deploy-steward — Claude Code skill: what it does, how to install
deploy-steward serializes PR merges into main, tracks CI gates, triggers deploys, and files repair tasks when checks fail.
MIT ★ 937 updated 2026-10-29
/plugin marketplace add SethGammon/Citadel
/plugin install citadel@citadel-localWhat deploy-steward does
The skill prevents landing stampedes when multiple agents finish work at once. It processes PRs sequentially using node scripts/deploy-steward.js. Input candidates come from .planning/pr-readiness/*.md reports or explicit URLs supplied via --enqueue-pr. State is saved to .planning/deploy-steward/queue.jsonl.
The steward acquires .planning/deploy-steward/lease.lock to prevent race conditions against main. It inspects PR state via gh pr view, updates branches behind base, waits for checks, and merges via gh pr merge. When checks fail or conflicts appear, it opens a repair task in .planning/intake/ and stops.
An authenticated GitHub CLI is mandatory. Running the steward halts without gh auth status passing, since raw tokens and arbitrary curl calls are rejected by design.
What it helps with
- serialize multiple pull requests into a safe mainline landing queue
- update out-of-date branches against main before CI validation
- execute custom deploy commands after each successful merge
- generate intake repair tasks when checks fail or conflicts arise
Example prompts for the agent
- Scan ready PRs and run the steward with the deploy command
npm run deploy - Enqueue PR
https://github.com/OWNER/REPO/pull/123and execute a dry run - Process the landing queue in merge-queue mode
How to install deploy-steward
1Plugin marketplace
/plugin marketplace add SethGammon/Citadel/plugin install citadel@citadel-localThis installs the whole citadel plugin with all its skills
2Skills CLI (Vercel)
npx skills add SethGammon/Citadel --skill deploy-steward -a claude-code3Manually: put the files in place
~/.claude/skills/deploy-steward/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/deploy-steward/
1Skills CLI (Vercel)
npx skills add SethGammon/Citadel --skill deploy-steward -a codex2Manually: put the files in place
.agents/skills/deploy-steward/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add SethGammon/Citadel --skill deploy-steward -a cursor2Manually: put the files in place
.cursor/skills/deploy-steward/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install SethGammon/Citadel skills/deploy-steward2Manually: put the files in place
.github/skills/deploy-steward/SKILL.md1Skills CLI (Vercel)
npx skills add SethGammon/Citadel --skill deploy-steward -a gemini-cli2Platform CLI
gemini skills install https://github.com/SethGammon/Citadel --path skills/deploy-stewardManually: put the files in place
.devin/skills/deploy-steward/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about deploy-steward
How to install deploy-steward in Claude Code?
Run /plugin marketplace add SethGammon/Citadel followed by /plugin install citadel@citadel-local. Alternatively, run npx skills add SethGammon/Citadel --skill deploy-steward -a claude-code.
What does deploy-steward do?
It turns parallel agent PRs into a single serial merge stream. It acquires a landing lease, verifies CI checks, updates stale branches, and invokes a deployment script after each safe merge.
What are the prerequisites for running deploy-steward?
You need an active GitHub CLI session (gh auth login) and either .planning/pr-readiness/ reports or specific PR URLs passed via --enqueue-pr.
Source text
SKILL.md · MIT
--- name: deploy-steward license: MIT description: >- Mainline deploy steward. Consumes ready PRs, owns a lease on mainline landing, refreshes PR state, updates stale branches, waits for CI and deploy gates, merges or queues one candidate at a time, and opens repair tasks for failures. user-invocable: true auto-trigger: false trigger_keywords: - deploy steward - deploy queue - merge steward - mainline steward - land prs - land PRs - deploy prs - deploy PRs - merge queue - release train effort: high last-updated: 2026-06-20 --- # /deploy-steward - Mainline Deploy Steward You are the deploy steward. You serialize mainline landing for parallel agent work. Your job is to turn many ready PRs into one safe mainline stream: refresh, update, wait for checks, merge, deploy, and open repair work when something fails. ## When to Use Use this when the user asks to: - land multiple ready PRs without a manual merge stampede - keep agent PRs rebased or branch-updated against latest main - merge only after required checks pass on current heads - run a deploy command after each safe serial merge - open repair tasks for failed checks, conflicts, or closed PRs Do not use this for code review quality findings; use `/review`. Do not use this for local worktree conflict analysis only; use `/merge-review`. ## Required Inputs The steward can start from either source: - `.planning/pr-readiness/*.md` reports generated by `node scripts/pr-ready.js` - explicit GitHub PR URLs passed with `--enqueue-pr` If `.planning/` does not exist yet, run the project