fix: bound idle evaluation model streams

.gitignore

@@ -26,6 +26,7 @@ router_session_parquet/
 # Environment files
 .env
 .env.*
+skills/*/.local/
 
 # Local benchmark outputs
 benchmark_artifacts/
@@ -42,4 +43,8 @@ benchmarks/data/manifest.json
 claude-code-sourcemap-main
 router_session_parquet
 
+# Locally managed annotation pages served by /annotations
+annotation-pages/
+
 标签定义
+.codex-tmp

backend/api/server.py

@@ -47,6 +47,12 @@ from src.agent_types import (
 )
 from src.bundled_skills import ALWAYS_ENABLED_HIDDEN_SKILL_NAMES, get_bundled_skills
 from src.data_agent_inputs import DataAgentInputError, load_input_sources
+from src.evaluation_runtime import (
+    DEFAULT_CONCURRENCY,
+    MAX_CONCURRENCY,
+    EvaluationError,
+    EvaluationRuntime,
+)
 from src.jupyter_runtime import (
     DEFAULT_JUPYTER_WORKSPACE_ROOT,
     JupyterRuntimeError,
@@ -586,6 +592,12 @@ class AgentState:
             self.model_config_for,
         )
         self.memory_manager.start()
+        self.evaluation_runtime = EvaluationRuntime(
+            root=self.session_directory.parent / 'evaluations',
+            cwd_for_account=lambda account_id: self.config_for(account_id).cwd,
+            model_config_for=self.model_config_for,
+            account_paths_for=self.account_paths,
+        )
 
     @property
     def cwd(self) -> Path:
@@ -1125,6 +1137,78 @@ class SkillSyncRequest(BaseModel):
     account_id: str | None = None
 
 
+class EvaluationDatasetCreateRequest(BaseModel):
+    account_id: str = Field(min_length=1)
+    name: str = ''
+    filename: str = Field(min_length=1)
+    content_base64: str | None = None
+    rows: list[dict[str, Any]] | None = None
+
+
+class EvaluationDatasetMappingRequest(BaseModel):
+    account_id: str = Field(min_length=1)
+    mapping: dict[str, Any]
+
+
+class EvaluationLabelMappingSuggestRequest(BaseModel):
+    account_id: str = Field(min_length=1)
+    mapping: dict[str, Any]
+    skill_name: str = 'label-master'
+    snapshot_id: str | None = None
+    model: str | None = None
+
+
+class EvaluationAnalyzeRequest(BaseModel):
+    account_id: str = Field(min_length=1)
+    query: str = Field(min_length=1, max_length=20_000)
+    skill_name: str = 'label-master'
+    snapshot_id: str | None = None
+    model: str | None = None
+    history: list[dict[str, Any]] = Field(default_factory=list)
+    context: dict[str, Any] = Field(default_factory=dict)
+    domain: str = ''
+    request_id: str = ''
+    metadata: dict[str, Any] = Field(default_factory=dict)
+
+
+class EvaluationExperimentCreateRequest(BaseModel):
+    account_id: str = Field(min_length=1)
+    dataset_id: str = Field(min_length=1)
+    name: str = ''
+    skill_name: str = 'label-master'
+    snapshot_id: str | None = None
+    model: str | None = None
+    concurrency: int = Field(
+        default=DEFAULT_CONCURRENCY,
+        ge=1,
+        le=MAX_CONCURRENCY,
+    )
+
+
+class EvaluationSkillVersionSaveRequest(BaseModel):
+    account_id: str = Field(min_length=1)
+    base_snapshot_id: str = Field(min_length=1)
+    version_name: str = Field(min_length=1, max_length=80)
+    note: str = Field(default='', max_length=1000)
+    files: dict[str, str]
+
+
+class EvaluationExperimentActionRequest(BaseModel):
+    account_id: str = Field(min_length=1)
+
+
+class EvaluationRetryRequest(BaseModel):
+    account_id: str = Field(min_length=1)
+    case_ids: list[str] | None = None
+    disagreements_only: bool = False
+
+
+class EvaluationReviewRequest(BaseModel):
+    account_id: str = Field(min_length=1)
+    review_status: str = Field(max_length=80)
+    review_note: str = Field(default='', max_length=4000)
+
+
 class SessionUpdate(BaseModel):
     title: str | None = Field(default=None, max_length=80)
     is_training: bool | None = None
@@ -1132,6 +1216,7 @@ class SessionUpdate(BaseModel):
 
 class FeishuAccountRequest(BaseModel):
     account_id: str | None = None
+    force: bool = False
 
 
 class FeishuOnlineDocRequest(BaseModel):
@@ -1177,6 +1262,7 @@ def create_app(state: AgentState) -> FastAPI:
             scanner_task.cancel()
             _watcher_manager.cancel_all()
             _bash_bg_manager.cancel_all()
+            state.evaluation_runtime.shutdown()
             state.event_loop = None
 
     app = FastAPI(title='Claw Code GUI', version='1.0', lifespan=lifespan)
@@ -1467,6 +1553,331 @@ def create_app(state: AgentState) -> FastAPI:
         )
         return result
 
+    # ------------- Skill evaluations ---------------------------------------
+    @app.get('/api/evaluations/metadata')
+    async def evaluation_metadata(account_id: str) -> dict[str, Any]:
+        try:
+            return state.evaluation_runtime.metadata(_safe_account_id(account_id))
+        except EvaluationError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    @app.get('/api/evaluations/skills/{skill_name}/versions')
+    async def list_evaluation_skill_versions(
+        skill_name: str,
+        account_id: str,
+    ) -> dict[str, Any]:
+        try:
+            return await asyncio.to_thread(
+                state.evaluation_runtime.list_skill_versions,
+                _safe_account_id(account_id),
+                skill_name,
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    @app.get(
+        '/api/evaluations/skills/{skill_name}/versions/{snapshot_id}'
+    )
+    async def get_evaluation_skill_version(
+        skill_name: str,
+        snapshot_id: str,
+        account_id: str,
+    ) -> dict[str, Any]:
+        try:
+            return state.evaluation_runtime.get_skill_version_manifest(
+                _safe_account_id(account_id),
+                skill_name,
+                snapshot_id,
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=404, detail=str(exc)) from exc
+
+    @app.get(
+        '/api/evaluations/skills/{skill_name}/versions/{snapshot_id}/file'
+    )
+    async def get_evaluation_skill_version_file(
+        skill_name: str,
+        snapshot_id: str,
+        account_id: str,
+        path: str,
+    ) -> dict[str, Any]:
+        try:
+            return state.evaluation_runtime.read_skill_version_file(
+                _safe_account_id(account_id),
+                skill_name,
+                snapshot_id,
+                path,
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=404, detail=str(exc)) from exc
+
+    @app.post('/api/evaluations/skills/{skill_name}/versions')
+    async def save_evaluation_skill_version(
+        skill_name: str,
+        payload: EvaluationSkillVersionSaveRequest,
+    ) -> dict[str, Any]:
+        try:
+            return await asyncio.to_thread(
+                state.evaluation_runtime.save_skill_version,
+                _safe_account_id(payload.account_id),
+                skill_name,
+                base_snapshot_id=payload.base_snapshot_id,
+                version_name=payload.version_name,
+                note=payload.note,
+                files=payload.files,
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    @app.post('/api/evaluations/analyze')
+    async def analyze_evaluation_case(
+        payload: EvaluationAnalyzeRequest,
+    ) -> dict[str, Any]:
+        try:
+            return await asyncio.to_thread(
+                state.evaluation_runtime.analyze_case,
+                account_id=_safe_account_id(payload.account_id),
+                query=payload.query,
+                skill_name=payload.skill_name,
+                snapshot_id=payload.snapshot_id,
+                model=payload.model,
+                history=payload.history,
+                context=payload.context,
+                domain=payload.domain,
+                request_id=payload.request_id,
+                metadata=payload.metadata,
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    @app.get('/api/evaluations/analyses')
+    async def list_evaluation_analyses(
+        account_id: str,
+    ) -> list[dict[str, Any]]:
+        return state.evaluation_runtime.list_single_analyses(
+            _safe_account_id(account_id)
+        )
+
+    @app.get('/api/evaluations/analyses/{analysis_id}')
+    async def get_evaluation_analysis(
+        analysis_id: str,
+        account_id: str,
+    ) -> dict[str, Any]:
+        try:
+            return state.evaluation_runtime.get_single_analysis(
+                analysis_id,
+                _safe_account_id(account_id),
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=404, detail=str(exc)) from exc
+
+    @app.get('/api/evaluations/datasets')
+    async def list_evaluation_datasets(account_id: str) -> list[dict[str, Any]]:
+        return state.evaluation_runtime.list_datasets(_safe_account_id(account_id))
+
+    @app.post('/api/evaluations/datasets')
+    async def create_evaluation_dataset(
+        payload: EvaluationDatasetCreateRequest,
+    ) -> dict[str, Any]:
+        try:
+            return await asyncio.to_thread(
+                state.evaluation_runtime.create_dataset,
+                account_id=_safe_account_id(payload.account_id),
+                name=payload.name,
+                filename=payload.filename,
+                content_base64=payload.content_base64,
+                rows=payload.rows,
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    @app.get('/api/evaluations/datasets/{dataset_id}')
+    async def get_evaluation_dataset(
+        dataset_id: str,
+        account_id: str,
+    ) -> dict[str, Any]:
+        try:
+            return state.evaluation_runtime.get_dataset(
+                dataset_id,
+                _safe_account_id(account_id),
+                include_rows=False,
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=404, detail=str(exc)) from exc
+
+    @app.patch('/api/evaluations/datasets/{dataset_id}/mapping')
+    async def update_evaluation_dataset_mapping(
+        dataset_id: str,
+        payload: EvaluationDatasetMappingRequest,
+    ) -> dict[str, Any]:
+        try:
+            return state.evaluation_runtime.update_dataset_mapping(
+                dataset_id,
+                _safe_account_id(payload.account_id),
+                payload.mapping,
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    @app.post('/api/evaluations/datasets/{dataset_id}/label-mapping/suggest')
+    async def suggest_evaluation_label_mapping(
+        dataset_id: str,
+        payload: EvaluationLabelMappingSuggestRequest,
+    ) -> dict[str, Any]:
+        try:
+            return await asyncio.to_thread(
+                state.evaluation_runtime.suggest_label_mapping,
+                dataset_id,
+                _safe_account_id(payload.account_id),
+                mapping=payload.mapping,
+                skill_name=payload.skill_name,
+                snapshot_id=payload.snapshot_id,
+                model=payload.model,
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    @app.get('/api/evaluations/experiments')
+    async def list_evaluation_experiments(
+        account_id: str,
+    ) -> list[dict[str, Any]]:
+        return state.evaluation_runtime.list_experiments(
+            _safe_account_id(account_id)
+        )
+
+    @app.post('/api/evaluations/experiments')
+    async def create_evaluation_experiment(
+        payload: EvaluationExperimentCreateRequest,
+    ) -> dict[str, Any]:
+        try:
+            return await asyncio.to_thread(
+                state.evaluation_runtime.create_experiment,
+                account_id=_safe_account_id(payload.account_id),
+                dataset_id=payload.dataset_id,
+                name=payload.name,
+                skill_name=payload.skill_name,
+                snapshot_id=payload.snapshot_id,
+                model=payload.model,
+                concurrency=payload.concurrency,
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    @app.get('/api/evaluations/experiments/{experiment_id}')
+    async def get_evaluation_experiment(
+        experiment_id: str,
+        account_id: str,
+    ) -> dict[str, Any]:
+        try:
+            return state.evaluation_runtime.get_experiment(
+                experiment_id,
+                _safe_account_id(account_id),
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=404, detail=str(exc)) from exc
+
+    @app.post('/api/evaluations/experiments/{experiment_id}/start')
+    async def start_evaluation_experiment(
+        experiment_id: str,
+        payload: EvaluationExperimentActionRequest,
+    ) -> dict[str, Any]:
+        try:
+            return state.evaluation_runtime.start(
+                experiment_id,
+                _safe_account_id(payload.account_id),
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    @app.post('/api/evaluations/experiments/{experiment_id}/pause')
+    async def pause_evaluation_experiment(
+        experiment_id: str,
+        payload: EvaluationExperimentActionRequest,
+    ) -> dict[str, Any]:
+        try:
+            return state.evaluation_runtime.pause(
+                experiment_id,
+                _safe_account_id(payload.account_id),
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    @app.post('/api/evaluations/experiments/{experiment_id}/resume')
+    async def resume_evaluation_experiment(
+        experiment_id: str,
+        payload: EvaluationExperimentActionRequest,
+    ) -> dict[str, Any]:
+        try:
+            return state.evaluation_runtime.resume(
+                experiment_id,
+                _safe_account_id(payload.account_id),
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    @app.post('/api/evaluations/experiments/{experiment_id}/cancel')
+    async def cancel_evaluation_experiment(
+        experiment_id: str,
+        payload: EvaluationExperimentActionRequest,
+    ) -> dict[str, Any]:
+        try:
+            return state.evaluation_runtime.cancel(
+                experiment_id,
+                _safe_account_id(payload.account_id),
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    @app.post('/api/evaluations/experiments/{experiment_id}/retry')
+    async def retry_evaluation_experiment(
+        experiment_id: str,
+        payload: EvaluationRetryRequest,
+    ) -> dict[str, Any]:
+        try:
+            return state.evaluation_runtime.retry(
+                experiment_id,
+                _safe_account_id(payload.account_id),
+                case_ids=payload.case_ids,
+                disagreements_only=payload.disagreements_only,
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    @app.patch('/api/evaluations/cases/{case_id}/review')
+    async def review_evaluation_case(
+        case_id: str,
+        payload: EvaluationReviewRequest,
+    ) -> dict[str, Any]:
+        try:
+            return state.evaluation_runtime.update_review(
+                case_id,
+                _safe_account_id(payload.account_id),
+                review_status=payload.review_status,
+                review_note=payload.review_note,
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
+    @app.get('/api/evaluations/experiments/{experiment_id}/export')
+    async def export_evaluation_experiment(
+        experiment_id: str,
+        account_id: str,
+    ) -> Response:
+        try:
+            filename, content = state.evaluation_runtime.export_csv(
+                experiment_id,
+                _safe_account_id(account_id),
+            )
+        except EvaluationError as exc:
+            raise HTTPException(status_code=404, detail=str(exc)) from exc
+        return Response(
+            content,
+            media_type='text/csv; charset=utf-8',
+            headers={
+                'Content-Disposition': f"attachment; filename*=UTF-8''{quote(filename)}"
+            },
+        )
+
     # ------------- memory ----------------------------------------------------
     @app.get('/api/memory/user')
     async def get_user_memory(account_id: str) -> dict[str, Any]:
@@ -1616,7 +2027,7 @@ def create_app(state: AgentState) -> FastAPI:
 
     @app.post('/api/integrations/feishu/login')
     async def feishu_login(payload: FeishuAccountRequest) -> dict[str, Any]:
-        return _start_feishu_login(state, payload.account_id)
+        return _start_feishu_login(state, payload.account_id, force=payload.force)
 
     @app.post('/api/integrations/feishu/logout')
     async def feishu_logout(payload: FeishuAccountRequest) -> dict[str, Any]:
@@ -1684,6 +2095,20 @@ def create_app(state: AgentState) -> FastAPI:
         except OSError as exc:
             raise HTTPException(status_code=400, detail=f'读取文件失败: {exc}')
         except Exception as exc:
+            if _is_feishu_auth_error(str(exc)):
+                _stop_feishu_login(payload.account_id)
+                logout_result = _run_feishu_cli(paths, ['logout'], timeout_seconds=30)
+                status = _feishu_status_payload(state, payload.account_id)
+                status['logout_output'] = logout_result.get('output')
+                raise HTTPException(
+                    status_code=409,
+                    detail={
+                        'code': 'feishu_auth_expired',
+                        'message': '飞书登录已过期或未登录，请重新授权后再转换。',
+                        'force_login': True,
+                        'status': status,
+                    },
+                )
             target_name = '飞书表格' if kind == 'sheet' else '飞书文档'
             raise HTTPException(status_code=502, detail=f'{target_name}创建失败: {exc}')
         url = _extract_first_url(rendered)
@@ -3169,8 +3594,42 @@ def _stored_display_messages(
     stored: StoredAgentSession,
 ) -> tuple[dict[str, Any], ...]:
     if stored.display_messages:
-        return tuple(dict(message) for message in stored.display_messages)
+        messages = tuple(dict(message) for message in stored.display_messages)
-    return tuple(dict(message) for message in stored.messages)
+    else:
+        messages = tuple(dict(message) for message in stored.messages)
+    return tuple(message for message in messages if _is_visible_display_message(message))
+
+
+_INTERNAL_DISPLAY_KINDS = {
+    'compact_boundary',
+    'compact_summary',
+    'continuation_request',
+    'file_history_replay',
+    'plugin_tool_runtime',
+    'runtime_context',
+    'snipped_message',
+    'system_context',
+}
+
+
+def _is_visible_display_message(message: dict[str, Any]) -> bool:
+    if message.get('role') == 'system':
+        return False
+    metadata = message.get('metadata')
+    if isinstance(metadata, dict):
+        kind = metadata.get('kind')
+        if isinstance(kind, str) and kind in _INTERNAL_DISPLAY_KINDS:
+            return False
+    content = message.get('content')
+    if isinstance(content, str):
+        stripped = content.strip()
+        if stripped.startswith('<system-reminder>'):
+            return False
+        if message.get('role') == 'user' and stripped.startswith(
+            'This session is being continued from a previous conversation'
+        ):
+            return False
+    return True
 
 
 def _stored_session_has_incomplete_tail(stored: StoredAgentSession) -> bool:
@@ -3572,10 +4031,7 @@ def _save_in_progress_session(
             stored = load_agent_session(safe_id, directory=directory)
         except (FileNotFoundError, OSError, json.JSONDecodeError):
             return None
-        messages = list(stored.messages)
         display_messages = list(_stored_display_messages(stored))
-        if messages and _is_pending_user_message(dict(messages[-1])):
-            messages.pop()
         if display_messages and _is_pending_user_message(dict(display_messages[-1])):
             display_messages.pop()
         pending_message = {
@@ -3585,9 +4041,6 @@ def _save_in_progress_session(
             'metadata': pending_metadata,
             'message_id': f'user_pending_{int(time.time() * 1000)}',
         }
-        messages.append(
-            dict(pending_message)
-        )
         display_messages.append(dict(pending_message))
         budget_state = (
             dict(stored.budget_state)
@@ -3599,7 +4052,6 @@ def _save_in_progress_session(
             save_agent_session(
                 replace(
                     stored,
-                    messages=tuple(messages),
                     display_messages=tuple(display_messages),
                     budget_state=budget_state,
                 ),
@@ -4352,16 +4804,36 @@ def _run_feishu_cli(
         }
 
 
+def _is_feishu_auth_error(text: str) -> bool:
+    lowered = text.lower()
+    markers = (
+        'not logged in',
+        'no credentials',
+        'unauthorized',
+        'unauthenticated',
+        'login expired',
+        'token expired',
+        'invalid token',
+        'access token expired',
+        'refresh token expired',
+        '登录已过期',
+        '授权已过期',
+        '未登录',
+        '未授权',
+        '凭证已过期',
+        '认证失败',
+        '请重新登录',
+        '重新授权',
+    )
+    return any(marker in lowered for marker in markers)
+
+
 def _feishu_status_payload(state: AgentState, account_id: str | None) -> dict[str, Any]:
     paths = _feishu_paths(state, account_id)
     _reap_feishu_login(account_id)
     result = _run_feishu_cli(paths, ['status'], timeout_seconds=30)
     output = str(result.get('output') or '')
-    lowered = output.lower()
+    logged_in = result.get('returncode') == 0 and not _is_feishu_auth_error(output)
-    logged_in = result.get('returncode') == 0 and not any(
-        marker in lowered
-        for marker in ('not logged in', '未登录', 'no credentials')
-    )
     pending = _feishu_login_snapshot(account_id)
     payload: dict[str, Any] = {
         'logged_in': logged_in,
@@ -4378,15 +4850,21 @@ def _feishu_status_payload(state: AgentState, account_id: str | None) -> dict[st
     return payload
 
 
-def _start_feishu_login(state: AgentState, account_id: str | None) -> dict[str, Any]:
+def _start_feishu_login(
-    status = _feishu_status_payload(state, account_id)
+    state: AgentState,
-    if status.get('logged_in'):
+    account_id: str | None,
-        return status
+    *,
+    force: bool = False,
+) -> dict[str, Any]:
+    if not force:
+        status = _feishu_status_payload(state, account_id)
+        if status.get('logged_in'):
+            return status
 
     key = _feishu_login_key(account_id)
     with _FEISHU_LOGIN_LOCK:
         existing = _FEISHU_LOGIN_PROCESSES.get(key)
-        if existing is not None and existing.process.poll() is None:
+        if not force and existing is not None and existing.process.poll() is None:
             return {
                 'logged_in': False,
                 'status': 'login_pending',
@@ -4395,6 +4873,8 @@ def _start_feishu_login(state: AgentState, account_id: str | None) -> dict[str,
         _FEISHU_LOGIN_PROCESSES.pop(key, None)
 
     paths = _feishu_paths(state, account_id)
+    if force:
+        _run_feishu_cli(paths, ['logout'], timeout_seconds=30)
     env = os.environ.copy()
     env.update(_feishu_env(paths))
     try:

docs/assets/planning-data-driven-loop-v1.svg

@@ -0,0 +1,148 @@
+<svg width="1600" height="900" viewBox="0 0 1600 900" xmlns="http://www.w3.org/2000/svg">
+  <defs>
+    <marker id="arrow" markerWidth="12" markerHeight="12" refX="10" refY="6" orient="auto">
+      <path d="M2,2 L10,6 L2,10 Z" fill="#64748b"/>
+    </marker>
+    <marker id="arrowBlue" markerWidth="12" markerHeight="12" refX="10" refY="6" orient="auto">
+      <path d="M2,2 L10,6 L2,10 Z" fill="#2563eb"/>
+    </marker>
+    <filter id="softShadow" x="-10%" y="-10%" width="120%" height="130%">
+      <feDropShadow dx="0" dy="8" stdDeviation="10" flood-color="#0f172a" flood-opacity="0.10"/>
+    </filter>
+    <style>
+      .title { font: 700 38px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #0f172a; }
+      .subtitle { font: 400 20px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #475569; }
+      .sectionTitle { font: 700 22px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #0f172a; }
+      .stepNo { font: 700 17px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #64748b; }
+      .stepTitle { font: 700 19px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #0f172a; }
+      .stepText { font: 400 14px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #475569; }
+      .badge { font: 700 13px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #9a3412; }
+      .baseTitle { font: 700 24px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #0f172a; }
+      .baseMain { font: 700 28px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #1e293b; }
+      .baseSub { font: 400 18px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #475569; }
+      .chip { font: 600 15px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #334155; }
+      .loopLabel { font: 600 17px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #2563eb; }
+    </style>
+  </defs>
+
+  <rect width="1600" height="900" fill="#f8fafc"/>
+
+  <text x="800" y="72" text-anchor="middle" class="title">planning 模型数据驱动迭代链路</text>
+  <text x="800" y="108" text-anchor="middle" class="subtitle">从问题发现到模型改进：数据开发 + AutoResearch + Agent 平台底座的全流程提效</text>
+
+  <!-- Main entry: problem discovery -->
+  <g filter="url(#softShadow)">
+    <rect x="68" y="220" width="190" height="130" rx="18" fill="#ffffff" stroke="#dbeafe" stroke-width="2"/>
+    <text x="95" y="253" class="stepNo">1</text>
+    <text x="163" y="283" text-anchor="middle" class="stepTitle">问题发现</text>
+    <text x="163" y="314" text-anchor="middle" class="stepText">线上 badcase</text>
+    <text x="163" y="337" text-anchor="middle" class="stepText">评测错误 / 指标异常</text>
+  </g>
+
+  <!-- Data development module -->
+  <g filter="url(#softShadow)">
+    <rect x="300" y="170" width="690" height="230" rx="24" fill="#eef6ff" stroke="#bfdbfe" stroke-width="2"/>
+    <text x="645" y="205" text-anchor="middle" class="sectionTitle">数据开发提效</text>
+    <text x="645" y="232" text-anchor="middle" class="stepText">把问题转成可训练、可评测的数据资产</text>
+  </g>
+
+  <!-- AutoResearch module -->
+  <g filter="url(#softShadow)">
+    <rect x="1040" y="170" width="455" height="230" rx="24" fill="#f5f3ff" stroke="#ddd6fe" stroke-width="2"/>
+    <text x="1267" y="205" text-anchor="middle" class="sectionTitle">AutoResearch</text>
+    <text x="1267" y="232" text-anchor="middle" class="stepText">planning 模型迭代与结构分析</text>
+  </g>
+
+  <!-- Step boxes: data development -->
+  <g>
+    <rect x="330" y="265" width="140" height="95" rx="16" fill="#ffffff" stroke="#bfdbfe" stroke-width="1.5"/>
+    <text x="350" y="292" class="stepNo">2</text>
+    <text x="400" y="316" text-anchor="middle" class="stepTitle">线上数据挖掘</text>
+    <text x="400" y="342" text-anchor="middle" class="stepText">rid / session</text>
+    <text x="400" y="362" text-anchor="middle" class="stepText">prompt / output</text>
+  </g>
+
+  <g>
+    <rect x="500" y="265" width="150" height="95" rx="16" fill="#fff7ed" stroke="#fed7aa" stroke-width="1.8"/>
+    <text x="520" y="292" class="stepNo">3</text>
+    <text x="575" y="316" text-anchor="middle" class="stepTitle">问题分析</text>
+    <text x="575" y="342" text-anchor="middle" class="stepText">归因 / 标签边界</text>
+    <rect x="536" y="352" width="78" height="24" rx="12" fill="#ffedd5" stroke="#fdba74"/>
+    <text x="575" y="369" text-anchor="middle" class="badge">人工 Review</text>
+  </g>
+
+  <g>
+    <rect x="680" y="265" width="140" height="95" rx="16" fill="#ffffff" stroke="#bfdbfe" stroke-width="1.5"/>
+    <text x="700" y="292" class="stepNo">4</text>
+    <text x="750" y="316" text-anchor="middle" class="stepTitle">数据构造</text>
+    <text x="750" y="342" text-anchor="middle" class="stepText">训练样本</text>
+    <text x="750" y="362" text-anchor="middle" class="stepText">评测样本 / 边界样例</text>
+  </g>
+
+  <g>
+    <rect x="850" y="265" width="115" height="95" rx="16" fill="#ffffff" stroke="#bfdbfe" stroke-width="1.5"/>
+    <text x="868" y="292" class="stepNo">5</text>
+    <text x="907" y="316" text-anchor="middle" class="stepTitle">格式转换</text>
+    <text x="907" y="342" text-anchor="middle" class="stepText">CSV / JSONL</text>
+    <text x="907" y="362" text-anchor="middle" class="stepText">eval set</text>
+  </g>
+
+  <!-- Step boxes: autoresearch -->
+  <g>
+    <rect x="1072" y="265" width="155" height="95" rx="16" fill="#ffffff" stroke="#ddd6fe" stroke-width="1.5"/>
+    <text x="1092" y="292" class="stepNo">6</text>
+    <text x="1149" y="316" text-anchor="middle" class="stepTitle">planning 模型迭代</text>
+    <text x="1149" y="342" text-anchor="middle" class="stepText">训练 / 评测</text>
+    <text x="1149" y="362" text-anchor="middle" class="stepText">配置实验</text>
+  </g>
+
+  <g>
+    <rect x="1262" y="265" width="200" height="95" rx="16" fill="#fff7ed" stroke="#fed7aa" stroke-width="1.8"/>
+    <text x="1282" y="292" class="stepNo">7</text>
+    <text x="1362" y="316" text-anchor="middle" class="stepTitle">结构分析 / 结果分析</text>
+    <text x="1362" y="342" text-anchor="middle" class="stepText">指标拆解 / 回归验证</text>
+    <rect x="1323" y="352" width="78" height="24" rx="12" fill="#ffedd5" stroke="#fdba74"/>
+    <text x="1362" y="369" text-anchor="middle" class="badge">人工 Review</text>
+  </g>
+
+  <!-- Straight arrows -->
+  <path d="M258 285 L318 285" stroke="#64748b" stroke-width="3" fill="none" marker-end="url(#arrow)"/>
+  <path d="M470 312 L492 312" stroke="#64748b" stroke-width="3" fill="none" marker-end="url(#arrow)"/>
+  <path d="M650 312 L672 312" stroke="#64748b" stroke-width="3" fill="none" marker-end="url(#arrow)"/>
+  <path d="M820 312 L842 312" stroke="#64748b" stroke-width="3" fill="none" marker-end="url(#arrow)"/>
+  <path d="M965 312 L1064 312" stroke="#64748b" stroke-width="3" fill="none" marker-end="url(#arrow)"/>
+  <path d="M1227 312 L1254 312" stroke="#64748b" stroke-width="3" fill="none" marker-end="url(#arrow)"/>
+
+  <!-- Feedback loop -->
+  <path d="M1362 402 C1362 500 1070 505 820 505 C465 505 163 485 163 365" stroke="#2563eb" stroke-width="3" stroke-dasharray="8 8" fill="none" marker-end="url(#arrowBlue)"/>
+  <text x="760" y="486" text-anchor="middle" class="loopLabel">新 badcase / 新边界 / 新样本 / 新策略回流</text>
+
+  <!-- Platform base -->
+  <g filter="url(#softShadow)">
+    <rect x="90" y="570" width="1410" height="210" rx="28" fill="#ffffff" stroke="#cbd5e1" stroke-width="2"/>
+    <text x="795" y="615" text-anchor="middle" class="baseTitle">ZK Data Agent 平台底座</text>
+    <text x="795" y="663" text-anchor="middle" class="baseMain">Agent Loop + Skill + 工具链 + 记忆 + Jupyter 工作区</text>
+    <text x="795" y="699" text-anchor="middle" class="baseSub">通过 Agent Loop 调度 Skill 和工具，连接线上数据、执行环境与产物沉淀</text>
+
+    <g>
+      <rect x="205" y="725" width="140" height="34" rx="17" fill="#f1f5f9"/>
+      <text x="275" y="748" text-anchor="middle" class="chip">线上日志工具</text>
+      <rect x="375" y="725" width="120" height="34" rx="17" fill="#f1f5f9"/>
+      <text x="435" y="748" text-anchor="middle" class="chip">标签知识</text>
+      <rect x="525" y="725" width="150" height="34" rx="17" fill="#f1f5f9"/>
+      <text x="600" y="748" text-anchor="middle" class="chip">数据格式转换</text>
+      <rect x="705" y="725" width="120" height="34" rx="17" fill="#f1f5f9"/>
+      <text x="765" y="748" text-anchor="middle" class="chip">会话沉淀</text>
+      <rect x="855" y="725" width="130" height="34" rx="17" fill="#f1f5f9"/>
+      <text x="920" y="748" text-anchor="middle" class="chip">运行态观测</text>
+      <rect x="1015" y="725" width="155" height="34" rx="17" fill="#f1f5f9"/>
+      <text x="1092" y="748" text-anchor="middle" class="chip">账号 / 工作区隔离</text>
+      <rect x="1200" y="725" width="160" height="34" rx="17" fill="#f1f5f9"/>
+      <text x="1280" y="748" text-anchor="middle" class="chip">训练评测工具链</text>
+    </g>
+  </g>
+
+  <!-- Thin connector from business flow to base -->
+  <path d="M795 526 L795 565" stroke="#94a3b8" stroke-width="2" stroke-dasharray="5 6"/>
+  <text x="825" y="548" class="stepText">平台能力支撑全链路</text>
+</svg>

