ANT

ANT is a CLI-first agent-native terminal experiment.

The first milestone is deliberately narrow:

ant diagnose tests

ANT detects a repository's test command, asks before running it, captures output, and asks an OpenAI model to explain failures. The MVP does not edit files and does not let the model generate shell commands.

MVP Principles

• Tasks are tracked as runs, not loose chat messages.
• Commands are deterministic and approval-gated.
• Run history is stored locally in ~/.ant/ant.db.
• The terminal stays transparent: command, cwd, risk, and data sharing are shown before execution.

Commands

ant doctor
ant diagnose tests [--workspace <path>] [--json]
ant run "diagnose failing tests" [--workspace <path>]
ant explain repo [--workspace <path>] [--json]
ant runs
ant inspect <run_id> [--json]

Environment

export OPENAI_API_KEY=...
export ANT_MODEL=gpt-5.4-mini
export ANT_DB_PATH=~/.ant/ant.db

Test Hooks

Automated tests use two explicit hooks so they do not need a live model or manual input:

ANT_MOCK_DIAGNOSIS=1
ANT_UNSAFE_ASSUME_YES_FOR_TESTS=1

Do not use the unsafe approval hook for normal development unless you deliberately want to bypass command approval.