Repository navigation
Opencode in linux cannot access tools or create files #729
Description
I am using opencode with devstral / ollama in ubuntu. Devstral keeps telling me it has no access to files on my computer, and keeps giving me "advice" on how to implement stuff myself. Am I missing some config?
Activity
JonBoyleCoding commentedon Jul 6, 2025
I had the same issue with many ollama models. I wonder whether it's an issue with the models being trained in such a way to not expect file manipulation - but it seemed so commonplace that I wondered otherwise. GPT4 through copilot also had the same issue for me - it wouldn't want to run commands or do anything and then just throw it back to me.
codyseally commentedon Jul 6, 2025
No idea, but I was watching THIS GUY showing how he used Devstral with opencode, so I said to give it a try. I get this:
I see that he is not using ollama but Unsloth for getting devstral. The whole point to me was to use local models, otherwise I'm not really interested in another API client.
LE: Tried the exact devstral + ollama setup with aider, works with no problems. The issue is clearly in opencode.
JonBoyleCoding commentedon Jul 6, 2025
Intriguing - I presume then this is explicitly an ollama issue with opencode. I have been using opencode on some of my open repositories that I don't care about with claude and github copilot and it works fine.
codyseally commentedon Jul 6, 2025
It most likely has to do with the way opencode formats tool use. Probably it is incompatible with ollama's tools use api, the same way aider is. I guess the json format in which the llm served with ollama calls a tool is not the expected json for opencode.
Sadly I have to stick to aider, although I really like opencode's interface, but it's useless to me if it does not work with ollama.
KodeMunkie commentedon Jul 7, 2025
I'm having the same issue. Qwen 30B (MoE) seems to work well in the likes of Codename Goose, but sadly on opencode (which tbh is much nicer to use) it "thinks" it's unable to write or create any files, which also affects /init
ajunca commentedon Jul 8, 2025
I can confirm with the same issue. I will attach an screenshot of running Devstral small 24b Q8 with a context window of 90k.
I also have to say, that in the configuration, the "limit" part is not working, as it defaults to 4k context window ignoring this parameters. I had to create an ollama model with devstral an default 90k context length. This is the config is ignoring the context:
So with a 90k window, all input can fit on the LLM, this is the results I'm getting. Devstral knows about the tools, and "calls them", but nothing happens. And in the example, "test.txt" file is not created (nor AGENTS.md).
abate commentedon Jul 9, 2025
Same same, I've tried with qwen3:8b . I'm not able to run
/initsuccessfully and to generate the AGENTS.md fileAlso some models out there like gemme3n:latest do not support
tools. Maybe we can compile a list of local models compatible with opencode ?adamdotdevin commentedon Jul 9, 2025
this is likely an issue with the default system prompt we use. dax is working on custom system prompts per model (and configurable) today.
codyseally commentedon Jul 9, 2025
But would you need a custom prompt per model or a tool-calling prompt that works with Ollama ? My guess is that Ollama has a fixed way in which all models call tools and that just needs to be made compatible with Opencode. Otherwise you would need to design new prompts every time a new model comes out.
KodeMunkie commentedon Jul 9, 2025
Whatever it is, please avoid if you can, whatever Codename Goose uses.
Goose is hit and miss whether in any one session tooling works without telling it (and reminding) and the massive prompts it have fill up the context (it's 3.6K tokens by default it seems, based on a simple hello prompt and viewing the Ollama logs).
skirmish commentedon Jul 9, 2025
I have a short list of ollama models that support tools calls that I have tried to use:
mistral:7b,qwen2.5-coder:3b,llama3.2:1banddevstral:24b. When you're looking at the ollama models, you can see which models support tools on the model listing page.But all of these do not appear to work with opencode with ollama. Looking at the logs from ollama, it does appear to be trying to use tools, but I haven't parsed the logs in any depth to troubleshoot yet.
ahmed-bekhet commentedon Jul 9, 2025
I am also facing the same issue, but I noticed something that is not the main issue but may trip people in the future and they wouldn't notice what's happening.
The default context size of ollama is 4096, the context size that you may set in opencode.json doesn't affect that.
I suspect this could also cause the model to not be aware of the tools available to it, if the tool description is part of the system prompt.
When I increased the context size, I started noticing that the model actually attempts to use the tools, before that it was just trying to give me instruction on how to do the things I requested myself.
Here is an example of ollama logs when it received requests from opencode
The solution I found for that was to create another model in ollama from my main model, you can find instructions for that here:
https://github.com/ollama/ollama/blob/main/docs/modelfile.md
Here is how I setup mine:
Create a file with any name
Then add these configuration to ollama:
ollama create <new-model-name> -f <path-to-model-file>Now when you use that new model name in opencode it will no longer run into the issue where it has to truncate the context as easily.
