alectivism

Cross Model Agent Delegation — AI skill for Claude Code

AI community

Deterministic delegation between Claude Code and the Codex CLI, in both directions.

How to install Cross Model Agent Delegation

This entry records only its repository, not the path inside it, so there is no exact command to give. Open alectivism/cross-model-agent-delegation and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Cross Model Agent Delegation does

Deterministic delegation between Claude Code and the Codex CLI, in both directions. Seven task classes per side, wrappers that own every flag, Codex models resolved from the live catalog (GPT-6 ready), model-pinned agent templates, guard hooks, and Teams/Enterprise org guidance.

Alternatives in AI

Skill Seekers Roadmap — Transform Skill Seekers into the easiest way to create Claude AI skills from any knowledge source - documentat 11.1k ★
Lumenx Model Onboarding — LumenX 模型接入与文档更新流程 - 厂商文档抓取、model_catalog 更新、运行时范围判断、端到端验证 1.2k ★
Simulink Agentic Toolkit — The Simulink Agentic Toolkit gives your AI agent both the tools and the expertise to work effectively with Sim 987 ★

README

Cross-Model Agent Delegation

Delegation between [Claude Code](https://code.claude.com) and the [OpenAI Codex CLI](https://developers.openai.com/codex/cli), in **both directions**, with a script owning every flag so the model classifies but never improvises.

This is the setup I run daily as Chief Agent Officer at the Marketing + Media Alliance, extracted and stripped of anything organization-specific. Every claim below was verified against a named CLI version on a stated date, and the failure modes in the "what this cost to get right" sections are ones that actually bit.

Claude Code → Codex. A Fable/Opus orchestrator keeps planning and synthesis; research, review, retrieval, and bulk edits go to Sonnet/Haiku subagents or to Codex workers on GPT-6 Sol (everyday work), GPT-6 Astra (review and the hardest calls), or GPT-6 Luna (mechanical work) at time of writing, resolved from Codex's live model catalog rather than pinned.
Codex → Claude Code. A Codex session sends the same shapes of work the other way, to Haiku/Sonnet/Opus workers.

Two reasons to run it either way:

It cuts expensive-model token burn. Ad-hoc subagents inherit the session model, so an Opus session burns Opus on grep work unless something pins it. Both wrappers pin it.
It spreads load across two subscriptions. Claude quota and ChatGPT quota are separate pools, so leaf work crossing the boundary makes both last longer. You also get the cross-family bonus: a GPT model reviewing Claude's work, or a Claude model reviewing GPT's, catches errors two same-family models agree on.

You don't invoke any of this by hand. You talk to whichever CLI you're in; the skill and rules make it classify and route on its own:

"Fix the flaky retry test" → the fix happens locally, then verification goes cross-family as a review-class call.
"What does our webhook layer actually do?" → an explore or ingest worker, with only the summary returning to your context.
"Get a second opinion