user_id=current_user.id,
         title=data.title,
         description=data.description,
         priority=data.priority,
         due_date=data.due_date,
-        tags=json.dumps(data.tags) if data.tags else None,
+        tags=_encode_tags(data.tags),
+        source=data.source,
+        conversation_id=data.conversation_id,
+        quadrant=data.quadrant,
+        assignee_type=data.assignee_type,
+        assignee_id=data.assignee_id,
+        status=data.status,
     )
+    _sync_task_completion(task)
+    if data.source == TaskSource.CHAT:
+        append_task_history(task, action="created_from_chat", new_value=task.title)
+    append_task_history(task, action="created", new_value=task.title)
+    for index, subtask_data in enumerate(data.subtasks):
+        subtask = TaskSubTask(
+            title=subtask_data.title,
+            description=subtask_data.description,
+            status=subtask_data.status,
+            order_index=index if subtask_data.order_index is None else subtask_data.order_index,
+            assignee_type=subtask_data.assignee_type,
+            assignee_id=subtask_data.assignee_id,
+        )
+        _sync_subtask_completion(subtask)
+        task.subtasks.append(subtask)
+        append_task_history(task, action="subtask_created", new_value=subtask.title)
     db.add(task)
     await db.commit()
-    await db.refresh(task)
+
-    return task
+    task = await _get_task_or_404(db, task_id=task.id, user_id=current_user.id)
+    if data.dispatch_to_commander:
+        await queue_task_dispatch(task, db=db)
+        task = await _get_task_or_404(db, task_id=task.id, user_id=current_user.id)
+    return _task_detail_to_out(task)
 
 
-@router.patch("/{task_id}", response_model=TaskOut)
+@router.get("/{task_id}", response_model=TaskDetailOut)
+async def get_task(
+    task_id: str,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    task = await _get_task_or_404(db, task_id=task_id, user_id=current_user.id)
+    return _task_detail_to_out(task)
+
+
+@router.patch("/{task_id}", response_model=TaskDetailOut)
 async def update_task(
     task_id: str,
     data: TaskUpdate,
     current_user: User = Depends(get_current_user),
     db: AsyncSession = Depends(get_db),
 ):
-    import json
+    task = await _get_task_or_404(db, task_id=task_id, user_id=current_user.id)
-    result = await db.execute(
+    payload = data.model_dump(exclude_none=True)
-        select(Task).where(Task.id == task_id, Task.user_id == current_user.id)
+    previous_assignee = (task.assignee_type, task.assignee_id)
-    )
-    task = result.scalar_one_or_none()
-    if not task:
-        raise HTTPException(status_code=404, detail="任务不存在")
 
-    for field, value in data.model_dump(exclude_none=True).items():
+    for field, value in payload.items():
+        previous = getattr(task, field)
         if field == "tags":
-            setattr(task, field, json.dumps(value))
+            task.tags = _encode_tags(value)
-        elif field == "status" and value == TaskStatus.DONE:
+            append_task_history(task, action="updated", old_value=_decode_tags(previous), new_value=value)
-            task.completed_at = datetime.now(UTC)
+            continue
-            setattr(task, field, value)
-        elif field == "status":
-            task.completed_at = None
         setattr(task, field, value)
+        if field == "status":
+            _sync_task_completion(task)
+            append_task_history(task, action="status_changed", old_value=previous, new_value=value)
+        elif previous != value:
+            append_task_history(task, action="updated", old_value=previous, new_value=value)
+
+    if ("assignee_type" in payload or "assignee_id" in payload) and previous_assignee != (task.assignee_type, task.assignee_id):
+        append_task_history(
+            task,
+            action="assigned",
+            old_value=f"{previous_assignee[0]}:{previous_assignee[1]}",
+            new_value=f"{task.assignee_type}:{task.assignee_id}",
+        )
 
     await db.commit()
-    await db.refresh(task)
+    task = await _get_task_or_404(db, task_id=task_id, user_id=current_user.id)
-    return task
+    return _task_detail_to_out(task)
 
 
 @router.delete("/{task_id}", status_code=204)
@@ -99,11 +260,171 @@ async def delete_task(
     current_user: User = Depends(get_current_user),
     db: AsyncSession = Depends(get_db),
 ):
-    result = await db.execute(
+    task = await _get_task_or_404(db, task_id=task_id, user_id=current_user.id)
-        select(Task).where(Task.id == task_id, Task.user_id == current_user.id)
-    )
-    task = result.scalar_one_or_none()
-    if not task:
-        raise HTTPException(status_code=404, detail="任务不存在")
     await db.delete(task)
     await db.commit()
+
+
+@router.post("/{task_id}/subtasks", status_code=201)
+async def create_subtask(
+    task_id: str,
+    data: TaskSubTaskCreate,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    task = await _get_task_or_404(db, task_id=task_id, user_id=current_user.id)
+    max_order = max((item.order_index for item in task.subtasks), default=-1)
+    subtask = TaskSubTask(
+        task_id=task.id,
+        title=data.title,
+        description=data.description,
+        status=data.status,
+        order_index=max_order + 1 if data.order_index is None else data.order_index,
+        assignee_type=data.assignee_type,
+        assignee_id=data.assignee_id,
+    )
+    _sync_subtask_completion(subtask)
+    task.subtasks.append(subtask)
+    append_task_history(task, action="subtask_created", new_value=data.title)
+    await db.commit()
+    db.expire_all()
+    task = await _get_task_or_404(db, task_id=task_id, user_id=current_user.id)
+    detail = _task_detail_to_out(task)
+    created_subtask = max(
+        (item for item in detail.subtasks if item.title == data.title),
+        key=lambda item: (item.order_index, item.created_at),
+        default=None,
+    )
+    if created_subtask is None:
+        raise HTTPException(status_code=500, detail="Created subtask could not be loaded")
+    return {
+        **created_subtask.model_dump(),
+        "task": detail.model_dump(),
+        "subtasks": [item.model_dump() for item in detail.subtasks],
+        "history": [item.model_dump() for item in detail.history],
+        "dispatch": detail.dispatch.model_dump(),
+        "dispatch_summary": detail.dispatch_summary.model_dump(),
+    }
+
+
+@router.patch("/{task_id}/subtasks/{subtask_id}", response_model=TaskDetailOut)
+async def update_subtask(
+    task_id: str,
+    subtask_id: str,
+    data: TaskSubTaskUpdate,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    task = await _get_task_or_404(db, task_id=task_id, user_id=current_user.id)
+    subtask = next((item for item in task.subtasks if item.id == subtask_id), None)
+    if subtask is None:
+        raise HTTPException(status_code=404, detail="Subtask not found")
+
+    payload = data.model_dump(exclude_none=True)
+    for field, value in payload.items():
+        previous = getattr(subtask, field)
+        setattr(subtask, field, value)
+        if field == "status":
+            _sync_subtask_completion(subtask)
+        if previous != value:
+            append_task_history(
+                task,
+                action="updated" if field != "status" else "status_changed",
+                old_value=f"{subtask.id}:{field}:{previous}",
+                new_value=f"{subtask.id}:{field}:{value}",
+            )
+
+    await db.commit()
+    db.expire_all()
+    task = await _get_task_or_404(db, task_id=task_id, user_id=current_user.id)
+    return _task_detail_to_out(task)
+
+
+@router.delete("/{task_id}/subtasks/{subtask_id}", response_model=TaskDetailOut)
+async def delete_subtask(
+    task_id: str,
+    subtask_id: str,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    task = await _get_task_or_404(db, task_id=task_id, user_id=current_user.id)
+    subtask = next((item for item in task.subtasks if item.id == subtask_id), None)
+    if subtask is None:
+        raise HTTPException(status_code=404, detail="Subtask not found")
+
+    append_task_history(task, action="updated", old_value="subtask_deleted", new_value=subtask.title)
+    await db.delete(subtask)
+    await db.commit()
+    db.expire_all()
+    task = await _get_task_or_404(db, task_id=task_id, user_id=current_user.id)
+    return _task_detail_to_out(task)
+
+
+@router.post("/{task_id}/subtasks/reorder", response_model=TaskDetailOut)
+async def reorder_subtasks(
+    task_id: str,
+    data: TaskSubTaskReorderRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    task = await _get_task_or_404(db, task_id=task_id, user_id=current_user.id)
+    lookup = {item.id: item for item in task.subtasks}
+    for item in data.items:
+        subtask = lookup.get(item.id)
+        if subtask is None:
+            raise HTTPException(status_code=404, detail=f"Subtask not found: {item.id}")
+        subtask.order_index = item.order_index
+
+    append_task_history(
+        task,
+        action="subtask_reordered",
+        new_value=",".join(f"{item.id}:{item.order_index}" for item in data.items),
+    )
+    await db.commit()
+    db.expire_all()
+    task = await _get_task_or_404(db, task_id=task_id, user_id=current_user.id)
+    return _task_detail_to_out(task)
+
+
+@router.post("/{task_id}/dispatch", response_model=TaskDispatchResponse)
+async def dispatch_task(
+    task_id: str,
+    data: TaskDispatchRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    if data.target != "commander":
+        raise HTTPException(status_code=400, detail="Only commander dispatch is supported")
+    task = await _get_task_or_404(db, task_id=task_id, user_id=current_user.id)
+    _, payload = await queue_task_dispatch(task, db=db)
+    task = await _get_task_or_404(db, task_id=task_id, user_id=current_user.id)
+    return TaskDispatchResponse(
+        status=task.dispatch_status,
+        run_id=task.dispatch_run_id,
+        task=_task_detail_to_out(task),
+        payload=payload,
+    )
+
+
+@router.post("/{task_id}/subtasks/{subtask_id}/dispatch", response_model=TaskDispatchResponse)
+async def dispatch_subtask(
+    task_id: str,
+    subtask_id: str,
+    data: TaskDispatchRequest,
+    current_user: User = Depends(get_current_user),
+    db: AsyncSession = Depends(get_db),
+):
+    if data.target != "commander":
+        raise HTTPException(status_code=400, detail="Only commander dispatch is supported")
+    task = await _get_task_or_404(db, task_id=task_id, user_id=current_user.id)
+    subtask = next((item for item in task.subtasks if item.id == subtask_id), None)
+    if subtask is None:
+        raise HTTPException(status_code=404, detail="Subtask not found")
+    _, payload = await queue_task_dispatch(task, db=db, subtask=subtask)
+    task = await _get_task_or_404(db, task_id=task_id, user_id=current_user.id)
+    return TaskDispatchResponse(
+        status=subtask.dispatch_status,
+        run_id=subtask.dispatch_run_id,
+        task=_task_detail_to_out(task),
+        payload=payload,
+    )

backend/app/routers/terminal.py

@@ -0,0 +1,79 @@
+"""
+Terminal WebSocket Router - 终端 WebSocket 端点
+"""
+
+from fastapi import APIRouter, WebSocket, WebSocketDisconnect
+
+from app.agents.tools.terminal_engine import pty_manager
+
+router = APIRouter(prefix="/ws/terminal", tags=["terminal"])
+
+
+class ConnectionManager:
+    """WebSocket 连接管理器"""
+
+    def __init__(self):
+        self.active_connections: dict[str, WebSocket] = {}
+
+    async def connect(self, session_id: str, websocket: WebSocket):
+        await websocket.accept()
+        self.active_connections[session_id] = websocket
+
+    def disconnect(self, session_id: str):
+        if session_id in self.active_connections:
+            del self.active_connections[session_id]
+
+    async def send(self, session_id: str, data: str):
+        if session_id in self.active_connections:
+            await self.active_connections[session_id].send_text(data)
+
+    def is_connected(self, session_id: str) -> bool:
+        return session_id in self.active_connections
+
+
+manager = ConnectionManager()
+
+
+@router.websocket("/{session_id}")
+async def terminal_websocket(websocket: WebSocket, session_id: str):
+    """终端 WebSocket 端点"""
+    await manager.connect(session_id, websocket)
+
+    try:
+        # 获取该 session 的输出队列
+        queue = pty_manager._output_queues.get(session_id)
+        if queue:
+            # 异步任务：转发 PTY 输出到 WebSocket
+            async def forward_output():
+                while True:
+                    try:
+                        data = await asyncio.wait_for(queue.get(), timeout=0.1)
+                        if data is None:
+                            await manager.send(session_id, "[SESSION_END]")
+                            break
+                        await manager.send(session_id, data)
+                    except asyncio.TimeoutError:
+                        # 检查连接是否还活跃
+                        if not manager.is_connected(session_id):
+                            break
+                    except Exception:
+                        break
+
+            import asyncio
+
+            forward_task = asyncio.create_task(forward_output())
+
+            # 主循环：接收用户输入
+            try:
+                while True:
+                    data = await websocket.receive_text()
+                    # 接收用户输入，写入 PTY
+                    await pty_manager.write(session_id, data + "\n")
+            except WebSocketDisconnect:
+                pass
+            finally:
+                forward_task.cancel()
+    except Exception:
+        pass
+    finally:
+        manager.disconnect(session_id)

