| name | babysit |
| description | Continuously babysit the user's open GitHub PRs from a dedicated loop session. Run /babysit in a fresh chat to start a perpetual loop - each cycle finds the user's open authored PRs, then for each one fixes failing CI, syncs/resolves base conflicts, addresses review-bot comments (Cursor Bugbot, Codex, Copilot), pushes, and replies on threads, then schedules the next cycle. Never merges, never opens duplicate PRs. Use when the user wants hands-off PR maintenance running on its own. Optional args - interval (default 15m) and a scope filter (org or owner/repo). |
babysit - recurring PR maintenance loop
Run this in a dedicated fresh chat. It turns the session into a perpetual PR
babysitter: each cycle does one maintenance pass over your open PRs, then
schedules the next cycle itself. Keep this session separate from the chat you
actually work in so its output never clutters your working session.
invocation
/babysit [interval] [scope]
interval - how often to run a cycle. e.g. 15m, 30m, 1h. default 15m.
each tick is capped at 1h (ScheduleWakeup max).scope - optional filter for which PRs to babysit: a GitHub org (freesolo-co)
or a single repo (freesolo-co/flash). default: all open PRs you authored.
parse these from the arguments given to /babysit. if none given, use the defaults.
each cycle (what THIS session does)
Keep this session lean. Do not do the heavy scanning/fixing inline here - delegate
each PR to a subagent so the noisy work stays isolated and this session only
accumulates short summaries. Steps:
-
discover the target PRs authored by the user:
gh search prs --author @me --state open --json repository,number,title,url,headRefName
if a scope arg was given, filter to that org or repo. de-dupe. if there are no
open PRs, print a one-line "no open PRs this cycle" and skip to scheduling.
-
for EACH open PR, spawn a subagent (Agent tool, general-purpose) and hand it the
per-PR cycle brief below along with the PR's repo (owner/name), number, url, and
head branch. run the subagents in parallel (one message, multiple Agent calls).
each subagent does the full fix cycle and returns ONE tight line: what it found,
what it did, and status.
-
print a compact summary for this cycle - one line per PR that had activity
(repo#num - what was done - status). PRs with nothing to do get no line. end with
a single counter line: cycle done HH:MM - N PRs checked, M acted on. do NOT
reprint a growing cumulative table each cycle; one cycle, one short block.
-
schedule the next cycle (see "scheduling" below).
per-PR cycle brief (what EACH subagent does)
You are babysitting one PR: <repo> #<num> on branch <head>. Fix what needs
fixing, push, reply to reviewers, and report back one line. Never merge.
-
gather everything that needs fixing on this PR:
a. review-bot feedback: Copilot review, Cursor Bugbot, Codex review, plus any
other PR review + inline comments. when reading inline comments
(gh api repos/<repo>/pulls/<num>/comments), capture each comment's id
alongside its path/line/body so you can post a threaded reply later (step 4).
b. failing CI / status checks: gh pr checks <num> --repo <repo>. for any
failing check, read its logs (gh run view <run-id> --log-failed) to find the
root cause.
c. build issues: check out the PR branch locally and confirm it builds/typechecks
with the repo's own tooling (bun install + bun run build / tsc / biome / lint /
tests). reproduce failures locally when possible.
d. base sync / merge conflicts:
gh pr view <num> --repo <repo> --json mergeable,mergeStateStatus,baseRefName.
use the PR's actual baseRefName (some repos target dev, not main).
mergeStateStatus BEHIND = base advanced with no conflicts; mergeable
CONFLICTING / mergeStateStatus DIRTY = real conflicts with base.
-
BEFORE making any code change (including resolving conflicts), recover the
originating Claude session that built this PR so you fix things with full intent
and in the same working style:
- build distinctive search strings: the branch name, the worktree path it was
developed in (PRs are built in
.claude/worktrees/<name> worktrees), changed
file paths, and distinctive symbols from the diff.
rg -l --glob '*.jsonl' "<branch>|<distinctive-symbol>|<changed-file>" /Users/minitom/.claude/projects/- transcripts live at
/Users/minitom/.claude/projects/<cwd-slug>/<session-uuid>.jsonl
(cwd-slug = working dir with / replaced by -). the most relevant are usually
under the PR's worktree slug and the repo's main slug. EXCLUDE this babysit
session's own transcript so you don't just re-read your own notes.
- read the matching transcript(s): extract the human's real instructions/intent
(lines with
"type":"user" that are real text, not tool_result or <...>
synthetic messages) and the assistant's key decisions. understand WHY the code
is shaped this way before changing it. if a finding contradicts a deliberate,
documented design choice, prefer flagging it over overriding it.
- if no originating session is found, note that and proceed from the PR
description and code comments.
-
apply fixes, guided by the recovered intent, in a checkout/worktree of the PR
branch:
- review bots: make the changes that are clearly correct and in scope.
hard rules for every subagent:
- never merge the PR (no
gh pr merge, no auto-merge, no /merge). fix, push, reply
only - the user does the final merge.
- never open a new PR for the same work; push all fixes to the existing PR branch.
you MAY note (not open) a genuinely orthogonal in-scope issue for the user.
- do not touch a PR that is already merged or closed.
scheduling the next cycle
After printing this cycle's summary, keep the loop alive: call ScheduleWakeup with
delaySeconds = the interval in seconds (15m = 900, 30m = 1800, 1h = 3600, capped at
3600), prompt = /babysit <interval> <scope> (pass the same interval and scope so
the next tick repeats identically), and a one-line reason like
"next PR babysit cycle in ". This is /loop dynamic mode - each tick does one
pass then reschedules, so running /babysit once keeps it going until you stop the
session. Always reschedule, even on a "no open PRs" cycle, so newly opened PRs get
picked up.
