ai-cli

Go CLI. PTY-based wrapper around AI coding CLIs with automatic limit detection and provider switching.

Structure

main.go — entry point, arg routing, ai switch (SIGUSR1 to all PIDs), clipboard helper
internal/config/ — YAML config loading, default limit patterns from real CLI source
internal/proxy/ — PTY proxy: spawns CLI in pseudo-terminal, monitors output, SIGUSR1/SIGQUIT signal handling, provider picker menu (raw mode single-keypress), multi-PID management
internal/session/ — session persistence, Claude Code JSONL session file extraction (~/.claude/projects/<project>/<id>.jsonl), git diff/log capture, continuation prompt builder
internal/logger/ — append-only log file
test/ — fake provider scripts for testing limit detection flow

go build -o /usr/local/bin/ai .

~/.config/ai-cli/config.yaml — provider chain, permission bypass args/env, limit patterns.

Full PTY passthrough — user sees actual TUI, no filtering
Output ring buffer (8KB) scanned for limit patterns (ANSI-stripped via regex)
ai switch sends SIGUSR1 to ALL running ai processes via PID files in ~/.local/share/ai-cli/pids/
Context extraction reads Claude Code JSONL session files directly, parses user/assistant messages
Extra CLI args only forwarded to initial provider, not fallbacks (prevents flag leakage)
Provider picker uses raw mode for single-keypress selection
Permission bypass baked into config per provider (different mechanisms: CLI flags for Claude/Codex, env var for OpenCode, none needed for Pi)

Claude Code: --dangerously-skip-permissions flag. Sessions stored as JSONL at ~/.claude/projects/
OpenCode: OPENCODE_PERMISSION='{"*":"allow"}' env var (TUI doesn't support --dangerously-skip-permissions, only opencode run does). Model format: provider/model (e.g. google/gemini-2.5-flash)
Codex: --full-auto flag. Model format: just model name (e.g. o3)
Pi: No permission system. Model format: pattern match without provider prefix (e.g. gemini-2.5-flash). Binary requires full nvm path in config due to lazy-loaded nvm.