Escape dyad tags inside thinking blocks (#229)

src/ipc/handlers/chat_stream_handlers.ts

@@ -215,6 +215,7 @@ export function registerChatStreamHandlers() {
           abortController,
           updatedChat,
         );
+        fullResponse = cleanThinkingByEscapingDyadTags(fullResponse);
       } else {
         // Normal AI processing for non-test prompts
         const settings = readSettings();
@@ -348,7 +349,10 @@ This conversation includes one or more image attachments. When the user uploads
           ...codebasePrefix,
           ...limitedMessageHistory.map((msg) => ({
             role: msg.role as "user" | "assistant" | "system",
-            content: msg.content,
+            // Why remove thinking tags?
+            // Thinking tags are generally not critical for the context
+            // and eats up extra tokens.
+            content: removeThinkingTags(msg.content),
           })),
         ];
 
@@ -411,6 +415,7 @@ This conversation includes one or more image attachments. When the user uploads
         try {
           for await (const textPart of textStream) {
             fullResponse += textPart;
+            fullResponse = cleanThinkingByEscapingDyadTags(fullResponse);
             if (
               fullResponse.includes("$$SUPABASE_CLIENT_CODE$$") &&
               updatedChat.app?.supabaseProjectId
@@ -707,3 +712,27 @@ async function prepareMessageWithAttachments(
     content: contentParts,
   };
 }
+
+function cleanThinkingByEscapingDyadTags(text: string): string {
+  // Extract content inside <think> </think> tags
+  const thinkRegex = /<think>([\s\S]*?)<\/think>/g;
+
+  return text.replace(thinkRegex, (match, content) => {
+    // We are replacing the opening tag with a look-alike character
+    // to avoid issues where thinking content includes dyad tags
+    // and are mishandled by:
+    // 1. FE markdown parser
+    // 2. Main process response processor
+    const processedContent = content
+      .replace(/<dyad/g, "＜dyad")
+      .replace(/<\/dyad/g, "＜/dyad");
+
+    // Return the modified think tag with processed content
+    return `<think>${processedContent}</think>`;
+  });
+}
+
+function removeThinkingTags(text: string): string {
+  const thinkRegex = /<think>([\s\S]*?)<\/think>/g;
+  return text.replace(thinkRegex, "").trim();
+}

testing/fake-llm-server/index.ts

@@ -32,9 +32,28 @@ function createStreamChunk(
   return `data: ${JSON.stringify(chunk)}\n\n${isLast ? "data: [DONE]\n\n" : ""}`;
 }
 
+const CANNED_MESSAGE = `
+  <think>
+  \`<dyad-write>\`:
+  I'll think about the problem and write a bug report.
+
+  <dyad-write>
+
+  <dyad-write path="file1.txt">
+  Fake dyad write
+  </dyad-write>
+  </think>
+  
+  <dyad-write path="file1.txt">
+  A file (2)
+  </dyad-write>
+  More
+  EOM`;
+
 // Handle POST requests to /v1/chat/completions
 app.post("/v1/chat/completions", (req, res) => {
   const { stream = false, messages = [] } = req.body;
+  console.log("* Received messages", messages);
 
   // Check if the last message contains "[429]" to simulate rate limiting
   const lastMessage = messages[messages.length - 1];
@@ -61,7 +80,7 @@ app.post("/v1/chat/completions", (req, res) => {
           index: 0,
           message: {
             role: "assistant",
-            content: "hello world",
+            content: CANNED_MESSAGE,
           },
           finish_reason: "stop",
         },
@@ -75,7 +94,7 @@ app.post("/v1/chat/completions", (req, res) => {
   res.setHeader("Connection", "keep-alive");
 
   // Split the "hello world" message into characters to simulate streaming
-  const message = "hello world";
+  const message = CANNED_MESSAGE;
   const messageChars = message.split("");
 
   // Stream each character with a delay
@@ -94,7 +113,7 @@ app.post("/v1/chat/completions", (req, res) => {
       clearInterval(interval);
       res.end();
     }
-  }, 100);
+  }, 10);
 });
 
 // Start the server