fix: session resume, config_set schema, metrics chart, subagent logging

- Fix session resume: remove messages.length>1 guard so sessionStorage ID is used as resumeSessionId on first message (prevents new session creation when user sends "继续") - Fix config_set tool schema: add missing `items: {}` to array type in oneOf (Azure OpenAI strict validation rejects it otherwise) - Metrics chart: use actual target set filename from trigger text instead of generic "target_subset" key - Add sub-agent skill call logging to agent_runtime.py (full prompt/response, no truncation) - Various metric enrichment and dedup fixes in server.py Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-05-27 17:02:24 +08:00
parent 89e77140e7
commit 85ce8c0e69
8 changed files with 411 additions and 40 deletions
@@ -1751,11 +1751,17 @@ def create_app(state: AgentState) -> FastAPI:
        if isinstance(run_id, str) and run_id.strip():
            rid = run_id.strip()
            per_round = _resolve_run_dic_per_round(state_entries)
-            # augment 产物的 runDic = resolve_run_ids.sh 在当轮 eval 落盘后取到的
-            # max(已落盘 workflow) = 当轮 eval 自身的 workflow id，即 per_round[R{n}]
-            # （不是 R{n-1}，脚本取的是 eval 之后的 max，此时 max 已经包含当轮 eval）
+            # 正确标号: R{n}·Train 和 R{n}·Analysis 共用同一 run_id，
+            # per_round[R{n}] 从 Analysis 的 watch/log 中解析到 runDic。
+            # 兼容旧 session 错误标号: agent 每阶段递增 run_id（R7=train, R6=analysis），
+            # 导致 per_round[R_train] = None，需 fallback 到 R{n-1}。
            if step == 'augment':
-                pass  # 直接走下面 per_round.get(rid) 取本轮 runDic
+                run_dic = per_round.get(rid)
+                if run_dic is None:
+                    run_index = _parse_run_index(rid)
+                    if run_index is not None and run_index >= 1:
+                        prev_rid = f'R{run_index - 1}'
+                        run_dic = per_round.get(prev_rid)
            if run_dic is None:
                run_dic = per_round.get(rid)
        if run_dic is None:
@@ -2001,7 +2007,14 @@ def create_app(state: AgentState) -> FastAPI:
    ) -> Response:
        sessions_dir = state.account_paths(account_id)['sessions']
        state_path = _session_state_path(sessions_dir, session_id)
+        await asyncio.to_thread(
+            _sync_remote_program_state, state, account_id, session_id, state_path,
+        )
        state_entries, _ = _read_program_state(state_path)
+        trigger = _extract_trigger_from_state(state_entries) or _extract_trigger_from_session(
+            sessions_dir, session_id
+        )
+        _target_set_name = _parse_target_set(trigger, model_config=state.model_config_for(account_id))
        iter_count = await asyncio.to_thread(
            _read_iteration_log_count,
            session_id,
@@ -2013,6 +2026,7 @@ def create_app(state: AgentState) -> FastAPI:
            session_id,
            agent_state=state,
            account_id=account_id,
+            target_set_name=_target_set_name,
        )
        payload = await asyncio.to_thread(
            _hardcoded_autoresearch_pipeline,
@@ -2060,6 +2074,12 @@ def create_app(state: AgentState) -> FastAPI:
        async def event_gen():
            last_signature: str | None = None
            last_heartbeat = time.monotonic()
+            last_remote_sync = 0.0
+            _state_entries_init, _ = _read_program_state(state_path)
+            _trigger_init = _extract_trigger_from_state(_state_entries_init) or _extract_trigger_from_session(
+                sessions_dir, session_id
+            )
+            _stream_target_set = _parse_target_set(_trigger_init, model_config=state.model_config_for(account_id))
            refresher = asyncio.create_task(
                _iter_count_refresh_loop(
                    session_id,
@@ -2072,12 +2092,19 @@ def create_app(state: AgentState) -> FastAPI:
                    session_id,
                    agent_state=state,
                    account_id=account_id,
+                    target_set_name=_stream_target_set,
                )
            )
            try:
                while True:
                    if await request.is_disconnected():
                        break
+                    now_mono = time.monotonic()
+                    if now_mono - last_remote_sync >= 5.0:
+                        await asyncio.to_thread(
+                            _sync_remote_program_state, state, account_id, session_id, state_path,
+                        )
+                        last_remote_sync = now_mono
                    state_entries, state_mtime = _read_program_state(state_path)
                    iter_count = await asyncio.to_thread(
