deploy-safely

name	deploy-safely
description	Deploy Your Project to production safely. Use before any deployment, after code changes, or when asked to push to prod.
metadata	{"version":"2.0.0","author":"Master Agent","category":"operations","updated":"2026-03-30T00:00:00.000Z"}

Deploy Safely — DevOps

You are responsible for getting code safely to production on Fly.io.

Before Starting

Check:

• Are there uncommitted changes? git status
• Did smoke tests pass? python3 smoke_test.py
• Is this approved by Master/S&QA?

How This Skill Works

Mode 1: Standard Deploy

Code changes committed, smoke test passing, deploy to Fly.io.

Mode 2: Hotfix Deploy

Critical bug found, need to deploy ASAP. Skip non-essential checks but ALWAYS smoke test.

Mode 3: Verify Only

No deploy — just run smoke tests and health checks. Report status.

Deploy Checklist

Pre-Deploy

• python3 -c "import ast; ast.parse(open('changed_file.py').read())" for each changed file
• python3 smoke_test.py — must be 48/48
• git add {specific files} — NEVER git add -A or git add .
• git commit -m "descriptive message"

Deploy

# Preferred (remote build — avoids local disk issues)
~/.fly/bin/flyctl deploy --remote-only

# Fallback (local Docker build)
~/.fly/bin/flyctl deploy --local-only

Post-Deploy

• curl -sL -o /dev/null -w "%{http_code}" https://your-project.fly.dev/health — expect 200
• Test specific endpoints if relevant
• Report to Master

Proactive Triggers

• pricing.py in changeset → needs git add -f (gitignored)
• mcp_server.py changed → remind Master that PyPI publish needed for installed clients
• db.py schema changed → migration will run on next startup, verify it's backward compatible
• Deploy fails "no space left" → run docker system prune -af, retry with --remote-only
• Post-deploy 500 errors → check Fly logs: ~/.fly/bin/flyctl logs --no-tail | tail -20

Gotchas

• Some networks block .ai TLD — use fly.dev for health checks
• fly.dev redirects non-/health paths to custom domain — use SSH for internal API testing
• Fly SSH tunnel can go down transiently — retry after 5 mins
• Metrics token warning is harmless — ignore it
• Smoke test may show connection resets for ~30s after deploy (app restarting)
• --local-only filled local disk during extended sessions — prefer --remote-only