[Desktop] Initial parent model gates custom subagent spawning (Luna/5.5 work, Sol/Terra fail)

Summary

In Codex Desktop, the initial parent model selected when a new conversation is created appears to determine whether a named custom subagent can be spawned successfully.

With the same local configuration and the same Ornith custom agent:

• conversations initially created with gpt-5.6-luna: Ornith works
• conversations initially created with gpt-5.5: Ornith works
• conversations initially created with gpt-5.6-sol: Ornith cannot spawn Ornith reliably
• conversations initially created with gpt-5.6-terra: Ornith cannot spawn Ornith reliably

Reproduction

1. Configure a custom agent under ~/.codex/agents/Ornith.toml.
2. Configure a local OpenAI-compatible provider with wire_api = "responses".
3. Create a new Codex Desktop conversation with gpt-5.6-sol or gpt-5.6-terra as the initial parent model.
4. Ask Codex to spawn the named Ornith agent.
5. Repeat from a new conversation initially created with gpt-5.6-luna or gpt-5.5; the same agent works.

The successful native call shape is:

await tools.multi_agent_v1__spawn_agent({
  agent_type: "Ornith",
  message: "Run a minimal read-only test. Do not modify files."
});

Tool-surface observation

The successful sessions expose and use the native multi_agent_v1__spawn_agent surface with an agent_type: "Ornith" parameter.

The affected Desktop sessions use the tool-backed collaboration.spawn_agent surface, which has a different wrapper/schema.

Additional observations

• Switching the parent model after the conversation has started does not rebuild the multi-agent capability snapshot. A Luna-started conversation continues to spawn Ornith successfully after switching models.
• Direct local Codex CLI execution with explicit LocalLLM and the Ornith model works.
• This is related to the Desktop tool-backed custom-agent issue reported in #32291 and is now closed as a duplicate of #34301.

Expected behavior

Custom subagent availability should not depend on the initial parent model. The same custom agent should be selectable consistently across parent models.

Environment:

• Codex Desktop on Windows
• codex-cli 0.146.0-alpha.3.1
• local OpenAI-compatible llama.cpp server
• custom agent provider: LocalLLM