setup/init flow first or create `.planning/pr-readiness/` before scanning. Treat a missing readiness directory as an empty queue, not as proof that there is nothing to land. The runtime requires: - GitHub CLI authenticated for the target repo - visible PR status checks unless `--allow-no-checks` is explicitly justified - a deploy command if the repo deploys outside the merge platform If `gh` is not installed or is not authenticated, stop before `--run` and ask the operator to run `gh auth status` / `gh auth login`. Do not substitute raw tokens or ad hoc curl commands inside the steward. ## Invocation Forms ```bash node scripts/deploy-steward.js --scan node scripts/deploy-steward.js --scan --run node scripts/deploy-steward.js --run --deploy-command "npm run deploy" node scripts/deploy-steward.js --enqueue-pr https://github.com/OWNER/REPO/pull/123 --run node scripts/deploy-steward.js --run --merge-mode merge-queue ``` Use `--dry-run` before first live use on a repo: ```bash node scripts/deploy-steward.js --scan --run --dry-run ``` ## Protocol ### Phase 1 - Build the Queue Run: ```bash node scripts/deploy-steward.js --scan ``` This reads `.planning/pr-readiness`, writes `.planning/deploy-steward/queue.jsonl`, and records each candidate as `ready`, `blocked`, or already terminal. For ad hoc PRs: ```bash node scripts/deploy-steward.js --enqueue-pr <url> ``` ### Phase 2 - Acquire the Steward Lease The run command acquires `.planning/deploy-steward/lease.lock` and mirrors the lease to `.planning/deploy-steward/lease.json`. If another steward has an active lease, stop instead of racing main. Use `--force-lease` only after confirming the old holder is dead. ### Phase 3 - Process One Mainline Candidate at a Time Run: ```bash node scripts/deploy-steward.js --run ``` For each processable queue item, the steward: 1. Refreshes live PR state with `gh pr view`. 2. Stops and writes a repair task if the PR is closed, conflicted, or unreadable. 3. Updates the branch against its base when GitHub reports it is behind. 4. Waits when checks are pending. 5. Opens a repair task when checks fail. 6. Merges with `gh pr merge --<method> --match-head-commit <sha>` when checks pass. 7. Runs the optional deploy command after a successful serial merge. If a branch update, pending check, merge queue wait, or repair task occurs, stop the run. Rerun after the external state advances. ### Phase 4 - Deploy For serial landing, pass the project deploy command: ```bash node scripts/deploy-steward.js --run --deploy-command "npm run deploy" ``` If the repo uses GitHub merge queue, prefer: ```bash node scripts/deploy-steward.js --run --merge-mode merge-queue ``` In merge-queue mode, the steward queues one PR with GitHub and marks it `landing`. Rerun after GitHub merges it so the next candidate can be assessed against the new mainline state. ## State Files - `.planning/deploy-steward/queue.jsonl` - durable landing queue - `.planning/deploy-steward/lease.json` - active steward lease mirror - `.planning/deploy-steward/runs/latest.md` - latest steward report - `.planning/intake/*deploy-steward*.md` - generated repair tasks ## Safety Rules - Never let multiple steward runs merge into the same target at once. - Never merge without a current PR refresh. - Never merge a PR with failing or pending checks. - Never merge a conflicted PR; create repair work instead. - Never force-push a contributor branch from the steward. - Use `--allow-no-checks` only when the repo truly has no CI. - Use `--require-fresh-readiness` when local readiness reports must remain the source of truth; otherwise live GitHub checks are authoritative after branch refreshes. ## Fringe Cases - **No `.planning/` directory**: stop setup and initialize Citadel state before writing queue or repair files. - **No readiness reports**: report an empty queue; accept explicit `--enqueue-pr` URLs if the operator wants to seed the queue manually. - **`gh` unavailable or unauthenticated**: do not run; surface the authentication problem and keep the queue unchanged. - **PR already merged**: mark the queue item `landed` and continue to the next candidate. - **PR closed without merge**: create a repair intake task and stop. - **Branch is behind base**: update the PR branch, mark it waiting for checks, and stop until CI reruns. - **Checks fail**: create a repair intake task and stop. - **Checks pending**: keep the item queued and stop until the external state advances. - **Merge conflict or dirty merge state**: create repair work; never merge or force-push from the steward. ## Quality Gates - `node scripts/test-deploy-steward.js` - `node scripts/skill-lint.js deploy-steward` - `node scripts/test-routing-sync.js` - `node scripts/test-all.js` ## Exit Protocol Return the steward report and a handoff: ```text ---HANDOFF--- - Queue: <N> candidates in .planning/deploy-steward/queue.jsonl - Last action: merged | updated-branch | waiting-for-checks | repair-needed | idle - Report: .planning/deploy-steward/runs/latest.md - Repair task: <path or none> --- ```
© SethGammon · MIT · SethGammon/Citadel