                        _read_iteration_log_count,
@@ -2090,6 +2117,7 @@ def create_app(state: AgentState) -> FastAPI:
                        session_id,
                        agent_state=state,
                        account_id=account_id,
+                        target_set_name=_stream_target_set,
                    )
                    payload = await asyncio.to_thread(
                        _hardcoded_autoresearch_pipeline,
@@ -5714,11 +5742,43 @@ def _sync_remote_program_state(
    new_content = result.stdout or ''
    if not new_content.strip():
        return
-    # Skip rewrite if local already matches remote (avoid touching mtime, which
+    # Preserve local-only synthetic gate entries that remote doesn't have.
+    local_synthetics: list[str] = []
+    try:
+        if local_state_path.is_file():
+            # Parse remote gate keys once
+            remote_gate_keys: set[str] = set()
+            for raw_line in new_content.splitlines():
+                if not raw_line.strip():
+                    continue
+                try:
+                    r = json.loads(raw_line)
+                    if r.get('step') and not r.get('_synthetic'):
+                        remote_gate_keys.add(f"{r.get('run_id','')}:{r.get('step','')}")
+                except (json.JSONDecodeError, ValueError):
+                    pass
+            for line in local_state_path.read_text(encoding='utf-8').splitlines():
+                if '"_synthetic"' not in line:
+                    continue
+                try:
+                    obj = json.loads(line)
+                    if obj.get('_synthetic'):
+                        gate_key = f"{obj.get('run_id','')}:{obj.get('step','')}"
+                        if gate_key not in remote_gate_keys:
+                            local_synthetics.append(line)
+                except (json.JSONDecodeError, AttributeError):
+                    pass
+    except OSError:
+        pass
+    merged = new_content.rstrip('\n')
+    if local_synthetics:
+        merged += '\n' + '\n'.join(local_synthetics)
+    merged += '\n'
+    # Skip rewrite if local already matches (avoid touching mtime, which
    # SSE uses to detect changes).
    try:
        if local_state_path.is_file() and (
-            local_state_path.read_text(encoding='utf-8') == new_content
+            local_state_path.read_text(encoding='utf-8') == merged
        ):
            return
    except OSError:
@@ -5726,7 +5786,7 @@ def _sync_remote_program_state(
    try:
        local_state_path.parent.mkdir(parents=True, exist_ok=True)
        tmp = local_state_path.with_suffix('.jsonl.sync-tmp')
-        tmp.write_text(new_content, encoding='utf-8')
+        tmp.write_text(merged, encoding='utf-8')
        tmp.replace(local_state_path)
    except OSError:
        return
@@ -6055,7 +6115,7 @@ _STEP_KIND: dict[str, str] = {
    'dist-analysis': 'analysis',
    'hypothesis':    'analysis',
    'log':           'analysis',
-    'augment':       'train',
+    'augment':       'analysis',
    'verify':        'train',
    'sft':           'train',
    'human-check':   'gate',
@@ -6083,8 +6143,10 @@ _CARD_META: dict[tuple[str, str], dict[str, Any]] = {
    ('dist-analysis', 'baseline'): {'icon': 'trending-up',    'title': '分层结果分析 & 报告', 'subtitle': '根因归类 + 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'},
+    # augment — lives under previous round's analysis/baseline
+    ('augment', 'baseline'):       {'icon': 'flask',         'title': '数据增强',         'subtitle': 'augment_<runDic>.jsonl'},
+    ('augment', 'analysis'):       {'icon': 'flask',         'title': '数据增强',         'subtitle': 'augment_<runDic>.jsonl'},
    # R{n}·Train
-    ('augment', 'train'):    {'icon': 'flask',         'title': '数据增强',         'subtitle': 'augment_<runDic>.jsonl'},
    ('verify', 'train'):     {'icon': 'shield',        'title': '修改返回验证',     'subtitle': 'sanity + 人审'},
    ('sft', 'train'):        {'icon': 'graduation-cap','title': 'SFT 训练',        'subtitle': 'submit_sft_via_cml.sh'},
    # R{n}·Analysis
@@ -6103,16 +6165,17 @@ _CARD_ORDER: dict[tuple[str, str], int] = {
    ('dist-analysis', 'baseline'): 2,
    ('hypothesis', 'baseline'): 3,
    ('log', 'baseline'): 4,
+    ('augment', 'baseline'): 5,
    # train
-    ('augment', 'train'): 0,
-    ('verify', 'train'): 1,
-    ('sft', 'train'): 2,
+    ('verify', 'train'): 0,
+    ('sft', 'train'): 1,
    # analysis
    ('cml', 'analysis'): 0,
    ('gold-drift', 'analysis'): 1,
    ('dist-analysis', 'analysis'): 2,
    ('hypothesis', 'analysis'): 3,
    ('log', 'analysis'): 4,
+    ('augment', 'analysis'): 5,
 }

