[BUG] Fetch failed after 5m0s regardless of configured timeout

[nalimcos]

Problem (one or two sentences)

Hi,
I am using Roo Code in VSCodium with Ollama as inference provider, on limited hardware resulting in long processing times.

This usually works fine at the beginning of a task (or with very small models), but as processing times climb with increased context size, this consistently ends with a 'fetch failed' error in Roo Code.

Context (who is affected and when)

Happens to users running Ollama as inference provider (and probably others), when their configuration leads to 5min+ response times.

Reproduction steps

Roo Code 3.51.1
VSCodium:
Version: 1.110.11607 Commit: 8c11aee2be8d4bcafcd9812191f979df54d2c007 Date: 2026-03-07T23:26:50.681Z Electron: 39.6.0 ElectronBuildId: undefined Chromium: 142.0.7444.265 Node.js: 22.22.0 V8: 14.2.231.22-electron.0 OS: Linux x64 6.9.8-amd64
This is on Debian testing.

Can be reproduced by running a slow Ollama server (large dense model with long context length, limited hardware such as CPU-only inference).

Expected result

The task keeps progressing

Actual result

The tasks shows progress at first, but at some point requests start failing with 'fetch failed'; recovery without switching to a faster model is rare.

Variations tried (optional)

5m0s is 300s, which apparently matches Undici / Node.js default fetch timeout.
Luckily, VSCodium has a "Use Electron Fetch" setting (http.electronFetch in json config); when enabled, I no longer get 'fetch failed' errors.

App Version

3.51.1

API Provider (optional)

Ollama

Model Used (optional)

N/A

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

My Ollama container docker logs show the following:

[GIN] 2026/03/16 - 00:50:26 | 500 |          5m0s |      172.22.0.1 | POST     "/api/chat"
time=2026-03-16T00:50:26.111Z level=INFO source=runner.go:682 msg="aborting completion request due to client closing the connection"

[roomote-v0]

I've analyzed this issue and here's my proposed implementation plan:

Root Cause: The NativeOllamaHandler in src/api/providers/native-ollama.ts creates an Ollama client without configuring any fetch timeout. The ollama npm package uses Node.js native fetch internally, which is powered by Undici and has a default timeout of 300 seconds (5 minutes). This matches exactly the 5m0s timeout the user observes. The comment in the code saying "The ollama npm package handles timeouts internally" is incorrect -- the package simply inherits the Undici default.

Meanwhile, all other OpenAI-based providers (OpenAI, LM Studio, base-openai-compatible) already respect the user-configurable apiRequestTimeout setting (default: 600s) via getApiRequestTimeout() in src/api/providers/utils/timeout-config.ts. The Ollama handler is the only provider that does not honor this setting.

Proposed Fix:

1. Modify NativeOllamaHandler.ensureClient() in src/api/providers/native-ollama.ts to pass a custom fetch wrapper to the Ollama client that applies the apiRequestTimeout setting via AbortSignal.timeout(). The ollama npm package's Config type accepts a fetch option for exactly this purpose.

2. The implementation would look roughly like:

   import { getApiRequestTimeout } from "./utils/timeout-config"
   
   // In ensureClient():
   const timeout = getApiRequestTimeout()
   if (timeout) {
     clientOptions.fetch = (url, init) => {
       return fetch(url, {
         ...init,
         signal: init?.signal ?? AbortSignal.timeout(timeout),
       })
     }
   }

3. Add/update tests in src/api/providers/__tests__/native-ollama.spec.ts to verify the timeout is applied to the Ollama client.

This approach ensures Ollama users with slow inference (5min+) benefit from the same configurable timeout that other providers already use, without breaking any existing behavior.

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

[nalimcos]

@roomote That sounds reasonable, proceed.

[roomote-v0]

Thanks for the approval! I'll proceed with the implementation now. I'll open a PR shortly to share progress.

[roomote-v0]

Opened a draft PR with my attempt at addressing this: #11933

The fix passes a custom fetch wrapper to the Ollama client that applies the user-configured apiRequestTimeout setting via AbortSignal.timeout(). This replaces the hardcoded 300s Undici default with the configurable timeout (default 600s). Feedback welcome!