backend/frontend/runtime: 训练面板 / session 隔离 / agent runtime 调整

- backend/api/server.py：训练 step-detail / pipeline 卡片排序逻辑
- frontend：thread-list / thread / training-pipeline-panel / step-detail-sheet 配套调整，新增 metrics-chart-panel
- src/agent_*：tool spec / runtime / prompting 配套
- .gitignore 加 .logs/

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

.gitignore

@@ -21,6 +21,7 @@ archive/
 scratchpad/
 tasks/
 router_session_parquet/
+.logs/
 
 # Environment files
 .env

backend/api/server.py

@@ -1696,12 +1696,31 @@ def create_app(state: AgentState) -> FastAPI:
     async def get_training_step_detail(
         session_id: str,
         step: str,
+        run_id: str | None = None,
         account_id: str | None = None,
     ) -> Response:
         sessions_dir = state.account_paths(account_id)['sessions']
         state_path = _session_state_path(sessions_dir, session_id)
         state_entries, _ = _read_program_state(state_path)
-        run_dic = _resolve_run_dic_from_state(state_entries)
+        # Per-round: if frontend passed run_id, use that round's actual runDic
+        # so historical step views (e.g., R1's augment) read the right
+        # workflow<N> directory. Fall back to "max in state" only when run_id
+        # is absent or the round has no resolvable runDic — preserves prior
+        # behaviour for unbundled callers.
+        run_dic: int | None = None
+        if isinstance(run_id, str) and run_id.strip():
+            rid = run_id.strip()
+            per_round = _resolve_run_dic_per_round(state_entries)
+            # augment 处理的是「上一轮评测的错例」，产物按 R{n-1} 的 runDic 命名
+            # （data_clean_<R{n-1}>/、augment_<R{n-1}>.jsonl ...），不是本轮 eval workflow
+            if step == 'augment':
+                m = re.match(r'^[Rr](\d+)$', rid)
+                if m and int(m.group(1)) >= 1:
+                    run_dic = per_round.get(f'R{int(m.group(1)) - 1}')
+            if run_dic is None:
+                run_dic = per_round.get(rid)
+        if run_dic is None:
+            run_dic = _resolve_run_dic_from_state(state_entries)
         runtime = _jupyter_runtime_for_session(state, account_id, session_id)
         artifact_paths = _step_artifact_paths(step, run_dic, runtime)
 
@@ -1850,6 +1869,33 @@ def create_app(state: AgentState) -> FastAPI:
                         'shown_rows': len(table_rows),
                     }
                 )
+            if step == 'dist-analysis':
+                # 软警告：dist-analysis 标 complete 时若 H1 候选 CSV 缺失，把 section
+                # 的兜底 "产物文件不存在" 换成更明确的提示。某些轮次（hparam 类假设、
+                # 无标签问题）确实可能没有候选，所以不 fail 这一步、不阻塞 human-check。
+                target_run = (run_id or '').strip()
+                last_status: str | None = None
+                for entry in state_entries:
+                    if not isinstance(entry, dict) or entry.get('step') != 'dist-analysis':
+                        continue
+                    eid = (entry.get('run_id') or '').strip()
+                    if target_run and eid and eid != target_run:
+                        continue
+                    if 'status' in entry:
+                        last_status = entry.get('status')
+                if last_status == 'complete':
+                    rd_label = str(run_dic) if run_dic is not None else '<runDic>'
+                    for sec in table_sections:
+                        if (
+                            str(sec.get('title', '')).startswith('H1 预计修改候选')
+                            and sec.get('kind') == 'missing'
+                        ):
+                            sec['error'] = (
+                                f'未检测到 output/relabel_candidates_{rd_label}.csv —— '
+                                f'若本轮无 H1 候选可忽略；如有候选请按 SKILL.md 「阶段一」'
+                                f'落盘 file,line,query,old_label,是否改(1/0),建议新label。'
+                            )
+                            break
             payload = {
                 'step': step,
                 'run_dic': run_dic,
@@ -1923,6 +1969,12 @@ def create_app(state: AgentState) -> FastAPI:
             agent_state=state,
             account_id=account_id,
         )
+        metrics_history = await asyncio.to_thread(
+            _read_iteration_log_metrics_history,
+            session_id,
+            agent_state=state,
+            account_id=account_id,
+        )
         payload = await asyncio.to_thread(
             _hardcoded_autoresearch_pipeline,
             session_id,
@@ -1931,6 +1983,8 @@ def create_app(state: AgentState) -> FastAPI:
             model_config=state.model_config_for(account_id),
             agent_state=state,
             account_id=account_id,
+            iteration_log_count=iter_count,
+            metrics_history=metrics_history,
         )
         failure_status = await asyncio.to_thread(
             _read_session_failure_status,
@@ -1964,48 +2018,82 @@ def create_app(state: AgentState) -> FastAPI:
         async def event_gen():
             last_signature: str | None = None
             last_heartbeat = time.monotonic()
-            while True:
+            refresher = asyncio.create_task(
-                if await request.is_disconnected():
+                _iter_count_refresh_loop(
-                    break
-                state_entries, state_mtime = _read_program_state(state_path)
-                iter_count = await asyncio.to_thread(
-                    _read_iteration_log_count,
                     session_id,
                     agent_state=state,
                     account_id=account_id,
                 )
-                payload = await asyncio.to_thread(
+            )
-                    _hardcoded_autoresearch_pipeline,
+            metrics_refresher = asyncio.create_task(
+                _metrics_history_refresh_loop(
                     session_id,
-                    sessions_dir=sessions_dir,
-                    state_entries=state_entries,
                     agent_state=state,
                     account_id=account_id,
                 )
-                failure_status = await asyncio.to_thread(
+            )
-                    _read_session_failure_status,
+            try:
-                    sessions_dir,
+                while True:
-                    session_id,
+                    if await request.is_disconnected():
-                )
+                        break
-                run_active = _is_run_active(state, account_id, session_id)
+                    state_entries, state_mtime = _read_program_state(state_path)
-                _apply_program_state(
+                    iter_count = await asyncio.to_thread(
-                    payload,
+                        _read_iteration_log_count,
-                    state_entries,
+                        session_id,
-                    iteration_log_count=iter_count or 0,
+                        agent_state=state,
-                    session_failure_status=failure_status,
+                        account_id=account_id,
-                    run_active=run_active,
+                    )
-                    session_id=session_id,
+                    metrics_history = await asyncio.to_thread(
-                )
+                        _read_iteration_log_metrics_history,
-                payload_str = json.dumps(payload, ensure_ascii=False)
+                        session_id,
-                signature = f'{state_mtime}|{hash(payload_str)}'
+                        agent_state=state,
-                if signature != last_signature:
+                        account_id=account_id,
-                    last_signature = signature
+                    )
-                    yield f'data: {payload_str}\n\n'
+                    payload = await asyncio.to_thread(
-                now = time.monotonic()
+                        _hardcoded_autoresearch_pipeline,
-                if now - last_heartbeat > 15:
+                        session_id,
-                    yield ': heartbeat\n\n'
+                        sessions_dir=sessions_dir,
-                    last_heartbeat = now
+                        state_entries=state_entries,
-                await asyncio.sleep(0.5)
+                        agent_state=state,
+                        account_id=account_id,
+                        iteration_log_count=iter_count,
+                        metrics_history=metrics_history,
+                    )
+                    failure_status = await asyncio.to_thread(
+                        _read_session_failure_status,
+                        sessions_dir,
+                        session_id,
+                    )
+                    run_active = _is_run_active(state, account_id, session_id)
+                    _apply_program_state(
+                        payload,
+                        state_entries,
+                        iteration_log_count=iter_count or 0,
+                        session_failure_status=failure_status,
+                        run_active=run_active,
+                        session_id=session_id,
+                    )
+                    payload_str = json.dumps(payload, ensure_ascii=False)
+                    signature = f'{state_mtime}|{hash(payload_str)}'
+                    if signature != last_signature:
+                        last_signature = signature
+                        yield f'data: {payload_str}\n\n'
+                    now = time.monotonic()
+                    if now - last_heartbeat > 15:
+                        yield ': heartbeat\n\n'
+                        last_heartbeat = now
+                    await asyncio.sleep(0.5)
+            finally:
+                refresher.cancel()
+                try:
+                    await refresher
+                except (asyncio.CancelledError, Exception):  # noqa: BLE001
+                    pass
+                metrics_refresher.cancel()
+                try:
+                    await metrics_refresher
+                except (asyncio.CancelledError, Exception):  # noqa: BLE001
+                    pass
 
         return StreamingResponse(
             event_gen(),
@@ -2886,7 +2974,7 @@ def _normalize_run_event(event: dict[str, object]) -> dict[str, Any] | None:
         'exceeds_soft_limit',
     }
     normalized = {
-        key: _json_safe_limited(value)
+        key: _json_safe_limited(value, parent_key=key)
         for key, value in event.items()
         if key in allowed and value is not None
     }
@@ -2894,20 +2982,25 @@ def _normalize_run_event(event: dict[str, object]) -> dict[str, Any] | None:
     return normalized
 
 
-def _json_safe_limited(value: Any, *, depth: int = 0) -> Any:
+_VERBOSE_STRING_KEYS = {'code', 'command', 'content', 'delta', 'script', 'stdout', 'output'}
+_VERBOSE_STRING_LIMIT = 20000
+
+
+def _json_safe_limited(value: Any, *, depth: int = 0, parent_key: str | None = None) -> Any:
     if depth > 4:
         return str(value)[:500]
     if isinstance(value, str):
-        return value[:2000]
+        limit = _VERBOSE_STRING_LIMIT if parent_key in _VERBOSE_STRING_KEYS else 2000
+        return value[:limit]
     if isinstance(value, (int, float, bool)) or value is None:
         return value
     if isinstance(value, dict):
         return {
-            str(key)[:200]: _json_safe_limited(item, depth=depth + 1)
+            str(key)[:200]: _json_safe_limited(item, depth=depth + 1, parent_key=str(key))
             for key, item in list(value.items())[:80]
         }
     if isinstance(value, (list, tuple)):
-        return [_json_safe_limited(item, depth=depth + 1) for item in value[:80]]
+        return [_json_safe_limited(item, depth=depth + 1, parent_key=parent_key) for item in value[:80]]
     try:
         json.dumps(value)
         return value
@@ -4885,6 +4978,35 @@ class _WatcherManager:
                         account_id=account_id,
                         session_id=session_id,
                     )
+                    # R{n>=1} 链路：SFT 训练完后 submit_sft_via_cml.sh 内置 nohup
+                    # watcher 自动起 R{n} 评测，agent 不在回路中，没人写
+                    # cml=running，UI 会从 sft=complete 直接跳到 cml=complete，
+                    # 中间 ~20min 评测期看起来像"流程卡死"。检测到同 session
+                    # 已注册 cml watch（说明 agent 提交 SFT 时预注册了下游评测
+                    # 的 watch）就顺手补写 cml=running 同 run_id。
+                    if (
+                        step == 'sft'
+                        and session_id is not None
+                        and self.is_watching(session_id, 'cml')
+                    ):
+                        await self._write_step(
+                            state_path,
+                            step='cml',
+                            status='running',
+                            run_id=run_id,
+                            agent_state=agent_state,
+                            account_id=account_id,
+                            session_id=session_id,
+                        )
+                        await self._write_log(
+                            state_path,
+                            run_id,
+                            '↪ sft 完成自动衔接 → step cml running '
+                            '(submit_sft_via_cml.sh 内置 watcher 已触发 R 评测)',
+                            agent_state=agent_state,
+                            account_id=account_id,
+                            session_id=session_id,
+                        )
                     return
                 if time.monotonic() - start > timeout:
                     await self._write_step(
@@ -5507,13 +5629,27 @@ def _scan_for_watchers(state: AgentState) -> None:
                 status = entry.get('status')
                 if isinstance(step, str) and step and isinstance(status, str):
                     last_status_per_step[step] = status
+            # See note in _scan_for_watchers_async: only the latest watch
+            # entry per step is current.
+            latest_watch_per_step: dict[str, dict[str, Any]] = {}
+            latest_running_run_id: dict[str, str] = {}
             for entry in entries:
                 watch = entry.get('watch')
-                if not isinstance(watch, dict):
+                if isinstance(watch, dict):
-                    continue
+                    w_step = watch.get('step')
-                step = watch.get('step')
+                    if isinstance(w_step, str) and w_step:
-                if not isinstance(step, str) or not step:
+                        latest_watch_per_step[w_step] = watch
                     continue
+                e_step = entry.get('step')
+                e_status = entry.get('status')
+                e_run = entry.get('run_id')
+                if (
+                    isinstance(e_step, str) and e_step
+                    and e_status == 'running'
+                    and isinstance(e_run, str) and e_run
+                ):
+                    latest_running_run_id[e_step] = e_run
+            for step, watch in latest_watch_per_step.items():
                 if last_status_per_step.get(step) in {
                     'complete',
                     'failed',
@@ -5535,6 +5671,9 @@ def _scan_for_watchers(state: AgentState) -> None:
                     interval = 30.0
                     timeout = 3600.0
                 run_id = str(watch.get('run_id') or '?')
+                running_run_id = latest_running_run_id.get(step)
+                if running_run_id and run_id != running_run_id:
+                    continue
                 _watcher_manager.register_file_exists(
                     session_id=session_id,
                     state_path=state_path,
@@ -5617,13 +5756,30 @@ async def _scan_for_watchers_async(state: AgentState) -> None:
                 status = entry.get('status')
                 if isinstance(step, str) and step and isinstance(status, str):
                     last_status_per_step[step] = status
+            # Only the LAST watch entry per step represents the current round's
+            # intent. Earlier watch entries from previous rounds must NOT be
+            # re-registered when a new round resets the step to running, or
+            # the resulting watcher will fire with a stale run_id and the UI
+            # will never see a complete tagged with the current round.
+            latest_watch_per_step: dict[str, dict[str, Any]] = {}
+            latest_running_run_id: dict[str, str] = {}
             for entry in entries:
                 watch = entry.get('watch')
-                if not isinstance(watch, dict):
+                if isinstance(watch, dict):
-                    continue
+                    w_step = watch.get('step')
-                step = watch.get('step')
+                    if isinstance(w_step, str) and w_step:
-                if not isinstance(step, str) or not step:
+                        latest_watch_per_step[w_step] = watch
                     continue
+                e_step = entry.get('step')
+                e_status = entry.get('status')
+                e_run = entry.get('run_id')
+                if (
+                    isinstance(e_step, str) and e_step
+                    and e_status == 'running'
+                    and isinstance(e_run, str) and e_run
+                ):
+                    latest_running_run_id[e_step] = e_run
+            for step, watch in latest_watch_per_step.items():
                 if last_status_per_step.get(step) in {
                     'complete',
                     'failed',
@@ -5645,6 +5801,13 @@ async def _scan_for_watchers_async(state: AgentState) -> None:
                     interval_s = 30.0
                     timeout_s = 3600.0
                 run_id = str(watch.get('run_id') or '?')
+                # If a newer round has written `step=<step>, status=running`
+                # but the matching watch entry hasn't landed yet, skip and
+                # wait for the next scan rather than spawning a watcher tied
+                # to the previous round's run_id.
+                running_run_id = latest_running_run_id.get(step)
+                if running_run_id and run_id != running_run_id:
+                    continue
                 _watcher_manager.register_file_exists(
                     session_id=session_id,
                     state_path=state_path,
@@ -5752,7 +5915,6 @@ _STEP_KIND: dict[str, str] = {
     'cml':           'analysis',
     'gold-drift':    'analysis',
     'dist-analysis': 'analysis',
-    'report':        'analysis',
     'hypothesis':    'analysis',
     'log':           'analysis',
     'augment':       'train',
@@ -5780,8 +5942,7 @@ _CARD_META: dict[tuple[str, str], dict[str, Any]] = {
     # R0·Baseline
     ('cml', 'baseline'):           {'icon': 'bar-chart',      'title': 'Baseline 评测',    'subtitle': 'CML workflow metric_diff'},
     ('gold-drift', 'baseline'):    {'icon': 'alert-triangle', 'title': 'Gold Drift 检查', 'subtitle': 'drift_<runDic>.json'},
-    ('dist-analysis', 'baseline'): {'icon': 'trending-up',    'title': '分层结果分析',     'subtitle': '需求集合 / 车载 / specific'},
+    ('dist-analysis', 'baseline'): {'icon': 'trending-up',    'title': '分层结果分析 & 报告', 'subtitle': '根因归类 + workflow<runDic>.md'},
-    ('report', 'baseline'):        {'icon': 'file-text',      'title': '问题分析 & 报告',  'subtitle': 'workflow<runDic>.md'},
     ('hypothesis', 'baseline'):    {'icon': 'dna',            'title': '形成假设',         'subtitle': 'iteration_log hypothesis 字段'},
     ('log', 'baseline'):           {'icon': 'clipboard-list', 'title': '记录迭代日志',     'subtitle': 'iteration_log.jsonl'},
     # R{n}·Train
@@ -5791,8 +5952,7 @@ _CARD_META: dict[tuple[str, str], dict[str, Any]] = {
     # R{n}·Analysis
     ('cml', 'analysis'):           {'icon': 'bar-chart',      'title': 'CML 评测',         'subtitle': 'workflow metric_diff'},
     ('gold-drift', 'analysis'):    {'icon': 'alert-triangle', 'title': 'Gold Drift 检查', 'subtitle': 'drift_<runDic>.json'},
-    ('dist-analysis', 'analysis'): {'icon': 'trending-up',    'title': '分层结果分析',     'subtitle': '需求集合 / 车载 / specific'},
+    ('dist-analysis', 'analysis'): {'icon': 'trending-up',    'title': '分层结果分析 & 报告', 'subtitle': '根因归类 + workflow<runDic>.md'},
-    ('report', 'analysis'):        {'icon': 'file-text',      'title': '问题分析 & 报告',  'subtitle': 'workflow<runDic>.md'},
     ('hypothesis', 'analysis'):    {'icon': 'dna',            'title': '形成假设',         'subtitle': 'iteration_log hypothesis 字段'},
     ('log', 'analysis'):           {'icon': 'clipboard-list', 'title': '记录迭代日志',     'subtitle': 'iteration_log.jsonl'},
 }
@@ -5803,9 +5963,8 @@ _CARD_ORDER: dict[tuple[str, str], int] = {
     ('cml', 'baseline'): 0,
     ('gold-drift', 'baseline'): 1,
     ('dist-analysis', 'baseline'): 2,
-    ('report', 'baseline'): 3,
+    ('hypothesis', 'baseline'): 3,
-    ('hypothesis', 'baseline'): 4,
+    ('log', 'baseline'): 4,
-    ('log', 'baseline'): 5,
     # train
     ('augment', 'train'): 0,
     ('verify', 'train'): 1,
@@ -5814,9 +5973,8 @@ _CARD_ORDER: dict[tuple[str, str], int] = {
     ('cml', 'analysis'): 0,
     ('gold-drift', 'analysis'): 1,
     ('dist-analysis', 'analysis'): 2,
-    ('report', 'analysis'): 3,
+    ('hypothesis', 'analysis'): 3,
-    ('hypothesis', 'analysis'): 4,
+    ('log', 'analysis'): 4,
-    ('log', 'analysis'): 5,
 }
 
