feat(chat): reconnect via WorkflowChatTransport instead of hand-rolled recovery

Merged
SW
4mo ago
+3573-856

Row 1 of the chat#1923 plan — the fix for the UI stall.

Replaces our client-side stream-recovery machinery with @workflow/ai's WorkflowChatTransport, which owns reconnection natively.

Why the transport, and not one more fix

The transport is a single generator:

for await (const chunk of postStream) { chunkIndex++; yield chunk; if (chunk.type === 'finish') gotFinish = true; }
if (!gotFinish) yield* this.reconnectToStreamIterator(options, runId, chunkIndex);
// …which loops: while (!gotFinish) { GET …/stream?startIndex=chunkIndex; … }

One continuous stream reaches useChat, so the turn settles exactly when finish arrives.

Our version tracked that same state across React refs — an in-flight flag, an edge-detected trigger, a probe schedule, a terminal-state branch — and every fix exposed another unverified state:

AttemptFailure
trigger on useChat statusfired once, never again
flag held across await resumeStream()blocked recovery for the whole ~120 s connection
final stream endfired nothing at all — 2 of 3 preview runs passed, the third hung after 9 recoveries

There is no trigger to miss here, and no ref to hold.

Deleted — 7 files

createChunkCountingFetch, buildStreamReconnectUrl, shouldRecoverStalledStream, useStreamRecovery, their tests, and the hook's wiring in useVercelChat. startIndex is no longer ours to compute.

Kept, because the SDK cannot know it

  • auth + sessionId / chatId via prepareSendMessagesRequest
  • the reconnect URL repointed at /api/chat/{chatId}/stream — the SDK defaults to keying by run id
  • a 204 override: our api answers "nothing to resume" with a bodyless 204, which the SDK surfaces as Failed to fetch chat: 204 and counts toward maxConsecutiveErrors. Mirrors upstream open-agents' wrapper.

Version choice

@workflow/ai@4.2.0 is stable and carries newer transport code than the 5.x beta upstream pins — I diffed them, and 4.2.0 has the mid-part-resume orphan filter (createOrphanFilter) that 5.0.0-beta.4 lacks. That matters here because we resume at arbitrary chunk indices.

Lockfile resolved with default resolution: 9 version changes, viem and everything else unmoved. --config.resolution-mode=lowest-direct was tried and rejected — it downgraded viem 2.52.0 → 2.23.2 and broke the coinbase types. "Lowest" is not "least change".

Checks

356 tests pass, tsc --noEmit clean, next build compiles.

Preview verification to follow in a comment. The gate is the one that has never held: reconnects chain through every stream end, and the turn ends coherently — across 3 runs.


Summary by cubic

Switches chat reconnection to @workflow/ai’s WorkflowChatTransport to stop UI stalls on long runs by resuming streams until the finish chunk. Removes custom recovery code and aligns reconnects with our API.

  • Refactors
    • Added createWorkflowChatTransport and switched useChatTransport; removed useStreamRecovery and related helpers/tests.
    • Reconnects call /api/chat/{chatId}/stream with auth; startIndex is computed by the transport and it streams until finish; treat 204 as “nothing to resume”.
    • Injected sessionId, chatId, and recoupAccessToken via prepareSendMessagesRequest.
    • Added @workflow/ai@4.2.0; lockfile updated.

Summary by CodeRabbit

  • New Features

    • Improved chat connection handling with automatic stream resumption.
    • Added secure session and access-token handling for chat requests and reconnects.
    • Added a single retry for reconnects caused by expired authentication.
    • Gracefully handles completed streams during reconnection.
  • Bug Fixes

    • Improved recovery from interrupted or incomplete chat responses.
    • Prevents unnecessary reconnect attempts when a stream has ended normally.