refactor MixinSession._raw_ask: simplify retry logic, skip yielding first-only error chunk

agent_loop.py

@@ -91,5 +91,5 @@ def agent_runner_loop(client, system_prompt, user_input, handler, tools_schema,
                 next_prompt += f"<tool_result>\n{datastr}\n</tool_result>\n\n"
             next_prompt += outcome.next_prompt
         next_prompt = handler.next_prompt_patcher(next_prompt, None, turn+1)
-        messages = [{"role": "user", "content": next_prompt}]
+        messages = [{"role": "user", "content": next_prompt}]   # just new message, history is kept in *Session
     return {'result': 'MAX_TURNS_EXCEEDED'}

assets/tools_schema.json

@@ -1,7 +1,7 @@
 [
   {"type": "function", "function": {
     "name": "code_run",
-    "description": "代码执行器。优先python，系统操作用powershell。禁同时调用多个。为免转义问题，代码放正文 ```python/powershell 块中。禁硬编码大量数据",
+    "description": "代码执行器。优先使用python。禁同时调用多个。为免转义问题，代码放正文 ```python/powershell 块中。禁硬编码大量数据",
     "parameters": {"type": "object", "properties": {
       "script": {"type": "string", "description": "[Optional] 要执行的代码。为免转义建议留空，改用正文代码块（与此参数互斥）"},
       "type": {"type": "string", "enum": ["python", "powershell"], "description": "代码类型", "default": "python"},

llmcore.py

@@ -814,26 +814,20 @@ class MixinSession:
     @property
     def primary(self): return self._sessions[0]
     def _raw_ask(self, *args, **kwargs):
-        last_err = None
         for attempt in range(self._retries + 1):
             gen = self._orig_raw_asks[attempt % len(self._sessions)](*args, **kwargs)
-            try: first = next(gen)
+            last_chunk, return_val, yielded = None, [], False
-            except StopIteration as e: return e.value or []
+            try:
-            if isinstance(first, str) and first.startswith('Error:'):
+                while True:
-                last_err = first
+                    chunk = next(gen); last_chunk = chunk
-                for _ in gen: pass  # drain
+                    if not yielded and isinstance(chunk, str) and chunk.startswith('Error:'): continue
-                if attempt < self._retries:
+                    yield chunk; yielded = True
-                    delay = min(30, self._base_delay * (2 ** attempt))
+            except StopIteration as e: return_val = e.value or []
-                    print(f'[MixinSession] {first[:80]}, retry {attempt+1}/{self._retries} in {delay:.1f}s')
+            if isinstance(last_chunk, str) and last_chunk.startswith('Error:') and attempt < self._retries:
-                    time.sleep(delay); continue
+                delay = min(30, self._base_delay * (2 ** attempt))
-            else:
+                print(f'[MixinSession] {last_chunk[:80]}, retry {attempt+1}/{self._retries} in {delay:.1f}s')
-                yield first
+                time.sleep(delay); continue
-                try:
+            return return_val
-                    while True: yield next(gen)
-                except StopIteration as e: return e.value or []
-        yield last_err or 'Error: all retries exhausted'
-        return [{'type': 'text', 'text': last_err}]
-
 
 class NativeToolClient:
     THINKING_PROMPT = """