fix: stabilize AI coach chat against truncation and empty replies

Co-authored-by: Cursor <cursoragent@cursor.com>

ai_client.py

@@ -124,11 +124,12 @@ def _openai_message_text(msg: dict) -> str:
     return text
 
 
-def _apply_max_tokens(body: dict, max_tokens: int | None) -> None:
+def _apply_max_tokens(body: dict, max_tokens: int | None, *, chat: bool = False) -> None:
     if max_tokens is not None and max_tokens > 0:
         mt = int(max_tokens)
         body["max_tokens"] = mt
-        body["max_completion_tokens"] = mt
+        if not chat:
+            body["max_completion_tokens"] = mt
 
 
 def _openai_chat_completion(
@@ -152,7 +153,7 @@ def _openai_chat_completion(
         "temperature": temperature,
         "stream": False,
     }
-    _apply_max_tokens(body, max_tokens)
+    _apply_max_tokens(body, max_tokens, chat=chat)
     r = requests.post(
         _openai_chat_url(),
         headers=headers,
@@ -167,9 +168,27 @@ def _openai_chat_completion(
     choice = choices[0] or {}
     msg = choice.get("message") or {}
     text = _openai_message_text(msg)
+    finish = str(choice.get("finish_reason") or "")
+    if not text and chat and max_tokens:
+        retry_body = dict(body)
+        retry_body.pop("max_completion_tokens", None)
+        r2 = requests.post(
+            _openai_chat_url(),
+            headers=headers,
+            json=retry_body,
+            timeout=_ai_timeout_seconds(image_count=image_count, chat=chat),
+        )
+        r2.raise_for_status()
+        data2 = r2.json()
+        choices2 = data2.get("choices") or []
+        if choices2:
+            msg2 = (choices2[0] or {}).get("message") or {}
+            text2 = _openai_message_text(msg2)
+            if text2:
+                return text2, str((choices2[0] or {}).get("finish_reason") or finish)
     if not text:
-        return "AI 生成失败：空内容", choice.get("finish_reason") or "error"
-    return text, str(choice.get("finish_reason") or "")
+        return "AI 生成失败：空内容", finish or "error"
+    return text, finish
 
 
 def _generate_openai(
@@ -264,7 +283,7 @@ _CHAT_CONTINUE_USER = (
 )
 _CHAT_END_CHARS = "。！？.!?\"」』）)>】"
 _INCOMPLETE_TAIL_RE = re.compile(
-    r"(会不会|是不是|够不够|能不能|要不要|如何|怎么|什么|哪里|多少|对吗|怎么样|"
+    r"(不会|不能|没有|会不会|是不是|够不够|能不能|要不要|如何|怎么|什么|哪里|多少|对吗|怎么样|"
     r"这个\.\.\.|这个…|\.\.\.\d+\.|\d+\.)$"
 )
 
@@ -291,7 +310,7 @@ def _should_continue(reason: str, full_text: str) -> bool:
 
 
 def _chat_continue_message(full_text: str) -> str:
-    tail = full_text[-900:] if len(full_text) > 900 else full_text
+    tail = full_text[-500:] if len(full_text) > 500 else full_text
     return (
         f"{_CHAT_CONTINUE_USER}\n\n"
         f"已写到最后这几句：\n「{tail}」\n\n"
@@ -299,6 +318,14 @@ def _chat_continue_message(full_text: str) -> str:
     )
 
 
+def _chat_continue_system(system: str) -> str:
+    return (
+        f"{system.strip()}\n\n"
+        "【续写模式】只输出断点后的剩余内容，不要重复前文；"
+        "列表每条单独一行；必须以句号、问号或感叹号收尾。"
+    )
+
+
 def ai_generate_chat(
     *,
     system: str,
@@ -306,9 +333,9 @@ def ai_generate_chat(
     temperature: float = 0.5,
     images_b64: Optional[Sequence[str]] = None,
     max_tokens: int = 8192,
-    max_continuations: int = 3,
+    max_continuations: int = 4,
 ) -> str:
-    """聊天专用：system/user 分消息；输出触顶时自动续写（携带已写全文）。"""
+    """聊天专用：system/user 分消息；输出触顶时轻量续写（不重复巨型上下文）。"""
     images = _collect_images(None, images_b64)
     max_rounds = max(1, int(max_continuations) + 1)
     try:
@@ -336,7 +363,7 @@ def ai_generate_chat(
                     messages,
                     temperature=temperature,
                     max_tokens=max_tokens,
-                    image_count=len(images),
+                    image_count=len(images) if attempt == 0 else 0,
                     chat=True,
                 )
                 if chunk.startswith("AI 调用失败") or chunk.startswith("AI 生成失败"):
@@ -346,8 +373,7 @@ def ai_generate_chat(
                 if not _should_continue(reason, full) or attempt >= max_rounds - 1:
                     break
                 messages = [
-                    {"role": "system", "content": system.strip()},
-                    base_user_msg,
+                    {"role": "system", "content": _chat_continue_system(system)},
                     {"role": "assistant", "content": full},
                     {"role": "user", "content": _chat_continue_message(full)},
                 ]
@@ -356,12 +382,14 @@ def ai_generate_chat(
         prompt = f"{system.strip()}\n\n---\n\n{user.strip()}"
         parts: list[str] = []
         for attempt in range(max_rounds):
-            current_prompt = prompt
             if parts:
                 full = "".join(parts)
                 current_prompt = (
-                    f"{prompt}\n\n【你已写道】\n{full}\n\n{_chat_continue_message(full)}"
+                    f"{_chat_continue_system(system)}\n\n"
+                    f"【你已写道】\n{full}\n\n{_chat_continue_message(full)}"
                 )
+            else:
+                current_prompt = prompt
             chunk, reason = _generate_ollama(
                 current_prompt,
                 images if not parts else [],
@@ -371,11 +399,13 @@ def ai_generate_chat(
             )
             if chunk.startswith("AI 生成失败") and not parts:
                 return chunk
+            if chunk.startswith("AI 生成失败"):
+                break
             parts.append(chunk)
             full = "".join(parts)
             if not _should_continue(reason, full) or attempt >= max_rounds - 1:
                 break
-        return "".join(parts).strip() or "AI 生成失败"
+        return "".join(parts).strip() or "AI 生成失败：空内容"
     except requests.HTTPError as e:
         detail = ""
         try: