plugin-custom-llm 1.3.1 → 1.3.2

package/dist/server/llm-providers/custom-llm.js

@@ -542,44 +542,51 @@ function fixEmptyToolProperties(model) {
   };
   return model;
 }
-function wrapWithToolCallIdSanitizer(model) {
-  var _a, _b;
-  const originalGenerate = (_a = model._generate) == null ? void 0 : _a.bind(model);
-  if (originalGenerate) {
-    model._generate = async function(...args) {
-      const result = await originalGenerate(...args);
-      for (const gen of (result == null ? void 0 : result.generations) ?? []) {
-        const msg = gen == null ? void 0 : gen.message;
-        if (msg == null ? void 0 : msg.tool_calls) {
-          for (const tc of msg.tool_calls) {
-            tc.id = sanitizeToolCallId(tc.id);
-          }
-        }
+function sanitizeGenerateResult(result) {
+  if (!result) return result;
+  for (const gen of (result == null ? void 0 : result.generations) ?? []) {
+    const msg = gen == null ? void 0 : gen.message;
+    if (msg == null ? void 0 : msg.tool_calls) {
+      for (const tc of msg.tool_calls) {
+        tc.id = sanitizeToolCallId(tc.id);
       }
-      return result;
-    };
+    }
   }
-  const streamMethod = typeof model._streamResponseChunks === "function" ? "_streamResponseChunks" : "_stream";
-  const originalStream = (_b = model[streamMethod]) == null ? void 0 : _b.bind(model);
-  if (originalStream) {
-    model[streamMethod] = async function* (...args) {
-      for await (const chunk of originalStream(...args)) {
-        const msg = chunk == null ? void 0 : chunk.message;
-        if (msg == null ? void 0 : msg.tool_call_chunks) {
-          for (const tc of msg.tool_call_chunks) {
-            tc.id = sanitizeToolCallId(tc.id);
-          }
-        }
-        if (msg == null ? void 0 : msg.tool_calls) {
-          for (const tc of msg.tool_calls) {
-            tc.id = sanitizeToolCallId(tc.id);
-          }
+  return result;
+}
+function sanitizeStreamChunk(chunk) {
+  const msg = chunk == null ? void 0 : chunk.message;
+  if (msg == null ? void 0 : msg.tool_call_chunks) {
+    for (const tc of msg.tool_call_chunks) {
+      tc.id = sanitizeToolCallId(tc.id);
+    }
+  }
+  if (msg == null ? void 0 : msg.tool_calls) {
+    for (const tc of msg.tool_calls) {
+      tc.id = sanitizeToolCallId(tc.id);
+    }
+  }
+  return chunk;
+}
+function createSanitizedChatClass(BaseClass) {
+  return class SanitizedChatModel extends BaseClass {
+    async _generate(messages, options, runManager) {
+      const result = await super._generate(messages, options, runManager);
+      return sanitizeGenerateResult(result);
+    }
+    async *_streamResponseChunks(messages, options, runManager) {
+      for await (const chunk of super._streamResponseChunks(messages, options, runManager)) {
+        yield sanitizeStreamChunk(chunk);
+      }
+    }
+    async *_stream(messages, options, runManager) {
+      if (typeof super._stream === "function") {
+        for await (const chunk of super._stream(messages, options, runManager)) {
+          yield sanitizeStreamChunk(chunk);
         }
-        yield chunk;
       }
-    };
-  }
-  return model;
+    }
+  };
 }
 class CustomLLMProvider extends import_plugin_ai.LLMProvider {
   get baseURL() {
@@ -618,7 +625,8 @@ class CustomLLMProvider extends import_plugin_ai.LLMProvider {
     if (reqConfig.extraBody && typeof reqConfig.extraBody === "object") {
       Object.assign(modelKwargs, reqConfig.extraBody);
     }
-    const ChatClass = enableReasoning ? createReasoningChatClass() : getChatOpenAI();
+    const BaseChatClass = enableReasoning ? createReasoningChatClass() : getChatOpenAI();
+    const ChatClass = createSanitizedChatClass(BaseChatClass);
     const config = {
       apiKey,
       ...this.modelOptions,
@@ -644,7 +652,6 @@ class CustomLLMProvider extends import_plugin_ai.LLMProvider {
     }
     let model = new ChatClass(config);
     model = fixEmptyToolProperties(model);
-    model = wrapWithToolCallIdSanitizer(model);
     if (streamKeepAlive && !disableStream) {
       return wrapWithStreamKeepAlive(model, {
         intervalMs: Number(keepAliveIntervalMs) || 5e3,

package/package.json

@@ -3,7 +3,7 @@
   "displayName": "AI LLM: Custom (OpenAI Compatible)",
   "displayName.zh-CN": "AI LLM：自定义（OpenAI 兼容）",
   "description": "OpenAI-compatible LLM provider with auto response format detection for external LLM services.",
-  "version": "1.3.1",
+  "version": "1.3.2",
   "main": "dist/server/index.js",
   "files": [
     "dist",

