fix: use user message instead of assistant prefill for continuation retry (#3027)

## Summary - Non-Anthropic providers (OpenAI, Google, Ollama, etc.) error with "The model does not support assistant message prefill" when the unclosed `dyad-write` continuation retry triggers - The continuation logic was using an assistant message as the last message (prefill pattern), which only Anthropic supports - Replaced with a user message that includes the partial response and asks the model to continue where it left off ## Test plan - Use a non-Anthropic provider (e.g. OpenAI, Gemini) in build mode - Trigger a response that produces an unclosed `<dyad-write>` tag (e.g. a large file write that gets cut off) - Verify the continuation retry no longer errors and the model completes the response 🤖 Generated with [Claude Code](https://claude.com/claude-code)  --- <a href="https://app.devin.ai/review/dyad-sh/dyad/pull/3027" target="_blank"> <picture> <source media="(prefers-color-scheme: dark)" srcset="https://static.devin.ai/assets/gh-open-in-devin-review-dark.svg?v=1"> <img src="https://static.devin.ai/assets/gh-open-in-devin-review-light.svg?v=1" alt="Open with Devin"> </picture> </a>  --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

fix: use user message instead of assistant prefill for continuation retry (#3027)
dd5dcda6 · Will Chen · GitHub · 4d7fa151 · dd5dcda6 · dd5dcda6
--- a/e2e-tests/snapshots/partial_response.spec.ts_partial-message-is-resumed-1.txt
+++ b/e2e-tests/snapshots/partial_response.spec.ts_partial-message-is-resumed-1.txt
@@ -132,3 +132,8 @@ role: assistant
 message: START OF MESSAGE
 <dyad-write path="src/new-file.ts" description="this file will be partially written">
 const a = "[[STRING_TO_BE_FINISHED]]
+===
+role: user
+message: Your previous response did not finish completely. Continue exactly where you left off without any preamble.
\ No newline at end of file
--- a/src/ipc/handlers/chat_stream_handlers.ts
+++ b/src/ipc/handlers/chat_stream_handlers.ts
@@ -1453,10 +1453,18 @@ ${formattedSearchReplaceIssues}`,
              continuationAttempts++;
              const { fullStream: contStream } = await simpleStreamText({
-                // Build messages: replay history then pre-fill assistant with current partial.
+                // Build messages: replay history, then ask the model to continue from the partial response.
                chatMessages: [
                  ...chatMessages,
-                  { role: "assistant", content: fullResponse },
+                  {
+                    role: "assistant",
+                    content: fullResponse,
+                  },
+                  {
+                    role: "user",
+                    content:
+                      "Your previous response did not finish completely. Continue exactly where you left off without any preamble.",
+                  },
                ],
                modelClient,
                files: files,

--- a/testing/fake-llm-server/chatCompletionHandler.ts
+++ b/testing/fake-llm-server/chatCompletionHandler.ts
@@ -59,7 +59,8 @@ export const createChatCompletionHandler =
    if (
      !localAgentFixture &&
      (userTextContent.includes("incomplete todo(s)") ||
-        userTextContent.includes("previous response stream was interrupted"))
+        userTextContent.includes("previous response stream was interrupted") ||
+        userTextContent.includes("did not finish completely"))
    ) {
      for (const msg of userMessages) {
        const textContent = getTextContent(msg);
@@ -334,11 +335,13 @@ export default Index;
      }
    }
+    // Continuation requests: the partial assistant output is in a preceding assistant
+    // message, then a user message asks to continue ("did not finish completely").
+    // Check any message for the marker. See chat_stream_handlers continuation prompt.
    if (
-      lastMessage &&
+      messages.some((m: any) =>
-      lastMessage.content &&
+        getTextContent(m).includes("[[STRING_TO_BE_FINISHED]]"),
-      typeof lastMessage.content === "string" &&
+      )
-      lastMessage.content.trim().endsWith("[[STRING_TO_BE_FINISHED]]")
    ) {
      messageContent = `[[STRING_IS_FINISHED]]";</dyad-write>\nFinished writing file.`;
      messageContent += "\n\n" + generateDump(req);