[eriktews]

Description

By default, karakeep will only wait for 5 minutes for HTTP headers, no matter which timeout is set.

This fixes the issue by setting the timeouts for the HTTP headers based on the interference job timeout.

How Has This Been Tested?

Tested manually with Ollama running locally.

Checklist:

• I have carefully read CONTRIBUTING.md
• I have performed a self-review of my own code
• I have made corresponding changes to the documentation if applicable
• I have no unrelated changes in the PR.
• I have confirmed that any new dependencies are strictly necessary.
• I have written tests for new code (if applicable)

Please describe to which degree, if any, an LLM was used in creating this pull request.

Code was partially generated using Opus 4.6.

[greptile-apps]

Greptile Summary

This PR fixes the undici default headersTimeout (5 minutes) overriding the configured inference fetch timeout by creating a custom undici Agent with headersTimeout/bodyTimeout set to the configured value, and adds a setTimeout-based fallback mechanism for non-undici fetch implementations.

• A mergeSignals helper is introduced to combine the internal AbortController signal with any caller-supplied signal, so callers can still cancel requests independently of the timeout.
• The undici Agent is resolved once at factory time via a dynamic import(\"undici\"), with a .catch(() => undefined) fallback so the code works in environments without undici.
• A manual header-timer setTimeout fires if headers are not received within timeoutMs, and a body-timer is set once headers arrive; the body-timer is never cleared on successful body consumption, leaving a dangling handle for up to timeoutMs after every successful request.

Confidence Score: 4/5

The change is safe to merge; the core fix (passing a properly-configured undici Agent) is correct and addresses the reported Ollama timeout issue.

The timeout logic is functionally correct for both the headers and body cases. The body timer never being cleared means a dangling handle lives for up to the full configured timeout after every successful inference call, but it fires harmlessly on a completed request and does not produce wrong results or data loss.

packages/shared/customFetch.ts — the body timeout timer cleanup path deserves a follow-up.

Important Files Changed

Filename	Overview
packages/shared/customFetch.ts	Replaces the simple AbortSignal.timeout approach with a dual-mechanism timeout (undici Agent + manual setTimeout), adding header and body timeout support. The body timer is never cleared on successful completion, leaving lingering handles, and the file is missing its trailing newline.

Prompt To Fix All With AI

Fix the following 2 code review issues. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 2
packages/shared/customFetch.ts:51-55
**Body timeout timer never cleared on success**

The body timer is set inside `.then()` but has no corresponding `clearTimeout` call. After the response body is fully consumed (e.g., Ollama finishes streaming tokens), the timer still fires `timeoutMs` later and calls `controller.abort()` on an already-completed request. For a 5-minute configured timeout, every successful request leaves a live timer handle for up to 5 minutes after completion. Under concurrent AI inference jobs this accumulates into unnecessary timer overhead. The fix requires wrapping the returned `response.body` stream to intercept the end-of-body event, or tracking whether the body is done via a flag cleared in a stream `transform`/`pipeTo` wrapper.

### Issue 2 of 2
packages/shared/customFetch.ts:65
Missing newline at end of file.

```suggestion
export const customFetch = createCustomFetch();
```

_{Reviews (1): Last reviewed commit: "Fix headers timeout for AI interference ..." | Re-trigger Greptile}

[eriktews]

I have updated this pull request so that fixes the header timeout for OpenAI compatible APIs as well.

[eriktews]

This fixes #2994

[MohamedBassem]

@pullfrog review

[pullfrog]

ℹ️ No critical issues — one minor note inline.

Reviewed changes

• packages/shared/customFetch.ts — replaced the AbortSignal.timeout wrapper with a helper that pairs npm undici's own fetch with an Agent/ProxyAgent whose headersTimeout/bodyTimeout are set to the supplied timeout, so undici's 5-minute defaults no longer override the configured value.
• packages/shared/inference.ts — the OpenAI client now wires OPENAI_TIMEOUT_SEC (falling back to OpenAI.DEFAULT_TIMEOUT) into that fetch, and the Ollama client uses INFERENCE_FETCH_TIMEOUT_SEC; the previous fetchOptions.dispatcher proxy plumbing is removed.
• docs/docs/03-configuration/01-environment-variables.md — clarified that both variables now map to HTTP header/body timeouts, and noted the OpenAI SDK's retry behavior.

The core fix is correct: bodyTimeout is undici's inactivity timer between body chunks, which matches the documented behavior, and OpenAI.DEFAULT_TIMEOUT keeps the fetch timeouts aligned with the SDK's own default when OPENAI_TIMEOUT_SEC is unset. Typecheck and the existing inference.test.ts suite pass.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}

[MohamedBassem]

Thanks @eriktews, I've simplified the PR a bit while hopefully should still fix the bug you faced.