# Codex Orchestrate

> Orchestrate a fleet of parallel `codex exec` workers with you as supervisor — one per isolated git worktree, dispatched headless, each verified INDEPENDENTLY, then PR/merge. The manual "codex-ultracode" fan-out pattern with the hard gotchas baked in (stdin blocking, background tracking, don't-trust-self-reports, writer isolation). Triggers: orchestrate codex, codex workers/fleet, spawn codex, delegate to codex in parallel, manual ultracode, 开 codex 小弟, 派 codex worker — NOT for a single cross-vendor opinion (second-opinion-codex agent) or web-UI worker decomposition (/orchestrate).

- Skill: `shenxingy/codex-orchestrate-2` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add shenxingy/codex-orchestrate-2`
- Raw SKILL.md: https://api.skillmd.com/api/skills/shenxingy/codex-orchestrate-2/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: shenxingy (https://skillmd.com/u/shenxingy)
- Updated: 2026-09-21
- Page: https://skillmd.com/skills/shenxingy/codex-orchestrate-2

---


# Codex Orchestrate — you supervise a fleet of `codex exec` workers

Use when the user wants Codex to actually DO work at scale (implement N gaps, research N sources, review N dimensions) and you drive it — the manual equivalent of Codex's native `ultra`/multi_agent, but with no slot cap and with YOUR verification gate on every result.

Read `prompt.md` for the full loop, the exact dispatch command, and the verification checklist. The four rules that make it work (each learned from a real failure):

1. **`< /dev/null`** on every `codex exec` — it blocks on stdin EOF otherwise, even with the prompt as an arg, and hangs until timeout.
2. **One isolated git worktree per worker** off the latest main — parallel writers to shared build/test state race; core-engine changes merge **sequentially** so each builds on the last.
3. **Verify independently — never trust the worker's self-report.** Re-run the tests yourself (a worker reported "996 passed"; the real number was 995 + 1 real failure). For core-engine changes run the FULL suite, not just the new test.
4. **Dispatch via a tracked background call** (`run_in_background`), not `&` inside another command, or you lose the completion signal.