用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/getsentry/outpost --skill fix-ci命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
Triage and respond to comments on a PR. Fix if actionable, reply either way. Load repo-setup first.
Triage and respond to comments on a PR. Fix if actionable, reply either way. Load repo-setup first.
Diagnose and fix failing CI on a PR. Load repo-setup first.
正在显示 SKILL.md
| name | fix-ci |
| description | Diagnose and fix failing CI on a PR. Load repo-setup first. |
| license | Apache-2.0 |
| metadata | {"audience":"autonomous-agents"} |
Fix failing CI on a PR the bot authored. Load repo-setup first.
Use fix-ci: attempt N comments as a diagnostic trail, not an excuse to stop.
Count existing attempts:
ATTEMPTS=$(gh api "repos/<owner>/<repo>/issues/<number>/comments" --paginate \
--jq '[.[] | select(.user.login == "'"$ME"'" and (.body | startswith("fix-ci: attempt")))] | length')
Post a short comment like "fix-ci: attempt 2 — looks like a type error in
foo.ts, investigating" before starting work. After three unsuccessful code
attempts, stop making speculative changes and re-establish the facts: fetch the
latest default branch, inspect the exact failed job and log again, and either
identify a new targeted fix or report the concrete external blocker. Do not
silently abandon, close, or mark the PR ready.
Never call a failure pre-existing or unrelated without evidence. Record the exact failed job, test or error, and log excerpt first. Then compare against the latest default branch: either show the same failure on that branch, or reproduce it after rebasing the PR onto it. If the latest default branch is green, treat the failure as PR-caused until evidence proves otherwise.
Never close a PR or say it is ready while required checks are failing. A flaky retry or an infrastructure blocker is not a green result; report its evidence and leave the PR open for the next actionable event.
Find failed checks for the PR head: gh pr checks <number> and gh run list --branch <branch> --status failure.
Read GitHub Actions logs with gh run view <id> --log-failed. If gh run view returns 404,
the failed check is external rather than a workflow run; inspect its check output instead:
gh api "repos/<owner>/<repo>/commits/<SHA>/check-runs" --paginate \
--jq '.check_runs[] | select(.conclusion == "failure") | {name,details_url,output}'
Record the exact failed job or check and its first relevant error, log excerpt, or check output in the attempt comment. Do not label an external check as infrastructure merely because it has no Actions log.
Categorize the failure — see references/failure-taxonomy.md for
the full taxonomy and decision tree. Categories: test failure,
type/lint error, build error, snapshot diff, flaky test, or infra
issue.
Flaky GitHub Actions run? Re-run once (gh run rerun <id> --failed) and stop.
Infra/dependency issue? BLOCKED.
Otherwise: make the smallest fix. Reproduce locally if possible.
Load deslop and review skills.
Commit and push. Verify the remote branch contains the same commit before you say the fix is ready:
git push origin HEAD
test "$(git rev-parse HEAD)" = "$(git rev-parse @{u})"
Then post one comment summarizing what you fixed and how. Write it like a teammate explaining the fix, not a status report.
Stop after pushing. Do not busy-wait in a polling loop for CI: the next CI completion webhook will wake you with the result.
Avoid modifying CI config unless the failure is specifically in it. Avoid bumping dependency versions — the fix should target the code, not the toolchain. Don't force-push. Don't merge.