Codex Desktop with Bedrock uses unsupported gpt-5.4-mini for helper turns

Summary

Codex Desktop is configured to use Amazon Bedrock with a supported model ID, but some internal Desktop helper/orchestration turns are sent to Bedrock using the bare model ID gpt-5.4-mini.

Amazon Bedrock rejects that model with a 404. In the UI, the user-visible chat shows the thinking indicator briefly, then it disappears and no assistant response is added. The turn appears to complete silently instead of surfacing an error.

Environment

• GitHub user filing: scoropeza
• Codex Desktop version: 26.601.21317
• App version seen in logs: 0.136.0-alpha.2
• Platform: macOS
• Provider: amazon-bedrock
• Configured model: openai.gpt-5.4
• Region: us-east-2

Relevant ~/.codex/config.toml excerpt:

model = "openai.gpt-5.4"
model_provider = "amazon-bedrock"
model_reasoning_effort = "medium"

[model_providers.amazon-bedrock.aws]
region = "us-east-2"

Expected behavior

When model_provider = "amazon-bedrock" is configured, all Desktop model calls should use Bedrock-compatible model IDs or inherit the configured model.

For example:

openai.gpt-5.4
openai.gpt-5.5

Internal helper turns such as title generation, skill orchestration, or other Desktop-side helper tasks should not use bare OpenAI-hosted model IDs that Bedrock does not support.

Actual behavior

Logs show internal helper turns using:

model=gpt-5.4-mini

Bedrock rejects the request:

unexpected status 404 Not Found: The model 'gpt-5.4-mini' does not exist,
url: https://bedrock-mantle.us-east-2.api.aws/openai/v1/responses

The user-visible result is a silent failure: the chat thinks briefly, then stops with no assistant message.

Reproduction pattern

This was observed when starting fresh Codex Desktop chats where Desktop appeared to run internal helper/orchestration work.

One reproducible trigger involved a local custom skill installed under ~/.codex/skills/..., but the issue appears to be the Desktop helper model selection rather than the skill content itself.

The helper prompt in logs included title-generation-style text such as:

You are a helpful assistant. You will be presented with a user prompt,
and your job is to provide a short title for a task...

That helper turn was submitted with:

model=gpt-5.4-mini

even though the configured model was:

openai.gpt-5.4

Relevant log excerpts

turn{... model=gpt-5.4-mini ...}:run_turn:run_sampling_request{... model=gpt-5.4-mini ...}

stream_request:model_client.stream_responses_api{
  model=gpt-5.4-mini
  wire_api=responses
  transport="responses_http"
  http.method="POST"
  api.path="responses"
}

Request completed method=POST
url=https://bedrock-mantle.us-east-2.api.aws/openai/v1/responses
status=404 Not Found

Turn error: unexpected status 404 Not Found:
The model 'gpt-5.4-mini' does not exist

Impact

This breaks some Codex Desktop flows for Bedrock users even when the main configured model works. The main assistant can successfully run on openai.gpt-5.4, but Desktop helper/orchestration turns that use gpt-5.4-mini fail.

Because the error is not surfaced in the chat, it looks like Codex simply stopped responding.

Suggested fix

When model_provider = "amazon-bedrock" is configured, Desktop helper/orchestration turns should either:

• inherit the configured model, or
• map internal mini/default helper models to a Bedrock-supported model such as openai.gpt-5.4.

It would also help if helper-turn failures surfaced a visible error instead of completing the chat turn with no assistant output.

Note

I realize the Desktop app UI may not be fully open source, but the failure appears to involve the local Codex/app-server model selection path used by Desktop helper turns. Please redirect if there is a better tracker.