plugin-custom-llm 1.3.0 → 1.3.2

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

@@ -66,6 +66,11 @@ function getChatOpenAICompletions() {
   }
   return _ChatOpenAICompletions;
 }
+function sanitizeToolCallId(id) {
+  if (!id || typeof id !== "string") return id;
+  const idx = id.indexOf("__thought__");
+  return idx !== -1 ? id.substring(0, idx) : id;
+}
 function getToolCallsKey(toolCalls = []) {
   return toolCalls.map((tc) => {
     var _a;
@@ -537,6 +542,52 @@ function fixEmptyToolProperties(model) {
   };
   return model;
 }
+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;
+}
+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);
+        }
+      }
+    }
+  };
+}
 class CustomLLMProvider extends import_plugin_ai.LLMProvider {
   get baseURL() {
     return null;
@@ -574,7 +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,
@@ -627,7 +679,7 @@ class CustomLLMProvider extends import_plugin_ai.LLMProvider {
       workContext
     };
     if (toolCalls) {
-      content.tool_calls = toolCalls;
+      content.tool_calls = Array.isArray(toolCalls) ? toolCalls.map((tc) => ({ ...tc, id: sanitizeToolCallId(tc.id) })) : toolCalls;
     }
     if (Array.isArray(content.content)) {
       const textBlocks = content.content.filter((block) => block.type === "text");

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": "1.3.0",
+  "version": "1.3.2",
   "main": "dist/server/index.js",
   "files": [
     "dist",

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

@@ -50,6 +50,17 @@ function getChatOpenAICompletions() {
   return _ChatOpenAICompletions;
 }
 
+/**
+ * Sanitize a tool call ID by stripping the `__thought__<base64>` suffix
+ * that Gemini models append during streaming. The suffix is excessively
+ * long and causes errors when langgraph reads messages back from history.
+ */
+function sanitizeToolCallId(id: string | undefined): string | undefined {
+  if (!id || typeof id !== 'string') return id;
+  const idx = id.indexOf('__thought__');
+  return idx !== -1 ? id.substring(0, idx) : id;
+}
+
 /**
  * Build tool_calls key for reasoning content map lookup.
  */
@@ -715,6 +726,69 @@ function fixEmptyToolProperties(model: any) {
   return model;
 }
 
+/**
+ * Sanitize all tool call IDs in a ChatResult (used after _generate).
+ */
+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;
+}
+
+/**
+ * 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;
+}
+
+/**
+ * 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 {
   get baseURL() {
     return null;
@@ -755,7 +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,
@@ -830,7 +908,9 @@ export class CustomLLMProvider extends LLMProvider {
     };
 
     if (toolCalls) {
-      content.tool_calls = toolCalls;
+      content.tool_calls = Array.isArray(toolCalls)
+        ? toolCalls.map((tc: any) => ({ ...tc, id: sanitizeToolCallId(tc.id) }))
+        : toolCalls;
     }
 
     if (Array.isArray(content.content)) {