Deliver
Deliver: $ARGUMENTS
An orchestrator. You approve the spec. Everything after that runs without asking, except the gates listed below. The run keeps its state in .claude/state/runs/<branch-slug>/tasks.md, so it survives compaction and a resumed session.
Stages
why-no-hook: skill workflow guidance; each step requires understanding the surrounding context (repo, task shape, prior state).
- Research: run
/research when the code is unfamiliar. Skip it for code already read this session (review-time: see section note)
- Spec: run
/spec. Discovery asks one question at a time. The spec's Open Questions must be empty before you approve it. Notify with ~/.agents/scripts/notify.sh "Spec ready for review" (review-time: see section note)
- Set up, after approval:
(review-time: see section note)
- Create the branch and worktree with
/worktree <slug>, and set the spec's branch: frontmatter (review-time: see section note)
- Write
tasks.md with one line per stage, and tick stages as they finish (review-time: see section note)
- Print this line for the user to paste as
/goal, which keeps the session working across turns: The PR for <spec> is open, every automated criterion passes at HEAD, and ~/.agents/scripts/pr-ready.sh prints READY or WAITING_ON_HUMAN for it. The run continues without it (review-time: see section note)
- Plan: run
/plan. It asks only about an architectural choice the spec left open (review-time: see section note)
- Build: run
/build, which runs each slice through its Slice Loop. A plan spanning two or more repos goes to /drive-fleet with the plan instead (review-time: see section note)
- Verify: spawn the
Spec Verifier for the whole branch, then run /verify-done. Fix and repeat until both pass (review-time: see section note)
- Open the PR: run
/mr. It reuses the step 6 pass when nothing changed since. List manual criteria under Blocked on me in the PR body (review-time: see section note)
- Watch: follow the
[pr-watch] steps until ~/.agents/scripts/pr-ready.sh prints READY or WAITING_ON_HUMAN. Resolve each bot thread after replying. Human-reviewer replies go to the user as drafts (review-time: see section note)
- Report: end with Blocked on me, Changed, and Found, plus the PR URL. Notify with
~/.agents/scripts/notify.sh "PR ready: <url>" (review-time: see section note)
Gates the user keeps
Stop and ask only for these (review-time: see section note):
- Approving the spec
- Posting a reply to a human reviewer
- Force-pushing, merging, or pushing to
main
- Deploying or applying to production
- Deleting data, or changing anything outside the repository and its PR
Budget and escalation
- Three review rounds per slice, from
/build (review-time: see section note)
- Three attempts at the same CI failure, from
/ci (review-time: see section note)
- The same failure twice on any other step
(review-time: see section note)
- A finding or conflict that invalidates the plan: update the plan, and ask only if the scope changes
(review-time: see section note)
On escalation, notify, add the item under Blocked on me in tasks.md, and keep working on slices that do not depend on it (review-time: see section note)
1---2name: deliver3description: "Deliver a feature end to end with one human gate: you approve the spec, then the agent plans, builds tests-first, verifies against the spec, opens the PR, and drives CI and review comments until it is ready to merge."4disable-model-invocation: true5metadata:6 source: https://skillmd.com/skills/domengabrovsek/deliver7---89# Deliver1011Deliver: $ARGUMENTS1213An orchestrator. You approve the spec. Everything after that runs without asking, except the gates listed below. The run keeps its state in `.claude/state/runs/<branch-slug>/tasks.md`, so it survives compaction and a resumed session.1415## Stages1617**why-no-hook:** skill workflow guidance; each step requires understanding the surrounding context (repo, task shape, prior state).18191. **Research**: run `/research` when the code is unfamiliar. Skip it for code already read this session `(review-time: see section note)`202. **Spec**: run `/spec`. Discovery asks one question at a time. The spec's Open Questions must be empty before you approve it. Notify with `~/.agents/scripts/notify.sh "Spec ready for review"` `(review-time: see section note)`213. **Set up**, after approval: `(review-time: see section note)`22 - Create the branch and worktree with `/worktree <slug>`, and set the spec's `branch:` frontmatter `(review-time: see section note)`23 - Write `tasks.md` with one line per stage, and tick stages as they finish `(review-time: see section note)`24 - Print this line for the user to paste as `/goal`, which keeps the session working across turns: `The PR for <spec> is open, every automated criterion passes at HEAD, and ~/.agents/scripts/pr-ready.sh prints READY or WAITING_ON_HUMAN for it.` The run continues without it `(review-time: see section note)`254. **Plan**: run `/plan`. It asks only about an architectural choice the spec left open `(review-time: see section note)`265. **Build**: run `/build`, which runs each slice through its Slice Loop. A plan spanning two or more repos goes to `/drive-fleet` with the plan instead `(review-time: see section note)`276. **Verify**: spawn the `Spec Verifier` for the whole branch, then run `/verify-done`. Fix and repeat until both pass `(review-time: see section note)`287. **Open the PR**: run `/mr`. It reuses the step 6 pass when nothing changed since. List manual criteria under **Blocked on me** in the PR body `(review-time: see section note)`298. **Watch**: follow the `[pr-watch]` steps until `~/.agents/scripts/pr-ready.sh` prints READY or WAITING_ON_HUMAN. Resolve each bot thread after replying. Human-reviewer replies go to the user as drafts `(review-time: see section note)`309. **Report**: end with **Blocked on me**, **Changed**, and **Found**, plus the PR URL. Notify with `~/.agents/scripts/notify.sh "PR ready: <url>"` `(review-time: see section note)`3132## Gates the user keeps3334Stop and ask only for these `(review-time: see section note)`:3536- Approving the spec37- Posting a reply to a human reviewer38- Force-pushing, merging, or pushing to `main`39- Deploying or applying to production40- Deleting data, or changing anything outside the repository and its PR4142## Budget and escalation4344- Three review rounds per slice, from `/build` `(review-time: see section note)`45- Three attempts at the same CI failure, from `/ci` `(review-time: see section note)`46- The same failure twice on any other step `(review-time: see section note)`47- A finding or conflict that invalidates the plan: update the plan, and ask only if the scope changes `(review-time: see section note)`4849On escalation, notify, add the item under **Blocked on me** in `tasks.md`, and keep working on slices that do not depend on it `(review-time: see section note)`