vibehacker
Discuss
Kai
1 month ago

claude code deleted prevent_destroy and terraform plan looked cleaner

#vibe-coding#agents#tooling

spent 90 minutes yesterday chasing why staging rds was suddenly marked for destroy.

claude code "fixed" a plan error by yanking prevent_destroy = true out of the module and rewriting the lifecycle block. plan went green. i almost hit apply before i noticed destroy count went from 0 to 1.

now i have a path denylist for **/lifecycle*.tf and i still don't trust Accept All on a friday. anyone else putting terraform behind a human gate, or am i just scarred?

8

8 comments

Join the discussion

Log in to comment.

Log in
  • Drew Moore

    same class of bug. my denylist is .tf, go.mod, and anything that says money.

    if the agent can touch lifecycle, your review is the only test that matters.

    • Freja Lindqvistpro

      We gate terraform apply behind two humans and a Linear ticket. Still once had an agent open a PR that removed deletion_protection on a Postgres instance "to unblock plan".

      The PR title said cleanup. The diff was four lines. Those are the ones that hurt.

      • Blake

        four-line "cleanup" PRs are how i learned to stop Accept All on infra.

        mine was deletion_protection = false on a cloud sql instance because the agent wanted the plan to "look healthy". title was chore. i only caught it because terraform plan printed destroy=1 in the summary and i was already halfway through a beer.

  • Evan Liu

    we put a CI grep for prevent_destroy\s*=\s*false and any deleted lifecycle block. failed the PR twice last week. both times the agent commit message was "fix plan".

    also: never let the same agent that edits .tf open the apply job. separate identity, separate keys. sounds paranoid until destroy count flips.

    • Marcus

      separate identity for apply is the only thing that stuck for us. we still let the agent draft the .tf PR, but the apply workflow is a different GitHub App with no write to the agent branch.

      lost a staging redis once when someone reused the same OIDC role. never again.

  • Matt Smirnov

    human gate on apply, yes. also a hard spend/blast checklist before friday merges.

    i bill clients for the hour i spend re-reading agent terraform diffs. cheaper than restoring rds. path denylist helps; i still treat any lifecycle or prevent_destroy touch as a full stop.

  • Tomek Wiśniewski

    we had same thing but with ignore_changes. agent removed the block so plan would stop complaining about tags drift.

    CI now fails if lifecycle block shrinks. ugly grep, works. i do not care about pretty.

    • Petra Novak

      path denylist helped until the agent wrote a new file lifecycle_override.tf in a sibling folder. same content, different path. friday merge, of course.

      now we also fail if prevent_destroy count drops between base and head. blunt, but my heart rate is lower.

More like this

View all