 _SECTION_LABELS: dict[str, str] = {
@@ -6514,6 +6672,13 @@ def _scan_run_history_max() -> int | None:
     return best if best >= 0 else None
 
 
+_ITER_COUNT_CACHE: dict[str, int | None] = {}
+
+
+def _iter_count_cache_key(session_id: str | None, account_id: str | None) -> str:
+    return f'{account_id or "_"}|{session_id or "_"}'
+
+
 def _read_iteration_log_count(
     session_id: str | None = None,
     *,
@@ -6523,10 +6688,53 @@ def _read_iteration_log_count(
     """Count non-empty lines in iteration_log.jsonl.
 
     Chat-root lives under the remote workspace_cwd, so reads MUST go through
-    the bound jupyter runtime — backend host typically can't read this path
+    the bound jupyter runtime — the round-trip is multi-second. To keep the
-    locally. Falls back to the legacy global path on the local fs only when
+    SSE pipeline loop responsive, this function returns the last cached value
-    no runtime is bound (session never connected, or transient lookup miss).
+    when one exists; only a cold cache pays the synchronous remote cost. The
+    cache is refreshed in the background by `_iter_count_refresh_loop`, which
+    the SSE handler spawns alongside its event loop.
     """
+    cache_key = _iter_count_cache_key(session_id, account_id)
+    if cache_key in _ITER_COUNT_CACHE:
+        return _ITER_COUNT_CACHE[cache_key]
+    value = _read_iteration_log_count_uncached(
+        session_id, agent_state=agent_state, account_id=account_id,
+    )
+    _ITER_COUNT_CACHE[cache_key] = value
+    return value
+
+
+async def _iter_count_refresh_loop(
+    session_id: str | None,
+    *,
+    agent_state: 'AgentState | None',
+    account_id: str | None,
+    interval_seconds: float = 3.0,
+) -> None:
+    """Background task that re-reads iteration_log_count every `interval_seconds`
+    and updates the in-memory cache. Cancelled when the caller (SSE handler)
+    exits."""
+    cache_key = _iter_count_cache_key(session_id, account_id)
+    while True:
+        try:
+            value = await asyncio.to_thread(
+                _read_iteration_log_count_uncached,
+                session_id,
+                agent_state=agent_state,
+                account_id=account_id,
+            )
+            _ITER_COUNT_CACHE[cache_key] = value
+        except Exception:  # noqa: BLE001
+            pass
+        await asyncio.sleep(interval_seconds)
+
+
+def _read_iteration_log_count_uncached(
+    session_id: str | None = None,
+    *,
+    agent_state: 'AgentState | None' = None,
+    account_id: str | None = None,
+) -> int | None:
     runtime = (
         _jupyter_runtime_for_session(agent_state, account_id, session_id)
         if session_id and agent_state is not None
@@ -6565,6 +6773,150 @@ def _read_iteration_log_count(
     return None
 
 
+_METRICS_HISTORY_CACHE: dict[str, dict[str, Any] | None] = {}
+
+
+def _read_iteration_log_metrics_history(
+    session_id: str | None = None,
+    *,
+    agent_state: 'AgentState | None' = None,
+    account_id: str | None = None,
+) -> dict[str, Any] | None:
+    """Cached read of full iteration_log.jsonl as metrics time-series."""
+    cache_key = _iter_count_cache_key(session_id, account_id)
+    if cache_key in _METRICS_HISTORY_CACHE:
+        return _METRICS_HISTORY_CACHE[cache_key]
+    value = _read_iteration_log_metrics_history_uncached(
+        session_id, agent_state=agent_state, account_id=account_id,
+    )
+    _METRICS_HISTORY_CACHE[cache_key] = value
+    return value
+
+
+async def _metrics_history_refresh_loop(
+    session_id: str | None,
+    *,
+    agent_state: 'AgentState | None',
+    account_id: str | None,
+    interval_seconds: float = 3.0,
+) -> None:
+    cache_key = _iter_count_cache_key(session_id, account_id)
+    while True:
+        try:
+            value = await asyncio.to_thread(
+                _read_iteration_log_metrics_history_uncached,
+                session_id,
+                agent_state=agent_state,
+                account_id=account_id,
+            )
+            _METRICS_HISTORY_CACHE[cache_key] = value
+        except Exception:  # noqa: BLE001
+            pass
+        await asyncio.sleep(interval_seconds)
+
+
+def _read_iteration_log_metrics_history_uncached(
+    session_id: str | None = None,
+    *,
+    agent_state: 'AgentState | None' = None,
+    account_id: str | None = None,
+) -> dict[str, Any] | None:
+    """Parse iteration_log.jsonl into {'metrics': [keys], 'rounds': [...]}.
+
+    Each entry's `results` dict contributes one round point; metric key union
+    is taken across all rounds (sorted by first-appearance for stable column
+    order). Missing values produce nulls so frontend line charts break the
+    series naturally."""
+    runtime = (
+        _jupyter_runtime_for_session(agent_state, account_id, session_id)
+        if session_id and agent_state is not None
+        else None
+    )
+    text: str | None = None
+    if runtime is not None:
+        remote_root = _chat_root_for_runtime(runtime)
+        if remote_root is not None:
+            remote_path = f'{remote_root}/results/iteration_log.jsonl'
+            cmd = (
+                f'test -f {shlex.quote(remote_path)} '
+                f'&& cat {shlex.quote(remote_path)} '
+                f'|| echo __MISSING__'
+            )
+            try:
+                result = runtime.run_command(
+                    cmd, timeout_seconds=10.0, max_output_chars=400_000,
+                )
+            except Exception:  # noqa: BLE001
+                return None
+            stdout = (result.stdout or '').strip() if result is not None else ''
+            if stdout and stdout != '__MISSING__':
+                text = stdout
+    if text is None:
+        legacy = _autoresearch_root() / 'results' / 'iteration_log.jsonl'
+        try:
+            if legacy.is_file():
+                text = legacy.read_text(encoding='utf-8', errors='replace')
+        except OSError:
+            return None
+    if not text:
+        return None
+
+    metrics_order: list[str] = []
+    metrics_seen: set[str] = set()
+    rounds: list[dict[str, Any]] = []
+    for line in text.splitlines():
+        line = line.strip()
+        if not line:
+            continue
+        try:
+            obj = json.loads(line)
+        except (ValueError, TypeError):
+            continue
+        if not isinstance(obj, dict):
+            continue
+        results = obj.get('results')
+        if not isinstance(results, dict):
+            continue
+        values: dict[str, float] = {}
+        for k, v in results.items():
+            if not isinstance(k, str):
+                continue
+            try:
+                fv = float(v)
+            except (TypeError, ValueError):
+                continue
+            values[k] = fv
+            if k not in metrics_seen:
+                metrics_seen.add(k)
+                metrics_order.append(k)
+        if not values:
+            continue
+        iteration = obj.get('iteration')
+        if isinstance(iteration, str) and iteration.lstrip('-').isdigit():
+            iteration = int(iteration)
+        if not isinstance(iteration, int):
+            iteration = len(rounds)
+        run_dic = obj.get('runDic') or obj.get('run_dic')
+        if isinstance(run_dic, str) and run_dic.isdigit():
+            run_dic = int(run_dic)
+        if not isinstance(run_dic, int):
+            run_dic = 0
+        timestamp = obj.get('timestamp')
+        if not isinstance(timestamp, str):
+            timestamp = ''
+        rounds.append({
+            'iteration': iteration,
+            'runDic': run_dic,
+            'label': f'R{iteration}',
+            'timestamp': timestamp,
+            'values': values,
+        })
+
+    if not rounds:
+        return None
+    return {'metrics': metrics_order, 'rounds': rounds}
+
+
 def _read_iteration_log_h1_label(
     runtime: 'JupyterRuntimeSession | None',
     run_dic: int | None,
@@ -7019,6 +7371,8 @@ def _compute_kpis(
     *,
     agent_state: 'AgentState | None' = None,
     account_id: str | None = None,
+    iteration_log_count: int | None = None,
+    metrics_history: dict[str, Any] | None = None,
 ) -> list[dict[str, Any]]:
     """Compute the 7 top KPIs from authoritative sources. Falls back to '—'.
 
