| name | coding-agent |
| description | Delegate coding tasks (build features, refactor, batch PR reviews, parallel issue fixes via Codex/Claude Code/Pi/OpenCode) by spawning the `coding-agent` sub-agent. Use when: (1) the work needs file exploration + multi-step edits, (2) you want a separate process so its tool noise stays out of your context, (3) you need parallel coding sessions. NOT for: simple one-liner edits (just edit), reading code (use read tools), thread-bound ACP requests (use sessions). |
| metadata | {"all_agents":{"emoji":"🧩","requires":{"anyBins":["claude","codex","opencode","pi"]}}} |
coding-agent (delegation skill)
This skill is a thin pointer: it tells you to spawn the coding-agent
sub-agent rather than running coding CLIs yourself.
The sub-agent lives at src/app/agents/coding-agent/ and owns the full
operational playbook (CLI selection, PTY rules, worktree etiquette,
report format). It runs in its own process, with its own tool budget,
so its shell output does not pollute the caller's context.
How to call it
subagent_spawn {
agent: "coding-agent",
message: "<one-line brief>\n\n<details: target dir, CLI preference, scope, success criteria>",
timeout: 1800
}
The brief should include:
- What to build / refactor / review.
- Where — absolute path to the target repo (the sub-agent will
cd there).
- Which CLI if you have a preference (
codex, claude, pi, opencode).
Default is codex.
- Scope guardrails — files / modules off-limits, branches not to touch.
- Success criteria — tests pass, PR opened, file produced, etc.
The sub-agent returns a short structured report (status / ran / workdir / result). Quote that report back to the user; don't paraphrase.
When NOT to spawn
- The change is one or two lines → just edit directly.
- The user only wants to read or search code → use file tools.
- The request is a chat-thread coding session bound to a Discord/IRC
thread → use the
sessions skill (subagent_spawn exits when done;
sessions keeps a long-lived ACP child).
- Work inside
~/.openclaw/ or the live OpenClaw checkout
~/Projects/openclaw/ → never spawn coding agents there.
Parallel work
For batch PR reviews or parallel issue fixes, spawn the sub-agent
multiple times in parallel — each call is a separate process. Mention in
the brief that the sub-agent should use git worktree so the runs don't
collide.
