Babysit PR
SkillMonitoring & opsShepherd the current user's open PR through base updates, CI failures, and review feedback without rewriting history or merging. Use when asked to babysit, monitor, or fix an existing PR.
Instructions available. Your AI can read the instructions. Execution depends on the setup they require.
Account requirements not reviewed. Check the skill instructions before use; ahel provides instructions and does not run this skill.
Add ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.
Then ask your AI: use the Babysit PR skill
What this skill tells your AI
The instructions your AI receives, as published by trailofbits/coop in .agents/skills/babysit-pr/SKILL.md and read by ahel’s review.
One pass surveys the PR, acts on safe mechanical work, then reports. Do not merge, close, reopen, dismiss reviews, rewrite published history, or edit PR metadata unless asked.
Preconditions and survey
Require an authenticated gh, a clean git worktree, an open PR for the current
branch, and ownership of its head branch. Stop on a dirty tree rather than
stashing. A merged PR is terminal; a closed-unmerged PR needs user direction.
Gather once:
gh pr viewwith state, draft status, head/base refs and SHAs, merge state, review decision, check rollup, reviews, and comments;- base divergence after fetching the base;
- inline
reviewThreadsthrough GraphQL (resolution is authoritative); - failed/pending run details and logs.
Treat top-level CHANGES_REQUESTED reviews and structured bot comments as
feedback even when no inline thread exists. Do not assume a green-looking list
is complete: compare the visible checks with the repository's required CI jobs.
Act in priority order
- Integrate a behind/conflicting base with
git merge --no-ff origin/<base>. Never rebase. Resolve semantically; stop if the conflict crosses scope. - Reproduce settled CI failures. Fix the underlying fmt, clippy, test, taplo, deny, or workflow issue. Surface flaky/environmental and real-VM integration failures rather than fabricating a code fix.
- Address unresolved review feedback only when the ask is concrete and mechanical. Verify the claim first. Escalate design, product, security, and scope decisions.
Make one logical change per commit. Run the narrow test plus format and clippy;
run taplo for TOML. Never use --no-verify. Push normally and stop on rejection
rather than force-pushing.
After a pushed commit fixes an inline thread, reply exactly Fixed in <sha>. or
Fixed in <sha>: <one-line>., then resolve that thread. Do not reply or resolve
before the fix is pushed, and never close a judgment-call thread for the user.
Re-survey once after changes. Report pushed commits, failed/pending checks, unresolved feedback, merge state, and absent integration/platform gates. If only an external signal remains, offer to monitor at an appropriate cadence; do not start recurring monitoring unless the user asked.
Signals
- GitHub stars
- 740
- Forks
- 38
- Last commit
- Oct 2026
Advanced
- Item type
- skill
- Key
babysit-pr-trailofbits- Source
- github.com/trailofbits/coop
More in Monitoring & ops
Skill · anthropics
More in Monitoring & opsagent-eval
Skill · affaan-m
More in Monitoring & opsarchitecture-decision-records
Skill · affaan-m
More in Monitoring & opsbabysit
Skill · thedotmack
More in Monitoring & opseng-runbook
Skill · nexu-io
More in Monitoring & opsweekly-update
Skill · nexu-io
More in Monitoring & ops