Release v0.4.0 - Auto-detection of shell commands vs AI prompts via `which` lookup on the first word - `!` prefix to force shell passthrough (e.g. `ait !git status`) - Default to AI prompt when first word is not a known binary - `AGENTS.md` loading from project git root — injected into system prompt for project-aware answers (`src/agents_md.rs`) - `find_git_root()` utility walking parent directories for `.git/` (`src/git.rs`) - `build_system_prompt()` prepends AGENTS.md content to base system prompt - `--debug` flag showing raw model response, input classification, AGENTS.md status, and parsed output mode - 19 new unit tests (cli detection, git root, agents_md loading, build_system_prompt, prose+command parsing) - Bumped `crossterm` 0.28 → 0.29, `toml` 0.8 → 1.0 - Input without `/ai` prefix now auto-classifies instead of always treating as shell command - Response parser now only treats pure bash blocks as `Command`; mixed prose+command responses are shown as `Chat` so the user sees the full context - System prompt hardened: bash blocks must contain executable commands only, never fabricated output; prefer prose when unsure