/ship-it — PRs → merged → deployed → verified

One idempotent tick (target ≤2 min of action; waiting happens between ticks). State lives in <repo>/.local/state/ship-it-<prd-slug>.json (create dir if missing; ensure .local/ is gitignored): PR list in stack order, per-PR status, deploys monitored, fixes attempted, soak results. Re-running against unchanged state must produce the same outcome.

Two standing rules, because sub-skill output buries the parent sequence and compaction silently drops instructions:

• After EVERY skill returns, restate position: "→ Returning to /ship-it: next is step N (<name>)."
• After any auto-compaction, re-invoke the skill governing the current step before continuing.

Goal (tick-scoped): set a session goal now (SlashCommand /goal if available, otherwise state it and treat it as binding): "This tick's fixes applied, state file updated, summary printed." Overall completion lives in the state file — the final tick announces DONE and tells Luc to stop the /loop.

Hard rules: everything goes through PRs — never mutate staging/production directly (no flyctl deploy/secrets set, no dashboard changes; flyctl/Sentry/PostHog are read-only investigation tools). Never merge over red or pending CI. CI/CD is law.

Tick

1. Resolve + Skill: `learn-existing` — focus pr:<next unmerged PR in stack order> (from $ARGUMENTS or state file; the first tick builds the PR list from the EXECUTION-REPORTs / plan/INDEX.md of the PRD's — or, when scoped to a single Initiative, just that unit's — brainstorm folders, records the scope (full-prd | initiative:<id> | pr-list) in the state file, and sets the PRD's Index.md row to Shipping — create Index.md if missing).
2. Skill: `babysit-pr` — one tick on that PR: sync, rebase, fix failing CI, address comments, queue auto-merge. Its strict no-admin gate stays intact.
3. Review-bypass extension (owned by /ship-it, not the skill) — if and only if: every check terminal-green (SUCCESS/SKIPPED/NEUTRAL), mergeable=MERGEABLE, and reviewDecision=REVIEW_REQUIRED is the sole reason the merge is blocked (i.e. mergeStateStatus=BLOCKED purely for the missing approval — nothing red, nothing pending) → gh pr merge <n> --admin --squash. Log it in the state file. This is the single sanctioned --admin use; anything red or pending waits for the next tick.
4. Skill: `control-deployment` — if a merge landed this tick or a deploy is in flight: monitor the pipeline → staging health → observability soak. Deploy issues become fix PRs, which join the state file's PR list (and may bypass review under the same green-CI-only rule).
5. Skill: `verify` — focus deployment, once a repo's deploys report LANDED: staging endpoints + e2e smoke against staging where the repo supports it.
6. Stacked-PR maintenance — after a merge, the next PR in the stack usually needs a rebase onto main (squash merges orphan stacked commits); let babysit-pr handle it next tick.
7. Update state + print the tick summary: per-PR table (PR | checks | merged? | deploy | soak), actions taken, blockers, what the next tick will do.

Completion (final tick)

When every PR in the state file is merged, every deploy LANDED, and the soak window is clean — what happens next depends on the recorded scope:

• `full-prd` (every Initiative's PRs accounted for): Index.md row → Done, move the PRD + its Initiative folder into 03-Product/PRDs/DONE/, set frontmatter status/last-updated.
• `initiative:<id>` or `pr-list` (partial run): set only that Initiative's frontmatter status: shipped + last-updated; the PRD row stays Shipping and NOTHING moves to DONE/ — a partial scope must never archive the whole PRD.

Announce DONE (<scope>) — stop the /loop.

If a tick discovers an unfixable-by-PR situation (e.g. requires Luc's credentials or a paid-plan change), write it to the state file as BLOCKED-HUMAN, surface it loudly in the summary, and keep ticking the rest.