babysit-pr
Babysit a PR to ready: review, repairs, CI, and a current-head verdict. Merging belongs to ship.
- License 许可
- MIT
- 最近更新
- 2026-11-11
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- babysit-pr
- description
- Babysit a PR to ready: review, repairs, CI, and a current-head verdict. Merging belongs to ship.
02 资源文件 4
- 参考文档agents/openai.yaml260 B
- 参考文档references/acceptance.md2.2 KB
- 参考文档references/real-surface-qa.md2.2 KB
- 参考文档references/repairs.md1.9 KB
03 SKILL.md 正文
Babysit PR
Carry the PR through review, in-scope repairs, required QA, and hosted checks to
a verdict supported by evidence for its current head. $open-pr prepares the
review presentation; Babysit owns the review and repair cycle.
User authorization
An explicit user invocation of $babysit-pr, or a direct user request to
babysit, shepherd, or finish a PR, authorizes the complete repair loop:
- read the repository and live PR state;
- make the smallest local changes needed to resolve substantiated blockers;
- commit those verified, in-scope repairs;
- non-force push those repairs to the PR's existing head branch and monitor the new head;
- resolve review threads whose concern is cleared on the current head; and
- mark that PR ready for review after its current head earns
READY.
Automatic model selection of this skill does not grant push authority. An explicit read-only or review-only request also narrows the authorization. Require separate approval for force pushes, rebases that rewrite the remote branch, retargeting or replacing the PR branch, materially broader changes, public replies, closing or reopening the PR, other lifecycle changes, and merge.
Stay in the current Codex task. Do not create, message, monitor, or manage other user-visible tasks unless the user explicitly asks. Consume existing reviewer or handoff results as evidence, and do not start a duplicate reviewer for a frozen scope that already has an adequate completed or running review.
When selected QA requires video, follow the video evidence contract. Keep optional demonstration clips separate from required acceptance evidence.
Track repairs, reviews, QA, and verification using the model contributions contract. Before issuing a verdict, refresh the PR's model contributions record when body edits are authorized; otherwise include the proposed update in the handoff.
Workflow
- Capture repository and PR identity, base and head SHAs, worktree state, linked acceptance criteria, changed user journeys, checks, mergeability, draft state, and review threads. Use the GitHub transport contract. Keep issue numbers, PR numbers, and branch names distinct. Refresh the snapshot after every push.
- Read existing review receipts and apply the DiffOwl coverage contract, the authority for full reviews, repair coverage, reuse, and independent review. Inspect comments with their threads, replies, resolution state, and current code; group duplicate findings and treat each as a claim to verify.
- When findings need disposition or a repair is required, follow repairs and review threads. Continue until the substantiated blockers are resolved or a specific authority or dependency prevents further progress.
- For changed user-visible journeys, follow real-surface QA. Otherwise record why that QA does not apply. Missing required evidence remains a gap.
- Verify hosted checks as described below. Fix in-scope failures and refresh affected evidence after repairs.
- Before issuing a verdict or changing draft state, follow acceptance. Recheck the live PR and return the shared coverage receipt plus current QA, checks, feedback, and mergeability.
Hosted checks
Use gh pr checks <number> or the equivalent platform view as the source of truth
for the entire PR-attached check set. Enumerate every check at the current head,
use native wait/watch or product automation for unsettled checks, and classify
each failure with platform evidence. Separate branch-caused failures from
infrastructure or provider flakes; do not dismiss a failure as a flake without
evidence or an authorized resolution. Never encode a fixed polling loop or stale
check list in the workflow.
Completion
Finish with READY only when acceptance passes for the current head and the PR
is open for review. Use NOT READY for established blockers and INCONCLUSIVE
for missing or unsettled evidence. Keep working while useful, authorized repairs
or pending checks can resolve the verdict. When blocked, report the exact missing
authority, dependency, or evidence and the completed work.
04 同仓库更多 Skills
why
Why code is this way: design rationale, regressions, and history from evidence. For runtime: how.
unslop
Edit prose to remove AI writing patterns when polishing text or asked to make it sound natural.
typescript-best-practices
Use when writing or reviewing .ts/.tsx code: types, validation boundaries, and API signatures.
sync-pstack
Analyze upstream cursor/plugins pstack changes and guide semantic ports into Codex-native Jstack without overwriting local skills. Use when the user says "sync pstack", "check pstack updates", or asks to update Jstack from upstream.
swarm
Fan out N parallel workers, drain them, and return one report. Use for /swarm, 'swarm this', or parallel coverage, races, gauntlets, and exploration.
statechart
Check a finite TLA+ model and explain its paths and rule failures as a statechart.