Ollama times out (regardless of ollama timeout setting) on long-prefill requests (~300s) due to default Undici timeout.

[hadsie]

Plugin Type

VSCode Extension

Cline Version

3.32

What happened?

When using the Ollama provider, chats that take longer than 5 minutes to produce the first token fail with "fetch failed". This is caused by Undici’s default headersTimeout (~300s).

Expected: The request should use the configured ollama timeout rather than abort at 300s.

Actual: The HTTP request is terminated around 300s with a transport error.

This seems related to:

• Improve Ollama Local Model Compatibility #4362
• Fix ollama timeouts #4310
• Ollama Provider: Timeout Handling for Long-Running Local Model Requests #4308

Steps to reproduce

1. Start a local ollama server and load a model that can have slow prefill with big prompts.
2. In Cline, select the ollama provider and a model with a large context window and a timeout >300 seconds.
3. Send a request with a very large system+user prompt (or otherwise force a long prefill) so that the model takes more than 300 seconds to reply with the first token.
4. Note that the fetch fails at 5 minutes, rather than whenever
   Observe that around the 5-minute mark the request fails with a network/headers timeout and no tokens are delivered.

Provider/Model

ollama/qwen3-30b-a3b-q5

System Information

Cline is running in vscode on a Mac M1 Air, Sequioa.
Ollama running on a GTX 3900 with 24GB VRAM.

[github-actions]

This issue is stale because it has been open for 60 days with no activity.

[linear]

CLINE-282

[github-actions]

Thanks for the detailed bug report and for linking the related issues!

You've correctly identified the root cause. The current Ollama handler (in src/core/api/providers/ollama.ts) implements a manual timeout using Promise.race() which respects the configured requestTimeoutMs. However, the underlying undici fetch dispatcher used in JetBrains/CLI environments has its own default timeouts (~300s for headersTimeout) that trigger before the user-configured timeout can take effect.

Looking at the code, when IS_STANDALONE === "true" (JetBrains/CLI), Cline creates an EnvHttpProxyAgent without customizing its timeout settings:

// src/shared/net.ts line 119
const agent = new EnvHttpProxyAgent({})  // Uses undici defaults

This means even if you set the Ollama timeout to 600s, the undici headersTimeout (default ~300s) will abort the connection first.

Workarounds:

1. Use smaller prompts/context - Break large requests into smaller chunks to reduce prefill time below 5 minutes
2. Optimize model loading - Ensure the model is already loaded in Ollama before sending large requests
3. Use VSCode extension - The VSCode build uses the global fetch which doesn't have the same timeout constraints

Fix direction: The fix would involve passing headersTimeout: 0 and bodyTimeout: 0 to the undici agent configuration in src/shared/net.ts, similar to what PR #4310 attempted (though that PR was closed without merging). This would allow the Ollama handler's manual timeout mechanism to work as intended.

SME	Reason
@dominiccooney	Authored the proxy/fetch infrastructure in #7659 and net.ts
@abeatrix	Recent contributor to Ollama provider
@celestial-vault	Contributor involved in timeout-related issues

Maintainer context: The fix location is src/shared/net.ts line 119. The EnvHttpProxyAgent constructor should include { headersTimeout: 0, bodyTimeout: 0 } to disable undici's internal timeouts and defer to the per-provider timeout handling.

---

Possible Duplicates:

• Support for custom timeout #3029 only fixes the issue for Ollama #6361 - Same underlying issue but requesting the fix for LM Studio/OpenAI Compatible provider (currently only Ollama has a timeout setting)
• When using a local model with longer context, I often start getting "Socket timeout" #3058 - Socket timeout issues with local models/large context (closed, consolidated into Improve Ollama Local Model Compatibility #4362)
• Improve Ollama Local Model Compatibility #4362 - Meta issue for Ollama local model compatibility (closed, but this specific timeout issue wasn't fully resolved)
• Ollama Provider: Timeout Handling for Long-Running Local Model Requests #4308 - Original timeout handling feature request (closed, PR Fix ollama timeouts #4310 attempted to fix but wasn't merged)

[ClineXDiego]

Closing this issue since it targets an older version.
Feel free to reopen or file a new issue after updating if the problem still occurs.

[hadsie]

I've tested this against the latest version and it's still timing out after 5 minutes. I've updated the related PR so that it works with the latest version of the codebase: #6554