autoship

Skill: Autoship — Automated Plan-to-PR

When to invoke: A plan already exists in docs/plans/ and you want to execute it all the way to an open PR without manually running each step.

Prerequisite: The plan must exist and be saved to docs/plans/[feature].md. If it doesn't exist yet, run /brainstorming then /plan first.

---

Modes

Mode	Trigger	When to use
Standard	"implement it with autopilot" / "autoship"	CI/CD handles deployment; Claude Code is on a developer machine
Deploy	"implement and deploy" / "autoship with deployment"	Claude Code is running directly on the staging server

---

What It Does

Standard mode:

Read plan → Health check → Self-assign → Worktree
  └─ For each task:
       Test Writer (RED) → RED gate → Implementer (GREEN+REFACTOR) → Reviews
  └─ After all tasks:
       /simplify → /review → Update docs → /pr

Deploy mode (everything above, plus):

  └─ After all tasks:
       /simplify → /review → Update docs
         → Detect deploy method → Deploy to staging
         → Health check loop → Integration/E2E tests
         → /pr (with deployment evidence in body)

The coordinator handles everything. You only intervene if:

• A task fails twice with no clear path forward
• /review finds Critical issues that require a design decision (not just a fix)
• Deployment or integration tests fail — never open a PR with a broken deployment

---

Protocol

Phase 0: Locate the Plan

Ask if not already provided:

Which plan should I execute? (e.g. docs/plans/user-auth.md)
Or describe the feature and I'll find the matching plan.

Read the plan file. Confirm it exists and has a task breakdown.

If deploy mode was triggered, also detect the deployment configuration now (see Phase 4D below) and confirm it before starting the task loop — better to discover a misconfigured deploy command before 30 minutes of implementation.

---

Phase 1: Plan Health Check (Before Any Code)

Review every task in the plan against these criteria. Fix problems now — they are much cheaper to fix in the plan than mid-implementation.

For each task, verify:

• Task description is unambiguous — one engineer reading it would implement the same thing as another
• FILES TO MODIFY lists exact paths (no "the service file" — must be backend/src/services/auth_service.py)
• Success criteria is verifiable ("test_X passes" or "endpoint returns 201") — not vague ("it works")
• A test can be written for this task (not auto-generated code or pure config)
• No task depends on a later task (execution order is correct)

If any task fails the health check:

• Fix it in the plan file directly
• Re-save docs/plans/[feature].md
• Report what was changed

Do not proceed to Phase 2 until every task passes the health check.

---

Phase 2: Setup

# Confirm clean state
git status  # must be clean

# Create worktree (if not already in one)
git worktree add -b feat/[feature-name] ../[project]-[feature-name] origin/main
cd ../[project]-[feature-name]

# Self-assign in checklist
# Edit docs/checklists/[feature]-checklist.md:
#   **Assignee:** autoship 🤖   **Status:** 🟡 In progress
git add docs/checklists/ && git commit -m "chore: self-assign [feature] for autoship"

---

Phase 3: Task Loop

Maintain a running status board and update it after every task:

## Autoship Status: [Feature]
| Task | Status | Commit |
|---|---|---|
| Task 1: ... | ✅ Done | abc1234 |
| Task 2: ... | 🔄 In progress | — |
| Task 3: ... | ⏳ Pending | — |

For each task:

Step A — Dispatch Test Writer (RED)

You are the Test Writer for Task N of [Feature] — RED phase only.

TASK: [exact task description]
TEST FILE: [exact path — e.g. backend/tests/unit/test_auth_service.py]
DESIRED BEHAVIOR: [one sentence]

Write ONE test: test_[function]_[condition]_[expected_outcome]
Run it. Paste exact output.

SUCCESS CRITERIA: FAILS with AssertionError, ImportError, AttributeError,
or TypeScript compilation error — NOT a syntax error.

STOP. Do not write production code.
SURGICAL CHANGES: Touch only the test file.

Step B — RED Gate

• Test exists at the correct path
• Test name follows test_[function]_[condition]_[expected_outcome]
• Output shows a meaningful failure (not SyntaxError)
• Test logic validates the actual desired behavior

Gate fails → retry once with specific feedback. Fails again → pause and report to user.