backend/app/routers/tools.py

@@ -0,0 +1,348 @@
+"""Tools API Router
+
+聚合两套工具体系的元数据：
+1. 注册层 (app/tools/) - YAML manifest 定义
+2. Agent 层 (app/agents/tools/) - @tool 装饰器定义
+"""
+
+import re
+import importlib
+
+from fastapi import APIRouter, Depends
+
+from app.routers.auth import get_current_user
+from app.models.user import User
+from app.schemas.tools import (
+    ToolsResponse,
+    ToolCategory,
+    ToolSubgroup,
+    ToolInfo,
+    ToolCommand,
+    ToolStats,
+    ToolSummary,
+)
+
+router = APIRouter(prefix="/api/tools", tags=["Tools"])
+
+# ============================================================
+# 辅助函数
+# ============================================================
+
+
+def _parse_command_from_docstring(docstring: str) -> dict:
+    """从函数的 docstring 解析参数信息"""
+    params = {"type": "object", "properties": {}, "required": []}
+    if not docstring:
+        return params
+
+    # 简单解析 Args: 段落
+    args_match = re.search(
+        r"Args:\s*(.*?)(?=\n\s*(?:Returns?|Raises?)|$", docstring, re.DOTALL | re.IGNORECASE
+    )
+    if args_match:
+        args_section = args_match.group(1)
+        # 匹配形如 "arg_name (type): description" 的行
+        for line in args_section.strip().split("\n"):
+            line = line.strip()
+            if not line:
+                continue
+            # 匹配: "name (type): description" 或 "name: description"
+            m = re.match(r"(\w+)\s*(?:\(\s*(\w+)\s*\))?\s*:", line)
+            if m:
+                param_name = m.group(1)
+                params["properties"][param_name] = {"type": "string", "description": line}
+                params["required"].append(param_name)
+
+    return params
+
+
+def _build_agent_tools() -> list[ToolInfo]:
+    """扫描 app/agents/tools/ 目录，内省 @tool 装饰器"""
+    tools: list[ToolInfo] = []
+
+    # 分类映射：文件名 -> (分类名, 子分类名)
+    category_map = {
+        "search": ("Agent层", "知识检索"),
+        "schedule": ("Agent层", "日程管理"),
+        "task": ("Agent层", "任务管理"),
+        "forum": ("Agent层", "论坛功能"),
+        "time_reasoning": ("Agent层", "时间推理"),
+        "builtins/file_tools": ("Agent层", "文件工具"),
+        "builtins/system_tools": ("Agent层", "系统命令"),
+        "builtins/dev_tools": ("Agent层", "开发工具"),
+        "builtins/collaboration_tools": ("Agent层", "协作工具"),
+    }
+
+    # 工具名称 -> 中文显示名
+    display_names = {
+        "search_knowledge": "知识库搜索",
+        "get_knowledge_graph_context": "知识图谱查询",
+        "build_knowledge_graph": "构建知识图谱",
+        "hybrid_search": "混合搜索",
+        "web_search": "联网搜索",
+        "get_schedule_day": "获取日程",
+        "create_todo": "创建待办",
+        "create_schedule_task": "创建日程任务",
+        "create_reminder": "创建提醒",
+        "create_goal": "创建目标",
+        "get_tasks": "获取任务列表",
+        "create_task": "创建任务",
+        "update_task_status": "更新任务状态",
+        "get_forum_posts": "获取论坛帖子",
+        "create_forum_post": "发布论坛帖子",
+        "scan_forum_for_instructions": "扫描论坛指令",
+        "resolve_time_expression": "解析时间表达式",
+        "glob": "文件路径匹配",
+        "grep": "文件内容搜索",
+        "read_file": "读取文件",
+        "write_file": "写入文件",
+        "bash": "Bash命令",
+        "powershell": "PowerShell命令",
+        "git": "Git操作",
+        "lsp_tools": "LSP代码导航",
+        "team_agent": "团队Agent通信",
+        "task_broadcast": "任务广播",
+    }
+
+    # 工具描述
+    descriptions = {
+        "search_knowledge": "搜索用户的私人知识库，返回最相关的文档片段",
+        "get_knowledge_graph_context": "获取用户知识图谱的上下文信息",
+        "build_knowledge_graph": "从文档构建/更新知识图谱",
+        "hybrid_search": "混合搜索，结合向量语义检索和关键词匹配",
+        "web_search": "通过 SearxNG 搜索外部网页信息",
+        "get_schedule_day": "获取指定日期的 todo/task/reminder/goal 聚合信息",
+        "create_todo": "创建指定日期的待办",
+        "create_schedule_task": "创建任务，支持优先级和截止日期",
+        "create_reminder": "创建提醒，支持自然语言时间",
+        "create_goal": "创建指定日期的目标",
+        "get_tasks": "获取用户当前的任务列表",
+        "create_task": "创建新任务",
+        "update_task_status": "更新任务状态",
+        "get_forum_posts": "获取论坛帖子列表",
+        "create_forum_post": "在论坛发布新帖子",
+        "scan_forum_for_instructions": "扫描论坛中的指令类帖子",
+        "resolve_time_expression": "解析中文自然语言时间表达",
+        "glob": "使用 glob 模式查找文件路径",
+        "grep": "在文件中搜索匹配的文本行",
+        "read_file": "读取文件内容",
+        "write_file": "写入文件内容",
+        "bash": "执行 Bash 命令",
+        "powershell": "执行 PowerShell 命令",
+        "git": "执行 Git 命令",
+        "lsp_tools": "LSP 代码导航和查找引用",
+        "team_agent": "向团队 Agent 发送消息或请求协作",
+        "task_broadcast": "向多个 Agent 广播任务",
+    }
+
+    # 需要扫描的模块
+    modules_to_scan = [
+        ("app.agents.tools.search", "search"),
+        ("app.agents.tools.schedule", "schedule"),
+        ("app.agents.tools.task", "task"),
+        ("app.agents.tools.forum", "forum"),
+        ("app.agents.tools.time_reasoning", "time_reasoning"),
+        ("app.agents.tools.builtins.file_tools", "builtins/file_tools"),
+        ("app.agents.tools.builtins.system_tools", "builtins/system_tools"),
+        ("app.agents.tools.builtins.dev_tools", "builtins/dev_tools"),
+        ("app.agents.tools.builtins.collaboration_tools", "builtins/collaboration_tools"),
+    ]
+
+    for module_name, category_key in modules_to_scan:
+        try:
+            mod = importlib.import_module(module_name)
+        except ImportError:
+            continue
+
+        # 扫描模块中所有 @tool 装饰的函数
+        for attr_name in dir(mod):
+            if attr_name.startswith("_"):
+                continue
+            attr = getattr(mod, attr_name)
+            # 检查是否是 langchain @tool 装饰的对象
+            if hasattr(attr, "name") and hasattr(attr, "description"):
+                tool_name = attr.name
+                tool_desc = attr.description or ""
+                # 清理 docstring 中的参数说明用于显示
+                display_desc = re.sub(r"\s*Args:\s*.*", "", tool_desc, flags=re.DOTALL).strip()
+                display_desc = re.sub(
+                    r"\s*Returns?:\s*.*", "", display_desc, flags=re.DOTALL
+                ).strip()
+
+                # 获取 category 和 subcategory
+                cat_info = category_map.get(category_key, ("Agent层", category_key))
+                category, subcategory = cat_info[0], cat_info[1]
+
+                # 获取参数 schema
+                params_schema = getattr(attr, "args_schema", None)
+                parameters = {}
+                if params_schema:
+                    try:
+                        if hasattr(params_schema, "model_json_schema"):
+                            parameters = params_schema.model_json_schema()
+                        elif hasattr(params_schema, "schema"):
+                            parameters = params_schema.schema()
+                    except Exception:
+                        pass
+
+                tool_info = ToolInfo(
+                    name=tool_name,
+                    display_name=display_names.get(tool_name, tool_name),
+                    description=descriptions.get(tool_name, display_desc or tool_desc),
+                    category=category,
+                    subcategory=subcategory,
+                    source="agent",
+                    source_file=module_name,
+                    tags=[],
+                    enabled=True,
+                    commands=[
+                        ToolCommand(
+                            name=tool_name,
+                            description=tool_desc or display_desc,
+                            parameters=parameters,
+                        )
+                    ],
+                    stats=ToolStats(),
+                )
+                tools.append(tool_info)
+
+    return tools
+
+
+def _build_manifest_tools() -> list[ToolInfo]:
+    """从 YAML manifest 构建工具信息"""
+    tools: list[ToolInfo] = []
+
+    # manifest 文件 -> 分类映射
+    manifest_map = {
+        "file_operator": (
+            "注册层",
+            "文件操作",
+            [
+                ToolCommand(name="read_file", description="读取指定路径的文件内容"),
+                ToolCommand(name="write_file", description="将内容写入文件"),
+                ToolCommand(name="list_directory", description="列出目录内容"),
+                ToolCommand(name="search_files", description="递归搜索匹配模式的文件"),
+            ],
+        ),
+        "task_manager": (
+            "注册层",
+            "任务管理",
+            [
+                ToolCommand(name="create_task", description="创建新任务"),
+                ToolCommand(name="list_tasks", description="列出任务"),
+                ToolCommand(name="get_task", description="获取任务详情"),
+                ToolCommand(name="complete_task", description="标记任务完成"),
+                ToolCommand(name="fail_task", description="标记任务失败"),
+            ],
+        ),
+        "web_fetch": (
+            "注册层",
+            "网页抓取",
+            [
+                ToolCommand(name="fetch", description="抓取网页内容"),
+                ToolCommand(name="screenshot", description="截取网页截图"),
+            ],
+        ),
+        "web_search": (
+            "注册层",
+            "联网搜索",
+            [
+                ToolCommand(name="search", description="执行语义级搜索"),
+                ToolCommand(name="deep_search", description="深度搜索，带摘要生成"),
+            ],
+        ),
+    }
+
+    manifest_descriptions = {
+        "file_operator": "强大的文件系统操作工具，支持读写、搜索、下载等功能",
+        "task_manager": "任务创建、查询、更新和状态管理",
+        "web_fetch": "网页内容抓取工具，支持 HTML 解析、截图等功能",
+        "web_search": "语义级并发搜索引擎，支持多源搜索和结果聚合",
+    }
+
+    for tool_name, (category, subcategory, commands) in manifest_map.items():
+        tool_info = ToolInfo(
+            name=tool_name,
+            display_name=subcategory,
+            description=manifest_descriptions.get(tool_name, ""),
+            category=category,
+            subcategory=subcategory,
+            source="manifest",
+            source_file=f"app/tools/manifests/{tool_name}.yaml",
+            tags=[],
+            enabled=True,
+            commands=commands,
+            stats=ToolStats(),
+        )
+        tools.append(tool_info)
+
+    return tools
+
+
+# ============================================================
+# 路由
+# ============================================================
+
+
+@router.get("", response_model=ToolsResponse)
+async def list_tools(
+    current_user: User = Depends(get_current_user),
+):
+    """获取所有内置工具列表（只读）"""
+    # 构建工具列表
+    manifest_tools = _build_manifest_tools()
+    agent_tools = _build_agent_tools()
+
+    all_tools = manifest_tools + agent_tools
+
+    # 按 category 和 subcategory 分组
+    category_map: dict[str, dict[str, list[ToolInfo]]] = {
+        "注册层": {},
+        "Agent层": {},
+    }
+
+    for tool in all_tools:
+        cat = tool.category
+        subcat = tool.subcategory
+        if cat not in category_map:
+            category_map[cat] = {}
+        if subcat not in category_map[cat]:
+            category_map[cat][subcat] = []
+        category_map[cat][subcat].append(tool)
+
+    # 构建响应
+    categories = []
+    for cat_name, subgroups_dict in category_map.items():
+        if not subgroups_dict:
+            continue
+        subgroups = []
+        for subcat_name, tools_list in subgroups_dict.items():
+            subgroups.append(
+                ToolSubgroup(
+                    name=subcat_name,
+                    display_name=subcat_name,
+                    tools=tools_list,
+                )
+            )
+        categories.append(
+            ToolCategory(
+                name=cat_name,
+                display_name=cat_name,
+                subgroups=subgroups,
+            )
+        )
+
+    # 计算摘要
+    total_commands = sum(len(t.commands) for t in all_tools)
+    active_commands = sum(len(t.commands) for t in all_tools if t.enabled)
+
+    summary = ToolSummary(
+        total_commands=total_commands,
+        active_commands=active_commands,
+        total_tools=len(all_tools),
+        manifest_tools=len(manifest_tools),
+        agent_tools=len(agent_tools),
+    )
+
+    return ToolsResponse(categories=categories, summary=summary)

backend/app/schemas/conversation.py

@@ -1,5 +1,7 @@
-from pydantic import BaseModel
 from datetime import datetime
+from typing import Literal
+
+from pydantic import BaseModel
 
 
 class MessageCreate(BaseModel):
@@ -37,6 +39,7 @@ class ChatRequest(BaseModel):
     conversation_id: str | None = None
     agent_id: str | None = None
     model_name: str | None = None
+    runtime: Literal["jarvis", "hermes"] | None = None
     file_ids: list[str] = []
 
 

backend/app/schemas/remote_mount.py

@@ -0,0 +1,58 @@
+from datetime import datetime
+from pydantic import BaseModel, Field, HttpUrl
+
+
+class RemoteMountCreate(BaseModel):
+    name: str = Field(..., min_length=1, max_length=255)
+    base_url: HttpUrl
+    username: str | None = Field(default=None, max_length=255)
+    password: str | None = Field(default=None, max_length=2000)
+    root_path: str = Field(default="/", min_length=1, max_length=1000)
+
+
+class RemoteMountOut(BaseModel):
+    id: str
+    name: str
+    mount_type: str
+    base_url: str
+    username: str | None
+    root_path: str
+    is_active: bool
+    last_sync_at: str | None
+    created_at: datetime
+    updated_at: datetime
+
+    model_config = {"from_attributes": True}
+
+
+class RemoteNodeOut(BaseModel):
+    path: str
+    name: str
+    is_dir: bool
+    size: int | None = None
+    modified_at: str | None = None
+    etag: str | None = None
+    children: list["RemoteNodeOut"] = []
+
+
+class RemoteMountTreeOut(BaseModel):
+    mount_id: str
+    root_path: str
+    nodes: list[RemoteNodeOut]
+
+
+class RemoteSyncRequest(BaseModel):
+    remote_path: str = Field(..., min_length=1, max_length=2000)
+    local_folder_id: str = Field(..., min_length=1, max_length=36)
+    mode: str = Field(default="file", pattern="^(file|folder)$")
+
+
+class RemoteSyncResultOut(BaseModel):
+    synced: int
+    skipped: int
+    failed: int
+    document_ids: list[str]
+    errors: list[str]
+
+
+RemoteNodeOut.model_rebuild()

backend/app/schemas/schedule_center.py

@@ -1,7 +1,8 @@
 from datetime import datetime
 
-from pydantic import BaseModel
+from pydantic import BaseModel, Field
 
+from app.models.task import TaskDispatchStatus, TaskPriority, TaskQuadrant, TaskStatus
 from app.schemas.goal import GoalOut
 from app.schemas.reminder import ReminderOut
 from app.schemas.task import TaskOut
@@ -18,6 +19,47 @@ class ScheduleCenterDaySummary(BaseModel):
     goal_total: int
 
 
+class ScheduleCenterFocusTaskOut(BaseModel):
+    id: str
+    title: str
+    status: TaskStatus
+    priority: TaskPriority
+    quadrant: TaskQuadrant | None = None
+    assignee_type: str | None = None
+    assignee_id: str | None = None
+    dispatch_status: TaskDispatchStatus
+    due_date: datetime | None = None
+
+
+class ScheduleCenterQuadrantTaskOut(BaseModel):
+    id: str
+    title: str
+    status: TaskStatus
+    priority: TaskPriority
+    dispatch_status: TaskDispatchStatus
+    assignee_type: str | None = None
+    assignee_id: str | None = None
+
+
+class ScheduleCenterQuadrantOut(BaseModel):
+    id: TaskQuadrant
+    title: str
+    subtitle: str
+    color: str
+    glow_color: str
+    icon: str
+    tasks: list[ScheduleCenterQuadrantTaskOut] = Field(default_factory=list)
+
+
+class ScheduleCenterCommanderSummaryOut(BaseModel):
+    total: int = 0
+    queued: int = 0
+    running: int = 0
+    completed: int = 0
+    failed: int = 0
+    overall_status: str | None = None
+
+
 class ScheduleCenterMonthOut(BaseModel):
     month: str
     days: list[ScheduleCenterDaySummary]
@@ -30,4 +72,9 @@ class ScheduleCenterDateOut(BaseModel):
     reminders: list[ReminderOut]
     goals: list[GoalOut]
     summary: ScheduleCenterDaySummary
+    focus_tasks: list[ScheduleCenterFocusTaskOut] = Field(default_factory=list)
+    quadrants: list[ScheduleCenterQuadrantOut] = Field(default_factory=list)
+    commander_summary: ScheduleCenterCommanderSummaryOut = Field(
+        default_factory=ScheduleCenterCommanderSummaryOut,
+    )
     generated_at: datetime

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/schemas/task.py

@@ -1,14 +1,146 @@
-from pydantic import BaseModel
+from __future__ import annotations
+
+import json
 from datetime import datetime
-from app.models.task import TaskStatus, TaskPriority
+
+from pydantic import BaseModel, Field
+from sqlalchemy import inspect
+from sqlalchemy.orm.attributes import NO_VALUE
+
+from app.models.task import (
+    Task,
+    TaskAssigneeType,
+    TaskDispatchStatus,
+    TaskHistory,
+    TaskPriority,
+    TaskQuadrant,
+    TaskSource,
+    TaskStatus,
+    TaskSubTask,
+)
+
+
+def _coerce_enum(value, enum_cls, default=None):
+    if value is None:
+        return default
+    if isinstance(value, enum_cls):
+        return value
+    if isinstance(value, str):
+        raw = value.strip()
+        if not raw:
+            return default
+        for item in enum_cls:
+            if raw == item.value or raw.lower() == item.value:
+                return item
+            if raw.upper() == item.name:
+                return item
+    return default
+
+
+def parse_tags(raw_tags: str | None) -> list[str]:
+    if not raw_tags:
+        return []
+    try:
+        parsed = json.loads(raw_tags)
+    except json.JSONDecodeError:
+        return []
+    if not isinstance(parsed, list):
+        return []
+    return [str(item) for item in parsed]
+
+
+def serialize_tags(tags: list[str] | None) -> str | None:
+    if not tags:
+        return None
+    return json.dumps([str(item) for item in tags], ensure_ascii=False)
+
+
+class TaskSubTaskCreate(BaseModel):
+    title: str
+    description: str | None = None
+    status: TaskStatus = TaskStatus.TODO
+    order_index: int | None = None
+    assignee_type: TaskAssigneeType | None = None
+    assignee_id: str | None = None
+
+
+class TaskSubTaskUpdate(BaseModel):
+    title: str | None = None
+    description: str | None = None
+    status: TaskStatus | None = None
+    order_index: int | None = None
+    assignee_type: TaskAssigneeType | None = None
+    assignee_id: str | None = None
+    dispatch_status: TaskDispatchStatus | None = None
+    dispatch_run_id: str | None = None
+    result_summary: str | None = None
+
+
+class TaskSubTaskReorderItem(BaseModel):
+    id: str
+    order_index: int
+
+
+class TaskSubTaskReorderRequest(BaseModel):
+    items: list[TaskSubTaskReorderItem] = Field(default_factory=list)
+
+
+class TaskSubTaskOut(BaseModel):
+    id: str
+    task_id: str
+    title: str
+    description: str | None
+    status: TaskStatus
+    order_index: int
+    assignee_type: TaskAssigneeType | None
+    assignee_id: str | None
+    dispatch_status: TaskDispatchStatus
+    dispatch_run_id: str | None
+    result_summary: str | None = None
+    completed_at: datetime | None
+    created_at: datetime
+    updated_at: datetime
+
+    model_config = {"from_attributes": True}
+
+
+class TaskHistoryOut(BaseModel):
+    id: str
+    task_id: str
+    action: str
+    old_value: str | None
+    new_value: str | None
+    created_at: datetime
+    updated_at: datetime
+
+    model_config = {"from_attributes": True}
+
+
+class TaskDispatchSummary(BaseModel):
+    status: TaskDispatchStatus
+    run_id: str | None = None
+    result_summary: str | None = None
+    started_at: datetime | None = None
+    last_synced_at: datetime | None = None
+    total_subtasks: int = 0
+    dispatched_subtasks: int = 0
+    subtask_dispatch_statuses: dict[str, int] = Field(default_factory=dict)
 
 
 class TaskCreate(BaseModel):
     title: str
     description: str | None = None
+    status: TaskStatus = TaskStatus.TODO
     priority: TaskPriority = TaskPriority.MEDIUM
     due_date: datetime | None = None
     tags: list[str] | None = None
+    source: TaskSource = TaskSource.MANUAL
+    conversation_id: str | None = None
+    quadrant: TaskQuadrant | None = None
+    assignee_type: TaskAssigneeType | None = None
+    assignee_id: str | None = None
+    subtasks: list[TaskSubTaskCreate] = Field(default_factory=list)
+    dispatch_to_commander: bool = False
 
 
 class TaskUpdate(BaseModel):
@@ -18,6 +150,16 @@ class TaskUpdate(BaseModel):
     priority: TaskPriority | None = None
     due_date: datetime | None = None
     tags: list[str] | None = None
+    source: TaskSource | None = None
+    conversation_id: str | None = None
+    quadrant: TaskQuadrant | None = None
+    assignee_type: TaskAssigneeType | None = None
+    assignee_id: str | None = None
+    dispatch_status: TaskDispatchStatus | None = None
+    dispatch_run_id: str | None = None
+    result_summary: str | None = None
+    started_at: datetime | None = None
+    last_synced_at: datetime | None = None
 
 
 class TaskOut(BaseModel):
@@ -28,12 +170,128 @@ class TaskOut(BaseModel):
     priority: TaskPriority
     due_date: datetime | None
     completed_at: datetime | None
-    tags: str | None
+    tags: list[str] = Field(default_factory=list)
+    source: TaskSource
+    conversation_id: str | None
+    quadrant: TaskQuadrant | None
+    assignee_type: TaskAssigneeType | None
+    assignee_id: str | None
+    dispatch_status: TaskDispatchStatus
+    dispatch_run_id: str | None
+    result_summary: str | None
+    started_at: datetime | None
+    last_synced_at: datetime | None
+    subtask_count: int = 0
     created_at: datetime
     updated_at: datetime
 
