A Retry Button, and the Case of the Review That Never Died

A customer wrote in this week with a frustrating one: his pull request review failed with "Timed out after 3 attempts," the dashboard showed a single hour-long failed run, and there was no way to try again short of pushing a throwaway commit. Three complaints in one email — and he was right on all three counts. Here's what we found and what shipped.

The review that never died

When we pulled the logs, the surprise was that his review had been perfectly healthy. It ran Claude, Gemini, and Grok side by side, and on a substantial PR, Grok's single big request can quietly work for fifteen-plus minutes without saying anything. Our stuck-job watchdog — the safety net that rescues reviews when a worker genuinely dies — took that silence for death, killed the working review, started it again, waited, killed it again, and after the third round reported a timeout. The hour-long "failed run" on the dashboard was really three healthy attempts stacked on top of each other, each one executed by an overzealous safety net.

The fix is boring in the best way: long-running model calls now check in every sixty seconds, so the watchdog can tell "working quietly" from "actually dead." It shipped the same day. The watchdog itself stays — it has rescued plenty of genuinely stuck reviews — it just no longer shoots the patient.

The Retry button

The second half of the complaint was just as fair: when a review does fail, you shouldn't need to invent a commit to get another one. Two fixes there:

• Every failed review now shows a Retry button on the Reviews dashboard. One click re-runs the review against the PR exactly as it stands. It uses the same machinery as the @codemouse review PR comment, so rapid duplicates collapse and you're never billed twice for the same head.
• Every failure now explains itself on the PR. A review that fails — provider out of credit, model retired, anything — leaves a comment with the reason in plain words and the retry instructions right there. No more silent deaths.

Also fixed while we were in there

Google retired an older Gemini model this week with a "no longer available" error, which briefly broke reviews for anyone who still had it selected. CodeMouse now maps retired Gemini models to their current successors automatically — your stored choice keeps working even when a provider renames the furniture.

All of this came from one support message. If something about CodeMouse annoys you, [email protected] goes straight to a human who can ship the fix — this week's average time from complaint to deploy was a few hours.

FAQ

Where is the Retry button?

Dashboard → Reviews. Any run marked failed shows a Retry link next to its status. Clicking it queues a fresh review of the PR at its current commit.

Does retrying cost anything extra?

Just the normal model usage on your own API key for the new review — same as any review. CodeMouse adds no markup and never bills for the failed attempt's unfinished work.

What if the review fails again?

The failure comment on the PR tells you why in plain words (for example, a provider account out of credits) and how to retry once you've fixed it. If the reason doesn't make sense, email [email protected] and a human will dig into the logs.

Do I still need to push a commit to trigger a re-review?

No — and you never did need a real change: the Retry button or a @codemouse review comment on the PR both work without touching your branch.