Flake Hunt

Run a flake hunt on a Biloba Ginkgo suite — run the whole browser suite many times (60 by default), every run to completion with its own JSON report, then read each spec's failure rate, the seeds, and Biloba's failure evidence out of those reports instead of re-running. Covers the hunt script (compile once, --procs at the measured knee, --randomize-all, --poll-progress-after, a per-run BILOBA_SCREENSHOTS_DIR), why --repeat and --until-it-fails can't measure a rate, what keeps a hunt a valid measurement (one hunt at a time on an idle machine, nothing editing the tree, focused hunts for iterating only, stale visual baselines), how many clean runs it takes to call a flake dead, reading the results (one systemic race behind many names, wedges, a long tail that clusters by run), and the performance record a hunt produces (spec timing, parallel efficiency, per-spec cost, drift between hunts). Use before declaring a flake fixed, after changing shared test helpers or fixtures, at the end of a batch of work, or to measure a suite's performance.

mkdir -p ~/.claude/skills/flake-hunt
curl -L https://claudskills.com/skills/flake-hunt/SKILL.md \
  -o ~/.claude/skills/flake-hunt/SKILL.md

@misc{flake-hunt-2026,
  author    = {onsi},
  title     = {Flake Hunt [Claude Code skill]},
  year      = {2026},
  publisher = {ClaudSkills},
  url       = {https://claudskills.com/skills/flake-hunt/}
}