fix: delete non-native gem & auto create gems

Author: sagitchu

README.md

@@ -145,25 +145,19 @@ When a gem is selected:
 
 - The server uses the gem definition's `model` as the actual Gemini model name for the upstream call.
 - The OpenAI-compatible response tries to echo back the client-provided `model` (i.e. `gem:<id>`) to keep client-side consistency.
+- **Auto-Sync**: On startup, the server checks if these gems exist on your Google account. If missing, it automatically creates them using the provided `id` (as title) and `system_prompt`.
 
 Example config:
 
 ```yaml
 gemini:
   gems:
-    - id: "default-gem"
-      model: "gemini-3.0-pro"
-      system_prompt: "You are a careful assistant."
+    - id: "coding-helper"
+      model: "gemini-2.0-flash"
+      system_prompt: "You are an expert software engineer."
       tool_policy: "allow" # allow | disallow | auto
       default_temperature: 0.2
       top_p: 0.8
-      max_output_tokens: 1024
-
-    # Native Gem (Use Gems created on gemini.google.com)
-    - id: "my-native-gem"
-      model: "gemini-2.0-flash"
-      is_native: true  # This flag passes the 'id' to Google's servers
-      default_temperature: 1.0 # Parameters can still be overridden locally
       max_output_tokens: 8192
 ```
 

README.zh.md

@@ -145,25 +145,19 @@ docker compose up -d
 
 - 服务端会使用 gem 定义中的 `model` 作为实际调用的 Gemini 模型名。
 - 对外返回的 OpenAI 兼容响应会尽量回显客户端传入的 `model`（即 `gem:<id>`），以保证客户端一致性。
+- **自动同步**：服务启动时会自动检查这些 Gem 是否存在于你的 Google 账号中。如果不存在，会自动根据提供的 `id`（作为标题）和 `system_prompt` 进行创建。
 
 配置示例：
 
 ```yaml
 gemini:
   gems:
-    - id: "default-gem"
-      model: "gemini-3.0-pro"
-      system_prompt: "You are a careful assistant."
+    - id: "coding-helper"
+      model: "gemini-2.0-flash"
+      system_prompt: "You are an expert software engineer."
       tool_policy: "allow" # allow | disallow | auto
       default_temperature: 0.2
       top_p: 0.8
-      max_output_tokens: 1024
-
-    # 原生 Gem (使用在 gemini.google.com 上创建的 Gems)
-    - id: "my-native-gem"
-      model: "gemini-2.0-flash"
-      is_native: true  # 此标志会将 'id' 传递给 Google 服务器
-      default_temperature: 1.0 # 仍然可以在本地覆盖参数
       max_output_tokens: 8192
 ```
 

app/server/chat.py

@@ -108,59 +108,6 @@ def _resolve_gem_from_model_or_400(
     return gem_id, gem, public_model, gem.model
 
 
-def _apply_gem_overrides(
-    request_obj: Any,
-    gem: GemDefinition,
-) -> None:
-    """Apply gem overrides onto a request object in-place.
-
-    Notes:
-        - This helper intentionally does NOT mutate `request_obj.model`.
-        - Upstream model selection is handled via `model=gem:<id>` parsing.
-    """
-
-    if hasattr(request_obj, "temperature"):
-        request_obj.temperature = gem.default_temperature
-
-    if gem.top_p is not None and hasattr(request_obj, "top_p"):
-        request_obj.top_p = gem.top_p
-
-    # ChatCompletions uses `max_tokens`, Responses uses `max_output_tokens`.
-    if hasattr(request_obj, "max_tokens"):
-        request_obj.max_tokens = gem.max_output_tokens
-    if hasattr(request_obj, "max_output_tokens"):
-        request_obj.max_output_tokens = gem.max_output_tokens
-
-    if gem.tool_policy == "disallow":
-        if hasattr(request_obj, "tools"):
-            request_obj.tools = None
-        if hasattr(request_obj, "tool_choice"):
-            request_obj.tool_choice = "none"
-    elif gem.tool_policy == "auto":
-        # Placeholder for future expansion.
-        pass
-    # Default behavior is `allow` (no changes).
-
-
-def _inject_gem_system_prompt(messages: list[Message], system_prompt: str) -> None:
-    """Inject a gem system prompt before model preparation.
-
-    The prompt is inserted as the first system message (or prepended to an
-    existing first system message) so it applies to the whole conversation.
-    """
-
-    if not system_prompt:
-        return
-
-    if messages and messages[0].role == "system" and isinstance(messages[0].content, str):
-        existing = messages[0].content or ""
-        separator = "\n\n" if existing else ""
-        messages[0].content = f"{system_prompt}{separator}{existing}"
-        return
-
-    messages.insert(0, Message(role="system", content=system_prompt))
-
-
 def _build_structured_requirement(
     response_format: dict[str, Any] | None,
 ) -> StructuredOutputRequirement | None:
@@ -704,12 +651,7 @@ async def create_chat_completion(
     gem_id, gem, public_model, actual_model = _resolve_gem_from_model_or_400(request.model)
     logger.info(f"[DEBUG_GEM] Resolved: gem_id={gem_id}, actual_model={actual_model}")
 