docs/assets/planning-data-driven-loop-v2.svg

@@ -0,0 +1,133 @@
+<svg width="1600" height="900" viewBox="0 0 1600 900" xmlns="http://www.w3.org/2000/svg">
+  <defs>
+    <marker id="arrow" markerWidth="12" markerHeight="12" refX="10" refY="6" orient="auto">
+      <path d="M2,2 L10,6 L2,10 Z" fill="#64748b"/>
+    </marker>
+    <marker id="arrowBlue" markerWidth="12" markerHeight="12" refX="10" refY="6" orient="auto">
+      <path d="M2,2 L10,6 L2,10 Z" fill="#2563eb"/>
+    </marker>
+    <marker id="arrowPurple" markerWidth="12" markerHeight="12" refX="10" refY="6" orient="auto">
+      <path d="M2,2 L10,6 L2,10 Z" fill="#7c3aed"/>
+    </marker>
+    <filter id="softShadow" x="-10%" y="-10%" width="120%" height="130%">
+      <feDropShadow dx="0" dy="8" stdDeviation="10" flood-color="#0f172a" flood-opacity="0.10"/>
+    </filter>
+    <style>
+      .title { font: 700 38px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #0f172a; }
+      .subtitle { font: 400 20px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #475569; }
+      .sectionTitle { font: 700 22px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #0f172a; }
+      .sectionSub { font: 400 14px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #64748b; }
+      .stepNo { font: 700 17px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #64748b; }
+      .stepTitle { font: 700 19px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #0f172a; }
+      .stepText { font: 400 14px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #475569; }
+      .badge { font: 700 13px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #9a3412; }
+      .loopLabel { font: 600 17px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #2563eb; }
+      .iterLabel { font: 700 16px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #6d28d9; }
+      .base { font: 600 18px -apple-system, BlinkMacSystemFont, "PingFang SC", "Microsoft YaHei", sans-serif; fill: #475569; }
+    </style>
+  </defs>
+
+  <rect width="1600" height="900" fill="#f8fafc"/>
+
+  <text x="800" y="76" text-anchor="middle" class="title">planning 模型数据驱动迭代链路</text>
+  <text x="800" y="112" text-anchor="middle" class="subtitle">从问题发现到模型改进：数据开发 + AutoResearch 的全流程提效</text>
+
+  <!-- Step 1 -->
+  <g filter="url(#softShadow)">
+    <rect x="70" y="265" width="185" height="120" rx="18" fill="#ffffff" stroke="#dbeafe" stroke-width="2"/>
+    <text x="94" y="297" class="stepNo">1</text>
+    <text x="162" y="323" text-anchor="middle" class="stepTitle">问题发现</text>
+    <text x="162" y="351" text-anchor="middle" class="stepText">线上 badcase</text>
+    <text x="162" y="373" text-anchor="middle" class="stepText">评测错误 / 指标异常</text>
+  </g>
+
+  <!-- Data development module -->
+  <g filter="url(#softShadow)">
+    <rect x="300" y="190" width="640" height="250" rx="24" fill="#eef6ff" stroke="#bfdbfe" stroke-width="2"/>
+    <text x="620" y="226" text-anchor="middle" class="sectionTitle">数据开发提效</text>
+    <text x="620" y="252" text-anchor="middle" class="sectionSub">把问题转成可训练、可评测的数据资产</text>
+  </g>
+
+  <!-- Data development steps -->
+  <g>
+    <rect x="335" y="300" width="150" height="105" rx="16" fill="#ffffff" stroke="#bfdbfe" stroke-width="1.5"/>
+    <text x="358" y="329" class="stepNo">2</text>
+    <text x="410" y="352" text-anchor="middle" class="stepTitle">线上数据挖掘</text>
+    <text x="410" y="379" text-anchor="middle" class="stepText">rid / session</text>
+    <text x="410" y="399" text-anchor="middle" class="stepText">prompt / output</text>
+  </g>
+
+  <g>
+    <rect x="535" y="300" width="155" height="105" rx="16" fill="#fff7ed" stroke="#fed7aa" stroke-width="1.8"/>
+    <text x="558" y="329" class="stepNo">3</text>
+    <text x="612" y="352" text-anchor="middle" class="stepTitle">问题分析</text>
+    <text x="612" y="379" text-anchor="middle" class="stepText">归因 / 标签边界</text>
+    <rect x="573" y="389" width="78" height="24" rx="12" fill="#ffedd5" stroke="#fdba74"/>
+    <text x="612" y="406" text-anchor="middle" class="badge">人工 Review</text>
+  </g>
+
+  <g>
+    <rect x="740" y="300" width="165" height="105" rx="16" fill="#ffffff" stroke="#bfdbfe" stroke-width="1.5"/>
+    <text x="763" y="329" class="stepNo">4</text>
+    <text x="822" y="352" text-anchor="middle" class="stepTitle">数据构造</text>
+    <text x="822" y="379" text-anchor="middle" class="stepText">样本生成 / 格式转换</text>
+    <text x="822" y="399" text-anchor="middle" class="stepText">训练集 / 评测集</text>
+  </g>
+
+  <!-- AutoResearch module -->
+  <g filter="url(#softShadow)">
+    <rect x="990" y="190" width="450" height="250" rx="24" fill="#f5f3ff" stroke="#ddd6fe" stroke-width="2"/>
+    <text x="1215" y="226" text-anchor="middle" class="sectionTitle">AutoResearch</text>
+    <text x="1215" y="252" text-anchor="middle" class="sectionSub">planning 模型迭代与结构分析</text>
+  </g>
+
+  <!-- AutoResearch steps -->
+  <g>
+    <rect x="1028" y="300" width="165" height="105" rx="16" fill="#ffffff" stroke="#ddd6fe" stroke-width="1.5"/>
+    <text x="1051" y="329" class="stepNo">5</text>
+    <text x="1110" y="352" text-anchor="middle" class="stepTitle">planning 模型迭代</text>
+    <text x="1110" y="379" text-anchor="middle" class="stepText">训练 / 评测</text>
+    <text x="1110" y="399" text-anchor="middle" class="stepText">配置实验</text>
+  </g>
+
+  <g>
+    <rect x="1248" y="300" width="160" height="105" rx="16" fill="#fff7ed" stroke="#fed7aa" stroke-width="1.8"/>
+    <text x="1271" y="329" class="stepNo">6</text>
+    <text x="1328" y="352" text-anchor="middle" class="stepTitle">结构分析 / 结果分析</text>
+    <text x="1328" y="379" text-anchor="middle" class="stepText">指标拆解 / 回归验证</text>
+    <rect x="1289" y="389" width="78" height="24" rx="12" fill="#ffedd5" stroke="#fdba74"/>
+    <text x="1328" y="406" text-anchor="middle" class="badge">人工 Review</text>
+  </g>
+
+  <!-- Step 7 online -->
+  <g filter="url(#softShadow)">
+    <rect x="1128" y="520" width="175" height="100" rx="18" fill="#ffffff" stroke="#cbd5e1" stroke-width="2"/>
+    <text x="1152" y="550" class="stepNo">7</text>
+    <text x="1215" y="576" text-anchor="middle" class="stepTitle">模型上线</text>
+    <text x="1215" y="604" text-anchor="middle" class="stepText">发布 / 观察 / 回流</text>
+  </g>
+
+  <!-- Main arrows -->
+  <path d="M255 325 L326 325" stroke="#64748b" stroke-width="3" fill="none" marker-end="url(#arrow)"/>
+  <path d="M485 352 L526 352" stroke="#64748b" stroke-width="3" fill="none" marker-end="url(#arrow)"/>
+  <path d="M690 352 L731 352" stroke="#64748b" stroke-width="3" fill="none" marker-end="url(#arrow)"/>
+  <path d="M905 352 L1019 352" stroke="#64748b" stroke-width="3" fill="none" marker-end="url(#arrow)"/>
+
+  <!-- Iteration loop between 5 and 6 -->
+  <path d="M1195 325 C1222 292 1250 292 1275 319" stroke="#7c3aed" stroke-width="3" fill="none" marker-end="url(#arrowPurple)"/>
+  <path d="M1276 385 C1248 424 1218 424 1194 391" stroke="#7c3aed" stroke-width="3" fill="none" marker-end="url(#arrowPurple)"/>
+  <text x="1218" y="286" text-anchor="middle" class="iterLabel">循环迭代</text>
+
+  <!-- From 6 to 7 -->
+  <path d="M1328 405 C1328 480 1260 498 1225 514" stroke="#64748b" stroke-width="3" fill="none" marker-end="url(#arrow)"/>
+
+  <!-- Feedback loop from 7 to 1 -->
+  <path d="M1128 570 C920 690 520 690 270 585 C175 545 142 470 154 395" stroke="#2563eb" stroke-width="3" stroke-dasharray="8 8" fill="none" marker-end="url(#arrowBlue)"/>
+  <text x="650" y="660" text-anchor="middle" class="loopLabel">新 badcase / 新边界 / 新样本 / 新策略回流</text>
+
+  <!-- Light platform support bar -->
+  <g>
+    <rect x="140" y="752" width="1320" height="58" rx="18" fill="#f1f5f9" stroke="#e2e8f0" stroke-width="1"/>
+    <text x="800" y="788" text-anchor="middle" class="base">ZK Data Agent 支撑：通过 Agent Loop 调度 Skill 和工具，连接线上数据、执行环境与产物沉淀</text>
+  </g>
+</svg>

docs/share-report-v1.md

@@ -0,0 +1,145 @@
+# 从个人 Agent 到团队 Agent：沉淀可复用的 AI 提效探索
+
+## 平台地址
+
+- 平台入口：http://10.189.47.6/
+- 平台文档：http://10.189.47.6/doc
+
+## Git 地址
+
+- 仓库：https://git.n.xiaomi.com/wuyang6/zk-data-agent.git
+
+现有 Skill 数量 12 个，涵盖线上数据挖掘、数据生成、日志分析、标签知识、模型打标、训练发起 / 评测等业务功能，提升组内工作效率。
+
+## 1. 为什么要做这个
+
+个人 AI 提效已经发生了，下一步要解决的是团队复用和流程沉淀。
+
+[图片]
+
+最初的目标，是提升中控 planning 模型迭代中的数据开发效率。
+
+- 日志挖掘、问题分析、数据构造、数据格式化....
+
+做的过程中发现，我们发现低效并不只是某个步骤慢，而是三类能力分散在个人手里：
+
+- 业务知识分散：哪些日志表可查、哪些字段有用、标签边界怎么判断、问题应该怎么归因。
+- 执行工具分散：脚本、参数、路径、格式转换逻辑和产物规范各自维护。
+- AI 经验分散：有效的 AI 流程、会话历史、修正经验、个人偏好难以共享和继承。
+
+类似情况也存在于评测、训练、结果分析、问题排查、线上配置修改等工作中，用于数据开发提效的逻辑也同样适用...
+
+所以，这个项目逐渐从一个数据开发提效实践，演化成一个中控数据 Agent：把已经验证有效的业务知识、执行工具和 AI 经验，整理成可以共享、组合和继续迭代的 Skill 与工具链。
+
+## 2. 目前做了什么
+
+中控数据 Agent 以 Agent Loop 作为执行底座，以业务 Skill 作为核心组织方式，再通过工作区、工具链和记忆机制，把个人经验沉淀成团队可以继续使用和改造的能力。
+
+系统的架构：http://10.189.47.6/doc
+
+[图片]
+
+### 2.1 Skill 热更新与同步
+
+平台支持在页面中更新 Skill：当有人新增或修改 Skill 后，可以通过 Skill 面板同步到当前服务，让新的能力进入模型可见列表。这样业务能力可以跟随实际需求快速迭代，不需要每次都改核心 Agent 代码。
+
+这一点适合团队协作：
+
+- 个人可以先在自己的分支里开发 Skill；
+- 稳定后同步到平台；
+- 其他人可以直接启用和复用；
+- Skill 可以继续被修改、补充和演进。
+
+[图片]
+
+### 2.2 记忆机制
+
+平台目前把记忆分成两类：用户记忆和 Skill 记忆。
+
+- 用户记忆：记录个人偏好、常用环境、输出习惯等。
+- Skill 记忆：记录某个 Skill 使用过程中积累的经验、边界和常见坑。
+
+这类记忆可以在页面中查看和编辑。使用次数越多，某个 Skill 相关的经验越容易被沉淀下来，后续类似任务可以少重复解释。
+
+和常见的“全局用户偏好记忆”相比，这里的重点是把记忆绑定到业务能力上。
+
+例如线上挖掘中，哪些字段更可靠、session 应该怎么抽、哪些输出格式容易出错，这些经验更适合沉淀到对应 Skill，而不是混在一个全局记忆里。
+
+[图片]
+
+典型记忆方法的对比：
+
+| 记忆方案 | 主要解决什么 | 典型做法 | 对我们的启发 | ZK Data Agent 的取舍 |
+|---|---|---|---|---|
+| ChatGPT Memory | 个人助手的长期个性化 | 从用户对话、历史聊天、文件、记忆摘要等来源形成个性化上下文；用户可以查看 Memory Sources、纠正、删除或关闭记忆。OpenAI 文档也强调，删除某条记忆并不等于删除所有来源，完整删除需要处理过去聊天、文件、连接应用等源头。(OpenAI Help Center) | 记忆必须用户可管理、可纠正、可删除、最好能追踪来源。 | 保留“用户记忆”，记录个人偏好、输出习惯、常用约定 |
+| Claude Code / CLAUDE.md + Auto Memory | 项目规则、工程约定、代码库经验沉淀 | CLAUDE.md 保存显式项目规则；Auto Memory 使用 MEMORY.md 作为索引，并把详细内容拆到 debugging.md、api-conventions.md 等主题文件。Claude Code 会在启动时加载 MEMORY.md 前 200 行或 25KB，主题文件按需读取；这些 memory 文件是 Markdown，可编辑、可删除。(Claude Platform Docs) | 记忆应该透明、文件化、可编辑、可版本化；同时要用“索引 + 主题文件”避免全量注入。 | 记忆正文落到 Markdown，用户可在页面编辑 |
+| LangGraph / LangMem | Agent 的会话状态、长期记忆、异步整理 | LangGraph 区分 thread-scoped short-term memory 和跨 session 的 long-term memory；长期记忆通过 namespace 隔离。LangMem 则提供从对话中抽取、合并、更新长期记忆的工具，并支持 hot path 写入和 background memory manager。(LangChain 文档) (LangChain AI) | 记忆需要分层：会话状态、用户长期记忆、Skill 规则、历史经验不是一类东西；写记忆也应区分同步和异步。 | 第一阶段不先上复杂检索框架，但借鉴其结构：主链路只记录候选记忆事件，后台异步整理；注入时按 user_id + skill_id + task_type 精准选择。 |
+| Mem0 | 通用 AI 应用 / Agent 的持久记忆层 | Mem0 把记忆分为 conversation、session、user、organizational 等层；通过 user_id、run_id、metadata 做作用域隔离，并在查询时合并召回不同层级的记忆。(Mem0) | 记忆的 scope 和生命周期应该是数据模型的一等字段，而不是临时写进 prompt。 | 当前先按用户和 Skill 精准注入，不先上复杂检索层 |
+| Zep | 企业级 Agent Memory、时序知识图谱、动态事实更新 | Zep 从聊天、业务数据、文档、JSON 等来源构建 temporal Context Graph；图中包含实体、关系和事实，事实变更后会 invalidating outdated facts，同时保留历史。它还会生成 token-efficient Context Block 给 Agent 使用。(Zep) | 对业务 Agent 很重要的一点是：记忆不能只追加，还要能处理事实过期、口径变化、历史保留、当前有效状态。 | 采用后台异步整理，避免影响主任务执行 |
+| ZK Data Agent | 团队业务 Skill 的持续复用 | 用户记忆 + Skill 核心记忆 + Skill 归档记忆 + 异步整理队列 + Markdown 可编辑。主任务执行时，只注入当前用户和当前 Skill 相关的高优先级记忆。 | 核心目标不是“什么都记住”，而是把反复纠正、反复复用、业务相关的经验沉淀到对应 Skill 上。 | 第一阶段采用轻量、可控方案：Markdown 可编辑、按 Skill 精准注入、后台异步整理、作用域隔离。 |
+
+### 2.3 Jupyter 工作区入口
+
+很多实际工作最终还是要落到远端环境里执行，例如：
+
+- 修改线上配置
+- 发起模型训练-cloudml / 分析训练结果
+- 读取 juiceFS 数据和产物
+
+平台支持把 Jupyter / 远端工作区绑定到当前用户或会话。Agent 在执行任务时，可以围绕当前工作区读写文件、运行脚本、整理结果。
+
+这个能力的价值在于：AI 不只停留在本地聊天和文本生成，而是可以进入真实工作环境，衔接已有的数据、脚本和训练流程。
+
+[图片]
+
+## 3. 我们在怎么用
+
+### 3.1 数据挖掘的例子
+
+会话链接：http://10.189.47.6/session/__LOCALID_0lj28j5i
+
+[图片]
+
+[图片]
+
+### 3.2 标签大师的例子
+
+会话链接：
+
+- http://10.189.47.6/session/__LOCALID_HP8pkRE
+- http://10.189.47.6/session/__LOCALID_bextxgj0
+
+[图片]
+
+### 3.3 线上正则干预的例子
+
+- http://10.189.47.6/session/__LOCALID_dgjr9shm
+
+## 4. 实际效果
+
+| 场景 | 原流程 | 当前流程 | 粗略提效 | 产出 |
+|---|---:|---:|---:|---|
+| 线上问题分析 / 评测集构建 | 1～2 天 | 30 分钟内 | 约 16～32 倍 | 目前已构造 5 个评测集，2000+ 数据 |
+| 数据构造与格式转换 | 1～2 小时 | 10 分钟内 | 约 6～12 倍 |  |
+| 线上配置修改（熟练工） | 20 分钟 | 2 分钟 | 约 10 倍 |  |
+
+[图片]
+
+## 5. 总结
+
+这次实践的起点很具体：提升中控 planning 模型迭代中的数据开发效率。
+
+做下来之后，我们验证了一个更通用的方向：很多 AI 提效能力，真正有复用价值的部分，往往不是某一次对话结果，而是背后的业务知识、执行脚本、流程经验和修正记录。
+
+中控数据 Agent 目前做的事情，就是把这些分散能力整理成团队可以复用的 Skill 和工具链：
+
+- 个人经验可以沉淀成 Skill。
+- Skill 可以被同步、启用和组合。
+- 使用过程中的修正经验可以进入记忆。
+- Jupyter 工作区让 Agent 能连接真实执行环境。
+
+后续继续迭代的重点可以放在三件事：
+
+- 扩展更多高频业务 Skill。
+- 提升 Skill 记忆和团队经验沉淀质量。
+- 把数据开发、评测、训练、线上问题分析这些流程打通得更稳定。

docs/technical-architecture/13-skill-evaluation-workbench.md

@@ -0,0 +1,451 @@
+# 13. Skill 评测实验台设计
+
+## 1. 背景
+
+当前快慢分流优化需要完成一条完整链路：
+
+```text
+已有评测集
+  -> 使用现成 Skill 判断快慢
+  -> 支持单条和批量运行
+  -> 对比人工标签和模型判断
+  -> 查看分歧及判定依据
+  -> 更新 Skill 或切换模型
+  -> 重新评测并比较变化
+```
+
+这里讨论的是离线优化和评估场景。评测过程允许 Agent 使用完整 Agent Loop，重点是判断质量、依据可追溯和标准持续收敛。线上快慢路由的耗时、成本和部署形态属于后续独立问题。
+
+本设计假设快慢分流 Skill 已经存在。实验台负责稳定调用和评估 Skill，不负责定义 Skill 内部的分类标准。
+
+## 2. 目标
+
+实验台需要支持：
+
+- 单条输入调用指定 Skill，返回判断、理由和执行记录。
+- 上传 CSV、XLSX 或 JSONL，配置字段映射后批量运行。
+- 选择模型和不可变的 Skill 版本。
+- 后台并发执行，支持暂停、继续、取消和失败重试。
+- 自动计算总体、分类别和分垂域指标。
+- 集中查看人工与模型分歧，并进行人工复核。
+- 查看模型声明的 Skill 判定依据和实际访问记录。
+- 对比不同 Skill 版本或模型版本的修正与退化。
+- 导出原始结果、复核结果和版本对比结果。
+
+实验台不承担：
+
+- 在线请求的实时路由。
+- 快系统能力或分类原则的编写。
+- 将每个普通 Skill 自动转成页面应用。
+- 修改现有聊天 Agent 的会话和运行流程。
+
+## 3. 总体架构
+
+```mermaid
+flowchart LR
+    UI["快慢分流评测页面"] --> EXP["Evaluation Service"]
+    API["Skill Run API"] --> EXP
+
+    EXP --> MAP["Dataset Mapping"]
+    EXP --> SNAP["Skill Snapshot"]
+    EXP --> SCHED["Batch Scheduler"]
+
+    SCHED --> RUNNER["Headless Skill Runner"]
+    RUNNER --> AGENT["LocalCodingAgent"]
+    RUNNER --> MODEL["指定模型"]
+    RUNNER --> SKILL["固定 Skill 快照"]
+
+    EXP --> STORE["Evaluation Store"]
+    STORE --> RESULT["指标 / 分歧 / 复核 / 版本对比"]
+```
+
+新增能力由五个部分组成：
+
+1. `Evaluation Service`：管理数据集、实验配置、运行和结果。
+2. `Dataset Mapping`：把不同格式的数据转换为统一 case。
+3. `Skill Snapshot`：固定每次实验使用的 Skill 内容。
+4. `Headless Skill Runner`：使用现有 Agent Core 在后台执行指定 Skill。
+5. `Batch Scheduler`：拆分、并发和恢复批量任务。
+
+## 4. 与现有 Agent 的隔离
+
+现有聊天链路保持不变：
+
+```text
+/api/chat
+  -> AgentState
+  -> account/session-scoped LocalCodingAgent
+  -> 聊天 session 和活动流
+```
+
+评测链路新增独立入口：
+
+```text
+/api/evaluations
+  -> EvaluationRuntime
+  -> evaluation-scoped LocalCodingAgent
+  -> 评测任务、case 结果和内部执行记录
+```
+
+两条链路可以复用：
+
+- `LocalCodingAgent`
+- `ModelConfig`
+- Skill loader
+- Tool handler
+- `OutputSchemaConfig`
+- 模型兼容层
+
+评测链路必须独立管理：
+
+- Agent 实例。
+- Session 命名空间。
+- 数据库记录。
+- 运行队列。
+- 并发和资源配额。
+- Skill 快照。
+
+评测产生的内部 Agent session 不进入左侧聊天会话列表，也不读取聊天历史、用户记忆或当前聊天 session 状态。
+
+代码隔离只能保证功能不互相污染。批量任务仍可能争用模型服务资源，因此评测任务需要独立 worker pool、并发上限和低于聊天请求的调度优先级。有条件时可以为评测配置独立模型地址。
+
+## 5. 数据集字段映射
+
+评测文件格式由数据来源决定，不应写进 Skill，也不应要求用户临时组织成提示词。
+
+上传文件后，系统先读取表头和样本，用户把源字段映射到统一字段：
+
+| 统一字段 | 说明 | 必填 |
+|---|---|---|
+| `case_id` | case 唯一标识 | 否，可自动生成 |
+| `query` | 当前用户 query | 是 |
+| `gold_route` | 人工快慢标签 | 评测时必填 |
+| `history` | 对话历史 | 否 |
+| `context` | 设备、位置等上下文 | 否 |
+| `domain` | 垂域 | 否 |
+| `request_id` | 线上 RID | 否 |
+| `metadata` | 其他保留字段 | 否 |
+| `source_row` | 原始行 | 自动保留 |
+
+页面需要支持：
+
+- 列名自动推荐。
+- 手动选择源字段。
+- JSON path，例如 `data.query`。
+- 标签值转换，例如 `快/慢`、`0/1`、`fast/slow`。
+- 必要的字段组合和简单转换。
+- 转换后样本预览。
+- 缺失字段、非法标签和重复 ID 校验。
+
+确认后的映射保存为可复用的 `Dataset Mapping Profile`。Skill Runner 只接收统一 case，不感知原始文件的列名和格式。
+
+统一 case 示例：
+
+```json
+{
+  "case_id": "001",
+  "query": "到目的地电量够不够",
+  "gold_route": "fast",
+  "history": [],
+  "context": {},
+  "domain": "地图导航",
+  "request_id": "",
+  "metadata": {},
+  "source_row": {}
+}
+```
+
+## 6. Skill 版本与快照
+
+实验不能只记录 Skill 名称。Skill 内容会持续更新，同一个名称在不同时间可能代表不同标准。
+
+每次创建实验时生成不可变快照，并记录：
+
+```json
+{
+  "skill_name": "fast-slow-routing",
+  "repository": "zk-data-agent",
+  "git_commit": "abc1234",
+  "content_hash": "sha256:...",
+  "snapshot_id": "skill_snapshot_xxx",
+  "status": "committed"
+}
+```
+
+版本来源可以是：
+
+- 当前部署版本。
+- 指定 Git commit。
+- 指定 Git tag。
+- 尚未提交的草稿快照。
+
+快照建议存放在：
+
+```text
+.port_sessions/evaluations/skill_snapshots/
+  <skill-name>/
+    <commit-or-content-hash>/
+```
+
+运行旧版本时不能切换当前项目分支，也不能覆盖线上 Skill。对于 Git 中的历史版本，可以通过 `git archive` 提取单个 Skill 目录到快照缓存。
+
+草稿版本允许参与实验，但页面必须明确标记“未提交”，并使用内容哈希保证同一次实验可复现。
+
+评测 Agent 始终从快照加载 Skill；聊天 Agent 继续从当前部署目录加载 Skill，两者互不影响。
+
+## 7. Headless Skill Runner
+
+`Headless` 表示没有聊天页面，不表示减少 Agent 能力。
+
+Runner 接收：
+
+```json
+{
+  "skill_snapshot_id": "skill_snapshot_xxx",
+  "model_id": "model_xxx",
+  "case": {
+    "case_id": "001",
+    "query": "到目的地电量够不够"
+  }
+}
+```
+
+Runner 的执行过程：
+
+1. 从快照读取指定 Skill。
+2. 创建独立的评测 Agent 实例和 session。
+3. 直接把指定 Skill 注入当前任务，不等待模型自行召回 Skill。
+4. 把统一 case 作为任务输入。
+5. 按 Skill 约定执行完整 Agent Loop。
+6. 保存最终结果、执行事件、工具调用和用量。
+7. 将最终输出归一化为评测结果。
+
+统一评测结果至少包含：
+
+```json
+{
+  "case_id": "001",
+  "prediction": "fast",
+  "reason": "当前快系统已有对应能力",
+  "confidence": 0.95,
+  "evidence_refs": [
+    "knowledge/地图能力.md#到达电量预估"
+  ],
+  "status": "completed",
+  "raw_output": "",
+  "agent_session_id": ""
+}
+```
+
+单条判断创建一个 Agent Run。批量判断把每条 case 拆成独立 Agent Run，避免样本间上下文污染。
+
+## 8. 批量调度
+
+批量调度器负责：
+
+- 把数据集拆成独立 case 任务。
+- 使用受控并发执行。
+- 保存每条 case 的状态和重试次数。
+- 单条失败不终止整个实验。
+- 支持暂停、继续和取消。
+- 服务重启后从数据库恢复未完成任务。
+- 只重跑失败项、分歧项或用户选中的 case。
+
+任务状态：
+
+```text
+pending
+running
+completed
+failed
+cancelled
+```
+
+聊天请求和评测请求需要独立并发池。评测批量任务默认低优先级，避免影响当前在线使用者。
+
+## 9. 判定依据与可追溯性
+
+页面需要回答“模型依据 Skill 的哪部分作出判断”，但不能把完整模型思考当作可靠因果证据。
+
+实验台展示两类事实：
+
+1. **实际访问记录**：Agent 读取了哪些 Skill 文件、知识文件和章节。
+2. **模型声明的判定依据**：最终结果中的 `evidence_refs`。
+
+生成 Skill 快照时，系统为 Markdown 标题和知识文件建立稳定引用，例如：
+
+```text
+SKILL.md#判断流程
+knowledge/地图能力.md#到达电量预估
+knowledge/边界.md#开放式行程规划
+```
+
+评测 Runner 的统一任务约定要求最终结果返回引用。页面将引用解析为可点击的原文片段。
+
+如果模型没有引用明确规则，页面展示“未引用具体依据”，并允许按此条件筛选。这类 case 本身就是标准缺失或模型未正确使用 Skill 的候选问题。
+
+## 10. 页面信息架构
+
+页面采用“实验配置、结果页签、单条详情”三层结构。
+
+### 10.1 实验配置
+
+顶部配置区包含：
+
+- 数据集。
+- 字段映射。
+- Skill 和版本。
+- 模型。
+- case 数量。
+- 并发配置。
+- 开始、暂停、继续、取消、克隆实验。
+
+运行开始后配置区折叠，避免占用结果空间。
+
+### 10.2 结果页签
+
+| 页签 | 主要内容 |
+|---|---|
+| 概览 | 运行进度、总体指标、混淆矩阵、快慢分布、分垂域指标 |
+| 分歧 | 人工标签与模型判断不一致、低置信、无明确依据的 case |
+| 全部 Case | 全量结果、状态、筛选、排序和导出 |
+| 版本对比 | 不同 Skill 或模型实验之间的修正、退化和指标变化 |
+
+### 10.3 单条详情抽屉
+
+点击一条 case 后展示：
+
+1. 原始数据和统一后的输入。
+2. 人工标签、模型判断、置信度和理由。
+3. Skill 判定依据及原文片段。
+4. 实际读取文件和工具调用。
+5. 默认折叠的完整 Agent 执行记录。
+6. 人工复核结果和备注。
+
+人工复核至少支持：
+
+- 人工原标签正确。
+- 模型判断正确。
+- 分类标准存在歧义。
+- 快系统能力信息缺失。
+- 暂不确定。
+
+## 11. 实验与版本对比
+
+一个实验由以下对象共同确定：
+
+```text
+数据集版本
++ 字段映射版本
++ Skill 快照
++ 模型版本
++ 执行参数
+```
+
+同一数据集可以克隆实验并替换 Skill 或模型：
+
+```text
+实验 A：Skill abc1234 + 模型 A
+实验 B：Skill def5678 + 模型 A
+实验 C：Skill def5678 + 模型 B
+```
+
+版本对比需要展示：
+
+- 总体指标变化。
+- 各垂域指标变化。
+- 从错误变正确的 case。
+- 从正确变错误的 case。
+- 判断未变化但理由变化的 case。
+- 新增或消失的无依据结果。
+
+这样才能形成“评测、发现分歧、更新 Skill、重新评测”的优化闭环。
+
+## 12. 数据存储
+
+评测运行状态适合使用独立 SQLite 数据库，例如：
+
+```text
+.port_sessions/evaluations/evaluations.db
+```
+
+核心实体：
+
+```text
+evaluation_datasets
+dataset_mapping_profiles
+skill_snapshots
+evaluation_experiments
+evaluation_cases
+evaluation_case_runs
+evaluation_results
+evaluation_reviews
+```
+
+数据库保存可查询状态、关系和人工复核结果。上传原文件、Skill 快照、导出文件和较大的 Agent transcript 保存在实验目录，数据库记录路径和摘要。
+
+## 13. API 轮廓
+
+```text
+POST /api/evaluations/datasets
+POST /api/evaluations/datasets/{id}/mapping
+
+POST /api/evaluations/skill-snapshots
+GET  /api/evaluations/skill-snapshots
+
+POST /api/evaluations
+GET  /api/evaluations/{id}
+POST /api/evaluations/{id}/start
+POST /api/evaluations/{id}/pause
+POST /api/evaluations/{id}/resume
+POST /api/evaluations/{id}/cancel
+
+GET  /api/evaluations/{id}/results
+POST /api/evaluations/{id}/retry
+POST /api/evaluations/{id}/reviews
+GET  /api/evaluations/{id}/export
+
+POST /api/evaluations/compare
+```
+
+API 同时服务 Web 页面和外部批量调用。单条调用可以创建只包含一个 case 的轻量实验，复用相同运行与记录机制。
+
+## 14. 实施顺序
+
+### P1：后台单条运行
+
+- Skill 快照。
+- Headless Skill Runner。
+- 单条输入和统一结果。
+- 独立于聊天 session 的运行记录。
+
+### P2：数据集与批量调度
+
+- 文件上传和字段映射。
+- 批量拆分、并发、暂停、恢复和重试。
+- 基础指标和导出。
+
+### P3：评测页面
+
+- 实验配置。
+- 概览、分歧、全部 Case。
+- 单条详情和人工复核。
+- Skill 依据引用。
+
+### P4：版本对比
+
+- 克隆实验。
+- Skill 和模型版本矩阵。
+- 修正、退化和理由变化分析。
+
+## 15. 验收标准
+
+- 开启评测任务后，现有聊天 Agent 可以正常创建、继续和停止会话。
+- 600 条批量任务不会出现在聊天会话列表。
+- 不同输入格式可以通过字段映射转换为统一 case。
+- 每次实验可以确认具体 Skill commit 或内容哈希。
+- Skill 更新后，历史实验仍能读取原快照并复现配置。
+- 单条失败不会中断全量任务，服务重启后可以恢复。
+- 页面可以定位人工与模型分歧，并查看简短理由和 Skill 引用。
+- 可以比较两个实验的修正项和退化项。
+- 批量评测的并发不会明显拖慢聊天 Agent。

docs/technical-architecture/README.md

@@ -16,6 +16,7 @@
 9. [外部系统 Skill：ELK、SQL、模型迭代](09-external-skills.md)
 10. [Agent 记忆机制调研与对比](10-memory-research.md)
 11. [运行中输入队列与 Runtime Guidance 注入](12-runtime-guidance-queue.md)
+12. [Skill 评测实验台设计](13-skill-evaluation-workbench.md)
 
