wtf is pi-leakguard?

pi-leakguard is a safety extension for pi.dev, a coding agent, that tries to stop it from accidentally reading, writing, or leaking secrets like passwords and API keys. The author describes it as an independent continuation of a similar project's ideas, not an official fork, and it keeps the same MIT license. The README is careful to describe it as a defense in depth seatbelt rather than an airtight sandbox: it catches accidental leaks and common tricks used to extract secrets, but for untrusted code or automation running without supervision, it recommends using a real isolation tool like Docker instead of relying on this alone. When one of its protective layers is triggered, it can do one of three things. It can block the action entirely unless you confirm it should proceed, it can redact a secret from the output before the model ever sees it while still letting the agent know a secret existed, or it can simply warn you without stopping anything. The tool covers many angles: blocking reads and writes to sensitive paths like environment files and SSH keys, scrubbing known secret patterns like API keys and database URLs from command output, detecting attempts to bypass protection through symlinks or hidden characters, catching shell commands designed to dump or smuggle out secrets, and scanning file writes and git commits for secret material before they happen. Notably, the README explains that it deliberately does not try to detect prompt injection attacks by pattern matching, arguing that attackers can phrase an injection in endless different ways, so a fixed pattern list would miss most real attempts while frequently flagging normal, harmless text. Instead of trying to catch the injection itself, it focuses on limiting the damage an injection could cause, by redacting secrets and blocking the network calls that would carry them out. Four modes are available, ranging from full blocking with confirmation prompts down to fully disabled, and settings persist to a local configuration file.