@@ -7054,17 +7408,33 @@ def _compute_kpis(
     target_metric = _format_metric_value(metrics, 'target_set_pass_rate')
     overall_metric = _format_metric_value(metrics, 'overall_car_pass_rate')
     specific_metric = _format_metric_value(metrics, 'specific_test_pass_rate')
-    iter_count = _read_iteration_log_count(
+    # Round number = MAX numeric `iteration` field across iteration_log.jsonl
-        session_id,
+    # entries, NOT the line count. iteration_log.jsonl carries multiple rows
-        agent_state=agent_state,
+    # per round (separate hypothesis-only / results / final-summary entries),
-        account_id=account_id,
+    # so line count overcounts. Falls back to line-count for backwards
-    )
+    # compatibility when metrics_history isn't available.
-    if iter_count is None:
+    max_iteration: int | None = None
-        iteration = 'R0-baseline'
+    if metrics_history and isinstance(metrics_history.get('rounds'), list):
-    elif iter_count == 0:
+        for r in metrics_history['rounds']:
-        iteration = 'R0-baseline'
+            it = r.get('iteration') if isinstance(r, dict) else None
+            if isinstance(it, int):
+                if max_iteration is None or it > max_iteration:
+                    max_iteration = it
+    if max_iteration is not None:
+        iteration = 'R0-baseline' if max_iteration == 0 else f'R{max_iteration}'
     else:
-        iteration = f'R{iter_count}'
+        if iteration_log_count is not None:
+            iter_count = iteration_log_count
+        else:
+            iter_count = _read_iteration_log_count(
+                session_id,
+                agent_state=agent_state,
+                account_id=account_id,
+            )
+        if iter_count is None or iter_count == 0:
+            iteration = 'R0-baseline'
+        else:
+            iteration = f'R{iter_count}'
     elapsed = _format_elapsed(_compute_step0_start_seconds(state_entries))
     target_value = (
         f'{target_set} · {target_metric}' if target_metric != '—' else target_set
@@ -7134,6 +7504,79 @@ def _resolve_run_dic_from_state(
     return best
 
 
+def _resolve_run_dic_per_round(
+    state_entries: list[dict[str, Any]],
+) -> dict[str, int]:
+    """Map each run_id (e.g., 'R2') to the runDic that round actually used.
+
+    Same priority as ``_resolve_run_dic_from_state`` but bucketed per run_id
+    so per-round step-detail views read the right artifact directory instead
+    of always using the latest workflow id seen anywhere in state."""
+    workflow_re = re.compile(r'workflow(\d+)')
+    rundic_re = re.compile(r'runDic\s*[=:取]?\s*(\d+)')
+
+    def _scan(obj: Any) -> int | None:
+        if isinstance(obj, str):
+            cands = [int(x) for x in workflow_re.findall(obj)]
+            cands += [int(x) for x in rundic_re.findall(obj)]
+            return max(cands) if cands else None
+        if isinstance(obj, dict):
+            best: int | None = None
+            for v in obj.values():
+                hit = _scan(v)
+                if hit is not None and (best is None or hit > best):
+                    best = hit
+            return best
+        if isinstance(obj, list):
+            best = None
+            for v in obj:
+                hit = _scan(v)
+                if hit is not None and (best is None or hit > best):
+                    best = hit
+            return best
+        return None
+
+    def _entry_run_id(entry: dict[str, Any]) -> str | None:
+        # 顶层 run_id 优先；否则从 watch.run_id / log.iter 兜底，让 watcher/log
+        # 这种 agent 不写顶层 run_id 但嵌套里带轮次信号的条目也能贡献 runDic。
+        rid = entry.get('run_id')
+        if isinstance(rid, str) and rid.strip():
+            return rid.strip()
+        watch = entry.get('watch')
+        if isinstance(watch, dict):
+            wrid = watch.get('run_id')
+            if isinstance(wrid, str) and wrid.strip():
+                return wrid.strip()
+        log = entry.get('log')
+        if isinstance(log, dict):
+            lit = log.get('iter')
+            if isinstance(lit, str) and lit.strip():
+                return lit.strip()
+        return None
+
+    explicit: dict[str, int] = {}
+    fallback: dict[str, int] = {}
+    for entry in state_entries:
+        run_id = _entry_run_id(entry)
+        if not run_id:
+            continue
+        rd = entry.get('runDic') or entry.get('run_dic') or entry.get('rundic')
+        if isinstance(rd, str) and rd.isdigit():
+            rd = int(rd)
+        if isinstance(rd, int):
+            explicit[run_id] = rd  # latest wins
+            continue
+        hit = _scan(entry)
+        if hit is not None:
+            cur = fallback.get(run_id)
+            if cur is None or hit > cur:
+                fallback[run_id] = hit
+
+    out = dict(fallback)
+    out.update(explicit)
+    return out
+
+
 def _step_artifact_paths(
     step: str,
     run_dic: int | None,
@@ -7176,9 +7619,6 @@ def _step_artifact_paths(
                 paths.append(
                     f'{runtime.binding.workspace_cwd}/output/relabel_candidates_{run_dic}.csv'
                 )
-    elif step == 'report':
-        if run_dic is not None:
-            paths.append(f'{chat_root}/results/workflow{run_dic}.md')
     elif step == 'hypothesis':
         paths.append(f'{chat_root}/results/iteration_log.jsonl')
     elif step == 'augment':
@@ -7342,6 +7782,8 @@ def _hardcoded_autoresearch_pipeline(
     *,
     agent_state: 'AgentState | None' = None,
     account_id: str | None = None,
+    iteration_log_count: int | None = None,
+    metrics_history: dict[str, Any] | None = None,
 ) -> dict[str, Any]:
     """Canonical autoresearch model-iteration pipeline structure.
 
@@ -7359,6 +7801,8 @@ def _hardcoded_autoresearch_pipeline(
             model_config=model_config,
             agent_state=agent_state,
             account_id=account_id,
+            iteration_log_count=iteration_log_count,
+            metrics_history=metrics_history,
         )
     else:
         # Defensive default if caller didn't pass sessions_dir.
@@ -7385,6 +7829,7 @@ def _hardcoded_autoresearch_pipeline(
         'items': [],
         'in_flight': None,
         'logs': [],
+        'metrics_history': metrics_history,
     }
 
 

frontend/app/app/api/claw/training/step-detail/route.ts

@@ -9,6 +9,7 @@ export async function GET(request: Request) {
 	const incoming = new URL(request.url);
 	const sessionId = incoming.searchParams.get("session_id") ?? "";
 	const step = incoming.searchParams.get("step") ?? "";
+	const runId = incoming.searchParams.get("run_id");
 	if (!sessionId || !step) {
 		return Response.json(
 			{ error: "session_id and step are required" },
@@ -20,6 +21,7 @@ export async function GET(request: Request) {
 	url.searchParams.set("account_id", account.id);
 	url.searchParams.set("session_id", sessionId);
 	url.searchParams.set("step", step);
+	if (runId) url.searchParams.set("run_id", runId);
 
 	const response = await fetch(url, { cache: "no-store" });
 	const payload = await response.text();

frontend/app/app/assistant.tsx

@@ -1,7 +1,11 @@
 "use client";
 
 import type { ExportedMessageRepository } from "@assistant-ui/core";
-import { AssistantRuntimeProvider, useAuiState } from "@assistant-ui/react";
+import {
+	AssistantRuntimeProvider,
+	useAui,
+	useAuiState,
+} from "@assistant-ui/react";
 import {
 	AssistantChatTransport,
 	useChatRuntime,
@@ -29,6 +33,7 @@ import {
 } from "@/components/ui/sidebar";
 import {
 	ACTIVE_SESSION_CHANGED_EVENT,
+	consumeFreshLocalId,
 	readActiveSessionId,
 	readPendingWorkspaceSessionId,
 	writeActiveSessionId,
@@ -91,6 +96,7 @@ export const Assistant = ({ initialSessionId }: AssistantProps = {}) => {
 	});
 	const replaySession = useCallback(
 		(sessionId: string, repository: ExportedMessageRepository) => {
+			console.log("[replay-session] called", { sessionId });
 			writeActiveSessionId(sessionId);
 			pushSessionUrl(sessionId);
 			// 后端已经落盘/结束后，用回放结果接管当前会话。
@@ -173,6 +179,22 @@ export const Assistant = ({ initialSessionId }: AssistantProps = {}) => {
 	);
 };
 
+// 共享 composer 草稿缓存里 newTask（无 sessionId）使用的 key。
+const NEWTASK_KEY = "__newtask__";
+
+// AssistantWorkspace cache effect 通知 ImeComposerInput 直接更新 textarea
+// 内容（绕过 store→local 间接路径，避免 switchToNewThread 时序问题）。
+export const COMPOSER_RESTORE_EVENT = "claw-composer-restore-draft";
+
+// __LOCALID_xxx 是侧栏在 newTask 状态下点工作区时临时分配的 placeholder
+// session id（让 jupyter bind 有 id 可传）。每个 LOCALID 拥有独立的 draft
+// key，避免多个 LOCALID 会话在 cache 里互相覆盖。newTask → 首次生成 LOCALID
+// 这一瞬的 textarea 闪空，由下面 save/restore effect 里的一次性继承处理。
+function getDraftKey(activeSessionId: string | null): string {
+	if (!activeSessionId) return NEWTASK_KEY;
+	return activeSessionId;
+}
+
 // 用户在对话里表达「我要开始/进行模型训练（这一轮迭代）」时命中的关键短语。
 // 要求动词或专属触发词必须存在；避免 "什么是模型训练" 之类无意义匹配。
 const APPLY_INTENT_PATTERNS: RegExp[] = [
@@ -197,6 +219,7 @@ function detectApplyIntent(text: string): boolean {
 }
 
 function AssistantWorkspace() {
+	const aui = useAui();
 	const [activeSessionId, setActiveSessionId] = useState<string | null>(() =>
 		readActiveSessionId(),
 	);
@@ -209,6 +232,12 @@ function AssistantWorkspace() {
 	const pipelineContainerRef = useRef<HTMLDivElement>(null);
 	const lastPipelineYRef = useRef<number | null>(null);
 	const [exitButtonVisible, setExitButtonVisible] = useState(false);
+	// Composer 是 per-thread 的，但本应用所有 backend session 都共用同一个
+	// assistant-ui thread（从不调 switchToThread），因此 composer.text 跨 session
+	// 共享。在 activeSessionId 切换时手动 save+restore，给每个 session 一份
+	// 独立的 draft。null sessionId（newTask）用 NEWTASK_KEY 落盘。
+	const composerCacheRef = useRef<Map<string, string>>(new Map());
+	const prevSessionKeyRef = useRef<string>(getDraftKey(activeSessionId));
 
 	const handlePipelineMouseMove = useCallback(
 		(event: React.MouseEvent<HTMLDivElement>) => {
@@ -244,11 +273,20 @@ function AssistantWorkspace() {
 	}, [showPipeline]);
 
 	useEffect(() => {
-		const update = () => setActiveSessionId(readActiveSessionId());
+		const update = () => {
+			const next = readActiveSessionId();
+			console.log("[active-session] focus-update", { next });
+			setActiveSessionId(next);
+		};
 		const handleChanged = (event: Event) => {
 			const detail = (event as CustomEvent<{ sessionId?: string | null }>)
 				.detail;
-			setActiveSessionId(detail?.sessionId ?? readActiveSessionId());
+			const next = detail?.sessionId ?? readActiveSessionId();
+			console.log("[active-session] event-changed", {
+				detailSessionId: detail?.sessionId,
+				next,
+			});
+			setActiveSessionId(next);
 		};
 		window.addEventListener(ACTIVE_SESSION_CHANGED_EVENT, handleChanged);
 		window.addEventListener("focus", update);
@@ -259,6 +297,59 @@ function AssistantWorkspace() {
 		};
 	}, []);
 
+	// 切 session 时保存当前 composer 草稿到旧 key、恢复新 key 的草稿。
+	useEffect(() => {
+		const cache = composerCacheRef.current;
+		const nextKey = getDraftKey(activeSessionId);
+		const prevKey = prevSessionKeyRef.current;
+		console.log("[draft-cache] effect-fire", {
+			prevKey,
+			nextKey,
+			activeSessionId,
+			willSkip: prevKey === nextKey,
+		});
+		if (prevKey === nextKey) return;
+		// 不能信任 aui store 的 composer.text——assistant-ui 内部 reducer 在
+		// 切换 / isEditing 翻转时会让 store 与 textarea 脱钩（典型现象：textarea
+		// 显示着 "hello" 但 store 已经是 ""），后果是这里把空串存进 cache、
+		// 用户的草稿被静默吞掉。直接读 DOM 拿 textarea 当前真实值。
+		// EditComposer 用的是 aui-edit-composer-input，不会撞 selector。
+		const composerEl = document.querySelector<HTMLTextAreaElement>(
+			"textarea.aui-composer-input",
+		);
+		const composerState = aui.composer().getState();
+		const savingText =
+			composerEl?.value ?? composerState.text ?? "";
+		cache.set(prevKey, savingText);
+		// newTask 里点工作区会**新生成**一个 LOCALID 并切过去，用户视角里
+		// 这是同一份草稿的延续——把 newTask 的文本继承到 LOCALID 名下，
+		// 避免 textarea 闪空。consumeFreshLocalId 只对前端刚生成的 LOCALID
+		// 命中，点击侧栏老 LOCALID 不命中，确保历史会话之间互不串。
+		const inheritFromNewTask =
+			prevKey === NEWTASK_KEY && consumeFreshLocalId(activeSessionId);
+		if (inheritFromNewTask) {
+			cache.set(nextKey, savingText);
+		}
+		const restored = cache.get(nextKey) ?? "";
+		console.log("[draft-cache] do-switch", {
+			prevKey,
+			nextKey,
+			savingText,
+			savingTextSource: composerEl ? "dom" : "store",
+			storeText: composerState.text,
+			restored,
+			inheritFromNewTask,
+		});
+		aui.composer().setText(restored);
+		window.dispatchEvent(
+			new CustomEvent(COMPOSER_RESTORE_EVENT, {
+				detail: { text: restored },
+			}),
+		);
+		console.log("[draft-cache] dispatched", { restored });
+		prevSessionKeyRef.current = nextKey;
+	}, [activeSessionId, aui]);
+
 	useEffect(() => {
 		closeActivityPanel();
 	}, [showPipeline, closeActivityPanel]);
@@ -280,36 +371,58 @@ function AssistantWorkspace() {
 		if (applied) return;
 		if (!activeSessionId) return;
 		let cancelled = false;
-		const url = `/api/claw/training/pipeline?session_id=${encodeURIComponent(activeSessionId)}`;
+		const pipelineUrl = `/api/claw/training/pipeline?session_id=${encodeURIComponent(activeSessionId)}`;
+		const runUrl = `/api/claw/runs/latest?session_id=${encodeURIComponent(activeSessionId)}`;
 		const tick = async () => {
 			try {
-				const res = await fetch(url, { cache: "no-store" });
+				const res = await fetch(pipelineUrl, { cache: "no-store" });
+				if (res.ok) {
+					const payload = await res.json();
+					if (cancelled) return;
+					if (payload) {
+						const items = Array.isArray(payload.items) ? payload.items : [];
+						const hasItemProgress = items.some(
+							(it: {
+								type?: string;
+								status?: string;
+								cards?: Array<{ status?: string }>;
+							}) => {
+								if (it.type === "gate") {
+									return it.status === "running" || it.status === "complete";
+								}
+								return (
+									Array.isArray(it.cards) &&
+									it.cards.some(
+										(c) =>
+											c.status === "running" || c.status === "complete",
+									)
+								);
+							},
+						);
+						// items[] 走 phase-complete-only 过滤；R0 只有一张卡在跑时
+						// items 是空的，进展信号在 in_flight 字段里——也算进展。
+						const hasInFlight = Boolean(payload.in_flight);
+						if (hasItemProgress || hasInFlight) {
+							setApplied(true);
+							return;
+						}
+					}
+				}
+			} catch {
+				/* ignore, fall through to runs/latest fallback */
+			}
+			// Pipeline 还没写出 items/in_flight（agent 处于 prep 阶段，比如在 watcher
+			// 里 sleep 等远端文件）时，pipeline 端点会返回 state=pending、items=[]，
+			// 但后端 run 其实已经在跑。直接看 run 状态兜底，免得用户起了训练但
+			// monitor 面板等到第一个 phase 才弹。
+			try {
+				const res = await fetch(runUrl, { cache: "no-store" });
 				if (!res.ok) return;
 				const payload = await res.json();
 				if (cancelled || !payload) return;
-				const items = Array.isArray(payload.items) ? payload.items : [];
+				if (payload.status === "running" || payload.status === "queued") {
-				const hasItemProgress = items.some(
+					setApplied(true);
-					(it: {
+				}
-						type?: string;
-						status?: string;
-						cards?: Array<{ status?: string }>;
-					}) => {
-						if (it.type === "gate") {
-							return it.status === "running" || it.status === "complete";
-						}
-						return (
-							Array.isArray(it.cards) &&
-							it.cards.some(
-								(c) =>
-									c.status === "running" || c.status === "complete",
-							)
-						);
-					},
-				);
-				// items[] 走 phase-complete-only 过滤；R0 只有一张卡在跑时
-				// items 是空的，进展信号在 in_flight 字段里——也算进展。
-				const hasInFlight = Boolean(payload.in_flight);
-				if (hasItemProgress || hasInFlight) setApplied(true);
 			} catch {
 				/* ignore */
 			}

