How to control Timeouts

[mrbm]

I am running ollama on my local network on a dedicaled m4 24gig with Qwen3:14b Quant with 65k context window.

I had it create a slightly more than basic react app and it created things, wrote some of the thngs in the dialog and wrote 50% of the things to disk, the other 50 it didn't I gave it a command to write the rest of the items...and then it did some, but not all of it, i did another command to finish...and now it just timesout.

I am reading the ollama serve logs and noticed that open code is sending a timeout singnal around 4 min ond each time.

So...the issue is opencode.

I didn't set a timeout in the config. Then i tried setting a timeout in the opencode config to 600 seconds, same results.

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "ollama-local-remote": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "http://192.168.XX.XXX:11434/v1",
        "timeout": 600000
      },
      "models": {
        "llama3:8b": {"name": "Ollama Llama3"},
        "qwen3:14b-65k": {
          "name": "qwen3:14b 65k (remote local)",
          "tools": true,
          "reasoning": true,
          "options": { "num_ctx": 65536 }
        },
        "qwen3:14b-q4_K_M": {
          "name": "qwen3:14b-q4_K_M (remote local)",
          "tools": true,
          "reasoning": true
        }

Is this the right way to do it? Or is this something that can't be changed. Thanks! OpenCode.ai special sauce is going to be for those homebuilds which are going to have less resources and as a result opencode needs to be able to wait longer. Love the community thoughts or is this coming soon. I think there was 3 releases yesterday so i know things are happening rapidly.

[thdxr]

hmm im not sure where this timeout is being set - i don't think that timeout option exists

[mrbm]

I am not attempted to be super familiar with the code base, but just a quick search brings me to this:

https://github.com/sst/opencode/blob/dev/packages/opencode/src/util/timeout.ts

Is there a way to set this project to wait a certain time length max or indefinitely? Or does this have a default waiting period occuring at some stage of the project? Maybe some &http.Client method has a default timeout? I'm in a dr office waiting with my eyes dialated so can't really see anymore.. but maybe these notes help

[douglarek]

Maybe same as #655, opencode does not complete a download service but ends abruptly. My scenario is upgrading software packages in the Gentoo overlay. The download stops halfway, causing the Manifest that records the package to write an incorrect checksum.

[mrbm]

https://github.com/sst/opencode/blob/60b55f9d921805a4a9d38c9eed05a060df08df82/packages/opencode/src/tool/bash.ts

i see timeouts here, i have no idea how this might be effecting http.client requests to an llm or requests for other things, but i think there is a possible issue somehere.

[mrbm]

@thdxr i think the real pillar of strength that claude and google can't take from this project will be the ability to use local models which are obvious always going to run slower. I do think testing on a slow response scenario where from a separate machine running ollama might not response for 7 min while it thinks is a worthy test case to backtrack and find where open code might be closing the connection pre-maturely.

[thdxr]

yep will look into this

[noahfraiture]

I agree that timeout would be very convenient, but also for other capabilities like bash execution

[ggaabe]

running into timeouts from slow local model responses as well. would be great to control.

[dc-tec]

Same here when running a larger local model, would be nice if we could configure the timeout.

[dysangel]

same. My model was trying to do multiple tool calls and eventually timing out. I'd be happy to just be able to disable the timeout for local use

[egyptianbman]

Same here, getting constant timeouts with ollama. When this happens, the request is not aborted so the model keeps churning and subsequent requests are more likely to fail as the model is busy.

[dysangel]

I've been looking into this the last few days. By default a lot of network requests time out after 5 minutes. I'm submitting a PR to mlx-lm to send back SSE comments which notify that the prompt is currently processing and keep the connection alive. If necessary we could make similar keepalive tweaks to ollama and llama.cpp etc to allow time for local models to process very large contexts without all the clients disconnecting

[jamviking]

As I am experiencing the same timeout issue when running OpenCode against a local LMStudio installation, us local users need a way to override the default NodeJS timeout as documented here: https://blog.appsignal.com/2023/11/08/how-to-use-timeouts-in-nodejs.html

Hope this helps as I need it...

[dysangel]

@jamviking if you're using an MLX model, then yes that issue is still present