update: 同步最新代码与配置

app/main.py

@@ -2,7 +2,6 @@ import asyncio
 from contextlib import asynccontextmanager
 
 from fastapi import FastAPI
-from fastapi.responses import ORJSONResponse
 from loguru import logger
 
 from .server.chat import router as chat_router
@@ -93,7 +92,7 @@ def create_app() -> FastAPI:
         description="OpenAI-compatible API for Gemini Web",
         version="1.0.0",
         lifespan=lifespan,
-        default_response_class=ORJSONResponse,
+
     )
 
     add_cors_middleware(app)

app/server/chat.py

@@ -1,6 +1,8 @@
+import asyncio
 import base64
 import hashlib
 import io
+import random
 import re
 import reprlib
 import uuid
@@ -258,7 +260,7 @@ def _process_llm_output(
         # Let's import the cleaning functions? They are internal to client.py.
         # Better: Use GeminiClientWrapper.extract_output with a dummy ModelOutput.
         
-        dummy_output = ModelOutput(text=visible_output, thoughts=None, images=[], metadata=[], candidates=[])
+        dummy_output = ModelOutput(metadata=[], candidates=[{"rcid": "dummy", "text": visible_output}])
         # We don't want to re-add thoughts (include_thoughts=False) because we handle them separately?
         # extract_output puts thoughts in <think> tags if include_thoughts=True.
         # Here `visible_output` DOES NOT have thoughts (they are in `raw_output_with_think`'s <think> part).
@@ -280,7 +282,7 @@ def _process_llm_output(
             # Apply to storage_output too? 
             # storage_output usually should match visible_output for history consistency.
             storage_output = GeminiClientWrapper.extract_output(
-                 ModelOutput(text=storage_output, thoughts=None, images=[], metadata=[], candidates=[]), 
+                 ModelOutput(metadata=[], candidates=[{"rcid": "dummy", "text": storage_output}]), 
                  include_thoughts=False
             )
         except Exception as e:
@@ -780,6 +782,40 @@ async def _send_with_split(
     Includes retry with exponential backoff for transient failures.
     """
     
+    async def _stream_with_retry(
+        content: str, file_list: list | None
+    ) -> AsyncGenerator[ModelOutput, None]:
+        """Manual retry logic for streaming."""
+        # Reduced retries to 1 to minimize ban risk
+        max_retries = 1
+        for attempt in range(max_retries + 1):
+            try:
+                gen = session.send_message_stream(content, files=file_list)
+                has_yielded = False
+                async for chunk in gen:
+                    yield chunk
+                    has_yielded = True
+                return
+            except Exception as e:
+                if has_yielded:
+                    raise e
+                
+                # Safety check: Do not retry if error indicates auth failure or rate limit
+                error_str = str(e).lower()
+                if "429" in error_str or "403" in error_str or "quota" in error_str:
+                    raise e
+
+                if attempt < max_retries:
+                    # Increased delay with randomness to avoid pattern detection
+                    base_delay = 3.0
+                    jitter = random.uniform(0.5, 1.5)
+                    delay = base_delay + jitter
+                    
+                    logger.warning(f"Stream failed to start (attempt {attempt+1}/{max_retries}). Retrying in {delay:.2f}s. Error: {e}")
+                    await asyncio.sleep(delay)
+                else:
+                    raise e
+
     @retry_with_backoff(
         max_retries=3,
         base_delay=1.0,
@@ -792,7 +828,7 @@ async def _send_with_split(
     ) -> AsyncGenerator[ModelOutput, None] | ModelOutput:
         """Internal function with retry logic."""
         if is_stream:
-            return session.send_message_stream(content, files=file_list)
+            return _stream_with_retry(content, file_list)
         return await session.send_message(content, files=file_list)
     
     if len(text) <= MAX_CHARS_PER_REQUEST:
@@ -1554,6 +1590,17 @@ def _create_responses_real_streaming_response(
                     if img.url not in seen_urls:
                         images.append(img)
                         seen_urls.add(img.url)
+        
+        # Check if image generation was forced via tool_choice (same logic as non-streaming)
+        image_generation_forced = (
+            request.tool_choice is not None 
+            and isinstance(request.tool_choice, ResponseToolChoice) 
+            and request.tool_choice.type == "image_generation"
+        )
+        if image_generation_forced and not images and not assistant_text:
+            logger.warning("Image generation was forced via tool_choice but no images or text were returned in stream.")
+            yield f"data: {orjson.dumps({**base_event, 'type': 'error', 'error': {'message': 'No images returned from forced image generation request.'}}).decode('utf-8')}\n\n"
+            return
 
         response_contents, image_call_items = [], []
         seen_hashes = set()
@@ -2000,6 +2047,22 @@ async def create_chat_completion(
 
