fix: 经验巡检评审失败重试——模型偶发输出解释文字不带 JSON（实测 4/11 条评审失败自动保留），失败后重试一次并强调只输出 JSON；仍失败才保守保留

2026-09-04 16:19:23 +08:00
parent 04d07b2c62
commit fbd01a056c
1 changed files with 38 additions and 32 deletions
@@ -286,38 +286,44 @@ def _ensure_audit_table():


 def _review_one_exp(cfg, prompt, recipe, tool_seq, hits):
-    """调模型评审单条经验，返回 (verdict, score, reason)；失败保守返回保留。"""
-    try:
-        import httpx
-        body = {
-            "model": cfg.get("model") or "deepseek-v4-flash-vision-exp",
-            "messages": [{"role": "user",
-                          "content": _AUDIT_PROMPT.format(
-                              prompt=(prompt or "")[:400],
-                              recipe=(recipe or "")[:600],
-                              tool_seq=(tool_seq or "")[:600],
-                              hits=hits or 0)}],
-            "max_tokens": 200,
-        }
-        headers = {"Authorization": f"Bearer {cfg.get('api_key', '')}",
-                   "Content-Type": "application/json"}
-        r = httpx.post(f"{(cfg.get('api_base') or 'https://api.deepseek.com').rstrip('/')}/chat/completions",
-                       json=body, headers=headers, timeout=20)
-        if r.status_code != 200:
-            raise RuntimeError(f"HTTP {r.status_code}")
-        text = (((r.json() or {}).get("choices") or [{}])[0]
-                .get("message") or {}).get("content") or ""
-        m = re.search(r"\{.*\}", text, re.S)
-        if not m:
-            raise RuntimeError("响应无 JSON")
-        j = json.loads(m.group(0))
-        keep = bool(j.get("keep", True))
-        score = max(0, min(10, int(j.get("score", 5) or 5)))
-        reason = str(j.get("reason", ""))[:200]
-        return ("keep" if keep else "delete"), score, reason
-    except Exception as e:
-        _log.warning(f"经验评审失败（按保留处理）: {e}")
-        return "keep", 0, f"评审失败自动保留: {str(e)[:80]}"
+    """调模型评审单条经验，返回 (verdict, score, reason)。
+
+    模型偶尔输出解释文字不带 JSON——重试一次（强调只输出 JSON），
+    仍失败则保守返回保留。
+    """
+    import httpx
+    base_prompt = _AUDIT_PROMPT.format(prompt=(prompt or "")[:400],
+                                       recipe=(recipe or "")[:600],
+                                       tool_seq=(tool_seq or "")[:600],
+                                       hits=hits or 0)
+    for attempt in (1, 2):
+        try:
+            content = base_prompt if attempt == 1 else (
+                base_prompt + "\n注意：只输出一个 JSON 对象，不要输出任何其它文字、解释或代码块标记。")
+            body = {
+                "model": cfg.get("model") or "deepseek-v4-flash-vision-exp",
+                "messages": [{"role": "user", "content": content}],
+                "max_tokens": 200,
+            }
+            headers = {"Authorization": f"Bearer {cfg.get('api_key', '')}",
+                       "Content-Type": "application/json"}
+            r = httpx.post(f"{(cfg.get('api_base') or 'https://api.deepseek.com').rstrip('/')}/chat/completions",
+                           json=body, headers=headers, timeout=20)
+            if r.status_code != 200:
+                raise RuntimeError(f"HTTP {r.status_code}")
+            text = (((r.json() or {}).get("choices") or [{}])[0]
+                    .get("message") or {}).get("content") or ""
+            m = re.search(r"\{.*\}", text, re.S)
+            if not m:
+                raise RuntimeError("响应无 JSON")
+            j = json.loads(m.group(0))
+            keep = bool(j.get("keep", True))
+            score = max(0, min(10, int(j.get("score", 5) or 5)))
+            reason = str(j.get("reason", ""))[:200]
+            return ("keep" if keep else "delete"), score, reason
+        except Exception as e:
+            _log.warning(f"经验评审第 {attempt} 次失败: {e}")
+    return "keep", 0, "评审失败自动保留（两次尝试均无有效 JSON）"


 def run_experience_audit():