codejunkie99/fable-orchestrator

What it solves

Fable orchestrator is a routing skill for Codex that separates the planning and adjudication phase of software development from the actual code implementation. This prevents the planning model from becoming bogged down in implementation details and ensures a clear separation of concerns between the high-level architect and the implementation agents.

How it works

Fable uses a specific model hierarchy: Claude Fable 5.1 acts as the planner and adjudicator, remaining outside the implementation graph. The actual coding work is delegated to OpenCode Go agents, specifically GPT-5.6 Luna for standard implementation and DeepSeek V4 Flash for high-throughput, iterative tasks. Codex serves as the runtime that validates the graph, manages parallel workers, and collects evidence for final adjudication by Fable.

Who it’s for

Developers using the Codex ecosystem who want a structured, multi-model orchestration layer to manage complex coding tasks through a bounded graph of implementation nodes.

Highlights

• Separation of Concerns: Planning and adjudication are handled by a separate model from the implementation agents.
• Multi-Model Routing: Specifically routes tasks to GPT-5.6 Luna or DeepSeek V4 Flash based on the task type.
• Local-First Design: Operates as a local routing skill without requiring its own proxy, dashboard, or credential store.
• Bounded Graph Execution: Codex validates and executes implementation nodes in parallel where possible.