Delegate bounded coding tasks to Command Code CLI and review diffs

Command Code Delegate hands a bounded coding task to the Command Code CLI (cmd), a separate implementer that writes code in your working tree while you stay the orchestrator writing the brief and reviewing the diff. Command Code's headless mode has exactly two states: the default (-p with no --yolo) allows read, grep, and glob but refuses every write, edit, and shell call, with no prompt to grant them mid-run; --yolo (alias --dangerously-skip-permissions) allows every tool everywhere the process can reach, with no filesystem sandbox or path restriction - --permission-mode auto-accept and --tools-all do not lift the headless write gate. A helper, scripts/relay.mjs, wraps cmd -p, captures the run, and writes a structured result.json with the implementer's finalMessage, its own gate claims, and touchedFiles; a run is finished when result.json exists with a status field, a missing cmd binary exits 127 but still writes a result with status commandcode_unavailable, and a pre-run usage error exits 2 with no result file. The relay never commits - the orchestrator re-runs the project's own gate commands and reviews the diff before committing the verified work.

Use it to delegate a bounded coding task you want implemented and then review yourself, or when the user explicitly asks for delegation to Command Code. Do not use it when the task is small enough to do inline, when the cmd CLI is not installed or not authenticated (cmd login), when you'd rather write the code yourself or only want a review (Command Code has its own /review), or on native Windows where cmdc --version fails - upstream recommends WSL there. Before any write-capable run, the skill calls for explaining the unsandboxed full-trust mode and getting explicit human acceptance - a request to delegate is not by itself consent to host-wide access.

Input is a written brief: the goal, current state, what to change and what to leave untouched, the project's actual gate commands (discovered from the repo's AGENTS.md/CLAUDE.md/Makefile, not assumed), and a report contract telling Command Code it will not commit - one task per brief. Prerequisites checked once beforehand: cmd --version succeeds and cmd status reports authenticated, the cmd (or cmdc on native Windows) binary is confirmed on PATH, and the target git repository's working tree is clean before dispatch. Output is a structured result.json carrying the implementer's finalMessage, touchedFiles, its gate claims, and, on a --read-only dispatch, a readOnlyViolation flag derived from Git-visible changes (null when git could not tell).

Wraps the commandcode CLI (cmd, or cmdc on native Windows) through a scripts/relay.mjs helper; that executable script is not bundled here - see the upstream project for the runtime. --read-only dispatches double as an adversarial second opinion: the brief lists agreed points and contested points for Command Code to defend or concede, touching no files. Delegation itself follows an authorization model - once a human has opted in, committing verified, gate-passing work is the agreed contract, with two limits: surface (not silently absorb) Command Code's own design decisions and defensible-but-unasked turns, and stop to ask rather than silently expand scope. Adapted from amElnagdy/delegate-skills (MIT).

Anyone orchestrating with Claude Code, OpenCode, or a comparable agent that can run a shell command and read a file, who wants to hand a bounded coding task to a separate implementer, review its diff against the brief and outside the paths it named, and land the verified result themselves.