-    model_config = {"from_attributes": True}
+
+class TaskDetailOut(TaskOut):
+    subtasks: list[TaskSubTaskOut] = Field(default_factory=list)
+    history: list[TaskHistoryOut] = Field(default_factory=list)
+    dispatch: TaskDispatchSummary
+    dispatch_summary: TaskDispatchSummary
+
+
+
+
+class TaskDispatchRequest(BaseModel):
+    target: str = "commander"
+    conversation_id: str | None = None
+    assignee_type: TaskAssigneeType | None = None
+    assignee_id: str | None = None
+
+
+class TaskDispatchResponse(BaseModel):
+    status: TaskDispatchStatus
+    run_id: str | None = None
+    task: TaskDetailOut
+    payload: dict[str, object] = Field(default_factory=dict)
 
 
 class DailyPlanRequest(BaseModel):
     user_id: str
+
+
+def build_task_out(task: Task) -> TaskOut:
+    subtasks_attr = inspect(task).attrs.subtasks.loaded_value
+    return TaskOut(
+        id=task.id,
+        title=task.title,
+        description=task.description,
+        status=_coerce_enum(task.status, TaskStatus, TaskStatus.TODO),
+        priority=_coerce_enum(task.priority, TaskPriority, TaskPriority.MEDIUM),
+        due_date=task.due_date,
+        completed_at=task.completed_at,
+        tags=parse_tags(task.tags),
+        source=_coerce_enum(task.source, TaskSource, TaskSource.MANUAL),
+        conversation_id=task.conversation_id,
+        quadrant=_coerce_enum(task.quadrant, TaskQuadrant, None),
+        assignee_type=_coerce_enum(task.assignee_type, TaskAssigneeType, None),
+        assignee_id=task.assignee_id,
+        dispatch_status=_coerce_enum(task.dispatch_status, TaskDispatchStatus, TaskDispatchStatus.IDLE),
+        dispatch_run_id=task.dispatch_run_id,
+        result_summary=task.result_summary,
+        started_at=task.started_at,
+        last_synced_at=task.last_synced_at,
+        subtask_count=0 if subtasks_attr is NO_VALUE else len(subtasks_attr or []),
+        created_at=task.created_at,
+        updated_at=task.updated_at,
+    )
+
+
+def build_task_detail_out(task: Task) -> TaskDetailOut:
+    normalized_task_dispatch = _coerce_enum(task.dispatch_status, TaskDispatchStatus, TaskDispatchStatus.IDLE)
+
+    normalized_subtasks = [
+        TaskSubTaskOut(
+            id=item.id,
+            task_id=item.task_id,
+            title=item.title,
+            description=item.description,
+            status=_coerce_enum(item.status, TaskStatus, TaskStatus.TODO),
+            order_index=item.order_index,
+            assignee_type=_coerce_enum(item.assignee_type, TaskAssigneeType, None),
+            assignee_id=item.assignee_id,
+            dispatch_status=_coerce_enum(item.dispatch_status, TaskDispatchStatus, TaskDispatchStatus.IDLE),
+            dispatch_run_id=item.dispatch_run_id,
+            result_summary=item.result_summary,
+            completed_at=item.completed_at,
+            created_at=item.created_at,
+            updated_at=item.updated_at,
+        )
+        for item in task.subtasks
+    ]
+
+    subtask_dispatch_statuses: dict[str, int] = {}
+    for item in normalized_subtasks:
+        key = item.dispatch_status.value
+        subtask_dispatch_statuses[key] = subtask_dispatch_statuses.get(key, 0) + 1
+
+    dispatched_subtasks = sum(1 for item in normalized_subtasks if item.dispatch_status != TaskDispatchStatus.IDLE)
+
+    return TaskDetailOut(
+        **build_task_out(task).model_dump(),
+        subtasks=normalized_subtasks,
+        history=[TaskHistoryOut.model_validate(item) for item in task.history],
+        dispatch=TaskDispatchSummary(
+            status=normalized_task_dispatch,
+            run_id=task.dispatch_run_id,
+            result_summary=task.result_summary,
+            started_at=task.started_at,
+            last_synced_at=task.last_synced_at,
+            total_subtasks=len(normalized_subtasks),
+            dispatched_subtasks=dispatched_subtasks,
+            subtask_dispatch_statuses=subtask_dispatch_statuses,
+        ),
+        dispatch_summary=TaskDispatchSummary(
+            status=normalized_task_dispatch,
+            run_id=task.dispatch_run_id,
+            result_summary=task.result_summary,
+            started_at=task.started_at,
+            last_synced_at=task.last_synced_at,
+            total_subtasks=len(normalized_subtasks),
+            dispatched_subtasks=dispatched_subtasks,
+            subtask_dispatch_statuses=subtask_dispatch_statuses,
+        ),
+    )

backend/app/schemas/tools.py

@@ -0,0 +1,76 @@
+"""Tools API Schemas"""
+
+from pydantic import BaseModel
+from typing import Optional
+
+
+class ToolCommand(BaseModel):
+    """单个工具命令"""
+
+    name: str
+    description: str
+    parameters: dict = {}
+
+
+class ToolStats(BaseModel):
+    """工具调用统计"""
+
+    call_count: int = 0
+    error_count: int = 0
+    total_duration_ms: int = 0
+    avg_duration_ms: int = 0
+    error_rate: float = 0.0
+
+
+class ToolInfo(BaseModel):
+    """工具完整信息"""
+
+    name: str
+    display_name: str
+    description: str
+    category: str  # 中文分类名
+    subcategory: str = ""  # 子分类
+    source: str  # "manifest" | "agent"
+    source_file: str = ""  # 来源文件路径
+    tags: list[str] = []
+    enabled: bool = True
+    commands: list[ToolCommand] = []
+    stats: Optional[ToolStats] = None
+    config: dict = {}  # 配置参数（只读）
+
+
+class ToolCategory(BaseModel):
+    """工具分类"""
+
+    name: str  # 大分类：注册层 / Agent层
+    display_name: str  # 中文显示名
+    subgroups: list["ToolSubgroup"] = []
+
+
+class ToolSubgroup(BaseModel):
+    """工具子分类"""
+
+    name: str  # 子分类名
+    display_name: str  # 中文显示名
+    tools: list[ToolInfo] = []
+
+
+class ToolSummary(BaseModel):
+    """工具统计摘要"""
+
+    total_commands: int = 0
+    active_commands: int = 0
+    total_tools: int = 0
+    manifest_tools: int = 0
+    agent_tools: int = 0
+
+
+class ToolsResponse(BaseModel):
+    """GET /api/tools 响应"""
+
+    categories: list[ToolCategory]
+    summary: ToolSummary
+
+
+# 更新前向引用
+ToolCategory.model_rebuild()

backend/app/services/agent_runtime/__init__.py

@@ -0,0 +1,9 @@
+from app.services.agent_runtime.hermes_runtime import HermesRuntimeAdapter, hermes_runtime_adapter
+from app.services.agent_runtime.jarvis_runtime import JarvisRuntimeAdapter, jarvis_runtime_adapter
+
+__all__ = [
+    "HermesRuntimeAdapter",
+    "hermes_runtime_adapter",
+    "JarvisRuntimeAdapter",
+    "jarvis_runtime_adapter",
+]

backend/app/services/agent_runtime/base.py

@@ -0,0 +1,37 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any, AsyncGenerator, Protocol
+
+from app.models.conversation import Conversation, Message
+from app.models.user import User
+
+
+RuntimeName = str
+
+
+@dataclass(slots=True)
+class RuntimePreparedContext:
+    user: User
+    conversation: Conversation
+    user_message: Message
+    assistant_message: Message
+    raw_message: str
+    full_message: str
+    file_ids: list[str]
+    model_name: str | None
+    memory_context: str | None
+
+
+class ChatRuntime(Protocol):
+    name: RuntimeName
+
+    async def chat_stream(
+        self,
+        prepared: RuntimePreparedContext,
+    ) -> AsyncGenerator[dict[str, Any], None]: ...
+
+    async def chat_once(
+        self,
+        prepared: RuntimePreparedContext,
+    ) -> tuple[str, str | None]: ...

backend/app/services/agent_runtime/hermes_runtime.py

@@ -0,0 +1,172 @@
+from __future__ import annotations
+
+import asyncio
+import importlib.util
+import sys
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import Any, AsyncGenerator
+
+from app.services.agent_runtime.base import ChatRuntime, RuntimePreparedContext
+from app.services.agent_runtime.hermes_session_manager import hermes_session_manager
+
+
+class HermesRuntimeAdapter(ChatRuntime):
+    name = "hermes"
+
+    def __init__(self) -> None:
+        self._repo_path = Path(__file__).resolve().parents[4] / ".tmp" / "hermes-agent"
+        self._agent_class = None
+
+    def probe(self) -> dict[str, Any]:
+        cli_path = self._repo_path / "cli.py"
+        run_agent_path = self._repo_path / "run_agent.py"
+        return {
+            "repo_path": str(self._repo_path),
+            "repo_exists": self._repo_path.exists(),
+            "cli_exists": cli_path.exists(),
+            "run_agent_exists": run_agent_path.exists(),
+            "supports_single_query": True,
+            "supports_resume": True,
+            "integration_mode": "python_ai_agent_bridge",
+        }
+
+    def _load_agent_class(self):
+        if self._agent_class is not None:
+            return self._agent_class
+
+        run_agent_path = self._repo_path / "run_agent.py"
+        if not run_agent_path.exists():
+            raise RuntimeError(f"Hermes run_agent.py 未找到: {run_agent_path}")
+
+        repo_path = str(self._repo_path)
+        if repo_path not in sys.path:
+            sys.path.insert(0, repo_path)
+
+        spec = importlib.util.spec_from_file_location("jarvis_hermes_run_agent", run_agent_path)
+        if spec is None or spec.loader is None:
+            raise RuntimeError("无法加载 Hermes run_agent 模块")
+
+        module = importlib.util.module_from_spec(spec)
+        spec.loader.exec_module(module)
+        self._agent_class = getattr(module, "AIAgent")
+        return self._agent_class
+
+    def _build_agent(self, prepared: RuntimePreparedContext, session_id: str):
+        agent_class = self._load_agent_class()
+        kwargs: dict[str, Any] = {
+            "session_id": session_id,
+            "platform": "jarvis",
+            "user_id": prepared.user.id,
+            "quiet_mode": True,
+            "persist_session": True,
+            "skip_context_files": True,
+            "max_iterations": 30,
+        }
+        if prepared.model_name:
+            kwargs["model"] = prepared.model_name
+        return agent_class(**kwargs)
+
+    def _build_system_message(self, prepared: RuntimePreparedContext) -> str:
+        parts = [
+            "You are Hermes running inside the Jarvis chat runtime.",
+            "Return normal assistant text for the user. Do not mention internal bridge details unless asked.",
+        ]
+        if prepared.memory_context:
+            parts.append(prepared.memory_context)
+        return "\n\n".join(parts)
+
+    async def chat_stream(
+        self,
+        prepared: RuntimePreparedContext,
+    ) -> AsyncGenerator[dict[str, Any], None]:
+        handle = hermes_session_manager.get_or_create(
+            conversation_id=prepared.conversation.id,
+            user_id=prepared.user.id,
+        )
+        async with handle.lock:
+            yield {
+                "type": "progress",
+                "stage": "planning",
+                "label": "Hermes 正在准备会话",
+                "agent": "hermes",
+                "step": "加载 Hermes runtime",
+                "steps": [
+                    "恢复会话上下文",
+                    "调用 Hermes AIAgent",
+                    "回传流式回复",
+                ],
+            }
+
+            queue: asyncio.Queue[dict[str, Any] | None] = asyncio.Queue()
+            loop = asyncio.get_running_loop()
+            result_box: dict[str, Any] = {"content": None, "error": None, "model": prepared.model_name or "hermes"}
+
+            def stream_callback(delta: str) -> None:
+                loop.call_soon_threadsafe(queue.put_nowait, {"type": "chunk", "content": delta})
+
+            def run_sync() -> None:
+                try:
+                    agent = self._build_agent(prepared, handle.hermes_session_id)
+                    result = agent.run_conversation(
+                        prepared.full_message,
+                        system_message=self._build_system_message(prepared),
+                        stream_callback=stream_callback,
+                    )
+                    result_box["content"] = str(result.get("final_response") or "")
+                    result_box["model"] = getattr(agent, "model", prepared.model_name or "hermes")
+                except Exception as exc:  # pragma: no cover - surfaced through queue
+                    result_box["error"] = f"Hermes 执行失败: {exc}"
+                    loop.call_soon_threadsafe(
+                        queue.put_nowait,
+                        {"type": "error", "error": result_box["error"]},
+                    )
+                finally:
+                    loop.call_soon_threadsafe(queue.put_nowait, None)
+
+            worker = asyncio.create_task(asyncio.to_thread(run_sync))
+            streamed_text = ""
+            while True:
+                event = await queue.get()
+                if event is None:
+                    break
+                if event.get("type") == "chunk":
+                    streamed_text += str(event.get("content", ""))
+                yield event
+
+            await worker
+            handle.last_used_at = datetime.now(UTC)
+            handle.metadata = {
+                "session_id": handle.hermes_session_id,
+                "model": result_box["model"],
+                "last_error": result_box["error"],
+            }
+
+            final_text = result_box["content"] or streamed_text
+            if final_text and final_text != streamed_text:
+                yield {"type": "chunk", "content": final_text}
+
+    async def chat_once(self, prepared: RuntimePreparedContext) -> tuple[str, str | None]:
+        handle = hermes_session_manager.get_or_create(
+            conversation_id=prepared.conversation.id,
+            user_id=prepared.user.id,
+        )
+
+        async with handle.lock:
+            agent = await asyncio.to_thread(self._build_agent, prepared, handle.hermes_session_id)
+            result = await asyncio.to_thread(
+                agent.run_conversation,
+                prepared.full_message,
+                self._build_system_message(prepared),
+            )
+            handle.last_used_at = datetime.now(UTC)
+            resolved_model = getattr(agent, "model", prepared.model_name or "hermes")
+            handle.metadata = {
+                "session_id": handle.hermes_session_id,
+                "model": resolved_model,
+                "last_error": None,
+            }
+            return str(result.get("final_response") or ""), resolved_model
+
+
+hermes_runtime_adapter = HermesRuntimeAdapter()

backend/app/services/agent_runtime/hermes_session_manager.py

@@ -0,0 +1,37 @@
+from __future__ import annotations
+
+import asyncio
+from dataclasses import dataclass, field
+from datetime import UTC, datetime
+from typing import Any
+
+
+@dataclass(slots=True)
+class HermesSessionHandle:
+    conversation_id: str
+    user_id: str
+    hermes_session_id: str
+    last_used_at: datetime = field(default_factory=lambda: datetime.now(UTC))
+    restart_count: int = 0
+    lock: asyncio.Lock = field(default_factory=asyncio.Lock)
+    metadata: dict[str, Any] = field(default_factory=dict)
+
+
+class HermesSessionManager:
+    def __init__(self) -> None:
+        self._sessions: dict[str, HermesSessionHandle] = {}
+
+    def get_or_create(self, *, conversation_id: str, user_id: str) -> HermesSessionHandle:
+        handle = self._sessions.get(conversation_id)
+        if handle is None:
+            handle = HermesSessionHandle(
+                conversation_id=conversation_id,
+                user_id=user_id,
+                hermes_session_id=f"jarvis-{conversation_id}",
+            )
+            self._sessions[conversation_id] = handle
+        handle.last_used_at = datetime.now(UTC)
+        return handle
+
+
+hermes_session_manager = HermesSessionManager()

backend/app/services/agent_runtime/jarvis_runtime.py

@@ -0,0 +1,21 @@
+from __future__ import annotations
+
+from typing import Any, AsyncGenerator
+
+from app.services.agent_runtime.base import ChatRuntime, RuntimePreparedContext
+
+
+class JarvisRuntimeAdapter(ChatRuntime):
+    name = "jarvis"
+
+    async def chat_stream(
+        self,
+        prepared: RuntimePreparedContext,
+    ) -> AsyncGenerator[dict[str, Any], None]:
+        raise NotImplementedError("Jarvis runtime is executed inside AgentService")
+
+    async def chat_once(self, prepared: RuntimePreparedContext) -> tuple[str, str | None]:
+        raise NotImplementedError("Jarvis runtime is executed inside AgentService")
+
+
+jarvis_runtime_adapter = JarvisRuntimeAdapter()

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
@@ -21,12 +22,29 @@ 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
+from app.services.agent_runtime.base import RuntimePreparedContext
+from app.services.agent_runtime.hermes_runtime import hermes_runtime_adapter
+from app.services.agent_runtime.hermes_session_manager import hermes_session_manager
 
 logger = logging.getLogger(__name__)
 
@@ -36,6 +54,7 @@ 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 +100,41 @@ 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()
@@ -327,10 +381,47 @@ class AgentService:
     def __init__(self, db: AsyncSession):
         self.db = db
 
