Sol Terra Luna Auto Router

Prioritize the common GPT Codex workflow: the user asks for repository work and continues directly in a visible Codex task running the selected model and effort. Apply the same routing policy in GPT Work and Codex CLI. Optimize the balance between result quality and total input, reasoning, rework, verification, and handoff tokens.

Preserve the authority boundary

• Treat invocation of this skill, explicit or implicit, as an instruction to classify the task and use the surface's native model-routing mechanism. The skill explicitly authorizes bounded model routing; it does not authorize unrelated side effects.
• Route only within the GPT-5.6 family: gpt-5.6-luna, gpt-5.6-terra, and gpt-5.6-sol. Never route to, downgrade to, or use a fallback from GPT-5.5 or another non-5.6 model. When the selected 5.6 profile is unavailable, choose another appropriate 5.6 profile or report the unavailable gate honestly.
• For a nontrivial Codex request, current model and effort must be positively known to exactly match the selected route before work may stay in place. Unknown current model or effort is a mismatch. When fork_thread, send_message_to_thread, and navigation are available, a mismatch requires exactly one visible same-directory routing fork.
• Never claim that the current host task changed models or reasoning effort. A profile label in prose is not execution evidence. Report a routed child only after the model-aware send succeeds; when settings readback is available, verify that readback before claiming the switch executed.
• If no model-aware adapter is available, continue in the current task using the selected operating profile and report the model and effort as a recommendation, not an executed switch.
• Do not infer authority for deployment, destructive commands, external messages, purchases, or other consequential actions from model-routing authority.

Establish context