codex-model-router

name	codex-model-router
description	Use before spawning Codex subagents or delegated Codex workflows to select the supported OpenAI model and reasoning effort for the work.
disable-model-invocation	true
user-invocable	false

Model Router

This policy governs model and reasoning selection for delegated Codex work. It does not change the model for a singular session that performs the work itself.

Picking the right models for workflows and subagents

• Honor a model or effort explicitly requested by the user. Otherwise use the lanes below.
• Select only models exposed by the active orchestration surface. When a lane's preferred model is unavailable, use gpt-5.6-sol at the lane's effort. If that model is also unavailable, use the strongest exposed coding model at the same effort instead of bypassing the surface to force a model.
• If a delegated agent may delegate again, tell it to use $codex-model-router before it selects or spawns its own agents.
• Do not select Terra unless the user explicitly overrides this repository policy in the current request.
• Do not encode pro as a reasoning effort. Use it only through a separate supported model or mode control exposed by the active product surface.

Model and Effort Lanes

Lane	Work	Model	Effort
Mechanical	Preprocessing or shell-heavy inventory	gpt-5.6-luna	low
Exploration	Bounded, parallel, read-only evidence gathering	gpt-5.6-luna	medium
Exploration summary	Compress completed evidence when Spark is supported	gpt-5.3-codex-spark	high
Exploration summary fallback	Compress completed evidence when Spark is unavailable	gpt-5.6-sol	high
Clear implementation	Clear-spec implementation, fixes, or review slices	gpt-5.6-luna	high
Judgment	Synthesis requiring judgment or independent review	gpt-5.6-sol	low
Ambiguous implementation	Multi-subsystem implementation with unresolved design choices	gpt-5.6-sol	medium
Security and architecture	Security analysis or architectural decisions	gpt-5.6-sol	high
Hardest	Only genuinely hardest cases	gpt-5.6-sol	max

For exploration, gather evidence before starting a separate summary pass. Spark summarizes completed evidence; it does not make architecture, security, or implementation decisions. When the result requires those decisions, use the matching judgment lane.