v2.0.1：修复推理型模型(Qwen3/DeepSeek reasoning_content)误报未收到输出+连接测试按连通判定+采样失败不中断+流式超时放宽60s/300s

llm_providers.py

@@ -56,6 +56,8 @@ def _parse_sse_line(line):
 
 def _metrics(start, first_token_at, end, prompt_tokens, output_tokens,
              cached_tokens, output_chars, prompt_chars):
+    if first_token_at is None:
+        first_token_at = end  # 未收到正文但请求完成（如纯 usage 响应）
     ttft_ms = (first_token_at - start) * 1000
     decode_ms = (end - first_token_at) * 1000
     total_ms = (end - start) * 1000
@@ -103,6 +105,7 @@ def stream_openai(cfg, prompt, gen, log, should_stop=None):
     first_token_at = None
     output_chars = 0
     prompt_tokens = output_tokens = cached_tokens = 0
+    event_count = 0
     resp = None
     try:
         while True:
@@ -126,13 +129,15 @@ def stream_openai(cfg, prompt, gen, log, should_stop=None):
         for obj in _iter_json(resp):
             if should_stop and should_stop():
                 raise StopRequested()
+            event_count += 1
             if obj.get("choices"):
                 delta = obj["choices"][0].get("delta") or {}
-                text = delta.get("content") or ""
-                if text:
+                # 兼容推理型模型：Qwen3/DeepSeek 思维链在 reasoning_content
+                piece = delta.get("content") or delta.get("reasoning_content") or ""
+                if piece:
                     if first_token_at is None:
                         first_token_at = time.time()
-                    output_chars += len(text)
+                    output_chars += len(piece)
             usage = obj.get("usage")
             if usage:
                 prompt_tokens = usage.get("prompt_tokens") or 0
@@ -148,8 +153,8 @@ def stream_openai(cfg, prompt, gen, log, should_stop=None):
         if resp is not None:
             resp.close()
 
-    if first_token_at is None:
-        raise ProviderError("未收到任何输出内容")
+    if event_count == 0:
+        raise ProviderError("未收到任何输出内容（HTTP 200 但响应流为空）")
     end = time.time()
     return _metrics(start, first_token_at, end, prompt_tokens, output_tokens,
                     cached_tokens, output_chars, len(prompt))
@@ -190,6 +195,7 @@ def stream_anthropic(cfg, prompt, gen, log, should_stop=None):
     first_token_at = None
     output_chars = 0
     prompt_tokens = output_tokens = 0
+    event_count = 0
     resp = None
     try:
         if should_stop and should_stop():
@@ -204,12 +210,15 @@ def stream_anthropic(cfg, prompt, gen, log, should_stop=None):
         for obj in _iter_json(resp):
             if should_stop and should_stop():
                 raise StopRequested()
+            event_count += 1
             etype = obj.get("type")
             if etype == "message_start":
                 usage = (obj.get("message") or {}).get("usage") or {}
                 prompt_tokens = usage.get("input_tokens") or 0
             elif etype == "content_block_delta":
-                text = (obj.get("delta") or {}).get("text") or ""
+                delta = obj.get("delta") or {}
+                # 兼容 extended thinking：thinking 文本也算输出
+                text = delta.get("text") or delta.get("thinking") or ""
                 if text:
                     if first_token_at is None:
                         first_token_at = time.time()
@@ -225,8 +234,8 @@ def stream_anthropic(cfg, prompt, gen, log, should_stop=None):
         if resp is not None:
             resp.close()
 
-    if first_token_at is None:
-        raise ProviderError("未收到任何输出内容")
+    if event_count == 0:
+        raise ProviderError("未收到任何输出内容（HTTP 200 但响应流为空）")
     end = time.time()
     return _metrics(start, first_token_at, end, prompt_tokens, output_tokens,
                     0, output_chars, len(prompt))
@@ -252,6 +261,7 @@ def stream_google(cfg, prompt, gen, log, should_stop=None):
     first_token_at = None
     output_chars = 0
     prompt_tokens = output_tokens = cached_tokens = 0
+    event_count = 0
     resp = None
     try:
         if should_stop and should_stop():
@@ -266,11 +276,13 @@ def stream_google(cfg, prompt, gen, log, should_stop=None):
         for obj in _iter_json(resp):
             if should_stop and should_stop():
                 raise StopRequested()
+            event_count += 1
             cands = obj.get("candidates") or []
             if cands:
                 parts = (cands[0].get("content") or {}).get("parts") or []
                 for part in parts:
-                    text = part.get("text") or ""
+                    # 兼容 thinking 模型：thought 文本也算输出
+                    text = part.get("text") or part.get("thought") or ""
                     if text:
                         if first_token_at is None:
                             first_token_at = time.time()
@@ -288,8 +300,8 @@ def stream_google(cfg, prompt, gen, log, should_stop=None):
         if resp is not None:
             resp.close()
 
-    if first_token_at is None:
-        raise ProviderError("未收到任何输出内容")
+    if event_count == 0:
+        raise ProviderError("未收到任何输出内容（HTTP 200 但响应流为空）")
     end = time.time()
     return _metrics(start, first_token_at, end, prompt_tokens, output_tokens,
                     cached_tokens, output_chars, len(prompt))