Experiences and recommendations for using CLI agents

Hi,

First of all, thank you for this great project. I've been exploring the use of CLI-based coding agents (like Pi Agent and OpenCode) with FastFlowLM as the backend, and I'd like to share my experience and ask for guidance.

What I'm trying to do

Use FastFlowLM as the LLM provider for coding agents like Pi Agent and OpenCode, which rely on tool calling (file read/write, bash execution, etc.) to perform code generation tasks.

What I've tried

I have tested several models with different results. Using gpt-oss:20b, tool calls are simply not triggered — the agent fails to execute any tool. With nanbeige4.1:3b, the model freezes constantly and responses never complete. gemma4-it:e4b shows similar behavior, freezing repeatedly and never finishing a response.

What I'd like to know

1.Has anyone successfully used FastFlowLM with a CLI coding agent (Pi Agent, OpenCode, or similar)?
2.Which models from the available list work best for tool calling and agent workflows?
3.Are there any specific server configurations or flags needed to enable proper function calling support?
4.Is tool use / function calling currently supported in FastFlowLM's OpenAI-compatible API, or is it planned?

Any experiences, recommendations, or a brief guide on the topic would be greatly appreciated. I'm happy to provide more details about my setup or run additional tests if needed.

FLM v0.9.43 on AMD RYZEN AI 7 350

flm serve --port 8080 --ctx-len 32768 --pmode turbo  --host 0.0.0.0 --cors 0

flm validate
[Linux]  Kernel: 7.1.0-rc6-1-cachyos-rc
[Linux]  NPU: /dev/accel/accel0 with 8 columns
[Linux]  NPU FW Version: 1.1.2.64
[Linux]  amdxdna version: 0.15
[Linux]  Memlock Limit: infinity

Thanks!

Start with the flm serve configuration and the OpenAI-compatible API behavior described in the issue; compare how tool calls behave with the named CLI agents and models. Done means documenting whether function calling is supported, which configurations or models work, and whether additional implementation is needed.