babysit-pr

Hold one PR until it is merge-ready: poll, classify, fix, push, poll again. Terminal states are merged, closed, needs-user. Nothing else ends the loop, not a push, not a green run, not a report.

Loop

Each pass:

Snapshot.
```
bash scripts/pr-state.sh <pr> [--since <lastPushAt>]
```
. Carry the previous pass's
```
lastPushAt
```
into
```
--since
```
, so only review items newer than your last push surface.
Terminal first.
```
state
```
is
```
MERGED
```
or
```
CLOSED
```
: report it and stop.
Review items before CI. A fix commit retriggers CI, so answering review first avoids spending a rerun on a SHA you are about to replace.
CI. For each check with a failing
```
conclusion
```
, run
```
bash scripts/classify-failure.sh <check-url>
```
before touching code.
Push. Commit on the PR head branch, push, then return to step 1 on the new SHA.
Wait.
```
bash scripts/watch.sh <pr>
```
blocks until the checks settle, then prints a fresh snapshot. Use it instead of sleeping. Cap it with
```
BABYSIT_MAX_MINUTES
```
, default 60.

Rules

Published and fresh only. Act on review items whose
```
submittedAt
```
or
```
createdAt
```
is newer than
```
lastPushAt
```
.
```
pr-state.sh
```
drops
```
PENDING
```
reviews: a reviewer still drafting has not spoken yet. An item answered before the last push stays answered.
Classify a failure, then act on the class.
```
classify-failure.sh
```
prints
```
flaky
```
or
```
real
```
from log evidence, then the first 20 failing lines.
```
real
```
: patch the branch.
```
flaky
```
: rerun it with
```
gh run rerun <run-id> --failed
```
, budget 3 reruns per failing check; budget spent means needs-user. While a failure reads
```
flaky
```
, leave tests, build scripts, CI config and dependency pins as they are. A patch that only silences infra hides the signal for everyone downstream.
Humans get the user's words, not yours. A human reviewer is any
```
reviews[].author
```
that is neither the PR author nor flagged
```
isBot
```
. Surface their item to the user with a suggested reply, and wait for explicit confirmation before posting anything to GitHub. Bot items you act on and resolve yourself.
Resolve what you fixed. Once the fix is pushed:
```
bash scripts/resolve-thread.sh <threadId>
```
, using
```
threads[].id
```
from the snapshot.
Merge is the user's call. The sign-off rule (
```
.claude/rules/workflow-sign-off.md
```
): staged means done, committed means signed off. Stop at green and report. Run
```
gh pr merge <pr> --auto --squash
```
only when invoked with
```
--merge
```
, or when the user says merge in this session.
Green is a milestone, not an exit. No failing check,
```
mergeable: MERGEABLE
```
, no unresolved thread: report "ready to merge", then keep polling. Late review items still land while the PR is open.

Reviewer set

pr-state.sh

reads bot logins from

.skilly/config.json

at the git root:

json

{ "review": { "bots": ["chatgpt-codex-connector[bot]", "some-reviewer"] } }

A login ending in

[bot]

also counts. Missing key or missing file means an empty list, and every reviewer but the PR author reads as human. Read the config each run; keep no login in your head.

CI

Checks are whatever the PR reports. The repo's tier in

.skilly/config.json

decides the set (the

verify

skill explains tiers).

verify

is the one check every tier runs, so a red

verify

is always a statement about this branch.

Commits

Conventional commit, prefix by content:

fix:

when a check was red,

chore:

for a lint or config nit. Add one body line, so the origin of the change stays greppable:

Babysit-pr: fixes review thread <thread-url>

Report

Close every pass with: head SHA, check tally, mergeability, what was pushed, reruns spent, what is still open.

babysit-pr

NPX Install

Tags

SKILL.md Content

Loop

Rules

Reviewer set

CI

Commits

Report