What problem does it solve? Choosing the right GPT-5.6 model and reasoning effort for every Codex task is repetitive manual work, and picking wrong wastes either quality or latency. This Skill automates that decision with a fail-open, benefit-gated router that recommends a route once and only spawns a model-specific subagent when the benefit clearly exceeds startup overhead. ## Core Features & Use Cases - Automatic model routing: Classifies each task as mechanical, ordinary, or complex and recommends Luna, Terra, or Sol with low-through-max reasoning effort based on benchmark-calibrated lanes. - Benefit-gated subagents: Automatically creates or reuses a bounded leaf agent running the recommended model only when route benefit clears measured startup and aggregation cost; otherwise work stays local. - Low-overhead tool concurrency: Runs independent safe tool and process calls concurrently in the coordinator without creating child agents. - Project-scoped exit: Disable, restore, or check the Skill per project via managed entries in the project's .codex/config.toml. - Use Case: A developer asks Codex to refactor a module; the router classifies it as bounded complex work, recommends Sol/medium, and either executes locally or delegates to a Sol leaf agent, showing a one-line routing notice in the user's language. ## Quick Start Ask Codex to install the codex-auto-model-router Skill from its GitHub repository, restart Codex, then make any code change request and the router will recommend and apply a model route automatically.
