deploy-guard

A small policy pack for teams that let an agent touch a repository which ships to production.

failproofai policies add chhhee10/deploy-guard

9 policies in three categories, 6 on by default. Requires failproofai.

Category	Policies	On by default
Deploys	3	2
Data	3	2
Repo hygiene	3	2

Look before you install

failproofai policies show chhhee10/deploy-guard

Manifest only — the code is never downloaded, so looking at a pack cannot run a pack.

Take part of it

failproofai policies add chhhee10/deploy-guard                          # the defaults
failproofai policies add chhhee10/deploy-guard --category deploys
failproofai policies add chhhee10/deploy-guard --policy block-prod-deploy
failproofai policies add chhhee10/deploy-guard --all

Scope it to particular agents — it guards all twelve otherwise:

failproofai policies add chhhee10/deploy-guard --cli claude codex

What is in it

Deploys

• block-prod-deploy — Block one-command deploys to production (vercel, fly, netlify, wrangler, serverless, eb)
• block-mutable-image-push — Block docker push of a mutable tag (:latest, :prod, :stable) — push an immutable tag instead
• warn-migration-deploy (opt-in) — Ask for a rollback plan before running a database migration

Data

• block-remote-db-shell — Block database shells pointed at a non-local host
• block-unscoped-sql-write — Block DROP / TRUNCATE and any DELETE or UPDATE with no WHERE clause
• warn-bulk-export (opt-in) — Flag a query that writes rows out to a file

Repo hygiene

• require-conventional-commit — Require a Conventional Commits prefix on every commit message
• block-lockfile-delete — Block deleting a lockfile — regenerating one silently moves every transitive dependency
• block-ci-config-edit (opt-in) — Block edits to CI workflow files — the thing that checks the work should not be edited by it

Why these

Each one is a thing that is fine for a person to do deliberately and wrong for an agent to do as a step inside something else. Shipping to production, opening a shell on a live database, deleting a lockfile, editing the CI config that checks the work — all reasonable actions, none of which should happen because a model decided they were the next move.

The three that are off by default are the opinionated ones: they warn about things that are often correct.

Editing them

git clone https://github.com/chhhee10/deploy-guard && cd deploy-guard
failproofai policies -i -c ./deploy-guard.mjs   # enforce it here, right now
failproofai publish --dry-run                   # build and validate, publish nothing
failproofai publish                             # ship it; the version counts itself

Licence

MIT.