cli-delegation

CLI subagent skills

Agent Skills that let a coding agent drive another coding CLI programmatically — for a second opinion, a code review, or a delegated worker run. Each delegate-to-* skill ships a read-only reviewer and an edit-capable worker/profile.

The important trick is not just "how to launch the CLI"; it is how to launch it without losing control of the main checkout. Prefer isolated branches/worktrees, machine-readable output, and explicit run state.

Skill	CLI	Purpose
delegate-to-claude-code	claude	Delegate reviewer/editor runs; --permission-mode plan vs acceptEdits / auto
claude-remote-control-server	claude remote-control	Run persistent per-repo Remote Control servers under systemd
delegate-to-codex	codex exec	Delegate reviewer/editor runs; -s read-only vs workspace-write
delegate-to-opencode	opencode run	Delegate reviewer/editor runs; primary/all agents with edit: deny vs edit: allow
fastcontext	fastcontext	Delegate read-only repository exploration; returns file:line citations without spending your context

Install

Each skill is a plain Agent Skills folder, so any SKILL.md-aware tool can consume it. Pick a rail:

skills CLI (installs into Claude Code, Codex, opencode, Cursor, and many others):

npx skills add T0mSIlver/skills                               # pick interactively
npx skills add T0mSIlver/skills -a claude-code -a codex -g -y # everything, globally

Claude Code plugin marketplace:

/plugin marketplace add T0mSIlver/skills
/plugin install cli-delegation@t0msilver-skills
/plugin install claude-rc-server@t0msilver-skills

Manual: copy any top-level skill directory into your agent's skills folder (~/.claude/skills/, ~/.codex/skills/, ~/.config/opencode/skills/, …).

Some skills need more than the folder copy — each declares its requirements in compatibility: frontmatter:

• fastcontext requires the separate fastcontext CLI on PATH; the skill checks for it at load and points you there if missing.
• claude-remote-control-server sets up a user systemd service via its bundled scripts/install-claude-rc-server-service.sh (nothing runs at install time — the skill walks the agent through it).
• delegate-to-claude-code prefers its scripts/claude-rc-spawn helper (needs tmux) on PATH for remote-visible sessions; plain claude -p delegation works without it.

Shared conventions

• Worktree first for edits. Launch edit workers in a new branch/worktree so long runs do not modify the main agent's checkout. Commit or patch in only the exact local state the worker needs; do not blindly git add -A unrelated work.
• Remote-visible Claude sessions. For Claude Code delegation, prefer claude-rc-spawn: it starts interactive Claude in detached tmux with Remote Control enabled, injects the prompt, and leaves a session the user can inspect from claude.ai/code.
• Prompt as a file. Write the brief to a markdown file with context, task, constraints, acceptance criteria, and required output shape. Pass its contents as the prompt argument, or attach it, instead of hand-writing a large inline string.
• Capture run state. Save the harness output, session id, branch, worktree, and prompt path. Long runs need a handle for polling, resume, cleanup, and review.
• Reviewer after worker. Treat the edit worker's final message as a claim. Run a fresh read-only review against the diff before merging, cherry-picking, or opening a PR.

Common gotchas