+    def _resolve_runtime(self, runtime: str | None) -> str:
+        return runtime or "jarvis"
+
     async def _try_auto_summarize_background(self, user_id: str, conversation_id: str) -> None:
         async with async_session() as session:
             await memory_service.try_auto_summarize(session, user_id, conversation_id)
 
+    # ———— M.4: 主动记忆提取 ————
+    async def _extract_memories_background(self, user_id: str, conversation_id: str) -> None:
+        """Background task to extract memories from conversation after response."""
+        from app.services.memory.memory_extractor import MemoryExtractor
+        from sqlalchemy import select
+        from app.models.conversation import Message
+
+        try:
+            async with async_session() as db:
+                # Load last 10 messages from conversation
+                result = await db.execute(
+                    select(Message)
+                    .where(Message.conversation_id == conversation_id)
+                    .order_by(Message.created_at.desc())
+                    .limit(10)
+                )
+                messages = list(result.scalars().all())
+
+                if len(messages) < 2:
+                    return
+
+                extractor = MemoryExtractor()
+                new_memories = await extractor.extract_from_conversation(
+                    db, user_id, conversation_id, messages
+                )
+
+                if new_memories:
+                    await extractor.save_memories(db, user_id, conversation_id, new_memories)
+                    logger.info(
+                        f"[MemoryExtractor] Extracted {len(new_memories)} new memories from conversation {conversation_id}"
+                    )
+        except Exception as e:
+            logger.exception(f"[MemoryExtractor] Extraction failed: {e}")
+
     def _build_progress_event(
         self,
         stage: str,
@@ -427,18 +518,27 @@ 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,
@@ -448,9 +548,119 @@ 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,
@@ -458,10 +668,12 @@ class AgentService:
         conversation_id: str | None = None,
         file_ids: list[str] | None = None,
         model_name: str | None = None,
+        runtime: str | None = None,
     ) -> tuple[str, str, AsyncGenerator[dict[str, Any], None]]:
         """
         处理对话请求（流式）
         """
+        runtime_name = self._resolve_runtime(runtime)
         user_llm_config = await self._get_user_llm_config(user_id, model_name)
         model_name_used = model_name
         if model_name and not user_llm_config:
@@ -543,11 +755,18 @@ class AgentService:
             self.db, user_id, conversation_id, message
         )
 
+        # M.5: Inject recall memories into context (before LLM call)
+        from app.services.memory.recall_injector import MemoryRecallInjector
+
+        recall_ctx = await MemoryRecallInjector().build_context(self.db, user_id, message)
+        if recall_ctx:
+            memory_ctx = f"{memory_ctx}\n{recall_ctx}" if memory_ctx else recall_ctx
+
         assistant_msg = Message(
             conversation_id=conversation_id,
             role="assistant",
             content="",
-            model=model_name_used or "jarvis",
+            model=(model_name_used or "jarvis") if runtime_name == "jarvis" else runtime_name,
             attachments=None,
         )
         self.db.add(assistant_msg)
@@ -562,28 +781,113 @@ class AgentService:
                 "title": "Assistant message",
                 "content_summary": content[:500],
                 "raw_excerpt": content[:2000],
-                "metadata_": {"role": "assistant"},
+                "metadata_": {"role": "assistant", "runtime": runtime_name},
                 "importance_signal": 0.8,
             }
 
+        if runtime_name == "hermes":
+            user = await self.db.get(User, user_id)
+            if user is None:
+                raise ValueError("用户不存在")
+
+            prepared = RuntimePreparedContext(
+                user=user,
+                conversation=conv,
+                user_message=user_msg,
+                assistant_message=assistant_msg,
+                raw_message=message,
+                full_message=full_message,
+                file_ids=file_ids or [],
+                model_name=model_name_used,
+                memory_context=memory_ctx,
+            )
+
+            async def run_hermes():
+                collected = ""
+                stream_failed = False
+                try:
+                    async for event in hermes_runtime_adapter.chat_stream(prepared):
+                        if event.get("type") == "chunk":
+                            collected += str(event.get("content", ""))
+                        elif event.get("type") == "error":
+                            stream_failed = True
+                        yield event
+                finally:
+                    try:
+                        session_handle = hermes_session_manager.get_or_create(
+                            conversation_id=conv.id,
+                            user_id=user_id,
+                        )
+                        assistant_msg.content = collected if collected else ("Hermes 执行失败，请检查运行配置。" if stream_failed else "")
+                        assistant_msg.model = str(session_handle.metadata.get("model") or "hermes")
+                        assistant_msg.attachments = [
+                            {
+                                "kind": "runtime_info",
+                                "runtime": "hermes",
+                                "session_id": session_handle.hermes_session_id,
+                                "model": session_handle.metadata.get("model"),
+                                "last_error": session_handle.metadata.get("last_error"),
+                            }
+                        ]
+                        conv.agent_state = {
+                            "runtime": "hermes",
+                            "runtime_state": {
+                                "hermes": {
+                                    "session_id": session_handle.hermes_session_id,
+                                    "message_id": assistant_msg.id,
+                                    "model": session_handle.metadata.get("model"),
+                                    "last_error": session_handle.metadata.get("last_error"),
+                                }
+                            },
+                        }
+                        await BrainService(self.db).create_event(
+                            user_id,
+                            **_build_assistant_event_payload(assistant_msg.content),
+                        )
+                        await self.db.commit()
+                        await self.db.refresh(assistant_msg)
+                    except Exception:
+                        logger.exception("save_hermes_assistant_message_failed")
+                    asyncio.create_task(self._try_auto_summarize_background(user_id, conversation_id))
+                    asyncio.create_task(self._extract_memories_background(user_id, conversation_id))
+
+            return conversation_id, assistant_msg.id, run_hermes()
+
         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))
 
@@ -708,7 +1012,7 @@ class AgentService:
                     if collected:
                         assistant_msg.content = collected
                         continuity_snapshot = _build_continuity_snapshot(state or {})
-                        assistant_msg.attachments = (
+                        attachments = (
                             [
                                 {
                                     "kind": "agent_continuity_state",
@@ -716,7 +1020,25 @@ 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 = (
                             {
@@ -730,11 +1052,23 @@ 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))
+                # M.4: Extract memories from conversation
+                asyncio.create_task(self._extract_memories_background(user_id, conversation_id))
 
         return conversation_id, assistant_msg.id, run_agent()
 
@@ -745,10 +1079,12 @@ class AgentService:
         conversation_id: str | None = None,
         file_ids: list[str] | None = None,
         model_name: str | None = None,
+        runtime: str | None = None,
     ) -> tuple[str, str, str, str | None]:
         """
         简单同步版对话
         """
+        runtime_name = self._resolve_runtime(runtime)
         user_llm_config = await self._get_user_llm_config(user_id, model_name)
         model_name_used = model_name
         if model_name and not user_llm_config:
@@ -785,7 +1121,7 @@ class AgentService:
             conversation_id=conversation_id,
             role="assistant",
             content="",
-            model=model_name_used or "jarvis",
+            model=(model_name_used or "jarvis") if runtime_name == "jarvis" else runtime_name,
             attachments=None,
         )
         self.db.add(assistant_msg)
@@ -807,20 +1143,107 @@ class AgentService:
             self.db, user_id, conversation_id, message
         )
 
+        # M.5: Inject recall memories into context (before LLM call)
+        from app.services.memory.recall_injector import MemoryRecallInjector
+
+        recall_ctx = await MemoryRecallInjector().build_context(self.db, user_id, message)
+        if recall_ctx:
+            memory_ctx = f"{memory_ctx}\n{recall_ctx}" if memory_ctx else recall_ctx
+
+        if runtime_name == "hermes":
+            user = await self.db.get(User, user_id)
+            if user is None:
+                raise ValueError("用户不存在")
+            prepared = RuntimePreparedContext(
+                user=user,
+                conversation=conv,
+                user_message=user_msg,
+                assistant_message=assistant_msg,
+                raw_message=message,
+                full_message=message,
+                file_ids=file_ids or [],
+                model_name=model_name_used,
+                memory_context=memory_ctx,
+            )
+            response_content, resolved_model_name = await hermes_runtime_adapter.chat_once(prepared)
+            assistant_msg.content = response_content
+            assistant_msg.model = resolved_model_name or "hermes"
+            assistant_msg.attachments = [{
+                "kind": "runtime_info",
+                "runtime": "hermes",
+                "session_id": hermes_session_manager.get_or_create(
+                    conversation_id=conv.id,
+                    user_id=user_id,
+                ).hermes_session_id,
+                "model": resolved_model_name,
+            }]
+            conv.agent_state = {
+                "runtime": "hermes",
+                "runtime_state": {
+                    "hermes": {
+                        "session_id": hermes_session_manager.get_or_create(
+                            conversation_id=conv.id,
+                            user_id=user_id,
+                        ).hermes_session_id,
+                        "message_id": assistant_msg.id,
+                        "model": resolved_model_name,
+                    }
+                },
+            }
+            await brain_service.create_event(
+                user_id,
+                source_type="conversation",
+                source_id=conversation_id,
+                event_type="message_created",
+                title="Assistant message",
+                content_summary=response_content[:500],
+                raw_excerpt=response_content[:2000],
+                metadata_={"role": "assistant", "runtime": "hermes"},
+                importance_signal=0.8,
+            )
+            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=None,
+            )
+            return conversation_id, assistant_msg.id, response_content, assistant_msg.model
+
         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=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)
@@ -850,7 +1273,7 @@ class AgentService:
         continuity_snapshot = (
             _build_continuity_snapshot(result_state) if "result_state" in locals() else None
         )
-        assistant_msg.attachments = (
+        attachments = (
             [
                 {
                     "kind": "agent_continuity_state",
@@ -858,7 +1281,25 @@ 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 = (
             {
@@ -868,7 +1309,17 @@ 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/document_service.py

@@ -5,6 +5,7 @@
 
 from pathlib import Path
 import tempfile
+import shutil
 from sqlalchemy.ext.asyncio import AsyncSession
 from sqlalchemy import select
 from fastapi import UploadFile
@@ -18,7 +19,6 @@ import json
 import os
 import re
 import aiofiles
-import uuid
 from dataclasses import dataclass, field
 
 
@@ -52,9 +52,9 @@ class DocumentService:
         if ext not in ALLOWED_EXTENSIONS:
             raise ValueError(f"不支持的文件类型: {ext}")
 
-        os.makedirs(settings.UPLOAD_DIR, exist_ok=True)
+        folder_path = await self._get_storage_directory(user_id, folder_id)
-        file_id = str(uuid.uuid4())
+        folder_path.mkdir(parents=True, exist_ok=True)
-        file_path = os.path.join(settings.UPLOAD_DIR, f"{file_id}{ext}")
+        file_path = self._resolve_unique_file_path(folder_path, file.filename)
 
         content = await file.read()
         file_size = len(content)
@@ -64,7 +64,7 @@ class DocumentService:
         async with aiofiles.open(file_path, "wb") as f:
             await f.write(content)
 
-        parsed = await self._parse_document(file_path, ext)
+        parsed = await self._parse_document(str(file_path), ext)
         parsed.structured_markdown = self._render_structured_markdown(parsed)
 
         doc = Document(
@@ -73,7 +73,7 @@ class DocumentService:
             filename=file.filename,
             file_type=ext[1:],
             file_size=file_size,
-            file_path=file_path,
+            file_path=str(file_path),
             summary=parsed.summary[:500] if len(parsed.summary) > 500 else parsed.summary,
             folder_id=folder_id,
             ingestion_status="uploaded",
@@ -171,6 +171,83 @@ class DocumentService:
 
         return "/" + "/".join(path_parts) if path_parts else None
 
+    async def ensure_folder_directory(self, user_id: str, folder_id: str | None) -> Path:
+        folder_path = await self._get_storage_directory(user_id, folder_id)
+        folder_path.mkdir(parents=True, exist_ok=True)
+        return folder_path
+
+    async def delete_folder_directory(self, user_id: str, folder_id: str) -> None:
+        folder_path = await self._get_storage_directory(user_id, folder_id)
+        if folder_path.exists():
+            shutil.rmtree(folder_path, ignore_errors=True)
+
+    async def rename_folder_directory(self, user_id: str, folder_id: str, old_name: str, new_name: str) -> None:
+        folder = await self.db.get(Folder, folder_id)
+        if folder is None:
+            return
+
+        parent_path = await self._get_storage_directory(user_id, folder.parent_id)
+        old_path = parent_path / self._sanitize_storage_name(old_name)
+        new_path = parent_path / self._sanitize_storage_name(new_name)
+
+        if old_path != new_path:
+            parent_path.mkdir(parents=True, exist_ok=True)
+            if old_path.exists():
+                old_path.rename(new_path)
+            else:
+                new_path.mkdir(parents=True, exist_ok=True)
+        else:
+            new_path.mkdir(parents=True, exist_ok=True)
+
+        document_result = await self.db.execute(
+            select(Document).where(Document.user_id == user_id)
+        )
+        for document in document_result.scalars().all():
+            try:
+                relative_path = Path(document.file_path).relative_to(old_path)
+            except ValueError:
+                continue
+            document.file_path = str(new_path / relative_path)
+
+    async def _get_storage_directory(self, user_id: str, folder_id: str | None) -> Path:
+        base_path = Path(settings.UPLOAD_DIR) / user_id
+        if not folder_id:
+            return base_path
+
+        folders = await self.db.execute(
+            select(Folder).where(Folder.user_id == user_id)
+        )
+        folder_map = {folder.id: folder for folder in folders.scalars().all()}
+
+        path_segments: list[str] = []
+        current_id = folder_id
+        while current_id:
+            folder = folder_map.get(current_id)
+            if folder is None:
+                raise ValueError("鐖舵枃浠跺す涓嶅瓨鍦?")
+            path_segments.insert(0, self._sanitize_storage_name(folder.name))
+            current_id = folder.parent_id
+
+        return base_path.joinpath(*path_segments)
+
+    def _resolve_unique_file_path(self, directory: Path, original_name: str) -> Path:
+        safe_name = self._sanitize_storage_name(Path(original_name).name, is_file=True)
+        stem = Path(safe_name).stem
+        suffix = Path(safe_name).suffix
+        candidate = directory / safe_name
+        counter = 2
+        while candidate.exists():
+            candidate = directory / f"{stem}-{counter}{suffix}"
+            counter += 1
+        return candidate
+
+    def _sanitize_storage_name(self, name: str, is_file: bool = False) -> str:
+        invalid_chars = '<>:"/\\|?*'
+        sanitized = ''.join('_' if char in invalid_chars or ord(char) < 32 else char for char in name).strip().rstrip('.')
+        if not sanitized:
+            return 'untitled' if is_file else 'folder'
+        return sanitized
+
     async def delete_document(self, user_id: str, document_id: str):
         result = await self.db.execute(
             select(Document).where(

backend/app/services/memory/__init__.py

@@ -0,0 +1,20 @@
+"""Memory Services Module"""
+
+from app.services.memory.frequency_tracker import FrequencyTracker
+from app.services.memory.emotion_analyzer import EmotionAnalyzer
+from app.services.memory.impact_evaluator import ImpactEvaluator
+from app.services.memory.importance_scorer import ImportanceScorer, ImportanceLevel
+from app.services.memory.forgetting_curve import ForgettingCurve
+from app.services.memory.memory_decay import MemoryDecay
+from app.services.memory.reinforcement import MemoryReinforcement
+
+__all__ = [
+    "FrequencyTracker",
+    "EmotionAnalyzer",
+    "ImpactEvaluator",
+    "ImportanceScorer",
+    "ImportanceLevel",
+    "ForgettingCurve",
+    "MemoryDecay",
+    "MemoryReinforcement",
+]

