| name | deploy-and-verify |
| description | Use when deploying an application and verifying the deployment works — reads DEPLOY.md, executes deployment, runs smoke tests or E2E tests, reports structured results. Triggers on "deploy", "deploy and verify", "上线", "验证", or after finishing a development branch. |
deploy-and-verify: Deploy and Verify
Deploy the application and verify the deployment works.
Trigger: User asks to deploy, deploy and verify, 上线, 验证. Or agent judges deployment is the natural next step after implementation completes — subagent tasks done, plan execution done, simple task done, or branch finish.
Workflow
1. Read Deployment Config
Read DEPLOY.md from the project root. If it exists, extract:
- Deploy commands (the steps under "Deployment" section)
- Verification commands or URLs (health endpoints, E2E test commands)
- Environment info (ports, URLs)
If DEPLOY.md does not exist:
- Check if the
deploy skill's template is appropriate — offer to create DEPLOY.md using that template
- If user declines, proceed with best-effort deployment (look for
package.json scripts, docker-compose.yml, Makefile targets)
2. Execute Deployment
Run the deploy commands from DEPLOY.md (or discovered equivalents) using Bash tool directly.
Pre-deploy cleanup: If the deploy involves starting a server, kill any existing process on the target port first:
lsof -ti:$PORT | xargs kill -9 2>/dev/null; true
Execute each deploy step sequentially. If any step fails, stop and report.
3. Verify Deployment
After deployment succeeds, verify it works. Try in order — use the first that applies:
- DEPLOY.md has verification commands → run them
package.json has test:e2e script → run npm run test:e2e- Deployed a HTTP service →
curl the health/root endpoint
- None of the above → report "SKIPPED: no verification configured"
4. Report Results
Report structured results:
## Deploy & Verify Results
- **Deploy:** SUCCESS | FAILED
- **Verify:** PASSED | FAILED | SKIPPED
- **URL:** <access URL if applicable>
- **Evidence:** <command output showing success or failure>
If FAILED at deploy or verify stage:
- Include the error output
- Suggest a likely cause (code bug / config issue / infrastructure)
- Do NOT automatically retry — let agent/human decide next step
Principles
- Linear steps. No state machine, no loop counting.
- Report, don't retry. Failure is information, not a trigger for automatic action.
- Works with Claude Code alone. No Agent-Orchestrator dependency.
- Reuses DEPLOY.md convention from the
deploy skill.
Red Flags
| Thought | Reality |
|---|
| "Deploy probably worked" | Verify with evidence — curl output, test results |
| "I'll skip verification" | Verification is the whole point of this skill |
| "Let me retry automatically" | Report failure, let agent/human decide |
| "I need ao spawn for this" | Use Bash tool directly — this is a single-agent operation |
Integration
Triggered by (any of):
- subagent-driven-development — after all tasks complete and two-stage review passes
- executing-plans — after all batches complete
- to-do-it — after simple task completes (suggestion only, not automatic)
- finishing-a-development-branch — after branch finish (optional, still valid)
- Manual — user says "deploy", "上线", "验证"
Works before: value-proof (optional)
Uses convention from: deploy skill (DEPLOY.md)
Relationship with deploy skill: The deploy skill handles DEPLOY.md awareness and maintenance — it reminds you to update DEPLOY.md when infra changes. This skill (deploy-and-verify) handles actual deployment execution and verification. Use deploy for context, deploy-and-verify for action.
If deployment is not needed (e.g., library package, skill-only repo): skip this skill and go directly to value-proof or simply declare the work done.