 ## 一句话定位
 

frontend/app/app/annotations/[name]/route.ts

@@ -0,0 +1,69 @@
+import { readFile } from "node:fs/promises";
+import { resolveAnnotationPage } from "@/lib/annotation-pages";
+
+export const runtime = "nodejs";
+export const dynamic = "force-dynamic";
+
+export async function GET(
+	_request: Request,
+	{ params }: { params: Promise<{ name: string }> },
+) {
+	const { name } = await params;
+	const filePath = resolveAnnotationPage(name);
+	if (!filePath)
+		return new Response("Invalid annotation page", { status: 400 });
+
+	try {
+		const content = await readFile(filePath, "utf8");
+		return new Response(injectSubmissionBridge(content, name), {
+			headers: {
+				"cache-control": "no-store",
+				"content-type": "text/html; charset=utf-8",
+				"x-content-type-options": "nosniff",
+			},
+		});
+	} catch {
+		return new Response("Annotation page not found", { status: 404 });
+	}
+}
+
+function injectSubmissionBridge(content: string, pageName: string) {
+	if (!content.includes("</body>") || !content.includes("mergedRows")) {
+		return content;
+	}
+	const safePageName = JSON.stringify(pageName).replaceAll("<", "\\u003c");
+	const bridge = `
+<script>
+  window.__annotationSubmissionBridge = {
+    pageName: ${safePageName},
+    read: () => mergedRows().map(row => ({
+      requestId: String(row.request_id || ""),
+      result: String(row.GSB || ""),
+      note: String(row["人工备注"] || "")
+    })),
+    apply: items => {
+      for (const row of rows || []) {
+        if (row?.request_id) delete saved[row.request_id];
+      }
+      for (const item of items || []) {
+        if (!item || !item.requestId) continue;
+        saved[item.requestId] = { GSB: item.result || "", note: item.note || "" };
+      }
+      persist();
+      render();
+    },
+    applyMissing: items => {
+      for (const item of items || []) {
+        if (!item || !item.requestId || saved[item.requestId]) continue;
+        saved[item.requestId] = { GSB: item.result || "", note: item.note || "" };
+      }
+      persist();
+      render();
+    },
+    notify: message => toast(message)
+  };
+</script>
+<script src="/annotation-submissions.js"></script>
+`;
+	return content.replace(/<\/body>/i, `${bridge}</body>`);
+}

frontend/app/app/annotations/page.tsx

@@ -0,0 +1,167 @@
+"use client";
+
+import {
+	ArrowLeftIcon,
+	ExternalLinkIcon,
+	FileTextIcon,
+	RefreshCwIcon,
+	SearchIcon,
+} from "lucide-react";
+import { useCallback, useEffect, useMemo, useState } from "react";
+import { Button } from "@/components/ui/button";
+import { Input } from "@/components/ui/input";
+
+type AnnotationPageInfo = {
+	name: string;
+	title: string;
+	size: number;
+	modifiedAt: string;
+	href: string;
+};
+
+export default function AnnotationsPage() {
+	const [items, setItems] = useState<AnnotationPageInfo[]>([]);
+	const [query, setQuery] = useState("");
+	const [isLoading, setIsLoading] = useState(true);
+	const [error, setError] = useState("");
+
+	const refresh = useCallback(async () => {
+		setIsLoading(true);
+		setError("");
+		try {
+			const response = await fetch("/api/annotations", { cache: "no-store" });
+			const payload = (await response.json()) as {
+				items?: AnnotationPageInfo[];
+				error?: string;
+			};
+			if (!response.ok) throw new Error(payload.error ?? "加载失败");
+			setItems(payload.items ?? []);
+		} catch (reason) {
+			setError(reason instanceof Error ? reason.message : "加载失败");
+		} finally {
+			setIsLoading(false);
+		}
+	}, []);
+
+	useEffect(() => {
+		void refresh();
+	}, [refresh]);
+
+	const filteredItems = useMemo(() => {
+		const keyword = query.trim().toLocaleLowerCase("zh-CN");
+		if (!keyword) return items;
+		return items.filter((item) =>
+			`${item.title}\n${item.name}`
+				.toLocaleLowerCase("zh-CN")
+				.includes(keyword),
+		);
+	}, [items, query]);
+
+	return (
+		<main className="min-h-dvh bg-[#f6f8fb] text-slate-900">
+			<header className="border-slate-200 border-b bg-white">
+				<div className="mx-auto flex w-full max-w-6xl items-center gap-3 px-5 py-4 sm:px-8">
+					<Button variant="ghost" size="icon" asChild title="返回工作台">
+						<a href="/">
+							<ArrowLeftIcon />
+						</a>
+					</Button>
+					<div className="min-w-0 flex-1">
+						<h1 className="font-semibold text-xl">标注页面</h1>
+						<p className="mt-0.5 text-slate-500 text-sm">
+							{isLoading ? "正在读取" : `共 ${items.length} 个页面`}
+						</p>
+					</div>
+					<Button
+						variant="outline"
+						size="sm"
+						onClick={() => void refresh()}
+						disabled={isLoading}
+					>
+						<RefreshCwIcon className={isLoading ? "animate-spin" : ""} />
+						刷新
+					</Button>
+				</div>
+			</header>
+
+			<section className="mx-auto w-full max-w-6xl px-5 py-6 sm:px-8">
+				<div className="relative mb-4 max-w-md">
+					<SearchIcon className="-translate-y-1/2 pointer-events-none absolute top-1/2 left-3 size-4 text-slate-400" />
+					<Input
+						value={query}
+						onChange={(event) => setQuery(event.target.value)}
+						placeholder="搜索页面"
+						className="bg-white pl-9"
+					/>
+				</div>
+
+				<div className="overflow-hidden border-slate-200 border-y bg-white sm:rounded-md sm:border">
+					<div className="hidden grid-cols-[minmax(0,1fr)_140px_100px_44px] gap-4 border-slate-200 border-b bg-slate-50 px-4 py-2 font-medium text-slate-500 text-xs sm:grid">
+						<span>页面</span>
+						<span>更新时间</span>
+						<span>大小</span>
+						<span />
+					</div>
+
+					{error ? (
+						<div className="px-4 py-10 text-center text-red-600 text-sm">
+							{error}
+						</div>
+					) : !isLoading && filteredItems.length === 0 ? (
+						<div className="px-4 py-14 text-center text-slate-500 text-sm">
+							没有匹配的标注页面
+						</div>
+					) : (
+						<ul className="divide-y divide-slate-100">
+							{filteredItems.map((item) => (
+								<li key={item.name}>
+									<a
+										href={item.href}
+										target="_blank"
+										rel="noreferrer"
+										className="group grid min-h-18 grid-cols-[36px_minmax(0,1fr)_32px] items-center gap-3 px-4 py-3 transition-colors hover:bg-blue-50/60 sm:grid-cols-[36px_minmax(0,1fr)_140px_100px_32px]"
+									>
+										<span className="grid size-9 place-items-center rounded-md bg-blue-50 text-blue-600">
+											<FileTextIcon className="size-4" />
+										</span>
+										<span className="min-w-0">
+											<strong className="block truncate font-medium text-sm">
+												{item.title}
+											</strong>
+											<span className="mt-1 block truncate text-slate-500 text-xs">
+												{item.name}
+											</span>
+										</span>
+										<span className="hidden text-slate-500 text-sm sm:block">
+											{formatDate(item.modifiedAt)}
+										</span>
+										<span className="hidden text-slate-500 text-sm sm:block">
+											{formatSize(item.size)}
+										</span>
+										<ExternalLinkIcon className="size-4 text-slate-400 transition-colors group-hover:text-blue-600" />
+									</a>
+								</li>
+							))}
+						</ul>
+					)}
+				</div>
+			</section>
+		</main>
+	);
+}
+
+function formatDate(value: string) {
+	return new Intl.DateTimeFormat("zh-CN", {
+		month: "2-digit",
+		day: "2-digit",
+		hour: "2-digit",
+		minute: "2-digit",
+		hour12: false,
+	}).format(new Date(value));
+}
+
+function formatSize(bytes: number) {
+	if (bytes < 1024) return `${bytes} B`;
+	if (bytes < 1024 * 1024) return `${(bytes / 1024).toFixed(1)} KB`;
+	return `${(bytes / 1024 / 1024).toFixed(1)} MB`;
+}

frontend/app/app/api/annotations/[name]/submissions/route.ts

@@ -0,0 +1,62 @@
+import {
+	createAnnotationSubmission,
+	readAnnotationSubmission,
+	readAnnotationSubmissions,
+} from "@/lib/annotation-pages";
+import { getCurrentAccount } from "@/lib/claw-auth";
+
+export const runtime = "nodejs";
+export const dynamic = "force-dynamic";
+
+export async function GET(
+	request: Request,
+	{ params }: { params: Promise<{ name: string }> },
+) {
+	try {
+		const { name } = await params;
+		const revisionId = new URL(request.url).searchParams.get("revision");
+		if (revisionId) {
+			const submission = await readAnnotationSubmission(name, revisionId);
+			if (!submission) return errorResponse("提交记录不存在", 404);
+			return jsonResponse({ submission });
+		}
+		return jsonResponse(await readAnnotationSubmissions(name));
+	} catch (error) {
+		return errorResponse(errorMessage(error), 404);
+	}
+}
+
+export async function POST(
+	request: Request,
+	{ params }: { params: Promise<{ name: string }> },
+) {
+	const account = await getCurrentAccount();
+	if (!account) return errorResponse("请先登录工作台再提交", 401);
+	try {
+		const { name } = await params;
+		const body = (await request.json()) as { annotations?: unknown };
+		const submission = await createAnnotationSubmission(
+			name,
+			account,
+			body.annotations,
+		);
+		return jsonResponse({ submission }, 201);
+	} catch (error) {
+		return errorResponse(errorMessage(error), 400);
+	}
+}
+
+function jsonResponse(value: unknown, status = 200) {
+	return Response.json(value, {
+		status,
+		headers: { "cache-control": "no-store" },
+	});
+}
+
+function errorResponse(error: string, status: number) {
+	return jsonResponse({ error }, status);
+}
+
+function errorMessage(error: unknown) {
+	return error instanceof Error ? error.message : "操作失败";
+}

frontend/app/app/api/annotations/route.ts

@@ -0,0 +1,21 @@
+import { listAnnotationPages } from "@/lib/annotation-pages";
+
+export const runtime = "nodejs";
+export const dynamic = "force-dynamic";
+
+export async function GET() {
+	try {
+		const items = await listAnnotationPages();
+		return Response.json(
+			{ items },
+			{ headers: { "cache-control": "no-store" } },
+		);
+	} catch (error) {
+		return Response.json(
+			{
+				error: error instanceof Error ? error.message : "无法读取标注页面目录",
+			},
+			{ status: 500, headers: { "cache-control": "no-store" } },
+		);
+	}
+}

frontend/app/app/api/claw/evaluations/[...path]/route.ts

@@ -0,0 +1,78 @@
+import { getCurrentAccount } from "@/lib/claw-auth";
+
+const CLAW_API_URL = process.env.CLAW_API_URL ?? "http://127.0.0.1:8765";
+
+type RouteContext = {
+	params: Promise<{ path: string[] }>;
+};
+
+export async function GET(request: Request, context: RouteContext) {
+	return proxyEvaluationRequest("GET", request, context);
+}
+
+export async function POST(request: Request, context: RouteContext) {
+	return proxyEvaluationRequest("POST", request, context);
+}
+
+export async function PATCH(request: Request, context: RouteContext) {
+	return proxyEvaluationRequest("PATCH", request, context);
+}
+
+async function proxyEvaluationRequest(
+	method: "GET" | "POST" | "PATCH",
+	request: Request,
+	context: RouteContext,
+) {
+	const account = await getCurrentAccount();
+	if (!account)
+		return Response.json({ detail: "unauthorized" }, { status: 401 });
+
+	const { path } = await context.params;
+	const incomingUrl = new URL(request.url);
+	const targetUrl = new URL(
+		`${CLAW_API_URL}/api/evaluations/${path.map(encodeURIComponent).join("/")}`,
+	);
+	for (const [key, value] of incomingUrl.searchParams.entries()) {
+		targetUrl.searchParams.append(key, value);
+	}
+
+	const headers: Record<string, string> = {};
+	let body: string | undefined;
+	if (method === "GET") {
+		targetUrl.searchParams.set("account_id", account.id);
+	} else {
+		const payload = (await request.json()) as Record<string, unknown>;
+		headers["content-type"] = "application/json";
+		body = JSON.stringify({ ...payload, account_id: account.id });
+	}
+
+	try {
+		const response = await fetch(targetUrl, {
+			method,
+			headers,
+			body,
+			cache: "no-store",
+		});
+		const responseHeaders = new Headers();
+		responseHeaders.set(
+			"content-type",
+			response.headers.get("content-type") ?? "application/json",
+		);
+		const disposition = response.headers.get("content-disposition");
+		if (disposition) responseHeaders.set("content-disposition", disposition);
+		return new Response(await response.arrayBuffer(), {
+			status: response.status,
+			headers: responseHeaders,
+		});
+	} catch (error) {
+		return Response.json(
+			{
+				detail:
+					error instanceof Error
+						? `Unable to reach Claw backend: ${error.message}`
+						: "Unable to reach Claw backend",
+			},
+			{ status: 502 },
+		);
+	}
+}

frontend/app/app/api/claw/integrations/feishu/route.ts

@@ -20,6 +20,7 @@ export async function POST(request: Request) {
 
 	const body = (await request.json().catch(() => ({}))) as {
 		action?: unknown;
+		force?: unknown;
 	};
 	const action = typeof body.action === "string" ? body.action : "login";
 	const endpoint =
@@ -30,7 +31,10 @@ export async function POST(request: Request) {
 	const response = await fetch(`${CLAW_API_URL}${endpoint}`, {
 		method: "POST",
 		headers: { "content-type": "application/json" },
-		body: JSON.stringify({ account_id: account.id }),
+		body: JSON.stringify({
+			account_id: account.id,
+			...(action !== "logout" && body.force === true ? { force: true } : {}),
+		}),
 		cache: "no-store",
 	});
 	return proxyResponse(response);

frontend/app/components/assistant-ui/activity-panel.tsx

@@ -574,10 +574,20 @@ function SessionFileRow({ file }: { file: SessionFile }) {
 				unknown
 			>;
 			if (response.status === 409) {
+				const detail = payload.detail;
+				const forceLogin =
+					Boolean(
+						detail &&
+							typeof detail === "object" &&
+							(detail as { force_login?: unknown }).force_login,
+					) ||
+					(detail &&
+						typeof detail === "object" &&
+						(detail as { code?: unknown }).code === "feishu_auth_expired");
 				const loginResponse = await fetch("/api/claw/integrations/feishu", {
 					method: "POST",
 					headers: { "content-type": "application/json" },
-					body: JSON.stringify({ action: "login" }),
+					body: JSON.stringify({ action: "login", force: forceLogin }),
 				});
 				const loginPayload = (await loginResponse
 					.json()

frontend/app/components/assistant-ui/thread-list.tsx

@@ -6,6 +6,7 @@ import {
 import { ThreadListPrimitive } from "@assistant-ui/react";
 import type { UIMessage } from "ai";
 import {
+	FlaskConicalIcon,
 	Loader2Icon,
 	MoreHorizontalIcon,
 	PencilIcon,
@@ -148,6 +149,7 @@ export const ThreadList: FC = () => {
 	return (
 		<ThreadListPrimitive.Root className="aui-root aui-thread-list-root flex min-h-0 flex-1 flex-col gap-1">
 			<ThreadListNew />
+			<ThreadListEvaluation />
 			<JupyterWorkspaceList />
 			<ClawSessionList />
 		</ThreadListPrimitive.Root>
@@ -174,6 +176,21 @@ const ThreadListNew: FC = () => {
 	);
 };
 
+const ThreadListEvaluation: FC = () => {
+	return (
+		<Button
+			variant="ghost"
+			className="h-9 w-full justify-start gap-2 rounded-lg px-3 text-sm"
+			asChild
+		>
+			<a href="/evaluations">
+				<FlaskConicalIcon className="size-4" />
+				Skill 评测
+			</a>
+		</Button>
+	);
+};
+
 type JupyterWorkspaceEntry = {
 	id: string;
 	label?: string;
@@ -1220,7 +1237,7 @@ function resolvePendingPrompt(
 			: "";
 	if (!pendingPrompt) return null;
 	for (const message of messages) {
-		if (!message || message.role !== "user") continue;
+		if (message?.role !== "user") continue;
 		const content = cleanStoredContent(message.content ?? "").trim();
 		if (!content || content.trimStart().startsWith("<system-reminder>"))
 			continue;

frontend/app/components/assistant-ui/threadlist-sidebar.tsx

@@ -5,6 +5,7 @@ import {
 	BrainIcon,
 	ChevronDownIcon,
 	Clock3Icon,
+	FlaskConicalIcon,
 	LogOutIcon,
 	PanelLeftOpenIcon,
 	SaveIcon,
@@ -215,6 +216,12 @@ function CollapsedSidebarRail({
 				>
 					<SquarePenIcon className="size-4" />
 				</CollapsedIconButton>
+				<CollapsedIconButton
+					label="Skill 评测"
+					onClick={() => window.location.assign("/evaluations")}
+				>
+					<FlaskConicalIcon className="size-4" />
+				</CollapsedIconButton>
 				<CollapsedSessionSearch />
 				<CollapsedRecentSessions />
 			</div>

frontend/app/lib/annotation-pages.ts

@@ -0,0 +1,297 @@
+import { createHash, randomUUID } from "node:crypto";
+import {
+	mkdir,
+	open,
+	readdir,
+	readFile,
+	rename,
+	stat,
+	writeFile,
+} from "node:fs/promises";
+import path from "node:path";
+
+export type AnnotationPageInfo = {
+	name: string;
+	title: string;
+	size: number;
+	modifiedAt: string;
+	href: string;
+};
+
+export type AnnotationValue = {
+	requestId: string;
+	result: "G" | "S" | "B" | "";
+	note: string;
+};
+
+export type AnnotationSubmission = {
+	id: string;
+	version: number;
+	pageName: string;
+	submittedAt: string;
+	submittedBy: { id: string; username: string };
+	summary: {
+		total: number;
+		completed: number;
+		G: number;
+		S: number;
+		B: number;
+	};
+	annotations: AnnotationValue[];
+};
+
+export type AnnotationSubmissionSummary = Omit<
+	AnnotationSubmission,
+	"annotations"
+>;
+
+const annotationPagesRoot = path.resolve(
+	/* turbopackIgnore: true */
+	process.env.ANNOTATION_PAGES_DIR?.trim() || "/home/mi/annotation-pages",
+);
+const annotationSubmissionsRoot = path.resolve(
+	/* turbopackIgnore: true */
+	process.env.ANNOTATION_SUBMISSIONS_DIR?.trim() ||
+		path.join(annotationPagesRoot, ".submissions"),
+);
+const submissionWriteQueues = new Map<string, Promise<unknown>>();
+
+export async function listAnnotationPages(): Promise<AnnotationPageInfo[]> {
+	await mkdir(annotationPagesRoot, { recursive: true });
+	const entries = await readdir(annotationPagesRoot, { withFileTypes: true });
+	const pages = await Promise.all(
+		entries
+			.filter((entry) => entry.isFile() && isSafeHtmlName(entry.name))
+			.map(async (entry) => {
+				const filePath = path.join(annotationPagesRoot, entry.name);
+				const [fileStat, title] = await Promise.all([
+					stat(filePath),
+					readHtmlTitle(filePath, entry.name),
+				]);
+				return {
+					name: entry.name,
+					title,
+					size: fileStat.size,
+					modifiedAt: fileStat.mtime.toISOString(),
+					href: `/annotations/${encodeURIComponent(entry.name)}`,
+				};
+			}),
+	);
+
+	return pages.sort(
+		(left, right) =>
+			Date.parse(right.modifiedAt) - Date.parse(left.modifiedAt) ||
+			left.name.localeCompare(right.name, "zh-CN"),
+	);
+}
+
+export function resolveAnnotationPage(name: string): string | null {
+	if (!isSafeHtmlName(name)) return null;
+	const resolved = path.resolve(annotationPagesRoot, name);
+	if (path.dirname(resolved) !== annotationPagesRoot) return null;
+	return resolved;
+}
+
+export async function readAnnotationSubmissions(pageName: string) {
+	const pagePath = resolveAnnotationPage(pageName);
+	if (!pagePath) throw new Error("无效的标注页面");
+	await stat(pagePath);
+
+	const pageRoot = submissionPageRoot(pageName);
+	const latest = await readSubmissionFile(path.join(pageRoot, "latest.json"));
+	const revisionsRoot = path.join(pageRoot, "revisions");
+	let names: string[] = [];
+	try {
+		names = await readdir(revisionsRoot);
+	} catch {
+		return { latest, history: [] as AnnotationSubmissionSummary[] };
+	}
+
+	const history = (
+		await Promise.all(
+			names
+				.filter((name) => name.endsWith(".json"))
+				.map((name) => readSubmissionFile(path.join(revisionsRoot, name))),
+		)
+	)
+		.filter((item): item is AnnotationSubmission => item !== null)
+		.sort((left, right) => right.version - left.version)
+		.slice(0, 100)
+		.map(withoutAnnotations);
+
+	return { latest, history };
+}
+
+export async function readAnnotationSubmission(
+	pageName: string,
+	submissionId: string,
+) {
+	if (!/^[a-zA-Z0-9_-]{8,120}$/.test(submissionId)) return null;
+	const pagePath = resolveAnnotationPage(pageName);
+	if (!pagePath) throw new Error("无效的标注页面");
+	await stat(pagePath);
+	return readSubmissionFile(
+		path.join(
+			submissionPageRoot(pageName),
+			"revisions",
+			`${submissionId}.json`,
+		),
+	);
+}
+
+export async function createAnnotationSubmission(
+	pageName: string,
+	account: { id: string; username: string },
+	input: unknown,
+) {
+	const pagePath = resolveAnnotationPage(pageName);
+	if (!pagePath) throw new Error("无效的标注页面");
+	await stat(pagePath);
+	const annotations = normalizeAnnotations(input);
+
+	return enqueueSubmissionWrite(pageName, async () => {
+		const pageRoot = submissionPageRoot(pageName);
+		const revisionsRoot = path.join(pageRoot, "revisions");
+		await mkdir(revisionsRoot, { recursive: true });
+		const latest = await readSubmissionFile(path.join(pageRoot, "latest.json"));
+		const submittedAt = new Date().toISOString();
+		const id = `${submittedAt.replace(/\D/g, "").slice(0, 17)}-${randomUUID().slice(0, 8)}`;
+		const submission: AnnotationSubmission = {
+			id,
+			version: (latest?.version ?? 0) + 1,
+			pageName,
+			submittedAt,
+			submittedBy: account,
+			summary: summarizeAnnotations(annotations),
+			annotations,
+		};
+		await atomicWriteJson(path.join(revisionsRoot, `${id}.json`), submission);
+		await atomicWriteJson(path.join(pageRoot, "latest.json"), submission);
+		return submission;
+	});
+}
+
+function isSafeHtmlName(name: string): boolean {
+	return (
+		name === path.basename(name) &&
+		!name.startsWith(".") &&
+		/\.html?$/i.test(name)
+	);
+}
+
+function submissionPageRoot(pageName: string) {
+	const digest = createHash("sha256")
+		.update(pageName)
+		.digest("hex")
+		.slice(0, 24);
+	return path.join(annotationSubmissionsRoot, digest);
+}
+
+function normalizeAnnotations(input: unknown): AnnotationValue[] {
+	if (!Array.isArray(input)) throw new Error("annotations 必须是数组");
+	if (input.length > 20_000) throw new Error("单次提交最多包含 20000 条标注");
+	const seen = new Set<string>();
+	return input.map((raw, index) => {
+		if (!raw || typeof raw !== "object") {
+			throw new Error(`第 ${index + 1} 条标注格式错误`);
+		}
+		const value = raw as Record<string, unknown>;
+		const requestId = String(value.requestId ?? value.request_id ?? "").trim();
+		const result = String(value.result ?? value.GSB ?? "").trim();
+		const note = String(value.note ?? value.人工备注 ?? "").trim();
+		if (!requestId || requestId.length > 200) {
+			throw new Error(`第 ${index + 1} 条缺少有效 request_id`);
+		}
+		if (seen.has(requestId)) throw new Error(`request_id 重复：${requestId}`);
+		if (!["", "G", "S", "B"].includes(result)) {
+			throw new Error(`第 ${index + 1} 条 GSB 只能是 G、S、B 或空`);
+		}
+		if (note.length > 5000) throw new Error(`第 ${index + 1} 条备注过长`);
+		seen.add(requestId);
+		return {
+			requestId,
+			result: result as AnnotationValue["result"],
+			note,
+		};
+	});
+}
+
+function summarizeAnnotations(annotations: AnnotationValue[]) {
+	const summary = {
+		total: annotations.length,
+		completed: 0,
+		G: 0,
+		S: 0,
+		B: 0,
+	};
+	for (const annotation of annotations) {
+		if (!annotation.result) continue;
+		summary.completed += 1;
+		summary[annotation.result] += 1;
+	}
+	return summary;
+}
+
+async function readSubmissionFile(filePath: string) {
+	try {
+		return JSON.parse(await readFile(filePath, "utf8")) as AnnotationSubmission;
+	} catch {
+		return null;
+	}
+}
+
+function withoutAnnotations(
+	submission: AnnotationSubmission,
+): AnnotationSubmissionSummary {
+	const { annotations: _annotations, ...summary } = submission;
+	return summary;
+}
+
+async function atomicWriteJson(filePath: string, value: unknown) {
+	const temporaryPath = `${filePath}.${process.pid}.${randomUUID()}.tmp`;
+	await writeFile(temporaryPath, `${JSON.stringify(value, null, 2)}\n`, "utf8");
+	await rename(temporaryPath, filePath);
+}
+
+function enqueueSubmissionWrite<T>(pageName: string, task: () => Promise<T>) {
+	const previous = submissionWriteQueues.get(pageName) ?? Promise.resolve();
+	const current = previous.then(task, task);
+	submissionWriteQueues.set(pageName, current);
+	const cleanup = () => {
+		if (submissionWriteQueues.get(pageName) === current) {
+			submissionWriteQueues.delete(pageName);
+		}
+	};
+	void current.then(cleanup, cleanup);
+	return current;
+}
+
+async function readHtmlTitle(filePath: string, fileName: string) {
+	const handle = await open(filePath, "r");
+	try {
+		const buffer = Buffer.alloc(16 * 1024);
+		const { bytesRead } = await handle.read(buffer, 0, buffer.length, 0);
+		const head = buffer.subarray(0, bytesRead).toString("utf8");
+		const match = head.match(/<title[^>]*>([\s\S]*?)<\/title>/i);
+		const title = match?.[1]
+			?.replace(/<[^>]+>/g, " ")
+			.replace(/\s+/g, " ")
+			.trim();
+		return title ? decodeHtmlEntities(title) : stripHtmlExtension(fileName);
+	} finally {
+		await handle.close();
+	}
+}
+
+function stripHtmlExtension(fileName: string) {
+	return fileName.replace(/\.html?$/i, "");
+}
+
+function decodeHtmlEntities(value: string) {
+	return value
+		.replaceAll("&amp;", "&")
+		.replaceAll("&lt;", "<")
+		.replaceAll("&gt;", ">")
+		.replaceAll("&quot;", '"')
+		.replaceAll("&#39;", "'");
+}

frontend/app/lib/claw-auth.ts

@@ -1,6 +1,6 @@
 import { spawn } from "node:child_process";
 import { randomBytes, scryptSync, timingSafeEqual } from "node:crypto";
-import { mkdir, readFile, writeFile } from "node:fs/promises";
+import { mkdir, readFile, rename, writeFile } from "node:fs/promises";
 import path from "node:path";
 import { cookies } from "next/headers";
 
@@ -14,6 +14,7 @@ const COOKIE_NAME = authCookieName();
 const SESSION_MAX_AGE_SECONDS = 60 * 60 * 24 * 30;
 const SESSION_REFRESH_INTERVAL_MS = 60 * 60 * 12 * 1000;
 const LINUX_ACCOUNT_WORKSPACE_NAME = "zk-agent";
+let sessionsWriteQueue: Promise<unknown> = Promise.resolve();
 
 type UserRecord = {
 	id: string;
@@ -131,9 +132,10 @@ export async function logoutAccount() {
 	const cookieStore = await cookies();
 	const token = cookieStore.get(COOKIE_NAME)?.value;
 	if (token) {
-		const sessionsFile = await readSessions();
+		await mutateSessions((sessionsFile) => {
-		delete sessionsFile.sessions[token];
+			delete sessionsFile.sessions[token];
-		await writeSessions(sessionsFile);
+			return sessionsFile;
+		});
 	}
 	cookieStore.delete(COOKIE_NAME);
 }
@@ -151,8 +153,12 @@ export async function getCurrentAccount(): Promise<AccountSession | null> {
 	const account = usersFile.users.find((user) => user.id === session.accountId);
 	if (!account) return null;
 	if (shouldRefreshSession(session.updatedAt ?? session.createdAt)) {
-		session.updatedAt = new Date().toISOString();
+		await mutateSessions((latestSessionsFile) => {
-		await writeSessions(sessionsFile);
+			const latestSession = latestSessionsFile.sessions[token];
+			if (!latestSession) return latestSessionsFile;
+			latestSession.updatedAt = new Date().toISOString();
+			return latestSessionsFile;
+		});
 		await setSessionCookie(token);
 	}
 	return { id: account.id, username: account.username };
@@ -199,13 +205,15 @@ export function accountSessionRoot(accountId: string, sessionId: string) {
 
 async function createSession(account: UserRecord) {
 	const token = randomBytes(32).toString("hex");
-	const sessionsFile = await readSessions();
+	const now = new Date().toISOString();
-	sessionsFile.sessions[token] = {
+	await mutateSessions((sessionsFile) => {
-		accountId: account.id,
+		sessionsFile.sessions[token] = {
-		createdAt: new Date().toISOString(),
+			accountId: account.id,
-		updatedAt: new Date().toISOString(),
+			createdAt: now,
-	};
+			updatedAt: now,
-	await writeSessions(sessionsFile);
+		};
+		return sessionsFile;
+	});
 
 	await setSessionCookie(token);
 	return { id: account.id, username: account.username };
@@ -257,7 +265,22 @@ async function readSessions(): Promise<SessionsFile> {
 
 async function writeSessions(payload: SessionsFile) {
 	await mkdir(AUTH_STATE_ROOT, { recursive: true });
-	await writeFile(SESSIONS_PATH, JSON.stringify(payload, null, 2), "utf8");
+	const tempPath = `${SESSIONS_PATH}.${process.pid}.${Date.now()}.${randomBytes(4).toString("hex")}.tmp`;
+	await writeFile(tempPath, JSON.stringify(payload, null, 2), "utf8");
+	await rename(tempPath, SESSIONS_PATH);
+}
+
+async function mutateSessions(
+	mutator: (payload: SessionsFile) => SessionsFile | Promise<SessionsFile>,
+) {
+	const run = sessionsWriteQueue.then(async () => {
+		const current = await readSessions();
+		const next = await mutator(current);
+		await writeSessions(next);
+		return next;
+	});
+	sessionsWriteQueue = run.catch(() => undefined);
+	return run;
 }
 
 function normalizeUsername(username: string) {

frontend/app/public/annotation-submissions.js

@@ -0,0 +1,234 @@
+(() => {
+	const bridge = window.__annotationSubmissionBridge;
+	if (!bridge?.pageName || typeof bridge.read !== "function") return;
+
+	const endpoint = `/api/annotations/${encodeURIComponent(bridge.pageName)}/submissions`;
+	const toolbar = document.querySelector(".toolbar");
+	if (!toolbar) return;
+
+	const style = document.createElement("style");
+	style.textContent = `
+    .annotation-submit-status { color: #687385; font-size: 12px; white-space: nowrap; }
+    .annotation-history-dialog { width: min(680px, calc(100vw - 32px)); max-height: min(720px, calc(100vh - 32px)); padding: 0; border: 1px solid #d9dee7; border-radius: 8px; box-shadow: 0 24px 80px rgba(15, 23, 42, .2); }
+    .annotation-history-dialog::backdrop { background: rgba(15, 23, 42, .32); }
+    .annotation-history-head { position: sticky; top: 0; z-index: 1; display: flex; align-items: center; justify-content: space-between; gap: 16px; padding: 16px 18px; border-bottom: 1px solid #e5e9ef; background: #fff; }
+    .annotation-history-head h2 { margin: 0; font-size: 17px; }
+    .annotation-history-close { width: 32px; height: 32px; border: 0; border-radius: 5px; background: #f1f4f8; font-size: 20px; color: #556274; }
+    .annotation-history-list { margin: 0; padding: 0; list-style: none; overflow: auto; }
+    .annotation-history-item { display: grid; grid-template-columns: minmax(0, 1fr) auto; gap: 14px; align-items: center; padding: 14px 18px; border-bottom: 1px solid #edf0f4; }
+    .annotation-history-item:last-child { border-bottom: 0; }
+    .annotation-history-title { font-weight: 650; font-size: 14px; }
+    .annotation-history-meta { margin-top: 5px; color: #687385; font-size: 12px; }
+    .annotation-history-summary { margin-top: 7px; display: flex; flex-wrap: wrap; gap: 10px; color: #4b5666; font-size: 12px; }
+    .annotation-history-actions { display: flex; gap: 7px; }
+    .annotation-history-empty { padding: 40px 18px; text-align: center; color: #687385; }
+    @media (max-width: 820px) { .annotation-submit-status { display: none; } }
+  `;
+	document.head.appendChild(style);
+
+	const historyButton = button("提交记录", "secondary");
+	const submitButton = button("提交", "primary");
+	const status = document.createElement("span");
+	status.className = "annotation-submit-status";
+	toolbar.append(status, historyButton, submitButton);
+
+	const dialog = document.createElement("dialog");
+	dialog.className = "annotation-history-dialog";
+	dialog.innerHTML = `
+    <div class="annotation-history-head">
+      <h2>提交记录</h2>
+      <button class="annotation-history-close" aria-label="关闭">×</button>
+    </div>
+    <ol class="annotation-history-list"></ol>
+  `;
+	document.body.appendChild(dialog);
+	dialog.querySelector(".annotation-history-close").onclick = () =>
+		dialog.close();
+	dialog.addEventListener("click", (event) => {
+		if (event.target === dialog) dialog.close();
+	});
+
+	let history = [];
+	historyButton.onclick = async () => {
+		dialog.showModal();
+		await refreshHistory();
+	};
+	submitButton.onclick = async () => {
+		submitButton.disabled = true;
+		submitButton.textContent = "提交中";
+		try {
+			const response = await fetch(endpoint, {
+				method: "POST",
+				headers: { "content-type": "application/json" },
+				body: JSON.stringify({ annotations: bridge.read() }),
+			});
+			const payload = await response.json();
+			if (!response.ok) throw new Error(payload.error || "提交失败");
+			history = [
+				summaryOf(payload.submission),
+				...history.filter((item) => item.id !== payload.submission.id),
+			];
+			setStatus(payload.submission);
+			setRevisionUrl(payload.submission.id);
+			bridge.notify?.(`已提交第 ${payload.submission.version} 版`);
+		} catch (error) {
+			bridge.notify?.(error instanceof Error ? error.message : "提交失败");
+		} finally {
+			submitButton.disabled = false;
+			submitButton.textContent = "提交";
+		}
+	};
+
+	async function refreshHistory() {
+		const list = dialog.querySelector(".annotation-history-list");
+		list.innerHTML = '<li class="annotation-history-empty">正在读取</li>';
+		try {
+			const payload = await fetchJson(endpoint);
+			history = payload.history || [];
+			renderHistory();
+		} catch (error) {
+			list.innerHTML = `<li class="annotation-history-empty">${escapeHtml(error instanceof Error ? error.message : "读取失败")}</li>`;
+		}
+	}
+
+	function renderHistory() {
+		const list = dialog.querySelector(".annotation-history-list");
+		if (!history.length) {
+			list.innerHTML =
+				'<li class="annotation-history-empty">还没有提交记录</li>';
+			return;
+		}
+		list.innerHTML = "";
+		for (const item of history) {
+			const li = document.createElement("li");
+			li.className = "annotation-history-item";
+			li.innerHTML = `
+        <div>
+          <div class="annotation-history-title">第 ${item.version} 版 · ${escapeHtml(item.submittedBy?.username || "未知用户")}</div>
+          <div class="annotation-history-meta">${formatTime(item.submittedAt)}</div>
+          <div class="annotation-history-summary">
+            <span>已标 ${item.summary?.completed || 0}/${item.summary?.total || 0}</span>
+            <span>G ${item.summary?.G || 0}</span>
+            <span>S ${item.summary?.S || 0}</span>
+            <span>B ${item.summary?.B || 0}</span>
+          </div>
+        </div>
+        <div class="annotation-history-actions">
+          <button class="btn secondary" data-action="copy" type="button">复制链接</button>
+          <button class="btn secondary" data-action="view" type="button">查看</button>
+        </div>
+      `;
+			li.querySelector('[data-action="copy"]').onclick = async () => {
+				const url = revisionUrl(item.id);
+				try {
+					await navigator.clipboard.writeText(url);
+					bridge.notify?.("版本链接已复制");
+				} catch {
+					window.prompt("复制版本链接", url);
+				}
+			};
+			li.querySelector('[data-action="view"]').onclick = async () => {
+				try {
+					const payload = await fetchJson(
+						`${endpoint}?revision=${encodeURIComponent(item.id)}`,
+					);
+					bridge.apply(payload.submission.annotations || []);
+					dialog.close();
+					setStatus(payload.submission, "正在查看");
+					setRevisionUrl(payload.submission.id);
+					bridge.notify?.(`已打开第 ${payload.submission.version} 版`);
+				} catch (error) {
+					bridge.notify?.(error instanceof Error ? error.message : "读取失败");
+				}
+			};
+			list.appendChild(li);
+		}
+	}
+
+	async function initialize() {
+		try {
+			const requestedRevision = new URLSearchParams(window.location.search).get(
+				"revision",
+			);
+			if (requestedRevision) {
+				const payload = await fetchJson(
+					`${endpoint}?revision=${encodeURIComponent(requestedRevision)}`,
+				);
+				bridge.apply(payload.submission.annotations || []);
+				setStatus(payload.submission, "正在查看");
+				return;
+			}
+			const payload = await fetchJson(endpoint);
+			history = payload.history || [];
+			if (payload.latest) {
+				bridge.applyMissing(payload.latest.annotations || []);
+				setStatus(payload.latest);
+			} else {
+				status.textContent = "尚未提交";
+			}
+		} catch {
+			status.textContent = "提交记录暂不可用";
+		}
+	}
+
+	function setStatus(submission, prefix = "最新") {
+		status.textContent = `${prefix}：第 ${submission.version} 版 · ${submission.submittedBy?.username || "未知用户"}`;
+	}
+
+	function revisionUrl(revisionId) {
+		const url = new URL(window.location.href);
+		url.searchParams.set("revision", revisionId);
+		return url.toString();
+	}
+
+	function setRevisionUrl(revisionId) {
+		window.history.replaceState(null, "", revisionUrl(revisionId));
+	}
+
+	function button(label, variant) {
+		const element = document.createElement("button");
+		element.type = "button";
+		element.className = `btn ${variant}`;
+		element.textContent = label;
+		return element;
+	}
+
+	async function fetchJson(url) {
+		const response = await fetch(url, { cache: "no-store" });
+		const payload = await response.json();
+		if (!response.ok) throw new Error(payload.error || "请求失败");
+		return payload;
+	}
+
+	function summaryOf(submission) {
+		const { annotations: _annotations, ...summary } = submission;
+		return summary;
+	}
+
+	function formatTime(value) {
+		return new Intl.DateTimeFormat("zh-CN", {
+			year: "numeric",
+			month: "2-digit",
+			day: "2-digit",
+			hour: "2-digit",
+			minute: "2-digit",
+			hour12: false,
+		}).format(new Date(value));
+	}
+
+	function escapeHtml(value) {
+		return String(value ?? "").replace(
+			/[&<>"']/g,
+			(character) =>
+				({
+					"&": "&amp;",
+					"<": "&lt;",
+					">": "&gt;",
+					'"': "&quot;",
+					"'": "&#39;",
+				})[character],
+		);
+	}
+
+	void initialize();
+})();

