implement-night

Implement Night — end-to-end feature pipeline (thin orchestrator)

Build an already-approved feature plan all the way to a validated, ready-to-merge PR. This skill is a thin orchestrator: the real work lives in the sub-skills it invokes by name — /implement-task, /review, /review-ci — plus the agents those skills launch.

$ARGUMENTS is the approved Tasks Plan reference (path or feature slug) produced by /plan. If empty, ask the human which plan to build (and remind them to run /plan first if none exists).

You are the orchestrator — a MANAGER. You sequence the sub-skills, route rollback tasks, and enforce the gates. You do NOT write code, review the diff, or run tests yourself.

Read AGENTS.md first (tracker mode, retry caps, the pipeline map, and the cross-cutting rules).

Input: an approved Tasks Plan + the feature branch/worktree /plan created. Output: a CI-validated feature PR, ready for the human to squash-merge.

The pipeline blocks on the human only at the very end (squash-merge), plus the one optional self-improve prompt. (Plan approval already happened in /plan.) Everything between is automated; failures route back into /implement-task as rollup tasks rather than stopping the pipeline. Caps stop the pipeline: Tester FAIL 5/task, PA REJECT 3, PR-Reviewer 3, On-Call 5 — when a cap fires, surface USER ACTION REQUIRED and stop.

Critical rules:

• Never rubber-stamp a sub-skill's result. When /review reports "no blockers" or /review-ci reports "green", confirm the evidence is real before advancing.
• By name, never by path. Invoke /implement-task, /review, /review-ci by their skill names so it works when installed as a plugin.
• The orchestrator never writes code, never merges, never squashes.

---

Step 0 — Locate the branch worktree + tasks

/plan created the feature branch feat/{slug} — either in a dedicated worktree at ../{repo}-{slug} or in the current working tree — and wrote the feature's task files (tasks/<NNN>-*.md, status: pending) on it. Find where that branch is checked out:

git worktree list          # find the working tree whose branch is feat/{slug}

Confirm that working tree exists and contains pending task files (tasks/<NNN>-*.md with status: pending); WORKTREE_PATH is its path (the dedicated worktree, or the repo root if the branch lives in the current tree). If there's no matching branch or no pending tasks, stop and tell the human to run /plan first. Pass Working directory: {WORKTREE_PATH} to every sub-skill / agent so all work happens there.

---

Step 1 — Build the whole plan

Invoke /implement-task once with the feature's pending tasks. It loops over every task — SWE → Tester (FAIL max 5/task) → commit on PASS (flipping each task's status pending → in-progress → done) → next. (Do NOT loop here yourself; the per-task iteration lives inside /implement-task.)

invoke /implement-task with: the feature's pending tasks (tasks/<NNN>-*.md, status: pending), Working directory: {WORKTREE_PATH}

If /implement-task hits its Tester FAIL cap on a task, it stops with USER ACTION REQUIRED — propagate that and stop.

---

Step 2 — Review (loop until clean)

invoke /review with: feature {title}, Working directory: {WORKTREE_PATH}

/review pushes, creates/updates the PR, and runs PA acceptance then PR-Reviewer.

• Returns "clean" (no blockers) → proceed to Step 3.
• Returns a rollup task (PA REJECT or PR-Reviewer Blockers) → invoke /implement-task on that one rollup task (it builds + commits the fix), then re-invoke /review. Repeat.
• The PA REJECT (3) and PR-Reviewer (3) caps live inside /review; if it surfaces USER ACTION REQUIRED, stop.

---

Step 3 — Review CI

invoke /review-ci with: PR #{N}, Working directory: {WORKTREE_PATH}

/review-ci watches CI and drives it green (On-Call diagnoses → SWE fixes → re-check, max 5, self-contained).

• CI green → the PR is validated and ready to merge. Proceed to the tail.
• USER ACTION REQUIRED (cap hit) → stop.

---

Tail — self-improve (human-gated) + hand-off

The validated PR is ready. The orchestrator does NOT squash — per-task commits stay on the branch; the human uses GitHub's "Squash and merge".

Prompt the human:

Feature {title} is validated and ready to squash-merge: {PR URL}
Optional: run self-improve to capture corrections from this run into AGENTS.md? (y/N)

• N (default) → skip to the final summary.
• y → invoke the self-improve skill on this run; it proposes AGENTS.md updates the human reviews/accepts. When it returns, continue.

Then report the final summary:

## /implement-night complete — {Feature title}

**PR:** {URL} (validated; ready to squash-merge)
**Branch:** feat/{slug} ({N} per-task commits — GitHub squashes on merge)
**Worktree:** {WORKTREE_PATH}

**Tasks delivered ({N}):** {table — Tester / PA accept / PR-Reviewer / CI}
**Rollup tasks ({M}):** {list, or "none"}
**Self-Improve:** {ran / skipped}

Next: review the PR, then GitHub's **Squash and merge**. Remove the worktree with `git worktree remove {WORKTREE_PATH}` after merge.

The human squash-merges via GitHub. /implement-night ends here.

---

Notes on shape

• Thin by design. This file owns only the outer sequence + rollback routing. The SWE↔Tester loop, push/PR/acceptance/review, and CI handling all live in the sub-skills and the agent contracts — edit those, not this.
• Rollups re-enter through /implement-task. A /review rollup is built and committed like any other task, then /review re-runs. /review-ci handles its own CI fixes internally and does not re-enter /review.
• One human gate (merge) + one optional prompt (self-improve). Plan approval already happened in /plan. Don't add mid-pipeline gates.