ethanplusai/codex-on-crack

Coordinate your models from Codex. Plan, delegate, use host tools, and review work across workspaces. Formerly Astra Flash Orchestrator.

What it solves

It eliminates the need to manually switch between different AI models or repeatedly explain project context and CLI settings in every prompt. It allows users to orchestrate multiple models to work together on a single assignment, leveraging the specific strengths of different models for planning, implementation, and review.

How it works

The system uses a "lead" Codex model to coordinate work and manage tools. This lead model can delegate tasks to "native workers" (configured via Codex Router) or a Claude Code client (specifically targeting Opus 5.5). The workflow typically follows a sequence: a lead model plans the work, a worker model implements it, and the lead model reviews the result. The system includes three core skills—crack-plan, crack-setup, and crack—to handle planning, worker configuration, and execution.

Who it’s for

Developers and power users of Codex who want to use a combination of high-reasoning models for planning and faster, more cost-effective models for implementation, or those who want to integrate Claude Code's execution capabilities into their Codex workflow.

Highlights

  • Multi-model orchestration: Coordinate different models (e.g., Opus 5.5 for planning and DeepSeek V4.1 Flash for implementation) within one conversation.
  • Integrated Review Process: A lead model validates the work produced by worker models before finalizing.
  • Visual Monitoring: An optional MCP App panel provides a real-time view of agent handoffs, tool activity, and per-run timing.
  • Flexible Arrangements: Supports various configurations, from solo model work to complex lead-worker-reviewer chains.

Related

  • Project
    liquidos-ai/AutoAgents

    A high-performance, modular multi-agent framework written in Rust that supports structured tool calling, local and cloud LLM backends, and secure agent execution.

  • Project
    dapr/dapr-agents

    Dapr Agents is a Python SDK built on the Dapr runtime that lets developers create scalable, fault‑tolerant AI agents. It provides durable workflows, actor‑based execution, built‑in data source bindings, multi‑agent communication, and automatic tool discovery, all with Kubernetes‑native deployment and strong security.

  • Project
    octos-org/octos

    Octos is a Rust‑written, embeddable AI‑agent kernel that provides execution loops, context & memory management, tool sandboxing, workflow orchestration, and multi‑agent coordination. Applications embed the crates (or use language bindings) and interact via OUP—a JSON‑RPC 2.0 protocol over WebSocket or stdio—allowing any client (terminal, web UI, or another agent) to start sessions, assign turns, monitor tool tasks, handle approvals, and persist/recover state.

  • Project
    sno-ai/sno-station

    Sno Station is a local-first orchestration layer that connects multiple AI coding agents into a single, memory-sharing squad capable of handovers, peer reviews, and automated self-improvement.