     # Process images for OpenAI non-streaming flow
     images = resp_or_stream.images or []
+    
+    # Log response details for debugging
+    logger.debug(f"Chat response: text_len={len(visible_output)}, images={len(images)}, tool_calls={len(tool_calls)}")
+    logger.debug(f"Raw response text (first 500 chars): {raw_with_t[:500] if raw_with_t else 'EMPTY'}")
+    
+    # Check if response is completely empty
+    if not visible_output and not images and not tool_calls:
+        logger.warning("Gemini returned an empty response for chat completion (no text, images, or tool calls)")
+        # Log more details about the response object
+        logger.debug(f"Response object type: {type(resp_or_stream)}")
+        logger.debug(f"Response has candidates: {hasattr(resp_or_stream, 'candidates')}")
+        if hasattr(resp_or_stream, 'candidates'):
+            logger.debug(f"Candidates: {resp_or_stream.candidates}")
+        if hasattr(resp_or_stream, 'text'):
+            logger.debug(f"Response text attribute: {repr(resp_or_stream.text)[:200] if resp_or_stream.text else 'None'}")
+    
     image_markdown = ""
     seen_hashes = set()
     for image in images:
@@ -2299,8 +2362,8 @@ async def create_response(
             remain = [messages[-1]]
         msgs = _prepare_messages_for_model(
             remain,
-            request.tools,
-            request.tool_choice,
+            standard_tools or None,
+            model_tool_choice,
             None,
             False,
         )
@@ -2366,10 +2429,28 @@ async def create_response(
 
     assistant_text, storage_output, tool_calls = _process_llm_output(raw_t, raw_c, struct_req)
     images = resp_or_stream.images or []
-    if (
-        request.tool_choice is not None and request.tool_choice.type == "image_generation"
-    ) and not images:
-        raise HTTPException(status_code=status.HTTP_502_BAD_GATEWAY, detail="No images returned.")
+    
+    # Log response details for debugging
+    logger.debug(f"Response: text_len={len(assistant_text)}, images={len(images)}, tool_calls={len(tool_calls)}")
+    
+    # Check if response is completely empty (no text, no images, no tool calls)
+    if not assistant_text and not images and not tool_calls:
+        logger.warning("Gemini returned an empty response (no text, images, or tool calls)")
+        # Check if there's an error in the response
+        if hasattr(resp_or_stream, 'candidates') and resp_or_stream.candidates:
+            logger.debug(f"Candidates: {resp_or_stream.candidates}")
+    
+    # Check if image generation was forced via tool_choice
+    # Only enforce image requirement if tool_choice explicitly requests image generation
+    # tools: [{"type": "image_generation"}] just declares the tool is available, not mandatory
+    image_generation_forced = (
+        request.tool_choice is not None 
+        and isinstance(request.tool_choice, ResponseToolChoice) 
+        and request.tool_choice.type == "image_generation"
+    )
+    if image_generation_forced and not images and not assistant_text:
+        logger.warning("Image generation was forced via tool_choice but no images or text were returned.")
+        raise HTTPException(status_code=status.HTTP_502_BAD_GATEWAY, detail="No images returned from forced image generation request.")
 
     contents, img_calls = [], []
     seen_hashes = set()

app/services/client.py

@@ -117,6 +117,8 @@ class GeminiClientWrapper(GeminiClient):
         auto_refresh: bool = cast(bool, _UNSET),
         refresh_interval: float = cast(float, _UNSET),
         verbose: bool = cast(bool, _UNSET),
+        watchdog_timeout: float = cast(float, _UNSET),
+        **kwargs: Any,
     ) -> None:
         """
         Inject default configuration values.
@@ -127,15 +129,21 @@ class GeminiClientWrapper(GeminiClient):
         refresh_interval = cast(float, _resolve(refresh_interval, config.refresh_interval))
         verbose = cast(bool, _resolve(verbose, config.verbose))
 
