claude code swapped my staging supabase url and friday deploy hit prod
asked claude code to "fix the auth redirect for staging" on a Next app. it rewrote .env.local and quietly pointed NEXT_PUBLIC_SUPABASE_URL at the prod project.
i only caught it because the friday vercel preview started writing test users into real tables. Activity Monitor was fine. my blood pressure was not.
do people keep env files out of the agent's write path, or is that just me being traumatized now?
5 comments
Join the discussion
Log in to comment.
repro checklist i run now before any agent env edit:
git diff -- .env*must be empty or i abort.env*in the agent sandbox if the tool supports ithit the same class of bug last month — badge said "connected", wrote into the wrong project id. status UI without a project ref is theater.
yeah i moved secrets to Vercel env only and keep a stub
.env.examplein the repo. agent can touch the example; it never sees real keys.still burned one Stripe test key that somehow lived in a leftover local file. cancelled that seat the same night. friday deploys with an agent in the loop are just expensive dice rolls.
same setup — vercel env + stub example. i also lock the preview branch to a dedicated supabase project with a red banner in the dashboard.
friday agent deploys without a project-ref screenshot in the PR description get a hard reject from me now. learned that the expensive way.
i keep
.env*in a deny list for every agent tool that supports it. still got burned once when cursor rewroteturbo.jsonand the env loader path changed under me.prod project ref in the URL is the only check i trust now. the dashboard name is cosplay.
i treat every agent env edit like untrusted input. after it runs i
git diffand require the project ref string to match a hardcoded allowlist in a tiny script.if the model invents a "staging" url that is just prod with a different subdomain, your eyes will miss it. been there.