scripts/deploy-ubuntu.sh

@@ -213,6 +213,7 @@ export OPENAI_API_KEY=$(printf "%q" "${api_key}")
 export OPENAI_BASE_URL=$(printf "%q" "${base_url}")
 export OPENAI_MODEL=$(printf "%q" "${model}")
 export OPENAI_TIMEOUT_SECONDS="${OPENAI_TIMEOUT_SECONDS:-3600}"
+export CLAW_MODEL_IDLE_TIMEOUT_SECONDS="${CLAW_MODEL_IDLE_TIMEOUT_SECONDS:-60}"
 
 export CLAW_DEPLOY_INSTANCE=$(printf "%q" "${DEPLOY_INSTANCE}")
 export CLAW_BACKEND_HOST=$(printf "%q" "${backend_host}")
@@ -264,6 +265,7 @@ ensure_env_file() {
 		echo "OPENAI_BASE_URL=${OPENAI_BASE_URL:-}"
 		echo "OPENAI_MODEL=${OPENAI_MODEL:-}"
 		echo "OPENAI_TIMEOUT_SECONDS=${OPENAI_TIMEOUT_SECONDS:-3600}"
+		echo "CLAW_MODEL_IDLE_TIMEOUT_SECONDS=${CLAW_MODEL_IDLE_TIMEOUT_SECONDS:-60}"
 		echo "CLAW_SERVICE_SCOPE=${SERVICE_SCOPE}"
 		echo "CLAW_ENABLE_LINUX_ACCOUNTS=${CLAW_ENABLE_LINUX_ACCOUNTS:-0}"
 		echo "OPENAI_API_KEY=已配置"

scripts/start-backend.sh

@@ -20,7 +20,7 @@ export OPENAI_API_KEY
 export OPENAI_BASE_URL="${OPENAI_BASE_URL:-http://model.mify.ai.srv/v1}"
 export OPENAI_MODEL="${OPENAI_MODEL:-tongyi/deepseek-v4-pro}"
 export OPENAI_TIMEOUT_SECONDS="${OPENAI_TIMEOUT_SECONDS:-3600}"
-export CLAW_MODEL_IDLE_TIMEOUT_SECONDS="${CLAW_MODEL_IDLE_TIMEOUT_SECONDS:-${OPENAI_TIMEOUT_SECONDS}}"
+export CLAW_MODEL_IDLE_TIMEOUT_SECONDS="${CLAW_MODEL_IDLE_TIMEOUT_SECONDS:-60}"
 
 # 后端也会调用 Node 生态工具，例如 feishu-mcp-pro。systemd 不读取交互式
 # shell 的 nvm 配置，所以需要把部署时记录的 Node 路径显式传给 Python 进程。

scripts/start-webui.sh

@@ -142,6 +142,7 @@ export OPENAI_API_KEY="${OPENAI_API_KEY:-}"
 export OPENAI_BASE_URL="${OPENAI_BASE_URL:-http://model.mify.ai.srv/v1}"
 export OPENAI_MODEL="${OPENAI_MODEL:-tongyi/deepseek-v4-pro}"
 export OPENAI_TIMEOUT_SECONDS="${OPENAI_TIMEOUT_SECONDS:-3600}"
+export CLAW_MODEL_IDLE_TIMEOUT_SECONDS="${CLAW_MODEL_IDLE_TIMEOUT_SECONDS:-60}"
 
 BACKEND_LOG="${RUN_DIR}/webui-backend.log"
 FRONTEND_LOG="${RUN_DIR}/webui-frontend.log"

skills/label-fast-slow-routing/SKILL.md

@@ -0,0 +1,150 @@
+---
+name: label-fast-slow-routing
+description: 依据 2026.7 产品标准，对单条或批量 query 进行“快 / 慢 / 模糊”三档可解释打标。先用精简标签知识召回垂域，再优先应用垂域规则；未覆盖时依次使用多指令、自动任务等结构规则和产品核心原则，并输出决策层级、规则 ID、证据、冲突和能力承接状态。用于快慢分流标注、评测、边界仲裁和 badcase 分析。
+---
+
+# 快慢分流打标
+
+## 目标
+
+按用户对响应速度和结果质量的预期判断 `快`、`慢`、`模糊`。业务垂域、结构信号和当前系统能力只参与解释与路由，不代替语义标签。
+
+## 渐进加载
+
+每条 case 按下面顺序读取，命中后及时停止，不要加载整个 `references/`：
+
+1. 第一动作读取 [references/index.md](references/index.md)，完成垂域召回和结构信号预判；垂域不明显或命中召回边界时，再读取 [references/preclassification.md](references/preclassification.md)。
+2. 读取一个主垂域卡；只有主垂域确实无法确定时，才读取第二个候选垂域卡。
+3. 出现多轮依赖、多指令或自动任务信号时，再读取对应结构卡。多轮先重建当前任务，自动任务还要读取自动化垂域卡。
+4. 垂域和结构卡都没有明确覆盖时，才读取 [references/core-policy.md](references/core-policy.md)。
+5. 命中标记为冲突的规则，或做规则维护时，再读取 [references/policy-conflicts.md](references/policy-conflicts.md)。
+6. 处理边界样本或校准时，按需读取 [references/golden-examples.md](references/golden-examples.md)。
+
+普通 case 通常只需读取索引和一个垂域卡。不要运行时读取相邻 `label-master`；本 Skill 已包含所需的精简知识快照。
+
+## 决策优先级
+
+按以下优先级裁决：
+
+1. 用户当前任务明确给出的已批准口径。
+2. 2026.7 产品文档中的明确垂域规则。
+3. 2026.7 跨垂域结构规则。
+4. 2026.7 产品核心原则。
+5. 快系统能力状态，只记录承接情况，不修改语义标签。
+
+垂域规则和核心原则冲突时，以垂域规则为最终结论，并记录 `policy_conflict=true` 和冲突证据。
+
+## 工作流
+
+1. 提取当前 query、前轮、设备、端侧、时间窗口和已提供的能力清单。
+2. 使用索引输出一个主垂域、最多两个候选垂域，以及多指令、自动任务、上下文依赖信号。
+3. 存在上下文依赖时读取 `references/structure-multi-turn.md`，识别上下文关系并重建当前完整任务；不得按上一轮实际路由直接继承标签。
+4. 读取主垂域卡并查找明确规则。命中时设置 `decision_level="domain"`。
+5. 未命中垂域规则时检查结构卡。命中时设置 `decision_level="structure"`。
+6. 仍未命中时使用核心原则，设置 `decision_level="core"`。
+7. 做反证检查，说明为什么没有选择其他档位。
+8. 只在有明确能力清单或验证结果时填写 `fast_system_pending`；未检查时填 `null`。
+9. 需要完整落盘时，按 schema 输出并运行校验脚本。
+
+## 输出
+
+默认返回简短结论：
+
+```text
+标签：慢
+判断层级：导航/生服垂域
+命中规则：NAV_SLOW_CONDITIONED_POI
+依据：单地点规划带明确筛选条件，命中导航垂域慢规则。
+```
+
+如果外层评测协议只接受 `prediction`、`reason` 等通用字段，把层级和规则 ID
+写在 `reason` 开头，例如：
+
+```text
+[domain:NAV_SLOW_CONDITIONED_POI] 单地点规划带明确筛选条件，按导航垂域给慢。
+```
+
+需要落盘、批量评测或用户要求完整记录时，输出符合 [schemas/label-record.schema.json](schemas/label-record.schema.json) 的对象：
+
+```json
+{
+  "query": "导航去附近评分4.5以上的泰国菜",
+  "label": "慢",
+  "default_route": "慢系统",
+  "domain": "导航/生服",
+  "candidate_domains": ["导航/生服"],
+  "decision_level": "domain",
+  "decision_rule_id": "NAV_SLOW_CONDITIONED_POI",
+  "decision_source": "references/domain-navigation-life.md",
+  "decision_path": [
+    "preclassification:导航/生服",
+    "domain:NAV_SLOW_CONDITIONED_POI"
+  ],
+  "reason": "单地点规划带明确筛选条件，命中导航垂域慢规则。",
+  "evidence": ["“评分4.5以上”是明确筛选条件"],
+  "counterevidence": ["核心快规则把目标明确的简单条件操作视为快"],
+  "structural_signals": {
+    "complex": null,
+    "multi_instruction": false,
+    "auto_task": false,
+    "context_dependent": false
+  },
+  "context_used": false,
+  "context_relation": "independent",
+  "fast_system_pending": null,
+  "route_override": null,
+  "policy_conflict": true,
+  "conflict_refs": ["CONFLICT_NAV_CONDITIONED_POI"],
+  "confidence": "high",
+  "review_required": false,
+  "uncertainties": [],
+  "policy_version": "FAST-SLOW-2026.7"
+}
+```
+
+新生成的多轮记录应填写 `context_relation`，取值和判断方法见
+`references/structure-multi-turn.md`。单轮或与历史无关时填写
+`"independent"`；为兼容旧产物，该字段在 schema 中保持可选。只有实际使用了
+历史内容时，`context_used` 和 `structural_signals.context_dependent` 才能为
+`true`。
+
+使用 `null` 表示未评估，不要猜测端侧、上下文、能力状态或复杂度。
+
+## 校验
+
+单条记录：
+
+```bash
+python scripts/validate_label.py --record '<JSON object>'
+```
+
+JSON 或 JSONL：
+
+```bash
+python scripts/validate_label.py --file output/fast_slow_labels.jsonl
+```
+
+校验器会检查字段不变量、决策层级、知识来源文件和规则 ID。结构通过不代表语义正确。
+
+## 复核门禁
+
+满足任一情况时设置 `review_required=true`：
+
+- 缺少会改变垂域或标签的前轮、端侧或设备信息。
+- 当前轮明显依赖前文，但无法唯一重建原任务。
+- 两个同优先级规则给出不同结论。
+- 自动任务端侧未知，且不同端侧策略不同。
+- 只能依赖未确认的能力清单。
+- `confidence="low"`。
+
+已由优先级解决的产品文档冲突只需记录，不自动要求人工复核。
+
+## 禁止事项
+
+- 不按字数、句长、关键词数量或既有标签比例直接定标。
+- 不因快系统当前不支持而把 `快`改成`慢`。
+- 不把 `complex=true` 直接等同于`慢`。
+- 不把所有多指令或自动任务无条件判慢，必须检查专项例外。
+- 不把多轮、指代、确认、取消或上一轮实际路由直接等同于快慢标签。
+- 不使用 2026.6、`FAST_DIRECT` 或 `SLOW_*` 旧策略。
+- 不输出没有规则 ID、证据和判断层级的裸标签。

skills/label-fast-slow-routing/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "快慢分流打标"
+  short_description: "依据 2026.7 标准按垂域、结构和核心原则分层打标"
+  default_prompt: "Use $label-fast-slow-routing to label these queries as 快、慢或模糊并给出可复核依据。"

skills/label-fast-slow-routing/references/core-policy.md

@@ -0,0 +1,51 @@
+# 2026.7 核心原则
+
+本文件只在垂域和结构规则没有明确覆盖时作为兜底。
+
+## 三档定义
+
+| 标签 | 用户状态 | 定义 | 默认路由 |
+| --- | --- | --- | --- |
+| 快 | 操控 | 期望说完即得，延迟敏感 | 快系统 |
+| 慢 | 委托 | 能接受等待，期待系统推理、规划或生成高质量结果 | 慢系统 |
+| 模糊 | 问询 | 速度和质量都重要，或难以明确归类 | 慢系统 |
+
+核心问题：用户说出这句话时，能接受等多久？
+
+## 快
+
+- `CORE_FAST_DIRECT_ACTION`：操作目标明确，动作和对象可直接识别，不需要先决定做什么。
+- `CORE_FAST_OPERATION_FLOW`：用户正在明确操作流中执行直接动作。
+- `CORE_FAST_DETERMINISTIC_QUERY`：高频、结果唯一、无需加工的确定性工具查询。
+
+条件修饰、感受描述和非标准叫法不会自动变慢。只要动作与目标仍然明确，保持快。
+
+多轮不会自动变慢。前文只补全唯一对象或简单槽位，当前轮仍是直接动作、明确操作流或确定性查询时，保持快。
+
+## 慢
+
+- `CORE_SLOW_INTENT_INFERENCE`：只表达感受、状态或不满，没有明确动作和功能，需要反推动作。
+- `CORE_SLOW_SCENE_MAPPING`：描述想达到的场景或氛围，需要映射成一组操作。
+- `CORE_SLOW_MULTI_STEP_PLANNING`：需要多步计算、换算、跨源聚合、上下文推理或先规划再执行。
+- `CORE_SLOW_TRUE_DISAMBIGUATION`：存在真实多候选或必需槽位缺失，需要澄清或推断。
+
+当前轮继续上一轮的推荐、分析、故障排查或规划，或者新增约束导致需要重新检索和规划时，按慢处理。详细关系先读取 `structure-multi-turn.md`。
+
+## 模糊
+
+- `CORE_AMBIGUOUS_OPEN_QA`：闲聊或不属于明确垂域的开放问答。
+- `CORE_AMBIGUOUS_PRODUCT_QA`：没有更具体垂域规则覆盖的产品知识、故障咨询。
+- `CORE_AMBIGUOUS_STATE_QUERY`：没有更具体垂域规则覆盖的设备或环境信息查询。
+- `CORE_AMBIGUOUS_POLICY_GAP`：快慢证据同时存在，当前规则没有明确优先级。
+
+`模糊`不代表模型没把握。规则明确规定的模糊场景可以是高置信结果。
+
+## 能力承接
+
+语义上应为快、但当前快系统未支持时：
+
+- `label` 保持 `快`。
+- 有能力清单或验证证据时设置 `fast_system_pending=true`。
+- 没有检查能力时设置 `fast_system_pending=null`。
+
+禁止根据当前 parser、白名单、接口或模型能力反推语义标签。

skills/label-fast-slow-routing/references/domain-automation.md

@@ -0,0 +1,27 @@
+# 自动化垂域
+
+使用前先读取 `references/structure-automation.md`，确认 query 确实是条件触发任务或自动化管理。
+
+## 音箱 / 家庭自动化
+
+- `AUTO_SLOW_HOME_AUTOMATION`：音箱端家庭自动化默认慢，包括明确触发动作和意向化自动化设计。
+
+## 车载自动化
+
+- `AUTO_FAST_CAR_REMINDER`：车载闹钟或条件提醒给快。
+- `AUTO_SLOW_CAR_REGULAR_TASK`：车载常规条件任务语义上按慢记录，并设置 `route_override="座舱AC"`。
+
+## 未知端侧
+
+- `AUTO_AMBIGUOUS_ENDPOINT_UNKNOWN`：端侧未知且音箱、车载策略会产生不同结论时，给模糊并设置 `review_required=true`。
+
+无条件创建普通闹钟、提醒或日程不属于自动任务，回到通用工具垂域。
+
+## 多轮
+
+先按 `structure-multi-turn.md` 合并当前轮补充的触发条件、动作和端侧，再应用本卡规则：
+
+- `AUTO_CONTEXT_RECONSTRUCTED`：多轮只负责重建完整自动任务，不因轮次直接改变标签。
+- 当前轮补全后可确定是车载提醒、车载常规任务或家庭自动化时，分别使用对应规则。
+- 历史和当前轮合并后端侧仍未知，且不同端侧结论不同，命中 `AUTO_AMBIGUOUS_ENDPOINT_UNKNOWN`。
+- 缺少会改变自动任务类型或端侧判断的历史时，设置 `review_required=true`。

skills/label-fast-slow-routing/references/domain-car-control.md

@@ -0,0 +1,30 @@
+# 车控垂域
+
+适用于车辆设备、驾驶功能和座舱能力的执行控制。
+
+## 慢
+
+- `CAR_SLOW_INTENT_EXPRESSION`：只表达冷、热、刺眼、拥挤等感受，没有明确功能或动作。
+- `CAR_SLOW_SCENE_EXPRESSION`：只描述会议、睡眠、节日氛围等场景目标，需要组合车辆功能。
+- `CAR_SLOW_FUZZY_FUNCTION`：不知道功能名称，以模糊描述指代车辆功能。
+- `CAR_SLOW_MULTI_POSITION_REASONING`：包含位置选择、排除或复杂设备逻辑。
+- `CAR_SLOW_CONFIRMATION_CHAIN`：当前轮回答的是场景映射、模糊功能识别或复杂位置选择等慢车控任务的澄清，重建后原任务仍命中车控慢规则。
+
+## 快
+
+- `CAR_FAST_EXPLICIT_CONTROL`：除上述慢场景外，动作、车辆设备或功能可直接识别。
+- `CAR_FAST_EXPLICIT_MULTI_CONTROL`：多个动作全部是明确车辆设备控制，且不需要规划依赖。
+- `CAR_FAST_CONTEXT_DIRECT`：前文只补全唯一车辆设备、位置或属性，当前轮合并后是明确直接控制。
+
+确认、取消或简短补槽不直接继承上一轮系统路由。先按 `structure-multi-turn.md` 重建完整车控任务。
+
+## 冲突标记
+
+`CAR_SLOW_FUZZY_FUNCTION` 与核心汇总中的“非标准叫法但可识别时给快”冲突。按垂域优先给慢，并记录：
+
+```text
+policy_conflict=true
+conflict_refs=["CONFLICT_CAR_FUZZY_FUNCTION"]
+```
+
+示例“那个压线噔噔的声音关掉”在本层级判慢。

skills/label-fast-slow-routing/references/domain-content.md

@@ -0,0 +1,29 @@
+# 内容垂域
+
+适用于音乐、视频、电台、新闻、歌单、内容搜索、播放控制和内容问答。
+
+## 慢
+
+- `CONTENT_SLOW_DESCRIPTION_SEARCH`：依赖描述性语言识别资源。
+- `CONTENT_SLOW_QA`：内容知识问答或需要生成 answer。
+- `CONTENT_SLOW_SCENE_RECOMMENDATION`：按场景进行内容推荐。
+- `CONTENT_SLOW_LYRIC_SEARCH`：根据歌词片段搜索资源。
+- `CONTENT_SLOW_FRESHNESS_OR_RANKING`：最新、最热、排行等时效或排序请求。
+- `CONTENT_SLOW_PLAYLIST_OR_FAVORITES`：歌单编排、收藏搜索和复杂个人资源检索。
+- `CONTENT_SLOW_NON_EXPLICIT_PLAY`：没有明确播放意图，需要先理解用户要什么。
+- `CONTENT_SLOW_MULTI_STEP`：需要联网、澄清或多步推理后才能满足。
+
+## 快
+
+- `CONTENT_FAST_EXPLICIT_PLAY`：播放对象、歌手、类型或资源明确。
+- `CONTENT_FAST_PLAYER_CONTROL`：暂停、继续、上一首、下一首、快进等直接播控。
+- `CONTENT_FAST_SIMPLE_SEARCH`：精准搜索、泛推荐或已有上下文中的直接资源切换。
+
+## 多轮
+
+- `CONTENT_SLOW_CONTEXT_PLANNING`：当前轮继续上一轮的内容推荐、知识问答、描述搜索、歌单编排或复杂个人资源检索。
+- `CONTENT_SLOW_CONTEXT_CONSTRAINT`：当前轮为既有内容请求新增时效、排行、场景、风格或其他需要重新检索的约束。
+- `CONTENT_FAST_CONTEXT_CONTROL`：前轮建立明确资源上下文后，当前轮执行直接播控。
+- `CONTENT_FAST_CONTEXT_SWITCH`：前轮已经唯一确定资源集合，当前轮直接选择、切换或播放其中明确的一项。
+
+90 秒时间窗口和前后轮关键词重叠只用于判断是否延续同一任务，不直接决定快慢。仅因当前快系统没有资源或能力失败，也不修改语义标签；失败后的实际升级属于执行策略。

skills/label-fast-slow-routing/references/domain-general-tools.md

@@ -0,0 +1,20 @@
+# 通用工具垂域
+
+适用于计算、时间、天气、翻译、闹钟、提醒、电话等确定性工具能力。
+
+## 慢
+
+- `TOOLS_SLOW_MULTI_STEP_REASONING`：需要多步推理、复杂计算或多次换算，不能直接查询得到。
+- `TOOLS_SLOW_EXTERNAL_PROCESSING`：需要结合上下文或外部信息进行加工。
+- `TOOLS_SLOW_MISSING_SLOT`：执行所需关键槽位不全，需要追问或推断补齐。
+- `TOOLS_SLOW_UNSTRUCTURED_REQUEST`：表达开放，无法稳定结构化为工具参数。
+
+## 快
+
+- `TOOLS_FAST_DETERMINISTIC_QUERY`：当前时间、简单天气、确定性换算等可直接查询或计算。
+- `TOOLS_FAST_EXPLICIT_ACTION`：时间、对象和动作明确的闹钟、提醒、电话等操作。
+- `TOOLS_FAST_CONTEXT_COMPLETED_ACTION`：前文补全唯一时间、联系人、地点或对象后，重建任务可直接结构化为一次工具调用。
+
+普通“明天提醒我开会”属于工具操作；带事件触发条件的提醒继续读取自动化规则。
+
+只有结合前文后关键槽位仍缺失，才命中 `TOOLS_SLOW_MISSING_SLOT`；不能因为当前轮单独看不完整就直接判慢。

skills/label-fast-slow-routing/references/domain-iot.md

@@ -0,0 +1,28 @@
+# IoT 垂域
+
+适用于家庭设备控制、房间环境查询和 IoT 设备信息查询。
+
+## 慢
+
+- `IOT_SLOW_INTENT_CONTROL`：只给舒适、温馨等目标状态，需要反推设备组合。
+- `IOT_SLOW_ENV_QUERY`：环境查询未明确指定房间或需要比较、聚合。
+- `IOT_SLOW_OTHER_QUERY`：其他 IoT 查询没有明确设备与属性。
+- `IOT_SLOW_CONTEXT_REASONING`：当前轮依赖前文继续场景设计、环境比较、设备选择或聚合判断。
+- `IOT_SLOW_CONTEXT_UNRESOLVED`：结合历史后设备、房间或属性仍不唯一。
+
+## 快
+
+- `IOT_FAST_EXPLICIT_CONTROL`：设备、位置和动作可直接识别。
+- `IOT_FAST_ROOM_ENV_QUERY`：环境查询明确指定房间。
+- `IOT_FAST_DEVICE_PROPERTY_QUERY`：明确指定设备和属性。
+- `IOT_FAST_EXPLICIT_MULTI_CONTROL`：多个动作全部是明确 IoT 设备控制。
+- `IOT_FAST_CONTEXT_DIRECT`：前文已经唯一确定设备、房间或属性，当前轮是直接控制或单属性查询。
+
+## 冲突标记
+
+IoT 查询的细粒度规则覆盖核心汇总的“设备/环境信息问答给模糊”。命中查询规则时记录：
+
+```text
+policy_conflict=true
+conflict_refs=["CONFLICT_IOT_QUERY_VS_AMBIGUOUS"]
+```

skills/label-fast-slow-routing/references/domain-navigation-life.md

@@ -0,0 +1,34 @@
+# 导航 / 生活服务垂域
+
+适用于导航、路线、地图操作、位置与路况问答、POI 搜索及相关生活服务请求。
+
+## 慢
+
+- `NAV_SLOW_CONDITIONED_POI`：单地点规划带明确筛选条件，如评分、品类属性或多个约束。
+- `NAV_SLOW_FUZZY_POI`：单地点规划使用隐喻、外形或模糊描述识别目标。
+- `NAV_SLOW_MULTI_PLACE`：一次规划多个地点或先后目的地。
+- `NAV_SLOW_MULTI_ACTION`：导航请求同时要求路线选择、沿途搜索、终点搜索等多个动作。
+- `NAV_SLOW_LOW_FREQUENCY_QA`：沿途城市、红绿灯数量等非高频复杂问答。
+- `NAV_SLOW_COMPLEX_EXPRESSION`：自纠正、非人机话语或描述性表达使目标需要额外推理。
+- `NAV_SLOW_CONTEXT_PLANNING`：当前轮继续上一轮的路线规划、沿途搜索或多地点任务。
+- `NAV_SLOW_CONTEXT_CONSTRAINT`：当前轮为既有目的地新增评分、品类属性、排除条件或多个约束，需要重新检索或规划。
+
+## 快
+
+- `NAV_FAST_SIMPLE_SINGLE_POI`：不带筛选条件的单地点规划，包括 A、A 的 B、A 附近的 B、顺路的 A、A 最近的 B。
+- `NAV_FAST_MAP_OPERATION`：删除途经点、地址设置、导航播报、路线切换等地图操作。
+- `NAV_FAST_HIGH_FREQUENCY_QA`：到目的地或途经点的时间、距离、位置，当前位置、路名和路况查询。
+- `NAV_FAST_CONTEXT_SIMPLE`：前文已经唯一确定目的地、途经点或路线，当前轮只是开始导航、查询时间距离、切换路线或其他简单续接。
+
+## 冲突标记
+
+`NAV_SLOW_CONDITIONED_POI` 与核心汇总中的“目标明确且带简单条件可快”冲突。按垂域优先给慢，并记录：
+
+```text
+policy_conflict=true
+conflict_refs=["CONFLICT_NAV_CONDITIONED_POI"]
+```
+
+“最近、附近、顺路”在本垂域被明确列为无筛选条件的快路径，不按普通筛选条件处理。
+
+上一轮实际走快系统还是慢系统不直接决定当前标签。先按 `structure-multi-turn.md` 重建任务，再判断当前续接是否需要重新规划。

skills/label-fast-slow-routing/references/domain-product-qa.md

@@ -0,0 +1,28 @@
+# 产品问答垂域
+
+适用于产品知识、故障诊断、设置入口和车辆动态信息查询。
+
+## 慢
+
+- `PRODUCT_SLOW_FAULT_DIAGNOSIS`：报告异常、询问原因、要求排查或处理建议。
+- `PRODUCT_SLOW_STATIC_KNOWLEDGE`：询问功能说明、概念、规则、规格、使用方法或知识库内容。
+- `PRODUCT_SLOW_COMPLEX_DYNAMIC_QUERY`：读取本车信号、档案或统计值，并进行条件判断、聚合、比较、换算或够否判断。
+- `PRODUCT_SLOW_CONTEXT_REASONING`：当前轮依赖前文继续知识问答、故障诊断、条件判断、比较、聚合或够否判断。
+- `PRODUCT_SLOW_CONTEXT_UNRESOLVED`：结合已有历史后对象或属性仍不唯一，需要追问或推断。
+
+## 快
+
+- `PRODUCT_FAST_SETTINGS_ENTRY`：定位设置项入口、询问在哪里设置或直接跳转设置页。
+- `PRODUCT_FAST_SIMPLE_DYNAMIC_QUERY`：直接读取一个当前值、档案值或统计值。
+- `PRODUCT_FAST_CONTEXT_SIMPLE_QUERY`：前文已经唯一确定对象或属性，当前轮只读取一个当前值、档案值或统计值。
+
+“依赖上一轮”本身不是慢规则。历史只完成唯一指代消解时，应按重建后的完整问题判断。
+
+## 冲突标记
+
+本垂域与核心汇总的模糊档存在两组冲突：
+
+- 静态知识和故障诊断按本垂域给慢：`CONFLICT_PRODUCT_QA_SLOW_VS_AMBIGUOUS`。
+- 简单动态状态查询按本垂域给快：`CONFLICT_PRODUCT_STATE_FAST_VS_AMBIGUOUS`。
+
+命中时设置 `policy_conflict=true` 并记录对应冲突 ID。

skills/label-fast-slow-routing/references/domain-system-app.md

@@ -0,0 +1,21 @@
+# 系统 / App 控制垂域
+
+适用于系统设置、应用操作、屏幕操作和系统功能控制。
+
+## 慢
+
+- `SYSTEM_SLOW_INTENT_EXPRESSION`：只表达刺眼、吵、费眼等状态，没有明确操作目标。
+- `SYSTEM_SLOW_SCENE_EXPRESSION`：描述会议、睡眠、专注等场景，需要映射多个系统功能。
+- `SYSTEM_SLOW_DISAMBIGUATION`：指代不清、存在多个候选或功能名称无法唯一识别，需要追问。
+- `SYSTEM_SLOW_COMPLEX_TASK`：包含选择、排除、条件判断或多步编排。
+
+## 快
+
+- `SYSTEM_FAST_EXPLICIT_CONTROL`：系统功能、应用、页面或屏幕元素可直接识别，动作明确。
+- `SYSTEM_FAST_OPERATION_FLOW`：当前操作流中的返回、确认、取消、点击、切换等直接动作。
+- `SYSTEM_FAST_EXPLICIT_MULTI_CONTROL`：多个动作全部是明确系统或设备控制，且没有规划依赖。
+- `SYSTEM_FAST_CONTEXT_DIRECT`：前文已经唯一确定应用、页面、功能或屏幕元素，当前轮执行直接操作。
+
+应用内搜索、内容消费或外部服务如果已有更具体垂域卡，优先进入对应垂域。
+
+当前轮虽然含“这个、那里、确认”等词，但历史能唯一解析且动作直接时给快；历史仍有多个候选时命中 `SYSTEM_SLOW_DISAMBIGUATION`。

