babysit-pr

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.

---

Before you start

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.

---

Loop (repeat until done)

Each loop runs Phases 1–4 in order. A loop ends after Phase 4.

---

Phase 1 — Fetch & Categorize

1A. Fetch active threads

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.

1B. Check the 3-attempt rule

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:

• Remove it from this loop's working set
• Say: "3-attempt rule triggered for [file]:[line]. This thread has been addressed 3 times and Code Rabbit keeps re-raising it. Removing from this cycle — needs YOUR_NAME's manual review."

1C. Spawn the analyst

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.

---

GATE 1 — Present & Approve

Output the analyst's full result under the heading:

Code Rabbit Analysis — PR #$ARGUMENTS — Loop [N]

Then say exactly:

---

GATE 1 — [X] fix items, [Y] reply items.

Review the table above. You can:

• Change category: "move #3 to reply" or "move #1 to fix"
• Edit a fix description: "fix #2 should also add an upper-bound check"
• Edit a reply draft: "change reply #1 to say ..."
• Skip an item entirely: "skip #4"

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.

---

Phase 2 — Apply Fixes

(Skip this phase entirely if there are zero fix items after Gate 1.)

2A. Increment attempt counter

For each fix item, increment its {file}:{lineStart}:{commentHash} count in the attempt tracker.

2B. Spawn the fixer

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.

2C. Show the fixer's report

Output the fixer's full report (build status, test status, table of changes, output tails).

If BUILD is FAIL:

• Attempt 1 or 2 → say: "Build failed. I can retry with the error context. Say 'retry' to try again, or 'debug' to invoke /debug."
• Attempt 3 → invoke /debug automatically. Do NOT offer a 4th attempt.
• Do NOT proceed to Gate 2 with a failing build.

---

GATE 2 — Review Fixes Before Committing

Show the diff:

cd YOUR_PROJECT_ROOT && git diff

Output the diff under the heading:

Changes to commit — PR #$ARGUMENTS

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".

2D. Commit and push

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 [e-mail]" 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.

---

Phase 3 — Reply & Resolve

(Skip this phase entirely if there are zero reply items after Gate 1.)

3A. Show reply drafts

Output all reply drafts as a numbered list:

Reply drafts — PR #$ARGUMENTS

1. Thread [ID] — [file]:[line]: "[draft reply text]"
2. Thread [ID] — general: "[draft reply text]"

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.

3B. Post replies and resolve

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).

---

Phase 4 — Wait & Repeat

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.

• If YOUR_NAME says "wait": run sleep 600, then go back to Phase 1. Increment loop counter.
• If YOUR_NAME says "check now": go back to Phase 1 immediately. 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

• If [] → "PR #$ARGUMENTS has zero active Code Rabbit threads. All done!"
• If threads remain → "[N] threads still active. These may be from other reviewers (not Code Rabbit), or Code Rabbit threads we replied to but that aren't auto-resolved. Listing them for manual review:" → list thread IDs, files, first line of content → stop.

---

Hard rules (never break these)

• Never skip a GATE — always wait for YOUR_NAME's explicit word ("go", "commit", "send", "wait", "check now")
• Never commit without "commit" at Gate 2
• Never post replies without "send" at Gate 3
• Never add "Co-Authored-By" to any commit message
• 3-attempt rule: same {file}:{lineStart}:{commentHash} seen 3 times across loops → skip and flag
• If YOUR_NAME says "stop" at any point → stop immediately, summarize current state (which threads are fixed/replied/pending) and wait
• Follow the commit format from tasks/lessons.md
• ONE step at a time — explain what you are about to do, then do it, then stop