ci-rerun

# CircleCI from a GitHub check A CircleCI check on a PR links to `https://circleci.com/gh/<owner>/<repo>/<job-number>`, and that number is all `gh pr checks` gives. `scripts/circleci.py` turns it into the job's workflow and acts on that. ```bash C="${AGENTRC_DIR:-$HOME/.claude}/skills/ci-rerun/scripts/circleci.py" python3 "$C" log 391824 # the failed steps' last 150 lines (--lines 0: all), to classify the failure python3 "$C" rerun 391824 391728 ... # each job's workflow, once, from its failed jobs ``` `rerun` prints one JSON line, `{"reruns": [{"workflow", "jobs", "newWorkflow"}], "running": [{"workflow", "jobs", "status"}], "errors": [...]}`, and exits 1 when any workflow could not be re-run. Several failed jobs of one workflow are one re-run; the new workflow id is what a watcher records. A workflow still `running`, `failing`, `on_hold` or `queued` cannot be re-run yet: the script reads its status first, lists it under `running` without re-running it, and exits 3 when that is all that stopped it; re-run it later, there is no need to check its status yourself. A status it cannot read is an error. The re-run goes through the installed `circleci` CLI and its token in `~/.circleci/cli.yml`; the job, workflow status and log lookups are public. # Read the Docs from a GitHub check A failed `docs/readthedocs.org:<project>` status links to its build, `https://app.readthedocs.org/projects/<project>/builds/<id>/`, whose web log is often behind a rate limit. `scripts/rtd.py` works from that URL through the API. ```bash R="${AGENTRC_DIR:-$HOME/.claude}/skills/ci-rerun/scripts/rtd.py" python3 "$R" log <build-url> # state, the failure reason RTD recorded, failed commands' tails python3 "$R" rerun <build-url>... # a new build of each build's version ``` `rerun` prints `{"reruns": [{"build", "version", "newBuild"}], "errors": [...]}` and exits 1 when any build could not be re-run; builds of one version share one new build. The token is `RTD_TOKEN`, from the environment or a login shell; without it the script exits 2. `scripts/collect.py` is pr-babysit's CI collector, not a tool for this skill's users: its docstring is the contract. # Judgment - **Classify first.** Read the log; re-run only a failure that is the infrastructure's, not the code's. A re-run that fails the same way is a real failure. - **Once.** One re-run per workflow or version per failure; a second attempt is a human's call. - **Re-running is not publishing.** It changes nothing on the branch or the PR and needs no grant.