[ai] Fix "start" and "finish" chunk events in DurableAgent

[TooTallNate]

We were always sending start and finish chunk more than necessary. This fixes it

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Preview	Comments	Updated (UTC)
example-nextjs-workflow-turbopack	Ready	Preview	Comment	Nov 20, 2025 10:18pm
example-nextjs-workflow-webpack	Ready	Preview	Comment	Nov 20, 2025 10:18pm
example-workflow	Ready	Preview	Comment	Nov 20, 2025 10:18pm
workbench-express-workflow	Ready	Preview	Comment	Nov 20, 2025 10:18pm
workbench-hono-workflow	Ready	Preview	Comment	Nov 20, 2025 10:18pm
workbench-nitro-workflow	Ready	Preview	Comment	Nov 20, 2025 10:18pm
workbench-nuxt-workflow	Ready	Preview	Comment	Nov 20, 2025 10:18pm
workbench-sveltekit-workflow	Ready	Preview	Comment	Nov 20, 2025 10:18pm
workbench-vite-workflow	Ready	Preview	Comment	Nov 20, 2025 10:18pm
workflow-docs	Ready	Preview	Comment	Nov 20, 2025 10:18pm

[changeset-bot]

🦋 Changeset detected

Latest commit: 7fda66f

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@workflow/ai	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[vercel]

Additional Suggestion:

The removal of the finish message chunk breaks downstream consumers that depend on this message to determine when a stream has completed. This will cause unnecessary reconnection attempts and incorrect behavior in chat applications.

View Details

📝 Patch Details

diff --git a/packages/ai/src/agent/do-stream-step.ts b/packages/ai/src/agent/do-stream-step.ts
index ba04f07..2aa33a5 100644
--- a/packages/ai/src/agent/do-stream-step.ts
+++ b/packages/ai/src/agent/do-stream-step.ts
@@ -73,6 +73,9 @@ export async function doStreamStep(
           controller.enqueue({
             type: 'finish-step',
           });
+          controller.enqueue({
+            type: 'finish',
+          });
         },
         transform: async (part, controller) => {
           const partType = part.type;

Analysis

Missing 'finish' chunk in doStreamStep breaks WorkflowChatTransport stream completion detection

What fails: doStreamStep() in packages/ai/src/agent/do-stream-step.ts removed the {type: 'finish'} chunk from its flush() method, causing downstream consumers like WorkflowChatTransport to incorrectly trigger reconnection logic when the stream completes normally.

How to reproduce:

1. Create a DurableAgent and call agent.stream() which uses doStreamStep()
2. Wire the output stream to a backend HTTP response handler that serializes it
3. Use WorkflowChatTransport on the client to consume the response stream
4. Expected: Stream completes normally with onChatEnd callback
5. Actual: Without 'finish' chunk, the gotFinish flag is never set, causing reconnectToStream() to be called and onChatEnd callback never fires

Code evidence:

• doStreamStep() transforms LLM stream parts to UIMessageChunks via a TransformStream
• The flush() method at packages/ai/src/agent/do-stream-step.ts:72-76 was modified to only emit 'finish-step'
• WorkflowChatTransport.sendMessagesIterator() at packages/ai/src/workflow-chat-transport.ts:232-241 explicitly checks for chunk.value.type === 'finish' to set gotFinish = true
• Without this check passing, line 248 triggers reconnection logic instead of completing
• Per AI SDK stream protocol documentation, the stream protocol includes a final 'finish' chunk

Verification: Created test scenario showing stream without 'finish' chunk triggers 2 fetch calls (POST + reconnection GET), while stream with 'finish' chunk triggers only 1 call (POST only)

Fix applied: Restored the controller.enqueue({type: 'finish'}) call to the flush() method in the TransformStream, ensuring the 'finish' chunk is always emitted when the stream ends.