package/src/server/llm-providers/custom-llm.ts

@@ -727,53 +727,66 @@ function fixEmptyToolProperties(model: any) {
 }
 
 /**
- * Wrap a chat model to sanitize tool call IDs in outputs.
- * Gemini models can return IDs like `call_xxx__thought__<long_base64>`
- * which are too long for langgraph to handle on message replay.
- * This strips the `__thought__...` suffix at the model output level
- * so downstream code (convertAIMessage, etc.) only sees clean IDs.
+ * Sanitize all tool call IDs in a ChatResult (used after _generate).
  */
-function wrapWithToolCallIdSanitizer(model: any) {
-  // Patch _generate (used by invoke / non-streaming)
-  const originalGenerate = model._generate?.bind(model);
-  if (originalGenerate) {
-    model._generate = async function (...args: any[]) {
-      const result = await originalGenerate(...args);
-      for (const gen of result?.generations ?? []) {
-        const msg = gen?.message;
-        if (msg?.tool_calls) {
-          for (const tc of msg.tool_calls) {
-            tc.id = sanitizeToolCallId(tc.id);
-          }
-        }
+function sanitizeGenerateResult(result: any): any {
+  if (!result) return result;
+  for (const gen of result?.generations ?? []) {
+    const msg = gen?.message;
+    if (msg?.tool_calls) {
+      for (const tc of msg.tool_calls) {
+        tc.id = sanitizeToolCallId(tc.id);
       }
-      return result;
-    };
+    }
   }
+  return result;
+}
 
-  // Patch _streamResponseChunks or _stream (used by streamEvents / streaming)
-  const streamMethod = typeof model._streamResponseChunks === 'function' ? '_streamResponseChunks' : '_stream';
-  const originalStream = model[streamMethod]?.bind(model);
-  if (originalStream) {
-    model[streamMethod] = async function* (...args: any[]) {
-      for await (const chunk of originalStream(...args)) {
-        const msg = chunk?.message;
-        if (msg?.tool_call_chunks) {
-          for (const tc of msg.tool_call_chunks) {
-            tc.id = sanitizeToolCallId(tc.id);
-          }
-        }
-        if (msg?.tool_calls) {
-          for (const tc of msg.tool_calls) {
-            tc.id = sanitizeToolCallId(tc.id);
-          }
-        }
-        yield chunk;
-      }
-    };
+/**
+ * Sanitize tool call IDs in a streaming chunk.
+ */
+function sanitizeStreamChunk(chunk: any): any {
+  const msg = chunk?.message;
+  if (msg?.tool_call_chunks) {
+    for (const tc of msg.tool_call_chunks) {
+      tc.id = sanitizeToolCallId(tc.id);
+    }
+  }
+  if (msg?.tool_calls) {
+    for (const tc of msg.tool_calls) {
+      tc.id = sanitizeToolCallId(tc.id);
+    }
   }
+  return chunk;
+}
 
-  return model;
+/**
+ * Create a subclass of the given ChatModel class that sanitizes tool call IDs
+ * in all outputs. Gemini models return IDs like `call_xxx__thought__<long_base64>`
+ * which are too long for langgraph. Using class-level overrides (instead of
+ * instance patching) ensures the sanitization survives bindTools/RunnableBinding.
+ */
+function createSanitizedChatClass(BaseClass: any) {
+  return class SanitizedChatModel extends BaseClass {
+    async _generate(messages: any[], options: any, runManager?: any) {
+      const result = await super._generate(messages, options, runManager);
+      return sanitizeGenerateResult(result);
+    }
+
+    async *_streamResponseChunks(messages: any[], options: any, runManager?: any) {
+      for await (const chunk of super._streamResponseChunks(messages, options, runManager)) {
+        yield sanitizeStreamChunk(chunk);
+      }
+    }
+
+    async *_stream(messages: any[], options: any, runManager?: any) {
+      if (typeof super._stream === 'function') {
+        for await (const chunk of super._stream(messages, options, runManager)) {
+          yield sanitizeStreamChunk(chunk);
+        }
+      }
+    }
+  };
 }
 
 export class CustomLLMProvider extends LLMProvider {
@@ -816,7 +829,11 @@ export class CustomLLMProvider extends LLMProvider {
     // Issue #4: Use ReasoningChatOpenAI when enableReasoning is set.
     // This ensures reasoning_content is preserved and patched back into
     // assistant messages during tool call round-trips (required by DeepSeek-R1, etc.)
-    const ChatClass = enableReasoning ? createReasoningChatClass() : getChatOpenAI();
+    // Wrap with tool call ID sanitizer at the class level — ensures
+    // __thought__<base64> suffixes from Gemini are stripped in all code paths
+    // (invoke, stream, bindTools bindings) via prototype chain.
+    const BaseChatClass = enableReasoning ? createReasoningChatClass() : getChatOpenAI();
+    const ChatClass = createSanitizedChatClass(BaseChatClass);
     const config: Record<string, any> = {
       apiKey,
       ...this.modelOptions,
@@ -855,9 +872,6 @@ export class CustomLLMProvider extends LLMProvider {
     // Fix empty tool properties for strict providers (Gemini, etc.)
     model = fixEmptyToolProperties(model);
 
-    // Sanitize Gemini's __thought__<base64> suffixes in tool call IDs
-    model = wrapWithToolCallIdSanitizer(model);
-
     // Wrap with keepalive proxy if enabled (and streaming is not disabled)
     if (streamKeepAlive && !disableStream) {
       return wrapWithStreamKeepAlive(model, {