[valstu]

What

A page that didn't start an answer can show it as it is written: a second tab, a page reloaded mid-answer, and the answer to a signal, a held message or a schedule. Optional and labelled experimental. Without liveStreams on the runtime nothing changes.

• liveStreams on the runtime (LiveStreams): publish / subscribe keep a running answer's stream where another request can read it; optional notify / listen carry per-chat events. Every turn's stream is copied into it under runId:fencingToken.
• memoryLiveStreams() in core, for one long-running server.
• Redis is a recipe in the guide, not a package: about 30 lines over resumable-stream and two pub/sub commands.
• runtime.streamChat / GET chats/:id/stream: the running answer from its start, as useChat reads it on resume. 204 when nothing is running.
• runtime.watchChat / GET chats/:id/events: server-sent events (hello, started, ended, changed), so a page needs no poll loop. 404 when the store can't notify.
• Docs: new guide guides/live-updates, reference and route tables, pointers from the queue, signals, chat UI, streaming and Vercel pages. Changeset: patch for core and server.

Tested

• packages/core/src/live-stream.test.ts: each case against the in-memory store and against resumable-stream over a stand-in pub/sub. Second reader gets the answer from its start; a held turn's answer is readable; events for started / ended / changed; a failed publish doesn't stop the answer.
• apps/example/src/live-stream.smoke.ts (pnpm --filter @intentface/latch-example smoke:live-stream, needs a Redis at REDIS_URL, default redis://127.0.0.1:6390): two runtimes with their own connections to a real Valkey and one database, three tabs driven by the AI SDK's chat client (AbstractChat + DefaultChatTransport). A tab on the other instance shows the first half of an answer before it is finished; the sending tab, also watching, shows each message once; a tab opened mid-answer catches up; a signal's answer is read live by all three. Not run in CI.

Not tested

• React itself: useChat and EventSource in a browser. The guide's snippet follows the logic the smoke runs, but has not been run.
• A real serverless host (function duration cut on the events connection, concurrency).
• Crash recovery: resuming a stream when the history already holds part of that answer.
• Held messages over real Redis (covered by the unit tests only).

Known limits

• Events sent while a page reconnects are not replayed; hello covers a running answer.
• If the instance writing an answer dies, its stream ends with it.
• Each open chat holds one request open.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CN7rKG2PbHZnCtCWmFAP2w

[vercel]

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

Project	Deployment	Actions	Updated
intentface-latch-docs	Ready	Preview	Nov 13, 2026 5:30pm UTC

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts