القائمة
Drive a PR to zero review threads — fetch, categorize, fix, reply, repeat. Use when handling Code Rabbit comments, fixing PR review feedback, or working through the CR cycle. Usage: /babysit-pr <PR_ID>
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Adversarial evaluation of code changes — runs build, tests, plan compliance, and security review. Use after implementing changes or before creating a PR. Usage: /evaluate [story-id] [--quick]
Build a feature from a tracker task (local task, GitHub issue, or Todoist task) or plain description — understand, plan, execute, evaluate, and PR in a streamlined flow, or loop back on a rejected PR with `--rework <PR#>`. Lighter than /story — designed for solo devs and small teams. Usage: /implement <issue-id, task-title, or description> [--discuss] [--research] [--quick] [--auto] [--full] [--autonomous] [--rework <PR#>]
Goal-driven story execution — understand → define goal → plan → execute → verify → e2e gate → PR. Done = goal met (acceptance + e2e gate green), not just compiles. Use when starting a sprint story, implementing a feature, or picking up a task. Usage: /story <story-id> [--auto] [--autonomous]
Reconcile merged PRs and completed work against open tracker items. Mode-aware: works in local, tracker, and both modes. Supports --dry-run and --import-backup [file]. Usage: /sync-tracker [--dry-run] [--import-backup [file]]
Execute XML tasks from the story plan wave by wave (Phase 3 only — no understand, no plan, no PR). Use when resuming a story that already has a task plan. Usage: /run-tasks <story-id> [--auto]
Root cause diagnosis when the 3-attempt rule triggers — stop retrying and diagnose. Builds a deterministic feedback loop first, then generates ranked falsifiable hypotheses. Mandates regression tests and instrumentation cleanup. Aliases: /debug, /diagnose. Usage: /debug
| name | babysit-pr |
| description | Drive a PR to zero review threads — fetch, categorize, fix, reply, repeat. Use when handling Code Rabbit comments, fixing PR review feedback, or working through the CR cycle. Usage: /babysit-pr <PR_ID> |
Core Philosophy: Drive a PR to zero review threads — analyze once, get your approval at every gate, fix and commit, and never post a reply or push code without your explicit word.
Triggers: "babysit PR 17009", "fix Code Rabbit comments", "handle PR review", "clear review threads", "work through the CR cycle"
You are babysitting PR #$ARGUMENTS through its Code Rabbit review cycle.
Your job: drive this PR to zero active Code Rabbit threads. You do this in loops of: fetch → analyze → fix → reply → wait 10 min → repeat.
Read YOUR_PROJECT_ROOT\tasks\lessons.md. Note the "PR Comment Review Process" and "Patterns Code Rabbit Flags" sections.
Check you are on the right branch:
cd YOUR_PROJECT_ROOT && git status && git branch --show-current
If the branch shown does not match PR #$ARGUMENTS, tell YOUR_NAME and stop.
Initialize an attempt tracker: a map of {file}:{lineStart}:{commentHash} → count. Starts empty. Persists for the entire session — you update it every loop.
The commentHash is the first 60 characters of the Code Rabbit comment body, lowercased and stripped of line numbers and whitespace. This ensures the same issue re-raised on a different line (after a rename or code move) is still tracked as the same attempt.
Each loop runs Phases 1–4 in order. A loop ends after Phase 4.
bash "YOUR_PROJECT_ROOT/.claude/code-platform/active/get-pr-review-threads.sh" $ARGUMENTS
If the result is [] or empty → say "Zero active Code Rabbit threads on PR #$ARGUMENTS. Done!" and stop entirely.
For each thread, compute its key: {file}:{lineStart}:{commentHash} (use "general" as the file if file is null).
Check the attempt tracker. If any thread's key has a count of 3 or more:
Spawn a babysit-pr-analyst (foreground) with this prompt:
PR ID: $ARGUMENTS Active threads JSON: [paste the full JSON from step 1A, excluding any threads removed in 1B] Categorize each thread as fix or reply. Produce the full analysis table.