backend/app/services/memory/daily_digest.py

@@ -0,0 +1,264 @@
+"""
+DailyDigestGenerator
+
+Generates daily summary of user's day including conversations, tasks, key memories.
+Generated at 22:00 daily via scheduler.
+"""
+
+import json
+from dataclasses import dataclass, field
+from datetime import date, datetime, timedelta, UTC
+from typing import Any
+
+from sqlalchemy import select, and_
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.models.memory import MemorySummary, UserMemory
+from app.models.task import Task
+
+
+@dataclass
+class DailyDigestData:
+    """Daily digest data structure."""
+
+    date: date
+    summary: str
+    key_points: list[dict] = field(default_factory=list)
+    pending_questions: list[dict] = field(default_factory=list)
+    suggestions: list[dict] = field(default_factory=list)
+
+
+class DailyDigestGenerator:
+    """Generate daily summary for a user."""
+
+    MAX_KEY_POINTS = 5
+    MAX_SUGGESTIONS = 3
+
+    async def generate(
+        self,
+        db: AsyncSession,
+        user_id: str,
+        target_date: date | None = None,
+    ) -> DailyDigestData:
+        """Generate daily digest for user.
+
+        1. Get today's conversation summaries
+        2. Get high importance memories from today
+        3. Get today's tasks
+        4. Generate summary using LLM
+        """
+        if target_date is None:
+            target_date = datetime.now(UTC).date()
+
+        start_of_day = datetime.combine(target_date, datetime.min.time()).replace(tzinfo=UTC)
+        end_of_day = datetime.combine(target_date, datetime.max.time()).replace(tzinfo=UTC)
+
+        # 1. Get conversation summaries from today
+        summaries = await self._get_today_summaries(db, user_id, start_of_day, end_of_day)
+
+        # 2. Get high importance memories from today
+        high_importance_memories = await self._get_high_importance_memories(
+            db, user_id, start_of_day, end_of_day
+        )
+
+        # 3. Get today's tasks
+        tasks = await self._get_today_tasks(db, user_id, start_of_day, end_of_day)
+
+        # 4. Generate summary using LLM
+        summary = await self._generate_summary(
+            summaries=summaries,
+            memories=high_importance_memories,
+            tasks=tasks,
+        )
+
+        # 5. Extract key points
+        key_points = self._extract_key_points(high_importance_memories, summaries, tasks)
+
+        # 6. Generate suggestions
+        suggestions = self._generate_suggestions(high_importance_memories, tasks)
+
+        return DailyDigestData(
+            date=target_date,
+            summary=summary,
+            key_points=key_points,
+            pending_questions=[],  # Filled by LLM analysis
+            suggestions=suggestions,
+        )
+
+    async def _get_today_summaries(
+        self,
+        db: AsyncSession,
+        user_id: str,
+        start: datetime,
+        end: datetime,
+    ) -> list[MemorySummary]:
+        """Get conversation summaries from today."""
+        result = await db.execute(
+            select(MemorySummary)
+            .where(
+                MemorySummary.user_id == user_id,
+                MemorySummary.summary_at >= start,
+                MemorySummary.summary_at <= end,
+            )
+            .order_by(MemorySummary.summary_at.desc())
+        )
+        return list(result.scalars().all())
+
+    async def _get_high_importance_memories(
+        self,
+        db: AsyncSession,
+        user_id: str,
+        start: datetime,
+        end: datetime,
+    ) -> list[UserMemory]:
+        """Get high importance memories accessed or created today."""
+        result = await db.execute(
+            select(UserMemory)
+            .where(
+                UserMemory.user_id == user_id,
+                UserMemory.importance_level == "high",
+                ((UserMemory.last_accessed_at >= start) | (UserMemory.last_accessed_at.is_(None))),
+            )
+            .order_by(UserMemory.importance_score.desc())
+            .limit(5)
+        )
+        return list(result.scalars().all())
+
+    async def _get_today_tasks(
+        self,
+        db: AsyncSession,
+        user_id: str,
+        start: datetime,
+        end: datetime,
+    ) -> list[Task]:
+        """Get tasks updated today."""
+        result = await db.execute(
+            select(Task)
+            .where(
+                Task.user_id == user_id,
+                Task.updated_at >= start,
+                Task.updated_at <= end,
+            )
+            .order_by(Task.priority.desc())
+        )
+        return list(result.scalars().all())
+
+    async def _generate_summary(
+        self,
+        summaries: list[MemorySummary],
+        memories: list[UserMemory],
+        tasks: list[Task],
+    ) -> str:
+        """Generate daily summary text using LLM."""
+        from app.services.llm_service import get_llm
+        from langchain_core.messages import HumanMessage, SystemMessage
+
+        summary_texts = "\n".join(f"- {s.summary_text}" for s in summaries)
+        memory_texts = "\n".join(f"- [{m.memory_type}] {m.content}" for m in memories[:3])
+        task_texts = "\n".join(f"- [{t.status}] {t.title}" for t in tasks[:5])
+
+        prompt = f"""今天的主要活动摘要：
+
+对话摘要：
+{summary_texts or "无"}
+
+高重要性记忆：
+{memory_texts or "无"}
+
+任务：
+{task_texts or "无"}
+
+请用1-2句话简洁总结今天的核心活动。不要过度发挥。"""
+        try:
+            llm = get_llm()
+            response = await llm.invoke(
+                [
+                    SystemMessage(
+                        content="你是一个记忆助手。请简洁总结用户今天的核心活动，不超过50字。"
+                    ),
+                    HumanMessage(content=prompt),
+                ]
+            )
+            return response.content.strip()
+        except Exception:
+            # Fallback: count activities
+            total = len(summaries) + len(memories) + len(tasks)
+            if total == 0:
+                return "今天没有明显的活动记录。"
+            return f"今天共处理了 {total} 项活动（{len(summaries)} 次对话、{len(memories)} 条重要记忆、{len(tasks)} 个任务）。"
+
+    def _extract_key_points(
+        self,
+        memories: list[UserMemory],
+        summaries: list[MemorySummary],
+        tasks: list[Task],
+    ) -> list[dict]:
+        """Extract key points from today's activities."""
+        key_points = []
+
+        for m in memories[: self.MAX_KEY_POINTS]:
+            key_points.append(
+                {
+                    "content": m.content[:100],
+                    "importance": m.importance_score or 0.5,
+                    "source": "memory",
+                }
+            )
+
+        for t in tasks[:3]:
+            if len(key_points) >= self.MAX_KEY_POINTS:
+                break
+            key_points.append(
+                {
+                    "content": t.title,
+                    "importance": t.priority / 10.0 if t.priority else 0.5,
+                    "source": "task",
+                }
+            )
+
+        key_points.sort(key=lambda x: x["importance"], reverse=True)
+        return key_points[: self.MAX_KEY_POINTS]
+
+    def _generate_suggestions(
+        self,
+        memories: list[UserMemory],
+        tasks: list[Task],
+    ) -> list[dict]:
+        """Generate suggestions based on today's activities."""
+        suggestions = []
+
+        # Suggest following up on high importance memories
+        for m in memories[:2]:
+            if len(suggestions) >= self.MAX_SUGGESTIONS:
+                break
+            suggestions.append(
+                {
+                    "text": f"可以继续聊聊「{m.content[:20]}」相关的话题",
+                    "reason": "这是你关心的高优先级话题",
+                }
+            )
+
+        # Suggest incomplete high-priority tasks
+        incomplete = [t for t in tasks if t.status != "done"]
+        for t in incomplete[:2]:
+            if len(suggestions) >= self.MAX_SUGGESTIONS:
+                break
+            suggestions.append(
+                {
+                    "text": f"继续推进：{t.title}",
+                    "reason": "这是未完成的高优先级任务",
+                }
+            )
+
+        return suggestions[: self.MAX_SUGGESTIONS]
+
+    async def get_recent_digests(
+        self,
+        db: AsyncSession,
+        user_id: str,
+        limit: int = 7,
+    ) -> list[DailyDigestData]:
+        """Get recent digests (stored as JSON in user metadata or separate table)."""
+        # For simplicity, return empty list - digests are stored separately
+        # In production, would query a daily_digests table
+        return []

backend/app/services/memory/emotion_analyzer.py

@@ -0,0 +1,149 @@
+"""
+EmotionAnalyzer
+
+Extracts emotional intensity from text and calculates emotion-based importance scores.
+"""
+
+import re
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from app.models.memory import UserMemory
+
+
+class EmotionAnalyzer:
+    """Analyze emotional keywords in memory content"""
+
+    # Emotion keyword weights (higher = more important)
+    EMOTION_KEYWORDS = {
+        # High intensity
+        "急": 1.0,
+        "紧急": 1.0,
+        "很重要": 0.9,
+        "非常重要": 1.0,
+        "必须": 0.9,
+        # Medium-high intensity
+        "困扰": 0.8,
+        "烦恼": 0.7,
+        "担心": 0.7,
+        "焦虑": 0.8,
+        "害怕": 0.8,
+        "恐惧": 0.9,
+        # Medium intensity
+        "想解决": 0.6,
+        "希望": 0.5,
+        "想要": 0.4,
+        "需要": 0.4,
+        "渴望": 0.6,
+        # Low intensity (casual/neutral)
+        "无所谓": 0.1,
+        "随便": 0.1,
+        "都行": 0.1,
+        "还好": 0.2,
+        # Negative valence
+        "讨厌": 0.6,
+        "不喜欢": 0.5,
+        "恨": 0.8,
+        "不喜欢": 0.5,
+        # Positive valence
+        "喜欢": 0.5,
+        "爱": 0.7,
+        "开心": 0.4,
+        "高兴": 0.4,
+    }
+
+    # Urgency patterns — only match when there's an explicit time-bound word
+    # NOTE: "今天" alone is too common and matches neutral sentences like "今天天气不错"
+    URGENCY_PATTERNS = [
+        (re.compile(r"马上|立刻|立即|赶紧"), 1.0),
+        (re.compile(r"今天内|今天必须|今日必须"), 0.8),
+        (re.compile(r"明天|明天之前|明日"), 0.6),
+        (re.compile(r"这周|本周"), 0.4),
+        (re.compile(r"尽快|早点"), 0.7),
+    ]
+
+    def extract(self, text: str) -> list[str]:
+        """Extract emotion keywords from text
+
+        Returns:
+            List of matched emotion keywords
+        """
+        if not text:
+            return []
+
+        matched = []
+        text_lower = text.lower()
+
+        for keyword, weight in self.EMOTION_KEYWORDS.items():
+            if keyword in text_lower:
+                matched.append(keyword)
+
+        # Check urgency patterns
+        for pattern, weight in self.URGENCY_PATTERNS:
+            if pattern.search(text):
+                matched.append(f"[URGENCY:{weight}]")
+
+        return matched
+
+    def calculate_score(self, memory: "UserMemory") -> float:
+        """Calculate emotion-based importance score (0.0 - 1.0)
+
+        Uses the highest-weighted emotion keyword found in the content.
+        """
+        content = memory.content or ""
+        emotion_tags = memory.emotion_tags or []
+
+        # Check emotion_tags first (pre-extracted)
+        if emotion_tags:
+            max_weight = 0.0
+            for tag in emotion_tags:
+                if tag in self.EMOTION_KEYWORDS:
+                    max_weight = max(max_weight, self.EMOTION_KEYWORDS[tag])
+            if max_weight > 0:
+                return max_weight
+
+        # Extract from content
+        matched = self.extract(content)
+        if not matched:
+            return 0.0
+
+        # Get highest weight
+        max_weight = 0.0
+        for keyword in matched:
+            if keyword.startswith("[URGENCY:"):
+                # Extract urgency weight
+                try:
+                    weight = float(keyword.split(":")[1].rstrip("]"))
+                    max_weight = max(max_weight, weight)
+                except (ValueError, IndexError):
+                    pass
+            elif keyword in self.EMOTION_KEYWORDS:
+                max_weight = max(max_weight, self.EMOTION_KEYWORDS[keyword])
+
+        return min(1.0, max_weight)
+
+    def get_emotion_profile(self, text: str) -> dict:
+        """Get detailed emotion profile for text
+
+        Returns:
+            Dict with matched keywords, max_weight, and sentiment
+        """
+        matched = self.extract(text)
+        weights = []
+        for keyword in matched:
+            if keyword.startswith("[URGENCY:"):
+                try:
+                    weights.append(float(keyword.split(":")[1].rstrip("]")))
+                except (ValueError, IndexError):
+                    pass
+            elif keyword in self.EMOTION_KEYWORDS:
+                weights.append(self.EMOTION_KEYWORDS[keyword])
+
+        return {
+            "matched_keywords": matched,
+            "max_weight": max(weights) if weights else 0.0,
+            "avg_weight": sum(weights) / len(weights) if weights else 0.0,
+            "sentiment": "positive"
+            if (weights and sum(weights) / len(weights) > 0.5)
+            else "neutral",
+        }

backend/app/services/memory/forgetting_curve.py

@@ -0,0 +1,70 @@
+"""
+ForgettingCurve
+
+Calculates memory decay based on Ebbinghaus forgetting curve.
+decay_score = exp(-days_since_access / half_life)
+
+Importance level affects half-life:
+- high: half_life = 30 * 3 = 90 days (slowest decay)
+- medium: half_life = 30 * 1 = 30 days
+- low: half_life = 30 * 0.5 = 15 days (fastest decay)
+"""
+
+import math
+from datetime import UTC, datetime
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from app.models.memory import UserMemory
+
+
+class ForgettingCurve:
+    """Calculate memory decay based on time and importance."""
+
+    BASE_HALF_LIFE_DAYS = 30
+
+    # Half-life multipliers by importance level
+    HALF_LIFE_MULTIPLIERS = {
+        "high": 3.0,
+        "medium": 1.0,
+        "low": 0.5,
+    }
+
+    def calculate_decay(self, memory: "UserMemory") -> float:
+        """Calculate decay score (0.0-1.0). Higher = more remembered.
+
+        Uses exponential decay: exp(-days_since_access / half_life)
+        """
+        last_accessed = getattr(memory, "last_accessed_at", None) or getattr(
+            memory, "last_recalled_at", None
+        )
+        if last_accessed is None:
+            return 1.0  # Never accessed = full retention
+
+        now = datetime.now(UTC)
+        if isinstance(last_accessed, datetime):
+            if last_accessed.tzinfo is None:
+                last_accessed = last_accessed.replace(tzinfo=UTC)
+            days_since = (now - last_accessed).total_seconds() / 86400
+        else:
+            days_since = 0
+
+        half_life = self.get_half_life(memory)
+        decay = math.exp(-days_since / half_life)
+        return min(1.0, max(0.0, decay))
+
+    def get_half_life(self, memory: "UserMemory") -> float:
+        """Get half-life in days based on importance level."""
+        importance_level = getattr(memory, "importance_level", "medium") or "medium"
+        multiplier = self.HALF_LIFE_MULTIPLIERS.get(importance_level, 1.0)
+        return self.BASE_HALF_LIFE_DAYS * multiplier
+
+    def should_archive(self, memory: "UserMemory") -> bool:
+        """decay < 0.2 → memory should be archived (cold storage)."""
+        decay = self.calculate_decay(memory)
+        return decay < 0.2
+
+    def should_deprioritize(self, memory: "UserMemory") -> bool:
+        """decay < 0.5 → memory should be deprioritized (not in active reminders)."""
+        decay = self.calculate_decay(memory)
+        return decay < 0.5

backend/app/services/memory/frequency_tracker.py

@@ -0,0 +1,84 @@
+"""
+FrequencyTracker
+
+Tracks how often a memory is recalled and calculates frequency/recency scores.
+"""
+
+from datetime import UTC, datetime
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from app.models.memory import UserMemory
+
+
+class FrequencyTracker:
+    """Track and score memory recall frequency"""
+
+    # Score weights
+    MAX_FREQUENCY = 10  # Cap frequency count for scoring
+    RECENCY_DECAY_DAYS = 30  # After 30 days, recency score drops significantly
+
+    def increment(self, memory: "UserMemory") -> "UserMemory":
+        """Increment recall count and update last recalled timestamp"""
+        memory.frequency_count = (memory.frequency_count or 0) + 1
+        memory.last_recalled_at = datetime.now(UTC)
+        return memory
+
+    def get_frequency_score(self, memory: "UserMemory") -> float:
+        """Calculate normalized frequency score (0.0 - 1.0)
+
+        Uses logarithmic scaling to prevent high-frequency memories
+        from dominating completely.
+        """
+        count = memory.frequency_count or 0
+        if count == 0:
+            return 0.0
+        # Logarithmic scaling: more recalls have diminishing returns
+        # log(1+x) / log(1+MAX) gives 0-1 range
+        import math
+
+        score = math.log(1 + count) / math.log(1 + self.MAX_FREQUENCY)
+        return min(1.0, max(0.0, score))
+
+    def get_recency_score(self, memory: "UserMemory") -> float:
+        """Calculate recency score (0.0 - 1.0)
+
+        Memory recalled recently scores higher. Uses exponential decay.
+        """
+        last_recalled = memory.last_recalled_at
+        if last_recalled is None:
+            return 0.0
+
+        now = datetime.now(UTC)
+        if isinstance(last_recalled, datetime):
+            if last_recalled.tzinfo is None:
+                last_recalled = last_recalled.replace(tzinfo=UTC)
+            days_since = (now - last_recalled).total_seconds() / 86400
+        else:
+            days_since = self.RECENCY_DECAY_DAYS
+
+        # Exponential decay: half-life of RECENCY_DECAY_DAYS
+        import math
+
+        decay = math.exp(-days_since / self.RECENCY_DECAY_DAYS)
+        return min(1.0, max(0.0, decay))
+
+    def get_time_decay(self, memory: "UserMemory") -> float:
+        """Calculate time-based decay factor for forgetting curve"""
+        last_accessed = getattr(memory, "last_accessed_at", None)
+        if last_accessed is None:
+            last_accessed = memory.last_recalled_at
+        if last_accessed is None:
+            return 1.0
+
+        now = datetime.now(UTC)
+        if isinstance(last_accessed, datetime):
+            if last_accessed.tzinfo is None:
+                last_accessed = last_accessed.replace(tzinfo=UTC)
+            days_since = (now - last_accessed).total_seconds() / 86400
+        else:
+            days_since = 0
+
+        import math
+
+        return math.exp(-days_since / self.RECENCY_DECAY_DAYS)

backend/app/services/memory/impact_evaluator.py

@@ -0,0 +1,52 @@
+"""
+ImpactEvaluator
+
+Evaluates the breadth of impact a memory has based on associated topics.
+"""
+
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from app.models.memory import UserMemory
+
+
+class ImpactEvaluator:
+    """Evaluate the impact breadth of a memory"""
+
+    # Threshold for maximum impact score
+    IMPACT_THRESHOLD = 5  # Number of associated topics for max impact
+
+    def evaluate(self, memory: "UserMemory") -> float:
+        """Calculate impact score (0.0 - 1.0)
+
+        The more associated topics a memory has, the higher its impact.
+        Topics represent "what this memory is about" — if it touches
+        many aspects of the user's life, it has high impact.
+        """
+        associated_topics = memory.associated_topics or []
+        if not associated_topics:
+            return 0.0
+
+        # Normalize: IMPACT_THRESHOLD topics = full impact (1.0)
+        raw_score = len(associated_topics) / self.IMPACT_THRESHOLD
+        return min(1.0, raw_score)
+
+    def get_topic_overlap(self, memory_a: "UserMemory", memory_b: "UserMemory") -> float:
+        """Calculate topic overlap between two memories (0.0 - 1.0)
+
+        Used for finding related memories.
+        """
+        topics_a = set(memory_a.associated_topics or [])
+        topics_b = set(memory_b.associated_topics or [])
+
+        if not topics_a or not topics_b:
+            return 0.0
+
+        intersection = topics_a & topics_b
+        union = topics_a | topics_b
+
+        return len(intersection) / len(union) if union else 0.0
+
+    def rank_by_impact(self, memories: list["UserMemory"]) -> list["UserMemory"]:
+        """Rank memories by impact score (descending)"""
+        return sorted(memories, key=lambda m: self.evaluate(m), reverse=True)

