The agent is constantly assuming it is in a particular and guessing wrong. We could conserve a lot of wasted tokens, time, and potentially even dangerous command executions if the agent is informed of the working directory before it writes a command.
The entire prompt line (or two) would be useful if you have the agent working in a DB CLI or it needs to activate a venv or conda env before activating a command.
In short, I need the agent to know its command context before it writes a command.