+        call_kwargs: dict[str, Any] = {
+            "timeout": timeout,
+            "auto_close": auto_close,
+            "close_delay": close_delay,
+            "auto_refresh": auto_refresh,
+            "refresh_interval": refresh_interval,
+            "verbose": verbose,
+        }
+        if watchdog_timeout is not _UNSET:
+            call_kwargs["watchdog_timeout"] = watchdog_timeout
+        if kwargs:
+            call_kwargs.update(kwargs)
+
         try:
-            await super().init(
-                timeout=timeout,
-                auto_close=auto_close,
-                close_delay=close_delay,
-                auto_refresh=auto_refresh,
-                refresh_interval=refresh_interval,
-                verbose=verbose,
-            )
+            await super().init(**call_kwargs)
         except Exception:
             logger.exception(f"Failed to initialize GeminiClient {self.id}")
             raise

config/config.yaml

@@ -2,7 +2,7 @@
 
 server:
   host: "0.0.0.0"          # Server bind address
-  port: 7860               # Server port
+  port: 8058               # Server port
   api_key: "miaolegewang"            # API key for authentication (null for no auth)
   https:
     enabled: false                     # Enable HTTPS
@@ -19,12 +19,12 @@ cors:
 gemini:
   clients:
     - id: "example-id-1"   # Arbitrary client ID
-      secure_1psid: "g.a0006wg0ISvzbHtAs0FzakIebdoqG5dGk9WFXw6e638gOIHnKKiv8YIi5uu5jaWX67qorJAznwACgYKAWASARYSFQHGX2MikWwtMeXo1Swux9m-XIyS2RoVAUF8yKoxSRH01AeWFX2soV9uEmfc0076"
-      secure_1psidts: "sidts-CjIB7I_69IrJuStE_UhVsJNkMkIpeJblzj8dNfGM2kReQgfvnw0QveII_X9QLqyoJ0liIhAA"
+      secure_1psid: "g.a0006ghLMuTMAgxpLCPEErGijJmoYdaeoQ8jQvccH014n1gYV_A6_9K7SUzPUYqsFZYeKCFj0QACgYKAe8SARcSFQHGX2MiQREdDmeRiaB-meb6rxWh1RoVAUF8yKqKLXLomJGmcEm899BsY1K90076"
+      secure_1psidts: "sidts-CjEB7I_69ImkFO--6ZoqrN9V6ThAw-I__6wzxDni5ELJR7Ia4najVtYeQCmjpk01XH0yEAA"
       proxy: null          # Optional proxy URL (null/empty means direct connection)
   timeout: 120             # Init timeout in seconds
   auto_refresh: true       # Auto-refresh session cookies
-  refresh_interval: 540    # Refresh interval in seconds
+  refresh_interval: 3600    # Refresh interval in seconds
   verbose: false           # Enable verbose logging for Gemini requests
   max_chars_per_request: 1000000     # Maximum characters Gemini Web accepts per request. Non-pro users might have a lower limit
   model_strategy: "append" # Strategy: 'append' (default + custom) or 'overwrite' (custom only)

run.py

@@ -13,9 +13,6 @@ if __name__ == "__main__":
     # Setup loguru logging
     setup_logging(level=g_config.logging.level)
 
-    # Override port from environment variable (for HuggingFace Spaces compatibility)
-    port = int(os.getenv("PORT", g_config.server.port))
-
     # Check HTTPS configuration
     if g_config.server.https.enabled:
         key_path = g_config.server.https.key_file
@@ -28,20 +25,20 @@ if __name__ == "__main__":
             )
             sys.exit(1)
 
-        logger.info(f"Starting server at https://{g_config.server.host}:{port} ...")
+        logger.info(f"Starting server at https://{g_config.server.host}:{g_config.server.port} ...")
         uvicorn.run(
             app,
             host=g_config.server.host,
-            port=port,
+            port=g_config.server.port,
             log_config=None,
             ssl_keyfile=key_path,
             ssl_certfile=cert_path,
         )
     else:
-        logger.info(f"Starting server at http://{g_config.server.host}:{port} ...")
+        logger.info(f"Starting server at http://{g_config.server.host}:{g_config.server.port} ...")
         uvicorn.run(
             app,
             host=g_config.server.host,
-            port=port,
+            port=g_config.server.port,
             log_config=None,
         )