skills/label-fast-slow-routing/references/golden-examples.md

@@ -0,0 +1,65 @@
+# 2026.7 黄金边界样例
+
+样例按“垂域优先、结构其次、核心兜底”口径校准。
+
+| Query / 上下文 | 标签 | 判断层级 | 规则 ID |
+| --- | --- | --- | --- |
+| 开空调 | 快 | 车控 | `CAR_FAST_EXPLICIT_CONTROL` |
+| 太冷了，调一下空调 | 快 | 车控 | `CAR_FAST_EXPLICIT_CONTROL` |
+| 我感觉好冷快冻死了，帮我调一下 | 慢 | 车控 | `CAR_SLOW_INTENT_EXPRESSION` |
+| 调一个浪漫点的氛围灯 | 快 | 车控 | `CAR_FAST_EXPLICIT_CONTROL` |
+| 帮我把车里布置得浪漫一点 | 慢 | 车控 | `CAR_SLOW_SCENE_EXPRESSION` |
+| 导航去附近评分 4.5 以上的泰国菜 | 慢 | 导航/生服 | `NAV_SLOW_CONDITIONED_POI` |
+| 导航去最近的加油站 | 快 | 导航/生服 | `NAV_FAST_SIMPLE_SINGLE_POI` |
+| 导航去大裤衩，再沿途找充电站 | 慢 | 导航/生服 | `NAV_SLOW_MULTI_ACTION` |
+| 删除途经点 | 快 | 导航/生服 | `NAV_FAST_MAP_OPERATION` |
+| 那个压线噔噔的声音关掉 | 慢 | 车控 | `CAR_SLOW_FUZZY_FUNCTION` |
+| 小米 SU7 支持哪些驾驶模式 | 慢 | 产品问答 | `PRODUCT_SLOW_STATIC_KNOWLEDGE` |
+| 我的车怎么突然没声音了 | 慢 | 产品问答 | `PRODUCT_SLOW_FAULT_DIAGNOSIS` |
+| 现在胎压是多少 | 快 | 产品问答 | `PRODUCT_FAST_SIMPLE_DYNAMIC_QUERY` |
+| 哨兵模式在哪里设置 | 快 | 产品问答 | `PRODUCT_FAST_SETTINGS_ENTRY` |
+| 播放周杰伦的歌 | 快 | 内容 | `CONTENT_FAST_EXPLICIT_PLAY` |
+| 播放最近短视频里很火的那首歌 | 慢 | 内容 | `CONTENT_SLOW_FRESHNESS_OR_RANKING` |
+| 现在几点了 | 快 | 通用工具 | `TOOLS_FAST_DETERMINISTIC_QUERY` |
+| 上午 9:18 到 12:18，再加下午 2 点到 6 点共多久 | 慢 | 通用工具 | `TOOLS_SLOW_MULTI_STEP_REASONING` |
+| 把空调调到 24 度，然后打开座椅加热 | 快 | 车控 | `CAR_FAST_EXPLICIT_MULTI_CONTROL` |
+| 查天气再决定要不要开窗 | 慢 | 多指令结构 | `MULTI_SLOW_CROSS_DOMAIN_WORKFLOW` |
+| 音箱端：每天晚上 8 点开灯开空调 | 慢 | 自动化 | `AUTO_SLOW_HOME_AUTOMATION` |
+| 车载端：车速超过 70 提醒我减速 | 快 | 自动化 | `AUTO_FAST_CAR_REMINDER` |
+| 车载端：每次上车就导航去公司 | 慢 | 自动化 | `AUTO_SLOW_CAR_REGULAR_TASK` |
+| 自动帮我设置一下，端侧未知 | 模糊 | 自动化 | `AUTO_AMBIGUOUS_ENDPOINT_UNKNOWN` |
+| 无聊了，陪我聊聊天 | 模糊 | 核心 | `CORE_AMBIGUOUS_OPEN_QA` |
+| 上轮：导航去公司；当前：换一条路线 | 快 | 导航/生服 | `NAV_FAST_CONTEXT_SIMPLE` |
+| 上轮：导航去附近餐厅；当前：要评分 4.5 以上并且方便停车 | 慢 | 导航/生服 | `NAV_SLOW_CONTEXT_CONSTRAINT` |
+| 上轮：公司和家选一个；当前：去公司 | 快 | 导航/生服 | `NAV_FAST_CONTEXT_SIMPLE` |
+| 上轮：播放周杰伦的歌；当前：下一首 | 快 | 内容 | `CONTENT_FAST_CONTEXT_CONTROL` |
+| 上轮：推荐通勤音乐；当前：安静一点但不能太困 | 慢 | 内容 | `CONTENT_SLOW_CONTEXT_CONSTRAINT` |
+| 上轮：要打开哪个车窗；当前：副驾 | 快 | 车控 | `CAR_FAST_CONTEXT_DIRECT` |
+| 上轮：帮我把车里布置得浪漫一点；系统询问是否继续；当前：确认 | 慢 | 车控 | `CAR_SLOW_CONFIRMATION_CHAIN` |
+| 上轮：现在胎压是多少；当前：后轮呢 | 快 | 产品问答 | `PRODUCT_FAST_CONTEXT_SIMPLE_QUERY` |
+| 上轮：为什么车机没声音；当前：刚升级系统以后开始的 | 慢 | 产品问答 | `PRODUCT_SLOW_CONTEXT_REASONING` |
+| 上轮：打开微信设置；当前：点通知 | 快 | 系统/App控制 | `SYSTEM_FAST_CONTEXT_DIRECT` |
+| 上轮：明天上午开会；当前：十点提醒我 | 快 | 通用工具 | `TOOLS_FAST_CONTEXT_COMPLETED_ACTION` |
+| 上轮：客厅和卧室都有空调；当前：打开那个 | 慢 | IoT | `IOT_SLOW_CONTEXT_UNRESOLVED` |
+| 当前：就按刚才那个办，但未提供任何历史（需复核） | 模糊 | 多轮结构 | `CONTEXT_REVIEW_HISTORY_MISSING` |
+
+重点检查以下紧邻对：
+
+```text
+太冷了，调一下空调             -> 快
+我感觉好冷，帮我调一下          -> 慢
+
+导航去最近的加油站              -> 快
+导航去评分 4.5 以上的泰国菜      -> 慢
+
+现在胎压是多少                  -> 快
+刚刚谁动了我的车                -> 慢
+
+多个明确设备控制                -> 快
+查询结果决定后续控制             -> 慢
+
+前文只补全唯一对象               -> 按完整任务判断，直接动作通常快
+当前轮增加检索或规划约束           -> 慢
+确认或取消                       -> 重建原任务，不按表面词定标
+缺少会改变结论的历史               -> 不猜测，进入复核
+```

skills/label-fast-slow-routing/references/index.md

@@ -0,0 +1,52 @@
+# 渐进加载索引
+
+本文件只负责召回知识，不直接给快慢结论。
+
+## 第一阶段输出
+
+先提取：
+
+- 当前 query 和可见上下文。
+- 操作、询问或聊天。
+- 对象、设备、资源、位置和端侧。
+- 主垂域和最多两个候选垂域。
+- 多指令、自动任务和上下文依赖信号。
+
+主垂域不明显、控制与问答边界不清或同时召回多个领域时，读取
+`references/preclassification.md`；明显的单垂域请求无需额外加载。
+
+## 垂域召回
+
+| 主要信号 | 主垂域 | 读取文件 |
+| --- | --- | --- |
+| 导航、路线、目的地、途经点、当前位置、路况、附近 POI、沿途 POI | 导航/生服 | `references/domain-navigation-life.md` |
+| 车窗、座椅、空调、雨刮、车灯、驾驶模式、智驾、泊车等车辆操作 | 车控 | `references/domain-car-control.md` |
+| 小米产品、汽车功能说明、故障、设置入口、本车动态状态 | 产品问答 | `references/domain-product-qa.md` |
+| 音乐、视频、电台、新闻、歌单、播放和内容搜索 | 内容 | `references/domain-content.md` |
+| 系统设置、应用打开关闭、屏幕操作、免打扰、亮度、音量 | 系统/App控制 | `references/domain-system-app.md` |
+| 计算、时间、天气、翻译、闹钟、提醒、电话等工具能力 | 通用工具 | `references/domain-general-tools.md` |
+| 家庭灯具、家电、扫地机、房间环境、IoT 设备状态 | IoT | `references/domain-iot.md` |
+| 当、如果、时候、之后、每、一...就、自动化、智能习惯、超级任务 | 自动化 | `references/domain-automation.md` |
+
+## 召回边界
+
+- 车辆“怎么设置、什么意思、为什么异常”优先产品问答；“打开、关闭、调节”优先车控。
+- 家庭设备“打开、关闭、调节”优先 IoT；系统自身的亮度、音量、应用和页面操作优先系统/App控制。
+- “附近有什么”同时可能命中导航和生活服务，先看用户要导航、查询还是消费服务。
+- 闹钟或提醒没有条件触发结构时属于通用工具；存在事件条件触发时同时召回自动化。
+- query 只表达通用知识、闲聊或开放问题时，不强行归入垂域，后续读取核心原则。
+
+## 结构信号
+
+| 信号 | 何时读取 |
+| --- | --- |
+| 多个独立动作、跨对象/跨垂域、并列连接、多方向调整 | `references/structure-multi-instruction.md` |
+| 条件触发 + 动作、自动化、智能习惯、超级任务 | `references/structure-automation.md`，并读取 `references/domain-automation.md` |
+| 当前轮含省略、指代、确认/取消、修正或补充，且结论可能依赖前轮 | 先读取 `references/structure-multi-turn.md` 重建任务，再读取主垂域卡 |
+
+## 停止条件
+
+- 主垂域卡命中明确规则后停止加载其他垂域卡。
+- 只有主垂域无法确定时才读取第二张候选垂域卡。
+- 垂域和结构规则均未覆盖时读取 `references/core-policy.md`。
+- 命中卡片中的冲突标记时读取 `references/policy-conflicts.md`。

skills/label-fast-slow-routing/references/policy-conflicts.md

@@ -0,0 +1,19 @@
+# 2026.7 规则冲突表
+
+本表记录产品文档内部已知冲突。当前口径为垂域规则优先，冲突不会静默覆盖。
+
+| 冲突 ID | 场景 | 垂域结论 | 核心汇总结论 | 当前结论 |
+| --- | --- | --- | --- | --- |
+| `CONFLICT_NAV_CONDITIONED_POI` | 带评分等明确条件的单地点导航 | 慢 | 快 | 慢 |
+| `CONFLICT_CAR_FUZZY_FUNCTION` | 非标准名称描述车辆功能 | 慢 | 快 | 慢 |
+| `CONFLICT_PRODUCT_QA_SLOW_VS_AMBIGUOUS` | 静态产品知识、故障诊断 | 慢 | 模糊 | 慢 |
+| `CONFLICT_PRODUCT_STATE_FAST_VS_AMBIGUOUS` | 简单车辆动态状态查询 | 快 | 模糊 | 快 |
+| `CONFLICT_IOT_QUERY_VS_AMBIGUOUS` | IoT 环境或设备信息查询 | 细分为快/慢 | 模糊 | 按 IoT 细则 |
+
+命中时：
+
+1. 使用垂域标签。
+2. 设置 `policy_conflict=true`。
+3. 把对应冲突 ID 写入 `conflict_refs`。
+4. 在 `counterevidence` 中简述另一层规则。
+5. 优先级已经解决时不自动设置 `review_required=true`。

skills/label-fast-slow-routing/references/preclassification.md

@@ -0,0 +1,31 @@
+# 垂域预分类知识
+
+本文件是从标签大师提炼的粗粒度召回知识，只帮助选择快慢分流垂域卡，不输出业务标签、function 或 Agent target。
+
+## 预分类顺序
+
+1. 先识别自动任务和多指令结构。
+2. 再判断请求是控制、查询、内容消费还是知识问答。
+3. 最后结合对象和端侧选择主垂域。
+
+## 粗粒度映射
+
+| 标签大师领域或典型标签 | 快慢分流垂域 |
+| --- | --- |
+| 地图导航、地图设置、地图问答、地址设置、走哪问哪、餐饮/旅游/汽车服务 POI | 导航/生服 |
+| 车载控制 | 车控 |
+| 小米产品帮助、手车互联、系统/车载/家用设备状态查询 | 产品问答 |
+| 音乐、视频、电台、新闻、播放控制、歌单、内容问答 | 内容 |
+| 系统控制、应用控制、屏幕操作、相机、声纹 | 系统/App控制 |
+| 时间、天气、计算、翻译、闹钟、提醒、电话等工具标签 | 通用工具 |
+| 设备控制、家庭 IoT 环境和设备查询 | IoT |
+| 自动任务、系统/应用/设备定时控制 | 自动化 |
+
+## 关键边界
+
+- 同一个设备名可能落入控制或产品问答。动作执行进入控制垂域，功能说明、故障和状态查询进入产品问答。
+- “附近餐厅”根据满足方式区分：要路线或地图结果进入导航/生服；要团购、订座或消费服务仍属于导航/生服大类，但保留候选业务标签。
+- “提醒我明天开会”是普通提醒；“到公司时提醒我开会”是自动任务。
+- 自动任务和多指令是结构信号，也可能同时存在业务垂域。
+
+预分类不做最终裁决。关键词冲突时，以用户真实满足方式和候选垂域卡的适用范围为准。

skills/label-fast-slow-routing/references/structure-automation.md

@@ -0,0 +1,29 @@
+# 自动任务结构
+
+本文件提炼自标签大师的自动任务知识，只判断是否存在条件触发结构。
+
+## 正例
+
+常见结构：
+
+- 当 / 如果 / 时候 / 之后 / 到达 / 每 / 一...就。
+- 条件状态 + 动作，如“电量低于 20% 时提醒我充电”。
+- 自动化、智能场景、智能习惯、超级任务等显式管理请求。
+- 持续或延迟条件，如“座椅加热十分钟”。
+
+`就`分隔条件和动作时，`就`之前的完整内容属于条件；`且`连接的多个状态都属于条件。条件默认向右作用，不影响前面的动作。
+
+## 负例
+
+- 无条件创建普通闹钟、提醒、日程或倒计时。
+- 车辆已有固定功能的描述或控制，如“打开车道偏离预警”。
+- 只有条件描述而没有要执行的动作。
+- 普通功能名称中带“自动”，但用户没有创建条件任务。
+
+## 输出信号
+
+- `auto_task=true`：确认存在条件触发任务或自动化管理。
+- `auto_task=false`：普通控制、查询或无条件工具操作。
+- `auto_task=null`：缺少端侧或上下文，无法可靠判断。
+
+确认 `auto_task=true` 后，继续读取 `references/domain-automation.md`。自动任务结构本身不直接决定快慢。

skills/label-fast-slow-routing/references/structure-multi-instruction.md

@@ -0,0 +1,28 @@
+# 多指令结构
+
+本文件提炼自标签大师的多指令知识。它只判断结构及快慢例外，不生成拆分后的业务 target。
+
+## 识别
+
+`multi_instruction=true` 需要至少两个可独立执行的子任务。常见信号：
+
+- 和、并、然后、再、顺便、同时、并且连接独立动作。
+- 多个对象分别执行不同操作。
+- 同一设备存在多个独立方向调整。
+- 跨垂域动作或前一步结果决定下一步。
+
+以下通常仍是单意图：
+
+- 同一设备的打开后设置属性，如“打开空调调到 26 度”。
+- 同一动作作用于同品类多个位置。
+- 打开应用后搜索、点击、播放等共同完成一个任务的顺序操作。
+- 口误、重复、改口和连续相反操作。
+- 多个并列查询目标。
+
+## 快慢规则
+
+- `MULTI_FAST_EXPLICIT_DEVICE_CONTROLS`：全部子任务都是明确设备控制动作，不含查询依赖、选择、排除或规划，给快。
+- `MULTI_SLOW_INDEPENDENT_ACTIONS`：多个独立动作需要编排，且不满足设备控制快例外，给慢。
+- `MULTI_SLOW_CROSS_DOMAIN_WORKFLOW`：跨垂域、跨工具，或前一步结果决定后一步，给慢。
+
+垂域卡已经明确覆盖时，保留多指令信号并服从垂域规则。结构规则只在垂域未覆盖时裁决。

skills/label-fast-slow-routing/references/structure-multi-turn.md

@@ -0,0 +1,79 @@
+# 多轮上下文结构
+
+本文件统一判断当前轮如何依赖前文。多轮本身不是快慢依据；先重建用户当前真正要完成的任务，再由主垂域裁决。
+
+## 第一步：判断是否真的依赖前文
+
+移除对话历史后，当前 query 的对象、动作、约束和满足方式仍然明确，则：
+
+- `CONTEXT_INDEPENDENT`：前文与当前结论无关，按当前 query 的普通垂域规则判断。
+- `context_dependent=false`
+- `context_used=false`
+- `context_relation="independent"`
+
+只有省略、指代、确认、修正或补充会改变当前任务解释时，才设置 `context_dependent=true` 和 `context_used=true`。
+
+时间相邻、同一 session、相同关键词或上一轮走过某个系统，只能作为同一任务的辅助证据，不能单独证明上下文依赖。
+
+## 第二步：重建当前任务
+
+只继承完成当前判断所需的最少信息：
+
+1. 找到仍在延续的最近一个有效用户任务。
+2. 解析当前轮是在直接执行、补槽、改约束、继续规划，还是回答澄清。
+3. 合并成语义完整的当前任务，但不得继承已经被用户取消、否定或改写的条件。
+4. 前文存在多个可继承候选时，不猜测用户指向。
+
+## 继承边界
+
+- 当前轮明确说出新对象、新任务或“不是 A，是 B”时，以当前轮为准；被替换的旧信息不得继续保留。
+- 当前轮已经能独立构成完整任务时，不因词面相似强行继承前文。
+- 助手上一轮的追问可以帮助理解用户简短回答，但助手自行猜测、推荐或生成的内容不能当作用户已确认事实。
+- 工具失败、能力不足、拒绝或上一轮实际路由不改变任务的语义快慢；只在能力承接字段中记录。
+- 多段历史都可能被指代且无法唯一选择时，按未消歧处理。
+
+## 上下文关系
+
+### 可快的关系
+
+- `CONTEXT_FAST_RESOLVED_DIRECT`：前文只提供唯一对象、位置、资源或简单槽位，当前轮是直接控制、确定性查询或明确资源切换。设置 `context_relation="resolved_direct"`。
+- `CONTEXT_FAST_OPERATION_CONTINUATION`：用户处于已经明确的操作流中，当前轮执行返回、确认、取消、暂停、继续、切换等直接动作。设置 `context_relation="operation_continuation"`。
+
+这两类只是“具备判快条件”。若重建后的完整任务命中垂域慢规则，仍按垂域规则给慢。
+
+### 应慢的关系
+
+- `CONTEXT_SLOW_CONSTRAINT_EXTENSION`：当前轮新增或修改筛选、排序、比较、排除、路线、场景等约束，需要重新检索或规划。设置 `context_relation="constraint_extension"`。
+- `CONTEXT_SLOW_PLANNING_CONTINUATION`：当前轮继续上一轮的分析、推荐、生成、故障排查或多步规划。设置 `context_relation="planning_continuation"`。
+- `CONTEXT_SLOW_TRUE_DISAMBIGUATION`：结合现有历史后仍有多个真实候选，或关键槽位仍缺失，需要追问或推断。设置 `context_relation="unresolved_reference"`。
+
+### 回答澄清
+
+- `CONTEXT_CLARIFICATION_COMPLETION`：当前轮只回答系统上一轮的澄清、确认或补槽问题。设置 `context_relation="clarification_completion"`。
+
+这类不按“是、不是、确认、取消”等表面词直接定标，而是重建完整任务：
+
+- 澄清完成后是明确直接操作或确定性查询，按对应快规则判断。
+- 澄清完成后仍是推荐、规划、故障分析或场景映射，按对应慢规则判断。
+- 无法确认系统究竟在澄清哪个任务时，进入复核。
+
+## 历史缺失与复核
+
+- `CONTEXT_REVIEW_HISTORY_MISSING`：当前轮明显依赖前文，但输入没有提供足够历史，且不同解释会改变垂域或标签。
+
+此时不得凭空补全：
+
+```text
+context_relation="unresolved_reference"
+confidence="low"
+review_required=true
+```
+
+## 裁决顺序
+
+1. 使用本文件识别 `context_relation` 并重建当前任务。
+2. 使用重建后的任务读取主垂域卡。
+3. 主垂域有明确多轮或业务规则时，以垂域规则为最终结论。
+4. 主垂域没有覆盖时，才使用本文件中的 `CONTEXT_FAST_*` 或 `CONTEXT_SLOW_*` 作为结构层结论。
+
+禁止根据上一轮实际走快系统还是慢系统直接继承标签。

skills/label-fast-slow-routing/schemas/label-record.schema.json

@@ -0,0 +1,257 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://local.skills/label-fast-slow-routing/label-record.schema.json",
+  "title": "2026.7 快慢分流分层打标记录",
+  "type": "object",
+  "additionalProperties": false,
+  "required": [
+    "query",
+    "label",
+    "default_route",
+    "domain",
+    "candidate_domains",
+    "decision_level",
+    "decision_rule_id",
+    "decision_source",
+    "decision_path",
+    "reason",
+    "evidence",
+    "structural_signals",
+    "context_used",
+    "fast_system_pending",
+    "route_override",
+    "policy_conflict",
+    "conflict_refs",
+    "confidence",
+    "review_required",
+    "uncertainties",
+    "policy_version"
+  ],
+  "properties": {
+    "id": {
+      "type": ["string", "number", "null"]
+    },
+    "query": {
+      "type": "string",
+      "minLength": 1
+    },
+    "label": {
+      "enum": ["快", "慢", "模糊"]
+    },
+    "default_route": {
+      "enum": ["快系统", "慢系统"]
+    },
+    "domain": {
+      "enum": [
+        "导航/生服",
+        "车控",
+        "产品问答",
+        "内容",
+        "系统/App控制",
+        "通用工具",
+        "IoT",
+        "自动化",
+        null
+      ]
+    },
+    "candidate_domains": {
+      "type": "array",
+      "uniqueItems": true,
+      "items": {
+        "enum": [
+          "导航/生服",
+          "车控",
+          "产品问答",
+          "内容",
+          "系统/App控制",
+          "通用工具",
+          "IoT",
+          "自动化"
+        ]
+      }
+    },
+    "decision_level": {
+      "enum": ["domain", "structure", "core"]
+    },
+    "decision_rule_id": {
+      "type": "string",
+      "pattern": "^[A-Z][A-Z0-9_]+$"
+    },
+    "decision_source": {
+      "type": "string",
+      "pattern": "^references/[a-z0-9-]+\\.md(?:#.*)?$"
+    },
+    "decision_path": {
+      "type": "array",
+      "minItems": 1,
+      "items": {
+        "type": "string",
+        "minLength": 1
+      }
+    },
+    "reason": {
+      "type": "string",
+      "minLength": 1
+    },
+    "evidence": {
+      "type": "array",
+      "minItems": 1,
+      "items": {
+        "type": "string",
+        "minLength": 1
+      }
+    },
+    "counterevidence": {
+      "type": "array",
+      "items": {
+        "type": "string",
+        "minLength": 1
+      }
+    },
+    "structural_signals": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": [
+        "complex",
+        "multi_instruction",
+        "auto_task",
+        "context_dependent"
+      ],
+      "properties": {
+        "complex": {
+          "type": ["boolean", "null"]
+        },
+        "multi_instruction": {
+          "type": ["boolean", "null"]
+        },
+        "auto_task": {
+          "type": ["boolean", "null"]
+        },
+        "context_dependent": {
+          "type": ["boolean", "null"]
+        }
+      }
+    },
+    "context_used": {
+      "type": "boolean"
+    },
+    "context_relation": {
+      "enum": [
+        "independent",
+        "resolved_direct",
+        "operation_continuation",
+        "constraint_extension",
+        "planning_continuation",
+        "clarification_completion",
+        "unresolved_reference",
+        null
+      ]
+    },
+    "fast_system_pending": {
+      "type": ["boolean", "null"]
+    },
+    "route_override": {
+      "type": ["string", "null"]
+    },
+    "policy_conflict": {
+      "type": "boolean"
+    },
+    "conflict_refs": {
+      "type": "array",
+      "uniqueItems": true,
+      "items": {
+        "type": "string",
+        "pattern": "^CONFLICT_[A-Z0-9_]+$"
+      }
+    },
+    "confidence": {
+      "enum": ["high", "medium", "low"]
+    },
+    "review_required": {
+      "type": "boolean"
+    },
+    "uncertainties": {
+      "type": "array",
+      "items": {
+        "type": "string",
+        "minLength": 1
+      }
+    },
+    "policy_version": {
+      "const": "FAST-SLOW-2026.7"
+    }
+  },
+  "allOf": [
+    {
+      "if": {
+        "properties": {
+          "label": {
+            "const": "快"
+          }
+        },
+        "required": ["label"]
+      },
+      "then": {
+        "properties": {
+          "default_route": {
+            "const": "快系统"
+          }
+        }
+      }
+    },
+    {
+      "if": {
+        "properties": {
+          "label": {
+            "enum": ["慢", "模糊"]
+          }
+        },
+        "required": ["label"]
+      },
+      "then": {
+        "properties": {
+          "default_route": {
+            "const": "慢系统"
+          },
+          "fast_system_pending": {
+            "enum": [false, null]
+          }
+        }
+      }
+    },
+    {
+      "if": {
+        "properties": {
+          "policy_conflict": {
+            "const": true
+          }
+        },
+        "required": ["policy_conflict"]
+      },
+      "then": {
+        "properties": {
+          "conflict_refs": {
+            "minItems": 1
+          }
+        }
+      }
+    },
+    {
+      "if": {
+        "properties": {
+          "confidence": {
+            "const": "low"
+          }
+        },
+        "required": ["confidence"]
+      },
+      "then": {
+        "properties": {
+          "review_required": {
+            "const": true
+          }
+        }
+      }
+    }
+  ]
+}

skills/label-fast-slow-routing/scripts/validate_label.py

@@ -0,0 +1,436 @@
+#!/usr/bin/env python3
+"""Validate 2026.7 hierarchical fast/slow routing records."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import sys
+from pathlib import Path
+from typing import Any
+
+
+POLICY_VERSION = "FAST-SLOW-2026.7"
+SKILL_ROOT = Path(__file__).resolve().parent.parent
+CONFLICT_FILE = SKILL_ROOT / "references" / "policy-conflicts.md"
+
+DEFAULT_ROUTE_BY_LABEL = {
+    "快": "快系统",
+    "慢": "慢系统",
+    "模糊": "慢系统",
+}
+CONFIDENCE_VALUES = {"high", "medium", "low"}
+DECISION_LEVELS = {"domain", "structure", "core"}
+DOMAIN_VALUES = {
+    "导航/生服",
+    "车控",
+    "产品问答",
+    "内容",
+    "系统/App控制",
+    "通用工具",
+    "IoT",
+    "自动化",
+}
+RULE_ID_PATTERN = re.compile(r"^[A-Z][A-Z0-9_]+$")
+CONFLICT_ID_PATTERN = re.compile(r"^CONFLICT_[A-Z0-9_]+$")
+
+REQUIRED_FIELDS = {
+    "query",
+    "label",
+    "default_route",
+    "domain",
+    "candidate_domains",
+    "decision_level",
+    "decision_rule_id",
+    "decision_source",
+    "decision_path",
+    "reason",
+    "evidence",
+    "structural_signals",
+    "context_used",
+    "fast_system_pending",
+    "route_override",
+    "policy_conflict",
+    "conflict_refs",
+    "confidence",
+    "review_required",
+    "uncertainties",
+    "policy_version",
+}
+OPTIONAL_FIELDS = {
+    "id",
+    "counterevidence",
+    "context_relation",
+}
+CONTEXT_RELATION_VALUES = {
+    "independent",
+    "resolved_direct",
+    "operation_continuation",
+    "constraint_extension",
+    "planning_continuation",
+    "clarification_completion",
+    "unresolved_reference",
+}
+
+
+def _is_nonempty_string(value: Any) -> bool:
+    return isinstance(value, str) and bool(value.strip())
+
+
+def _check_string_list(
+    value: Any,
+    *,
+    field: str,
+    errors: list[str],
+    require_nonempty: bool = False,
+    unique: bool = False,
+) -> None:
+    if not isinstance(value, list):
+        errors.append(f"{field} 必须是数组")
+        return
+    if require_nonempty and not value:
+        errors.append(f"{field} 至少包含一项")
+    if any(not _is_nonempty_string(item) for item in value):
+        errors.append(f"{field} 的每一项都必须是非空字符串")
+    if unique and all(isinstance(item, str) for item in value):
+        if len(value) != len(set(value)):
+            errors.append(f"{field} 不允许重复项")
+
+
+def _validate_decision_source(
+    record: dict[str, Any],
+    *,
+    errors: list[str],
+) -> None:
+    source = record.get("decision_source")
+    rule_id = record.get("decision_rule_id")
+    level = record.get("decision_level")
+    if not _is_nonempty_string(source):
+        errors.append("decision_source 必须是非空字符串")
+        return
+    relative_path = source.split("#", 1)[0]
+    if not relative_path.startswith("references/") or not relative_path.endswith(".md"):
+        errors.append("decision_source 必须指向 references/*.md")
+        return
+    resolved = (SKILL_ROOT / relative_path).resolve()
+    references_root = (SKILL_ROOT / "references").resolve()
+    if references_root not in resolved.parents:
+        errors.append("decision_source 不允许跳出 references 目录")
+        return
+    if not resolved.is_file():
+        errors.append(f"decision_source 文件不存在: {relative_path}")
+        return
+    try:
+        source_text = resolved.read_text(encoding="utf-8")
+    except OSError as exc:
+        errors.append(f"无法读取 decision_source: {exc}")
+        return
+    if _is_nonempty_string(rule_id) and rule_id not in source_text:
+        errors.append(f"decision_rule_id 未出现在来源文件中: {rule_id}")
+
+    filename = resolved.name
+    if level == "domain" and not filename.startswith("domain-"):
+        errors.append("decision_level=domain 时来源文件必须是 domain-*.md")
+    if level == "structure" and not filename.startswith("structure-"):
+        errors.append("decision_level=structure 时来源文件必须是 structure-*.md")
+    if level == "core" and filename != "core-policy.md":
+        errors.append("decision_level=core 时来源文件必须是 core-policy.md")
+
+
+def _validate_conflicts(
+    record: dict[str, Any],
+    *,
+    errors: list[str],
+) -> None:
+    policy_conflict = record.get("policy_conflict")
+    conflict_refs = record.get("conflict_refs")
+    if not isinstance(policy_conflict, bool):
+        errors.append("policy_conflict 必须是布尔值")
+    _check_string_list(
+        conflict_refs,
+        field="conflict_refs",
+        errors=errors,
+        unique=True,
+    )
+    if not isinstance(conflict_refs, list):
+        return
+    if policy_conflict is True and not conflict_refs:
+        errors.append("policy_conflict=true 时 conflict_refs 至少包含一项")
+    if policy_conflict is False and conflict_refs:
+        errors.append("policy_conflict=false 时 conflict_refs 必须为空")
+
+    try:
+        known_conflicts = CONFLICT_FILE.read_text(encoding="utf-8")
+    except OSError as exc:
+        errors.append(f"无法读取冲突表: {exc}")
+        return
+    for conflict_id in conflict_refs:
+        if not isinstance(conflict_id, str):
+            continue
+        if not CONFLICT_ID_PATTERN.fullmatch(conflict_id):
+            errors.append(f"冲突 ID 格式非法: {conflict_id!r}")
+        elif conflict_id not in known_conflicts:
+            errors.append(f"冲突 ID 未登记: {conflict_id}")
+
+
+def validate_record(record: Any, index: int) -> dict[str, Any]:
+    errors: list[str] = []
+    warnings: list[str] = []
+
+    if not isinstance(record, dict):
+        return {
+            "index": index,
+            "valid": False,
+            "errors": ["记录必须是 JSON 对象"],
+            "warnings": [],
+        }
+    if "__parse_error__" in record:
+        return {
+            "index": index,
+            "valid": False,
+            "errors": [str(record["__parse_error__"])],
+            "warnings": [],
+        }
+
+    missing = sorted(REQUIRED_FIELDS - set(record))
+    if missing:
+        errors.append(f"缺少必填字段: {', '.join(missing)}")
+    unknown = sorted(set(record) - REQUIRED_FIELDS - OPTIONAL_FIELDS)
+    if unknown:
+        errors.append(f"存在未知字段: {', '.join(unknown)}")
+
+    if not _is_nonempty_string(record.get("query")):
+        errors.append("query 必须是非空字符串")
+
+    label = record.get("label")
+    if not isinstance(label, str) or label not in DEFAULT_ROUTE_BY_LABEL:
+        errors.append("label 必须是 快、慢、模糊 之一")
+    expected_route = DEFAULT_ROUTE_BY_LABEL.get(label) if isinstance(label, str) else None
+    if expected_route is not None and record.get("default_route") != expected_route:
+        errors.append(f"label={label} 时 default_route 必须是 {expected_route}")
+
+    domain = record.get("domain")
+    if domain is not None and (
+        not isinstance(domain, str) or domain not in DOMAIN_VALUES
+    ):
+        errors.append("domain 不是受支持的粗粒度垂域")
+    _check_string_list(
+        record.get("candidate_domains"),
+        field="candidate_domains",
+        errors=errors,
+        unique=True,
+    )
+    candidate_domains = record.get("candidate_domains")
+    if isinstance(candidate_domains, list):
+        unknown_domains = [
+            value
+            for value in candidate_domains
+            if not isinstance(value, str) or value not in DOMAIN_VALUES
+        ]
+        if unknown_domains:
+            errors.append(f"candidate_domains 包含未知垂域: {unknown_domains}")
+
+    level = record.get("decision_level")
+    if not isinstance(level, str) or level not in DECISION_LEVELS:
+        errors.append("decision_level 必须是 domain、structure、core 之一")
+    if level == "domain" and domain is None:
+        errors.append("decision_level=domain 时 domain 不能为空")
+
+    rule_id = record.get("decision_rule_id")
+    if not _is_nonempty_string(rule_id) or not RULE_ID_PATTERN.fullmatch(rule_id):
+        errors.append("decision_rule_id 必须是大写下划线规则 ID")
+
+    _check_string_list(
+        record.get("decision_path"),
+        field="decision_path",
+        errors=errors,
+        require_nonempty=True,
+    )
+    decision_path = record.get("decision_path")
+    if (
+        isinstance(decision_path, list)
+        and _is_nonempty_string(rule_id)
+        and not any(rule_id in str(item) for item in decision_path)
+    ):
+        errors.append("decision_path 必须包含 decision_rule_id")
+
+    _validate_decision_source(record, errors=errors)
+
+    if not _is_nonempty_string(record.get("reason")):
+        errors.append("reason 必须是非空字符串")
+    _check_string_list(
+        record.get("evidence"),
+        field="evidence",
+        errors=errors,
+        require_nonempty=True,
+    )
+    if "counterevidence" in record:
+        _check_string_list(
+            record.get("counterevidence"),
+            field="counterevidence",
+            errors=errors,
+        )
+
+    structural_signals = record.get("structural_signals")
+    expected_signal_keys = {
+        "complex",
+        "multi_instruction",
+        "auto_task",
+        "context_dependent",
+    }
+    if not isinstance(structural_signals, dict):
+        errors.append("structural_signals 必须是对象")
+    else:
+        actual_keys = set(structural_signals)
+        if actual_keys != expected_signal_keys:
+            errors.append(
+                "structural_signals 必须且只能包含 "
+                "complex、multi_instruction、auto_task、context_dependent"
+            )
+        for key in expected_signal_keys:
+            value = structural_signals.get(key)
+            if value is not None and not isinstance(value, bool):
+                errors.append(f"structural_signals.{key} 必须是布尔值或 null")
+
+    if not isinstance(record.get("context_used"), bool):
+        errors.append("context_used 必须是布尔值")
+    context_relation = record.get("context_relation")
+    if context_relation is not None and (
+        not isinstance(context_relation, str)
+        or context_relation not in CONTEXT_RELATION_VALUES
+    ):
+        errors.append("context_relation 不是受支持的上下文关系")
+    if context_relation == "independent" and record.get("context_used") is True:
+        errors.append("context_relation=independent 时 context_used 必须为 false")
+    if context_relation not in {None, "independent"}:
+        if record.get("context_used") is not True:
+            errors.append("依赖上下文的 context_relation 要求 context_used=true")
+        if isinstance(structural_signals, dict) and (
+            structural_signals.get("context_dependent") is not True
+        ):
+            errors.append(
+                "依赖上下文的 context_relation 要求 "
+                "structural_signals.context_dependent=true"
+            )
+
+    fast_system_pending = record.get("fast_system_pending")
+    if fast_system_pending is not None and not isinstance(fast_system_pending, bool):
+        errors.append("fast_system_pending 必须是布尔值或 null")
+    if fast_system_pending is True and label != "快":
+        errors.append("只有 label=快 时 fast_system_pending 才能为 true")
+
+    route_override = record.get("route_override")
+    if route_override is not None and not _is_nonempty_string(route_override):
+        errors.append("route_override 必须是非空字符串或 null")
+
+    _validate_conflicts(record, errors=errors)
+
+    confidence = record.get("confidence")
+    if not isinstance(confidence, str) or confidence not in CONFIDENCE_VALUES:
+        errors.append("confidence 必须是 high、medium、low 之一")
+    review_required = record.get("review_required")
+    if not isinstance(review_required, bool):
+        errors.append("review_required 必须是布尔值")
+    if confidence == "low" and review_required is not True:
+        errors.append("confidence=low 时 review_required 必须为 true")
+
+    _check_string_list(
+        record.get("uncertainties"),
+        field="uncertainties",
+        errors=errors,
+    )
+    uncertainties = record.get("uncertainties")
+    if isinstance(uncertainties, list) and uncertainties and review_required is False:
+        warnings.append("uncertainties 非空但 review_required=false，请确认不确定点不影响标签")
+
+    if record.get("policy_version") != POLICY_VERSION:
+        errors.append(f"policy_version 必须是 {POLICY_VERSION}")
+
+    return {
+        "index": index,
+        "id": record.get("id"),
+        "valid": not errors,
+        "errors": errors,
+        "warnings": warnings,
+    }
+
+
+def _load_records_from_file(path: Path) -> list[Any]:
+    text = path.read_text(encoding="utf-8-sig")
+    if path.suffix.lower() == ".jsonl":
+        records: list[Any] = []
+        for line_number, line in enumerate(text.splitlines(), start=1):
+            if not line.strip():
+                continue
+            try:
+                records.append(json.loads(line))
+            except json.JSONDecodeError as exc:
+                records.append(
+                    {
+                        "__parse_error__": (
+                            f"JSONL 第 {line_number} 行解析失败: {exc.msg}"
+                        )
+                    }
+                )
+        return records
+
+    payload = json.loads(text)
+    if isinstance(payload, list):
+        return payload
+    if isinstance(payload, dict) and isinstance(payload.get("records"), list):
+        return payload["records"]
+    return [payload]
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="校验 FAST-SLOW-2026.7 分层打标 JSON/JSONL"
+    )
+    source_group = parser.add_mutually_exclusive_group(required=True)
+    source_group.add_argument("--record", help="单条 JSON 对象字符串")
+    source_group.add_argument("--file", type=Path, help="JSON 或 JSONL 文件路径")
+    parser.add_argument(
+        "--strict",
+        action="store_true",
+        help="把 warning 也视为校验失败",
+    )
+    return parser.parse_args()
+
+
+def main() -> int:
+    args = _parse_args()
+    try:
+        if args.record is not None:
+            records = [json.loads(args.record)]
+        else:
+            records = _load_records_from_file(args.file)
+    except (OSError, json.JSONDecodeError) as exc:
+        result = {
+            "valid": False,
+            "total": 0,
+            "invalid": 1,
+            "warning_count": 0,
+            "errors": [str(exc)],
+            "results": [],
+        }
+        print(json.dumps(result, ensure_ascii=False, indent=2))
+        return 1
+
+    results = [validate_record(record, index) for index, record in enumerate(records)]
+    invalid = sum(not result["valid"] for result in results)
+    warning_count = sum(len(result["warnings"]) for result in results)
+    valid = invalid == 0 and (not args.strict or warning_count == 0)
+    summary = {
+        "valid": valid,
+        "total": len(records),
+        "invalid": invalid,
+        "warning_count": warning_count,
+        "results": results,
+    }
+    print(json.dumps(summary, ensure_ascii=False, indent=2))
+    return 0 if valid else 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())

