Design the Agent-Computer Interface (ACI)

PROJECT CONTEXT:
My agent has shell access and file tools but flails: it gets lost in long outputs, misuses commands, and misses errors. SWE-agent's research showed the interface between agent and computer is a measurable performance lever.

GOAL:
Redesign my agent's tool surface as a deliberate ACI: constrained commands, model-friendly viewers, and feedback formats that guide the next action.

REQUIREMENTS:
- An audit of current tools against ACI principles (simplicity, feedback quality, guardrails)
- A file viewer with windowed output, line numbers, and navigation state
- Edit commands with built-in validation feedback (lint/syntax check on write)
- Search tools that return concise, ranked results instead of raw dumps
- Every tool response formatted to guide the model's next action

CONSTRAINTS:
- Prefer few, purpose-built commands over exposing the entire shell
- Every tool output must fit a bounded token budget

DEFINITION OF DONE:
- The agent completes a bug-fix task with fewer turns than the raw-shell baseline
- File viewing shows a window with position indicators, never whole large files
- Failed edits return actionable errors (line, reason) and the file is unchanged
- Search results are capped, ranked, and deduplicated

COMMON FAILURES TO AVOID:
- Giving the model a bare shell and calling it an interface
- Tool outputs that dump thousands of lines and drown the signal
- Edits that silently corrupt files because nothing validated them
- Error messages written for humans (stack spew) instead of models (cause + next step)