project-recipes

Project recipes

Per-repo ops live in .claude/op.json, committed, never secret. SSH aliases (e.g. myhost) are fine; tokens and passwords never enter this file.

Schema

All keys optional. A command that needs a missing key runs Discovery; gated actions (deploy, publish) refuse rather than guess.

{
  "serve": "python3 -m http.server 8080",
  "build": "./build.sh",
  "test": "python3 -m pytest -q",
  "deploy": "ssh myhost 'cd /srv/www/<proj> && git pull && ./build.sh'",
  "sync": { "db": "<cmd>", "images": "<cmd>", "feed": "<cmd>" },
  "release": { "versions": ["package.json"], "publish": "npm publish", "verify": "<url>" },
  "live_url": "https://example.com",
  "cache_check": "curl -s \"$URL?v=$BUST\" | grep <marker>",
  "smoke": ["<url1>", "<url2>"],
  "rollback": "<cmd>"
}

$URL and $BUST are the only substitutions: the caller replaces them with live_url and a fresh cache-buster value at run time.

Discovery (on missing file or key)

1. Read AGENTS.md / CLAUDE.md, scripts/, package.json, Makefile, and any existing .claude/skills/ in the repo (per-repo ops skills often already document deploy and serve).
2. Draft the key(s) and present them as an arrow-choice: proposed value / edit / skip.
3. Write the confirmed keys to .claude/op.json (create the file if absent, merge if present). Show the diff.

Self-heal (on a failing recipe command)

1. Show the failing command and its error.
2. Rediscover that key (step 1 above), propose the fix as an arrow-choice.
3. Update the file on confirm, then retry once.

Hard stop: never invent a deploy, publish, or sync target that is not in the file or explicitly confirmed by the user this session.