fix: qwen3 - weird token output - reasoning content should not be in completion request (#4983)

* fix: qwen3 - weird token output  - reasoning content should not be in completion request

* chore: bump engine version to llama.cpp b5219

extensions/engine-management-extension/rolldown.config.mjs

@@ -15,7 +15,7 @@ export default defineConfig([
         `http://127.0.0.1:${process.env.CORTEX_API_PORT ?? '39291'}`
       ),
       PLATFORM: JSON.stringify(process.platform),
-      CORTEX_ENGINE_VERSION: JSON.stringify('v0.1.55'),
+      CORTEX_ENGINE_VERSION: JSON.stringify('v0.1.56'),
       DEFAULT_REMOTE_ENGINES: JSON.stringify(engines),
       DEFAULT_REMOTE_MODELS: JSON.stringify(models),
       DEFAULT_REQUEST_PAYLOAD_TRANSFORM: JSON.stringify(
@@ -38,7 +38,7 @@ export default defineConfig([
       file: 'dist/node/index.cjs.js',
     },
     define: {
-      CORTEX_ENGINE_VERSION: JSON.stringify('v0.1.55'),
+      CORTEX_ENGINE_VERSION: JSON.stringify('v0.1.56'),
     },
   },
 ])

extensions/inference-cortex-extension/download.bat

@@ -2,7 +2,7 @@
 set BIN_PATH=./bin
 set SHARED_PATH=./../../electron/shared
 set /p CORTEX_VERSION=<./bin/version.txt
-set ENGINE_VERSION=0.1.55
+set ENGINE_VERSION=0.1.56
 
 @REM Download cortex.llamacpp binaries
 set DOWNLOAD_URL=https://github.com/menloresearch/cortex.llamacpp/releases/download/v%ENGINE_VERSION%/cortex.llamacpp-%ENGINE_VERSION%-windows-amd64

extensions/inference-cortex-extension/download.sh

@@ -2,7 +2,7 @@
 
 # Read CORTEX_VERSION
 CORTEX_VERSION=$(cat ./bin/version.txt)
-ENGINE_VERSION=0.1.55
+ENGINE_VERSION=0.1.56
 CORTEX_RELEASE_URL="https://github.com/menloresearch/cortex.cpp/releases/download"
 ENGINE_DOWNLOAD_URL="https://github.com/menloresearch/cortex.llamacpp/releases/download/v${ENGINE_VERSION}/cortex.llamacpp-${ENGINE_VERSION}"
 CUDA_DOWNLOAD_URL="https://github.com/menloresearch/cortex.llamacpp/releases/download/v${ENGINE_VERSION}"

extensions/inference-cortex-extension/rolldown.config.mjs

@@ -19,7 +19,7 @@ export default defineConfig([
       CORTEX_SOCKET_URL: JSON.stringify(
         `ws://127.0.0.1:${process.env.CORTEX_API_PORT ?? '39291'}`
       ),
-      CORTEX_ENGINE_VERSION: JSON.stringify('v0.1.55'),
+      CORTEX_ENGINE_VERSION: JSON.stringify('v0.1.56'),
     },
   },
   {

web/utils/messageRequestBuilder.ts

@@ -131,11 +131,31 @@ export class MessageRequestBuilder {
     return this
   }
 
+  reasoningTagHandle = (
+    message: ChatCompletionMessage
+  ): ChatCompletionMessageContent => {
+    let content =
+      typeof message.content === 'string'
+        ? message.content
+        : (message.content?.[0]?.text ?? '')
+    // Reasoning content should not be sent to the model
+    if (content.includes('<think>')) {
+      const match = content.match(/<think>([\s\S]*?)<\/think>/)
+      if (match?.index !== undefined) {
+        const splitIndex = match.index + match[0].length
+        content = content.slice(splitIndex).trim()
+      }
+    }
+    return content
+  }
+
   normalizeMessages = (
     messages: ChatCompletionMessage[]
   ): ChatCompletionMessage[] => {
     const stack = new Stack<ChatCompletionMessage>()
     for (const message of messages) {
+      // Handle message content such as reasoning tags
+      message.content = this.reasoningTagHandle(message)
       if (stack.isEmpty()) {
         stack.push(message)
         continue