Cannot overwrite subagent's model when spawning via LLM

[breezewish]

What version of the Codex App are you using (From “About Codex” dialog)?

Version 26.506.31421 (2620)

What subscription do you have?

Pro 20x

What platform is your computer?

Linux 6.17.0-1010-aws aarch64 aarch64 (Codex App + Connection)

What issue are you seeing?

(issue is summarized via codex)

When calling spawn_agent with both agent_type and an explicit model, the spawned subagent run with the parent/default configured model instead of the requested model.

What steps can reproduce the bug?

Given a global config like:

model = "gpt-5.4"

[agents.potter_worker]
description = "Worker role"
config_file = "./agents/potter_worker.toml"

And ~/.codex/agents/potter_worker.toml contains role instructions but no model field.

Then spawn an agent with:

{
  "agent_type": "potter_worker",
  "message": "Work according to this handoff file: ...",
  "model": "gpt-5.2",
  "reasoning_effort": "xhigh"
}

Expected behavior

The spawned subagent should run with:

model: gpt-5.2
reasoning effort: xhigh

Actual behavior

The spawned subagent runs with the configured default model, for example:

model: gpt-5.4
reasoning effort: xhigh

The parent rollout records the spawn_agent function call with "model":"gpt-5.2", but the child rollout turn_context records "model":"gpt-5.4".

Root cause

The v1 spawn_agent handler applies requested model overrides before applying the role:

apply_requested_spawn_agent_model_overrides(...);
apply_role_to_config(&mut config, role_name).await?;

However, apply_role_to_config rebuilds the config using the config layer stack:

Config::load_config_with_layer_stack(
    ...,
    reload_overrides(config, preserve_current_provider),
    ...
)

The reload_overrides helper preserves some runtime fields such as cwd and model provider, but it does not preserve config.model or config.model_reasoning_effort. During config loading, the final model is resolved from:

model override -> active profile model -> config.toml model

Because the requested spawn model is not passed as a config override, it is lost unless the role TOML itself explicitly sets model.

Impact

This makes the model argument on spawn_agent unreliable for role-based agents. Callers can believe they started a subagent with one model while the actual thread runs with another. This is especially confusing for orchestrators that intentionally alternate models across rounds.

It also makes the spawn result misleading: the request contains the desired model, but the actual child thread must be inspected through turn_context to know what really ran.

Suggested fix

When applying a role that does not explicitly set model or model_reasoning_effort, preserve the already-applied requested spawn overrides through the role config reload.

Concretely, reload_overrides or the role application path should carry forward the current in-memory config.model and config.model_reasoning_effort unless the role layer explicitly owns those fields.

Existing behavior where a role config explicitly declares model should continue to override the requested model.