backend/app/services/memory/importance_scorer.py

@@ -0,0 +1,103 @@
+"""
+ImportanceScorer
+
+Composite importance scoring combining frequency, recency, emotion, and impact.
+"""
+
+from enum import Enum
+from typing import TYPE_CHECKING
+
+from app.services.memory.frequency_tracker import FrequencyTracker
+from app.services.memory.emotion_analyzer import EmotionAnalyzer
+from app.services.memory.impact_evaluator import ImpactEvaluator
+
+if TYPE_CHECKING:
+    from app.models.memory import UserMemory
+
+
+class ImportanceLevel(str, Enum):
+    """Importance level classification"""
+
+    HIGH = "high"
+    MEDIUM = "medium"
+    LOW = "low"
+
+
+class ImportanceScorer:
+    """Calculate composite importance score for memories
+
+    Score formula:
+        frequency_score * 0.35 +
+        recency_score * 0.20 +
+        emotion_score * 0.25 +
+        impact_score * 0.20
+    """
+
+    # Score weights
+    WEIGHT_FREQUENCY = 0.35
+    WEIGHT_RECENCY = 0.20
+    WEIGHT_EMOTION = 0.25
+    WEIGHT_IMPACT = 0.20
+
+    # Escalation threshold
+    HIGH_THRESHOLD = 0.8
+    MEDIUM_THRESHOLD = 0.5
+    LOW_THRESHOLD = 0.0
+
+    def __init__(self):
+        self.tracker = FrequencyTracker()
+        self.emotion_analyzer = EmotionAnalyzer()
+        self.impact_evaluator = ImpactEvaluator()
+
+    def calculate_score(self, memory: "UserMemory") -> float:
+        """Calculate composite importance score (0.0 - 1.0)"""
+        frequency = self.tracker.get_frequency_score(memory) * self.WEIGHT_FREQUENCY
+        recency = self.tracker.get_recency_score(memory) * self.WEIGHT_RECENCY
+        emotion = self.emotion_analyzer.calculate_score(memory) * self.WEIGHT_EMOTION
+        impact = self.impact_evaluator.evaluate(memory) * self.WEIGHT_IMPACT
+
+        total = frequency + recency + emotion + impact
+        return round(min(1.0, max(0.0, total)), 3)
+
+    def get_importance_level(self, score: float) -> ImportanceLevel:
+        """Classify importance score into level"""
+        if score >= self.HIGH_THRESHOLD:
+            return ImportanceLevel.HIGH
+        elif score >= self.MEDIUM_THRESHOLD:
+            return ImportanceLevel.MEDIUM
+        else:
+            return ImportanceLevel.LOW
+
+    def should_escalate(self, memory: "UserMemory") -> bool:
+        """Check if a memory should be escalated (promoted to higher importance)
+
+        A memory should escalate if:
+        - Score exceeds HIGH_THRESHOLD
+        - Emotion analysis shows high intensity keywords
+        """
+        score = self.calculate_score(memory)
+        if score >= self.HIGH_THRESHOLD:
+            return True
+
+        # Check emotion intensity
+        emotion_score = self.emotion_analyzer.calculate_score(memory)
+        if emotion_score >= 0.9:
+            return True
+
+        return False
+
+    def score_and_classify(self, memory: "UserMemory") -> tuple[float, ImportanceLevel]:
+        """Calculate score and classify in one call"""
+        score = self.calculate_score(memory)
+        level = self.get_importance_level(score)
+        return score, level
+
+    def update_memory_importance(self, memory: "UserMemory") -> "UserMemory":
+        """Update memory's importance_score and importance_level fields
+
+        Returns the updated memory.
+        """
+        score, level = self.score_and_classify(memory)
+        memory.importance_score = score
+        memory.importance_level = level.value if isinstance(level, ImportanceLevel) else level
+        return memory

backend/app/services/memory/memory_decay.py

@@ -0,0 +1,81 @@
+"""
+MemoryDecay
+
+Handles memory archiving, deprioritization, and restoration.
+"""
+
+from datetime import UTC, datetime
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from app.models.memory import UserMemory
+
+
+class MemoryDecay:
+    """Handle memory archiving and deprioritization decisions."""
+
+    ARCHIVE_THRESHOLD = 0.2
+    DEPRIORITIZE_THRESHOLD = 0.5
+
+    def evaluate(self, memory: "UserMemory") -> dict:
+        """Evaluate memory and return action recommendation.
+
+        Returns:
+            dict with keys: decay_score, should_archive, should_deprioritize, action
+        """
+        from app.services.memory.forgetting_curve import ForgettingCurve
+
+        curve = ForgettingCurve()
+        decay_score = curve.calculate_decay(memory)
+        archive = decay_score < self.ARCHIVE_THRESHOLD
+        deprioritize = decay_score < self.DEPRIORITIZE_THRESHOLD
+
+        if archive:
+            action = "archive"
+        elif deprioritize:
+            action = "deprioritize"
+        else:
+            action = "keep_active"
+
+        return {
+            "decay_score": decay_score,
+            "should_archive": archive,
+            "should_deprioritize": deprioritize,
+            "action": action,
+        }
+
+    def archive_memory(self, memory: "UserMemory") -> "UserMemory":
+        """Archive a memory (set is_archived=True, reset decay_score to low value).
+
+        Archived memories are moved to cold storage and not included in
+        active reminders or context injection.
+        """
+        memory.is_archived = True
+        memory.decay_score = 0.1  # Very low, will be restored on access
+        memory.archive_at = datetime.now(UTC)
+        return memory
+
+    def deprioritize_memory(self, memory: "UserMemory") -> "UserMemory":
+        """Mark a memory as deprioritized (excluded from active reminders).
+
+        Unlike archival, the memory is still accessible and included in
+        context injection if relevant.
+        """
+        # Just update decay_score, the importance_level already encodes priority
+        from app.services.memory.forgetting_curve import ForgettingCurve
+
+        curve = ForgettingCurve()
+        memory.decay_score = curve.calculate_decay(memory)
+        return memory
+
+    def restore_from_archive(self, memory: "UserMemory") -> "UserMemory":
+        """Restore a memory from archive.
+
+        Resets is_archived=False and decay_score=0.8 (strong retention).
+        The memory is moved back to hot storage.
+        """
+        memory.is_archived = False
+        memory.decay_score = 0.8  # Strong retention after restore
+        memory.last_accessed_at = datetime.now(UTC)
+        memory.archive_at = None
+        return memory

backend/app/services/memory/memory_extractor.py

@@ -0,0 +1,239 @@
+"""
+MemoryExtractor
+
+Automatically extracts memories from conversations using LLM.
+Extracts 5 types: fact, preference, goal, pain_point, event.
+Deduplicates against existing memories (similarity > 0.85 → reinforce instead of create).
+"""
+
+import json
+import logging
+from dataclasses import dataclass, field
+from typing import Any
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.models.conversation import Message
+from app.models.memory import UserMemory
+
+logger = logging.getLogger(__name__)
+
+MEMORY_TYPES = ("fact", "preference", "goal", "pain_point", "event")
+
+EXTRACT_PROMPT = """从以下对话中提取用户的记忆信息，以 JSON 格式返回。
+
+对话内容：
+{conversation_text}
+
+提取以下类型（只提取明确信息，不要猜测）：
+- fact: 关于用户的客观事实（职业、 location、技能、健康状况等）
+- preference: 用户的偏好和习惯（回答风格偏好、沟通偏好、生活习惯等）
+- goal: 用户提到的目标或计划（想做什么、计划做什么、目标是什么）
+- pain_point: 反复出现或明显困扰用户的问题
+- event: 今天发生的重要事件
+
+输出格式（只输出 JSON，不要其他内容）：
+[
+  {{"type": "fact", "content": "...", "confidence": 0.9}},
+  {{"type": "goal", "content": "...", "confidence": 0.7}}
+]"""
+
+
+@dataclass
+class ExtractedMemory:
+    """A memory extracted from conversation."""
+
+    memory_type: str  # "fact" | "preference" | "goal" | "pain_point" | "event"
+    content: str
+    confidence: float  # 0.0-1.0
+    source_conversation_id: str | None = None
+
+
+class MemoryExtractor:
+    """Extract memories from conversations using LLM."""
+
+    SIMILARITY_THRESHOLD = 0.85
+
+    async def extract_from_conversation(
+        self,
+        db: AsyncSession,
+        user_id: str,
+        conversation_id: str,
+        messages: list[Message],
+    ) -> list[ExtractedMemory]:
+        """Extract memories from conversation messages.
+
+        1. Build conversation text
+        2. Call LLM to extract memories
+        3. Parse JSON response
+        4. Deduplicate against existing memories
+        5. Return new memories
+        """
+        if len(messages) < 2:
+            return []
+
+        # 1. Build conversation text
+        conversation_text = "\n".join(f"[{m.role}] {m.content}" for m in messages[-10:])
+
+        # 2. Call LLM
+        extracted = await self._call_llm_extract(conversation_text)
+        if not extracted:
+            return []
+
+        # 3. Build ExtractedMemory objects
+        new_memories = [
+            ExtractedMemory(
+                memory_type=m["type"],
+                content=m["content"],
+                confidence=m.get("confidence", 0.5),
+                source_conversation_id=conversation_id,
+            )
+            for m in extracted
+            if m.get("type") in MEMORY_TYPES and m.get("content")
+        ]
+
+        # 4. Deduplicate
+        new_memories = await self._deduplicate(db, user_id, new_memories)
+
+        return new_memories
+
+    async def _call_llm_extract(self, conversation_text: str) -> list[dict]:
+        """Call LLM to extract memories from conversation text."""
+        from app.services.llm_service import get_llm
+        from langchain_core.messages import HumanMessage, SystemMessage
+
+        prompt = EXTRACT_PROMPT.format(conversation_text=conversation_text)
+
+        try:
+            llm = get_llm()
+            response = await llm.invoke(
+                [
+                    SystemMessage(
+                        content="你是一个记忆提取助手。从对话中提取用户的记忆信息，只返回JSON数组，不要其他内容。"
+                    ),
+                    HumanMessage(content=prompt),
+                ]
+            )
+            content = response.content.strip()
+
+            # Try to extract JSON from response
+            if content.startswith("["):
+                return json.loads(content)
+            # Try to find JSON in response
+            start = content.find("[")
+            end = content.rfind("]") + 1
+            if start != -1 and end != 0:
+                return json.loads(content[start:end])
+            return []
+        except (json.JSONDecodeError, Exception) as e:
+            logger.warning(f"Memory extraction LLM call failed: {e}")
+            return []
+
+    async def _deduplicate(
+        self,
+        db: AsyncSession,
+        user_id: str,
+        new_memories: list[ExtractedMemory],
+    ) -> list[ExtractedMemory]:
+        """Filter duplicates against existing UserMemory.
+
+        Similarity > 0.85 → reinforce existing instead of creating new.
+        Returns only truly new memories.
+        """
+        if not new_memories:
+            return []
+
+        result = await db.execute(
+            select(UserMemory)
+            .where(
+                UserMemory.user_id == user_id,
+                UserMemory.is_archived == False,
+            )
+            .limit(20)
+        )
+        existing = list(result.scalars().all())
+
+        deduplicated = []
+        for new_mem in new_memories:
+            is_duplicate = False
+            for existing_mem in existing:
+                if self._is_similar(new_mem.content, existing_mem.content):
+                    # Reinforce existing memory instead of creating new
+                    await self._reinforce_existing(db, existing_mem)
+                    is_duplicate = True
+                    break
+            if not is_duplicate:
+                deduplicated.append(new_mem)
+
+        return deduplicated
+
+    def _is_similar(self, text1: str, text2: str) -> bool:
+        """Simple similarity check using keyword overlap.
+
+        In production would use embedding similarity.
+        """
+        # Simple word overlap ratio
+        words1 = set(text1.lower().split())
+        words2 = set(text2.lower().split())
+        if not words1 or not words2:
+            return False
+
+        overlap = len(words1 & words2)
+        union = len(words1 | words2)
+        jaccard = overlap / union if union > 0 else 0
+
+        # Also check substring
+        if jaccard > 0.5:
+            return True
+        if len(text1) > 5 and len(text2) > 5:
+            if text1[:20].lower() == text2[:20].lower():
+                return True
+        return False
+
+    async def _reinforce_existing(
+        self,
+        db: AsyncSession,
+        memory: UserMemory,
+    ) -> None:
+        """Reinforce an existing memory instead of creating a duplicate."""
+        from app.services.memory.reinforcement import MemoryReinforcement
+
+        reinforcement = MemoryReinforcement()
+        reinforcement.trigger(memory)
+        await db.commit()
+
+    async def save_memories(
+        self,
+        db: AsyncSession,
+        user_id: str,
+        conversation_id: str,
+        memories: list[ExtractedMemory],
+    ) -> list[UserMemory]:
+        """Save extracted memories as UserMemory records."""
+        from app.services.memory.importance_scorer import ImportanceScorer
+
+        saved = []
+        scorer = ImportanceScorer()
+
+        for mem in memories:
+            user_mem = UserMemory(
+                user_id=user_id,
+                memory_type=mem.memory_type,
+                content=mem.content,
+                source_conversation_id=mem.source_conversation_id,
+                importance_score=0.5,  # Will be updated by scorer
+                importance_level="medium",
+            )
+            # Update importance based on content
+            scorer.update_memory_importance(user_mem)
+
+            db.add(user_mem)
+            saved.append(user_mem)
+
+        if saved:
+            await db.commit()
+            for mem in saved:
+                await db.refresh(mem)
+
+        return saved

backend/app/services/memory/proactive_informer.py

@@ -0,0 +1,136 @@
+"""
+ProactiveInformer
+
+Checks conversation context and proactively informs user of relevant memories.
+Only fires probabilistically based on trigger type.
+"""
+
+import random
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from sqlalchemy.ext.asyncio import AsyncSession
+
+
+# Trigger types and their firing probabilities
+TRIGGERS = {
+    "high_importance_topic": {"probability": 0.8},
+    "repeat_question": {"probability": 1.0},
+    "forgotten_context": {"probability": 0.5},
+    "pending_goal": {"probability": 0.3},
+}
+
+# Message style templates for proactive informing
+INFORM_STYLE = {
+    "casual": "对了，你之前提到...",
+    "gentle": "不知道你有没有注意到...",
+    "helpful": "我记起你关心这个，要不看看...",
+}
+
+# Keywords that indicate different trigger types
+TRIGGER_KEYWORDS = {
+    "high_importance_topic": ["关于", "提到", "说过", "记得"],
+    "repeat_question": ["之前", "上次", "以前", "好像问过"],
+    "forgotten_context": ["忘了", "不记得", "记不清", "之前聊过"],
+    "pending_goal": ["目标", "计划", "想做", "打算", "想学"],
+}
+
+
+class ProactiveInformer:
+    """Proactively inform users of relevant memories based on conversation context."""
+
+    def should_inform(self, trigger_type: str) -> bool:
+        """Check if should inform based on probability."""
+        if trigger_type not in TRIGGERS:
+            return False
+        probability = TRIGGERS[trigger_type]["probability"]
+        return random.random() < probability
+
+    def detect_trigger(self, message: str) -> str | None:
+        """Detect which trigger type (if any) this message corresponds to."""
+        msg_lower = message.lower()
+        for trigger_type, keywords in TRIGGER_KEYWORDS.items():
+            for keyword in keywords:
+                if keyword in msg_lower:
+                    return trigger_type
+        return None
+
+    def get_inform_message(self, trigger_type: str, context: dict) -> str:
+        """Generate natural proactive message based on trigger type."""
+        style = INFORM_STYLE.get(context.get("style", "casual"), INFORM_STYLE["casual"])
+
+        if trigger_type == "high_importance_topic":
+            memory_content = context.get("memory_content", "")
+            return f"{style}「{memory_content[:30]}」这个话题你很关心，要深入聊聊吗？"
+
+        elif trigger_type == "repeat_question":
+            return "你之前问过类似的问题，我之前的回答可能还有参考价值。"
+
+        elif trigger_type == "forgotten_context":
+            memory_content = context.get("memory_content", "")
+            return f"这个话题你一个月前聊过：「{memory_content[:30]}」。要恢复一下吗？"
+
+        elif trigger_type == "pending_goal":
+            goal_content = context.get("goal_content", "")
+            return f"你之前说要「{goal_content[:30]}」，有进展了吗？"
+
+        return ""
+
+    async def check_and_inform(
+        self,
+        db: "AsyncSession",
+        user_id: str,
+        current_message: str,
+    ) -> str | None:
+        """Main entry point.
+
+        Returns proactive message if should inform, else None.
+        """
+        # 1. Detect trigger type
+        trigger_type = self.detect_trigger(current_message)
+        if not trigger_type:
+            return None
+
+        # 2. Check probability
+        if not self.should_inform(trigger_type):
+            return None
+
+        # 3. Fetch relevant context
+        context = await self._fetch_trigger_context(db, user_id, trigger_type)
+        if not context:
+            return None
+
+        # 4. Generate message
+        return self.get_inform_message(trigger_type, context)
+
+    async def _fetch_trigger_context(
+        self,
+        db: "AsyncSession",
+        user_id: str,
+        trigger_type: str,
+    ) -> dict | None:
+        """Fetch relevant context for the trigger type."""
+        from sqlalchemy import select
+        from app.models.memory import UserMemory
+
+        # Get most recent high-importance memory
+        result = await db.execute(
+            select(UserMemory)
+            .where(
+                UserMemory.user_id == user_id,
+                UserMemory.importance_level == "high",
+                UserMemory.is_archived == False,
+            )
+            .order_by(UserMemory.last_accessed_at.desc())
+            .limit(1)
+        )
+        memory = result.scalar_one_or_none()
+
+        if not memory:
+            return None
+
+        return {
+            "memory_content": memory.content,
+            "style": "casual",
+            "memory_id": str(memory.id),
+        }

