nyldn/fable5-optimizer

Fable 5 Optimizer is a Claude Code add-on that helps decide which of two AI models should handle a given piece of work. It is built for people using Claude Code alongside Codex, which runs on GPT-5.5, and it installs as either an on-demand skill or an always-loaded policy file so Claude Code follows the same routing rule every time. The idea behind the project is that Fable 5, the model, should be the one making judgment calls such as choosing an architecture, deciding on a risky change, or handling anything touching an API contract, security surface, or user-facing interface. Codex is cheaper to run and better suited to grinding through repetitive, well-defined work like a migration or a review pass, but it should not be the one setting direction. This package gives Claude Code a routing table that spells out exactly where that line sits, plus a set of signals that always bump a task back to Fable 5 regardless of how mechanical it looks on the surface. Beyond the routing table, it includes a preparedness gate, which tells Codex to gather relevant context into a packet first for complex tasks, so that when Fable 5 steps in it can start reasoning immediately instead of rebuilding context from scratch. It also ships ready to use codex commands for review, bounded implementation, and browser or computer-use verification, all reporting results in the same consistent format, along with guardrails such as checkable acceptance criteria before delegating work and rules for when an autonomous run should stop and wait for a human. You can install it as a per-user skill that loads automatically whenever a task looks like it involves Fable 5 or Codex routing, as a project-scoped skill, or as a permanent block inside a project's CLAUDE.md file. Installation is done with a single curl command piped to bash, or by cloning the repository and running its install script. The routing guidance itself still works even without the Codex CLI installed, since Claude Code will simply note that Codex is unavailable and continue using its own tools. The project is released under the MIT license.