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

@@ -24,6 +24,7 @@ import {
 import { Button } from "@/components/ui/button";
 import {
 	clearActiveSessionId,
+	markFreshLocalId,
 	readActiveSessionId,
 	writeActiveSessionId,
 	writePendingWorkspaceSessionId,
@@ -163,6 +164,7 @@ function ensureSidebarWorkspaceSessionId(): {
 		typeof crypto !== "undefined" && "randomUUID" in crypto
 			? `__LOCALID_${crypto.randomUUID().replace(/-/g, "").slice(0, 8)}`
 			: `__LOCALID_${Math.random().toString(36).slice(2, 10)}`;
+	markFreshLocalId(generated);
 	writeActiveSessionId(generated);
 	return { sessionId: generated, created: true };
 }
@@ -438,6 +440,10 @@ const ClawSessionList: FC = () => {
 									type="button"
 									className="aui-thread-list-item-trigger flex h-full min-w-0 flex-1 items-center truncate px-3 text-start text-sm"
 									onClick={async () => {
+										console.log("[session-list] click", {
+											sessionId: session.session_id,
+											preview: session.preview,
+										});
 										setOpenMenuSessionId(null);
 										writeActiveSessionId(session.session_id);
 										pushSessionUrl(session.session_id);

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

@@ -84,6 +84,7 @@ import {
 	ACTIVE_SESSION_CHANGED_EVENT,
 	clearActiveSessionId,
 	clearPendingWorkspaceSessionId,
+	markFreshLocalId,
 	PENDING_WORKSPACE_SESSION_CHANGED_EVENT,
 	readActiveSessionId,
 	readPendingWorkspaceSessionId,
@@ -261,8 +262,14 @@ function useRefreshCurrentRun() {
 		includePending: true,
 		includeActive: true,
 	});
-	const activeRunRef = useRef<string | null>(null);
+	// Per-session bookkeeping: switching sessions (sidebar history) and coming
-	const appliedTerminalRef = useRef<string | null>(null);
+	// back must NOT re-trigger replaySession on a session that has already been
+	// applied — replaySession rebuilds the thread tree, which causes a layout
+	// shift / scroll jitter in the chat UI. Single-value refs would get
+	// overwritten when visiting another session and forget that this session
+	// was already settled.
+	const activeRunRefMap = useRef<Map<string, string>>(new Map());
+	const appliedTerminalRefMap = useRef<Map<string, string>>(new Map());
 
 	useEffect(() => {
 		if (!sessionId) return;
@@ -274,8 +281,8 @@ function useRefreshCurrentRun() {
 			if (cancelled) return;
 			if (runStatus?.status === "queued" || runStatus?.status === "running") {
 				const runKey = runStatus.run_id ?? "active";
-				const previousRunKey = activeRunRef.current;
+				const previousRunKey = activeRunRefMap.current.get(sessionId) ?? null;
-				activeRunRef.current = runKey;
+				activeRunRefMap.current.set(sessionId, runKey);
 				// Watcher auto-resume: backend started a new run via _maybe_auto_resume
 				// without a local SSE stream driving the UI. We need to replay once so
 				// the active-run placeholder (pending prompt + 思考中 bubble) appears.
@@ -294,14 +301,15 @@ function useRefreshCurrentRun() {
 				}
 				return;
 			}
-			if (!runtimeRunning && !activeRunRef.current) return;
+			const sessionActiveRun = activeRunRefMap.current.get(sessionId) ?? null;
+			if (!runtimeRunning && !sessionActiveRun) return;
 			const terminalKey = [
-				runStatus?.run_id ?? activeRunRef.current ?? "unknown",
+				runStatus?.run_id ?? sessionActiveRun ?? "unknown",
 				runStatus?.status ?? "idle",
 				runStatus?.finished_at ?? "",
 				runStatus?.elapsed_ms ?? "",
 			].join(":");
-			if (appliedTerminalRef.current === terminalKey) return;
+			if (appliedTerminalRefMap.current.get(sessionId) === terminalKey) return;
 			const response = await fetch(`/api/claw/sessions/${sessionId}`, {
 				cache: "no-store",
 			});
@@ -311,8 +319,8 @@ function useRefreshCurrentRun() {
 				sessionId,
 				toReplayRepository(payload, sessionId, runStatus),
 			);
-			activeRunRef.current = null;
+			activeRunRefMap.current.delete(sessionId);
-			appliedTerminalRef.current = terminalKey;
+			appliedTerminalRefMap.current.set(sessionId, terminalKey);
 			scheduleSessionListRefresh();
 		}
 
@@ -775,6 +783,7 @@ function ensureWorkspaceSessionId(current: string | null) {
 		typeof crypto !== "undefined" && "randomUUID" in crypto
 			? `__LOCALID_${crypto.randomUUID().replace(/-/g, "").slice(0, 8)}`
 			: `__LOCALID_${Math.random().toString(36).slice(2, 10)}`;
+	markFreshLocalId(generated);
 	writeActiveSessionId(generated);
 	scheduleSessionListRefresh();
 	return generated;
@@ -2393,6 +2402,11 @@ const ImeComposerInput: FC<ComponentProps<typeof TextareaAutosize>> = ({
 	const justSubmittedRef = useRef(false);
 	const isDisabled = runtimeDisabled || disabled;
 
+	// Pull store→local on external store changes (session switch, paste-from-attachment,
+	// submit-clear, etc). The reverse direction (local→store) is pushed eagerly by
+	// onChange/handleInsert/handleKeyDown/onCompositionEnd via syncComposerText, so we
+	// don't need a symmetric effect — and adding one creates a feedback loop that
+	// re-renders the composer at ~1ms intervals (visible as shake).
 	useEffect(() => {
 		if (!isComposingRef.current) {
 			if (justSubmittedRef.current && storeText) return;
@@ -2401,16 +2415,6 @@ const ImeComposerInput: FC<ComponentProps<typeof TextareaAutosize>> = ({
 		}
 	}, [storeText]);
 
-	useEffect(() => {
-		if (
-			!isComposingRef.current &&
-			localText !== storeText &&
-			aui.composer().getState().isEditing
-		) {
-			aui.composer().setText(localText);
-		}
-	}, [aui, localText, storeText]);
-
 	useEffect(() => {
 		if (!autoFocus || isDisabled) return;
 		const textarea = textareaRef.current;
@@ -2433,6 +2437,21 @@ const ImeComposerInput: FC<ComponentProps<typeof TextareaAutosize>> = ({
 		};
 	}, [aui]);
 
+	// AssistantWorkspace 切 session 时通过事件通知我们恢复 draft——直接
+	// 控制 localText，绕过 store→local 路径上的 justSubmittedRef 屏蔽和
+	// switchToNewThread 引起的异步覆盖。
+	useEffect(() => {
+		const handler = (event: Event) => {
+			const detail = (event as CustomEvent<{ text?: string }>).detail;
+			const text = detail?.text ?? "";
+			justSubmittedRef.current = false;
+			setLocalText(text);
+		};
+		window.addEventListener("claw-composer-restore-draft", handler);
+		return () =>
+			window.removeEventListener("claw-composer-restore-draft", handler);
+	}, []);
+
 	const syncComposerText = useCallback(
 		(value: string) => {
 			if (!aui.composer().getState().isEditing) return;

frontend/app/components/training/metrics-chart-panel.tsx

@@ -0,0 +1,214 @@
+"use client";
+
+import {
+	CartesianGrid,
+	Line,
+	LineChart,
+	ResponsiveContainer,
+	Tooltip,
+	XAxis,
+	YAxis,
+} from "recharts";
+import { cn } from "@/lib/utils";
+
+export type Round = {
+	iteration: number;
+	runDic: number;
+	label: string;
+	timestamp: string;
+	values: Record<string, number | null>;
+};
+
+export type MetricsHistory = {
+	metrics: string[];
+	rounds: Round[];
+};
+
+const METRIC_LABELS: Record<string, string> = {
+	req_set_car: "需求集合(车载)",
+	dapan_car: "大盘车载",
+	specific_test: "Specific Test",
+	triage_err_rate: "Triage 错误率",
+	bvt_nav: "导航BVT",
+	talkable_controllable: "可聊可控",
+	target_subset_count: "目标子集数量",
+	target_subset_wrong: "目标子集错误数",
+	target_subset_correct: "目标子集正确数",
+	target_subset_acc: "目标子集准确率",
+	icl_test_overall: "ICL 测试整体",
+	specific_test_overall: "专项测试整体",
+	overrall: "大盘集",
+	overrall_che_zai: "大盘集车载",
+	overrall_shou_ji: "大盘集手机",
+	overrall_dian_shi: "大盘集电视",
+	overrall_yan_jing: "大盘集眼镜",
+	overrall_you_ping: "大盘集有屏",
+	overrall_wu_ping: "大盘集无屏",
+};
+
+const LINE_COLORS = [
+	"#34d399", // emerald
+	"#38bdf8", // sky
+	"#a78bfa", // violet
+	"#fbbf24", // amber
+	"#f472b6", // rose
+	"#22d3ee", // cyan
+	"#a3e635", // lime
+	"#fb923c", // orange
+];
+
+function labelOf(key: string): string {
+	return METRIC_LABELS[key] ?? key;
+}
+
+function formatValue(v: number | null | undefined): string {
+	if (v === null || v === undefined || Number.isNaN(v)) return "—";
+	return v.toFixed(4);
+}
+
+function formatDelta(curr: number | null | undefined, base: number | null | undefined): {
+	text: string;
+	tone: "up" | "down" | "flat";
+} {
+	if (
+		curr === null ||
+		curr === undefined ||
+		base === null ||
+		base === undefined ||
+		Number.isNaN(curr) ||
+		Number.isNaN(base)
+	) {
+		return { text: "—", tone: "flat" };
+	}
+	const d = curr - base;
+	if (Math.abs(d) < 1e-6) return { text: "±0", tone: "flat" };
+	const sign = d > 0 ? "+" : "";
+	return { text: `${sign}${d.toFixed(4)}`, tone: d > 0 ? "up" : "down" };
+}
+
+function chartRowsFor(history: MetricsHistory, key: string): Array<Record<string, unknown>> {
+	return history.rounds.map((r) => {
+		const v = r.values[key];
+		return {
+			label: r.label,
+			[key]: v === undefined || v === null ? null : v,
+		};
+	});
+}
+
+export function MetricsChartPanel({ history }: { history: MetricsHistory | null }) {
+	if (!history || history.rounds.length === 0) {
+		return (
+			<div className="flex h-full items-center justify-center text-zinc-500 text-xs">
+				暂无指标数据，等首轮分析完成
+			</div>
+		);
+	}
+
+	const latest = history.rounds[history.rounds.length - 1];
+	const baseline = history.rounds[0];
+	const prev =
+		history.rounds.length >= 2 ? history.rounds[history.rounds.length - 2] : baseline;
+
+	const colorFor = (key: string): string => {
+		const idx = history.metrics.indexOf(key);
+		return LINE_COLORS[idx >= 0 ? idx % LINE_COLORS.length : 0];
+	};
+
+	return (
+		<div className="h-full min-h-0 overflow-y-auto pr-1">
+			<div className="grid gap-2 [grid-template-columns:repeat(auto-fit,minmax(280px,1fr))]">
+				{history.metrics.map((k) => {
+					const curr = latest.values[k];
+					const dBase = formatDelta(curr, baseline.values[k]);
+					const dPrev = formatDelta(curr, prev.values[k]);
+					const color = colorFor(k);
+					const rows = chartRowsFor(history, k);
+					return (
+						<div
+							key={k}
+							className="flex flex-col rounded-md border border-zinc-800 bg-zinc-900/40 p-2"
+						>
+							<div className="flex shrink-0 items-baseline justify-between gap-2">
+								<div className="flex items-center gap-1.5 min-w-0">
+									<span
+										className="size-1.5 shrink-0 rounded-full"
+										style={{ backgroundColor: color }}
+									/>
+									<span className="truncate font-medium text-[11px] text-zinc-200">
+										{labelOf(k)}
+									</span>
+								</div>
+								<span className="font-mono font-semibold text-[14px] text-zinc-50 leading-none">
+									{formatValue(curr)}
+								</span>
+							</div>
+							<div className="mt-0.5 flex shrink-0 gap-2 font-mono text-[10px]">
+								<span className={deltaClass(dBase.tone)}>
+									Δbase {dBase.text}
+								</span>
+								<span className={deltaClass(dPrev.tone)}>
+									Δprev {dPrev.text}
+								</span>
+							</div>
+							<div className="mt-1.5 h-32">
+								<ResponsiveContainer width="100%" height="100%">
+									<LineChart
+										data={rows}
+										margin={{ top: 4, right: 8, left: -12, bottom: 0 }}
+									>
+										<CartesianGrid stroke="#27272a" strokeDasharray="3 3" />
+										<XAxis
+											dataKey="label"
+											stroke="#a1a1aa"
+											tick={{ fontSize: 10 }}
+										/>
+										<YAxis
+											stroke="#a1a1aa"
+											tick={{ fontSize: 10 }}
+											tickFormatter={(v: number) =>
+												typeof v === "number" ? v.toFixed(4) : v
+											}
+											domain={["auto", "auto"]}
+											width={60}
+										/>
+										<Tooltip
+											contentStyle={{
+												background: "#18181b",
+												border: "1px solid #3f3f46",
+												borderRadius: 6,
+												fontSize: 11,
+											}}
+											labelStyle={{ color: "#e4e4e7" }}
+											itemStyle={{ color: "#e4e4e7" }}
+											formatter={(value: number, name: string) => [
+												typeof value === "number" ? value.toFixed(4) : value,
+												labelOf(name),
+											]}
+										/>
+										<Line
+											type="monotone"
+											dataKey={k}
+											stroke={color}
+											strokeWidth={1.5}
+											dot={{ r: 2.5 }}
+											activeDot={{ r: 4 }}
+											connectNulls={false}
+											isAnimationActive={false}
+										/>
+									</LineChart>
+								</ResponsiveContainer>
+							</div>
+						</div>
+					);
+				})}
+			</div>
+		</div>
+	);
+}
+
+function deltaClass(tone: "up" | "down" | "flat"): string {
+	if (tone === "up") return "text-emerald-400";
+	if (tone === "down") return "text-rose-400";
+	return "text-zinc-500";
+}

