KingGyuSuh/ultracodex

ultracodex

English · 한국어 · 日本語 · 简体中文

Claude orchestrates. Codex cross-checks. Blend Codex headless (codex exec) nodes into Claude Code's own Workflow (ultracode) orchestration — so a genuinely different model family verifies, judges, and second-guesses at the points where same-model agreement is least trustworthy.

ultracodex is an open-source Claude Code plugin — a single skill, codex-workflow — that teaches Claude how to mix Codex headless nodes into a Workflow: Claude Code's agent() / pipeline() / parallel() orchestration tool, a.k.a. ultracode. The orchestration logic stays in Claude's JavaScript; only who does the work changes at chosen nodes. This is "Pattern A" — and it is additive, not a model-switcher. Both families run in one Workflow.

Workflow (Claude JS orchestration)
 ├─ find / generate / synthesize .. Claude agent()  ← broad, fast, cache-warm
 └─ verify / judge / 2nd-opinion ... codex node     ← GPT, independent failure modes

Requires ultracode mode. Because the skill authors and runs Workflows, the Workflow orchestration tool must be available — which in Claude Code means ultracode mode (enable it with /effort → ultracode; it turns on dynamic workflow orchestration). Without it the plugin still loads and the skill is found, but there is no Workflow tool to drive. See Prerequisites.

Why it exists — the philosophy

More verifiers from the same model tend to share that model's blind spots: they re-confirm the same correlated false positives. A different model family fails on different inputs, so it can refute what same-model reviewers would rubber-stamp. That is the whole idea:

• Diversity, not replacement. The win comes from disagreement between independent models, so the goal is to keep both families in one run — never to swa