Problem

We stream the LLM output server-side to multiple consumers (several users viewing the same thread). But the user prompt is never put on the stream — it's only written to the DB. So a second viewer of a shared thread doesn't see the other user's message until a DB refetch happens (reconnect or a thread.status event), which feels laggy on a live thread.

Fix

Publish the materialized request message onto the same per-thread NATS JetStream subject as a transient data-user-message control chunk, at POST time, before the run's assistant chunks (so it sorts first).

Key point — large prompts are handled for free: the publish goes through the existing publishRawChunk → serializeChunk path, which already fragments anything over MAX_PUBLISH_BYTES (768 KiB) and drops over MAX_CHUNKED_BYTES (32 MiB). File attachments are already offloaded to object-storage URLs by uploadFileParts, so the payload is essentially text. No new large-payload machinery — this reuses the harness stream codec.

Server

• user-message-stream.ts (new) — buildUserMessageChunk / isUserMessageControlChunk / publishUserMessage, mirroring run-status-stage.ts. Publish is best-effort (never fails the POST).
• routes.ts — publish right after emitRequestMessage, in the same shouldPersistRequestMessage gate.
• project-chunks.ts — filter the chunk out of the durable assistant projection. fenceFilter already drops it (it's fence-less), but the explicit type filter matches the data-run-status precedent and future-proofs against it ever being published with a run fence.

Client

• thread-connection.ts — intercept data-user-message in handleChunk before the AI-SDK fold (the reassembler hardcodes role:"assistant"), and route it into the existing applyLocalMessage → mergeAndSort/upsertById path. Dedup is free: the author's optimistic row shares the same id, and a later DB refetch reconciles created_at via preserveCreatedAt.

Testing

• New unit tests for the helpers (user-message-stream.test.ts).
• Extended project-chunks.test.ts to assert the chunk never leaks into projected assistant parts while normal assistant text still folds.
• bun test (affected), tsc --noEmit (0 errors), bun run lint (0 errors) all pass.

Not covered / follow-up

Live two-viewer end-to-end wasn't driven headlessly here — worth a manual check (or an e2e in packages/e2e) with two sessions on one thread to confirm the prompt renders live for the non-author.

Summary by cubic

Mirror the user's prompt onto the per-thread run stream so other viewers see it instantly, fixing lag in shared threads. Uses a transient data-user-message control chunk and keeps the durable projection unchanged.

• Bug Fixes
  • Server: publish a best-effort data-user-message chunk at POST time (before assistant chunks) via existing publishRawChunk fragmentation; filter it out of the assistant projection; gated with the same shouldPersistRequestMessage check.
  • Client: intercept data-user-message before the assistant fold and upsert via applyLocalMessage so the author's optimistic message dedupes and other viewers see it live.

^{Written for commit 4f24dff. Summary will update on new commits.}