frontend/app/components/training/step-detail-sheet.tsx

@@ -3,10 +3,12 @@
 import {
 	ExternalLinkIcon,
 	FileIcon,
+	Maximize2Icon,
+	Minimize2Icon,
 	RefreshCwIcon,
 	XIcon,
 } from "lucide-react";
-import { useEffect, useMemo, useState } from "react";
+import { useEffect, useMemo, useRef, useState } from "react";
 import ReactMarkdown from "react-markdown";
 import remarkGfm from "remark-gfm";
 import { Button } from "@/components/ui/button";
@@ -52,6 +54,7 @@ type StepDetail = {
 type Props = {
 	sessionId: string | null;
 	stepKey: string | null;
+	runId?: string | null;
 	stepTitle?: string;
 	stepStatus?: string;
 	open: boolean;
@@ -61,6 +64,7 @@ type Props = {
 export function StepDetailSheet({
 	sessionId,
 	stepKey,
+	runId,
 	stepTitle,
 	stepStatus,
 	open,
@@ -69,13 +73,41 @@ export function StepDetailSheet({
 	const [detail, setDetail] = useState<StepDetail | null>(null);
 	const [loading, setLoading] = useState(false);
 	const [reloadCount, setReloadCount] = useState(0);
+	const [width, setWidth] = useState(672);
+	const [isFullscreen, setIsFullscreen] = useState(false);
+	const dragRef = useRef<{ startX: number; startW: number } | null>(null);
+
+	const onResizePointerDown = (e: React.PointerEvent<HTMLDivElement>) => {
+		if (isFullscreen) return;
+		e.currentTarget.setPointerCapture(e.pointerId);
+		dragRef.current = { startX: e.clientX, startW: width };
+	};
+	const onResizePointerMove = (e: React.PointerEvent<HTMLDivElement>) => {
+		if (!dragRef.current) return;
+		const dx = dragRef.current.startX - e.clientX;
+		const max =
+			typeof window !== "undefined"
+				? Math.max(360, window.innerWidth - 32)
+				: 1200;
+		setWidth(Math.max(360, Math.min(max, dragRef.current.startW + dx)));
+	};
+	const onResizePointerUp = (e: React.PointerEvent<HTMLDivElement>) => {
+		if (!dragRef.current) return;
+		dragRef.current = null;
+		e.currentTarget.releasePointerCapture(e.pointerId);
+	};
 
 	useEffect(() => {
 		if (!open || !sessionId || !stepKey) return;
 		let cancelled = false;
 		setLoading(true);
 		setDetail(null);
-		const url = `/api/claw/training/step-detail?session_id=${encodeURIComponent(sessionId)}&step=${encodeURIComponent(stepKey)}`;
+		const params = new URLSearchParams({
+			session_id: sessionId,
+			step: stepKey,
+		});
+		if (runId) params.set("run_id", runId);
+		const url = `/api/claw/training/step-detail?${params.toString()}`;
 		fetch(url, { cache: "no-store" })
 			.then((res) => (res.ok ? res.json() : null))
 			.then((payload) => {
@@ -89,15 +121,39 @@ export function StepDetailSheet({
 		return () => {
 			cancelled = true;
 		};
-	}, [open, sessionId, stepKey, reloadCount]);
+	}, [open, sessionId, stepKey, runId, reloadCount]);
 
 	return (
 		<Sheet open={open} onOpenChange={onOpenChange}>
 			<SheetContent
 				side="right"
 				hideClose
-				className="flex w-[min(50rem,calc(100vw-2rem))] flex-col gap-0 p-0 sm:max-w-2xl"
+				onEscapeKeyDown={(e) => {
+					if (isFullscreen) {
+						e.preventDefault();
+						setIsFullscreen(false);
+					}
+				}}
+				className="flex max-w-none flex-col gap-0 p-0 sm:max-w-none"
+				style={
+					isFullscreen
+						? { width: "100vw", maxWidth: "100vw" }
+						: { width: `${width}px`, maxWidth: "none" }
+				}
 			>
+				{!isFullscreen ? (
+					<div
+						onPointerDown={onResizePointerDown}
+						onPointerMove={onResizePointerMove}
+						onPointerUp={onResizePointerUp}
+						onPointerCancel={onResizePointerUp}
+						className="group absolute inset-y-0 left-0 z-50 w-1.5 cursor-ew-resize hover:bg-sky-500/40 active:bg-sky-500/60"
+						title="拖动调整宽度"
+						aria-label="拖动调整宽度"
+					>
+						<div className="-translate-y-1/2 absolute top-1/2 left-[2px] h-10 w-[2px] rounded-full bg-muted-foreground/30 group-hover:bg-sky-400" />
+					</div>
+				) : null}
 				<header className="flex shrink-0 items-start justify-between gap-3 border-b px-5 py-3">
 					<div className="min-w-0">
 						<div className="flex items-center gap-2">
@@ -119,6 +175,21 @@ export function StepDetailSheet({
 						) : null}
 					</div>
 					<div className="mt-0.5 flex shrink-0 items-center gap-1">
+						<Button
+							type="button"
+							variant="ghost"
+							size="icon"
+							className="size-7"
+							onClick={() => setIsFullscreen((v) => !v)}
+							title={isFullscreen ? "退出全屏 (Esc)" : "全屏"}
+							aria-label={isFullscreen ? "退出全屏" : "全屏"}
+						>
+							{isFullscreen ? (
+								<Minimize2Icon className="size-3.5" />
+							) : (
+								<Maximize2Icon className="size-3.5" />
+							)}
+						</Button>
 						<Button
 							type="button"
 							variant="ghost"

frontend/app/components/training/training-pipeline-panel.tsx

@@ -8,20 +8,25 @@ import {
 	ClipboardListIcon,
 	ClockIcon,
 	DnaIcon,
-	ExternalLinkIcon,
 	FileTextIcon,
 	FlaskConicalIcon,
 	GraduationCapIcon,
 	HourglassIcon,
 	Loader2Icon,
 	type LucideIcon,
+	Maximize2Icon,
 	MicroscopeIcon,
+	Minimize2Icon,
 	RefreshCwIcon,
 	ShieldCheckIcon,
 	TrendingUpIcon,
 	UserCheckIcon,
 } from "lucide-react";
-import { useEffect, useRef, useState } from "react";
+import { type ReactNode, useEffect, useRef, useState } from "react";
+import {
+	MetricsChartPanel,
+	type MetricsHistory,
+} from "@/components/training/metrics-chart-panel";
 import { StepDetailSheet } from "@/components/training/step-detail-sheet";
 import { cn } from "@/lib/utils";
 
@@ -103,6 +108,7 @@ type PipelinePayload = {
 	items: PipelineItem[];
 	in_flight?: InFlight | null;
 	logs: LogLine[];
+	metrics_history?: MetricsHistory | null;
 };
 
 const ICON_MAP: Record<string, LucideIcon> = {
@@ -145,7 +151,9 @@ export function TrainingPipelinePanel({
 		key: string;
 		title: string;
 		status: string;
+		runId?: string;
 	} | null>(null);
+	const [bottomTab, setBottomTab] = useState<"log" | "metrics">("log");
 
 	useEffect(() => {
 		setLoading(true);
@@ -182,7 +190,7 @@ export function TrainingPipelinePanel({
 	};
 
 	return (
-		<div className="flex min-w-0 flex-1 flex-col overflow-hidden border-r bg-muted/20">
+		<div className="relative flex min-w-0 flex-1 flex-col overflow-hidden border-r bg-muted/20">
 			<div className="min-h-0 flex-1 overflow-y-auto px-6 py-5">
 				<HeaderRow status={data?.status ?? null} loading={loading} />
 				<KpiRow kpis={data?.kpis ?? []} />
@@ -222,6 +230,7 @@ export function TrainingPipelinePanel({
 															key: stripNamespace(card.key),
 															title: card.title,
 															status: card.status,
+															runId: item.run_id,
 														})
 													}
 												/>
@@ -238,6 +247,7 @@ export function TrainingPipelinePanel({
 														key: stripNamespace(item.key),
 														title: item.title,
 														status: item.status,
+														runId: item.run_id,
 													})
 												}
 											/>
@@ -250,11 +260,16 @@ export function TrainingPipelinePanel({
 					})()}
 				</div>
 			</div>
-			<LegendBar />
+			<BottomPanel
-			<LogStream logs={data?.logs ?? []} />
+				logs={data?.logs ?? []}
+				history={data?.metrics_history ?? null}
+				tab={bottomTab}
+				onTabChange={setBottomTab}
+			/>
 			<StepDetailSheet
 				sessionId={sessionId}
 				stepKey={openCard?.key ?? null}
+				runId={openCard?.runId ?? null}
 				stepTitle={openCard?.title}
 				stepStatus={openCard?.status}
 				open={openCard !== null}
@@ -774,96 +789,165 @@ function MiniStatus({ status }: { status: CardStatus }) {
 	);
 }
 
-function LegendBar() {
+const BOTTOM_PANEL_DEFAULT_HEIGHT = 288;
-	const items = [
+const BOTTOM_PANEL_MIN_HEIGHT = 120;
-		{
+
-			icon: BarChart3Icon,
+function BottomPanel({
-			label: "Measure",
+	logs,
-			sub: "Quantify performance",
+	history,
-			color:
+	tab,
-				"bg-sky-500/10 text-sky-700 dark:text-sky-300 border-sky-500/30",
+	onTabChange,
-		},
+}: {
-		{
+	logs: LogLine[];
-			icon: DnaIcon,
+	history: MetricsHistory | null;
-			label: "Improve",
+	tab: "log" | "metrics";
-			sub: "Data, training, quality",
+	onTabChange: (next: "log" | "metrics") => void;
-			color:
+}) {
-				"bg-emerald-500/10 text-emerald-700 dark:text-emerald-300 border-emerald-500/30",
+	const [height, setHeight] = useState(BOTTOM_PANEL_DEFAULT_HEIGHT);
-		},
+	const [isFullscreen, setIsFullscreen] = useState(false);
-		{
+	const dragRef = useRef<{ startY: number; startH: number } | null>(null);
-			icon: MicroscopeIcon,
+
-			label: "Understand",
+	useEffect(() => {
-			sub: "Analyze & diagnose",
+		if (!isFullscreen) return;
-			color:
+		const onKey = (e: KeyboardEvent) => {
-				"bg-violet-500/10 text-violet-700 dark:text-violet-300 border-violet-500/30",
+			if (e.key === "Escape") setIsFullscreen(false);
-		},
+		};
-		{
+		window.addEventListener("keydown", onKey);
-			icon: RefreshCwIcon,
+		return () => window.removeEventListener("keydown", onKey);
-			label: "Iterate",
+	}, [isFullscreen]);
-			sub: "Repeat & grow",
+
-			color:
+	const onHandlePointerDown = (e: React.PointerEvent<HTMLDivElement>) => {
-				"bg-amber-500/10 text-amber-700 dark:text-amber-300 border-amber-500/30",
+		if (isFullscreen) return;
-		},
+		e.currentTarget.setPointerCapture(e.pointerId);
-	];
+		dragRef.current = { startY: e.clientY, startH: height };
+	};
+	const onHandlePointerMove = (e: React.PointerEvent<HTMLDivElement>) => {
+		if (!dragRef.current) return;
+		const dy = dragRef.current.startY - e.clientY;
+		const max =
+			typeof window !== "undefined"
+				? Math.max(BOTTOM_PANEL_MIN_HEIGHT, window.innerHeight - 100)
+				: 800;
+		setHeight(
+			Math.max(
+				BOTTOM_PANEL_MIN_HEIGHT,
+				Math.min(max, dragRef.current.startH + dy),
+			),
+		);
+	};
+	const onHandlePointerUp = (e: React.PointerEvent<HTMLDivElement>) => {
+		if (!dragRef.current) return;
+		dragRef.current = null;
+		e.currentTarget.releasePointerCapture(e.pointerId);
+	};
+
 	return (
-		<div className="flex shrink-0 items-center gap-4 border-t bg-background/80 px-6 py-2.5 overflow-x-auto">
+		<div
-			{items.map(({ icon: Icon, label, sub, color }) => (
+			className={cn(
-				<div key={label} className="flex items-center gap-2 shrink-0">
+				"flex flex-col bg-zinc-950",
-					<span
+				isFullscreen
-						className={cn(
+					? "absolute inset-0 z-50 border-t"
-							"flex size-7 items-center justify-center rounded-full border",
+					: "shrink-0 border-t",
-							color,
+			)}
-						)}
+			style={isFullscreen ? undefined : { height: `${height}px` }}
-					>
+		>
-						<Icon className="size-3.5" />
+			{!isFullscreen && (
-					</span>
+				<div
-					<div className="leading-tight">
+					onPointerDown={onHandlePointerDown}
-						<div className="text-[11px] font-semibold text-foreground">
+					onPointerMove={onHandlePointerMove}
-							{label}
+					onPointerUp={onHandlePointerUp}
-						</div>
+					onPointerCancel={onHandlePointerUp}
-						<div className="text-[10px] text-muted-foreground">{sub}</div>
+					className="group h-1.5 shrink-0 cursor-ns-resize bg-zinc-900 hover:bg-sky-500/40 active:bg-sky-500/60"
-					</div>
+					title="拖动调整高度"
+				>
+					<div className="mx-auto mt-[2px] h-[2px] w-10 rounded-full bg-zinc-700 group-hover:bg-sky-400" />
 				</div>
-			))}
+			)}
+			<div className="flex min-h-0 flex-1 flex-col px-5 py-3">
+				<div className="mb-2 flex shrink-0 items-center justify-between">
+					<div className="inline-flex rounded-full bg-zinc-900/80 p-0.5">
+						<PillTab active={tab === "log"} onClick={() => onTabChange("log")}>
+							Live Log
+						</PillTab>
+						<PillTab
+							active={tab === "metrics"}
+							onClick={() => onTabChange("metrics")}
+						>
+							Metrics
+						</PillTab>
+					</div>
+					<button
+						type="button"
+						onClick={() => setIsFullscreen((v) => !v)}
+						className="inline-flex items-center gap-1 text-[11px] text-zinc-400 transition-colors hover:text-zinc-200"
+						title={isFullscreen ? "退出 (Esc)" : "查看全部"}
+					>
+						查看全部
+						{isFullscreen ? (
+							<Minimize2Icon className="size-3" />
+						) : (
+							<Maximize2Icon className="size-3" />
+						)}
+					</button>
+				</div>
+				<div className="min-h-0 flex-1">
+					{tab === "log" ? (
+						<LogStreamBody logs={logs} />
+					) : (
+						<MetricsChartPanel history={history} />
+					)}
+				</div>
+			</div>
 		</div>
 	);
 }
 
