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
| Part |
This loop |
| trigger |
an open PR of yours with potential review activity; scheduled via /loop 5m /babysit, or an explicit /babysit (optionally a PR number) |
| action |
ONE pass: fetch new review findings (Copilot + human) via git-pr-feedback → triage → address one root-cause finding (or rebase if base moved) → re-run the project gate → push → reconcile the PR thread. One change per pass; re-check the whole PR state after |
| proof |
the gh PR review/status query returns zero OPEN actionable/blocking comments AND mergeable == "MERGEABLE" (up to date with base, CI green) — default-FAIL: assume findings are unaddressed until the fetch proves otherwise (treat mergeable == "UNKNOWN" as not-yet-proven and re-check) |
| memory |
a per-PR fix log (findings addressed, findings deferred + why), git history, and the PR thread itself — all durable across the session-scoped /loop fires |
| stop |
zero blocking findings AND mergeable OR poll/iteration cap OR no-progress for 3 rounds (same finding unaddressed 3×) OR budget cap OR an HITL checkpoint is hit |
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).
1---2name: babysit3version: 1.1.04description: >-5 Keep one of your open PRs healthy on a schedule: poll for new review activity,6 auto-address bot/Copilot nits and routine rebases, re-run the project gate, and7 reconcile the PR thread — looping until the PR has zero blocking findings and is8 mergeable, with HITL checkpoints on anything irreversible. This is the9 scheduled, guardrailed loop around git-pr-feedback (which does the per-pass10 fetch/triage/fix). Run it as Boris Cherny's daily-driver "/loop 5m /babysit",11 or invoke /babysit once. Use when you want a PR babysat while you work12 elsewhere, want review comments auto-handled on a cadence, or want a PR kept13 rebased and green. Trigger on "babysit my PR", "babysit prs", "watch this PR",14 "keep my PR green", "auto-address review comments on a schedule", "loop 5m15 babysit", "keep the PR rebased", "poll the PR for new feedback". Never16 auto-merges and never force-pushes a shared branch unattended. A17 configuration of loop-controller.18requires_claude_code: true19min_plan: starter20disable-model-invocation: true21allowed-tools: ["Read", "Edit", "Bash", "Grep", "Glob"]22composes_with: ["loop-controller", "git-pr-feedback", "git-pr", "git-commit", "loop"]23spawned_by: []24metadata:25 source: https://skillmd.com/skills/ivy00johns/babysit26---2728# babysit2930> **A configuration of [`loop-controller`].** That skill supplies the loop31> machinery — primitive selection, the full guardrail stack, state32> externalization. This skill supplies the three things specific to "keep a PR33> healthy": the **per-pass recipe** (it schedules [`git-pr-feedback`], it does34> not re-implement it), the **mechanical proof** (zero open blocking findings +35> mergeable), and the **HITL boundary** that decides what's AFK-safe vs what36> pauses. Read `loop-controller` for the guardrails; they're inherited, not37> repeated here.38>39> **Why `disable-model-invocation`:** this loop commits, pushes, rebases, and40> touches a live PR thread on its own, on a clock. It is user-driven — you want41> to *type* `/loop 5m /babysit` (or `/babysit`), not have Claude silently start42> tending a PR because a reviewer happened to comment.4344## The 5-part contract4546| Part | This loop |47|---|---|48| **trigger** | an open PR of yours with potential review activity; scheduled via `/loop 5m /babysit`, or an explicit `/babysit` (optionally a PR number) |49| **action** | ONE pass: fetch new review findings (Copilot + human) via [`git-pr-feedback`] → triage → address **one root-cause finding** (or rebase if base moved) → re-run the project gate → push → reconcile the PR thread. One change per pass; re-check the whole PR state after |50| **proof** | the `gh` PR review/status query returns **zero OPEN actionable/blocking comments AND `mergeable == "MERGEABLE"`** (up to date with base, CI green) — default-FAIL: assume findings are unaddressed until the fetch proves otherwise (treat `mergeable == "UNKNOWN"` as not-yet-proven and re-check) |51| **memory** | a per-PR fix log (findings addressed, findings deferred + why), git history, and the PR thread itself — all durable across the session-scoped `/loop` fires |52| **stop** | zero blocking findings **AND** mergeable **OR** poll/iteration cap **OR** no-progress for 3 rounds (same finding unaddressed 3×) **OR** budget cap **OR** an HITL checkpoint is hit |5354## The proof: zero blocking findings AND mergeable, default-FAIL5556"Healthy" is not "I replied to the last comment." It is **two conditions57observed together from `gh`**: the PR has zero *open, actionable, blocking*58review comments, **and** GitHub reports it mergeable (rebased onto base, CI59green). Assume the PR is **not** healthy until a fresh fetch proves both — that's60the default-FAIL stance, and it's why every pass *re-fetches* rather than61trusting last pass's verdict. A loop that stops after answering one comment ships62a PR that picked up three new ones and went stale behind a base push.6364Name the artifact explicitly: a `gh pr view --json mergeable,reviewDecision`65plus the open-comments count from `git-pr-feedback`'s fetch (see66`references/scheduling.md` for the exact query). Both must read clean *in the67same pass*.6869## Step 1 — One pass = one git-pr-feedback cycle7071The per-pass work **is** [`git-pr-feedback`]: identify the PR, fetch inline +72issue comments + review