Currently we can only explicitly instruct Agent to run a completion or tool call by entering a user request in the UI. If we tell it to do X, then Y, then Z it will re-interpret those commands as it sees fit. The instruction likely gets batched in with the first LLM completion call, instead of injecting at the agent control level.
But agent does respond to the results of lint/test runs with a new LLM completion.
Could we provide a simple terminal command to respond with the action to take next? This would allow the agent to run off of a plan of our own choosing.
Come to think of it, this could be as simple as a shell command to slice the last line from an instructions file.