fix: 多Agent协作 JSON 截断导致主管规划解析失败

- 根因: deepseek-v4-flash 规划输出超 max_tokens=2000 被截断，JSON 不完整解析失败
- agents.py: 新增 _chat_json() 统一处理 JSON 类调用，解析失败自动加大 max_tokens 重试(×1/×2/×3)
- 主管规划/辩论质询/辩论裁决/评审 全部改用 _chat_json 并提高初始上限(3000/1200/2500/2000)
- 实测: 主管(计算器网页版4子任务)✅ 评审(92分)✅ 辩论(指定阵容)✅

agents.py

@@ -92,6 +92,23 @@ def _extract_score(text):
     return 60.0, text
 
 
+def _chat_json(worker, messages, max_tokens=2000, temperature=None):
+    """调用 LLM 并解析 JSON；解析失败自动加大 max_tokens 重试（防止长文截断）。
+    返回 (data, raw_text, usage)；全部失败抛最后异常。"""
+    last_err = None
+    for attempt in range(3):
+        mt = max_tokens * (attempt + 1)  # 2000 → 4000 → 6000
+        text, usage = _chat_worker(worker, messages, temperature=temperature, max_tokens=mt)
+        try:
+            data = _extract_json(text)
+            if isinstance(data, dict) and not data:
+                raise ValueError('空 JSON')
+            return data, text, usage
+        except Exception as e:
+            last_err = e
+    raise last_err or ValueError('JSON 解析失败')
+
+
 # ---------------------------------------------------------------------------
 # 主管模式
 # ---------------------------------------------------------------------------
@@ -114,17 +131,17 @@ SUPERVISOR_SYNTH_PROMPT = (
 def run_supervisor(run_id, run, workers):
     _log(run_id, 'supervisor', None, 'plan',
          f'主管开始规划：{run["topic"][:200]}')
-    # 1) 主管拆解
+    # 1) 主管拆解（JSON 解析失败自动加大 max_tokens 重试）
-    plan_text, u1 = _chat_worker(
+    try:
+        data, _, u1 = _chat_json(
             workers[0],
             [{'role': 'system', 'content': '你只输出 JSON。'},
              {'role': 'user', 'content': SUPERVISOR_PLAN_PROMPT.format(
                  topic=run['topic'], context=run['context'] or '无')}],
-        temperature=0.3, max_tokens=2000)
+            max_tokens=3000, temperature=0.3)
-    try:
+        subtasks = data.get('subtasks') or data.get('tasks') or []
-        subtasks = _extract_json(plan_text)
+        if isinstance(data, list):
-        if isinstance(subtasks, dict):
+            subtasks = data
-            subtasks = subtasks.get('subtasks') or subtasks.get('tasks') or []
     except Exception as e:
         _update_run(run_id, status='failed', error=f'主管规划解析失败: {e}')
         _log(run_id, 'supervisor', None, 'plan', f'❌ 规划解析失败: {e}')
@@ -216,15 +233,19 @@ def run_review(run_id, run, workers):
         # 评审
         review_usage = None
         try:
-            review_text, u2 = _chat_worker(
+            review_data, review_text, u2 = _chat_json(
                 reviewer_w,
                 [{'role': 'system', 'content': '你只输出 JSON。'},
                  {'role': 'user', 'content': REVIEWER_PROMPT.format(
                      topic=run['topic'], output=final_output, rubric=rubric)}],
-                temperature=0.2, max_tokens=1500)
+                max_tokens=2000, temperature=0.2)
-            score, judgment = _extract_score(review_text)
             review_usage = u2
             total_tokens += u2['total_tokens']; total_cost += u2['cost']
+            if isinstance(review_data, dict):
+                score = float(review_data.get('score', review_data.get('总分', 60)))
+                judgment = review_data.get('judgment') or review_data.get('意见') or review_text
+            else:
+                score, judgment = _extract_score(review_text)
         except Exception as e:
             score, judgment = 0, f'评审调用失败: {e}'
         final_score = score
@@ -325,14 +346,13 @@ def run_debate(run_id, run, workers):
                 f"【{v['stance']}】(#{k})：{v['view'][:500]}"
                 for k, v in views.items() if k != w['id'])
             try:
-                text, u = _chat_worker(
+                data, _, u = _chat_json(
                     w,
                     [{'role': 'system', 'content': '你只输出 JSON。'},
                      {'role': 'user', 'content': DEBATE_REBUT_PROMPT.format(
                          stance=mine['stance'], topic=run['topic'],
                          my_view=mine['view'], others=others)}],
-                    temperature=0.7, max_tokens=800)
+                    max_tokens=1200, temperature=0.7)
-                data = _extract_json(text)
                 rebut = data.get('rebuttal') if isinstance(data, dict) else text
                 mine['view'] = mine['view'] + '\n\n【质询回应】' + str(rebut)
                 total_tokens += u['total_tokens']; total_cost += u['cost']
@@ -346,12 +366,11 @@ def run_debate(run_id, run, workers):
         f"【{v['stance']}】{v['view']}" for v in views.values())
     _log(run_id, 'judge', judge_w['id'], 'verdict', '裁判综合裁决中…')
     try:
-        verdict, u4 = _chat_worker(
+        data, verdict, u4 = _chat_json(
             judge_w,
             [{'role': 'system', 'content': '你只输出 JSON。你是公正严明的首席裁判。'},
              {'role': 'user', 'content': JUDGE_PROMPT.format(topic=run['topic'], transcript=transcript)}],
-            temperature=0.3, max_tokens=1500)
+            max_tokens=2500, temperature=0.3)
-        data = _extract_json(verdict)
         if isinstance(data, dict):
             consensus = data.get('consensus') or data.get('结论') or verdict
             summary = data.get('summary') or data.get('摘要') or ''