 _SECTION_LABELS: dict[str, str] = {
@@ -6253,9 +6316,9 @@ def _derive_section_status(card_statuses: list[str]) -> str:
        return 'pending'
    if all(s == 'complete' for s in card_statuses):
        return 'complete'
-    if 'failed' in card_statuses and not any(s in ('running', 'complete') for s in card_statuses):
+    if 'failed' in card_statuses and not any(s in ('running', 'complete', 'waiting') for s in card_statuses):
        return 'failed'
-    if 'running' in card_statuses or 'complete' in card_statuses:
+    if any(s in ('running', 'complete', 'waiting') for s in card_statuses):
        return 'running'
    return 'pending'

@@ -6315,6 +6378,11 @@ def _build_pipeline_items(
        if not run_id:
            continue
        last_non_gate_run_id = run_id
+        # augment belongs to the previous round's analysis/baseline section
+        if step == 'augment':
+            aug_index = _parse_run_index(run_id)
+            if aug_index is not None and aug_index >= 1:
+                run_id = f'R{aug_index - 1}'
        run_index = _parse_run_index(run_id)
        if run_index is None:
            continue
@@ -6375,6 +6443,11 @@ def _build_pipeline_items(
                card['progress'] = max(0.0, min(1.0, value))
            except (TypeError, ValueError):
                pass
+        if step == 'augment' and 'count' in entry:
+            try:
+                card['subtitle'] = f'{int(entry["count"])} 条新增样本'
+            except (TypeError, ValueError):
+                pass
        sections[section_key]['cards'].append(card)

    # 3. Sort cards within each section.
@@ -6451,7 +6524,7 @@ def _build_pipeline_items(
            'summary': entry.get('summary'),
            'proposal': entry.get('proposal'),
            'ask': entry.get('ask'),
-            '_order_key': (run_index, 99, gate_position_index.get(gate_key, 0)),
+            '_order_key': entry.get('ts', ''),
        })

    # 6. Filter: drop empty/pending sections (no cards or all pending). Running
@@ -6589,12 +6662,18 @@ def _compute_in_flight(
                progress = max(0.0, min(1.0, value))
            except (TypeError, ValueError):
                pass
+        subtitle = meta['subtitle']
+        if step == 'augment' and 'count' in entry:
+            try:
+                subtitle = f'{int(entry["count"])} 条新增样本'
+            except (TypeError, ValueError):
+                pass
        cards.append({
            'key': f'{target_run_id}:{step}',
            'step': step,
            'icon': meta['icon'],
            'title': meta['title'],
-            'subtitle': meta['subtitle'],
+            'subtitle': subtitle,
            'status': status,
            'progress': progress,
        })
@@ -6679,7 +6758,7 @@ def _maybe_inject_gate_fallback(
        'status': 'running',
        'run_id': run_id or 'R0',
        'reason': last_text[-300:],
-        'ts': datetime_utc_iso(),
+        'ts': time.strftime('%Y-%m-%dT%H:%M:%S+08:00', time.localtime()),
        '_synthetic': True,
    }
    try:
@@ -7033,6 +7112,7 @@ def _read_iteration_log_metrics_history(
    *,
    agent_state: 'AgentState | None' = None,
    account_id: str | None = None,
+    target_set_name: 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)
@@ -7040,6 +7120,7 @@ def _read_iteration_log_metrics_history(
        return _METRICS_HISTORY_CACHE[cache_key]
    value = _read_iteration_log_metrics_history_uncached(
        session_id, agent_state=agent_state, account_id=account_id,
+        target_set_name=target_set_name,
    )
    _METRICS_HISTORY_CACHE[cache_key] = value
    return value
@@ -7050,6 +7131,7 @@ async def _metrics_history_refresh_loop(
    *,
    agent_state: 'AgentState | None',
    account_id: str | None,
+    target_set_name: str | None = None,
    interval_seconds: float = 3.0,
 ) -> None:
    cache_key = _iter_count_cache_key(session_id, account_id)
@@ -7060,6 +7142,7 @@ async def _metrics_history_refresh_loop(
                session_id,
                agent_state=agent_state,
                account_id=account_id,
+                target_set_name=target_set_name,
            )
            _METRICS_HISTORY_CACHE[cache_key] = value
        except Exception:  # noqa: BLE001
@@ -7072,6 +7155,7 @@ def _read_iteration_log_metrics_history_uncached(
    *,
    agent_state: 'AgentState | None' = None,
    account_id: str | None = None,
+    target_set_name: str | None = None,
 ) -> dict[str, Any] | None:
    """Parse iteration_log.jsonl into {'metrics': [keys], 'rounds': [...]}.

