Native subagent ignores explicit model_provider override while model override works

What version of Codex CLI is running?

codex-cli 0.149.1

What subscription do you have?

plus

Which model were you using?

Parent: gpt-5.6-luna (xhigh) Subagent: MiniMax-M3

What platform is your computer?

Linux 6.17.0-29-generic x86_64 x86_64

What terminal emulator and version are you using (if applicable)?

GNOME Terminal 3.52.0 using VTE 0.76.0 +BIDI +GNUTLS +ICU +SYSTEMD

Codex doctor report

What issue are you seeing?

Native subagent roles appear to ignore an explicit model_provider override while correctly applying the model and model_reasoning_effort overrides.

My parent session uses ChatGPT/OpenAI:

• Parent model: gpt-5.6-luna
• Parent provider: OpenAI / ChatGPT account

The subagent role explicitly specifies:

model = "MiniMax-M3"
model_provider = "minimax"
model_reasoning_effort = "high"

### What steps can reproduce the bug?

1. Register a custom Responses API provider in ~/.codex/config.toml:

[features]
multi_agent = true

[agents]
enabled = true
default_subagent_model = "gpt-5.6-luna"
default_subagent_reasoning_effort = "xhigh"

[model_providers.minimax]
name = "MiniMax"
base_url = "https://api.minimaxi.com/v1"
experimental_bearer_token = "<redacted>"
wire_api = "responses"

Do NOT globally set:

model_provider = "minimax"

The parent session should remain authenticated through ChatGPT/OpenAI.

2. Create a native subagent role at:

~/.codex/agents/chain-test.toml

with the following configuration:

name = "chain_test"

description = """
Diagnostic subagent for testing cross-provider delegation.
"""

model = "MiniMax-M3"
model_provider = "minimax"
model_reasoning_effort = "high"

sandbox_mode = "read-only"

developer_instructions = """
Inspect the current repository.
Actually execute:
- pwd
- git status
- ls

Do not modify any files.
"""

3. Start the normal ChatGPT-authenticated Codex CLI:

codex

4. Ask the parent agent:

You must use the chain_test subagent for this task.
Do not perform the task yourself.

Inspect the current repository and actually run:
- pwd
- git status
- ls

Wait for chain_test to finish and return its result.

5. Codex successfully resolves the role and spawns the child:

Spawned <thread-id> (MiniMax-M3 high)

However, the child fails before running the requested commands with:

The 'MiniMax-M3' model is not supported when using Codex with a ChatGPT account.

6. As a control test, use the SAME ~/.codex configuration and SAME ChatGPT-authenticated environment, but override the provider at the session level:

codex exec \
  -C "$PWD" \
  --sandbox danger-full-access \
  -c 'model_provider="minimax"' \
  -c 'model="MiniMax-M3"' \
  -c 'model_reasoning_effort="high"' \
  "Actually execute pwd and return the result."

This works correctly.

Codex reports:

model: MiniMax-M3
provider: minimax

The MiniMax model then successfully performs the shell tool call.

Therefore, the custom provider itself works correctly from the same ChatGPT-authenticated CODEX_HOME.

The failure appears specific to applying the explicit model_provider override from a native subagent role: the model override is applied, but the provider appears to remain inherited from the OpenAI/ChatGPT parent.

### What is the expected behavior?

When a native subagent role explicitly specifies both:

model = "MiniMax-M3"
model_provider = "minimax"

the spawned child should use both overrides.

Expected configuration:

Parent:
  model: gpt-5.6-luna
  provider: openai / ChatGPT

Child:
  model: MiniMax-M3
  provider: minimax

The parent provider should only be inherited when the subagent role does not explicitly specify its own model_provider.

This should allow native subagents to use a different custom provider from the parent session.

### Additional information

I can also reproduce the same behavior in the ChatGPT/Codex desktop app.

The desktop app is using its bundled Codex runtime:

/usr/lib/chatgpt/resources/codex --version
codex-cli 0.150.0-alpha.8

The desktop native subagent also applies the MiniMax-M3 model override, but then fails with the same ChatGPT-account error instead of using the explicitly configured minimax provider.

The custom provider itself is confirmed to work.

Using the same normal ChatGPT-authenticated ~/.codex environment, a session-level override works successfully:

codex exec \
  -C "$PWD" \
  --sandbox danger-full-access \
  -c 'model_provider="minimax"' \
  -c 'model="MiniMax-M3"' \
  -c 'model_reasoning_effort="high"' \
  "Actually execute pwd and return the result."

Codex reports:

model: MiniMax-M3
provider: minimax

and MiniMax-M3 successfully performs the shell tool call.

So this does not appear to be a general custom-provider, authentication, or MiniMax compatibility issue.

The behavior appears specific to native subagent configuration:

- Subagent model override works.
- Subagent reasoning-effort override works.
- Explicit subagent model_provider override does not appear to take effect.

A practical use case is keeping an OpenAI/ChatGPT model as the parent/orchestrator while delegating implementation, testing, review, or exploration to specialized or lower-cost models on other providers.

For example:

OpenAI / ChatGPT parent
        |
        +-- MiniMax backend subagent
        +-- MiniMax frontend subagent
        +-- MiniMax QA subagent
        +-- MiniMax reviewer subagent

Start with ~/.codex/config.toml and ~/.codex/agents/chain-test.toml, then reproduce the native subagent launch using the documented ChatGPT parent and minimax provider configuration. Compare the child’s effective model and provider with the working session-level -c overrides. Done means an explicitly configured subagent uses minimax while the parent remains on ChatGPT/OpenAI and the child executes the requested commands.