[Bug] Agent mode does not work with directories or files when using Ollama (v1.2.5)

[Amuwa]

1. Press Cmd+Shift+P in Void, and type Help: About. Paste the information here!

Void Version: 1.2.5
Release: 25107
VSCode Version: 1.99.0
Commit: 6f55e6900b6fce0e331074811f02bd68f15f293f
Date: 2025-04-17T06:02:24.513Z
Electron: 34.3.2
ElectronBuildId: undefined
Chromium: 132.0.6834.210
Node.js: 20.18.3
V8: 13.2.152.41-electron.0
OS: Darwin arm64 24.3.0

2. when switched to the agent mode and ask Void to create some code (via ollama), all commands and codes just go within the chat window. Even when click the little "apply" button above them, the commands and codes go nowhere, until you create those directories and files manually and open it with current tab.

In the configuration, "fast apply" and "auto-approve" are chosen.

[andrewpareles]

What Ollama model are you using?

[Amuwa]

What Ollama model are you using?

I've tried the following models:

deepcoder:latest qwq:latest
gemma3:27b
phi4:latest
deepseek-r1:32b
mistral-small:latest
llama3.2:latest

[bonswouar]

I've only tried with deepcoder but I confirm the same behaviour.

Also, when linking some existing files, it sometimes seemed to have only the filepath and not the actual file content (not sure how to reproduce, it wasn't happening at first)

[hokimtam]

Same problem, any idea to fix it?

[hlesesne]

+1

VSCode Version: 1.99.30001 Void Version: 1.2.8 Commit: df721c5370330e8595904681fccac163f11e4cef Date: 2025-04-29T09:25:38.597Z Electron: 34.3.2 ElectronBuildId: undefined Chromium: 132.0.6834.210 Node.js: 20.18.3 V8: 13.2.152.41-electron.0 OS: Linux x64 6.11.0-21-generic

[sidster-io]

Hi, I have the same issue, I asked @graphcms.api.ts Add comments to all the functions, it does provide appropriate comments on the chat window even though I selected Agent mode. I'm using local qwen2.5-coder:7b-instruct-fp16 model. I press ▶️ Apply, it attempts to run but does nothing after a while.

Note: Using Fast Apply

[karim2001]

Same issue here

[michalszmidt]

Try with this model https://huggingface.co/bartowski/all-hands_openhands-lm-32b-v0.1-GGUF There is project called openhands with similar goal thus could learn from their experience see for instance this, they trained their model. Tough they say it shouldn't be used as gguf. Dunno why.

I had partial-success with it. Could you share your experience with it plz?

UPDATE: This is why it's partial: