prompt.md is the model prompt seed. harness.py loads the Cerebras-compatible OpenAI client, calls the model, and exposes one tool: shell. main.py is presentation only: it prints the transcript and can watch for changes.

runtime

Set CEREBRAS_BASE_URL, CEREBRAS_API_KEY, and CEREBRAS_MODEL_STRING in .env.
Run one turn with uv run main.py --run-once.
Start interactive chat loop with uv run main.py.
Watch prompt.md while chatting with uv run main.py --watch.
Watch-only (no stdin chat) with uv run main.py --watch --no-chat.

chat controls

Type a message and press Enter to append a user message and run a turn.
Use /run to force a turn without adding a user message.
Use /exit or /quit to stop.

session

The session file is append-only JSONL. Each event is either a model message, a tool result, or an error event. That keeps the loop dumb and easy to inspect.

shape

harness.py: model client, tool execution, transcript append.
main.py: formatting and file watching.
prompt.md: the behavior seed for the model.