Getting flm server to play ball with Codex?

I have one of these AMD-based NPUs and using Windows, and I wanted to see how well this NPU-based implementation works. This is probably a skill issue, but I am struggling to get flm to work with Codex.

The first weird thing is the Codex ignored my set environment variables and it defaulted to OpenAI servers. OK, I can live with that. I have set these as:

Get-ChildItem Env:

OPENAI_API_BASE                http://localhost:52625
OPENAI_API_KEY                 flm

So I changed the config file and created my custom profile. Here is the reelevant part of .codex\config.toml

model = "qwen3.6-moe:35b-a3b"
model_provider = "flm"

[model_providers.flm]
name = "FastFlowLM"
base_url = "http://localhost:52625/v1"
env_key = "OPENAI_API_KEY"
wire_api = "responses"

In one terminal window, I run flm serve qwen3.6-moe:35b-a3b --ctx-len 262144. then, in another terminal window, I run: codex --profile flm

Codex loads, I send in a prompt, the model loads, and I can see the output of the server.

But, I don't actually get a response in Codex, but a warning and an error message:

⚠ Model metadata for `qwen3.6-moe:35b-a3b` not found. Defaulting to fallback metadata; this can degrade performance and cause issues.

■ unexpected status 404 Not Found: {"error":"Not Found"}, url: http://localhost:52625/v1/responses

So, can anyone tell me what am I doing wrong?

(edit: codex CLI no longer supports 'chat' and 'completions', I have to use 'responses', and I see in the documentation that 'responses' endpoint is not supported - so probably that's it)

Start by reproducing the Windows setup with flm serve qwen3.6-moe:35b-a3b --ctx-len 262144 and the supplied .codex\config.toml. Check the documented API endpoints against Codex's request to /v1/responses; done means the supported integration path is identified and either works or the incompatibility is clearly documented.