Command Code (cmd)

cmd (Command Code) is a standalone terminal coding agent — a sibling to Claude Code / Codex — that continuously learns the user's taste of writing code. It has its own model access, permission system, session store, MCP support, and a "taste" personalization layer. This skill teaches you to (a) run it interactively, and (b) drive it headlessly as a sub-agent to implement work in parallel.

The high-value use is the second one: cmd is excellent at being dispatched — you write a precise spec, hand it an isolated worktree, and let it implement, test, and commit autonomously while you do other work.

The one invocation to memorize

For any unattended / headless / background dispatch, ALWAYS use:

cmd -p "<detailed self-contained prompt>" --yolo --max-turns 100000 --skip-onboarding

• -p/--print runs non-interactive: cmd executes the prompt and exits (no REPL). Required for scripting/background.
• --yolo bypasses every permission prompt (alias for --dangerously-skip-permissions). A headless cmd cannot answer prompts — without this it will hang forever the first time it wants to run a command or edit a file. This is mandatory for background runs.
• --max-turns 100000 effectively removes the turn cap. The default is only 10, and cmd exits with code 8 the moment it hits the cap — mid-implementation, leaving a half-done branch. Real implementation tasks need many turns; set it absurdly high so the task finishes on its own merits, not on an arbitrary limit.
• --skip-onboarding skips the interactive taste-onboarding demo, which would otherwise block a fresh headless run.