[jamescbury]

roughdraft open waits for Done Reviewing with a single unbounded long-poll to /api/review-events/watch. Node's built-in fetch (undici) applies a default 300-second headersTimeout, so when a reviewer takes more than ~5 minutes the CLI crashes with TypeError: fetch failed / UND_ERR_HEADERS_TIMEOUT. The server also clamps its own wait to 300s in review-events.ts, so an unbounded request was never going to be honored anyway.

What this changes

• runWatch now polls in bounded 240-second chunks — safely under both undici's 300s headers timeout and the server's own 300s clamp — with a per-request abort guard.
• Each poll resumes from the previous response's nextSequence (sent as afterSequence), so a Done Reviewing event that lands between polls is never missed; the server already retains the last 100 events for exactly this.
• Transient fetch failures now retry (up to 5 consecutive, 2 seconds apart) instead of crashing the CLI mid-review.
• --timeout semantics are unchanged, and overall timeouts longer than 300s now actually work (previously the server silently capped them).

Testing

• New tests reproduce the original crash (a UND_ERR_HEADERS_TIMEOUT-style fetch failure mid-watch now retries and completes instead of crashing), and cover the bounded-poll contract, afterSequence resumption across a server-side poll timeout, and --timeout expiry.
• pnpm check and pnpm test:smoke pass.
• Live verification: roughdraft open sat on "Waiting for Done Reviewing..." for 6 minutes 40 seconds — past the old 300s crash point — then received the Done Reviewing event and exited 0.

Fixes #127, fixes #108, fixes #93. I believe this also resolves #113 (the same 5-minute exit, attributed there to an SSE timeout), and it may explain some of the reports in #119.

🤖 Generated with Claude Code

[vercel]

@jamescbury is attempting to deploy a commit to the Lex Team on Vercel.

A member of the Team first needs to authorize it.