Cjbuilds/Codex-Orchestration

Codex Orchestration – A plug‑in for the Codex AI assistant

What it is – A Codex plug‑in that lets you attach other large‑language‑model (LLM) services (e.g., Anthropic Claude Fable 5, Claude Opus 5, OpenRouter models) to a Codex task and assign each model a specific role:

• Planner – drafts and iterates the task plan.
• Advisor – reviews the plan, points out gaps and finally approves it.
• Designer – creates a bounded visual/UX/IA hand‑off (optional).
• Executor – implements the approved plan (required).

Codex remains the top‑level orchestrator: it passes work between the roles, validates every result, and returns the final answer.

Why you might want it

• Use the strongest model for each phase (planning, review, design, implementation) instead of a single model.
• Get a more vetted plan before any code changes are made.
• Parallelise independent work steps, potentially cutting execution time by up to 2×.
• Reduce usage of premium‑model quotas because the heavy‑lifting is split across cheaper models.

Installation

# Add the plug‑in from the Codex marketplace
codex plugin marketplace add Cjbuilds/Codex-Orchestration
# Install the plug‑in itself
codex plugin add codex-orchestration@codex-orchestration

Python 3.11+ is required.

Quick start – all commands are entered as skill prompts inside a Codex chat, not as shell commands.

# Set up a full workflow (Planner + Advisor + Designer + Executor)
$codex-orchestration:codex-orchestration setup \
    planner: Claude Fable 5 High, \
    advisor: GPT-5.6 Sol High, \
    designer: GPT-5.6 Terra High, \
    executor: GPT-5.6 Luna Extra High

# Or a minimal workflow (only Advisor and Executor)
$codex-orchestration:codex-orchestration setup \
    advisor: Claude Opus 5 XHigh, \
    executor: GPT-5.6 Luna Extra High

After the setup prompt, start a normal Codex task – the saved workflow will be applied automatically.

Key concepts & limits

• Planner & Advisor may iterate up to eight review cycles; if a plan never receives PLAN_APPROVED, execution stops and the latest plan is shown.
• Designer only edits design artefacts handed to it by Codex; it never touches implementation code.
• Only one Claude subscription seat (Fable 5 or Opus 5) can be configured at a time, and it can only serve as Planner or Advisor.
• External models are added as roles via a secure, credential‑store‑backed flow; the plug‑in never asks you to paste API keys into chat.
• The plug‑in never changes the root model or the model picker – Codex stays the ultimate authority.

Managing the plug‑in

• … status – shows which roles are active and their readiness.
• … repair – fixes mismatched saved state without touching credentials.
• … disable – restores the previous routing configuration (does not delete custom roles).
• … --update – asks the plug‑in to update itself via Codex’s native plugin manager (available from v0.7.0).
• Uninstall: run the disable prompt, then remove the plug‑in with codex plugin remove ….

Development & quality

• Uses a standard Python development stack (requirements‑dev.txt, ruff, unittest).
• Includes a production‑readiness audit, security policy, and release process.

License – MIT.

---

All details above are taken directly from the repository’s README; no additional features have been inferred.