Agent Skills › rokartur/BetterCmdTab › babysit-pr

babysit-pr

GitHub

自动监控PR的CI状态并触发人工审查,根据审查结果修复代码或回复意见,直至获得批准。

.agents/skills/babysit-pr/SKILL.md rokartur/BetterCmdTab

Trigger Scenarios

用户要求监视PR 用户要求babysit PR

Install

npx skills add rokartur/BetterCmdTab --skill babysit-pr -g -y
More Options

Non-standard path

npx skills add https://github.com/rokartur/BetterCmdTab/tree/main/.agents/skills/babysit-pr -g -y

Use without installing

npx skills use rokartur/BetterCmdTab@babysit-pr

指定 Agent (Claude Code)

npx skills add rokartur/BetterCmdTab --skill babysit-pr -a claude-code -g -y

安装 repo 全部 skill

npx skills add rokartur/BetterCmdTab --all -g -y

预览 repo 内 skill

npx skills add rokartur/BetterCmdTab --list

SKILL.md

Frontmatter
{
    "name": "babysit-pr",
    "description": "Babysit a BetterCmdTab pull request through rokartur-review and CI until it is ready to merge. Use when the user asks to babysit, watch, or monitor a PR."
}

Babysit PR

rokartur-review is the house review bot (~/developer/review-bot). It reviews only on request: a review request from rokartur starts one run on Opus, billed to Artur's subscription, capped at 30 minutes. The bot then posts one GitHub review (APPROVED or CHANGES_REQUESTED, numbered findings inline) and drops its own request. Its next run reads the PR's conversation comments, not replies inside review threads.

Asking for babysitting authorizes commits and pushes to this PR's branch, --force-with-lease after a rebase. Merging and closing stay with the user.

Loop

  1. CI. gh pr checks <pr> --watch. App CI runs only when app paths change; "no checks reported" is a pass. Fix a red repo failure. Rerun a runner flake once with gh run rerun <id> --failed.
  2. Request. Once CI is green on the head commit, request a review: gh pr edit <pr> --add-reviewer rokartur-review.
  3. Wait. Run .agents/skills/babysit-pr/wait-review.sh <pr> in a background terminal (load_tools terminals) and poll its output. It prints the verdict on the head commit, or exits 1 when the bot's run failed: report that and stop. Put status notes in the same message as the next tool call: a message without one ends the turn and the babysitting stops.
  4. Findings. Read the review: gh api repos/rokartur/BetterCmdTab/pulls/<pr>/reviews for the summary, .../pulls/<pr>/comments for the inline findings. Check every finding against the source. Fix the real ones in one round. Answer the rejected ones in a single PR comment, one line per finding number with the reason, so the next run sees it.
  5. Repeat from 1 after pushing the fixes. Request the next review only after every finding of the last one is fixed or answered.

Stop when the bot's latest review is APPROVED on the head commit and CI is green, then report that the PR is ready.

Rules

  • Keep the PR to its original goal. Findings that ask for more go to the user as follow-ups.

  • When main moves, rebase and rerun from 1. If an overlapping PR makes this one obsolete, stop and report it.

  • Post only when something changed, and sign every comment:

    [MODEL-SLUG] RESPONDING ON BEHALF OF ARTUR
    -----
    [actual reply]
    

Version History

  • 3df8c68 Current 2026-10-18 19:36

Same Skill Collection

.agents/skills/add-preference/SKILL.md
.agents/skills/localize-strings/SKILL.md
.agents/skills/marketing-screenshots/SKILL.md
.agents/skills/release-changelog/SKILL.md
.agents/skills/release/SKILL.md
.agents/skills/verify/SKILL.md

Metadata

Files
0
Version
3df8c68
Hash
05c3fc28
Indexed
2026-10-18 19:36

inicio - Wiki
Copyright © 2011-2026 iteam. Current version is 2.155.2. UTC+08:00, 2026-10-21 03:12
浙ICP备14020137号-1