ship-pr

name	ship-pr
description	Babysit a PR: iterate with AI reviewers and CI until green, get it ready, optionally squash-merge as Kody and watch the deploy, then send a Discord summary. Medium risk waits for AI reviewer(s) and addresses valid feedback. Use when a pull request needs to be shepherded to done.

Ship PR

Risk → merge authority

Self-assess; user policy overrides.

• Low — green CI; nits ignorable; squash-merge when policy allows.
• Medium — wait for AI reviewer(s); address valid feedback (ignore insignificant nits / already-fixed / wrong); then merge when policy allows.
• High — leave ready-for-review unless the user granted merge authority.

Loop

1. Mark ready — kody:@kentcdodds/github/pr/set-review-status { prUrl, status: 'ready' } (or owner/repo/prNumber).
2. Wait for CI — gh pr checks (or compose loop-on-ci / fix-ci).
3. Fix failures; for medium+, wait on AI reviewer(s) and address valid feedback. Rebase only when actually unmergeable. For medium+, also run npm run preview:manual-test as the seeded user with data for this change (--request / session cookie; see preview-manual-test skill).
4. Green + (medium+: valid feedback cleared) → break.
5. Push → repeat.

Gates ≠ CI

Blocked on soak / parity CHECK / calendar gate → end the run and schedule a wake (execute + workflows.create({ runAt, idempotencyKey }) calling createRun). Don't sleep-poll or code-thrash an intentional time window. Leftovers that wait on that gate need a GitHub issue (Cleanup: title); see cleanup-after-migrations.

Batch related expand steps into fewer PRs when risk posture allows.

Merge / deploy

When policy + risk allow: squash-merge via kody:@kentcdodds/github/pr/merge { prUrl, mergeMethod: 'squash' }, watch deploy. Useful: pr/get-checks, request, graphql on the same package.

Done → Discord

Always summarize (merged or not) by calling kody:@kentcdodds/discord/send-shipped-pr with structured fields. Do not use raw post-message and do not invent token cost or model id — the export fetches billed Cursor usage for cost, and you pass model from metadata.

agentId (required):

• In a Cursor Cloud Agent VM, read it from the metadata socket: curl -fsS --unix-socket "${CURSOR_AGENT_SOCKET:-/run/cursor/api.sock}" http://cursor-agent/v1/meta-data/agent/id
• Otherwise pass the bc- id from the agent URL you were launched as (https://cursor.com/agents/{id}).

model (deterministic — never infer):

• In a Cursor Cloud Agent VM, read the model that served this turn: curl -fsS --unix-socket "${CURSOR_AGENT_SOCKET:-/run/cursor/api.sock}" http://cursor-agent/v1/meta-data/turn/model. Auto resolves to the concrete model that served, not Auto. See https://cursor.com/docs/cloud-agent/metadata
• Outside a managed VM, pass the model.id used at create/launch if recorded.
• If missing, omit model so the export posts Model pending — do not invent one.

import sendShippedPr from 'kody:@kentcdodds/discord/send-shipped-pr'

export default async function main() {
	return sendShippedPr({
		agentId, // bc- id from the metadata socket or launch URL
		model, // turn/model from metadata socket (deterministic)
		title: 'PR title',
		summary: 'What shipped / parked / blocked and why.',
		prUrl: 'https://github.com/owner/repo/pull/1',
		repo: 'owner/repo',
		extras: ['CI green', 'preview verified'],
		status: 'Shipped', // or 'Parked' | 'Blocked'
	})
}