-function LogStream({ logs }: { logs: LogLine[] }) {
+function PillTab({
+	active,
+	onClick,
+	children,
+}: {
+	active: boolean;
+	onClick: () => void;
+	children: ReactNode;
+}) {
+	return (
+		<button
+			type="button"
+			onClick={onClick}
+			className={cn(
+				"rounded-full px-3 py-1 font-semibold text-[11px] uppercase tracking-[0.12em] transition-colors",
+				active
+					? "bg-zinc-700 text-zinc-50"
+					: "text-zinc-400 hover:text-zinc-200",
+			)}
+		>
+			{children}
+		</button>
+	);
+}
+
+function LogStreamBody({ logs }: { logs: LogLine[] }) {
 	const ref = useRef<HTMLDivElement>(null);
 	useEffect(() => {
 		if (ref.current) ref.current.scrollTop = ref.current.scrollHeight;
 	}, [logs]);
 	return (
-		<div className="flex h-72 shrink-0 flex-col border-t bg-zinc-950 px-5 py-3">
+		<div
-			<div className="mb-2 flex shrink-0 items-center justify-between">
+			ref={ref}
-				<span className="font-semibold text-xs uppercase tracking-[0.18em] text-zinc-300">
+			className="h-full min-h-0 overflow-y-auto font-mono text-[12px] text-zinc-100 leading-6"
-					Live Log
+		>
-				</span>
+			{logs.length === 0 ? (
-				<button
+				<div className="text-zinc-500">暂无日志</div>
-					type="button"
+			) : (
-					className="inline-flex items-center gap-1 text-[11px] text-zinc-400 transition-colors hover:text-zinc-200"
+				logs.map((line, idx) => (
-				>
+					<div key={idx} className="flex gap-3">
-					查看全部
+						<span className="text-zinc-500">{line.ts}</span>
-					<ExternalLinkIcon className="size-3" />
+						<span className="text-emerald-400">[iter={line.iter}]</span>
-				</button>
+						<span>{line.text}</span>
-			</div>
+					</div>
-			<div
+				))
-				ref={ref}
+			)}
-				className="min-h-0 flex-1 overflow-y-auto font-mono text-[12px] leading-6 text-zinc-100"
-			>
-				{logs.length === 0 ? (
-					<div className="text-zinc-500">暂无日志</div>
-				) : (
-					logs.map((line, idx) => (
-						<div key={idx} className="flex gap-3">
-							<span className="text-zinc-500">{line.ts}</span>
-							<span className="text-emerald-400">[iter={line.iter}]</span>
-							<span>{line.text}</span>
-						</div>
-					))
-				)}
-			</div>
 		</div>
 	);
 }

frontend/app/lib/claw-active-session.ts

@@ -87,3 +87,19 @@ function normalizeSessionId(value: unknown) {
 	const trimmed = value.trim();
 	return trimmed || null;
 }
+
+// 记录"前端刚生成的 LOCALID"。AssistantWorkspace 的 draft cache effect 用它
+// 区分：从 newTask 切到一个**新生成**的 LOCALID（草稿应继承）vs 切到一个
+// **侧栏点击**的老 LOCALID（草稿不应继承）。
+const freshLocalIds = new Set<string>();
+
+export function markFreshLocalId(sessionId: string) {
+	freshLocalIds.add(sessionId);
+}
+
+export function consumeFreshLocalId(sessionId: string | null): boolean {
+	if (!sessionId) return false;
+	if (!freshLocalIds.has(sessionId)) return false;
+	freshLocalIds.delete(sessionId);
+	return true;
+}

