A small set of modules. No monolith.

The agent has to justify every edit or write before running it.

File edits wait for approval before they're applied.

Post-edit safety net. Suspends on rejection. Forces a re-review before the agent tries again.

AST-based bash interception. Blocks destructive commands before they run.

Lightweight isolation for sensitive data and network access.

Connects your local Pi instance to the self-hosted Bastille Console. Sends governance events for centralized audit, shared team policy, and team admin.

Full capture of HTTP requests and responses, every turn. The audit trail of what the agent actually did.

Cuts output token volume by roughly 50% or more.

Lossless session compaction. No LLM call. Nothing lost.

Auto-redirects pip, poetry, python, and venv to uv.

Spins up and manages agents and long-running processes through PTY sessions, not tmux.

High-leverage configs that enforce a methodology without capping the intelligence.

Manage your BASTILLE license from the Pi command palette - status, refresh, and clear.