what is lia-trust fr?

LIA Trust Kernel is a safety and accountability layer for AI coding agents. Instead of trusting an agent to always do the right thing, this tool sits between the agent and the actions it tries to take, checks each action against a set of rules, and refuses to allow anything that fails those checks by default. Every decision it makes is written to a tamper evident log, signed cryptographically so that anyone can later verify the log has not been altered, even without needing to trust the machine that produced it. Installing it is a single script that builds a small program called lia and wires it into whichever supported coding tools are found on your machine, including Claude Code, Codex, Gemini CLI, and Cursor. Once installed, it hooks into each tool's own extension system so it can review actions before they happen. The install can also be limited to just installing the program without touching any configuration, or run as a dry run first to preview what it would change. Beyond basic gatekeeping, the project includes a feature for wrapping an entire agent process with a time limit, so if an agent runs too long without finishing, it gets stopped automatically and the outcome is recorded. On Linux systems that support certain modern isolation features, it can also run a wrapped process inside a restricted sandbox that blocks network access and limits which files can be written, and it can hand a short lived, limited credential to that sandboxed process without ever exposing it as a plain environment variable. The project is explicit that this protection is not complete: it only covers actions that actually pass through its hooks, and there are known gaps, such as processes that were already running before it started, or other programs running under the same user account. The authors are careful throughout the documentation to label exactly what has been measured and tested versus what is a stated goal, and they provide a log rotation system so that a long running history of decisions can be archived and independently verified without keeping every single record active.