Merge branch 'wsh_dev' into 'main'

fix: pipeline排序/runDic映射/label-master repeatable/达标优先级

See merge request wuyang6/zk-data-agent!4

backend/api/server.py

@@ -1768,12 +1768,11 @@ 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 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
+            # 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）
             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}')
+                pass  # 直接走下面 per_round.get(rid) 取本轮 runDic
             if run_dic is None:
                 run_dic = per_round.get(rid)
         if run_dic is None:
@@ -6395,9 +6394,20 @@ def _build_pipeline_items(
         for c in sec['cards']:
             c.pop('_order', None)
 
-    # 4. Build round items in display order.
-    # Order: by run_index ascending, then within same round: train before analysis.
-    section_order = {'baseline': 0, 'train': 0, 'analysis': 1}
+    # 3b. If a gate is running for a given run_id, downgrade running cards
+    # in that run to 'waiting' (the step is blocked on human review).
+    runs_with_gate_running: set[str] = set()
+    for entry in latest_gate_by_key.values():
+        if entry.get('status') == 'running':
+            runs_with_gate_running.add(entry['_run_id'])
+    if runs_with_gate_running:
+        for sec in sections.values():
+            if sec['run_id'] in runs_with_gate_running:
+                for card in sec['cards']:
+                    if card['status'] == 'running':
+                        card['status'] = 'waiting'
+
+    # 4. Build round items in display order (chronological by first_ts).
     section_items: list[dict[str, Any]] = []
     for (run_id, section_type), sec in sections.items():
         run_index = _parse_run_index(run_id)
@@ -6415,7 +6425,7 @@ def _build_pipeline_items(
             'description': _SECTION_DESCRIPTIONS[section_type],
             'status': round_status,
             'cards': sec['cards'],
-            '_order_key': (run_index, section_order.get(section_type, 9), sec.get('first_ts', '')),
+            '_order_key': sec.get('first_ts', ''),
         })
 
     # 5. Build gate items.
@@ -6499,7 +6509,9 @@ def _compute_in_flight(
     seen_keys: set[str] = set()
     for idx in range(len(state_entries) - 1, -1, -1):
         entry = state_entries[idx]
-        if entry.get('kpi') or entry.get('log'):
+        if entry.get('kpi'):
+            continue
+        if entry.get('log') and not entry.get('step'):
             continue
         step = entry.get('step')
         if not isinstance(step, str) or not step:
@@ -6531,9 +6543,22 @@ def _compute_in_flight(
     if not target_run_id or not target_section:
         return None
 
+    # Check if a gate (human-check/human-review) is running for the target run_id
+    gate_running_for_run = False
+    for entry in state_entries:
+        step = entry.get('step')
+        if not isinstance(step, str) or not step:
+            continue
+        if _STEP_KIND.get(step) == 'gate' and entry.get('status') == 'running':
+            entry_run_id = entry.get('run_id')
+            if entry_run_id == target_run_id:
+                gate_running_for_run = True
+
     latest_per_step: dict[str, dict[str, Any]] = {}
     for idx, entry in enumerate(state_entries):
-        if entry.get('kpi') or entry.get('log'):
+        if entry.get('kpi'):
+            continue
+        if entry.get('log') and not entry.get('step'):
             continue
         step = entry.get('step')
         if not isinstance(step, str) or not step:
@@ -6559,6 +6584,8 @@ def _compute_in_flight(
         status = entry.get('status')
         if status not in ('complete', 'running'):
             continue
+        if status == 'running' and gate_running_for_run:
+            status = 'waiting'
         meta = _CARD_META.get((step, target_section), {
             'icon': 'clock',
             'title': step,
@@ -6634,7 +6661,9 @@ def _apply_program_state(
     # whenever a running section is hidden.
     latest_per_step: dict[str, str] = {}
     for entry in state_entries:
-        if entry.get('kpi') or entry.get('log'):
+        if entry.get('kpi'):
+            continue
+        if entry.get('log') and not entry.get('step'):
             continue
         step = entry.get('step')
         status = entry.get('status')
@@ -6655,15 +6684,19 @@ def _apply_program_state(
     statuses = list(latest_per_step.values())
     if not statuses:
         payload['status']['state'] = 'pending'
-    elif non_gate_running:
+    elif non_gate_running and not gate_running:
         payload['status']['state'] = 'running'
     elif gate_running:
-        # Only HiTL gates are 'running' — agent has handed control back to the
-        # user. Surface as 'waiting' so the header pill says "等待人工" instead
-        # of the spinning "Running" badge.
+        # HiTL gate is running — agent has handed control back to the user.
+        # Even if a parent step (e.g. augment) is still technically 'running',
+        # the gate blocks progress. Surface as 'waiting' so the header pill
+        # says "等待人工" instead of the spinning "Running" badge.
         payload['status']['state'] = 'waiting'
     elif all(s == 'complete' for s in statuses):
-        payload['status']['state'] = 'complete'
+        if run_active:
+            payload['status']['state'] = 'running'
+        else:
+            payload['status']['state'] = 'complete'
     elif 'complete' in statuses:
         payload['status']['state'] = 'running'
     elif 'failed' in statuses:
@@ -6702,6 +6735,7 @@ def _apply_program_state(
         and payload['status']['state'] == 'running'
         and running_steps
         and not all_watched
+        and not gate_running
     ):
         session_failure_status = 'failed'