summaries (paginated), triage each by author and73category, address what's clear, reply on GitHub. **Do not re-document or74re-implement that** — invoke it. babysit adds only the loop discipline around it:75one root-cause change per pass, then re-check the *whole* PR state. If the base76branch moved and the only blocker is staleness, the pass is a routine rebase (not77a finding fix). The full per-pass algorithm and the handoff are in78`references/scheduling.md`.7980## Step 2 — Schedule it with /loop (the watcher, not /goal)8182Per `loop-controller` Step 1: this is a **watch/poll** job — you wait for review83activity to *change* on a cadence — so the primitive is **`/loop`**, the84scheduler, **not** `/goal` (which pushes to a finish line). The recipe is the85documented daily driver:8687```88/loop 5m /babysit89```9091`/loop` is **session-scoped, expires after ~3 days, and does not catch up** on92missed intervals — those mechanics (and the `loop.md` self-paced variant) live in93`loop-controller`'s `references/primitives.md`; don't re-document them. What goes94in `loop.md` and how to scope the run to one PR are in95`references/scheduling.md`.9697## HITL is load-bearing for this loop9899babysit runs **only inside the reversible boundary** unattended. The irreversible100or human-facing actions are **HITL checkpoints — never autonomous**:101102- **Force-push / rebasing a shared branch** — pause. A fast-forward push of your103 own commits is fine; a history rewrite of a branch others may have pulled is104 not.105- **Resolving or replying to a *human* reviewer's conversation** — pause. A106 bot/Copilot nit can be answered and resolved unattended; a human's thread is a107 conversation, and **you do not argue with a reviewer inside a loop.** On any108 substantive disagreement with a human reviewer, stop and surface it.109- **Merging** — never. babysit does **not** auto-merge, full stop. Reaching the110 proof (zero blocking + mergeable) means *ready for the human to merge*, not111 "merge it."112113Within the boundary — addressing Copilot/bot nits, routine fast-forward rebases,114re-running the gate, fast-forward pushes — it is AFK-safe. The full AFK-safe vs115pause table is in `references/scheduling.md`.116117## Guardrails specific to this loop118119Inherits the full stack from `loop-controller` → `references/safety.md`. The caps120this loop sets:121122- **Poll / iteration cap** — `/loop`'s ~3-day session expiry is the outer bound;123 set an inner per-PR pass cap (default ~20) so a wedged PR doesn't burn the124 whole window. Hitting it is a *stop-and-escalate*, not a reason to loosen the125 proof.126- **No-progress detection** — if the **same blocking finding** survives **3127 consecutive passes** (addressed but the reviewer/check still flags it, or it128 can't be resolved within the reversible boundary), stop and surface it. Three129 passes on one finding means it needs a human, not a fourth attempt.130- **Budget cap** — a watch loop that fires every 5 minutes for days adds up;131 enforce a token/cost ceiling that *terminates* the loop (read it from132 `.claude/profile.yaml` when present), not just warns.133- **Never cheat the proof.** Don't mark a thread resolved without addressing it,134 don't dismiss a human reviewer's comment to clear the count, and don't merge to135 make "mergeable" moot. A green that came from silencing a reviewer is a136 *finding*, not a win.137138## Choosing the driver primitive139140`/loop 5m /babysit` is the scheduler. The *per-pass* exit (this pass made one141clean, reversible change and re-checked the PR) is provable from `gh` output, so a142pass can run under auto mode unattended within the HITL boundary above. The loop143as a whole has **no `/goal` finish line** — it watches until you stop it or a stop144condition fires.145146## Long-run hygiene (wired per loop-controller Step 6)147148This loop runs unattended for days, so the Claude 5 long-run rules are part of149its contract, not inherited ambience (drop-in text: `model-adaptation` →150`references/long-run-hygiene.md`):151152- **Evidence-backed progress** — every per-pass report line (finding addressed,153 rebase done, gate green) traces to a tool result from *that pass*: the `gh`154 query output, the gate's exit code, the push result. Never narrate a fix the155 re-fetch didn't confirm.156- **Don't end a pass on a promise** — a pass ends with the change made and the157 whole PR state re-checked, or at an explicit HITL stop; never with "I'll158 rebase next pass" and no action. The autonomous-operation reminder applies159 while the user is away — the HITL checkpoints above are the only sanctioned160 pauses.161- **Budget is a harness decision** — the cost ceiling reads from externalized162 state (`.claude/profile.yaml`); don't surface a remaining-token countdown to163 the working pass, and never summarize-and-quit on a phantom context worry.164- **Effort per pass** — routine nit/rebase passes run at `medium`/`high`;165 reserve `xhigh` for a genuinely hard finding (tiering: `model-adaptation`).166167## Reference files168169- `references/scheduling.md` — the per-pass algorithm and the `git-pr-feedback`170 handoff, the scheduling recipe (`/loop 5m /babysit`, what to put in `loop.md`,171 scoping to one PR, the exact `gh` proof query), and the HITL boundary table172 (what's AFK-safe vs what pauses).173174[`loop-controller`]: ../loop-controller/SKILL.md175[`git-pr-feedback`]: ../../git/git-pr-feedback/SKILL.md176[`git-pr`]: ../../git/git-pr/SKILL.md177[`git-commit`]: ../../git/git-commit/SKILL.md