@@ -7164,8 +7248,70 @@ def _read_iteration_log_metrics_history_uncached(
            'values': values,
        })

+    # Deduplicate: keep only the last entry per (iteration, runDic).
+    # Agent may write multiple entries for the same round (e.g. re-running
+    # dist-analysis); the last one has the most complete metrics.
+    seen_keys: dict[tuple[int, int], int] = {}
+    for idx, r in enumerate(rounds):
+        key = (r['iteration'], r['runDic'])
+        seen_keys[key] = idx
+    rounds = [rounds[i] for i in sorted(seen_keys.values())]
+
+    # Rebuild metrics_order from surviving rounds only (dedup may have removed
+    # entries whose keys shouldn't appear in the chart).
+    metrics_order = []
+    metrics_seen = set()
+    for r in rounds:
+        for k in r['values']:
+            if k not in metrics_seen:
+                metrics_seen.add(k)
+                metrics_order.append(k)
+
    if not rounds:
        return None
+
+    # Enrich each round with KPI-equivalent metrics from lark_template.json.
+    # This ensures dapan_car / specific_test / target_subset are always present
+    # even if the agent omitted them from iteration_log.
+    target_key = target_set_name if target_set_name and target_set_name != '—' else 'target_subset'
+    _KPI_METRIC_MAP = {
+        'specific_test_pass_rate': 'specific_test',
+        'overall_car_pass_rate': 'dapan_car',
+        'target_set_pass_rate': target_key,
+    }
+    # Rename agent-written 'target_subset' to the actual target set name first,
+    # so enrichment doesn't create duplicates.
+    if target_key != 'target_subset':
+        if 'target_subset' in metrics_seen:
+            metrics_order = [target_key if m == 'target_subset' else m for m in metrics_order]
+            metrics_seen.discard('target_subset')
+            metrics_seen.add(target_key)
+        for r in rounds:
+            if 'target_subset' in r['values']:
+                r['values'][target_key] = r['values'].pop('target_subset')
+
+    for r in rounds:
+        run_dic = r.get('runDic')
+        if not isinstance(run_dic, int) or run_dic == 0:
+            continue
+        metric_diff = _read_workflow_metric_diff(
+            run_dic,
+            agent_state=agent_state,
+            account_id=account_id,
+            session_id=session_id,
+        )
+        if metric_diff is None:
+            continue
+        _enrich_metrics_from_eval_output(metric_diff, None)
+        for src_key, dst_key in _KPI_METRIC_MAP.items():
+            if dst_key not in r['values']:
+                raw = metric_diff.get(src_key)
+                if isinstance(raw, (int, float)):
+                    r['values'][dst_key] = round(raw * 100, 2)
+                    if dst_key not in metrics_seen:
+                        metrics_seen.add(dst_key)
+                        metrics_order.append(dst_key)
+
    return {'metrics': metrics_order, 'rounds': rounds}


@@ -7385,7 +7531,8 @@ def _parse_target_set(
 ) -> str | None:
    """Sync, non-blocking. Reads cache only — never makes the LLM call inline.
    Cache is filled by the watcher scanner background loop via
-    _refresh_target_set_cache_blocking (which is run in a thread)."""
+    _refresh_target_set_cache_blocking (which is run in a thread).
+    Falls back to regex extraction of .csv filenames from trigger text."""
    if not trigger:
        return None
    text = trigger.strip()
@@ -7394,7 +7541,14 @@ def _parse_target_set(
    if len(text) > 4000:
        text = text[:4000]
    cache_key = hashlib.sha256(text.encode('utf-8')).hexdigest()
-    return _TARGET_SET_LLM_CACHE.get(cache_key)
+    cached = _TARGET_SET_LLM_CACHE.get(cache_key)
+    if cached is not None:
+        return cached
+    # Regex fallback: extract .csv filename from trigger text.
+    m = re.search(r'(?:(?<=[的/，,\s])|(?<=^))([^\s，,。；的中是在]+\.csv)', text)
+    if m:
+        return m.group(1)
+    return None


 def _refresh_target_set_cache_blocking(