mr-babysit

Babysit the MR(s)

Drive one or more merge requests toward mergeable: keep each rebased, fix failing CI, answer review comments that don't need product decisions, and wait for automated reviewers — passing back to the user only for genuine product/design calls or when everything solvable is solved.

This command is the author side of an MR. It is the mirror image of /code-review:watch, which is the reviewer side and never changes code.

This is not a rubber stamp. Review feedback — especially from AI review bots — is sometimes correct, sometimes wrong, sometimes pedantic, and sometimes proposes a fix that creates a new problem. Every finding is critically evaluated before any change lands, and this command is where that evaluation happens.

You are the orchestrator. You never edit code, never poll or triage the MR yourself, and never sleep. Two kinds of subagent do the work, and you keep the ledger:

• The watcher (glab:mr-watch, an agent from the glab plugin) — one, background, alive for the whole run. It runs the state script in a loop, interprets every change on every MR in the set, and messages you through SendMessage without ending its turn: a pipeline finishing, a push, new notes, a draft flip, a ticket move, the handshake reading, approvals, merged or closed. It is read-only and posts nothing. It is what makes you wait: nothing else in this command watches anything.
• The actor — the persistent implementation subagent when babysit runs as the last step of a /sdlc:write-plan plan and it still has context, otherwise a fresh subagent on opus. Background, resumed per batch, ending its turn after each. It reads ${CLAUDE_PLUGIN_ROOT}/skills/mr-babysit/references/actor-brief.md first; the brief holds every rule it shares across tasks. A triage batch turns what the watcher reported into proposed rows and changes nothing. An implementation batch does the approved rows: rebases, fixes, posts the fix replies and the dismissals, reclaims or marks ready, resolves, pushes, and reports by row.
• The ledger is yours. ./.claude/review-report/<topic>.babysit.md holds the set and what is or is not solved; you rewrite it after every decision. The watcher keeps its own ledger of movement and messages, so yours records no timeline.

At most two subagents are alive at once: the watcher and one actor. Four things re-invoke you, and nothing else: a watcher message, an actor returning, a 30-minute watchdog cron, and the user. Between them you are idle.