OpenAI API implementation is broken - shows up only for agentic coding

[Expro]

For about a month, LocalAI fails to work correctly with agentic coding tools - issue is visible across several different tools: Visual Studio Code + Kilo Code, Visual Studio Code + Roo Code, Zed.

Issue manifest as empty chat responses from agent - just list of API Requests. Assistants performs works over code, but says nothing.

I have initially assumed that it may be issue with llama.cpp, but since there has been tens of llama.cpp updates. I have also hosted model externally via vLLM and those tools works correctly with it, so that eliminates agentic coding tools as source of issues.

How to reproduce:

1. Install Visual Studio Code with Kilo Code extension OR install Visual Studio Code with Roo Code Extension OR install Zed.
2. Host any model on LocalAI.
3. Add model from point 2. to tool from point 1. via OpenAI API Compatible provider.
4. Ask agent to perform any work on project - doesn't really matter what.

[mudler]

Could you specify which models are you using for tests and how are you installing these? I've been using Opencode + Qwen3 Next + LocalAI with llama.cpp backend and works (as a proof, actually, #8552 was opened by LocalAGI/LocalAI configured with opencode and Qwen Next coder). I have occasional issues but it's being tracked upstream here ggml-org/llama.cpp#19382

[Expro]

I have tested following models: MiniMax-M2.1, Qwen3-VL-32B-Instruct, Nemotron-3-Nano-30B-A3B, MiniMax-M2.1-REAP-172B-A10B, Qwen3-Coder-Next. They all are from Huggingface (I think all of them are from Unsloth), not gallery, so that may be a factor?