The context size can also be set in the request to ollama and it would be nice if opencode can support that as it will give more flexibility.
JonBoyleCoding commentedon Jul 10, 2025
I had thought the solution would be more technical than just the custom prompts as I've seen people use other backends for similar models and have largely different responses. I hadn't thought to try what @ahmed-bekhet tried. When I have a chance I'll see if I can confirm it works for me.
ahmed-bekhet commentedon Jul 10, 2025
Oh, I didn't mean to imply that this fixed the issue, it didn't. It just allows the full context of the "/Init" to be passed on to the model.
Even after receiving the full context and the tools' details it still doesn't really call tools with any consistency. I checked the http traffic and the tool calls weren't there, I played around a little with temperature and with context length and max output tokens, but it didn't really help.
I ended up switching to LM Studio with the exact same LLM and it's working and calling the tools correctly.
8 remaining items
ajunca commentedon Jul 17, 2025
@codyseally Devstral Small 2507 with enough context length is working for me in LMStudio, so the issue here is with Ollama. Check that you have enough context length. If you don't have, you will see on the LMStudio logs something on the line that had to cut the input because was bigger than the context.
Check also the config is okey, for reference, this is mine for LMStudio provider/model:
Also for reference, I've configured the model in LMStudio with this parameters:
PD: I have 48GB VRAM (two 3090), so if you have 24GB VRAM probably can stretch it to 64000 context length.
renich commentedon Jul 18, 2025
OK. ollama user here in Fedora 42.
I was looking for this solution for a while. This is why I'm posting it here (and all issues) for all to know:
@p-lemonish discovered why and the fix is easy: https://github.com/p-lemonish/ollama-x-opencode
In short, download a model that has tool capabilities:
After that, you're good to go. Just try it out:
With a bit of luck, your file will be there.
Credits:
Pre-requisites
My configuration looks like this:
{ "$schema": "https://opencode.ai/config.json", "provider": { "ollama": { "npm": "@ai-sdk/openai-compatible", "options": { "baseURL": "http://localhost:11434/v1" }, "models": { "granite3.3:8b": { "tools": true }, "mistral-small3.2:24b": { "tools": true }, "qwen3:latest": { "tools": true, "reasoning": true } } } } }Notes
renich commentedon Jul 18, 2025
Try what p-lemonish suggests. Just save the model with the
num_ctxset and you'll succeed.abate commentedon Jul 21, 2025
I confirm that this works !
I noticed this in the ollama logs :
using Qwen3 8B
Sntx626 commentedon Jul 25, 2025
I'm replying here, since this is issue has seen the most activity on what I suspect is opencode not properly interpreting ollama tool call responses.
These are issues that seem to share the problem: #234 #526 #590 #1034 #1068
I recommend continuing the discussion here.
Example Session
Conversation History
User (2025-07-25 11:58:48)
Summarize the project.
Assistant (2025-07-25 11:58:48)
This project is a Rust-based simulation with separate crates for physics and simulator components. The main crate appears to be the executor, which calls into the physics library. The structure suggests that any code running on the GPU should go in the
executorcrate rather than thephysicscrate, since this will allow it to remain platform agnostic while still being able to use platform specific optimizations.User (2025-07-25 12:00:15)
Can you add a README.md?
Assistant (2025-07-25 12:00:15)
[TOOL_CALLS]todowrite[ARGS]{"todos":[{"content":"Add initial README.md file with project overview", "status":"pending", "priority":"high"}]}
User (2025-07-25 12:01:23)
Please analyze this codebase and create an AGENTS.md file containing:
The file you create will be given to agentic coding agents (such as yourself) that operate in this repository. Make it about 20 lines long.
If there are Cursor rules (in .cursor/rules/ or .cursorrules) or Copilot rules (in .github/copilot-instructions.md), make sure to include them.
If there's already an AGENTS.md, improve it if it's located in /home/sntx/Documents/projects/2025-07-23_stars
Assistant (2025-07-25 12:01:23)
I'll help you create the AGENTS.md file. First, let me search for any existing build/lint/test commands or style guidelines in the repository:[TOOL_CALLS]bash[ARGS]{"command": "cd /home/sntx/Documents/projects/2025-07-23_stars && find . -name "Cargo.toml" -o -name ".cursorrules" -o -name "Makefile" -o -name "*.md" | grep -i -E "(build|lint|test|style|copilot|cursor)"", "description": "Find configuration files that might contain build, lint, test commands or style guidelines"}
User (2025-07-25 12:02:52)
What tools can you call?