skills/label-master/knowledge/标签/屏幕操作/屏幕操作.md

@@ -0,0 +1,35 @@
+# 屏幕操作
+
+## 标注输出
+
+- 旧 tag 标签：屏幕操作
+- Agent 形式（推荐最终输出）：Agent(tag="屏幕操作")
+- 输出形态：intent 包装，直接 Agent(tag="屏幕操作")，非 function program
+
+## 功能抽象
+
+屏幕内已呈现 GUI 元素的点击/点选操作
+
+## 适用范围
+
+针对当前屏幕上已呈现的 UI 元素（按钮、图标、卡片、海报、列表项、开关、菜单项等）执行点击、点选、选中，如"点击订购信息""点我的历史""点某电影海报""点某图标""点会员卡片"。是对已可见界面元素的直接操作。
+
+## 典型 Query
+
+点击订购信息 / 点一下账号管理 / 我想点我的历史 / 帮我点我的想看 / 点击设置 / 点一下通知中心 / 点某电影海报 / 点某节目封面 / 点某开关图标 / 点小米影视VIP会员卡片
+
+## 易混淆标签
+
+- 应用控制：打开/关闭/卸载某 App 或跳转到某页面（"打开页面"），屏幕操作是"点击页面内已有元素"
+- 系统控制：系统级设置项开关调节（亮度/音量/护眼等），非 GUI 元素点击
+- 图片问答（VisionQA）：对屏幕内容做识别/问答，非点击操作
+
+## 划分原则
+
+- 已呈现界面元素的点击/点选 → 屏幕操作
+- 打开/进入某应用或某设置页 → 应用控制 / 系统控制
+- query 明确"点/点击/点一下/选中" + 界面元素名 → 屏幕操作
+
+## 未解决问题
+
+待补充。

skills/label-master/knowledge/标签总览.md

@@ -28,6 +28,7 @@
 | 小米产品问答 | 小米产品、设备状态、车机/家居产品知识等 | `标签/小米产品问答/` |
 | 内容和媒体播放 | 音乐、视频、电台、新闻、古诗、笑话等播放类能力 | `标签/内容和媒体播放/` |
 | 应用控制和搜索 | 应用搜索、应用打开、浏览器搜索等 | `标签/应用控制和搜索/` |
+| 屏幕操作 | 当前屏幕已呈现 GUI 元素的点击、点选、选中 | `标签/屏幕操作/` |
 | AI创作 | 文本、图片、视频、代码等创作类能力 | `标签/AI创作/` |
 | 工具类 | 翻译、计算、日历、提醒、计时、查询等工具型能力 | `标签/工具类/` |
 

skills/label-master/knowledge/索引/label_manifest.json

@@ -54,6 +54,7 @@
     "小憩模式",
     "小爱帮助",
     "小米产品帮助",
+    "屏幕操作",
     "工资税收计算",
     "平台比价->商品比价",
     "应用定时控制",
@@ -213,10 +214,10 @@
     }
   ],
   "counts": {
-    "agent_tags": 134,
+    "agent_tags": 135,
     "functions": 13,
     "intents": 83,
-    "labels": 125,
+    "labels": 126,
     "objects": 14
   },
   "description": "标签大师的中控标签知识库机器索引；由 scripts/build_label_manifest.py 从 Markdown 生成。",
@@ -1597,6 +1598,19 @@
       "recommended_output_shape": "待确认。",
       "scope": "针对小米汽车特有的，区别与其他小米终端类型的设置项查询，限定在车载范围。只有在车载端给小米产品问答，在其他端是QA()，通用问答会根据知识回答其他非小米产品或通用车载功能的问题。"
     },
+    {
+      "agent_candidate": "",
+      "confusing_labels": "- 应用控制：打开/关闭/卸载某 App 或跳转到某页面（\"打开页面\"），屏幕操作是\"点击页面内已有元素\"\n- 系统控制：系统级设置项开关调节（亮度/音量/护眼等），非 GUI 元素点击\n- 图片问答（VisionQA）：对屏幕内容做识别/问答，非点击操作",
+      "domain": "屏幕操作",
+      "examples": "点击订购信息 / 点一下账号管理 / 我想点我的历史 / 帮我点我的想看 / 点击设置 / 点一下通知中心 / 点某电影海报 / 点某节目封面 / 点某开关图标 / 点小米影视VIP会员卡片",
+      "function_candidate": "",
+      "name": "屏幕操作",
+      "old_tag": "屏幕操作",
+      "path": "knowledge/标签/屏幕操作/屏幕操作.md",
+      "principles": "- 已呈现界面元素的点击/点选 → 屏幕操作\n- 打开/进入某应用或某设置页 → 应用控制 / 系统控制\n- query 明确\"点/点击/点一下/选中\" + 界面元素名 → 屏幕操作",
+      "recommended_output_shape": "",
+      "scope": "针对当前屏幕上已呈现的 UI 元素（按钮、图标、卡片、海报、列表项、开关、菜单项等）执行点击、点选、选中，如\"点击订购信息\"\"点我的历史\"\"点某电影海报\"\"点某图标\"\"点会员卡片\"。是对已可见界面元素的直接操作。"
+    },
     {
       "agent_candidate": "待确认。",
       "confusing_labels": "音乐",

skills/label-master/knowledge/索引/候选召回索引.md

@@ -21,6 +21,7 @@
 | 介绍、解释、是什么、为什么、来源、原理、百科、搜索一下 | QA、问答、搜索、具体问答垂域 | [问答](../标签/通用问答/问答.md)，[函数目录 QA](../输出能力/函数目录.md) |
 | 写、生成、创作、润色、改写、作文、文案、剧本、诗歌 | 文本创作、作文、Generate、QA | [文本创作](../标签/AI创作/文本创作.md)，[函数目录 Generate](../输出能力/函数目录.md) |
 | 打开、关闭、调节、设置 + 音量、亮度、蓝牙、护眼、关机、截图、朗读屏幕 | 系统控制、应用控制、相机、声纹 | [系统控制-设备控制-车载控制](../边界/高频混淆/系统控制-设备控制-车载控制.md)，[系统控制](../标签/系统控制和IOT设备控制/系统控制.md) |
+| 点、点击、点一下、选中 + 当前屏幕已呈现的按钮、图标、卡片、海报、列表项、菜单项 | 屏幕操作、应用控制、系统控制、图片问答 | [屏幕操作](../标签/屏幕操作/屏幕操作.md)，[应用控制](../标签/应用控制和搜索/应用控制.md)，[图片问答](../标签/通用问答/图片问答.md) |
 | 打开、关闭、调节、设置 + 家居设备、灯、电视、扫地机、洗衣机、主卧/客厅/厨房 | 设备控制、设备定时控制、系统控制 | [系统控制-设备控制-车载控制](../边界/高频混淆/系统控制-设备控制-车载控制.md)，[设备控制](../标签/系统控制和IOT设备控制/设备控制.md) |
 | 打开、关闭、调节、设置 + 车窗、座椅、雨刮、前备箱、智驾、泊车、哨兵、驾驶模式 | 车载控制、车载设备状态查询、自动任务、系统控制 | [系统控制-设备控制-车载控制](../边界/高频混淆/系统控制-设备控制-车载控制.md)，[车载控制](../标签/系统控制和IOT设备控制/车载控制.md) |
 | 当、如果、时候、到达、之后、每、一...就、自动化、智能习惯、超级任务 | 自动任务、提醒、设备控制、车载控制 | [自动任务判断](../判断维度/自动任务判断.md)，[自动任务](../标签/系统控制和IOT设备控制/自动任务.md) |

skills/label-master/knowledge/索引/标签索引.md

@@ -78,6 +78,7 @@
 | 系统控制和IOT设备控制 | 小憩模式 | 小憩模式开关、休息意图触发（"我要休息"/"我要睡觉"等）、小憩状态查询 | Agent(tag="小憩模式") | x0 = Alarm(type="RESTMODE")Search(alarm=[x0]) | [小憩模式](../标签/系统控制和IOT设备控制/小憩模式.md) |
 | 应用控制和搜索 | 应用定时控制 | 应用定时控制 | Agent(tag="应用定时控制") | 待确认 | [应用定时控制](../标签/应用控制和搜索/应用定时控制.md) |
 | 应用控制和搜索 | 应用控制 | 应用控制 | Agent(tag="应用控制") | function定义：参数定义：app： 对应的应用名page: 打开的对应的页面Action: 具体的操作，如：ope... | [应用控制](../标签/应用控制和搜索/应用控制.md) |
+| 屏幕操作 | 屏幕操作 | 屏幕操作 | Agent(tag="屏幕操作") | 无，intent 包装，非 function program | [屏幕操作](../标签/屏幕操作/屏幕操作.md) |
 | 应用控制和搜索 | 搜索\|应用名(搜索\|QQ音乐) | 搜索\|应用名(搜索\|QQ音乐) | Agent(tag="搜索\|应用名(搜索\|QQ音乐)") | function定义：参数定义：app： 对应的应用名content: 需要搜索的内容示例：Query:打开快手搜索情侣... | [搜索\|应用名(搜索\|QQ音乐)](../标签/应用控制和搜索/搜索-应用名-QQ音乐.md) |
 | 应用控制和搜索 | 浏览器搜索 | 浏览器搜索 | Agent(tag="浏览器搜索") | 待确认 | [浏览器搜索](../标签/应用控制和搜索/浏览器搜索.md) |
 | 生活服务 | 交通购票 | 交通购票 | Agent(tag="交通购票") | function定义：参数定义： | [交通购票](../标签/生活服务/交通购票.md) |

skills/model-training-lite/SKILL.md

@@ -1,8 +1,8 @@
 ---
 name: model-training-lite
 name_zh: 轻量模型训练
-description: 从 Codex、Claude Code 或其他 Agent 通过用户提供的 Jupyter 环境和数据版本发起一次模型训练，聚焦“准备远端工作区、同步数据、提交 CML SFT、返回 JobID”，不包含完整自主迭代闭环。
+description: 从 Codex、Claude Code 或其他 Agent 通过用户提供的 Jupyter 环境和数据版本发起一次模型训练，并可选发起评测和生成轻量结果摘要；不包含自动数据增强闭环。
-when_to_use: 当用户已经准备好训练数据，想让 Agent 使用 Jupyter/CloudML 发起一次模型训练、复用当前仓库的训练脚本和模板、或询问如何把模型训练流程沉淀成轻量 skill 时使用。
+when_to_use: 当用户已经准备好训练数据，想让 Agent 使用 Jupyter/CloudML 基于某个数据分支或 commit 发起一次模型训练、训练后评测、汇总结果，或询问如何把模型训练流程沉淀成轻量 skill 时使用。
 aliases: jupyter-model-training, sft-submit, 轻量训练, 模型训练提交
 examples:
   - 用这个 Jupyter 和 ai-planning 当前分支重新发起训练
@@ -14,7 +14,17 @@ allowed_tools: read_file, write_file, grep_search, glob_search, ask_user_questio
 
 # 轻量模型训练
 
-本 skill 只负责一次训练提交：
+本 skill 是人机协同的训练/评测执行器：人负责目标和边界裁决，AI 负责候选、统计、证据整理和确定性执行。
+
+```text
+人提出目标问题
+  -> AI 整理候选、统计分布、生成 review 表
+  -> 人确认边界、保留样本、数据是否入库
+  -> AI 提交训练、发起评测、汇总结果
+  -> 人判断是否继续下一轮
+```
+
+训练评测执行链路：
 
 ```text
 Jupyter 授权环境
@@ -23,26 +33,67 @@ Jupyter 授权环境
   -> 同步训练脚本与数据版本
   -> 检查 CML / 数据 / zk_trainer
   -> 提交 SFT
-  -> 返回 JobID、产物路径、下一步评测入口
+  -> 可选：SFT 成功后提交 CML 评测
+  -> 可选：读取 workflow metric_diff 生成轻量结果摘要
+  -> 返回 JobID、workflow ID、产物路径、摘要路径
 ```
 
-不要把它扩展成 `model-iteration` 那种完整“baseline -> 分析 -> 增强 -> 训练 -> 评测 -> 再分析”的自主循环。
+不要把它扩展成 `model-iteration` 那种完整“baseline -> 分析 -> 增强 -> 训练 -> 评测 -> 再分析”的自主循环。本 skill 可以辅助人完成每一步决策前的信息准备，但不自动裁决边界、不自动修改数据、不自动生成增强样本、不自动进入下一轮训练。
 
 ## 必要输入
 
 开始前必须拿到：
 
-- **Jupyter 地址和授权方式**：URL；密码、token、cookie/session，或说明当前环境已有可用登录态。
+- **Jupyter 地址和授权方式**：URL；密码、token、cookie/session，或使用本地默认配置。
-- **数据版本**：git repo + branch + commit，或远端工作区中已存在的数据路径。
+- **数据版本**：git repo + branch + commit，或 git repo + branch，或远端工作区中已存在的数据路径。
 - **训练配方**：如果是当前 ZK 中控 SFT，默认复用 `skills/model-iteration/scripts/`；如果不是，必须让用户提供训练脚本、模板或命令。
 
 通常用户只需要显式给 `Jupyter 地址` 和 `数据版本`。当前项目的默认训练配方可从本仓库继承，不必每次追问。
 
+组内默认 Jupyter 配置路径：
+
+```text
+skills/model-training-lite/jupyter_defaults.json
+```
+
+这个文件随 skill 提交，组内默认可直接使用。脚本输出只显示认证是否已配置，不回显密码。
+
+个人覆盖配置路径：
+
+```text
+skills/model-training-lite/.local/jupyter_defaults.json
+```
+
+`.local/` 已加入 gitignore。需要临时覆盖组内默认配置时，使用 `.local/` 或环境变量，不要改共享配置。
+
+配置格式可参考：
+
+```text
+skills/model-training-lite/jupyter_defaults.example.json
+```
+
+脚本也支持环境变量：
+
+```text
+MODEL_TRAINING_LITE_JUPYTER_URL
+MODEL_TRAINING_LITE_JUPYTER_PASSWORD
+MODEL_TRAINING_LITE_JUPYTER_TOKEN
+MODEL_TRAINING_LITE_JUPYTER_COOKIE
+MODEL_TRAINING_LITE_JUPYTER_AUTH_TYPE
+```
+
+配置优先级：
+
+```text
+脚本输入 > 环境变量 > .local/jupyter_defaults.json > jupyter_defaults.json
+```
+
 可选但推荐确认：
 
 - `owner`：远端输出目录使用的用户前缀，例如 `wuyang6`。
 - `run_name`：本次训练工作区名，例如 `manual_YYYYMMDD_zk_intent_xxx`。
-- 是否训练后立刻发起评测。默认只提交训练，不自动串接评测。
+- 是否训练后立刻发起评测。默认只提交训练；用户要求“训练后评测/看效果/分析结果”时启用。
+- `model_old`：评测工作流中的对照模型路径；未提供时使用默认线上基线。
 
 ## 默认 ZK SFT 配方
 
@@ -62,6 +113,22 @@ skills/model-iteration/assets/config.yaml
 
 ## 推荐流程
 
+### 0. 人机协同边界
+
+AI 应主动辅助人完成信息准备，但不能替人拍板：
+
+| 阶段 | AI 负责 | 人负责 |
+|---|---|---|
+| 目标定义 | 把问题拆成可评测口径，列出候选集合和风险 | 确认目标问题和主指标 |
+| 数据准备 | 调用相关 skill 生成候选、统计、review 表 | 判断边界样本是否保留 |
+| 数据构造 | 参考 `product-data` 生成训练/评测格式 | 确认数据是否入库 |
+| 标签判断 | 参考 `label-master` 校验 tag / function / complex | 确认争议规则 |
+| 线上挖掘 | 参考 `online-mining-v2` 获取 rid/session/prompt/output | 确认挖掘口径 |
+| 批量打标 | 参考 `model-labeling` 请求线上模型，生成基线 | 确认宽口径和人工 override |
+| 训练评测 | 本 skill 提交 SFT、评测、轻量摘要 | 决定是否继续下一轮 |
+
+需要全自动假设驱动迭代时，使用 `model-iteration`，不要把本 skill 临时扩展成自动闭环。
+
 ### 1. 先生成提交计划
 
 优先执行 portable script：
@@ -80,12 +147,16 @@ python skills/model-training-lite/scripts/render_sft_submission_plan.py --input
   "data_commit": "b92be709",
   "owner": "wuyang6",
   "run_name": "manual_20260528_zk_intent_clean_train",
-  "recipe": "zk_sft"
+  "recipe": "zk_sft",
+  "run_eval": true,
+  "model_old": "/mnt/wangsenhao/verl_zk/qwen4b_cispo_wokl_add_bvt_2/global_step_5/actor/huggingface"
 }
 ```
 
 脚本只生成计划和命令草案，不连接 Jupyter，不提交训练。
 
+如果只提供 `data_branch` 不提供 `data_commit`，脚本会生成 `git reset --hard origin/<branch>`，并在结果中写 warning。正式训练报告里必须记录实际 `git rev-parse HEAD`，避免后续无法复现。
+
 ### 2. 连接 Jupyter
 
 如果已经有 session/cookie，可直接调用 Jupyter REST API：
@@ -117,6 +188,12 @@ zk_trainer/
 
 把默认 ZK SFT 配方里的 `scripts/` 和 `config.yaml` 同步到远端工作区。再 clone 或更新数据仓库，并 checkout 到用户指定 commit。
 
+除 `model-iteration` 默认脚本外，还要同步本 skill 的轻量分析脚本：
+
+```text
+skills/model-training-lite/scripts/summarize_eval_result.py
+```
+
 ### 4. 提交前检查
 
 提交训练前必须输出：
@@ -158,10 +235,49 @@ eval "$(./scripts/resolve_run_ids.sh)"
 
 再查一次 `cml custom_train describe <JOB_ID>`，确认状态不是提交后立即失败。
 
+### 7. 可选：训练后提交评测
+
+只有用户要求“训练后评测/看效果/分析结果”时执行。
+
+训练成功标记存在后：
+
+```bash
+source ~/.cloudml-cli/.profile
+export AUTORESEARCH_CHAT_ROOT=<remote_workspace>
+export AUTORESEARCH_ROOT="$AUTORESEARCH_CHAT_ROOT"
+cd "$AUTORESEARCH_CHAT_ROOT"
+eval "$(./scripts/resolve_run_ids.sh)"
+./scripts/submit_cml_eval.sh "$EVAL_RUNDIC" "$AUTORESEARCH_CHAT_ROOT/sft_output" "<model_old>"
+```
+
+提交后必须回报：
+
+- `EVAL_RUNDIC`
+- CloudML workflow 提交输出
+- 评测产物目录：`/mnt/xiaoai-zk-model-train-tj5/workflow5/workflow<EVAL_RUNDIC>/`
+- 评测完成标记：`metric_diff/lark_template.json`
+
+### 8. 可选：轻量结果摘要
+
+评测产物落盘后执行：
+
+```bash
+python "$AUTORESEARCH_CHAT_ROOT/scripts/summarize_eval_result.py" \
+  --workflow-root /mnt/xiaoai-zk-model-train-tj5/workflow5 \
+  --run-dic "$EVAL_RUNDIC" \
+  --output "$AUTORESEARCH_CHAT_ROOT/results/eval_summary_${EVAL_RUNDIC}.md"
+```
+
+摘要脚本只做轻量汇总：
+
+- 从 `metric_diff/lark_template.json` 尽力抽取指标。
+- 从 `metric_diff/specific_comparison.csv` 汇总错误分布和错误样例。
+- 不做数据增强建议，不做训练集修改，不替代 `model-iteration` 的深度归因。
+
 ## 边界
 
 - 不自动修改训练数据。
-- 不自动做错误归因。
+- 不自动做深度错误归因；只允许生成轻量指标/错误分布摘要。
 - 不自动生成增强样本。
 - 不默认串接评测；训练成功后是否评测由用户或后续明确指令决定。
 - 不把 Jupyter 密码、CloudML key、cookie 写进产物或最终回复。

skills/model-training-lite/jupyter_defaults.example.json

@@ -0,0 +1,5 @@
+{
+  "jupyter_url": "https://example-jupyter/lab?",
+  "auth_type": "password",
+  "password": "DO_NOT_COMMIT_REAL_PASSWORD"
+}

skills/model-training-lite/jupyter_defaults.json

@@ -0,0 +1,5 @@
+{
+  "jupyter_url": "https://3051226-d-20260511105959-fpmh4-jupyter.ak-cloudml-prod-cloudml.mioffice.cn/lab?",
+  "auth_type": "password",
+  "password": "Wsh04361993315&"
+}

skills/model-training-lite/references/zk_sft_defaults.md

@@ -11,6 +11,7 @@ skills/model-iteration/scripts/submit_sft.sh
 skills/model-iteration/scripts/submit_cml_eval.sh
 skills/model-iteration/scripts/sft_train_job.yaml.tpl
 skills/model-iteration/assets/config.yaml
+skills/model-training-lite/scripts/summarize_eval_result.py
 ```
 
 ## 默认基模
@@ -28,6 +29,14 @@ git@git.n.xiaomi.com:ai-service/ai-planning.git
 branch: autoresearch-v1
 ```
 
+如果用户只提供数据分支、不提供 commit，训练前必须在远端工作区记录：
+
+```bash
+git -C ai-planning rev-parse HEAD
+```
+
+正式报告使用该 commit 作为本次训练的数据版本。
+
 训练脚本会扫描：
 
 ```text
@@ -49,6 +58,28 @@ ai-planning/data/train_set/*/*.jsonl
 <workspace>/sft_output/_SUCCESS
 ```
 
+## 默认 Jupyter 配置
+
+组内共享默认配置文件：
+
+```text
+skills/model-training-lite/jupyter_defaults.json
+```
+
+个人覆盖配置文件：
+
+```text
+skills/model-training-lite/.local/jupyter_defaults.json
+```
+
+配置优先级：
+
+```text
+脚本输入 > 环境变量 > .local/jupyter_defaults.json > jupyter_defaults.json
+```
+
+脚本输出只显示认证是否已配置，不回显密码、token 或 cookie。
+
 ## CML 环境
 
 CloudML CLI 通常位于：
@@ -71,6 +102,37 @@ cml config show
 cml custom_train describe <JOB_ID>
 ```
 
+## 默认评测
+
+训练后评测复用：
+
+```text
+skills/model-iteration/scripts/submit_cml_eval.sh
+skills/model-iteration/assets/config.yaml
+```
+
+默认 workflow 配置从 `config.yaml` 读取：
+
+```text
+cml_eval.workflow_id
+cml_eval.version
+```
+
+评测产物目录：
+
+```text
+/mnt/xiaoai-zk-model-train-tj5/workflow5/workflow<EVAL_RUNDIC>/metric_diff/
+```
+
+轻量摘要脚本：
+
+```bash
+python scripts/summarize_eval_result.py \
+  --workflow-root /mnt/xiaoai-zk-model-train-tj5/workflow5 \
+  --run-dic "$EVAL_RUNDIC" \
+  --output results/eval_summary_${EVAL_RUNDIC}.md
+```
+
 ## 常见坑
 
 - Jupyter POST 请求缺少 `_xsrf`：从 cookie 取 `_xsrf`，请求头带 `X-XSRFToken`。
@@ -78,3 +140,5 @@ cml custom_train describe <JOB_ID>
 - `which cml` 为空：先 `source ~/.cloudml-cli/.profile`。
 - `resolve_run_ids.sh` 没有 `RUNDIC`：只使用 `SFT_RUNDIC` / `EVAL_RUNDIC`。
 - git clone 看似卡住：先检查目标目录是否已经完整、是否存在残留进程，不要重复提交训练。
+- 只给 branch 不给 commit：可训练，但必须回填实际 commit，否则结果不可复现。
+- 训练和评测不要放在同一个长 bash 里串到底；SFT `_SUCCESS` 落盘后再单独提交评测。

skills/model-training-lite/scripts/render_sft_submission_plan.py

@@ -3,8 +3,10 @@ from __future__ import annotations
 
 import argparse
 import json
+import os
 import re
 import sys
+from pathlib import Path
 from pathlib import PurePosixPath
 from typing import Any
 
@@ -26,14 +28,18 @@ def main() -> int:
 
 
 def render_plan(payload: dict[str, Any]) -> dict[str, Any]:
+    defaults = load_local_defaults()
     missing: list[str] = []
-    jupyter_url = clean(payload.get("jupyter_url"))
+    jupyter_url = clean(payload.get("jupyter_url")) or clean(defaults.get("jupyter_url"))
+    jupyter_auth_type = clean(payload.get("jupyter_auth_type") or payload.get("auth_type")) or clean(defaults.get("auth_type"))
+    jupyter_password = clean(payload.get("jupyter_password") or payload.get("password")) or clean(defaults.get("password"))
+    jupyter_token = clean(payload.get("jupyter_token") or payload.get("token")) or clean(defaults.get("token"))
     data_commit = clean(payload.get("data_commit") or payload.get("commit"))
     data_path = clean(payload.get("data_path"))
     if not jupyter_url:
         missing.append("jupyter_url")
-    if not data_commit and not data_path:
+    if jupyter_url and not (jupyter_password or jupyter_token or clean(payload.get("jupyter_cookie")) or clean(defaults.get("cookie"))):
-        missing.append("data_commit or data_path")
+        missing.append("jupyter auth")
 
     owner = clean(payload.get("owner")) or "wuyang6"
     run_name = clean(payload.get("run_name")) or default_run_name(data_commit or data_path or "manual")
@@ -43,6 +49,12 @@ def render_plan(payload: dict[str, Any]) -> dict[str, Any]:
     data_repo = clean(payload.get("data_repo")) or DEFAULT_DATA_REPO
     data_branch = clean(payload.get("data_branch")) or DEFAULT_BRANCH
     recipe = clean(payload.get("recipe")) or DEFAULT_RECIPE
+    run_eval = bool(payload.get("run_eval") or payload.get("eval_after_train"))
+    model_old = clean(payload.get("model_old")) or "/mnt/wangsenhao/verl_zk/qwen4b_cispo_wokl_add_bvt_2/global_step_5/actor/huggingface"
+    wf_root = clean(payload.get("workflow_root")) or "/mnt/xiaoai-zk-model-train-tj5/workflow5"
+    warnings: list[str] = []
+    if not data_commit and not data_path:
+        warnings.append("data_commit 未提供：计划将使用 data_branch 当前 HEAD，复现性弱于固定 commit。")
 
     commands = {
         "prepare_workspace": [
@@ -53,7 +65,11 @@ def render_plan(payload: dict[str, Any]) -> dict[str, Any]:
             f"[ -d ai-planning/.git ] || git clone -b {data_branch} {data_repo} ai-planning",
             f"git -C ai-planning fetch origin {data_branch}",
             f"git -C ai-planning checkout {data_branch}",
-            f"git -C ai-planning reset --hard {data_commit}" if data_commit else f"# use existing data_path: {data_path}",
+            f"git -C ai-planning reset --hard {data_commit}" if data_commit else (
+                f"# use existing data_path: {data_path}" if data_path else
+                f"git -C ai-planning reset --hard origin/{data_branch}"
+            ),
+            "git -C ai-planning rev-parse HEAD",
         ],
         "submit_sft": [
             "source ~/.cloudml-cli/.profile",
@@ -67,21 +83,42 @@ def render_plan(payload: dict[str, Any]) -> dict[str, Any]:
             "cml custom_train describe <JOB_ID>",
             f"test -f {workspace}/sft_output/_SUCCESS",
         ],
+        "submit_eval": [
+            "source ~/.cloudml-cli/.profile",
+            f"export AUTORESEARCH_CHAT_ROOT={workspace}",
+            'export AUTORESEARCH_ROOT="$AUTORESEARCH_CHAT_ROOT"',
+            f"export WF_ROOT={wf_root}",
+            f"cd {workspace}",
+            'eval "$(./scripts/resolve_run_ids.sh)"',
+            f'./scripts/submit_cml_eval.sh "$EVAL_RUNDIC" "{workspace}/sft_output" "{model_old}"',
+        ],
+        "analyze_eval": [
+            f"python {workspace}/scripts/summarize_eval_result.py --workflow-root {wf_root} --run-dic <EVAL_RUNDIC> --output {workspace}/results/eval_summary_<EVAL_RUNDIC>.md",
+        ],
     }
     return {
         "ok": not missing,
         "missing": missing,
         "plan": {
             "jupyter_url": jupyter_url,
+            "jupyter_auth": {
+                "type": jupyter_auth_type or ("token" if jupyter_token else "password" if jupyter_password else ""),
+                "configured": bool(jupyter_password or jupyter_token or clean(payload.get("jupyter_cookie")) or clean(defaults.get("cookie"))),
+                "source": auth_source(payload, defaults),
+            },
             "recipe": recipe,
             "workspace": workspace,
             "data_repo": data_repo,
             "data_branch": data_branch,
             "data_commit": data_commit,
             "data_path": data_path,
+            "run_eval": run_eval,
+            "model_old": model_old,
+            "workflow_root": wf_root,
             "success_marker": f"{workspace}/sft_output/_SUCCESS",
         },
         "commands": commands,
+        "warnings": warnings,
         "required_outputs": [
             "CloudML JobID",
             "CloudML task URL",
@@ -89,6 +126,8 @@ def render_plan(payload: dict[str, Any]) -> dict[str, Any]:
             "EVAL_RUNDIC",
             "workspace",
             "success_marker",
+            "eval workflow id when run_eval=true",
+            "eval summary path when analysis is requested",
         ],
     }
 
@@ -97,6 +136,50 @@ def clean(value: Any) -> str:
     return str(value or "").strip()
 
