Regression: luna cannot be specified in spawn_agent tool model anymore, e.g. for Explorer

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #32988
• #33126

Powered by Codex Action

Potential duplicates detected. Please review them and close your issue if it is a duplicate. [[Release blocker] Backport #32749: stable Codex breaks model-routed subagent plugins #32988](https://github.com/openai/codex/issues/32988) Subagents being launched as same model as parent agent with no change option #33126 _Powered by Codex Action_

This issue is not a duplicate of them. It is specifically about luna. Spawning sub-agents when specifying other models works for me.

Can confirm this is happening to me as well. It seems currenlty the Sol and Terra Models are affected

This is not Windows-specific.

The rejecting behavior was introduced by commit 92938d8 / [PR #32751](https://github.com/openai/codex/pull/32751), “Restrict spawned-agent models to the active backend.”

That commit added this V2 compatibility gate:

multi_agent_version != MultiAgentVersion::V2
    || model.multi_agent_version == Some(multi_agent_version)

It is applied both when advertising spawn_agent model overrides and when validating the requested model. The current catalog marks Sol and Terra as v2, but Luna as v1, so a V2 parent rejects Luna before reaching any platform-specific code.

This is not limited to Luna. With this gate, a V2 parent also rejects gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.3-codex-spark, and gpt-5.2 as explicit spawn_agent overrides. In practice, Sol and Terra parents are restricted to spawning Sol or Terra.

Independent confirmation from a Windows Codex Desktop session.

Environment:

• Host CLI: codex-cli 0.144.6
• codex debug models lists gpt-5.6-luna with visibility=list, supported_in_api=true, and supported efforts low, medium, high, xhigh, max.
• multi_agent=true and multi_agent_v2=true.
• Local config contains no model allowlist; it only sets a Terra root default and agent thread/depth limits.

Observed behavior:

• The native spawn_agent parameter named model is a per-child model override, not a declaration of which models support subagents.
• In this V2 session the tool exposes the override, but its allowed values are only gpt-5.6-sol and gpt-5.6-terra; gpt-5.6-luna is excluded before a spawn can be requested.
• Thus Luna is locally available and eligible in the model catalog, but unavailable specifically as a child-model override.

This independently matches the cross-backend filter introduced by #32751 / commit 92938d8. Please consider allowing cross-backend child routing (or dispatching the child through its own compatible backend) instead of filtering Luna out of the schema. If cross-backend routing is intentionally unsupported, the UI/tool documentation should state that explicitly; silent schema omission makes model-specialized subagent workflows impossible.

Would be great to get this patched. Being forced to dispatch subagents with Terra instead of Luna is much slower and usage-hungrier.

Indeed.

Workaround I'm using:

I have a custom model catalog (with custom dev prompts) that explicitly sets all models to support multi-agent-v2.
Since 0.145.0 got enough feature parity from multi-agent-v2 with multi-agent v1, this was the minimal fix required that replaced my old "use v1 everywhere" approach.

1. Dump the model cache with codex debug models > ~/.codex/model-catalog-v2.json
2. Update each model to multi_agent_version = "v2"
3. Add model_catalog_json = "/home/<your-name>/.codex/model-catalog-v2.json" (or windows equivalent) to the top of ~/.codex/config.toml

use tmux and communicate between panes simply. much easier

use tmux and communicate between panes simply. much easier

At this point you can use other harness. It's not a solution and it does not help