Codex Model Routing

Codex Model Routing

Principle

Always choose the cheapest GPT 5.6 route that can reliably do the delegated job, and make the model and effort explicit. The main thread keeps judgment and synthesis; subagents return compact evidence, patches, or verdicts.

GPT 5.6 reasoning efforts do not map one-to-one to GPT 5.5. Start lower than the similarly named GPT 5.5 effort: gpt-5.6-sol at low is the default starting point for a Codex conversation or low-risk main-thread judgment. Escalate only when the task provides evidence that it needs more reasoning.

Use only the GPT 5.6 family exposed by Codex CLI 0.144 or newer: gpt-5.6-luna, gpt-5.6-terra, and gpt-5.6-sol. If these ids are unavailable, stop and ask for the runtime to be upgraded rather than silently routing to an older family.

Routing table

Start with the cheapest reliable route. For primary Codex conversations, start at Sol low; for delegated workers, start with the applicable Luna or Terra route. Escalate only when uncertainty, failed validation, conflicting evidence, or task risk justifies it. Avoid max effort, Terra xhigh, and broad Sol fan-out because their quality gain usually does not justify the usage.

Prefer one strong agent over several weaker agents repeating the same work. Launch the critical-path child first, parallelize only independent units whose combined value justifies the usage, and normally cap concurrency at 2 to 3 children. Pipeline dependent checks as soon as prerequisites finish. Async reduces elapsed time, not token or quota usage; use blocked time for independent work and call wait only when the next result is required.

Spawn policy

Every Codex delegation plan should state:

• model
• reasoning_effort
• role
• exact paths
• what to inspect first
• user constraints
• expected return shape

Do not omit the model unless the platform forbids model overrides. If the platform forbids overrides, say that in the plan and keep the task bounded.

Explore defaults

For Codex Explore-style work, use Luna low only for bounded mechanical search:

• repo or file reconnaissance
• finding examples of a pattern
• reading logs or terminal output
• collecting candidate files and line references

Escalate to Luna high when exploration requires synthesis, Luna xhigh for quality-sensitive normal analysis, Terra high for larger multi-file work, Sol medium for ambiguous substantive work, and Sol high only when the result requires architecture or difficult judgment.

Prompt shape

Use self-contained prompts. Include enough context that the subagent does not need the main thread.

{
  "model": "gpt-5.6-luna",
  "reasoning_effort": "low",
  "role": "Explore agent",
  "task": "Find every place this project defines subagent model routing. Return paths, approximate lines, and a compact summary.",
  "paths": [
    "/absolute/path/to/repo"
  ],
  "constraints": [
    "Locate by content; line numbers are approximate.",
    "Return evidence only; do not edit files."
  ]
}

For maker/checker/fixer loops:

• Maker: Luna high for bounded work, Terra high for larger multi-file work, Sol medium for ambiguous substantive work.
• Checker: Luna xhigh for normal independent review, Sol high for hard final judgment.
• Fixer: Luna high for exact findings, Terra high when fixes require broad cross-file reasoning.
• Verifier: shell checks first, Luna low only when model judgment is needed.

Completion check

Before dispatching, verify that every worker has:

• explicit model and reasoning effort
• a bounded task
• exact inputs
• a return format that is smaller than the context it consumes
• no need to ask clarifying questions