Reviews keep hitting the 50 minute timeout. The coordinator agent that was
supposed to launch reviewers, watch the clock, stop stragglers, validate and
write the result does none of that reliably: reviewers get stopped and lose
their findings, foreground validators can't be stopped, deadlines get
overshot by minutes, and one leftover agent keeps the process alive until
the job fails. Piling more rules into its prompt hasn't helped.
Do the orchestration in the shell instead. Run one bounded claude invocation
per review lens in parallel under timeout(1), each returning its findings as
structured output, merge them with jq and run one bounded validator over the
candidates. A lens that times out is just reported as failed. The post job
builds the tracking comment from what it actually posted, so the model no
longer authors the summary or the review bookkeeping. Agents run from an
empty directory with all inputs outside it, so the sandbox keeps both the
checkout and the inputs read-only.
Daan De Meyer
c549d00e9f