backend/app/services/memory/recall_injector.py

@@ -0,0 +1,168 @@
+"""
+MemoryRecallInjector
+
+Injects relevant memories into LLM system prompt before response generation.
+Token budget: 800 by default (configurable).
+Priority: pain_point > goal > preference > fact > event
+"""
+
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from sqlalchemy.ext.asyncio import AsyncSession
+    from app.models.memory import UserMemory
+
+
+MEMORY_TYPE_PRIORITY = {
+    "pain_point": 1,
+    "goal": 2,
+    "preference": 3,
+    "fact": 4,
+    "event": 5,
+}
+
+DEFAULT_TOKEN_BUDGET = 800
+
+
+class MemoryRecallInjector:
+    """Inject relevant memories into system prompt with token budget control."""
+
+    def __init__(self, token_budget: int = DEFAULT_TOKEN_BUDGET):
+        self.token_budget = token_budget
+
+    async def build_context(
+        self,
+        db: "AsyncSession",
+        user_id: str,
+        current_message: str,
+        token_budget: int | None = None,
+    ) -> str:
+        """Build memory context string for injection into system prompt.
+
+        1. Recall relevant memories (top_k=20)
+        2. Filter out archived memories
+        3. Rank by importance × relevance × type priority
+        4. Select within token budget
+        5. Format as system prompt fragment
+        """
+        budget = token_budget or self.token_budget
+
+        # 1. Recall candidates using existing memory service
+        candidates = await recall_user_memories_for_injection(
+            db, user_id, current_message, top_k=20
+        )
+
+        # 2. Filter archived and non-UserMemory
+        active = [
+            m
+            for m in candidates
+            if isinstance(m, UserMemory) and not getattr(m, "is_archived", False)
+        ]
+
+        # 3. Rank
+        ranked = self._rank(active, current_message)
+
+        # 4. Budget select
+        selected = self._budget_select(ranked, budget)
+
+        # 5. Format
+        return self._format(selected)
+
+    def _rank(
+        self,
+        memories: list["UserMemory"],
+        query: str,
+    ) -> list["UserMemory"]:
+        """Rank memories by: relevance * 0.6 + importance * 0.4 * type_boost.
+
+        pain_point/goal get 1.0 type boost, others get 0.8.
+        """
+
+        def score(m: "UserMemory") -> float:
+            relevance = (
+                getattr(m, "similarity_score", 0.5) if hasattr(m, "similarity_score") else 0.5
+            )
+            importance = getattr(m, "importance_score", 0.5) or 0.5
+            mem_type = getattr(m, "memory_type", None) or "fact"
+            type_boost = 1.0 if mem_type in ("goal", "pain_point") else 0.8
+            return relevance * 0.6 + importance * 0.4 * type_boost
+
+        return sorted(memories, key=score, reverse=True)
+
+    def _budget_select(
+        self,
+        memories: list["UserMemory"],
+        token_budget: int,
+    ) -> list["UserMemory"]:
+        """Greedy selection until token budget runs out.
+
+        Rough estimate: 1 token ≈ 2 characters, fixed overhead = 20 tokens.
+        """
+        selected = []
+        used = 20  # "[关于你的记忆]\n"
+        for m in memories:
+            content = getattr(m, "content", "") or ""
+            cost = len(content) // 2 + 10
+            if used + cost > token_budget:
+                break
+            selected.append(m)
+            used += cost
+        return selected
+
+    def _format(self, memories: list["UserMemory"]) -> str:
+        """Format memories as system prompt fragment."""
+        if not memories:
+            return ""
+        lines = ["[关于你的记忆]"]
+        for m in memories:
+            mem_type = getattr(m, "memory_type", None) or ""
+            content = getattr(m, "content", "") or ""
+            type_label = f"[{mem_type}]" if mem_type else ""
+            lines.append(f"- {type_label} {content}".strip())
+        return "\n".join(lines)
+
+
+async def recall_user_memories_for_injection(
+    db: "AsyncSession",
+    user_id: str,
+    query: str,
+    top_k: int = 5,
+) -> list:
+    """Recall user memories for injection (used by MemoryRecallInjector).
+
+    This is a simplified version of recall_user_memories that returns
+    UserMemory objects directly instead of dicts.
+    """
+    import re
+    from sqlalchemy import select
+    from app.models.memory import UserMemory
+
+    def _extract_query_tokens(q: str) -> list[str]:
+        normalized = (q or "").lower()
+        tokens = [token for token in re.findall(r"[a-z0-9]+", normalized) if len(token) >= 3]
+        for chunk in re.findall(r"[\u4e00-\u9fff]+", q or ""):
+            stripped = chunk.strip()
+            if len(stripped) >= 4:
+                tokens.append(stripped)
+        return list(dict.fromkeys(tokens))
+
+    query_tokens = _extract_query_tokens(query)
+
+    result = await db.execute(
+        select(UserMemory)
+        .where(UserMemory.user_id == user_id)
+        .order_by(UserMemory.importance_score.desc(), UserMemory.created_at.desc())
+    )
+    memories = list(result.scalars().all())
+
+    if query_tokens:
+        matched = [
+            m
+            for m in memories
+            if any(token in ((getattr(m, "content", "") or "").lower()) for token in query_tokens)
+        ]
+        if matched:
+            return matched[:top_k]
+        return memories[:top_k]
+
+    return memories[:top_k]

backend/app/services/memory/reinforcement.py

@@ -0,0 +1,72 @@
+"""
+MemoryReinforcement
+
+Triggers memory reinforcement on recall and handles auto-reinforcement.
+"""
+
+from datetime import UTC, datetime
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from app.models.memory import UserMemory
+
+
+class MemoryReinforcement:
+    """Reinforce memories on recall to prevent forgetting."""
+
+    MAX_FREQUENCY = 10
+    AUTO_REINFORCE_BOOST = 1.1  # 10% boost per week
+
+    def trigger(self, memory: "UserMemory") -> "UserMemory":
+        """Called when memory is recalled: reset decay_score, increment frequency.
+
+        This is the core reinforcement mechanism - each recall makes the memory
+        stickier by resetting its decay curve and incrementing frequency.
+        """
+        from app.services.memory.forgetting_curve import ForgettingCurve
+
+        # Increment frequency count (capped at MAX_FREQUENCY)
+        current_freq = getattr(memory, "frequency_count", 0) or 0
+        memory.frequency_count = min(current_freq + 1, self.MAX_FREQUENCY)
+
+        # Update last accessed time
+        now = datetime.now(UTC)
+        memory.last_accessed_at = now
+        memory.last_recalled_at = now
+
+        # Reset decay score to near 1.0 (fully retained)
+        curve = ForgettingCurve()
+        memory.decay_score = min(0.95, curve.calculate_decay(memory) + 0.1)
+
+        return memory
+
+    def auto_reinforce(self, memories: list["UserMemory"]) -> list["UserMemory"]:
+        """Weekly auto-reinforce for high-importance memories.
+
+        Applies a 10% boost to frequency_count for high-importance memories
+        that haven't been accessed recently, keeping them fresh.
+        """
+        reinforced = []
+        now = datetime.now(UTC)
+
+        for memory in memories:
+            importance_level = getattr(memory, "importance_level", "medium") or "medium"
+            if importance_level != "high":
+                continue
+
+            current_freq = getattr(memory, "frequency_count", 0) or 0
+            if current_freq >= self.MAX_FREQUENCY:
+                continue
+
+            # Apply 10% boost, capped at MAX_FREQUENCY
+            new_freq = min(int(current_freq * self.AUTO_REINFORCE_BOOST + 1), self.MAX_FREQUENCY)
+            memory.frequency_count = new_freq
+
+            # Slightly improve decay score
+            current_decay = getattr(memory, "decay_score", 0.5) or 0.5
+            memory.decay_score = min(0.95, current_decay * 1.05)
+
+            memory.last_accessed_at = now
+            reinforced.append(memory)
+
+        return reinforced

backend/app/services/memory/reminder_scheduler.py

@@ -0,0 +1,113 @@
+"""
+ReminderScheduler
+
+Schedules and manages user reminders.
+"""
+
+from datetime import UTC, datetime, timedelta
+from typing import TYPE_CHECKING
+
+from sqlalchemy import select, and_
+
+from app.models.reminder import Reminder
+
+if TYPE_CHECKING:
+    from sqlalchemy.ext.asyncio import AsyncSession
+
+
+class ReminderScheduler:
+    """Schedule and manage user reminders."""
+
+    async def create_reminder(
+        self,
+        db: "AsyncSession",
+        user_id: str,
+        content: str,
+        trigger_at: datetime,
+        trigger_type: str = "time",
+        context_memory_id: str | None = None,
+    ) -> Reminder:
+        """Create a new reminder."""
+        reminder = Reminder(
+            user_id=user_id,
+            content=content,
+            trigger_type=trigger_type,
+            trigger_at=trigger_at,
+            context_memory_id=context_memory_id,
+            status="pending",
+        )
+        db.add(reminder)
+        await db.commit()
+        await db.refresh(reminder)
+        return reminder
+
+    async def get_due_reminders(self, db: "AsyncSession", user_id: str) -> list[Reminder]:
+        """Get reminders that are due (status=pending, trigger_at <= now, not snoozed)."""
+        now = datetime.now(UTC)
+        result = await db.execute(
+            select(Reminder)
+            .where(
+                Reminder.user_id == user_id,
+                Reminder.status == "pending",
+                Reminder.trigger_at <= now,
+                ((Reminder.snoozed_until.is_(None)) | (Reminder.snoozed_until <= now)),
+            )
+            .order_by(Reminder.trigger_at.asc())
+        )
+        return list(result.scalars().all())
+
+    async def snooze(
+        self,
+        db: "AsyncSession",
+        reminder_id: int,
+        minutes: int,
+    ) -> Reminder | None:
+        """Snooze reminder for N minutes."""
+        result = await db.execute(select(Reminder).where(Reminder.id == reminder_id))
+        reminder = result.scalar_one_or_none()
+        if not reminder:
+            return None
+
+        reminder.status = "snoozed"
+        reminder.snoozed_until = datetime.now(UTC) + timedelta(minutes=minutes)
+        await db.commit()
+        await db.refresh(reminder)
+        return reminder
+
+    async def dismiss(self, db: "AsyncSession", reminder_id: int) -> bool:
+        """Mark reminder as dismissed."""
+        result = await db.execute(select(Reminder).where(Reminder.id == reminder_id))
+        reminder = result.scalar_one_or_none()
+        if not reminder:
+            return False
+
+        reminder.status = "dismissed"
+        await db.commit()
+        return True
+
+    async def mark_sent(self, db: "AsyncSession", reminder_id: int) -> bool:
+        """Mark reminder as sent."""
+        result = await db.execute(select(Reminder).where(Reminder.id == reminder_id))
+        reminder = result.scalar_one_or_none()
+        if not reminder:
+            return False
+
+        reminder.status = "sent"
+        await db.commit()
+        return True
+
+    async def get_pending_reminders(
+        self,
+        db: "AsyncSession",
+        user_id: str,
+    ) -> list[Reminder]:
+        """Get all pending reminders for a user."""
+        result = await db.execute(
+            select(Reminder)
+            .where(
+                Reminder.user_id == user_id,
+                Reminder.status.in_(["pending", "snoozed"]),
+            )
+            .order_by(Reminder.trigger_at.asc())
+        )
+        return list(result.scalars().all())

backend/app/services/memory_service.py

@@ -7,6 +7,7 @@ Jarvis 记忆系统 (基于 Mem0)
 import logging
 import os
 import re
+import json
 from datetime import UTC, datetime
 from typing import Optional, Any
 from sqlalchemy import select, desc, func
@@ -15,6 +16,13 @@ from app.models.conversation import Conversation, Message
 from app.models.memory import UserMemory
 from app.models.user import User
 from app.services.brain_service import BrainService
+from app.services.memory.frequency_tracker import FrequencyTracker
+from app.services.memory.emotion_analyzer import EmotionAnalyzer
+from app.services.memory.impact_evaluator import ImpactEvaluator
+from app.services.memory.importance_scorer import ImportanceScorer
+from app.services.memory.forgetting_curve import ForgettingCurve
+from app.services.memory.memory_decay import MemoryDecay
+from app.services.memory.reinforcement import MemoryReinforcement
 from app.config import settings as _settings
 
 try:
@@ -312,8 +320,7 @@ def _extract_memory_query_tokens(query: str) -> list[str]:
             tokens.append(stripped_chunk)
             if len(stripped_chunk) > 6:
                 tokens.extend(
-                    stripped_chunk[index:index + 4]
+                    stripped_chunk[index : index + 4] for index in range(len(stripped_chunk) - 3)
-                    for index in range(len(stripped_chunk) - 3)
                 )
 
     return list(dict.fromkeys(tokens))
@@ -344,16 +351,21 @@ async def recall_user_memories(
 
     query_tokens = _extract_memory_query_tokens(query)
     statement = select(UserMemory).where(UserMemory.user_id == user_id)
-    result = await db.execute(statement.order_by(UserMemory.importance.desc(), UserMemory.created_at.desc()))
+    result = await db.execute(
+        statement.order_by(UserMemory.importance_score.desc(), UserMemory.created_at.desc())
+    )
     fallback_memories = list(result.scalars().all())
 
-    if _contains_hint(_normalize_query(query), MEMORY_QUERY_HINTS) or _matches_memory_query_pattern(_normalize_query(query)):
+    if _contains_hint(_normalize_query(query), MEMORY_QUERY_HINTS) or _matches_memory_query_pattern(
+        _normalize_query(query)
+    ):
         return fallback_memories[:top_k]
 
     if query_tokens:
         matched_memories = [
-            memory for memory in fallback_memories
+            memory
-            if any(token in (memory.content or '').lower() for token in query_tokens)
+            for memory in fallback_memories
+            if any(token in (memory.content or "").lower() for token in query_tokens)
         ]
         return matched_memories[:top_k]
 
@@ -361,13 +373,31 @@ async def recall_user_memories(
 
 
 async def _mark_memories_recalled(db: AsyncSession, memories: list[UserMemory]) -> None:
+    """Mark memories as recalled and update importance score + reinforce them."""
+    from app.services.memory.frequency_tracker import FrequencyTracker
+    from app.services.memory.importance_scorer import ImportanceScorer
+    from app.services.memory.reinforcement import MemoryReinforcement
+
     recalled_at = datetime.now(UTC)
+    tracker = FrequencyTracker()
+    scorer = ImportanceScorer()
+    reinforcement = MemoryReinforcement()
     updated = False
+
     for memory in memories:
         memory.is_recalled = True
         memory.recall_count = (memory.recall_count or 0) + 1
         memory.last_recalled_at = recalled_at
+        memory.frequency_count = memory.recall_count  # Keep in sync
+
+        # Update importance score on recall
+        scorer.update_memory_importance(memory)
+
+        # M.2: Reinforce memory on recall (reset decay, increment frequency)
+        reinforcement.trigger(memory)
+
         updated = True
+
     if updated:
         await db.commit()
 
