하나의 에이전트는 다른 유형입니다. delegate-router는 에이전트 기술입니다 (SKILL.md, 에이전트 Skills format) Claude Code 및 OpenAI Codex CLI는 LLM 라우터로 작동 코딩 작업. 그것은 손 크게, 이미 decided 일 (boilerplate, 비계, 시험 stubs, 많은 파일, 읽기 미디어 및 긴 문서를 통해 기계 편집 이미 지불: ** Codex** (GPT 모델), Gemini를 통해 Antigravity agy CLI, 또는 a **Claude Opus ** 시약. 호스트는 디자인과 검증을 유지하고, 자신의 저장 할당량 및 컨텍스트.

• **작업으로, 습관이 아닙니다. ** ROUTING.md maps 각 종류의 지도 모델의 주문 목록에 작업, 독립적 인 벤치 마크 및 작업 당 비용에서 선택. 그것은 당신이 가지고 있지 않는 모든 CLI를 건너 뛰고, 주인의 자신의 모델로 다시 경로를 결코 (excepts Opus 하위 시약 Opus 호스트, 이는 메인 컨텍스트를 깨끗하게 유지), 모델을 선호 구독의 또 다른 부분에, 그래서 당신이 지불하는 계획을 통해 작업을 확산.
• 모든 handoff 전에 판단. 4 개의 게이트 (bulk, 결정, verifiable, 포함) 및 undelegate list (디버깅, 아키텍처, 보안 코드, ~40 라인 이하).
• ** 호스트는 항상 확인합니다.

delegate-router: multi-model task delegation for Claude Code and Codex CLI

One agent thinks, another types. delegate-router is an agent skill (SKILL.md, the Agent Skills format) for Claude Code and OpenAI Codex CLI that works as an LLM router for coding work. It hands bulky, already-decided work (boilerplate, scaffolding, test stubs, mechanical edits across many files, reading media and long documents) to another model you already pay for: Codex (GPT models), Gemini through the Antigravity agy CLI, or a Claude Opus subagent. The host keeps the design and the verification, and saves its own quota and context.

• Routing by task, not by habit. ROUTING.md maps each kind of task to an ordered list of models, chosen from independent benchmarks and cost per task. It skips any CLI you don’t have, never routes back to the host’s own model (except an Opus subagent under an Opus host, which keeps the main context clean), and prefers a model on another of your subscriptions, so the work spreads across the plans you pay for.
• A judge before every handoff. Four gates (bulk, decided, verifiable, contained) and a never-delegate list (debugging, architecture, security code, anything under ~40 lines).
• The host always verifies. A delegate’s “done” isn’t evidence: the host runs the check and reads the diff.
• Routes that keep up on their own. Every 12 h, model_watch.py notices new models in your CLIs and the vendors’ release notes, re-checks the routes against a fresh benchmark snapshot, and tells you when a route should change. A tracker rescaling its scores moves nothing; a model that really beats the holder does.

New in v1.4: work goes to a model on another of your subscriptions first, so your main agent’s quota isn’t spent twice (how). Each route is judged on its own benchmark, and a route switches only on a cost per task measured on our bench (how). See the changelog.

Interactive diagram: docs/workflow.html (open it locally).

Requirements

• macOS or Linux, git, python3 3.9+.
• **Your main agent is Claude Code (claude) or