Wait for it to return the categorized result.
Output the analyst's full result under the heading:
Then say exactly:
GATE 1 — [X] fix items, [Y] reply items.
Review the table above. You can:
Say "go" when the plan looks right.
Do NOT proceed until YOUR_NAME says "go". If corrections are given, apply them to the analyst's output and re-show the updated table, then wait again.
Seed the live progress checklist for this loop. Once YOUR_NAME says "go", create (or, on later
loops, extend) a TodoWrite list with one item per approved thread — labelled fix or reply — so
YOUR_NAME can watch each thread get resolved before anything is committed. Mark items completed as
fixes are committed (2D) and replies posted (3B). See rules/progress-tracking.md.
(Skip this phase entirely if there are zero fix items after Gate 1.)
For each fix item, increment its {file}:{lineStart}:{commentHash} count in the attempt tracker.
Spawn a babysit-pr-fixer (foreground) with this prompt:
Apply these fix items to the YOUR_ORG codebase, then run full build and all tests.
Fix items: [paste the "Fix items — detailed descriptions" section from the analyst output, with any edits from Gate 1 applied]
Wait for it to return.
Output the fixer's full report (build status, test status, table of changes, output tails).
If BUILD is FAIL:
/debug automatically. Do NOT offer a 4th attempt.Show the diff:
cd YOUR_PROJECT_ROOT && git diff
Output the diff under the heading:
Then say exactly:
GATE 2 — Fixes applied. Build: [PASS/FAIL]. Tests: [PASS/FAIL].
Review the diff above. Say "commit" to commit and push, or tell me what to change first.
Do NOT commit until YOUR_NAME says "commit".
When YOUR_NAME says "commit":
cd YOUR_PROJECT_ROOT && git add -A && git commit -m "Fix Code Rabbit review comments" && git push
CRITICAL: Never add "Co-Authored-By: Claude Sonnet 4.6 noreply@anthropic.com" or any Co-Authored-By line to the commit message.
Confirm: "Committed and pushed. Code Rabbit will re-analyze within ~10 minutes." Mark each committed fix item completed in the TodoWrite list.
(Skip this phase entirely if there are zero reply items after Gate 1.)
Output all reply drafts as a numbered list:
Then say exactly:
GATE 3 — [Y] reply drafts above.
Edit any draft before sending: "change reply #2 to say ..."
Say "send" to post all replies and resolve the threads.
Do NOT post anything until YOUR_NAME says "send". If edits are given, update the drafts and re-show, then wait again.
For each approved reply item, run in sequence:
bash "YOUR_PROJECT_ROOT/.claude/code-platform/active/reply-pr-thread.sh" $ARGUMENTS <THREAD_ID> "<REPLY_TEXT>"
bash "YOUR_PROJECT_ROOT/.claude/code-platform/active/resolve-pr-thread.sh" $ARGUMENTS <THREAD_ID>
After all are done, confirm: "[Y] threads replied to and resolved." Mark each replied thread completed in the TodoWrite list.
Also increment the attempt counter for each reply item (same as fix items — these are addressed attempts).
If fixes were pushed in Phase 2 (new code was pushed this loop):
Say exactly:
GATE 4 — Fixes were pushed this loop. Code Rabbit needs ~10 minutes to re-analyze.
Say "wait" to start the 10-minute timer, or "check now" if you know Code Rabbit has already re-analyzed.
Do NOT proceed until YOUR_NAME says "wait" or "check now". Do NOT skip this gate even if Code Rabbit appears to have already responded.
sleep 600, then go back to Phase 1. Increment loop counter.If NO fixes were pushed (only replies, or neither):
There is nothing new for Code Rabbit to analyze. Do a final check:
bash "YOUR_PROJECT_ROOT/.claude/code-platform/active/get-pr-review-threads.sh" $ARGUMENTS
[] → "PR #$ARGUMENTS has zero active Code Rabbit threads. All done!"{file}:{lineStart}:{commentHash} seen 3 times across loops → skip and flagtasks/lessons.md