Multi-agent orchestration skill for herdr — run Claude Code as a conductor that spawns and coordinates worker agents (codex, grok, gemini, another claude, …) in herdr panes over the socket CLI.
한국어 스킬입니다. herdr 안에서 실행된 Claude Code가 다른 CLI 에이전트들을 pane에 띄우고 협업·토론·작업 분배를 지휘하게 해줍니다. (Skill body is written in Korean; the commands and patterns are language-agnostic.)
- Conductor architecture — one Claude Code pane controls everything; workers need zero setup
- Orchestration patterns — parallel query, debate, pipeline, divide-and-conquer, cross-verify, state machine
- Field-tested Windows preview notes (2026-07-03, build
0.7.1-preview.2026-06-30) — measured behavior that differs from the docs:herdr wait output/wait agent-statusfail on this build → pollpane getinstead (reported upstream: herdr#963)pane read --source recent/recent-unwrappedreturn empty → always use--source visible(reported upstream: herdr#962)- Pane ID format is
w1:p1(not1-1), default pane shell is PowerShell 5.1 - Per-agent quirks: grok TUI scroll-out (use headless
-p), gemini free-tier, codex trust dialog
- Copy-paste helper functions —
agent_send/agent_read/agent_wait/agent_ask
npx skills add inbeomheo/herdr-orchestra --skill herdr-orchestra -gOr manually: copy SKILL.md to ~/.claude/skills/herdr-orchestra/SKILL.md.
Also install the official herdr skill for the base CLI reference:
npx skills add ogulcancelik/herdr --skill herdr -g- herdr running, with Claude Code inside a herdr pane (
HERDR_ENV=1) - Worker agent CLIs you want to conduct:
codex,grok,gemini,claude, …
The skill refuses to run when HERDR_ENV is not 1, so a Claude Code session outside herdr will not try to control your panes.
MIT