open-coder-ai Block Destructive Commands
Best-effort guard against destructive commands: rm -rf targeting absolute, home ($HOME/~) or root-adjacent paths (and the PowerShell Remove-Item -Recurse equivalent); git push --force (not --force-with-lease); git reset --hard; git clean -f; kubectl delete; terraform destroy; aws s3 rm --recursive / rb --force; dropdb; helm uninstall/delete; docker volume rm/prune and system prune; gcloud ... delete. Destructive verbs are matched position-aware, so a bucket, path or object NAMED like a verb (aws s3 cp ... rm, docker volume inspect rm, helm list delete) is allowed. Known bypass classes include aliases, quoted arguments, non-standard clients, and scripts that invoke these commands indirectly. This is friction, not a security boundary.
block(destructive_command @position-aware): rm_-rf(/|~|$HOME|.)|Remove-Item_-Recurse, git_push_--force, git_reset_--hard, git_checkout_., git_clean_-f, kubectl_delete, terraform_destroy, aws_s3(rm_--recursive|rb_--force), dropdb, helm(uninstall|delete), docker_volume(rm|prune)|system_prune, gcloud_delete
require_approval: reset_hard|rm_-rf|branch_-D; prefer: stash|soft_reset|force-with-lease|dry-run
This skill is advisory: the client reading it has no mechanism to enforce it, and this policy stays advisory even when compiled by chock -- it ships rule text, not a blocking hook. See https://github.com/open-coder-ai/chock