frontend/app/package-lock.json

@@ -27,6 +27,7 @@
 				"react": "^19.2.5",
 				"react-dom": "^19.2.5",
 				"react-markdown": "^10.1.0",
+				"recharts": "^2.15.4",
 				"remark-gfm": "^4.0.1",
 				"tailwind-merge": "^3.5.0",
 				"tw-animate-css": "^1.4.0",
@@ -4691,6 +4692,69 @@
 				"tailwindcss": ">=3.0.0 || insiders || >=4.0.0-alpha.20 || >=4.0.0-beta.1"
 			}
 		},
+		"node_modules/@types/d3-array": {
+			"version": "3.2.2",
+			"resolved": "https://registry.npmjs.org/@types/d3-array/-/d3-array-3.2.2.tgz",
+			"integrity": "sha512-hOLWVbm7uRza0BYXpIIW5pxfrKe0W+D5lrFiAEYR+pb6w3N2SwSMaJbXdUfSEv+dT4MfHBLtn5js0LAWaO6otw==",
+			"license": "MIT"
+		},
+		"node_modules/@types/d3-color": {
+			"version": "3.1.3",
+			"resolved": "https://registry.npmjs.org/@types/d3-color/-/d3-color-3.1.3.tgz",
+			"integrity": "sha512-iO90scth9WAbmgv7ogoq57O9YpKmFBbmoEoCHDB2xMBY0+/KVrqAaCDyCE16dUspeOvIxFFRI+0sEtqDqy2b4A==",
+			"license": "MIT"
+		},
+		"node_modules/@types/d3-ease": {
+			"version": "3.0.2",
+			"resolved": "https://registry.npmjs.org/@types/d3-ease/-/d3-ease-3.0.2.tgz",
+			"integrity": "sha512-NcV1JjO5oDzoK26oMzbILE6HW7uVXOHLQvHshBUW4UMdZGfiY6v5BeQwh9a9tCzv+CeefZQHJt5SRgK154RtiA==",
+			"license": "MIT"
+		},
+		"node_modules/@types/d3-interpolate": {
+			"version": "3.0.4",
+			"resolved": "https://registry.npmjs.org/@types/d3-interpolate/-/d3-interpolate-3.0.4.tgz",
+			"integrity": "sha512-mgLPETlrpVV1YRJIglr4Ez47g7Yxjl1lj7YKsiMCb27VJH9W8NVM6Bb9d8kkpG/uAQS5AmbA48q2IAolKKo1MA==",
+			"license": "MIT",
+			"dependencies": {
+				"@types/d3-color": "*"
+			}
+		},
+		"node_modules/@types/d3-path": {
+			"version": "3.1.1",
+			"resolved": "https://registry.npmjs.org/@types/d3-path/-/d3-path-3.1.1.tgz",
+			"integrity": "sha512-VMZBYyQvbGmWyWVea0EHs/BwLgxc+MKi1zLDCONksozI4YJMcTt8ZEuIR4Sb1MMTE8MMW49v0IwI5+b7RmfWlg==",
+			"license": "MIT"
+		},
+		"node_modules/@types/d3-scale": {
+			"version": "4.0.9",
+			"resolved": "https://registry.npmjs.org/@types/d3-scale/-/d3-scale-4.0.9.tgz",
+			"integrity": "sha512-dLmtwB8zkAeO/juAMfnV+sItKjlsw2lKdZVVy6LRr0cBmegxSABiLEpGVmSJJ8O08i4+sGR6qQtb6WtuwJdvVw==",
+			"license": "MIT",
+			"dependencies": {
+				"@types/d3-time": "*"
+			}
+		},
+		"node_modules/@types/d3-shape": {
+			"version": "3.1.8",
+			"resolved": "https://registry.npmjs.org/@types/d3-shape/-/d3-shape-3.1.8.tgz",
+			"integrity": "sha512-lae0iWfcDeR7qt7rA88BNiqdvPS5pFVPpo5OfjElwNaT2yyekbM0C9vK+yqBqEmHr6lDkRnYNoTBYlAgJa7a4w==",
+			"license": "MIT",
+			"dependencies": {
+				"@types/d3-path": "*"
+			}
+		},
+		"node_modules/@types/d3-time": {
+			"version": "3.0.4",
+			"resolved": "https://registry.npmjs.org/@types/d3-time/-/d3-time-3.0.4.tgz",
+			"integrity": "sha512-yuzZug1nkAAaBlBBikKZTgzCeA+k1uy4ZFwWANOfKw5z5LRhV0gNA7gNkKm7HoK+HRN0wX3EkxGk0fpbWhmB7g==",
+			"license": "MIT"
+		},
+		"node_modules/@types/d3-timer": {
+			"version": "3.0.2",
+			"resolved": "https://registry.npmjs.org/@types/d3-timer/-/d3-timer-3.0.2.tgz",
+			"integrity": "sha512-Ps3T8E8dZDam6fUyNiMkekK3XUsaUEik+idO9/YjPtfj2qruF8tFBXS7XhtE4iIXBLxhmLjP3SXpLhVf21I9Lw==",
+			"license": "MIT"
+		},
 		"node_modules/@types/debug": {
 			"version": "4.1.13",
 			"resolved": "https://registry.npmjs.org/@types/debug/-/debug-4.1.13.tgz",
@@ -4963,6 +5027,127 @@
 			"resolved": "https://registry.npmjs.org/csstype/-/csstype-3.2.3.tgz",
 			"integrity": "sha512-z1HGKcYy2xA8AGQfwrn0PAy+PB7X/GSj3UVJW9qKyn43xWa+gl5nXmU4qqLMRzWVLFC8KusUX8T/0kCiOYpAIQ=="
 		},
+		"node_modules/d3-array": {
+			"version": "3.2.4",
+			"resolved": "https://registry.npmjs.org/d3-array/-/d3-array-3.2.4.tgz",
+			"integrity": "sha512-tdQAmyA18i4J7wprpYq8ClcxZy3SC31QMeByyCFyRt7BVHdREQZ5lpzoe5mFEYZUWe+oq8HBvk9JjpibyEV4Jg==",
+			"license": "ISC",
+			"dependencies": {
+				"internmap": "1 - 2"
+			},
+			"engines": {
+				"node": ">=12"
+			}
+		},
+		"node_modules/d3-color": {
+			"version": "3.1.0",
+			"resolved": "https://registry.npmjs.org/d3-color/-/d3-color-3.1.0.tgz",
+			"integrity": "sha512-zg/chbXyeBtMQ1LbD/WSoW2DpC3I0mpmPdW+ynRTj/x2DAWYrIY7qeZIHidozwV24m4iavr15lNwIwLxRmOxhA==",
+			"license": "ISC",
+			"engines": {
+				"node": ">=12"
+			}
+		},
+		"node_modules/d3-ease": {
+			"version": "3.0.1",
+			"resolved": "https://registry.npmjs.org/d3-ease/-/d3-ease-3.0.1.tgz",
+			"integrity": "sha512-wR/XK3D3XcLIZwpbvQwQ5fK+8Ykds1ip7A2Txe0yxncXSdq1L9skcG7blcedkOX+ZcgxGAmLX1FrRGbADwzi0w==",
+			"license": "BSD-3-Clause",
+			"engines": {
+				"node": ">=12"
+			}
+		},
+		"node_modules/d3-format": {
+			"version": "3.1.2",
+			"resolved": "https://registry.npmjs.org/d3-format/-/d3-format-3.1.2.tgz",
+			"integrity": "sha512-AJDdYOdnyRDV5b6ArilzCPPwc1ejkHcoyFarqlPqT7zRYjhavcT3uSrqcMvsgh2CgoPbK3RCwyHaVyxYcP2Arg==",
+			"license": "ISC",
+			"engines": {
+				"node": ">=12"
+			}
+		},
+		"node_modules/d3-interpolate": {
+			"version": "3.0.1",
+			"resolved": "https://registry.npmjs.org/d3-interpolate/-/d3-interpolate-3.0.1.tgz",
+			"integrity": "sha512-3bYs1rOD33uo8aqJfKP3JWPAibgw8Zm2+L9vBKEHJ2Rg+viTR7o5Mmv5mZcieN+FRYaAOWX5SJATX6k1PWz72g==",
+			"license": "ISC",
+			"dependencies": {
+				"d3-color": "1 - 3"
+			},
+			"engines": {
+				"node": ">=12"
+			}
+		},
+		"node_modules/d3-path": {
+			"version": "3.1.0",
+			"resolved": "https://registry.npmjs.org/d3-path/-/d3-path-3.1.0.tgz",
+			"integrity": "sha512-p3KP5HCf/bvjBSSKuXid6Zqijx7wIfNW+J/maPs+iwR35at5JCbLUT0LzF1cnjbCHWhqzQTIN2Jpe8pRebIEFQ==",
+			"license": "ISC",
+			"engines": {
+				"node": ">=12"
+			}
+		},
+		"node_modules/d3-scale": {
+			"version": "4.0.2",
+			"resolved": "https://registry.npmjs.org/d3-scale/-/d3-scale-4.0.2.tgz",
+			"integrity": "sha512-GZW464g1SH7ag3Y7hXjf8RoUuAFIqklOAq3MRl4OaWabTFJY9PN/E1YklhXLh+OQ3fM9yS2nOkCoS+WLZ6kvxQ==",
+			"license": "ISC",
+			"dependencies": {
+				"d3-array": "2.10.0 - 3",
+				"d3-format": "1 - 3",
+				"d3-interpolate": "1.2.0 - 3",
+				"d3-time": "2.1.1 - 3",
+				"d3-time-format": "2 - 4"
+			},
+			"engines": {
+				"node": ">=12"
+			}
+		},
+		"node_modules/d3-shape": {
+			"version": "3.2.0",
+			"resolved": "https://registry.npmjs.org/d3-shape/-/d3-shape-3.2.0.tgz",
+			"integrity": "sha512-SaLBuwGm3MOViRq2ABk3eLoxwZELpH6zhl3FbAoJ7Vm1gofKx6El1Ib5z23NUEhF9AsGl7y+dzLe5Cw2AArGTA==",
+			"license": "ISC",
+			"dependencies": {
+				"d3-path": "^3.1.0"
+			},
+			"engines": {
+				"node": ">=12"
+			}
+		},
+		"node_modules/d3-time": {
+			"version": "3.1.0",
+			"resolved": "https://registry.npmjs.org/d3-time/-/d3-time-3.1.0.tgz",
+			"integrity": "sha512-VqKjzBLejbSMT4IgbmVgDjpkYrNWUYJnbCGo874u7MMKIWsILRX+OpX/gTk8MqjpT1A/c6HY2dCA77ZN0lkQ2Q==",
+			"license": "ISC",
+			"dependencies": {
+				"d3-array": "2 - 3"
+			},
+			"engines": {
+				"node": ">=12"
+			}
+		},
+		"node_modules/d3-time-format": {
+			"version": "4.1.0",
+			"resolved": "https://registry.npmjs.org/d3-time-format/-/d3-time-format-4.1.0.tgz",
+			"integrity": "sha512-dJxPBlzC7NugB2PDLwo9Q8JiTR3M3e4/XANkreKSUxF8vvXKqm1Yfq4Q5dl8budlunRVlUUaDUgFt7eA8D6NLg==",
+			"license": "ISC",
+			"dependencies": {
+				"d3-time": "1 - 3"
+			},
+			"engines": {
+				"node": ">=12"
+			}
+		},
+		"node_modules/d3-timer": {
+			"version": "3.0.1",
+			"resolved": "https://registry.npmjs.org/d3-timer/-/d3-timer-3.0.1.tgz",
+			"integrity": "sha512-ndfJ/JxxMd3nw31uyKoY2naivF+r29V+Lc0svZxe1JvvIRmi8hUsrMvdOwgS1o6uBHmiz91geQ0ylPP0aj1VUA==",
+			"license": "ISC",
+			"engines": {
+				"node": ">=12"
+			}
+		},
 		"node_modules/debug": {
 			"version": "4.4.3",
 			"resolved": "https://registry.npmjs.org/debug/-/debug-4.4.3.tgz",
@@ -4979,6 +5164,12 @@
 				}
 			}
 		},
+		"node_modules/decimal.js-light": {
+			"version": "2.5.1",
+			"resolved": "https://registry.npmjs.org/decimal.js-light/-/decimal.js-light-2.5.1.tgz",
+			"integrity": "sha512-qIMFpTMZmny+MMIitAB6D7iVPEorVw6YQRWkvarTkT4tBeSLLiHzcwj6q0MmYSFCiVpiqPJTJEYIrpcPzVEIvg==",
+			"license": "MIT"
+		},
 		"node_modules/decode-named-character-reference": {
 			"version": "1.3.0",
 			"resolved": "https://registry.npmjs.org/decode-named-character-reference/-/decode-named-character-reference-1.3.0.tgz",
@@ -5025,6 +5216,16 @@
 				"url": "https://github.com/sponsors/wooorm"
 			}
 		},
+		"node_modules/dom-helpers": {
+			"version": "5.2.1",
+			"resolved": "https://registry.npmjs.org/dom-helpers/-/dom-helpers-5.2.1.tgz",
+			"integrity": "sha512-nRCa7CK3VTrM2NmGkIy4cbK7IZlgBE/PYMn55rrXefr5xXDP0LdtfPnblFDoVdcAfslJ7or6iqAUnx0CCGIWQA==",
+			"license": "MIT",
+			"dependencies": {
+				"@babel/runtime": "^7.8.7",
+				"csstype": "^3.0.2"
+			}
+		},
 		"node_modules/enhanced-resolve": {
 			"version": "5.21.0",
 			"resolved": "https://registry.npmjs.org/enhanced-resolve/-/enhanced-resolve-5.21.0.tgz",
@@ -5058,6 +5259,12 @@
 				"url": "https://opencollective.com/unified"
 			}
 		},
+		"node_modules/eventemitter3": {
+			"version": "4.0.7",
+			"resolved": "https://registry.npmjs.org/eventemitter3/-/eventemitter3-4.0.7.tgz",
+			"integrity": "sha512-8guHBZCwKnFhYdHr2ysuRWErTwhoN2X8XELRlrRwpmfeY2jjuUN4taQMsULKUVo1K4DvZl+0pgfyoysHxvmvEw==",
+			"license": "MIT"
+		},
 		"node_modules/eventsource-parser": {
 			"version": "3.0.8",
 			"resolved": "https://registry.npmjs.org/eventsource-parser/-/eventsource-parser-3.0.8.tgz",
@@ -5071,6 +5278,15 @@
 			"resolved": "https://registry.npmjs.org/extend/-/extend-3.0.2.tgz",
 			"integrity": "sha512-fjquC59cD7CyW6urNXK0FBufkZcoiGG80wTuPujX590cB5Ttln20E2UB4S/WARVqhXffZl2LNgS+gQdPIIim/g=="
 		},
+		"node_modules/fast-equals": {
+			"version": "5.4.0",
+			"resolved": "https://registry.npmjs.org/fast-equals/-/fast-equals-5.4.0.tgz",
+			"integrity": "sha512-jt2DW/aNFNwke7AUd+Z+e6pz39KO5rzdbbFCg2sGafS4mk13MI7Z8O5z9cADNn5lhGODIgLwug6TZO2ctf7kcw==",
+			"license": "MIT",
+			"engines": {
+				"node": ">=6.0.0"
+			}
+		},
 		"node_modules/get-nonce": {
 			"version": "1.0.1",
 			"resolved": "https://registry.npmjs.org/get-nonce/-/get-nonce-1.0.1.tgz",
@@ -5137,6 +5353,15 @@
 			"resolved": "https://registry.npmjs.org/inline-style-parser/-/inline-style-parser-0.2.7.tgz",
 			"integrity": "sha512-Nb2ctOyNR8DqQoR0OwRG95uNWIC0C1lCgf5Naz5H6Ji72KZ8OcFZLz2P5sNgwlyoJ8Yif11oMuYs5pBQa86csA=="
 		},
+		"node_modules/internmap": {
+			"version": "2.0.3",
+			"resolved": "https://registry.npmjs.org/internmap/-/internmap-2.0.3.tgz",
+			"integrity": "sha512-5Hh7Y1wQbvY5ooGgPbDaL5iYLAPzMTUrjMulskHLH6wnv/A+1q5rgEaiuqEjB+oxGXIVZs1FF+R/KPN3ZSQYYg==",
+			"license": "ISC",
+			"engines": {
+				"node": ">=12"
+			}
+		},
 		"node_modules/is-alphabetical": {
 			"version": "2.0.1",
 			"resolved": "https://registry.npmjs.org/is-alphabetical/-/is-alphabetical-2.0.1.tgz",
@@ -5197,6 +5422,12 @@
 				"jiti": "lib/jiti-cli.mjs"
 			}
 		},
+		"node_modules/js-tokens": {
+			"version": "4.0.0",
+			"resolved": "https://registry.npmjs.org/js-tokens/-/js-tokens-4.0.0.tgz",
+			"integrity": "sha512-RdJUflcE3cUzKiMqQgsCu06FPu9UdIJO0beYbPhHN4k6apgJtifcoCtT9bcxOpYBtpD2kCM6Sbzg4CausW/PKQ==",
+			"license": "MIT"
+		},
 		"node_modules/json-schema": {
 			"version": "0.4.0",
 			"resolved": "https://registry.npmjs.org/json-schema/-/json-schema-0.4.0.tgz",
@@ -5451,6 +5682,12 @@
 				"url": "https://opencollective.com/parcel"
 			}
 		},
+		"node_modules/lodash": {
+			"version": "4.18.1",
+			"resolved": "https://registry.npmjs.org/lodash/-/lodash-4.18.1.tgz",
+			"integrity": "sha512-dMInicTPVE8d1e5otfwmmjlxkZoUpiVLwyeTdUsi/Caj/gfzzblBcCE5sRHV/AsjuCmxWrte2TNGSYuCeCq+0Q==",
+			"license": "MIT"
+		},
 		"node_modules/longest-streak": {
 			"version": "3.1.0",
 			"resolved": "https://registry.npmjs.org/longest-streak/-/longest-streak-3.1.0.tgz",
@@ -5460,6 +5697,18 @@
 				"url": "https://github.com/sponsors/wooorm"
 			}
 		},
+		"node_modules/loose-envify": {
+			"version": "1.4.0",
+			"resolved": "https://registry.npmjs.org/loose-envify/-/loose-envify-1.4.0.tgz",
+			"integrity": "sha512-lyuxPGr/Wfhrlem2CL/UcnUc1zcqKAImBDzukY7Y5F/yQiNdko6+fRLevlw1HgMySw7f611UIY408EtxRSoK3Q==",
+			"license": "MIT",
+			"dependencies": {
+				"js-tokens": "^3.0.0 || ^4.0.0"
+			},
+			"bin": {
+				"loose-envify": "cli.js"
+			}
+		},
 		"node_modules/lucide-react": {
 			"version": "1.14.0",
 			"resolved": "https://registry.npmjs.org/lucide-react/-/lucide-react-1.14.0.tgz",
@@ -6394,6 +6643,15 @@
 				"node": "^10 || ^12 || >=14"
 			}
 		},
+		"node_modules/object-assign": {
+			"version": "4.1.1",
+			"resolved": "https://registry.npmjs.org/object-assign/-/object-assign-4.1.1.tgz",
+			"integrity": "sha512-rJgTQnkUnH1sFw8yT6VSU3zD3sWmu6sZhIseY8VX+GRu3P6F7Fu+JNDoXfklElbLJSnc3FUQHVe4cU5hj+BcUg==",
+			"license": "MIT",
+			"engines": {
+				"node": ">=0.10.0"
+			}
+		},
 		"node_modules/parse-entities": {
 			"version": "4.0.2",
 			"resolved": "https://registry.npmjs.org/parse-entities/-/parse-entities-4.0.2.tgz",
@@ -6482,6 +6740,23 @@
 				"node": "^10 || ^12 || ^13.7 || ^14 || >=15.0.1"
 			}
 		},
+		"node_modules/prop-types": {
+			"version": "15.8.1",
+			"resolved": "https://registry.npmjs.org/prop-types/-/prop-types-15.8.1.tgz",
+			"integrity": "sha512-oj87CgZICdulUohogVAR7AjlC0327U4el4L6eAvOqCeudMDVU0NThNaV+b9Df4dXgSP1gXMTnPdhfe/2qDH5cg==",
+			"license": "MIT",
+			"dependencies": {
+				"loose-envify": "^1.4.0",
+				"object-assign": "^4.1.1",
+				"react-is": "^16.13.1"
+			}
+		},
+		"node_modules/prop-types/node_modules/react-is": {
+			"version": "16.13.1",
+			"resolved": "https://registry.npmjs.org/react-is/-/react-is-16.13.1.tgz",
+			"integrity": "sha512-24e6ynE2H+OKt4kqsOvNd8kBpV65zoxbA4BVsEOB3ARVWQki/DHzaUoC5KuON/BiccDaCCTZBuOcfZs70kR8bQ==",
+			"license": "MIT"
+		},
 		"node_modules/property-information": {
 			"version": "7.1.0",
 			"resolved": "https://registry.npmjs.org/property-information/-/property-information-7.1.0.tgz",
@@ -6687,6 +6962,12 @@
 				"react": "^19.2.5"
 			}
 		},
+		"node_modules/react-is": {
+			"version": "18.3.1",
+			"resolved": "https://registry.npmjs.org/react-is/-/react-is-18.3.1.tgz",
+			"integrity": "sha512-/LLMVyas0ljjAtoYiPqYiL8VWXzUUdThrmU5+n20DZv+a+ClRoevUzw5JxU+Ieh5/c87ytoTBV9G1FiKfNJdmg==",
+			"license": "MIT"
+		},
 		"node_modules/react-markdown": {
 			"version": "10.1.0",
 			"resolved": "https://registry.npmjs.org/react-markdown/-/react-markdown-10.1.0.tgz",
@@ -6758,6 +7039,21 @@
 				}
 			}
 		},
+		"node_modules/react-smooth": {
+			"version": "4.0.4",
+			"resolved": "https://registry.npmjs.org/react-smooth/-/react-smooth-4.0.4.tgz",
+			"integrity": "sha512-gnGKTpYwqL0Iii09gHobNolvX4Kiq4PKx6eWBCYYix+8cdw+cGo3do906l1NBPKkSWx1DghC1dlWG9L2uGd61Q==",
+			"license": "MIT",
+			"dependencies": {
+				"fast-equals": "^5.0.1",
+				"prop-types": "^15.8.1",
+				"react-transition-group": "^4.4.5"
+			},
+			"peerDependencies": {
+				"react": "^16.8.0 || ^17.0.0 || ^18.0.0 || ^19.0.0",
+				"react-dom": "^16.8.0 || ^17.0.0 || ^18.0.0 || ^19.0.0"
+			}
+		},
 		"node_modules/react-style-singleton": {
 			"version": "2.2.3",
 			"resolved": "https://registry.npmjs.org/react-style-singleton/-/react-style-singleton-2.2.3.tgz",
@@ -6795,6 +7091,54 @@
 				"react": "^16.8.0 || ^17.0.0 || ^18.0.0 || ^19.0.0"
 			}
 		},
+		"node_modules/react-transition-group": {
+			"version": "4.4.5",
+			"resolved": "https://registry.npmjs.org/react-transition-group/-/react-transition-group-4.4.5.tgz",
+			"integrity": "sha512-pZcd1MCJoiKiBR2NRxeCRg13uCXbydPnmB4EOeRrY7480qNWO8IIgQG6zlDkm6uRMsURXPuKq0GWtiM59a5Q6g==",
+			"license": "BSD-3-Clause",
+			"dependencies": {
+				"@babel/runtime": "^7.5.5",
+				"dom-helpers": "^5.0.1",
+				"loose-envify": "^1.4.0",
+				"prop-types": "^15.6.2"
+			},
+			"peerDependencies": {
+				"react": ">=16.6.0",
+				"react-dom": ">=16.6.0"
+			}
+		},
+		"node_modules/recharts": {
+			"version": "2.15.4",
+			"resolved": "https://registry.npmjs.org/recharts/-/recharts-2.15.4.tgz",
+			"integrity": "sha512-UT/q6fwS3c1dHbXv2uFgYJ9BMFHu3fwnd7AYZaEQhXuYQ4hgsxLvsUXzGdKeZrW5xopzDCvuA2N41WJ88I7zIw==",
+			"license": "MIT",
+			"dependencies": {
+				"clsx": "^2.0.0",
+				"eventemitter3": "^4.0.1",
+				"lodash": "^4.17.21",
+				"react-is": "^18.3.1",
+				"react-smooth": "^4.0.4",
+				"recharts-scale": "^0.4.4",
+				"tiny-invariant": "^1.3.1",
+				"victory-vendor": "^36.6.8"
+			},
+			"engines": {
+				"node": ">=14"
+			},
+			"peerDependencies": {
+				"react": "^16.0.0 || ^17.0.0 || ^18.0.0 || ^19.0.0",
+				"react-dom": "^16.0.0 || ^17.0.0 || ^18.0.0 || ^19.0.0"
+			}
+		},
+		"node_modules/recharts-scale": {
+			"version": "0.4.5",
+			"resolved": "https://registry.npmjs.org/recharts-scale/-/recharts-scale-0.4.5.tgz",
+			"integrity": "sha512-kivNFO+0OcUNu7jQquLXAxz1FIwZj8nrj+YkOKc5694NbjCvcT6aSZiIzNzd2Kul4o4rTto8QVR9lMNtxD4G1w==",
+			"license": "MIT",
+			"dependencies": {
+				"decimal.js-light": "^2.4.1"
+			}
+		},
 		"node_modules/remark-gfm": {
 			"version": "4.0.1",
 			"resolved": "https://registry.npmjs.org/remark-gfm/-/remark-gfm-4.0.1.tgz",
@@ -7051,6 +7395,12 @@
 				"url": "https://github.com/sponsors/sindresorhus"
 			}
 		},
+		"node_modules/tiny-invariant": {
+			"version": "1.3.3",
+			"resolved": "https://registry.npmjs.org/tiny-invariant/-/tiny-invariant-1.3.3.tgz",
+			"integrity": "sha512-+FbBPE1o9QAYvviau/qC5SE3caw21q3xkvWKBtja5vgqOWIHHJ3ioaq1VPfn/Szqctz2bU/oYeKd9/z5BL+PVg==",
+			"license": "MIT"
+		},
 		"node_modules/trim-lines": {
 			"version": "3.0.1",
 			"resolved": "https://registry.npmjs.org/trim-lines/-/trim-lines-3.0.1.tgz",
@@ -7322,6 +7672,28 @@
 				"url": "https://opencollective.com/unified"
 			}
 		},
+		"node_modules/victory-vendor": {
+			"version": "36.9.2",
+			"resolved": "https://registry.npmjs.org/victory-vendor/-/victory-vendor-36.9.2.tgz",
+			"integrity": "sha512-PnpQQMuxlwYdocC8fIJqVXvkeViHYzotI+NJrCuav0ZYFoq912ZHBk3mCeuj+5/VpodOjPe1z0Fk2ihgzlXqjQ==",
+			"license": "MIT AND ISC",
+			"dependencies": {
+				"@types/d3-array": "^3.0.3",
+				"@types/d3-ease": "^3.0.0",
+				"@types/d3-interpolate": "^3.0.1",
+				"@types/d3-scale": "^4.0.2",
+				"@types/d3-shape": "^3.1.0",
+				"@types/d3-time": "^3.0.0",
+				"@types/d3-timer": "^3.0.0",
+				"d3-array": "^3.1.6",
+				"d3-ease": "^3.0.1",
+				"d3-interpolate": "^3.0.1",
+				"d3-scale": "^4.0.2",
+				"d3-shape": "^3.1.0",
+				"d3-time": "^3.0.0",
+				"d3-timer": "^3.0.1"
+			}
+		},
 		"node_modules/zod": {
 			"version": "4.4.3",
 			"resolved": "https://registry.npmjs.org/zod/-/zod-4.4.3.tgz",

