RLM subagents cannot use GPT-5.6 Luna from a GPT-5.6 Sol parent session

[cskwork]

Summary

In a chat whose active model is openai-codex/gpt-5.6-sol, RLM children inherit Sol. There does not appear to be an accessible GPT-5.6 Luna selector for spawning a child from that session, so the only apparent workaround is to switch the entire parent chat to GPT-5.6 Luna (max effort) with /model and then spawn children without an explicit selector.

This makes Luna subagents effectively unavailable from a Sol chat and can also lead to accidentally launching Sol children when Luna was intended.

Steps to reproduce

1. Start or use a parent chat running openai-codex/gpt-5.6-sol.
2. Attempt to spawn an RLM child intended to run GPT-5.6 Luna (max effort).
3. Observe that no usable explicit Luna model selector is available (or Luna cannot be selected for the child).
4. Spawn without an explicit selector.
5. Observe that the child inherits openai-codex/gpt-5.6-sol.

Actual behavior

The children run Sol. In my session, two children intended for Luna inherited Sol and had to be cancelled immediately.

Expected behavior

One of the following should work:

• GPT-5.6 Luna (including max effort) is exposed as a valid explicit model selector for rlm(...), regardless of the parent model; or
• the UI/API provides a documented way to choose a different child model without switching the parent chat.

At minimum, the selected/resolved child model should be clearly confirmed before work begins so unintended-model runs can be avoided.

Workaround

Switch the parent chat with /model to GPT-5.6 Luna (max effort), then spawn children without an explicit selector so they inherit Luna.

Environment / evidence

• Parent session reported: openai-codex/gpt-5.6-sol
• Affected feature: RLM child/subagent spawning and model inheritance
• Two unintended Sol children were cancelled immediately