Local models via llamacpp's responses API are no longer working

Resolved 💬 4 comments Opened Feb 4, 2026 by Mushoz Closed Feb 5, 2026

💡 Likely answer: A maintainer (etraut-openai, contributor) responded on this thread — see the highlighted reply below.

What version of Codex CLI is running?

v0.94.0

What subscription do you have?

Using a local model

Which model were you using?

gpt-oss-120b

What platform is your computer?

Archlinux

What terminal emulator and version are you using (if applicable)?

Gnome's default console

What issue are you seeing?

When using a model hosted by llamacpp with the following configuration:

[model_providers.llamacpp]
name = "llamacpp"
base_url = "http://127.0.0.1:3001/v1"
api_key = "sk-local"
wire_api = "responses"

[profiles.gpt-oss-120b]
model_provider = "llamacpp"
model = "gpt-oss-120b"
context_length = 131072

I am now seeing the following error messages:

{"error":{"code":400,"message":"'type' of tool must be 'function'","type":"invalid_request_error"}}

This used to work fine before. I updated both Codex & llamacpp, so I am not quite sure where the regression is.

What steps can reproduce the bug?

Configure a llamacpp backend with gpt-oss-120b
Point codex at that endpoint with the proper configuration
Run a query that requires a tool call.

What is the expected behavior?

Tool calls succeed

Additional information

_No response_

View original on GitHub ↗

4 Comments

Mushoz · 6 months ago

Just to add: It also fails when simply sending a "Hi"

Mushoz · 6 months ago

Another finding: The regression is definitely in Codex:

0.91.0 -> Works
0.92.0 -> Not available in Arch's repo, so cannot test
0.93.0 -> Does not work

etraut-openai contributor · 6 months ago

This is likely a problem with the responses API implementation in the llamacpp backend. Please report the issue to the maintainers of llamacpp.

The fact that it worked with a previous version of Codex is not an indication that this is a bug in Codex. We frequently change how we call the responses endpoint. The same calls to the OpenAI responses endpoint work with the latest version of the CLI.

Mushoz · 6 months ago

For future visitors, this thread explains how to fix it: https://github.com/ggml-org/llama.cpp/issues/19343