What problem does it solve? Autonomous and parallel AI agents can edit files without investigating dependencies, run destructive commands like rm -rf or git push --force, write outside their assigned scope, and declare work done without recording it. Reasoning-based self-checks cannot catch these failures because self-evaluation shares the biases that produced the output. ## Core Features & Use Cases - Fact-forcing gate: Before the first edit to a file, the agent must produce concrete facts — importing files, affected public functions, actual data shapes, and the user's verbatim instruction. - Destructive-command guard: Every rm -rf, git reset --hard, DROP TABLE, kubectl delete, or --no-verify invocation requires an impact list and a rollback procedure before running. - Write-scope freeze and stop-gate: Freeze Mode locks writes to a named subtree for parallel agents, and a stop-gate blocks "done" until project-memory files were actually updated, verified via mtimes and counters. - Use Case: When dispatching several agents in parallel, each owning a slice of the repository, apply Gates 1-3 per agent so out-of-scope writes are refused at the moment they happen instead of discovered afterwards. ## Quick Start Ask the agent to apply the safe-execution gates before making any edits or running destructive commands in this session.
