| name | babysit-prs |
| description | Drive every open PR to merge-ready on a self-paced loop, with no `/loop` wrapper. Each tick triages through `pr-feedback --auto confirmed`, which answers and folds through `pr-respond` and `fixup`, then waits for CI and new reviews and goes again until the PR is green. Use when asked to babysit, watch, or surveiller open PRs, or to keep them moving until they can merge. |
| argument-hint | [--once] [--every <interval>] |
Babysit PRs
Keep open PRs moving without polling them by hand. This skill owns the scan, the waiting, and the stopping. Every judgment about the code belongs to the skills it calls.
Each tick
-
List the PRs still in flight:
gh pr list --author "@me" --state open --json number,url,title,isDraft,headRefName
None left? Say so and stop the loop.
-
Per PR, spawn one subagent to run the triage. Send them all in a single message so they run concurrently:
Agent({
subagent_type: "general-purpose",
model: "opus",
name: "triage-pr-<n>",
description: "Triage PR #<n>",
prompt: "Think hard. Invoke the `pr-feedback` skill on PR #<n> with `--auto confirmed`. Adjudicating each review item — is this claim real, or refuted? — is the part that matters; default to refuted when unsure, and spend the reasoning there rather than on restating the diff. Return only: items answered this tick (one line each), items held for the user (one gist each), and any `fixup` blocker."
})
The subagent keeps every thread body, diff, and reply draft out of this context, which is what makes many open PRs affordable, and it gets a model that can actually weigh a review claim. Don't fetch threads, classify, or draft replies here — the main loop only collects status.
-
Report the tick as one table, one row per PR:
| PR | Issue | Mergeable | CI | Answered | Held | New |
|---|
[#num title](url) | [<KEY> title](https://linear.app/issue/<KEY>), — if the branch carries no key | mergeable + mergeStateStatus | gh pr checks <n> rollup, failures first | from the subagent's report | from the subagent's report, one gist each | new comments and reviews since last tick, — if none |
Take the Linear key from headRefName and the title from linear-cli once per PR, then reuse it for the rest of the session — it doesn't change between ticks.
A PR that is CLEAN, green on every check, and holding nothing is merge-ready. Notify via PushNotification (ToolSearch "select:PushNotification") and drop it from the loop. The merge itself is the user's.
-
Wait. End the turn with ScheduleWakeup, always, including a turn that ends on a held item or a fixup blocker. A turn that ends without an armed wakeup kills the loop.
ScheduleWakeup({ prompt: "/babysit-prs", delaySeconds: 300, noop: <nothing changed?>, reason: "<what you're waiting on>" })
Use 300 while CI runs or reviews are expected. That is the cadence this loop is built around. Stretch to 1200 or 1800 once every remaining PR is held on the user, since nothing will move until they answer. --every <interval> pins it.
Set noop: true on a tick where nothing changed, false when something moved.
Stopping
Call ScheduleWakeup({ stop: true }) and don't re-arm when every PR is merge-ready or closed, when --once was passed (never arm at all), or when the user says stop.
Everything else keeps ticking. A held item and a rebase conflict both mean report and come back, not give up.
What lives where
| Concern | Skill |
|---|
| Fetching threads, verdicts, P1/P2/Nit | pr-feedback, in a per-PR Opus subagent |
| Code changes, replies, reactions, resolving | pr-respond |
| Finding the introducing commit, fold, force-push | fixup |
| Restacking children after a fold | gh-stack |
| Scanning, PR status, cadence, stopping | here |
Loop state lives in this conversation, not on disk, so the ticks have to stay in one session. A tick that has forgotten the last one re-triages from scratch, which costs tokens, not correctness.