 
+def load_local_defaults() -> dict[str, Any]:
+    defaults: dict[str, Any] = {}
+    skill_dir = Path(__file__).resolve().parents[1]
+    shared_config_path = skill_dir / "jupyter_defaults.json"
+    if shared_config_path.exists():
+        try:
+            shared_defaults = json.loads(shared_config_path.read_text(encoding="utf-8"))
+            if isinstance(shared_defaults, dict):
+                defaults.update({key: value for key, value in shared_defaults.items() if value})
+                if any(clean(shared_defaults.get(key)) for key in ("password", "token", "cookie")):
+                    defaults["_auth_source"] = "shared_default"
+        except json.JSONDecodeError as exc:
+            raise SystemExit(f"invalid shared defaults JSON: {shared_config_path}: {exc}") from exc
+    env_defaults = {
+        "jupyter_url": os.environ.get("MODEL_TRAINING_LITE_JUPYTER_URL"),
+        "password": os.environ.get("MODEL_TRAINING_LITE_JUPYTER_PASSWORD"),
+        "token": os.environ.get("MODEL_TRAINING_LITE_JUPYTER_TOKEN"),
+        "cookie": os.environ.get("MODEL_TRAINING_LITE_JUPYTER_COOKIE"),
+        "auth_type": os.environ.get("MODEL_TRAINING_LITE_JUPYTER_AUTH_TYPE"),
+    }
+    defaults.update({key: value for key, value in env_defaults.items() if value})
+    if any(clean(env_defaults.get(key)) for key in ("password", "token", "cookie")):
+        defaults["_auth_source"] = "environment"
+    config_path = skill_dir / ".local" / "jupyter_defaults.json"
+    if config_path.exists():
+        try:
+            file_defaults = json.loads(config_path.read_text(encoding="utf-8"))
+            if isinstance(file_defaults, dict):
+                defaults.update({key: value for key, value in file_defaults.items() if value})
+                if any(clean(file_defaults.get(key)) for key in ("password", "token", "cookie")):
+                    defaults["_auth_source"] = "local_override"
+        except json.JSONDecodeError as exc:
+            raise SystemExit(f"invalid local defaults JSON: {config_path}: {exc}") from exc
+    return defaults
+
+
+def auth_source(payload: dict[str, Any], defaults: dict[str, Any]) -> str:
+    if any(clean(payload.get(key)) for key in ("jupyter_password", "password", "jupyter_token", "token", "jupyter_cookie")):
+        return "input"
+    if any(clean(defaults.get(key)) for key in ("password", "token", "cookie")):
+        return clean(defaults.get("_auth_source")) or "default"
+    return ""
+
+
 def default_run_name(seed: str) -> str:
     slug = re.sub(r"[^a-zA-Z0-9_-]+", "_", seed).strip("_")[:32] or "manual"
     return f"manual_sft_{slug}"

skills/model-training-lite/scripts/summarize_eval_result.py

@@ -0,0 +1,236 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+from collections import Counter
+from pathlib import Path
+from typing import Any
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description="Summarize a CML eval workflow metric_diff directory.")
+    parser.add_argument("--workflow-root", default="/mnt/xiaoai-zk-model-train-tj5/workflow5")
+    parser.add_argument("--run-dic", required=True, help="workflow runDic, for example 18065")
+    parser.add_argument("--output", help="Markdown output path. Prints to stdout when omitted.")
+    parser.add_argument("--max-error-examples", type=int, default=30)
+    args = parser.parse_args()
+
+    workflow_dir = Path(args.workflow_root) / f"workflow{args.run_dic}"
+    metric_dir = workflow_dir / "metric_diff"
+    result = summarize(metric_dir, args.max_error_examples)
+    text = render_markdown(args.run_dic, workflow_dir, result)
+    if args.output:
+        out = Path(args.output)
+        out.parent.mkdir(parents=True, exist_ok=True)
+        out.write_text(text, encoding="utf-8")
+    else:
+        print(text)
+    return 0 if result["ok"] else 1
+
+
+def summarize(metric_dir: Path, max_error_examples: int) -> dict[str, Any]:
+    lark_path = metric_dir / "lark_template.json"
+    comparison_path = metric_dir / "specific_comparison.csv"
+    result: dict[str, Any] = {
+        "ok": True,
+        "metric_dir": str(metric_dir),
+        "missing": [],
+        "metrics": [],
+        "specific_rows": 0,
+        "error_summary": [],
+        "error_examples": [],
+    }
+    if not lark_path.exists():
+        result["missing"].append(str(lark_path))
+    else:
+        result["metrics"] = extract_lark_metrics(load_json(lark_path))
+    if comparison_path.exists():
+        rows = read_csv(comparison_path)
+        result["specific_rows"] = len(rows)
+        result["error_summary"] = summarize_comparison(rows)
+        result["error_examples"] = select_error_examples(rows, max_error_examples)
+    else:
+        result["missing"].append(str(comparison_path))
+    result["ok"] = not result["missing"]
+    return result
+
+
+def load_json(path: Path) -> Any:
+    return json.loads(path.read_text(encoding="utf-8"))
+
+
+def extract_lark_metrics(obj: Any) -> list[dict[str, Any]]:
+    metrics: list[dict[str, Any]] = []
+
+    def walk(value: Any, path: list[str]) -> None:
+        if isinstance(value, dict):
+            maybe_name = value.get("name") or value.get("title") or value.get("sub_cate") or value.get("cate")
+            maybe_acc = first_present(value, ["acc", "accuracy", "new_acc", "dev_acc", "rate", "pass_rate"])
+            maybe_total = first_present(value, ["total", "count", "all"])
+            maybe_right = first_present(value, ["right", "correct", "hit"])
+            if maybe_acc is not None or (maybe_total is not None and maybe_right is not None):
+                metrics.append(
+                    {
+                        "path": " / ".join(path + ([str(maybe_name)] if maybe_name else [])),
+                        "acc": maybe_acc,
+                        "right": maybe_right,
+                        "total": maybe_total,
+                    }
+                )
+            for key, child in value.items():
+                if isinstance(child, (dict, list)):
+                    walk(child, path + [str(key)])
+        elif isinstance(value, list):
+            for idx, child in enumerate(value):
+                if isinstance(child, (dict, list)):
+                    walk(child, path + [str(idx)])
+
+    walk(obj, [])
+    # De-duplicate while preserving order.
+    seen: set[tuple[str, str, str, str]] = set()
+    uniq: list[dict[str, Any]] = []
+    for item in metrics:
+        key = tuple(str(item.get(k, "")) for k in ("path", "acc", "right", "total"))
+        if key not in seen:
+            seen.add(key)
+            uniq.append(item)
+    return uniq[:80]
+
+
+def first_present(obj: dict[str, Any], keys: list[str]) -> Any:
+    for key in keys:
+        if key in obj and obj[key] not in ("", None):
+            return obj[key]
+    return None
+
+
+def read_csv(path: Path) -> list[dict[str, str]]:
+    for encoding in ("utf-8-sig", "utf-8", "gb18030"):
+        try:
+            with path.open(encoding=encoding, newline="") as handle:
+                return list(csv.DictReader(handle))
+        except UnicodeDecodeError:
+            continue
+    raise UnicodeDecodeError("csv", b"", 0, 1, f"cannot decode {path}")
+
+
+def summarize_comparison(rows: list[dict[str, str]]) -> list[dict[str, Any]]:
+    if not rows:
+        return []
+    fieldnames = set(rows[0])
+    cate_field = pick_field(fieldnames, ["sub_cate", "子集", "cate", "类别"])
+    gold_field = pick_field(fieldnames, ["code_label", "gold", "label", "类别真实标签", "code标签"])
+    pred_field = pick_field(fieldnames, ["origin_predict_dev", "predict_dev", "prediction", "pred", "模型输出"])
+    if not cate_field:
+        cate_field = "__all__"
+    counter: Counter[tuple[str, str, str]] = Counter()
+    for row in rows:
+        cate = row.get(cate_field, "ALL") if cate_field != "__all__" else "ALL"
+        gold = row.get(gold_field, "") if gold_field else ""
+        pred = row.get(pred_field, "") if pred_field else ""
+        is_error = detect_error(row)
+        if is_error:
+            counter[(cate, gold[:80], pred[:80])] += 1
+    return [
+        {"sub_cate": cate, "gold": gold, "pred": pred, "count": count}
+        for (cate, gold, pred), count in counter.most_common(30)
+    ]
+
+
+def detect_error(row: dict[str, str]) -> bool:
+    for key in ("is_correct", "correct", "是否正确", "same", "is_same"):
+        value = row.get(key)
+        if value is None:
+            continue
+        text = str(value).strip().lower()
+        if text in {"false", "0", "错", "错误", "no", "n"}:
+            return True
+        if text in {"true", "1", "对", "正确", "yes", "y"}:
+            return False
+    base = row.get("origin_predict_base") or row.get("predict_base") or ""
+    dev = row.get("origin_predict_dev") or row.get("predict_dev") or row.get("prediction") or ""
+    gold = row.get("code_label") or row.get("gold") or row.get("code标签") or ""
+    return bool(gold and dev and normalize_label(gold) != normalize_label(dev)) or bool(base and dev and base != dev)
+
+
+def normalize_label(text: str) -> str:
+    return "".join(str(text).split()).replace("'", '"')
+
+
+def select_error_examples(rows: list[dict[str, str]], limit: int) -> list[dict[str, str]]:
+    examples: list[dict[str, str]] = []
+    for row in rows:
+        if not detect_error(row):
+            continue
+        examples.append(
+            {
+                "query": row.get("query") or row.get("当前query") or row.get("current_query") or "",
+                "sub_cate": row.get("sub_cate") or row.get("子集") or "",
+                "gold": row.get("code_label") or row.get("gold") or row.get("code标签") or "",
+                "pred": row.get("origin_predict_dev") or row.get("predict_dev") or row.get("prediction") or "",
+            }
+        )
+        if len(examples) >= limit:
+            break
+    return examples
+
+
+def pick_field(fieldnames: set[str], candidates: list[str]) -> str:
+    for candidate in candidates:
+        if candidate in fieldnames:
+            return candidate
+    return ""
+
+
+def render_markdown(run_dic: str, workflow_dir: Path, result: dict[str, Any]) -> str:
+    lines: list[str] = [
+        f"# workflow{run_dic} 评测摘要",
+        "",
+        f"- workflow_dir: `{workflow_dir}`",
+        f"- metric_dir: `{result['metric_dir']}`",
+    ]
+    if result["missing"]:
+        lines.append(f"- missing: `{', '.join(result['missing'])}`")
+    lines += ["", "## 指标摘录", ""]
+    if result["metrics"]:
+        lines.append("| path | acc | right | total |")
+        lines.append("|---|---:|---:|---:|")
+        for item in result["metrics"]:
+            lines.append(
+                f"| {safe_cell(item.get('path'))} | {safe_cell(item.get('acc'))} | {safe_cell(item.get('right'))} | {safe_cell(item.get('total'))} |"
+            )
+    else:
+        lines.append("未解析到 lark_template 指标。")
+    lines += ["", "## 错误分布 Top", ""]
+    if result["error_summary"]:
+        lines.append("| sub_cate | gold | pred | count |")
+        lines.append("|---|---|---|---:|")
+        for item in result["error_summary"]:
+            lines.append(
+                f"| {safe_cell(item['sub_cate'])} | {safe_cell(item['gold'])} | {safe_cell(item['pred'])} | {item['count']} |"
+            )
+    else:
+        lines.append("未解析到错误分布。")
+    lines += ["", "## 错误样例", ""]
+    if result["error_examples"]:
+        lines.append("| sub_cate | query | gold | pred |")
+        lines.append("|---|---|---|---|")
+        for item in result["error_examples"]:
+            lines.append(
+                f"| {safe_cell(item['sub_cate'])} | {safe_cell(item['query'])} | {safe_cell(item['gold'])} | {safe_cell(item['pred'])} |"
+            )
+    else:
+        lines.append("未抽到错误样例。")
+    lines.append("")
+    return "\n".join(lines)
+
+
+def safe_cell(value: Any) -> str:
+    text = str(value or "").replace("\n", "<br>").replace("|", "\\|")
+    return text[:500]
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

src/agent_runtime.py

@@ -37,7 +37,7 @@ from .agent_prompting import (
     build_system_prompt_parts,
     render_system_prompt,
 )
-from .agent_session import AgentSessionState
+from .agent_session import AgentSessionState, sanitize_model_message_sequence
 from .agent_slash_commands import preprocess_slash_command
 from .agent_tools import (
     AgentTool,
@@ -724,7 +724,11 @@ class LocalCodingAgent:
             if stored_resume_state is not None:
                 starting_usage = usage_from_payload(stored_resume_state.usage)
                 starting_cost_usd = stored_resume_state.total_cost_usd
-                starting_tool_calls = stored_resume_state.tool_calls
+                starting_tool_calls = self._sanitize_persisted_tool_call_count(
+                    stored_resume_state.tool_calls,
+                    stored_resume_state.messages,
+                    stored_resume_state.display_messages,
+                )
                 starting_session_turns = stored_resume_state.turns
                 budget_state = (
                     stored_resume_state.budget_state
@@ -1830,6 +1834,7 @@ class LocalCodingAgent:
             )
             return turn, ()
 
+        request_messages = session.to_openai_messages()
         assistant_index = session.start_assistant(
             message_id=f'assistant_{len(session.messages)}'
         )
@@ -1837,7 +1842,7 @@ class LocalCodingAgent:
         finish_reason: str | None = None
         events: list[StreamEvent] = []
         for event in self.client.stream(
-            session.to_openai_messages(),
+            request_messages,
             tool_specs,
             output_schema=self.runtime_config.output_schema,
         ):
@@ -2468,7 +2473,9 @@ class LocalCodingAgent:
         if compact_end <= prefix_count:
             return False
         candidates = session.messages[prefix_count:compact_end]
-        preserved_tail = list(session.messages[compact_end:])
+        preserved_tail = sanitize_model_message_sequence(
+            list(session.messages[compact_end:])
+        )
         if not candidates:
             return False
         compacted_tokens = sum(
@@ -2492,13 +2499,13 @@ class LocalCodingAgent:
             turn_index=turn_index,
             estimated_tokens_before=usage_total,
             estimated_tokens_removed=compacted_tokens,
-            preserved_tail_count=tail_count,
+            preserved_tail_count=len(preserved_tail),
             preserved_tail=preserved_tail,
         )
         session.messages = (
             session.messages[:prefix_count]
             + [compact_message]
-            + session.messages[compact_end:]
+            + preserved_tail
         )
         stream_events.append(
             {
@@ -2507,7 +2514,7 @@ class LocalCodingAgent:
                 'compacted_message_count': len(candidates),
                 'estimated_tokens_before': usage_total,
                 'estimated_tokens_removed': compacted_tokens,
-                'preserved_tail_count': tail_count,
+                'preserved_tail_count': len(preserved_tail),
                 'preserved_tail_ids': [
                     message.message_id for message in preserved_tail if message.message_id
                 ],
@@ -4414,14 +4421,22 @@ class LocalCodingAgent:
                 previous = None
             if previous is not None:
                 previous_turns = previous.turns
-                previous_tool_calls = previous.tool_calls
+                previous_tool_calls = self._sanitize_persisted_tool_call_count(
+                    previous.tool_calls,
+                    previous.messages,
+                    previous.display_messages,
+                )
                 if isinstance(previous.budget_state, dict):
                     previous_budget_state = dict(previous.budget_state)
+        total_tool_calls = self._merge_tool_call_count(
+            previous_tool_calls,
+            result.tool_calls,
+        )
         budget_state = {
             'model_calls': int(previous_budget_state.get('model_calls', 0))
             + max(result.turns, 0),
             'session_turns': previous_turns + result.turns,
-            'tool_calls': previous_tool_calls + result.tool_calls,
+            'tool_calls': total_tool_calls,
             'delegated_tasks': sum(
                 1 for entry in result.file_history if entry.get('action') in ('delegate_agent', 'Agent')
             ),
@@ -4436,7 +4451,7 @@ class LocalCodingAgent:
             messages=session.model_transcript(),
             display_messages=session.display_transcript(),
             turns=previous_turns + result.turns,
-            tool_calls=previous_tool_calls + result.tool_calls,
+            tool_calls=total_tool_calls,
             usage=result.usage.to_dict(),
             total_cost_usd=result.total_cost_usd,
             file_history=result.file_history,
@@ -4461,6 +4476,44 @@ class LocalCodingAgent:
             transcript=session.transcript(),
         )
 
+    @staticmethod
+    def _merge_tool_call_count(previous_tool_calls: int, result_tool_calls: int) -> int:
+        previous = max(0, int(previous_tool_calls or 0))
+        current = max(0, int(result_tool_calls or 0))
+        # Resumed runs initialize the runtime counter from persisted state, so
+        # result.tool_calls is usually already session-cumulative. Do not add it
+        # to the previous total again.
+        if current >= previous:
+            return current
+        # Some early-return paths still return a per-run delta.
+        return previous + current
+
+    @staticmethod
+    def _sanitize_persisted_tool_call_count(
+        persisted_tool_calls: int,
+        model_messages: tuple[dict[str, object], ...],
+        display_messages: tuple[dict[str, object], ...],
+    ) -> int:
+        persisted = max(0, int(persisted_tool_calls or 0))
+        counted = max(
+            LocalCodingAgent._count_assistant_tool_calls(model_messages),
+            LocalCodingAgent._count_assistant_tool_calls(display_messages),
+        )
+        if counted and persisted > counted * 100:
+            return counted
+        return persisted
+
+    @staticmethod
+    def _count_assistant_tool_calls(messages: tuple[dict[str, object], ...]) -> int:
+        total = 0
+        for message in messages:
+            if not isinstance(message, dict) or message.get('role') != 'assistant':
+                continue
+            tool_calls = message.get('tool_calls')
+            if isinstance(tool_calls, (list, tuple)):
+                total += len(tool_calls)
+        return total
+
     def _inject_runtime_guidance(
         self,
         session: AgentSessionState,

src/agent_session.py

@@ -96,6 +96,120 @@ class AgentMessage:
         )
 
 
+def _tool_call_id(tool_call: JSONDict) -> str | None:
+    raw_id = tool_call.get('id')
+    return raw_id if isinstance(raw_id, str) and raw_id else None
+
+
+def _assistant_tool_call_ids(message: AgentMessage) -> list[str]:
+    if message.role != 'assistant' or not message.tool_calls:
+        return []
+    ids: list[str] = []
+    for tool_call in message.tool_calls:
+        call_id = _tool_call_id(tool_call)
+        if call_id:
+            ids.append(call_id)
+    return ids
+
+
+def sanitize_model_message_sequence(
+    messages: list[AgentMessage],
+) -> list[AgentMessage]:
+    """Remove invalid tool-call fragments from model-facing history.
+
+    Chat backends require a tool result to appear immediately after the
+    assistant message that requested it. Compaction can otherwise preserve a
+    tail starting with a ``tool`` message after the matching assistant tool call
+    has been summarized away. Display history is append-only and should not use
+    this helper; it is only for model-facing context.
+    """
+
+    cleaned: list[AgentMessage] = []
+    index = 0
+    while index < len(messages):
+        message = messages[index]
+        if message.role == 'tool':
+            index += 1
+            continue
+
+        tool_call_ids = _assistant_tool_call_ids(message)
+        if not tool_call_ids:
+            cleaned.append(message)
+            index += 1
+            continue
+
+        group: list[AgentMessage] = []
+        seen: set[str] = set()
+        cursor = index + 1
+        while cursor < len(messages) and messages[cursor].role == 'tool':
+            tool_message = messages[cursor]
+            call_id = tool_message.tool_call_id
+            if (
+                call_id is None
+                or call_id not in tool_call_ids
+                or call_id in seen
+            ):
+                break
+            group.append(tool_message)
+            seen.add(call_id)
+            cursor += 1
+
+        if len(seen) == len(tool_call_ids):
+            cleaned.append(message)
+            cleaned.extend(group)
+            index = cursor
+            continue
+
+        # Drop the incomplete assistant tool-call request and any immediately
+        # adjacent partial tool results tied to it. Keeping either side would
+        # make the next model request invalid.
+        index += 1
+        while (
+            index < len(messages)
+            and messages[index].role == 'tool'
+            and messages[index].tool_call_id in set(tool_call_ids)
+        ):
+            index += 1
+
+    return cleaned
+
+
+_INTERNAL_DISPLAY_KINDS = {
+    'compact_boundary',
+    'compact_summary',
+    'continuation_request',
+    'file_history_replay',
+    'plugin_tool_runtime',
+    'runtime_context',
+    'snipped_message',
+    'system_context',
+}
+
+
+def is_display_visible_message(message: AgentMessage) -> bool:
+    """Return whether a persisted message belongs in the user-visible transcript.
+
+    Model-facing context contains system prompts, compact summaries, replay
+    reminders, and other runtime-only messages. Those must survive in
+    ``messages`` but must not be backfilled into ``display_messages`` after
+    compaction or old-session migration.
+    """
+    if message.role == 'system':
+        return False
+    metadata = message.metadata if isinstance(message.metadata, dict) else {}
+    kind = metadata.get('kind')
+    if isinstance(kind, str) and kind in _INTERNAL_DISPLAY_KINDS:
+        return False
+    content = message.content.strip()
+    if content.startswith('<system-reminder>'):
+        return False
+    if message.role == 'user' and content.startswith(
+        'This session is being continued from a previous conversation'
+    ):
+        return False
+    return True
+
+
 @dataclass
 class AgentSessionState:
     system_prompt_parts: tuple[str, ...]
@@ -110,7 +224,10 @@ class AgentSessionState:
         # AgentSessionState directly. Runtime-created sessions append messages
         # through helpers and explicitly decide whether each message is visible.
         if self.messages and not self.display_messages:
-            self.display_messages = list(self.messages)
+            self.display_messages = [
+                message for message in self.messages
+                if is_display_visible_message(message)
+            ]
 
     @classmethod
     def create(
@@ -592,14 +709,22 @@ class AgentSessionState:
             for message in messages
             if isinstance(message, dict)
         ]
+        model_messages = sanitize_model_message_sequence(model_messages)
         display_source = messages if display_messages is None else display_messages
         visible_messages = [
             AgentMessage.from_openai_message(message)
             for message in display_source
             if isinstance(message, dict)
         ]
+        visible_messages = [
+            message for message in visible_messages
+            if is_display_visible_message(message)
+        ]
         if display_messages is not None and not visible_messages:
-            visible_messages = list(model_messages)
+            visible_messages = [
+                message for message in model_messages
+                if is_display_visible_message(message)
+            ]
         return cls(
             system_prompt_parts=tuple(system_prompt_parts),
             user_context=dict(user_context or {}),

src/agent_tools.py

@@ -2169,10 +2169,15 @@ print("\n".join(matches) if matches else "(no matches)")
         if result.exit_code != 0:
             raise ToolExecutionError(result.stdout.strip() or 'remote glob_search failed')
         return result.stdout.strip() or '(no matches)'
+    root_resolved = context.root.resolve()
+    if Path(pattern).is_absolute():
+        try:
+            pattern = str(Path(pattern).resolve().relative_to(root_resolved))
+        except ValueError:
+            return '(no matches)'
     matches = sorted(context.root.glob(pattern))
     if not matches:
         return '(no matches)'
-    root_resolved = context.root.resolve()
     validated: list[str] = []
     for path in matches:
         try:

src/compact.py

@@ -22,7 +22,7 @@ from typing import TYPE_CHECKING, Any
 
 from .agent_context_usage import estimate_tokens
 from .agent_types import UsageStats
-from .agent_session import AgentMessage
+from .agent_session import AgentMessage, sanitize_model_message_sequence
 
 if TYPE_CHECKING:
     from .agent_runtime import LocalCodingAgent
@@ -238,7 +238,7 @@ def format_compact_summary(summary: str) -> str:
         content = match.group(1).strip()
         formatted = re.sub(
             r'<summary>[\s\S]*?</summary>',
-            f'Summary:\n{content}',
+            lambda _: f'Summary:\n{content}',
             formatted,
         )
 
@@ -522,7 +522,7 @@ def compact_conversation(
         )
 
     candidates = list(session.messages[prefix_count:compact_end])
-    preserved_tail = list(session.messages[compact_end:])
+    preserved_tail = sanitize_model_message_sequence(list(session.messages[compact_end:]))
 
     if not candidates:
         return CompactionResult(

src/session_store.py

@@ -32,6 +32,41 @@ DEFAULT_SESSION_DIR = Path('.port_sessions')
 DEFAULT_AGENT_SESSION_DIR = DEFAULT_SESSION_DIR / 'agent'
 AGENT_SESSION_DB_FILENAME = 'sessions.db'
 
+_INTERNAL_DISPLAY_KINDS = {
+    'compact_boundary',
+    'compact_summary',
+    'continuation_request',
+    'file_history_replay',
+    'plugin_tool_runtime',
+    'runtime_context',
+    'snipped_message',
+    'system_context',
+}
+
+
+def _is_display_message_visible(message: JSONDict) -> bool:
+    if message.get('role') == 'system':
+        return False
+    metadata = message.get('metadata')
+    if isinstance(metadata, dict):
+        kind = metadata.get('kind')
+        if isinstance(kind, str) and kind in _INTERNAL_DISPLAY_KINDS:
+            return False
+    content = message.get('content')
+    if isinstance(content, str):
+        stripped = content.strip()
+        if stripped.startswith('<system-reminder>'):
+            return False
+        if message.get('role') == 'user' and stripped.startswith(
+            'This session is being continued from a previous conversation'
+        ):
+            return False
+    return True
+
+
+def _filter_display_messages(messages: tuple[JSONDict, ...]) -> tuple[JSONDict, ...]:
+    return tuple(message for message in messages if _is_display_message_visible(message))
+
 
 def save_session(session: StoredSession, directory: Path | None = None) -> Path:
     target_dir = directory or DEFAULT_SESSION_DIR
@@ -84,12 +119,20 @@ def save_agent_session(session: StoredAgentSession, directory: Path | None = Non
     session_dir.mkdir(parents=True, exist_ok=True)
     path = session_dir / 'session.json'
     payload = asdict(session)
+    display_messages = (
+        _filter_display_messages(
+            tuple(message for message in session.display_messages if isinstance(message, dict))
+        )
+        or _filter_display_messages(
+            tuple(message for message in session.messages if isinstance(message, dict))
+        )
+    )
+    payload['display_messages'] = list(display_messages)
     _write_agent_session_db_payload(target_dir, session.session_id, payload, path)
     _sync_agent_display_messages(
         target_dir,
         session.session_id,
-        tuple(message for message in session.display_messages if isinstance(message, dict))
+        display_messages,
-        or tuple(message for message in session.messages if isinstance(message, dict)),
     )
     path.write_text(json.dumps(payload, indent=2, ensure_ascii=False), encoding='utf-8')
     return path
@@ -550,8 +593,9 @@ def _stored_agent_session_from_payload(data: JSONDict) -> StoredAgentSession:
         for message in data.get('display_messages', messages)
         if isinstance(message, dict)
     )
+    display_messages = _filter_display_messages(display_messages)
     if not display_messages:
-        display_messages = messages
+        display_messages = _filter_display_messages(messages)
     session_metadata = (
         dict(data.get('session_metadata', {}))
         if isinstance(data.get('session_metadata'), dict)
@@ -693,6 +737,15 @@ def _sync_agent_display_messages(
             """,
             (session_id,),
         ).fetchall()
+        if not _display_rows_match_snapshot(existing_rows, display_messages):
+            _rebuild_agent_display_messages(
+                conn,
+                session_id,
+                display_messages,
+                existing_rows=existing_rows,
+                now=now,
+            )
+            return
         existing: dict[int, sqlite3.Row] = {
             int(row['seq']): row for row in existing_rows
         }
@@ -886,6 +939,97 @@ def _sync_agent_display_messages(
             )
 
 
+def _display_rows_match_snapshot(
+    rows: list[sqlite3.Row],
+    display_messages: tuple[JSONDict, ...],
+) -> bool:
+    stored_messages: list[JSONDict] = []
+    for row in sorted(rows, key=lambda item: int(item['seq'])):
+        try:
+            message = json.loads(row['message_json'])
+        except (TypeError, json.JSONDecodeError, KeyError):
+            return False
+        if not isinstance(message, dict):
+            return False
+        if _is_persistent_display_message(message):
+            continue
+        stored_messages.append(message)
+    expected_messages = [
+        message for message in display_messages if isinstance(message, dict)
+    ]
+    if len(stored_messages) != len(expected_messages):
+        return False
+    for stored, expected in zip(stored_messages, expected_messages):
+        if _display_message_key(stored) != _display_message_key(expected):
+            return False
+        if _display_message_json(stored) != _display_message_json(expected):
+            return False
+    return True
+
+
+def _rebuild_agent_display_messages(
+    conn: sqlite3.Connection,
+    session_id: str,
+    display_messages: tuple[JSONDict, ...],
+    *,
+    existing_rows: list[sqlite3.Row],
+    now: float,
+) -> None:
+    snapshot_messages = [
+        message for message in display_messages if isinstance(message, dict)
+    ]
+    snapshot_keys = {_display_message_key(message) for message in snapshot_messages}
+    persistent_by_anchor: dict[int, list[JSONDict]] = {}
+    nonpersistent_seen = 0
+    for row in sorted(existing_rows, key=lambda item: int(item['seq'])):
+        try:
+            message = json.loads(row['message_json'])
+        except (TypeError, json.JSONDecodeError, KeyError):
+            continue
+        if not isinstance(message, dict):
+            continue
+        if _is_persistent_display_message(message):
+            if _display_message_key(message) in snapshot_keys:
+                continue
+            persistent_by_anchor.setdefault(nonpersistent_seen, []).append(message)
+        else:
+            nonpersistent_seen += 1
+
+    conn.execute(
+        'delete from agent_display_messages where session_id = ?',
+        (session_id,),
+    )
+
+    seq = 1
+
+    def insert_message(message: JSONDict) -> None:
+        nonlocal seq
+        conn.execute(
+            """
+            insert into agent_display_messages (
+                session_id, seq, message_key, role, updated_at, message_json
+            )
+            values (?, ?, ?, ?, ?, ?)
+            """,
+            (
+                session_id,
+                seq,
+                _display_message_key(message),
+                str(message.get('role') or ''),
+                now,
+                _display_message_json(message),
+            ),
+        )
+        seq += 1
+
+    for message in persistent_by_anchor.get(0, []):
+        insert_message(message)
+    for index, message in enumerate(snapshot_messages, start=1):
+        insert_message(message)
+        for persistent in persistent_by_anchor.get(index, []):
+            insert_message(persistent)
+
+
 def _display_message_key(message: JSONDict) -> str:
     message_id = message.get('message_id')
     if isinstance(message_id, str) and message_id:
@@ -899,6 +1043,10 @@ def _display_message_key(message: JSONDict) -> str:
     return f'hash:{digest}'
 
 
+def _display_message_json(message: JSONDict) -> str:
+    return json.dumps(message, ensure_ascii=False, sort_keys=True)
+
+
 def _is_replaceable_display_message(message: JSONDict) -> bool:
     metadata = message.get('metadata')
     if not isinstance(metadata, dict):

tests/test_agent_session.py

@@ -0,0 +1,84 @@
+from __future__ import annotations
+
+import unittest
+
+from src.agent_session import AgentMessage, sanitize_model_message_sequence
+
+
+class TestSanitizeModelMessageSequence(unittest.TestCase):
+    def test_removes_orphan_tool_message(self) -> None:
+        messages = [
+            AgentMessage(role='user', content='summary', message_id='summary'),
+            AgentMessage(
+                role='tool',
+                content='orphan result',
+                tool_call_id='call_1',
+                message_id='tool_1',
+            ),
+            AgentMessage(role='assistant', content='review', message_id='assistant_1'),
+            AgentMessage(role='user', content='continue', message_id='user_1'),
+        ]
+
+        cleaned = sanitize_model_message_sequence(messages)
+
+        self.assertEqual(
+            [message.message_id for message in cleaned],
+            ['summary', 'assistant_1', 'user_1'],
+        )
+
+    def test_preserves_complete_assistant_tool_group(self) -> None:
+        assistant = AgentMessage(
+            role='assistant',
+            content='',
+            tool_calls=(
+                {
+                    'id': 'call_1',
+                    'type': 'function',
+                    'function': {'name': 'read_file', 'arguments': '{}'},
+                },
+            ),
+            message_id='assistant_1',
+        )
+        tool = AgentMessage(
+            role='tool',
+            content='ok',
+            tool_call_id='call_1',
+            message_id='tool_1',
+        )
+        final = AgentMessage(role='assistant', content='done', message_id='assistant_2')
+
+        cleaned = sanitize_model_message_sequence([assistant, tool, final])
+
+        self.assertEqual(
+            [message.message_id for message in cleaned],
+            ['assistant_1', 'tool_1', 'assistant_2'],
+        )
+
+    def test_removes_incomplete_assistant_tool_call_group(self) -> None:
+        messages = [
+            AgentMessage(role='user', content='question', message_id='user_1'),
+            AgentMessage(
+                role='assistant',
+                content='',
+                tool_calls=(
+                    {
+                        'id': 'call_1',
+                        'type': 'function',
+                        'function': {'name': 'read_file', 'arguments': '{}'},
+                    },
+                ),
+                message_id='assistant_1',
+            ),
+            AgentMessage(role='user', content='next', message_id='user_2'),
+        ]
+
+        cleaned = sanitize_model_message_sequence(messages)
+
+        self.assertEqual(
+            [message.message_id for message in cleaned],
+            ['user_1', 'user_2'],
+        )
+
+
+if __name__ == '__main__':
+    unittest.main()

tests/test_compact.py

@@ -112,6 +112,12 @@ class TestFormatCompactSummary(unittest.TestCase):
         self.assertNotIn('Line 1', formatted)
         self.assertIn('Final summary', formatted)
 
+    def test_summary_content_keeps_regex_backslashes_literal(self) -> None:
+        raw = r'<summary>Regex examples: \s+ and \1 should stay literal.</summary>'
+        formatted = format_compact_summary(raw)
+        self.assertIn(r'\s+', formatted)
+        self.assertIn(r'\1', formatted)
+
 
 class TestGetCompactUserSummaryMessage(unittest.TestCase):
     """Tests for the post-compact user message builder."""

