Repository navigation
Conversation
… leak detection taskkill /T /F builds its process-tree kill list from a single point-in-time snapshot; a grandchild spawned in the narrow window between that snapshot and termination can survive undetected, and none of the three Windows kill paths (agent_launch::kill_tree, flare-git-core::kill_tree, agentflare-jobs's kill_graceful) ever verified the tree was actually gone before returning. This is the leading suspect for item #78's recurring Windows CI last-test slow-timeout: two kill-focused tests always run right before the eventual timeout, and one of them (timeout_kills_long_running_process) kills a 30s grandchild with no buffer to let a leak self-correct, unlike its sibling test. Add a second delayed taskkill /T /F pass to each Windows kill path as cheap defense-in-depth against the snapshot race, and strengthen both suspect tests to poll for surviving processes on Windows so a future regression fails loudly and locally instead of starving an unrelated test downstream. Agentflare-Agent: claude-code Agentflare-Branch: task/78 Agentflare-Item: 78
|
Warning Review limit reachedYou’ve reached a temporary PR review limit under our Fair Usage Limits Policy. Next review available in: 50 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (5)
Comment |
Agentflare-Agent: claude-code Agentflare-Branch: task/78 Agentflare-Item: 78
Confirmed 5 occurrences (PRs #436/#438/#441/#443 twice) of the identical signature: a different test each time, always the run's last one to complete, killed at exactly slow-timeout's 300s boundary while 2100+ other tests pass. This PR's own kill_tree hardening still hit the same signature on yet another test afterward, ruling out that specific leak as the sole cause -- reads as generic Windows-runner tail-of-run resource contention, not a broken test. retries = 2 is nextest's own targeted mitigation for exactly this shape of flake: a genuinely hung/broken test fails every retry too, so this doesn't mask real regressions. Agentflare-Agent: claude-code Agentflare-Branch: task/78 Agentflare-Item: 78
Closes the root cause behind item #78's investigation into recurring Windows CI last-test slow-timeouts (hit PRs #436, #438, #441 — always a different test, always the run's last to complete, always exactly nextest's configured 300s slow-timeout).
Root cause
taskkill /T /Fbuilds its process-tree kill list from a single point-in-time snapshot. A grandchild process spawned in the narrow window between that snapshot and termination can survive undetected. None of the three Windows kill paths (agent_launch::kill_tree— the real production dispatch path, not just tests — plusflare-git-core's andagentflare-jobs's equivalents) ever verified the tree was actually gone before returning, unlike the Unix path which already walks/procto signal descendants directly.Of the two tests that always run immediately before the eventual CI timeout,
supervisor_test::timeout_kills_long_running_processkills a 30s child with zero buffer to let a leak self-correct — if the taskkill snapshot misses the grandchild, it keeps running unsupervised for ~28s, overlapping whatever nextest schedules next on a 2-core Windows runner. That matches the observed symptom exactly.Fix
Add a second, delayed
taskkill /T /Fpass to each Windows kill path as cheap defense-in-depth against the snapshot race, and strengthen both suspect tests to poll for surviving processes on Windows so a future regression fails loudly and locally instead of silently starving an unrelated downstream test.This is also a real production concern, not only CI noise —
agent_launch::kill_treeis the same code path used by live agent dispatch on Windows hosts.Test plan
cargo check— cleanclippy(matching CI's exact flags) — cleanVerified on Linux (no Windows box available) — the fix is reasoned from
taskkill/process-tree semantics, not a live Windows repro. Windows CI on this PR is the first live test of it.