Step C — Dispatch Implementer (GREEN + REFACTOR)

You are the Implementer for Task N of [Feature] — GREEN then REFACTOR.

Failing test at [test file path]. Do NOT modify it.

TASK: [exact task description]
FILES TO MODIFY: [production files only]
SUCCESS CRITERIA: [test name] passes. No new failures in full suite.

GREEN: minimum code to make test pass — no gold-plating.
REFACTOR: only if code is unclear — run tests after every step.

SURGICAL CHANGES: production files only.
CONVENTIONS: [key items from CLAUDE.md]

Commit: git commit -m "feat/fix/refactor: [behavior]"
Report: test output, files changed, commit hash.

Step D — Task Reviews

Spec compliance:

• Code matches every requirement in the task spec
• Test passes, full suite clean
• No stubs (pass, TODO, NotImplementedError)
• Surgical check: only listed files modified

Code quality: invoke code-review skill

• Type safety, no security issues, follows conventions

Both pass → mark task ✅ in status board. Proceed to next task.

On Failure

• Task fails spec review → re-dispatch Implementer with specific failure feedback (retry once)
• Task fails twice → pause, update status board to ❌, report to user with exact blocker
• Do not skip a failing task and continue — later tasks may depend on it

---

Phase 4: Post-Implementation

Once all tasks are ✅:

Step 1 — Simplify

/simplify

Apply all suggestions. Run tests to confirm nothing broke.

Step 2 — Self-Review

/review

• Fix every Critical finding before continuing
• For Warnings: fix if straightforward, note in PR body if not
• Suggestions: optional

If Critical fixes are significant enough to require a new task, add it to the status board and execute it using Phase 3 before continuing.

Step 3 — Update Docs

Run /repo-docs — mandatory on every autoship run, no exceptions.

/repo-docs

/repo-docs updates only the sections affected by this branch's changes if docs already exist, or generates all docs from scratch if this is the first time. Commit any changes it makes:

git add docs/
git commit -m "docs: update project docs for [feature]"

---

Phase 4D: Deploy + Integration Tests (Deploy mode only)

Skip this phase entirely in standard mode. Only run when deploy mode was triggered.

Step 4D-1 — Detect Deployment Method

Check for these in order and use the first match:

Signal	Deploy command
docker-compose.yml or docker-compose.yaml	docker compose pull && docker compose up -d
Makefile with a deploy target	make deploy
package.json with a "deploy" script	npm run deploy
Procfile	foreman start or platform-specific command
None found	Ask the user for the exact deploy command before continuing

Record the deploy command. If it was inferred (not explicitly provided by the user), confirm it once before running:

Deploy command detected: docker compose pull && docker compose up -d
Proceeding — interrupt if this is wrong.

Step 4D-2 — Deploy

Run the deploy command. Capture stdout and stderr.

# Example for docker compose
docker compose pull && docker compose up -d
docker compose logs --tail=30

If the command exits non-zero: stop, report the exact error output, do not continue to health check or PR.

Step 4D-3 — Health Check

After deploying, confirm the app is responding. Try each endpoint in order until one returns 2xx:

curl -sf http://localhost:[PORT]/health      ||
curl -sf http://localhost:[PORT]/api/health  ||
curl -sf http://localhost:[PORT]/healthz     ||
curl -sf http://localhost:[PORT]/ping        ||
curl -sf http://localhost:[PORT]/

Retry every 5 seconds for up to 60 seconds (12 attempts). Record which endpoint responded and the HTTP status.

If still failing after 60 seconds:

1. Run docker compose logs --tail=50 (or equivalent) and report the output
2. Suggest rollback (see below)
3. Do not open a PR — the deployment is broken

Rollback suggestion on health check failure:

# Docker compose — restart with previous image
docker compose down && git stash && docker compose up -d

# Systemd service
sudo systemctl restart [service-name]

# PM2
pm2 restart all

Step 4D-4 — Integration and E2E Tests

Detect which integration/E2E test suites exist and run all of them:

Signal	Command
tests/integration/ directory	pytest tests/integration/ --tb=short -q
tests/e2e/ directory	pytest tests/e2e/ --tb=short -q
pytest.ini or pyproject.toml with integration marker	pytest -m integration --tb=short -q
playwright.config.ts or playwright.config.js	npx playwright test
cypress.config.ts or cypress.config.js	npx cypress run
package.json with "test:e2e" script	npm run test:e2e
package.json with "test:integration" script	npm run test:integration
Makefile with test-integration target	make test-integration

Run every suite that exists. Report pass/fail counts for each.

