Models provided via ollama always respond that are read-only

[WJKPK]

I'm using following versions of opencode and ollama

$ opencode --version
0.5.28
$ ollama --version
ollama version is 0.11.4

Here is my config.

It look similar for both - devstral & gpt-oss models. If i try to generate /init it also fails in similar way. Last message from screenshot is from grok and this agent is working properly. Messages about read-only mode are independent from mode chosen in TUI (plan/build/review).

How can I provide you more verbose information to help?

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Local Ollama tool calling either not calling or failing outright #1034: Local Ollama tool calling either not calling or failing outright - describes similar issues where Ollama models generate tool calls but they are not executed, instead appearing as plain text
• not being able to write files when using local models #590: not being able to write files when using local models - reports that Ollama models generate tool call JSON but files are not actually written
• Tool use with Ollama models. #1068: Tool use with Ollama models - shows tool call executions being aborted with Ollama models, similar behavior to your read-only responses

These issues all describe similar symptoms where Ollama models appear to generate tool calls but they don't execute properly, which could manifest as the "read-only" behavior you're experiencing.

Feel free to ignore if none of these address your specific case.

[rekram1-node]

Hm you may need to set "tool_call": true in your opencode.json, can you share yours by chance?

ex:

"provider": {
    "ollama": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Ollama (local)",
      "options": {
        "baseURL": "http://localhost:11434/v1"
      },
      "models": {
        "llama2": {
          "name": "Llama 2",
          "tool_call": true
        }
      }
    }
  }

[WJKPK]

I already added my config. Add tool_call helped, but models now have problems with tool calling.

Should we continue here, or should I create new issue?

[rekram1-node]

you may have to prompt for that, most models don't have that issue and you are using a pretty small model there

[WJKPK]

I tried prompting, I tried fixes from #1068 - it doesn't work. I need a tool that works well with local models. Sincere thanks for the open source tool, but my use case is not really supported.

[rekram1-node]

I have heard people get 20b working so I will try to get back to you on if they did anything special

[rekram1-node]

Can you set OLLAMA_CONTEXT_LENGTH=12288 I heard some people say ollama basically doesnt work without it in many cases

[johnjelinek]

note: you may need to tweak the context based on your available memory. 8192 is a good starting point and then increment by 1024 until you get something that works (or memory runs out).

[rekram1-node]

thank you @johnjelinek for jumping in!

[WJKPK]

Hmm, great! After set up of this variable it started working with gpt-oss.
As I am a Nix user, for those interested, here is my ollama with nvidia GPU declaration.

  services = {
    ollama = {
      enable = true;
      package = pkgs.ollama-cuda;
      acceleration = "cuda";
      environmentVariables = {
        OLLAMA_CONTEXT_LENGTH = "16384";
      };
    };

[rekram1-node]

thank you both of you I will try to point people here if they encounter ollama issues local inference issues

[johnjelinek]

I think this mustn't be unique to ollama. I'd be surprised if you didn't have to set the context window likewise in llama-server