-    if gem:
-        _apply_gem_overrides(request, gem)
-        if gem.system_prompt:
-            _inject_gem_system_prompt(request.messages, gem.system_prompt)
-
-    native_gem_id = gem_id if (gem and gem.is_native) else None
+    native_gem_id = gem_id if gem else None
 
     try:
         model = Model.from_name(actual_model)
@@ -982,10 +924,8 @@ async def create_response(
     base_messages, normalized_input = _response_items_to_messages(request_data.input)
 
     gem_id, gem, public_model, actual_model = _resolve_gem_from_model_or_400(request_data.model)
-    if gem:
-        _apply_gem_overrides(request_data, gem)
 
-    native_gem_id = gem_id if (gem and gem.is_native) else None
+    native_gem_id = gem_id if gem else None
 
     structured_requirement = _build_structured_requirement(request_data.response_format)
     if structured_requirement and request_data.stream:
@@ -1029,8 +969,6 @@ async def create_response(
         logger.debug("Image generation support enabled for /v1/responses request.")
 
     preface_messages = _instructions_to_messages(request_data.instructions)
-    if gem and gem.system_prompt:
-        _inject_gem_system_prompt(preface_messages, gem.system_prompt)
 
     conversation_messages = base_messages
     if preface_messages:

app/services/client.py

@@ -81,6 +81,45 @@ async def init(
             logger.exception(f"Failed to initialize GeminiClient {self.id}")
             raise
 
+    async def sync_gems(self) -> None:
+        """Ensure configured gems exist on the server."""
+        from ..utils import g_config
+
+        if not g_config.gemini.gems:
+            return
+
+        try:
+            gem_jar = await self.fetch_gems()
+            # GemJar behaves like a dict-like object where keys are ID and values are Gem objects
+            # Or it might be a list. Let's assume we can iterate it.
+            # To be safe against unknown structure, we'll try to inspect one if possible or just rely on 'title' attribute.
+            
+            # Since we can't easily debug, let's look at the probe output again.
+            # It has 'gems' attribute on client too. Maybe that's cached?
+            
+            existing_titles = set()
+            if gem_jar:
+                # Assuming gem_jar is iterable yielding Gem objects
+                for g in gem_jar:
+                    # Try to find the name/title attribute
+                    title = getattr(g, "title", getattr(g, "name", None))
+                    if title:
+                        existing_titles.add(title)
+
+            for gem_def in g_config.gemini.gems:
+                if gem_def.id not in existing_titles:
+                    logger.info(f"Creating missing gem for client {self.id}: {gem_def.id}")
+                    await self.create_gem(
+                        name=gem_def.id,
+                        prompt=gem_def.system_prompt or "",
+                        description=f"Auto-generated gem for {gem_def.id}",
+                    )
+                else:
+                    logger.debug(f"Gem already exists: {gem_def.id}")
+
+        except Exception as e:
+            logger.error(f"Failed to sync gems for client {self.id}: {e}")
+
     def running(self) -> bool:
         return self._running
 

app/services/pool.py

@@ -45,6 +45,8 @@ async def init(self) -> None:
                         verbose=g_config.gemini.verbose,
                         refresh_interval=g_config.gemini.refresh_interval,
                     )
+                    if client.running():
+                        await client.sync_gems()
                 except Exception:
                     logger.exception(f"Failed to initialize client {client.id}")
 
@@ -97,6 +99,8 @@ async def _ensure_client_ready(self, client: GeminiClientWrapper) -> bool:
                     verbose=g_config.gemini.verbose,
                     refresh_interval=g_config.gemini.refresh_interval,
                 )
+                if client.running():
+                    await client.sync_gems()
                 logger.info(f"Restarted Gemini client {client.id} after it stopped.")
                 return True
             except Exception:

app/utils/config.py

@@ -55,10 +55,6 @@ class GemDefinition(BaseModel):
 
     id: str = Field(..., description="Unique identifier for the gem")
     model: str = Field(..., description="Model name to use when this gem is selected")
-    is_native: bool = Field(
-        default=False,
-        description="If True, this gem ID is passed upstream to Google's servers. If False, it is a local configuration preset.",
-    )
     system_prompt: Optional[str] = Field(
         default=None,
         description="Optional system prompt injected before user messages",