If any integration or E2E test fails:

1. Report the exact failing test names and output
2. Diagnose: is this a test environment issue, a data fixture issue, or a real regression?
3. If it's a real regression — fix it as a new task (add to status board, run Phase 3 loop)
4. Do not open a PR until all integration and E2E tests pass

Step 4D-5 — Record Evidence

Capture results for the PR body:

## Deployment Verified ✅
- Environment: staging (hostname: [hostname])
- Deployed at: [timestamp]
- Deploy command: [command used]
- Health check: ✅ [endpoint] → [HTTP status]
- Unit tests: [N/N passed]
- Integration tests: [N/N passed]  (or "none found")
- E2E tests: [N/N passed]  (or "none found")

---

Phase 4: Open PR

/pr

Standard mode — PR body includes:

• Summary of what was built (bullets per task)
• Link to docs/plans/[feature].md and docs/checklists/[feature]-checklist.md
• Autoship status board (copy the final table)
• Any warnings or known limitations from /review

Deploy mode — PR body also includes:

• The "Deployment Verified" evidence block from Step 4D-5
• Any integration/E2E test output worth highlighting

---

Phase 5: Wrap-Up

After PR is open:

1. Update checklist: all tasks ✅, status 🟢 Done, **PR:** #[number]

2. Commit checklist update to the branch:

   git add docs/checklists/
   git commit -m "chore: mark [feature] complete, PR #[number]"
   git push
   

3. Report to user:

   Standard mode:

   ## Autoship Complete ✅
   Feature: [name]
   Tasks completed: N/N
   PR: #[number] — [title]
   Review findings: [none / N warnings noted in PR body]
   

   Deploy mode:

   ## Autoship Complete ✅ (with deployment)
   Feature: [name]
   Tasks completed: N/N
   Deployed to: staging ([hostname])
   Health: ✅ [endpoint]
   Integration tests: N/N passed
   E2E tests: N/N passed
   PR: #[number] — [title]
   Review findings: [none / N warnings noted in PR body]
   

---

Model Selection

Phase	Model
Plan health check	Sonnet — requires judgment about ambiguity
Test Writer	Haiku — formulaic
Implementer (1-2 files)	Haiku
Implementer (multi-file)	Sonnet
Implementer (architecture)	Opus
Simplify + Review	Sonnet
Deploy diagnosis (on failure)	Sonnet — log analysis requires judgment

---

Hard Rules

• ❌ Never skip the plan health check — ambiguous plans waste more time mid-loop than fixing them upfront
• ❌ Never dispatch the Implementer before the RED gate passes
• ❌ Never skip a failing task and continue to the next
• ❌ Never open a PR with Critical review findings unresolved
• ❌ Never open a PR with failing unit tests
• ❌ Never open a PR with a failed deployment or failing integration/E2E tests (deploy mode)
• ❌ Never skip the doc update step — update or generate docs before calling /pr
• ❌ Never work from main — always from a worktree
• ❌ Never push manually — always exit through /pr

---

Failure Recovery

What failed	Recovery
Task fails twice	Pause, report exact blocker to user, wait for instruction
Deploy command exits non-zero	Report exact error, suggest checking logs, do not open PR
Health check times out (60s)	Print last 50 log lines, suggest rollback command, do not open PR
Integration test fails	Diagnose real regression vs env issue; fix before PR if real
E2E test fails	Same as integration — fix or document as known limitation with user approval

---

When to Use the Built-in /autopilot Instead

Situation	Use
Plan already written in docs/plans/	/autoship (this skill)
No plan yet, want everything from scratch	Built-in /autopilot
Single small task, no plan needed	/tdd directly
Complex plan, want adversarial plan critique before implementation	Built-in /autopilot to generate + critique the plan, then /autoship to execute it