plugin-custom-llm 1.3.0 → 1.3.1

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 +542,45 @@ 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);
+          }
+        }
+      }
+      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);
+          }
+        }
+        yield chunk;
+      }
+    };
+  }
+  return model;
+}
 class CustomLLMProvider extends import_plugin_ai.LLMProvider {
   get baseURL() {
     return null;
@@ -600,6 +644,7 @@ 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,
@@ -627,7 +672,7 @@ 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": "1.3.1",
   "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 +726,56 @@ function fixEmptyToolProperties(model: any) {
   return model;
 }
 
+/**
+ * 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.
+ */
+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);
+          }
+        }
+      }
+      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;
+      }
+    };
+  }
+
+  return model;
+}
+
 export class CustomLLMProvider extends LLMProvider {
   get baseURL() {
     return null;
@@ -794,6 +855,9 @@ 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, {
@@ -830,7 +894,9 @@ 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)) {