spawn_agent rejects gpt-5.6-luna while parent runtime and another project can use Luna

Summary

spawn_agent appears to have a model registry mismatch: the parent Codex runtime can access gpt-5.6-luna, but the child-agent dispatch path rejects the same model.

Environment

• Codex: 0.147.0-alpha.6.6
• Platform: Windows 11
• Client: Codex Desktop

Evidence

Direct Luna probe in the same environment:

DIRECT_LUNA_PROBE: PASS
LUNA_MODEL: gpt-5.6-luna

However, custom agent dispatch fails:

spawn_agent only exposes:
gpt-5.6-sol
gpt-5.6-terra

Error:
Unknown model gpt-5.6-luna

A/B reproduction

A previously successful project/session using the same Codex version and Desktop source:

agent_type: luna_verifier
model: gpt-5.6-luna
fork_turns: none
model_reasoning_effort: medium
sandbox_mode: read-only

Result:
child agent starts successfully and returns verification output

Current project:

agent_type: luna_verifier
model: gpt-5.6-luna
fork_turns: none
model_reasoning_effort: medium
sandbox_mode: read-only

Result:
Unknown model gpt-5.6-luna
Available:
- gpt-5.6-sol
- gpt-5.6-terra

The custom agent TOML schema is identical between the working and failing cases.

Expected behavior

The model availability exposed to spawn_agent should be consistent with the parent runtime model catalog, or there should be a clear capability API indicating child-agent model availability.

Actual behavior

Parent runtime and child dispatch runtime expose different model sets:

Parent runtime:
  gpt-5.6-luna available

Child dispatch runtime:
  only gpt-5.6-sol / gpt-5.6-terra

No project configuration changes were used as a workaround.

Start by tracing the spawn_agent child-dispatch path and comparing its exposed model catalog with the parent runtime catalog, using the supplied gpt-5.6-luna reproduction and custom-agent TOML settings. Done means child-agent availability matches the parent catalog or exposes a clear capability result, with the reported mismatch covered by a regression check.