@@ -417,9 +447,7 @@ MEMORY_QUERY_HINTS = (
     "偏好",
     "习惯",
 )
-MEMORY_QUERY_PATTERNS = (
+MEMORY_QUERY_PATTERNS = (re.compile(r"\bremember\s+(?:that\s+)?i\b"),)
-    re.compile(r"\bremember\s+(?:that\s+)?i\b"),
-)
 GROUNDING_QUERY_HINTS = (
     "根据文档",
     "严格根据",
@@ -634,3 +662,77 @@ async def update_memory(
     except Exception as e:
         print(f"Mem0 update error: {e}")
         return False
+
+
+# ———— M.2: 遗忘曲线处理 ————
+
+
+async def process_memory_decay(db: AsyncSession, user_id: str) -> dict:
+    """
+    处理用户所有记忆的衰减：
+    1. 计算每条记忆的 decay_score
+    2. 归档 decay < 0.2 的记忆
+    3. 降权 decay < 0.5 的记忆
+    """
+    from sqlalchemy import select, update
+
+    result = await db.execute(
+        select(UserMemory).where(
+            UserMemory.user_id == user_id,
+            UserMemory.is_archived == False,
+        )
+    )
+    memories = list(result.scalars().all())
+
+    archived_count = 0
+    deprioritized_count = 0
+    curve = ForgettingCurve()
+    decay_mgr = MemoryDecay()
+
+    for memory in memories:
+        evaluation = decay_mgr.evaluate(memory)
+        decay_score = evaluation["decay_score"]
+
+        # Update decay_score on the memory
+        memory.decay_score = decay_score
+
+        if evaluation["should_archive"]:
+            decay_mgr.archive_memory(memory)
+            archived_count += 1
+        elif evaluation["should_deprioritize"]:
+            decay_mgr.deprioritize_memory(memory)
+            deprioritized_count += 1
+
+    if archived_count > 0 or deprioritized_count > 0:
+        await db.commit()
+
+    return {
+        "archived": archived_count,
+        "deprioritized": deprioritized_count,
+        "total": len(memories),
+    }
+
+
+async def process_weekly_reinforcement(db: AsyncSession, user_id: str) -> int:
+    """
+    每周自动强化高重要性记忆。
+    Returns number of reinforced memories.
+    """
+    from sqlalchemy import select
+
+    result = await db.execute(
+        select(UserMemory).where(
+            UserMemory.user_id == user_id,
+            UserMemory.importance_level == "high",
+            UserMemory.is_archived == False,
+        )
+    )
+    memories = list(result.scalars().all())
+
+    reinforcement = MemoryReinforcement()
+    reinforced = reinforcement.auto_reinforce(memories)
+
+    if reinforced:
+        await db.commit()
+
+    return len(reinforced)

backend/app/services/remote_sync_service.py

@@ -0,0 +1,108 @@
+from io import BytesIO
+from datetime import UTC, datetime
+
+from sqlalchemy import and_, select
+from sqlalchemy.ext.asyncio import AsyncSession
+from starlette.datastructures import UploadFile
+
+from app.models.folder import Folder
+from app.models.remote_mount import RemoteMount, RemoteSyncItem
+from app.services.document_service import DocumentService
+from app.services.webdav_service import WebDavNode, WebDavService
+
+
+class RemoteSyncService:
+    def __init__(self, db: AsyncSession, user_id: str):
+        self.db = db
+        self.user_id = user_id
+
+    async def sync_remote_path(
+        self,
+        mount: RemoteMount,
+        remote_path: str,
+        local_folder_id: str,
+        mode: str = "file",
+    ) -> dict:
+        folder = await self.db.execute(
+            select(Folder).where(and_(Folder.id == local_folder_id, Folder.user_id == self.user_id))
+        )
+        if folder.scalar_one_or_none() is None:
+            raise ValueError("本地目标文件夹不存在")
+
+        webdav = WebDavService(mount)
+        document_service = DocumentService(self.db, self.user_id)
+
+        synced = 0
+        skipped = 0
+        failed = 0
+        document_ids: list[str] = []
+        errors: list[str] = []
+
+        if mode == "folder":
+            nodes = await webdav.list_tree(remote_path)
+            targets = self._flatten_files(nodes)
+        else:
+            name = remote_path.rstrip("/").split("/")[-1] or "remote-file"
+            targets = [WebDavNode(path=remote_path, name=name, is_dir=False)]
+
+        for node in targets:
+            try:
+                content, filename = await webdav.download_file(node.path)
+                upload = UploadFile(filename=filename, file=BytesIO(content))
+                document = await document_service.upload_document(self.user_id, upload, folder_id=local_folder_id)
+                await self._upsert_sync_item(mount.id, node, local_folder_id, document.id)
+                document_ids.append(document.id)
+                synced += 1
+            except Exception as exc:  # noqa: BLE001
+                failed += 1
+                errors.append(f"{node.path}: {exc}")
+                await self._upsert_sync_item(mount.id, node, local_folder_id, None, status="failed", error=str(exc))
+
+        mount.last_sync_at = datetime.now(UTC).isoformat()
+        await self.db.commit()
+        return {
+            "synced": synced,
+            "skipped": skipped,
+            "failed": failed,
+            "document_ids": document_ids,
+            "errors": errors,
+        }
+
+    def _flatten_files(self, nodes: list[WebDavNode]) -> list[WebDavNode]:
+        results: list[WebDavNode] = []
+        for node in nodes:
+            if node.is_dir:
+                results.extend(self._flatten_files(node.children))
+            else:
+                results.append(node)
+        return results
+
+    async def _upsert_sync_item(
+        self,
+        mount_id: str,
+        node: WebDavNode,
+        local_folder_id: str,
+        local_document_id: str | None,
+        status: str = "synced",
+        error: str | None = None,
+    ) -> None:
+        result = await self.db.execute(
+            select(RemoteSyncItem).where(
+                and_(RemoteSyncItem.mount_id == mount_id, RemoteSyncItem.remote_path == node.path)
+            )
+        )
+        sync_item = result.scalar_one_or_none()
+        if sync_item is None:
+            sync_item = RemoteSyncItem(
+                mount_id=mount_id,
+                remote_path=node.path,
+            )
+            self.db.add(sync_item)
+
+        sync_item.remote_etag = node.etag
+        sync_item.remote_modified_at = node.modified_at
+        sync_item.local_folder_id = local_folder_id
+        sync_item.local_document_id = local_document_id
+        sync_item.sync_status = status
+        sync_item.last_error = error
+        sync_item.last_synced_at = datetime.now(UTC).isoformat()

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/scheduler_service.py

@@ -20,7 +20,137 @@ logger = logging.getLogger(__name__)
 scheduler = AsyncIOScheduler(timezone="Asia/Shanghai")
 
 
-# ===================== 定时任务函数 =====================
+# ===================== M.2: 遗忘曲线任务 =====================
+
+
+async def daily_forgetting_check():
+    """
+    每日遗忘检查 (03:00)
+    - 计算所有记忆的 decay_score
+    - 归档 decay < 0.2 的记忆
+    - 降权 decay < 0.5 的记忆
+    """
+    from app.services.memory_service import process_memory_decay
+    from sqlalchemy import select
+
+    logger.info("[Scheduler] 开始执行每日遗忘检查...")
+
+    async with async_session() as db:
+        from app.models.user import User
+
+        result = await db.execute(select(User).where(User.is_active == True))
+        users = result.scalars().all()
+
+        total_archived = 0
+        total_deprioritized = 0
+        for user in users:
+            try:
+                decay_result = await process_memory_decay(db, user.id)
+                total_archived += decay_result["archived"]
+                total_deprioritized += decay_result["deprioritized"]
+            except Exception as e:
+                logger.error(f"[Scheduler] 用户 {user.id} 遗忘检查失败: {e}")
+
+        logger.info(
+            f"[Scheduler] 每日遗忘检查完成，归档 {total_archived} 条，降权 {total_deprioritized} 条"
+        )
+
+
+async def weekly_reinforcement_task():
+    """
+    每周自动强化 (周一 04:00)
+    对 high 重要性记忆做轻量强化
+    """
+    from app.services.memory_service import process_weekly_reinforcement
+    from sqlalchemy import select
+
+    logger.info("[Scheduler] 开始执行每周强化任务...")
+
+    async with async_session() as db:
+        from app.models.user import User
+
+        result = await db.execute(select(User).where(User.is_active == True))
+        users = result.scalars().all()
+
+        total_reinforced = 0
+        for user in users:
+            try:
+                count = await process_weekly_reinforcement(db, user.id)
+                total_reinforced += count
+            except Exception as e:
+                logger.error(f"[Scheduler] 用户 {user.id} 强化任务失败: {e}")
+
+        logger.info(f"[Scheduler] 每周强化完成，共强化 {total_reinforced} 条记忆")
+
+
+# ===================== M.3: 主动提醒任务 =====================
+
+
+async def daily_digest_generation():
+    """
+    每日摘要生成 (22:00)
+    为所有活跃用户生成每日摘要
+    """
+    from app.services.memory.daily_digest import DailyDigestGenerator
+    from sqlalchemy import select
+
+    logger.info("[Scheduler] 开始执行每日摘要生成...")
+
+    async with async_session() as db:
+        from app.models.user import User
+
+        result = await db.execute(select(User).where(User.is_active == True))
+        users = result.scalars().all()
+
+        generated = 0
+        generator = DailyDigestGenerator()
+        for user in users:
+            try:
+                from datetime import date
+
+                digest = await generator.generate(db, user.id, target_date=date.today())
+                # In production, would save digest to database
+                generated += 1
+            except Exception as e:
+                logger.error(f"[Scheduler] 用户 {user.id} 摘要生成失败: {e}")
+
+        logger.info(f"[Scheduler] 每日摘要生成完成，共生成 {generated} 条")
+
+
+async def reminder_check_task():
+    """
+    提醒检查 (每15分钟)
+    检查到期的提醒并标记为 sent
+    """
+    from sqlalchemy import select
+
+    logger.info("[Scheduler] 开始检查到期提醒...")
+
+    async with async_session() as db:
+        from app.models.reminder import Reminder
+        from app.services.memory.reminder_scheduler import ReminderScheduler
+
+        scheduler = ReminderScheduler()
+        result = await db.execute(
+            select(Reminder).where(
+                Reminder.status == "pending",
+            )
+        )
+        reminders = result.scalars().all()
+
+        sent_count = 0
+        for reminder in reminders:
+            try:
+                due = await scheduler.get_due_reminders(db, reminder.user_id)
+                for due_reminder in due:
+                    await scheduler.mark_sent(db, due_reminder.id)
+                    sent_count += 1
+            except Exception as e:
+                logger.error(f"[Scheduler] 提醒检查失败: {e}")
+
+        if sent_count > 0:
+            logger.info(f"[Scheduler] 提醒检查完成，发送 {sent_count} 条提醒")
+
 
 async def daily_task_analysis():
     """
@@ -37,15 +167,13 @@ async def daily_task_analysis():
         yesterday = datetime.now(UTC).date() - timedelta(days=1)
 
         # 统计昨日任务完成情况
-        result = await db.execute(
+        result = await db.execute(select(Task).where(Task.updated_at >= yesterday))
-            select(Task).where(Task.updated_at >= yesterday)
-        )
         tasks = result.scalars().all()
 
         completed = [t for t in tasks if t.status == "done"]
         pending = [t for t in tasks if t.status != "done"]
 
-        report = f"""## 每日任务报告 - {yesterday.strftime('%Y-%m-%d')}
+        report = f"""## 每日任务报告 - {yesterday.strftime("%Y-%m-%d")}
 
 ### 完成情况
 - 总任务数: {len(tasks)}
@@ -65,6 +193,7 @@ async def daily_task_analysis():
 
         # 发布到论坛
         from app.models.forum import ForumPost
+
         post = ForumPost(
             title=f"每日报告 - {yesterday.strftime('%Y-%m-%d')}",
             content=report,
@@ -97,11 +226,14 @@ async def forum_scan_task():
 
     async with async_session() as db:
         from sqlalchemy import select
+
         result = await db.execute(
-            select(ForumPost).where(
+            select(ForumPost)
+            .where(
                 ForumPost.category == "instruction",
                 ForumPost.is_executed == False,
-            ).limit(5)
+            )
+            .limit(5)
         )
         posts = result.scalars().all()
 
@@ -165,9 +297,9 @@ async def tag_generation_task():
             tag_service = TagService(db, llm_client)
 
             result = await db.execute(
-                select(KGNode.user_id).distinct().where(
+                select(KGNode.user_id)
-                    KGNode.entity_type.in_(["conversation", "document", "chunk"])
+                .distinct()
-                )
+                .where(KGNode.entity_type.in_(["conversation", "document", "chunk"]))
             )
             user_ids = result.scalars().all()
 
@@ -211,8 +343,75 @@ async def daily_todo_generation():
             logger.error(f"[Scheduler] 每日待办生成失败: {e}")
 
 
+# ———— M.4: 主动记忆提取 ————
+async def check_idle_conversations():
+    """
+    每30分钟检查空闲超过30分钟的对话，提取记忆
+    M.4: 主动记忆提取
+    """
+    from datetime import timedelta, datetime, UTC
+    from app.models.conversation import Conversation, Message
+    from app.services.memory.memory_extractor import MemoryExtractor
+
+    logger.info("[Scheduler] 开始检查空闲对话...")
+
+    async with async_session() as db:
+        try:
+            # Find conversations idle > 30 minutes (no recent messages)
+            cutoff = datetime.now(UTC) - timedelta(minutes=30)
+
+            # Subquery to find last message time per conversation
+            from sqlalchemy import func
+
+            subq = (
+                select(Message.conversation_id, func.max(Message.created_at).label("last_message"))
+                .group_by(Message.conversation_id)
+                .having(func.max(Message.created_at) < cutoff)
+            ).subquery()
+
+            result = await db.execute(
+                select(Conversation)
+                .join(subq, Conversation.id == subq.c.conversation_id)
+                .where(Conversation.updated_at >= datetime.now(UTC) - timedelta(hours=24))
+                .limit(10)
+            )
+            idle_conversations = list(result.scalars().all())
+
+            extractor = MemoryExtractor()
+            total_extracted = 0
+
+            for conv in idle_conversations:
+                try:
+                    # Get conversation messages
+                    msg_result = await db.execute(
+                        select(Message)
+                        .where(Message.conversation_id == conv.id)
+                        .order_by(Message.created_at.desc())
+                        .limit(10)
+                    )
+                    messages = list(msg_result.scalars().all())
+
+                    if len(messages) >= 2:
+                        new_memories = await extractor.extract_from_conversation(
+                            db, conv.user_id, conv.id, messages
+                        )
+                        if new_memories:
+                            await extractor.save_memories(db, conv.user_id, conv.id, new_memories)
+                            total_extracted += len(new_memories)
+                except Exception as e:
+                    logger.warning(
+                        f"[MemoryExtractor] Failed to process conversation {conv.id}: {e}"
+                    )
+
+            if total_extracted > 0:
+                logger.info(f"[Scheduler] 空闲对话记忆提取完成，共提取 {total_extracted} 条记忆")
+        except Exception as e:
+            logger.error(f"[Scheduler] 空闲对话检查失败: {e}")
+
+
 # ===================== 调度器管理 =====================
 
+
 def start_scheduler():
     """启动调度器，注册所有定时任务"""
     if scheduler.running:
@@ -264,6 +463,54 @@ def start_scheduler():
         replace_existing=True,
     )
 
+    # ———— M.2: 遗忘曲线系统 ————
+    # 每天凌晨 03:00 执行遗忘检查
+    scheduler.add_job(
+        daily_forgetting_check,
+        CronTrigger(hour=3, minute=0, timezone="Asia/Shanghai"),
+        id="daily_forgetting_check",
+        name="每日遗忘检查",
+        replace_existing=True,
+    )
+
+    # 每周一 04:00 执行自动强化
+    scheduler.add_job(
+        weekly_reinforcement_task,
+        CronTrigger(day_of_week="mon", hour=4, minute=0, timezone="Asia/Shanghai"),
+        id="weekly_reinforcement",
+        name="每周记忆强化",
+        replace_existing=True,
+    )
+
+    # ———— M.3: 主动提醒系统 ————
+    # 每天 22:00 生成每日摘要
+    scheduler.add_job(
+        daily_digest_generation,
+        CronTrigger(hour=22, minute=0, timezone="Asia/Shanghai"),
+        id="daily_digest_generation",
+        name="每日摘要生成",
+        replace_existing=True,
+    )
+
+    # 每15分钟检查到期提醒
+    scheduler.add_job(
+        reminder_check_task,
+        IntervalTrigger(minutes=15),
+        id="reminder_check",
+        name="提醒检查",
+        replace_existing=True,
+    )
+
+    # ———— M.4: 主动记忆提取 ————
+    # 每30分钟检查空闲对话并提取记忆
+    scheduler.add_job(
+        check_idle_conversations,
+        IntervalTrigger(minutes=30),
+        id="check_idle_conversations",
+        name="空闲对话记忆提取",
+        replace_existing=True,
+    )
+
     scheduler.start()
     logger.info("[Scheduler] 定时任务调度器已启动")
 
@@ -282,10 +529,12 @@ def get_scheduler_status() -> dict:
 
     jobs = []
     for job in scheduler.get_jobs():
-        jobs.append({
+        jobs.append(
+            {
                 "id": job.id,
                 "name": job.name,
                 "next_run": str(job.next_run_time) if job.next_run_time else None,
-        })
+            }
+        )
 
     return {"status": "running", "jobs": jobs}

backend/app/services/secret_service.py

@@ -0,0 +1,24 @@
+import base64
+import hashlib
+
+from cryptography.fernet import Fernet
+
+from app.config import settings
+
+
+def _build_fernet() -> Fernet:
+    digest = hashlib.sha256(settings.SECRET_KEY.encode("utf-8")).digest()
+    key = base64.urlsafe_b64encode(digest)
+    return Fernet(key)
+
+
+def encrypt_secret(value: str | None) -> str | None:
+    if not value:
+        return None
+    return _build_fernet().encrypt(value.encode("utf-8")).decode("utf-8")
+
+
+def decrypt_secret(value: str | None) -> str | None:
+    if not value:
+        return None
+    return _build_fernet().decrypt(value.encode("utf-8")).decode("utf-8")

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

@@ -1,8 +1,12 @@
 from datetime import datetime, UTC
 from time import monotonic
+import os
 import platform
 import socket
 import subprocess
+import time
+
+import httpx
 
 try:
     import psutil
@@ -14,6 +18,15 @@ 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
+        from app.config import settings
+        self._settings = settings
 
     def _get_network_rates(self) -> tuple[float, float]:
         counters = psutil.net_io_counters()
@@ -127,3 +140,96 @@ class SystemService:
             **gpu_status,
             'timestamp': datetime.now(UTC).isoformat(),
         }
+
+    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': location,
+            **weather,
+            'weather_cached': False,
+            'weather_cached_at': now,
+        }