Assistant (2025-07-25 12:02:52)
I can use a variety of tools to assist with software engineering tasks, including:
Code Editing and File Operations:
Search and Analysis:
Version Control:
Project Management:
Build and Test Operations:
Documentation and Reference:
User (2025-07-25 12:04:29)
Display the project structure.
Assistant (2025-07-25 12:04:29)
[TOOL_CALLS]bash[ARGS]{"command": "cd /home/sntx/Documents/projects/2025-07-23_stars && ls -la", "description": "List all files and directories in the root folder"}
User (2025-07-25 12:05:26)
Can you also list all relevant subdirs?
Assistant (2025-07-25 12:05:26)
[TOOL_CALLS]bash[ARGS]{"command": "cd /home/sntx/Documents/projects/2025-07-23_stars && find . -type d | grep -v "\.git"", "description": "Find all directories in the repository, excluding .git"}
Devstral Modelfile
I slightly adapted the system prompt to include opencode instead of openhands and reduced the dependence on GitHub.
$XDG_CONFIG_DIR/opencode/config.json
{ "$schema": "https://opencode.ai/config.json", "provider": { "ollama": { "npm": "@ai-sdk/openai-compatible", "options": { "baseURL": "http://10.0.1.1:11434/v1" }, "models": { "devstral:24b-small-2507-q4_K_M": { "tool_call": true }, "mistral-small3.2:24b": { "tool_call": true } } } } }As we can see from multiple user reports in this issue, the llm often try to call tools12, however I assume opencode fails to parse the tool calls.
@thdxr Does my assumption hold?
Another thing of note is that specifially mistral models follow this cool_calls specification, is that supported by opencode?
Footnotes
Example 1 ↩
Example 2 ↩
jaswdr commentedon Jul 25, 2025
which models in the end you used with LM Studio that worked?
KodeMunkie commentedon Jul 27, 2025
I just wanted to add I'd be interested in knowing what was used in LM Studio that worked.
I've been using hunyuan-a13b-instruct, with a 32768 context length, but it still doesn't work, so the context length is only part of the problem.
It's quite intelligent - it does attempt to call tools (saying it does), no errors are shown, but actually gets confused when it tries to check the directory - for example where it's saved AGENTS.md - and then doesn't find anything, reasoning in so far as suggesting there's a tool bug (if only we could get it to fix it 😆 )
ajunca commentedon Jul 27, 2025
@KodeMunkie you will find what worked for me in LM Studio in my last entry:
PilotDevo commentedon Aug 20, 2025
The below also worked for me... I still think im going to move away from Ollama and leverage vLLM or something similar and spread across multiple gpu's for better performance/higher context.... just doesn't perform as well for heavier lifting.
jasperf commentedon Nov 17, 2025
First with Ollama installed qwen3:8B it said it could not write
...Since I can't access the file system.... Then it said the path was not clear though init command gave that. And that with just using basic/initTool: write
my
opencode.json:{ "$schema": "https://opencode.ai/config.json", "provider": { "ollama": { "npm": "@ai-sdk/openai-compatible", "name": "Ollama (local)", "options": { "baseURL": "http://localhost:11434/v1" }, "models": { "qwen3:8b": { "name": "Qwen3 8B" }, "granite3.1-moe:latest": { "name": "Granite 3.1 MoE" }, "qwen3:4b": { "name": "Qwen3 4B" } } } } }and Ollama list
Ollama server logs look fine:
but main issue seems to be
and feedback is
Update
Did manage to do do some basic tasks with Qhen 8B 16K context and Open Code. Was able to create a file with basic todo list. The
/initcommand still fails, also when fileAGENTS.mdexistts . Keep on complaining it is a binary file. Added details to https://github.com/imagewize/ollama-opencode-setup . Not done testing yet, but progressDigital-Yeti commentedon Jan 16, 2026
I have successfully gotten
devstral-small-2:latestto perform the/initcommand and create the agents.md file. I set the num_ctx parameter to 40960, using the method discussed earlier, and saved the model. Just increasing the context window seems to have resolved the lack of tool calling for me, at least for this model. I am unsure how much VRAM is required for this but it works on my 24GB 3090.thdxr commentedon Mar 25, 2026
To stay organized issues are automatically closed after 90 days of no activity. If the issue is still relevant please open a new one.
feat: add click-path-audit skill — finds state interaction bugs (anom…
fix(sidebar): preserve path case for directory queries (anomalyco#729)