Babysit

babysit

A configuration of loop-controller. That skill supplies the loop machinery — primitive selection, the full guardrail stack, state externalization. This skill supplies the three things specific to "keep a PR healthy": the per-pass recipe (it schedules git-pr-feedback, it does not re-implement it), the mechanical proof (zero open blocking findings + mergeable), and the HITL boundary that decides what's AFK-safe vs what pauses. Read loop-controller for the guardrails; they're inherited, not repeated here.

Why disable-model-invocation: this loop commits, pushes, rebases, and touches a live PR thread on its own, on a clock. It is user-driven — you want to type /loop 5m /babysit (or /babysit), not have Claude silently start tending a PR because a reviewer happened to comment.

The 5-part contract

The proof: zero blocking findings AND mergeable, default-FAIL

"Healthy" is not "I replied to the last comment." It is two conditions observed together from gh: the PR has zero open, actionable, blocking review comments, and GitHub reports it mergeable (rebased onto base, CI green). Assume the PR is not healthy until a fresh fetch proves both — that's the default-FAIL stance, and it's why every pass re-fetches rather than trusting last pass's verdict. A loop that stops after answering one comment ships a PR that picked up three new ones and went stale behind a base push.

Name the artifact explicitly: a gh pr view --json mergeable,reviewDecision plus the open-comments count from git-pr-feedback's fetch (see references/scheduling.md for the exact query). Both must read clean in the same pass.

Step 1 — One pass = one git-pr-feedback cycle

The per-pass work is git-pr-feedback: identify the PR, fetch inline + issue comments + review summaries (paginated), triage each by author and category, address what's clear, reply on GitHub. Do not re-document or re-implement that — invoke it. babysit adds only the loop discipline around it: one root-cause change per pass, then re-check the whole PR state. If the base branch moved and the only blocker is staleness, the pass is a routine rebase (not a finding fix). The full per-pass algorithm and the handoff are in references/scheduling.md.

Step 2 — Schedule it with /loop (the watcher, not /goal)

Per loop-controller Step 1: this is a watch/poll job — you wait for review activity to change on a cadence — so the primitive is /loop, the scheduler, not /goal (which pushes to a finish line). The recipe is the documented daily driver:

/loop 5m /babysit

/loop is session-scoped, expires after ~3 days, and does not catch up on missed intervals — those mechanics (and the loop.md self-paced variant) live in loop-controller's references/primitives.md; don't re-document them. What goes in loop.md and how to scope the run to one PR are in references/scheduling.md.

HITL is load-bearing for this loop

babysit runs only inside the reversible boundary unattended. The irreversible or human-facing actions are HITL checkpoints — never autonomous:

• Force-push / rebasing a shared branch — pause. A fast-forward push of your own commits is fine; a history rewrite of a branch others may have pulled is not.
• Resolving or replying to a human reviewer's conversation — pause. A bot/Copilot nit can be answered and resolved unattended; a human's thread is a conversation, and you do not argue with a reviewer inside a loop. On any substantive disagreement with a human reviewer, stop and surface it.
• Merging — never. babysit does not auto-merge, full stop. Reaching the proof (zero blocking + mergeable) means ready for the human to merge, not "merge it."

Within the boundary — addressing Copilot/bot nits, routine fast-forward rebases, re-running the gate, fast-forward pushes — it is AFK-safe. The full AFK-safe vs pause table is in references/scheduling.md.

Guardrails specific to this loop

Inherits the full stack from loop-controller → references/safety.md. The caps this loop sets:

• Poll / iteration cap — /loop's ~3-day session expiry is the outer bound; set an inner per-PR pass cap (default ~20) so a wedged PR doesn't burn the whole window. Hitting it is a stop-and-escalate, not a reason to loosen the proof.
• No-progress detection — if the same blocking finding survives 3 consecutive passes (addressed but the reviewer/check still flags it, or it can't be resolved within the reversible boundary), stop and surface it. Three passes on one finding means it needs a human, not a fourth attempt.
• Budget cap — a watch loop that fires every 5 minutes for days adds up; enforce a token/cost ceiling that terminates the loop (read it from .claude/profile.yaml when present), not just warns.
• Never cheat the proof. Don't mark a thread resolved without addressing it, don't dismiss a human reviewer's comment to clear the count, and don't merge to make "mergeable" moot. A green that came from silencing a reviewer is a finding, not a win.

Choosing the driver primitive

/loop 5m /babysit is the scheduler. The per-pass exit (this pass made one clean, reversible change and re-checked the PR) is provable from gh output, so a pass can run under auto mode unattended within the HITL boundary above. The loop as a whole has no /goal finish line — it watches until you stop it or a stop condition fires.

Long-run hygiene (wired per loop-controller Step 6)

This loop runs unattended for days, so the Claude 5 long-run rules are part of its contract, not inherited ambience (drop-in text: model-adaptation → references/long-run-hygiene.md):

• Evidence-backed progress — every per-pass report line (finding addressed, rebase done, gate green) traces to a tool result from that pass: the gh query output, the gate's exit code, the push result. Never narrate a fix the re-fetch didn't confirm.
• Don't end a pass on a promise — a pass ends with the change made and the whole PR state re-checked, or at an explicit HITL stop; never with "I'll rebase next pass" and no action. The autonomous-operation reminder applies while the user is away — the HITL checkpoints above are the only sanctioned pauses.
• Budget is a harness decision — the cost ceiling reads from externalized state (.claude/profile.yaml); don't surface a remaining-token countdown to the working pass, and never summarize-and-quit on a phantom context worry.
• Effort per pass — routine nit/rebase passes run at medium/high; reserve xhigh for a genuinely hard finding (tiering: model-adaptation).

Reference files

• references/scheduling.md — the per-pass algorithm and the git-pr-feedback handoff, the scheduling recipe (/loop 5m /babysit, what to put in loop.md, scoping to one PR, the exact gh proof query), and the HITL boundary table (what's AFK-safe vs what pauses).