Repeated evidence before agent handoff
One red run is not a flake.
Separate mixed histories from deterministic failures, preserve uncertainty, and calculate how many reruns a credible reproduction packet actually needs.
Repeated-run CSV
One commit, one row per test/run
The source announces grok_rescue but provides no plugin schema or API. This local tool does not install, call, monitor, or authenticate Grok, Codex, a repository, or a CI provider.
CI Flake Rescue Evidence / completed repeated-run diagnosis
—
—
—
Run-by-test signal
Analyze the sample to expose shared flaky-failure runs.
Why seven becomes 21
n = ceil(log(1 − .95) / log(1 − p))
At p = 37.5%, seven attempts give a 95% chance of observing at least one failure. Using the 13.7% lower Wilson bound instead yields 21 conservative attempts.
auth/refresh fails in all eight runs, so the supplied evidence supports a deterministic regression label.
r4 and r6 connect two flakes by correlation; the edge does not identify why either test failed.
A rescue packet should preserve what happened, and what remains uncertain.
The classifier requires one commit and a shared run universe. A complete history that both passes and fails is observed flaky; all-fail and all-pass histories remain separate; missing or single-run histories are insufficient. The original rows never disappear behind the label.
Wilson bounds and reproduction attempts describe sampling uncertainty, not causality. Co-failure edges show only repeated shared failed runs. Replace the hypothetical fixture with real exports and keep runner configuration, environment, logs, and code changes alongside this packet before asking any human or tool to rescue the failure.