Parameters setting like temperature do not work with responses api

It seems setting stuff like temperature does not work properly. When set it gets ignored and passed to localai with other settings instead.

{{
  "input": [
    {
      "role": "user",
      "content": "This is a test prompt."
    }
  ],
  "model": "MyAgent",
  "temperature": 1.0,
  "frequency_penalty": 0.0,
  "presence_penalty": 0.0,
  "repeat_penalty": 1.0,
  "top_p": 0.95,
  "top_k": 64,
  "microstat": 0,
  "microstat_eta": 0.2,
  "microstat_tau": 4.0
}}

When viewed in localai with debug flag it shows me this:

9:05AM DBG GRPC(gemma-3-12b-it-127.0.0.1:52164): stdout   "stream": false,
9:05AM DBG GRPC(gemma-3-12b-it-127.0.0.1:52164): stdout   "cache_prompt": false,
9:05AM DBG GRPC(gemma-3-12b-it-127.0.0.1:52164): stdout   "n_predict": -1,
9:05AM DBG GRPC(gemma-3-12b-it-127.0.0.1:52164): stdout   "top_k": 40,
9:05AM DBG GRPC(gemma-3-12b-it-127.0.0.1:52164): stdout   "top_p": 0.949999988079071,
9:05AM DBG GRPC(gemma-3-12b-it-127.0.0.1:52164): stdout   "typical_p": 1.0,
9:05AM DBG GRPC(gemma-3-12b-it-127.0.0.1:52164): stdout   "temperature": 0.8999999761581421,
9:05AM DBG GRPC(gemma-3-12b-it-127.0.0.1:52164): stdout   "repeat_last_n": 0,
9:05AM DBG GRPC(gemma-3-12b-it-127.0.0.1:52164): stdout   "repeat_penalty": 0.0,
9:05AM DBG GRPC(gemma-3-12b-it-127.0.0.1:52164): stdout   "frequency_penalty": 0.0,
9:05AM DBG GRPC(gemma-3-12b-it-127.0.0.1:52164): stdout   "presence_penalty": 0.0,
9:05AM DBG GRPC(gemma-3-12b-it-127.0.0.1:52164): stdout   "mirostat": 0,
9:05AM DBG GRPC(gemma-3-12b-it-127.0.0.1:52164): stdout   "mirostat_tau": 5.0,
9:05AM DBG GRPC(gemma-3-12b-it-127.0.0.1:52164): stdout   "mirostat_eta": 0.10000000149011612,
9:05AM DBG GRPC(gemma-3-12b-it-127.0.0.1:52164): stdout   "n_keep": 0,
9:05AM DBG GRPC(gemma-3-12b-it-127.0.0.1:52164): stdout   "seed": 48382911,

As you can see stuff like top_k is getting ignored, even temperature is not set to 1.0
I found these settings somewhere claiming that they'd be good for gemma-3 but I am unable to test them by sending them via prompt. It should work though, since openai also defines the property "temperature" to set it with each call. If I send a prompt directly to localai via chat completions api it does work with temperature so the bug seems to be in responses api.

Start by tracing the Responses API request handling and compare how generation parameters are forwarded with the chat completions API path. Reproduce the request using temperature, top_k, and the other listed settings, then verify that the values received by LocalAI match the request.