libardo667/kenshi-agent-env

Kenshi is a sandbox role-playing game set in a harsh desert world. This project builds a structured environment for running an AI agent inside that game. The goal is to let researchers experiment with whether a language model can play Kenshi, while keeping perception, planning, input, and game-integration failures clearly separated so they can be debugged on their own terms. The repository centers on a versioned observation protocol. The agent sees the game through screenshots and structured telemetry data, then decides what to do next. Every action, observation, and decision gets logged in a structured format that can be replayed later. Memory is stored in a local database, and plans follow strict schemas that bound what the agent can attempt at each step. Safety is a recurring theme. The default mode only uses ordinary Windows keyboard and mouse input, mimicking a human player. An optional native-assisted mode can also use a narrow internal command bridge, but it requires explicit acknowledgement and is never mixed silently with the interface-only path. Two independent gates must pass before real input is sent. A dedicated safety supervisor watches the stream and can cancel stalled actions, pause uncertain states, and report success only after confirming the result through a later telemetry update. The mock environment lets you run the full pipeline without owning the game. It shares the same reset, observe, step, and close interface as live mode, so tests written against the mock transfer directly. A heuristic baseline can complete a bundled one-day survival benchmark. The project also includes a continuous planning mode where one strategic call can authorize a bounded, multi-action plan with revision checks, retries, budgets, and cancellation handled by the executor. The native telemetry plugin has passed an initial in-game smoke test, emitting valid snapshots at two hertz and tracking character position, movement, pause, speed, and money. Broader field and input validation remains incomplete. The full README is longer than what was shown.