frontend/app/package.json

@@ -31,6 +31,7 @@
 		"react": "^19.2.5",
 		"react-dom": "^19.2.5",
 		"react-markdown": "^10.1.0",
+		"recharts": "^2.15.4",
 		"remark-gfm": "^4.0.1",
 		"tailwind-merge": "^3.5.0",
 		"tw-animate-css": "^1.4.0",

src/agent_prompting.py

@@ -162,8 +162,8 @@ def get_doing_tasks_section() -> str:
         '不要为了单次操作创建 helper 或抽象。优先使用能完整解决问题的最简单实现。',
         '除非确实需要新文件，否则优先编辑现有文件。',
         '对于文件解析、表格转换、JSON 或 JSONL 处理、日志分析、批量校验、数据抽样等任务，如果小型 Python 脚本比手工文本处理更可靠，可以编写小型 Python 脚本。',
-        '小脚本应短小、可读，并明确输入和输出。需要执行 Python 代码或 Python 脚本时，使用 python_exec；不要通过 bash 执行 python、python3 或 .venv/bin/python。',
+        '小脚本应短小、可读，并明确输入和输出。需要执行 Python 时，把代码写到 session/scratchpad 下的 .py 文件，再用 bash 跑 `python3 -u <script> 2>&1 | tee <log>`；`-u` 关 stdout 缓冲、`tee` 让进度行实时落盘，超时被 kill 时仍能 tail 日志看到死在哪。不要把多行 Python 代码以 `-c` 字符串方式塞进 bash command。',
-        '一次性 Python 分析优先直接传给 python_exec.code，不要为了临时分析创建项目文件。',
+        '一次性 Python 分析也走脚本文件 + bash，不要为了临时分析在项目根目录创建脚本，scratchpad 是默认临时区。',
         '如果确实需要临时脚本、缓存或中间产物，必须写入当前 session/scratchpad；交付产物必须优先写入当前 session/output。',
         '只有用户明确要求长期复用或该脚本属于产品代码时，才把 Python 脚本加入项目源码目录。',
         '当事情失败时，先诊断原因再改变方向。不要对同一个失败动作反复循环。',

src/agent_runtime.py

@@ -1620,9 +1620,9 @@ class LocalCodingAgent:
                         '请回复“继续”，我会要求模型重新生成合法 JSON 参数；'
                         '如果是 write_file 写短文本/Markdown，应改用 content_lines；'
                         '如果是很小的 JSON/JSONL/CSV，可用 json_content、jsonl_records 或 csv_rows；'
-                        '如果是脚本、长文本、大 JSON、JSONL、CSV 或生成数据，应直接改用 python_exec 写文件；'
+                        '如果是脚本、长文本、大 JSON、JSONL、CSV 或生成数据，应改用 bash 配合 quoted heredoc 写文件，'
-                        '如果是生成/转换结构化数据，应优先调用对应 skill 脚本或 python_exec，'
+                        '或先 write_file 落地一个小 .py 再用 bash 跑；'
-                        '不要把大段内容直接塞进工具参数；bash 仅作为最后兜底。'
+                        '不要把大段内容直接塞进工具参数。'
                     ) from exc
                 if not isinstance(arguments, dict):
                     raise OpenAICompatError(

src/agent_tool_specs/execution.py

@@ -10,51 +10,6 @@ def build_execution_tools(handlers: Mapping[str, ToolHandler]) -> list[AgentTool
     """构建本地命令、Python 执行与短等待工具声明。"""
 
     return [
-        AgentTool(
-            name='python_exec',
-            description=(
-                '优先用本工具执行小型 Python 代码或项目内 Python 脚本，用于结构化文件分析、'
-                'JSON/JSONL 处理、批量校验、数据抽样和快速计算。默认使用当前用户独立 Python venv；'
-                '不要用 bash 运行 python/python3/.venv/bin/python。不要用本工具安装依赖。'
-                '缺包时应向用户确认后再处理依赖。一次性分析优先传 code；项目或 skill 内已有脚本优先传 script_path，'
-                '不要在 code 里用 subprocess 二次调用 Python 脚本。如需写临时文件，'
-                '必须写入环境变量 PYTHON_EXEC_SCRATCHPAD 指向的会话隔离目录，'
-                '不要在项目根目录创建临时 .py 脚本。'
-            ),
-            parameters={
-                'type': 'object',
-                'properties': {
-                    'code': {
-                        'type': 'string',
-                        'description': '要通过 python -c 执行的 Python 代码。code 和 script_path 必须二选一；一次性分析优先使用 code。',
-                    },
-                    'script_path': {
-                        'type': 'string',
-                        'description': '工作区内已有、需要长期复用的 Python 脚本路径。code 和 script_path 必须二选一；不要为一次性分析在项目根目录新建脚本。',
-                    },
-                    'args': {
-                        'type': 'array',
-                        'items': {'type': 'string'},
-                        'description': '传给脚本的命令行参数。仅在 script_path 模式下使用。',
-                    },
-                    'stdin': {
-                        'type': 'string',
-                        'description': '可选标准输入内容。',
-                    },
-                    'timeout_seconds': {
-                        'type': 'number',
-                        'minimum': 1,
-                        'description': '可选超时时间，默认 30 秒。读大 CSV/JSONL、pandas 分析、批量校验、训练评测脚本等可能 >30s 的任务必须显式传值（常规 120-300、训练/长跑 600+）。',
-                    },
-                    'max_output_chars': {
-                        'type': 'integer',
-                        'minimum': 100,
-                        'description': '可选输出截断长度，默认使用当前会话输出限制。',
-                    },
-                },
-            },
-            handler=resolve_handler(handlers, 'python_exec', 'execution'),
-        ),
         AgentTool(
             name='python_package',
             description=(
@@ -94,10 +49,13 @@ def build_execution_tools(handlers: Mapping[str, ToolHandler]) -> list[AgentTool
         AgentTool(
             name='bash',
             description=(
-                '运行真实 shell 命令，例如系统命令、进程控制、git 只读检查或用户已确认的依赖安装。'
+                '运行真实 shell 命令，包括系统命令、进程控制、git 只读检查、用户已确认的依赖安装，'
-                '不要用 bash 执行 python、python3、pip 或 .venv/bin/python；结构化文件分析、'
+                '以及所有 Python 脚本执行。'
-                'JSON/JSONL 处理、批量校验、数据抽样和快速计算必须优先使用 python_exec。'
+                '执行 Python 时，把代码先写到 scratchpad 下的 .py 文件，'
-                'Python 包安装必须优先使用 python_package。'
+                '再用本工具跑 `python3 -u <script> 2>&1 | tee <log>`；'
+                '`-u` 关掉 stdout 缓冲，`tee` 让进度行实时落盘，超时被 kill 时仍能 `tail` 日志看到死在哪个 phase。'
+                '不要把多行 Python 代码以 `-c` 字符串方式塞进 command。'
+                'Python 包安装必须优先使用 python_package；不要通过 bash 执行 pip。'
                 '需要等待远端长任务（>30s 的训练、文件落盘、外部 API 回调）时，'
                 '设置 run_in_background=true 让命令在后台 detach 跑、立刻返回 task_id；'
                 '默认 wait_for_completion=true 时进程结束后会自动起新一轮把结果送回。'

src/agent_tool_specs/files.py

@@ -41,8 +41,8 @@ def build_file_tools(handlers: Mapping[str, ToolHandler]) -> list[AgentTool]:
             description=(
                 'Write or append a SMALL UTF-8 file inside the workspace. Creates parent directories when needed. '
                 'Use this for short notes, lightweight markdown, and small config files. '
-                'For Python scripts, JSONL, large JSON/CSV, generated datasets, or quote-heavy/multi-line content, prefer python_exec '
+                'For Python scripts, JSONL, large JSON/CSV, generated datasets, or quote-heavy/multi-line content, prefer bash with a quoted heredoc '
-                'to write the file with pathlib/json/csv; do not pack large content into tool arguments.'
+                "(e.g. `cat > scratchpad/foo.py <<'EOF' ... EOF`); do not pack large content into tool arguments."
             ),
             parameters={
                 'type': 'object',
@@ -51,7 +51,7 @@ def build_file_tools(handlers: Mapping[str, ToolHandler]) -> list[AgentTool]:
                     'content_lines': {
                         'type': 'array',
                         'items': {'type': 'string'},
-                        'description': 'Lines to join with "\\n"; use only for short text/markdown or small scripts. For long scripts use python_exec to create the file.',
+                        'description': 'Lines to join with "\\n"; use only for short text/markdown or small scripts. For long scripts use bash with a quoted heredoc to create the file.',
                     },
                     'content': {
                         'type': 'string',
@@ -63,12 +63,12 @@ def build_file_tools(handlers: Mapping[str, ToolHandler]) -> list[AgentTool]:
                     },
                     'json_content': {
                         'type': 'object',
-                        'description': 'Small JSON value to serialize with ensure_ascii=false and indent=2. For large JSON use python_exec streaming write.',
+                        'description': 'Small JSON value to serialize with ensure_ascii=false and indent=2. For large JSON write a small Python script to scratchpad and run it via bash.',
                     },
                     'jsonl_records': {
                         'type': 'array',
                         'items': {'type': 'object'},
-                        'description': 'Small list of objects to serialize as compact JSONL. For more than a few dozen records or generated datasets use python_exec.',
+                        'description': 'Small list of objects to serialize as compact JSONL. For more than a few dozen records or generated datasets, write a Python script to scratchpad and run it via bash.',
                     },
                     'csv_headers': {
                         'type': 'array',

src/agent_tools.py

@@ -117,7 +117,6 @@ def default_tool_registry() -> dict[str, AgentTool]:
             'grep_search': _grep_search,
         }),
         *build_execution_tools({
-            'python_exec': _run_python_exec,
             'python_package': _run_python_package,
             'bash': _run_bash,
             'bash_status': _run_bash_status,