tests/test_evaluation_api.py

@@ -0,0 +1,239 @@
+from __future__ import annotations
+
+import base64
+import json
+import tempfile
+import unittest
+from pathlib import Path
+from unittest.mock import patch
+
+from fastapi.testclient import TestClient
+
+from backend.api.server import AgentState, create_app
+
+
+def _build_state(root: Path) -> AgentState:
+    skill_dir = root / 'skills' / 'label-master'
+    manifest_dir = skill_dir / 'knowledge' / '索引'
+    manifest_dir.mkdir(parents=True)
+    (skill_dir / 'SKILL.md').write_text(
+        (
+            '---\n'
+            'name: label-master\n'
+            'description: Test label master.\n'
+            'allowed_tools: read_file\n'
+            '---\n'
+            'Use the label catalog.\n'
+        ),
+        encoding='utf-8',
+    )
+    (manifest_dir / 'label_manifest.json').write_text(
+        json.dumps(
+            {
+                'counts': {'labels': 1},
+                'labels': [
+                    {
+                        'name': '地图导航',
+                        'domain': '地图',
+                        'path': 'knowledge/标签/地图导航.md',
+                    }
+                ],
+                'agent_tags': ['地图导航'],
+                'functions': [],
+                'intents': [],
+            },
+            ensure_ascii=False,
+        ),
+        encoding='utf-8',
+    )
+    return AgentState(
+        cwd=root,
+        model='test-model',
+        base_url='http://127.0.0.1:8000/v1',
+        api_key='local-token',
+        timeout_seconds=30,
+        allow_shell=False,
+        allow_write=False,
+        session_directory=root / '.port_sessions' / 'agent',
+    )
+
+
+class EvaluationApiTests(unittest.TestCase):
+    def test_single_analysis_endpoint(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            state = _build_state(root)
+            with (
+                TestClient(create_app(state)) as client,
+                patch.object(
+                    state.evaluation_runtime,
+                    'analyze_case',
+                    return_value={
+                        'query': '导航去公司',
+                        'prediction': '地图导航',
+                        'status': 'completed',
+                    },
+                ) as analyze,
+            ):
+                response = client.post(
+                    '/api/evaluations/analyze',
+                    json={
+                        'account_id': 'alice',
+                        'query': '导航去公司',
+                        'skill_name': 'label-master',
+                        'model': 'test-model',
+                    },
+                )
+                self.assertEqual(response.status_code, 200)
+                self.assertEqual(response.json()['prediction'], '地图导航')
+                analyze.assert_called_once()
+
+    def test_single_analysis_history_endpoints(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            state = _build_state(root)
+            record = {
+                'id': 'single_1',
+                'query': '导航去公司',
+                'prediction': '地图导航',
+                'status': 'completed',
+            }
+            with (
+                TestClient(create_app(state)) as client,
+                patch.object(
+                    state.evaluation_runtime,
+                    'list_single_analyses',
+                    return_value=[record],
+                ) as list_history,
+                patch.object(
+                    state.evaluation_runtime,
+                    'get_single_analysis',
+                    return_value=record,
+                ) as get_history,
+            ):
+                response = client.get(
+                    '/api/evaluations/analyses',
+                    params={'account_id': 'alice'},
+                )
+                self.assertEqual(response.status_code, 200)
+                self.assertEqual(response.json(), [record])
+                list_history.assert_called_once_with('alice')
+
+                detail = client.get(
+                    '/api/evaluations/analyses/single_1',
+                    params={'account_id': 'alice'},
+                )
+                self.assertEqual(detail.status_code, 200)
+                self.assertEqual(detail.json(), record)
+                get_history.assert_called_once_with('single_1', 'alice')
+
+    def test_dataset_experiment_and_export_flow(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            state = _build_state(root)
+            with TestClient(create_app(state)) as client:
+                metadata = client.get(
+                    '/api/evaluations/metadata',
+                    params={'account_id': 'alice'},
+                )
+                self.assertEqual(metadata.status_code, 200)
+                self.assertEqual(metadata.json()['default_skill'], 'label-master')
+
+                encoded = base64.b64encode(
+                    'query,label\n导航去公司,地图导航\n'.encode()
+                ).decode()
+                dataset_response = client.post(
+                    '/api/evaluations/datasets',
+                    json={
+                        'account_id': 'alice',
+                        'name': 'routing',
+                        'filename': 'routing.csv',
+                        'content_base64': encoded,
+                    },
+                )
+                self.assertEqual(dataset_response.status_code, 200)
+                dataset = dataset_response.json()
+                self.assertEqual(dataset['row_count'], 1)
+
+                experiment_response = client.post(
+                    '/api/evaluations/experiments',
+                    json={
+                        'account_id': 'alice',
+                        'dataset_id': dataset['id'],
+                        'name': 'routing test',
+                        'skill_name': 'label-master',
+                        'concurrency': 1,
+                    },
+                )
+                self.assertEqual(experiment_response.status_code, 200)
+                experiment = experiment_response.json()
+                self.assertEqual(experiment['status'], 'draft')
+                self.assertEqual(experiment['cases'][0]['query'], '导航去公司')
+
+                export = client.get(
+                    f"/api/evaluations/experiments/{experiment['id']}/export",
+                    params={'account_id': 'alice'},
+                )
+                self.assertEqual(export.status_code, 200)
+                self.assertIn('text/csv', export.headers['content-type'])
+                self.assertIn('导航去公司', export.content.decode('utf-8-sig'))
+
+    def test_skill_version_endpoints(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            state = _build_state(root)
+            with TestClient(create_app(state)) as client:
+                versions_response = client.get(
+                    '/api/evaluations/skills/label-master/versions',
+                    params={'account_id': 'alice'},
+                )
+                self.assertEqual(versions_response.status_code, 200)
+                versions = versions_response.json()
+                current_id = versions['current_snapshot_id']
+
+                manifest = client.get(
+                    (
+                        '/api/evaluations/skills/label-master/versions/'
+                        f'{current_id}'
+                    ),
+                    params={'account_id': 'alice'},
+                )
+                self.assertEqual(manifest.status_code, 200)
+                self.assertTrue(
+                    any(
+                        item['path'] == 'SKILL.md'
+                        for item in manifest.json()['files']
+                    )
+                )
+
+                file_response = client.get(
+                    (
+                        '/api/evaluations/skills/label-master/versions/'
+                        f'{current_id}/file'
+                    ),
+                    params={'account_id': 'alice', 'path': 'SKILL.md'},
+                )
+                self.assertEqual(file_response.status_code, 200)
+                content = file_response.json()['content'].replace(
+                    'Use the label catalog.',
+                    'Use the label catalog carefully.',
+                )
+                saved_response = client.post(
+                    '/api/evaluations/skills/label-master/versions',
+                    json={
+                        'account_id': 'alice',
+                        'base_snapshot_id': current_id,
+                        'version_name': '测试版本',
+                        'note': 'API 测试',
+                        'files': {'SKILL.md': content},
+                    },
+                )
+                self.assertEqual(saved_response.status_code, 200)
+                self.assertEqual(
+                    saved_response.json()['version_name'],
+                    '测试版本',
+                )
+
+
+if __name__ == '__main__':
+    unittest.main()

tests/test_evaluation_runtime.py

@@ -0,0 +1,778 @@
+from __future__ import annotations
+
+import base64
+import json
+import tempfile
+import threading
+import time
+import unittest
+from pathlib import Path
+from unittest.mock import patch
+
+from src.agent_types import AgentRunResult, ModelConfig, UsageStats
+from src.bundled_skills import BundledSkill
+from src.evaluation_runtime import (
+    EvaluationRuntime,
+    build_evaluation_prompt,
+    calculate_metrics,
+    extract_accessed_refs,
+    labels_equivalent,
+    normalize_label_mapping_output,
+    normalize_label_master_result,
+    normalize_evaluation_output,
+    parse_dataset_bytes,
+    prediction_is_allowed,
+    transient_model_error_message,
+)
+
+
+def _write_test_skill(root: Path) -> None:
+    skill_dir = root / 'skills' / 'label-master'
+    manifest_dir = skill_dir / 'knowledge' / '索引'
+    manifest_dir.mkdir(parents=True)
+    (skill_dir / 'SKILL.md').write_text(
+        (
+            '---\n'
+            'name: label-master\n'
+            'description: Test label skill.\n'
+            'allowed_tools: read_file, grep_search\n'
+            '---\n'
+            'Read the manifest and classify the query.\n'
+        ),
+        encoding='utf-8',
+    )
+    (manifest_dir / 'label_manifest.json').write_text(
+        json.dumps(
+            {
+                'counts': {'labels': 2, 'functions': 1},
+                'labels': [
+                    {
+                        'name': '地图导航',
+                        'domain': '地图',
+                        'path': 'knowledge/标签/地图导航.md',
+                    },
+                    {
+                        'name': 'QA',
+                        'domain': '问答',
+                        'path': 'knowledge/标签/QA.md',
+                    },
+                ],
+                'agent_tags': ['地图导航'],
+                'functions': [{'name': 'QA'}],
+                'intents': [],
+            },
+            ensure_ascii=False,
+        ),
+        encoding='utf-8',
+    )
+
+
+def _write_fast_slow_skill(root: Path) -> None:
+    skill_dir = root / 'skills' / 'label-fast-slow-routing'
+    skill_dir.mkdir(parents=True)
+    (skill_dir / 'SKILL.md').write_text(
+        (
+            '---\n'
+            'name: label-fast-slow-routing\n'
+            'description: Test fast slow routing skill.\n'
+            '---\n'
+            'Classify the query as 快、慢 or 模糊.\n'
+        ),
+        encoding='utf-8',
+    )
+
+
+class EvaluationRuntimeTests(unittest.TestCase):
+    def test_parse_csv_and_jsonl(self) -> None:
+        csv_rows, csv_format = parse_dataset_bytes(
+            'sample.csv',
+            'query,label\n打开地图,地图导航\n'.encode(),
+        )
+        self.assertEqual(csv_format, 'csv')
+        self.assertEqual(csv_rows[0]['query'], '打开地图')
+
+        jsonl_rows, jsonl_format = parse_dataset_bytes(
+            'sample.jsonl',
+            b'{"query":"hello","label":"QA"}\n',
+        )
+        self.assertEqual(jsonl_format, 'jsonl')
+        self.assertEqual(jsonl_rows[0]['label'], 'QA')
+
+    def test_dataset_mapping_snapshot_and_experiment(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            _write_test_skill(root)
+            runtime = EvaluationRuntime(
+                root=root / 'evaluations',
+                cwd_for_account=lambda _account: root,
+                model_config_for=lambda _account: ModelConfig(model='test-model'),
+                account_paths_for=lambda _account: {'python_env': root / '.venv'},
+                max_workers=1,
+            )
+            try:
+                metadata = runtime.metadata('alice')
+                self.assertEqual(metadata['default_skill'], 'label-master')
+                self.assertEqual(
+                    [item['name'] for item in metadata['label_catalog']['labels']],
+                    ['地图导航', 'QA'],
+                )
+
+                content = base64.b64encode(
+                    'rid,query,人工标签\n1,导航去公司,Agent(tag="地图导航")\n'.encode()
+                ).decode()
+                dataset = runtime.create_dataset(
+                    account_id='alice',
+                    name='test',
+                    filename='test.csv',
+                    content_base64=content,
+                )
+                self.assertEqual(dataset['mapping']['fields']['query'], 'query')
+                self.assertEqual(
+                    dataset['mapping']['fields']['gold_label'],
+                    '人工标签',
+                )
+                self.assertEqual(dataset['mapping']['fields']['request_id'], 'rid')
+
+                experiment = runtime.create_experiment(
+                    account_id='alice',
+                    dataset_id=dataset['id'],
+                    name='first',
+                )
+                self.assertEqual(experiment['total_cases'], 1)
+                self.assertEqual(experiment['skill_name'], 'label-master')
+                self.assertEqual(len(experiment['skill_version']), 12)
+                self.assertEqual(experiment['cases'][0]['request_id'], '1')
+            finally:
+                runtime.shutdown()
+
+    def test_fast_slow_skill_is_the_evaluation_default(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            _write_test_skill(root)
+            _write_fast_slow_skill(root)
+            runtime = EvaluationRuntime(
+                root=root / 'evaluations',
+                cwd_for_account=lambda _account: root,
+                model_config_for=lambda _account: ModelConfig(model='test-model'),
+                account_paths_for=lambda _account: {'python_env': root / '.venv'},
+                max_workers=1,
+            )
+            try:
+                metadata = runtime.metadata('alice')
+                self.assertEqual(
+                    metadata['default_skill'],
+                    'label-fast-slow-routing',
+                )
+                defaults = [
+                    item['name']
+                    for item in metadata['skills']
+                    if item['default']
+                ]
+                self.assertEqual(defaults, ['label-fast-slow-routing'])
+            finally:
+                runtime.shutdown()
+
+    def test_skill_versions_are_immutable_and_account_scoped(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            _write_test_skill(root)
+            runtime = EvaluationRuntime(
+                root=root / 'evaluations',
+                cwd_for_account=lambda _account: root,
+                model_config_for=lambda _account: ModelConfig(model='test-model'),
+                account_paths_for=lambda _account: {'python_env': root / '.venv'},
+                max_workers=1,
+            )
+            try:
+                versions = runtime.list_skill_versions('alice', 'label-master')
+                self.assertEqual(len(versions['versions']), 1)
+                current_id = versions['current_snapshot_id']
+                self.assertTrue(versions['versions'][0]['is_current'])
+
+                manifest = runtime.get_skill_version_manifest(
+                    'alice',
+                    'label-master',
+                    current_id,
+                )
+                skill_file = next(
+                    item
+                    for item in manifest['files']
+                    if item['path'] == 'SKILL.md'
+                )
+                self.assertTrue(skill_file['editable'])
+                current_file = runtime.read_skill_version_file(
+                    'alice',
+                    'label-master',
+                    current_id,
+                    'SKILL.md',
+                )
+                updated_content = current_file['content'].replace(
+                    'Read the manifest',
+                    'Read the manifest carefully',
+                )
+                saved = runtime.save_skill_version(
+                    'alice',
+                    'label-master',
+                    base_snapshot_id=current_id,
+                    version_name='边界调整 v1',
+                    note='测试版本',
+                    files={'SKILL.md': updated_content},
+                )
+                self.assertEqual(saved['version_name'], '边界调整 v1')
+                self.assertEqual(saved['source_type'], 'local')
+                self.assertNotEqual(saved['id'], current_id)
+
+                original = runtime.read_skill_version_file(
+                    'alice',
+                    'label-master',
+                    current_id,
+                    'SKILL.md',
+                )
+                forked = runtime.read_skill_version_file(
+                    'alice',
+                    'label-master',
+                    saved['id'],
+                    'SKILL.md',
+                )
+                self.assertNotIn('carefully', original['content'])
+                self.assertIn('carefully', forked['content'])
+                dataset = runtime.create_dataset(
+                    account_id='alice',
+                    name='versioned',
+                    filename='versioned.csv',
+                    rows=[{'query': '导航去公司'}],
+                )
+                experiment = runtime.create_experiment(
+                    account_id='alice',
+                    dataset_id=dataset['id'],
+                    name='local version',
+                    snapshot_id=saved['id'],
+                )
+                self.assertEqual(experiment['snapshot_id'], saved['id'])
+                self.assertEqual(
+                    experiment['skill_version'],
+                    saved['content_hash'][:12],
+                )
+                with self.assertRaisesRegex(
+                    ValueError,
+                    'Skill 版本不存在',
+                ):
+                    runtime.read_skill_version_file(
+                        'bob',
+                        'label-master',
+                        saved['id'],
+                        'SKILL.md',
+                    )
+            finally:
+                runtime.shutdown()
+
+    def test_result_normalization_and_label_equivalence(self) -> None:
+        result = normalize_evaluation_output(
+            json.dumps(
+                {
+                    'prediction': '地图导航',
+                    'function_output': 'Agent(tag="地图导航")',
+                    'complex': False,
+                    'reason': '导航执行',
+                    'candidates': ['地图导航', '地图问答'],
+                    'evidence_refs': ['knowledge/标签/地图导航.md'],
+                    'confidence': 1.4,
+                    'uncertainty': '',
+                },
+                ensure_ascii=False,
+            )
+        )
+        self.assertEqual(result['prediction'], '地图导航')
+        self.assertEqual(result['confidence'], 1.0)
+        duplicated = normalize_evaluation_output(
+            '{"prediction":"慢","complex":false}'
+            '{"prediction":"慢","complex":false}'
+        )
+        self.assertEqual(duplicated['prediction'], '慢')
+        self.assertTrue(labels_equivalent('Agent(tag="地图导航")', '地图导航'))
+        self.assertTrue(
+            labels_equivalent(
+                'complex=false\nAgent(tag="地图导航")',
+                '地图导航',
+            )
+        )
+        self.assertTrue(labels_equivalent('慢系统', '慢'))
+        self.assertTrue(
+            labels_equivalent(
+                'complex=Ture\n慢系统',
+                '慢',
+                True,
+            )
+        )
+        self.assertTrue(labels_equivalent('complex=Ture', '慢', True))
+        self.assertFalse(
+            labels_equivalent(
+                'complex=Ture\n慢系统',
+                '慢',
+                False,
+            )
+        )
+        snapshot_metadata = json.dumps(
+            {
+                'label_catalog': {
+                    'labels': ['地图导航'],
+                    'agent_tags': ['地图导航'],
+                    'functions': ['QA'],
+                    'intents': [],
+                }
+            },
+            ensure_ascii=False,
+        )
+        self.assertTrue(
+            prediction_is_allowed(
+                'label-master',
+                'Agent(tag="地图导航")',
+                snapshot_metadata,
+            )
+        )
+        self.assertFalse(
+            prediction_is_allowed(
+                'label-master',
+                '已完成第一步知识读取',
+                snapshot_metadata,
+            )
+        )
+        self.assertTrue(
+            prediction_is_allowed(
+                'label-fast-slow-routing',
+                '模糊',
+                '{}',
+            )
+        )
+        self.assertFalse(
+            prediction_is_allowed(
+                'label-fast-slow-routing',
+                'fast',
+                '{}',
+            )
+        )
+        normalized_label = normalize_label_master_result(
+            {
+                **result,
+                'prediction': 'Agent(tag="地图导航")',
+                'function_output': '一段解释文字',
+            },
+            snapshot_metadata,
+        )
+        self.assertEqual(normalized_label['prediction'], '地图导航')
+        self.assertEqual(
+            normalized_label['function_output'],
+            'Agent(tag="地图导航")',
+        )
+        mapping = normalize_label_mapping_output(
+            json.dumps(
+                {
+                    'label_map': {
+                        '慢系统': '慢',
+                        'complex=Ture': 'complex=true',
+                    },
+                    'unmapped': [],
+                    'reason': '统一快慢和复杂度格式',
+                },
+                ensure_ascii=False,
+            ),
+            raw_labels=['慢系统', 'complex=Ture', '未知'],
+            skill_name='label-fast-slow-routing',
+        )
+        self.assertEqual(mapping['label_map']['慢系统'], '慢')
+        self.assertEqual(mapping['label_map']['complex=Ture'], 'complex=true')
+        self.assertEqual(mapping['unmapped'], ['未知'])
+        self.assertIn(
+            'HTTP 429',
+            transient_model_error_message(
+                'HTTP 429 from local model backend: Too many requests'
+            ),
+        )
+        self.assertIn(
+            '响应流意外中断',
+            transient_model_error_message('IncompleteRead(199 bytes read)'),
+        )
+        self.assertIn(
+            '响应流意外中断',
+            transient_model_error_message(
+                'Unable to reach local model backend: timed out'
+            ),
+        )
+
+    def test_transient_model_failure_retries_before_validation(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            runtime = EvaluationRuntime(
+                root=root / 'evaluations',
+                cwd_for_account=lambda _account: root,
+                model_config_for=lambda _account: ModelConfig(model='test-model'),
+                account_paths_for=lambda _account: {'python_env': root / '.venv'},
+                max_workers=1,
+            )
+            transient = AgentRunResult(
+                final_output='HTTP 429: Too many requests',
+                turns=0,
+                tool_calls=0,
+                transcript=(),
+                usage=UsageStats(),
+            )
+            success = AgentRunResult(
+                final_output='{"prediction":"快"}',
+                turns=1,
+                tool_calls=0,
+                transcript=(),
+                usage=UsageStats(),
+            )
+
+            class StubAgent:
+                def __init__(self) -> None:
+                    self.calls = 0
+
+                def run(self, *_args: object, **_kwargs: object) -> AgentRunResult:
+                    self.calls += 1
+                    return transient if self.calls == 1 else success
+
+            agent = StubAgent()
+            try:
+                with patch.object(runtime, '_extend_model_cooldown') as cooldown:
+                    result = runtime._run_agent_with_transient_backoff(
+                        agent,  # type: ignore[arg-type]
+                        'prompt',
+                        session_id='session',
+                        runtime_context='context',
+                        cancel_event=threading.Event(),
+                    )
+                self.assertIs(result, success)
+                self.assertEqual(agent.calls, 2)
+                cooldown.assert_called_once()
+            finally:
+                runtime.shutdown()
+
+    def test_fast_slow_prompt_uses_skill_contract_and_relative_paths(self) -> None:
+        skill = BundledSkill(
+            name='label-fast-slow-routing',
+            description='Test fast slow routing skill.',
+            source='directory',
+            get_prompt=lambda _agent, _args: 'Read references/policy.md.',
+        )
+        prompt, runtime_context = build_evaluation_prompt(
+            skill=skill,
+            snapshot_root=Path('/tmp/evaluation-snapshot'),
+            canonical_case={
+                'query': '打开空调',
+                'gold_label': '快',
+                'source_row': {
+                    'query': '打开空调',
+                    '人工标签': '快',
+                    'secret': 'should-not-leak',
+                },
+            },
+        )
+        self.assertIn('填写“快”“慢”或“模糊”', prompt)
+        self.assertNotIn('填写 fast 或 slow', prompt)
+        self.assertIn('`references/...`', runtime_context)
+        self.assertNotIn(
+            '不要添加 `skills/label-fast-slow-routing/` 前缀',
+            prompt,
+        )
+        self.assertNotIn('should-not-leak', prompt)
+        self.assertNotIn('"人工标签"', prompt)
+
+    def test_label_mapping_suggestion_uses_one_agent_run(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            _write_fast_slow_skill(root)
+            runtime = EvaluationRuntime(
+                root=root / 'evaluations',
+                cwd_for_account=lambda _account: root,
+                model_config_for=lambda _account: ModelConfig(model='test-model'),
+                account_paths_for=lambda _account: {'python_env': root / '.venv'},
+                max_workers=1,
+            )
+            try:
+                dataset = runtime.create_dataset(
+                    account_id='alice',
+                    name='fast-slow',
+                    filename='fast-slow.csv',
+                    rows=[
+                        {'query': '打开空调', 'label': '快系统'},
+                        {'query': '帮我规划路线', 'label': '慢系统'},
+                    ],
+                )
+                fake_result = AgentRunResult(
+                    final_output=json.dumps(
+                        {
+                            'label_map': {
+                                '快系统': '快',
+                                '慢系统': '慢',
+                            },
+                            'unmapped': [],
+                            'reason': '统一路由标签',
+                        },
+                        ensure_ascii=False,
+                    ),
+                    turns=1,
+                    tool_calls=1,
+                    transcript=(),
+                    usage=UsageStats(input_tokens=10, output_tokens=5),
+                )
+                with patch(
+                    'src.evaluation_runtime.LocalCodingAgent.run',
+                    return_value=fake_result,
+                ) as run:
+                    suggestion = runtime.suggest_label_mapping(
+                        dataset['id'],
+                        'alice',
+                        mapping=dataset['mapping'],
+                        skill_name='label-fast-slow-routing',
+                    )
+                self.assertEqual(run.call_count, 1)
+                self.assertEqual(
+                    suggestion['label_map'],
+                    {'快系统': '快', '慢系统': '慢'},
+                )
+            finally:
+                runtime.shutdown()
+
+    def test_single_analysis_uses_snapshot_without_dataset_records(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            _write_test_skill(root)
+            runtime = EvaluationRuntime(
+                root=root / 'evaluations',
+                cwd_for_account=lambda _account: root,
+                model_config_for=lambda _account: ModelConfig(model='test-model'),
+                account_paths_for=lambda _account: {'python_env': root / '.venv'},
+                max_workers=1,
+            )
+            fake_result = AgentRunResult(
+                final_output=json.dumps(
+                    {
+                        'prediction': '地图导航',
+                        'function_output': 'Agent(tag="地图导航")',
+                        'complex': False,
+                        'reason': '导航执行',
+                        'candidates': ['地图导航'],
+                        'evidence_refs': [
+                            'knowledge/索引/label_manifest.json'
+                        ],
+                        'confidence': 0.98,
+                        'uncertainty': '',
+                    },
+                    ensure_ascii=False,
+                ),
+                turns=1,
+                tool_calls=1,
+                transcript=(
+                    {
+                        'role': 'assistant',
+                        'tool_calls': [
+                            {
+                                'function': {
+                                    'name': 'read_file',
+                                    'arguments': json.dumps(
+                                        {
+                                            'path': (
+                                                'knowledge/索引/'
+                                                'label_manifest.json'
+                                            )
+                                        },
+                                        ensure_ascii=False,
+                                    ),
+                                }
+                            }
+                        ],
+                    },
+                ),
+                usage=UsageStats(input_tokens=10, output_tokens=5),
+            )
+            try:
+                with patch(
+                    'src.evaluation_runtime.LocalCodingAgent.run',
+                    return_value=fake_result,
+                ):
+                    result = runtime.analyze_case(
+                        account_id='alice',
+                        query='导航去公司',
+                    )
+                self.assertEqual(result['prediction'], '地图导航')
+                self.assertEqual(result['skill_name'], 'label-master')
+                self.assertEqual(result['model'], 'test-model')
+                self.assertEqual(runtime.list_datasets('alice'), [])
+                self.assertEqual(runtime.list_experiments('alice'), [])
+                history = runtime.list_single_analyses('alice')
+                self.assertEqual(len(history), 1)
+                self.assertEqual(history[0]['id'], result['id'])
+                self.assertEqual(history[0]['query'], '导航去公司')
+                self.assertEqual(history[0]['prediction'], '地图导航')
+                self.assertNotIn('raw_output', history[0])
+                detail = runtime.get_single_analysis(result['id'], 'alice')
+                self.assertEqual(detail['prediction'], '地图导航')
+                self.assertEqual(detail['canonical']['query'], '导航去公司')
+            finally:
+                runtime.shutdown()
+
+    def test_background_runner_completes_cases(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            _write_test_skill(root)
+            runtime = EvaluationRuntime(
+                root=root / 'evaluations',
+                cwd_for_account=lambda _account: root,
+                model_config_for=lambda _account: ModelConfig(model='test-model'),
+                account_paths_for=lambda _account: {'python_env': root / '.venv'},
+                max_workers=1,
+            )
+            try:
+                dataset = runtime.create_dataset(
+                    account_id='alice',
+                    name='test',
+                    filename='test.csv',
+                    rows=[
+                        {'query': '导航去公司', 'label': '地图导航'},
+                        {'query': '导航回家', 'label': '地图导航'},
+                    ],
+                )
+                experiment = runtime.create_experiment(
+                    account_id='alice',
+                    dataset_id=dataset['id'],
+                    name='runner',
+                    concurrency=1,
+                )
+                fake_result = AgentRunResult(
+                    final_output=json.dumps(
+                        {
+                            'prediction': '地图导航',
+                            'function_output': 'Agent(tag="地图导航")',
+                            'complex': False,
+                            'reason': '导航执行',
+                            'candidates': ['地图导航'],
+                            'evidence_refs': ['knowledge/标签/地图导航.md'],
+                            'confidence': 0.98,
+                            'uncertainty': '',
+                        },
+                        ensure_ascii=False,
+                    ),
+                    turns=1,
+                    tool_calls=1,
+                    transcript=(
+                        {
+                            'role': 'assistant',
+                            'tool_calls': [
+                                {
+                                    'function': {
+                                        'name': 'read_file',
+                                        'arguments': json.dumps(
+                                            {
+                                                'path': (
+                                                    'knowledge/索引/'
+                                                    'label_manifest.json'
+                                                )
+                                            },
+                                            ensure_ascii=False,
+                                        ),
+                                    }
+                                }
+                            ],
+                        },
+                    ),
+                    usage=UsageStats(input_tokens=10, output_tokens=5),
+                )
+                with patch(
+                    'src.evaluation_runtime.LocalCodingAgent.run',
+                    return_value=fake_result,
+                ):
+                    runtime.start(experiment['id'], 'alice')
+                    deadline = time.time() + 3
+                    current = runtime.get_experiment(experiment['id'], 'alice')
+                    while (
+                        current['status'] not in {'completed', 'completed_with_errors'}
+                        and time.time() < deadline
+                    ):
+                        time.sleep(0.02)
+                        current = runtime.get_experiment(
+                            experiment['id'],
+                            'alice',
+                        )
+                self.assertEqual(current['status'], 'completed')
+                self.assertEqual(current['completed_cases'], 2)
+                self.assertEqual(current['metrics']['accuracy'], 1.0)
+            finally:
+                runtime.shutdown()
+
+    def test_accessed_refs_only_include_snapshot_paths(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            snapshot_root = Path(directory)
+            evidence = snapshot_root / 'skills' / 'label-master' / 'knowledge.md'
+            evidence.parent.mkdir(parents=True)
+            evidence.write_text('evidence', encoding='utf-8')
+            transcript = (
+                {
+                    'role': 'assistant',
+                    'tool_calls': [
+                        {
+                            'function': {
+                                'name': 'read_file',
+                                'arguments': json.dumps(
+                                    {
+                                        'path': (
+                                            'skills/label-master/knowledge.md'
+                                        )
+                                    }
+                                ),
+                            }
+                        },
+                        {
+                            'function': {
+                                'name': 'read_file',
+                                'arguments': json.dumps({'path': '/etc/hosts'}),
+                            }
+                        },
+                    ],
+                },
+            )
+            self.assertEqual(
+                extract_accessed_refs(
+                    transcript,
+                    snapshot_root=snapshot_root,
+                ),
+                [
+                    {
+                        'tool': 'read_file',
+                        'path': 'skills/label-master/knowledge.md',
+                    }
+                ],
+            )
+
+    def test_metrics_count_disagreements(self) -> None:
+        metrics = calculate_metrics(
+            [
+                {
+                    'status': 'completed',
+                    'gold_label': 'A',
+                    'prediction': 'A',
+                    'correct': True,
+                    'confidence': 0.9,
+                    'evidence_refs': ['a.md'],
+                    'domain': 'one',
+                },
+                {
+                    'status': 'completed',
+                    'gold_label': 'A',
+                    'prediction': 'B',
+                    'correct': False,
+                    'confidence': 0.5,
+                    'evidence_refs': [],
+                    'domain': 'one',
+                },
+            ]
+        )
+        self.assertEqual(metrics['labeled'], 2)
+        self.assertEqual(metrics['disagreements'], 1)
+        self.assertEqual(metrics['accuracy'], 0.5)
+        self.assertEqual(metrics['low_confidence'], 1)
+        self.assertEqual(metrics['no_evidence'], 1)
+
+
+if __name__ == '__main__':
+    unittest.main()

tests/test_extended_tools.py

@@ -99,6 +99,45 @@ class ExtendedToolTests(unittest.TestCase):
         self.assertIn('[line truncated,', result.content)
         self.assertLess(len(result.content), 1200)
 
+    def test_glob_search_accepts_absolute_path_inside_workspace(self) -> None:
+        registry = default_tool_registry()
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            workspace = Path(tmp_dir)
+            target_dir = workspace / 'output'
+            target_dir.mkdir()
+            (target_dir / 'records.jsonl').write_text('{}\n', encoding='utf-8')
+            context = build_tool_context(
+                AgentRuntimeConfig(cwd=workspace),
+                tool_registry=registry,
+            )
+            result = execute_tool(
+                registry,
+                'glob_search',
+                {'pattern': str(target_dir / '*.jsonl')},
+                context,
+            )
+
+        self.assertTrue(result.ok)
+        self.assertIn('output/records.jsonl', result.content)
+
+    def test_glob_search_external_absolute_path_returns_no_matches(self) -> None:
+        registry = default_tool_registry()
+        with tempfile.TemporaryDirectory() as workspace_dir, tempfile.TemporaryDirectory() as external_dir:
+            external = Path(external_dir) / '*.jsonl'
+            context = build_tool_context(
+                AgentRuntimeConfig(cwd=Path(workspace_dir)),
+                tool_registry=registry,
+            )
+            result = execute_tool(
+                registry,
+                'glob_search',
+                {'pattern': str(external)},
+                context,
+            )
+
+        self.assertTrue(result.ok)
+        self.assertEqual(result.content, '(no matches)')
+
     def test_read_file_can_read_explicit_external_path(self) -> None:
         registry = default_tool_registry()
         with tempfile.TemporaryDirectory() as workspace_dir, tempfile.TemporaryDirectory() as external_dir:

tests/test_feishu_integration.py

@@ -109,6 +109,34 @@ class FeishuIntegrationTests(unittest.TestCase):
             )
             self.assertEqual(online_docs['files'][str(file_path)]['kind'], 'doc')
 
+    def test_create_online_doc_returns_login_required_when_mcp_token_expired(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            client, state = _build_client(Path(tmp_dir))
+            file_path = _write_account_file(state, 'alice', 'draft.md', '# hello')
+
+            with patch.object(
+                gui_server,
+                '_feishu_status_payload',
+                return_value={'logged_in': True, 'status': 'logged_in'},
+            ), patch.object(
+                gui_server.MCPRuntime,
+                'call_tool',
+                side_effect=RuntimeError('登录已过期或未登录，请重新授权'),
+            ), patch.object(
+                gui_server,
+                '_run_feishu_cli',
+                return_value={'returncode': 0, 'output': 'logged out'},
+            ):
+                response = client.post(
+                    '/api/files/online-doc',
+                    json={'account_id': 'alice', 'path': str(file_path)},
+                )
+
+        self.assertEqual(response.status_code, 409)
+        detail = response.json()['detail']
+        self.assertEqual(detail['code'], 'feishu_auth_expired')
+        self.assertTrue(detail['force_login'])
+
     def test_create_online_doc_converts_csv_to_feishu_sheet(self) -> None:
         with tempfile.TemporaryDirectory() as tmp_dir:
             client, state = _build_client(Path(tmp_dir))