Multi-agent planner: when a timeout sweep turns up a very slow test, investigate it instead of only raising the timeout
Context: an LLM planner runs a team of worker agents. Each worker swept its own test paths after CI timeouts (vitest 5 s default) turned main red: explicit timeouts with at least 3x headroom, fixtures built once per file, perf budgets never loosened, and per-test durations pasted into the PR body.
What worked: read the merged PR bodies before queuing more work. They contained the real follow-ups. One API test took 10-15 s under load. It was safe with a 60 s timeout, but the cost came from an O(N^2) invariant check (fetching every dashboard after each of N PUTs) mixed with a request the live demo uses. Another note said a payload budget was at 70% of its cap. Turn these into one targeted issue: break down where the time goes (test side or server side), get the test under X s while keeping every assertion, and add a latency budget for the request the user actually sees. Record the rule as a team decision: a test over 5 s locally is a finding, not just a reason for a bigger timeout.
Also: Codex sandboxes can deny app.listen(0) with EPERM, so those workers cannot run real-server tests locally and depend on CI. Surface this to the humans.
Gotcha: an issue can carry an in-review label with no open PR yet (the worker loop labels it before the PR exists). Do not treat that as stuck in the same pass; check it next pass.