更新了直接回答情况下的文本返回格式

agent.py

@@ -23,6 +23,8 @@ Rules:
 7. To finish, output:
    {"thought":"...","action":"final_answer","answer":"...","confidence":"high|medium|low"}
 8. Final answers must obey the requested format exactly. Do not include explanations.
+9. Never answer in plain text. Even if the answer is obvious, return JSON.
+10. For self-contained text/table/regex/list questions, call direct_answer_tool first.
 """.strip()
 
 
@@ -157,7 +159,9 @@ class GaiaAgent:
             + tool_prompt()
             + "\n\nTask:\n"
             + json.dumps(task_payload, ensure_ascii=False)
-            + "\n\nChoose the first tool call or return final_answer. Output JSON only."
+            + "\n\nChoose the first tool call or return final_answer. Output JSON only.\n"
+            + "Example tool call: {\"thought\":\"self-contained rule question\",\"action\":\"direct_answer_tool\",\"args\":{}}\n"
+            + "Example final answer: {\"thought\":\"tool returned the answer\",\"action\":\"final_answer\",\"answer\":\"right\",\"confidence\":\"high\"}"
         )
         return [
             {"role": "system", "content": SYSTEM_PROMPT},

tools/llm_client.py

@@ -27,23 +27,28 @@ def call_hf_chat(
     model: str = HF_TEXT_MODEL,
     max_tokens: int = 512,
     temperature: float = 0.1,
+    response_format: dict[str, Any] | None = None,
 ) -> str:
     token = get_hf_token()
     if not token:
         raise RuntimeError("未配置 HF_TOKEN。")
 
+    payload: dict[str, Any] = {
+        "model": model,
+        "messages": messages,
+        "temperature": temperature,
+        "max_tokens": max_tokens,
+    }
+    if response_format:
+        payload["response_format"] = response_format
+
     response = requests.post(
         HF_ROUTER_URL,
         headers={
             "Authorization": f"Bearer {token}",
             "Content-Type": "application/json",
         },
-        json={
-            "model": model,
-            "messages": messages,
-            "temperature": temperature,
-            "max_tokens": max_tokens,
-        },
+        json=payload,
         timeout=180,
     )
     if not response.ok:
@@ -51,7 +56,10 @@ def call_hf_chat(
             f"HF chat 请求失败：status={response.status_code}, body={response.text[:500]}"
         )
     data = response.json()
-    return data["choices"][0]["message"]["content"]
+    content = data["choices"][0]["message"].get("content")
+    if content is None:
+        return ""
+    return content
 
 
 def extract_json_object(text: str) -> dict[str, Any]:
@@ -94,9 +102,56 @@ def parse_agent_action(text: str) -> AgentAction:
     )
 
 
+def call_planner_model(messages: list[dict[str, Any]]) -> str:
+    try:
+        return call_hf_chat(
+            messages,
+            max_tokens=700,
+            temperature=0.0,
+            response_format={"type": "json_object"},
+        )
+    except RuntimeError as exc:
+        error_text = str(exc).lower()
+        if "response_format" not in error_text and "json_object" not in error_text:
+            raise
+        return call_hf_chat(
+            messages,
+            max_tokens=700,
+            temperature=0.0,
+        )
+
+
 def plan_next_action(messages: list[dict[str, Any]]) -> AgentAction:
-    raw = call_hf_chat(messages, max_tokens=700, temperature=0.0)
-    return parse_agent_action(raw)
+    raw = call_planner_model(messages)
+    try:
+        return parse_agent_action(raw)
+    except Exception as first_error:
+        repair_messages = messages + [
+            {
+                "role": "assistant",
+                "content": raw or "<empty response>",
+            },
+            {
+                "role": "user",
+                "content": (
+                    "Your previous response was invalid because it was not one JSON object. "
+                    "Return exactly one JSON object now, with no markdown and no prose. "
+                    "Valid tool-call schema: "
+                    "{\"thought\":\"...\",\"action\":\"tool_name\",\"args\":{}}. "
+                    "Valid final-answer schema: "
+                    "{\"thought\":\"...\",\"action\":\"final_answer\",\"answer\":\"...\",\"confidence\":\"high|medium|low\"}."
+                ),
+            },
+        ]
+        repaired_raw = call_planner_model(repair_messages)
+        try:
+            return parse_agent_action(repaired_raw)
+        except Exception as second_error:
+            raise ValueError(
+                "planner 连续两次没有输出合法 JSON。"
+                f" first_error={first_error}; first_raw={raw[:300]!r};"
+                f" second_error={second_error}; second_raw={repaired_raw[:300]!r}"
+            ) from second_error
 
 
 def answer_with_light_model(question: str, evidence: str) -> str: