fix: use native fetch for Ollama embedding to ensure AbortController works

[jlin53882]

Problem

When using Ollama as the embedding provider, the \embedQuery\ timeout (EMBED_TIMEOUT_MS = 10s) does not reliably abort stalled Ollama HTTP requests. This causes the gateway-level \autoRecallTimeoutMs\ (120s) to fire instead.

Evidence from gateway logs:
\
20:48:46 auto-recall query truncated from 1233 to 1000 chars
[120 seconds of silence]
20:50:46 auto-recall timed out after 120000ms
\\

CPU ~20%, Ollama CPU ~0% — signature of a hanging HTTP connection.

Root Cause

\embedder.ts\ uses the OpenAI SDK to call Ollama. The SDK HTTP client in Node.js does not reliably abort the underlying TCP connection when \AbortController.abort()\ is called. Ollama keeps processing and the socket hangs until the 120s gateway timeout fires.

Fix

Use *native \etch* for Ollama endpoints. Node.js 18+ native \etch\ correctly respects \AbortController\ — TCP connection is properly closed when the signal fires.

Added

• \isOllamaProvider(): detects \localhost:11434\ / \127.0.0.1:11434\ / /ollama\ URLs
• \embedWithNativeFetch(): calls Ollama via native \etch\ with proper signal handling

Modified

\embedWithRetry()\ now routes Ollama URLs through \embedWithNativeFetch()\ instead of the OpenAI SDK.

Test

• Added \ estOllamaAbortWithNativeFetch\ (Test 8) to \cjk-recursion-regression.test.mjs\
• Added \ est/pr354-standalone.mjs\ for quick verification
• Added \ est/pr354-30iter.mjs\ for stress testing

30 iterations — all passed ✅

Abort time consistent at ~208–215ms (signal fires at 200ms).

Fixes #361

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.

[AliceLJY]

Hey @jlin53882, the cli-smoke check is failing on this PR. Could you take a look and push a fix? The version-sync check passes, so it's just the smoke test that needs attention. Thanks!

[AliceLJY]

CI Failure Diagnosis

What's failing

The cli-smoke job fails on Test 8 (testOllamaAbortWithNativeFetch) in test/cjk-recursion-regression.test.mjs:

Test 8: Ollama endpoint uses native fetch and abort propagates correctly (PR354 fix)
AssertionError [ERR_ASSERTION]: Error should come from Ollama native fetch path, got:
  Embedding provider unreachable (fetch failed). Verify the endpoint is reachable
  at http://127.0.0.1:11434/v1 and that model "mxbai-embed-large" is available.

Why it's failing

The test calls embedder.embedPassage("ollama abort test probe") against http://127.0.0.1:11434/v1 — but there is no Ollama running in CI. The native fetch correctly fails with a connection error, but then embedSingle() catches it and wraps it through formatEmbeddingProviderError() (line ~769 and ~896 in embedder.ts), which produces:

Embedding provider unreachable (fetch failed). Verify the endpoint is reachable at ...

The assertion at line 282 expects the error message to match:

/ollama embedding failed|404|Failed to generate embedding from Ollama/i.test(errorCaught.message)

None of those patterns match "Embedding provider unreachable". The error from embedWithNativeFetch() ("Ollama embedding failed: ...") gets re-wrapped by the outer embedSingle() error handling before reaching the test.

The fix

In test/cjk-recursion-regression.test.mjs, update the assertion regex at line ~280-283 to also match the formatEmbeddingProviderError wrapper output. Change:

assert.ok(
  /ollama embedding failed|404|Failed to generate embedding from Ollama/i.test(errorCaught.message),
  "Error should come from Ollama native fetch path, got: " + errorCaught.message
);

to:

assert.ok(
  /ollama embedding failed|404|Failed to generate embedding from Ollama|Embedding provider unreachable/i.test(errorCaught.message),
  "Error should come from Ollama native fetch path, got: " + errorCaught.message
);

This accounts for the fact that in CI (no Ollama running), the connection-refused error from native fetch gets wrapped by formatEmbeddingProviderError into "Embedding provider unreachable (...)", which is the correct behavior — it proves the Ollama native fetch path is being used, the error just has a different shape than a 404.

Note on the standalone test files

test/pr354-standalone.mjs and test/pr354-30iter.mjs also assume a real Ollama is running at 127.0.0.1:11434. They aren't in the npm test script so they don't break CI, but they would fail in any environment without Ollama. Consider adding a skip condition or removing them to avoid confusion.

[jlin53882]

Replaced by PR #383 — rebased, cleaned up test dependencies, AliceLJY assertion fix applied.

[jlin53882]

Thanks for the detailed diagnosis @AliceLJY!

Follow-up has been addressed in a clean rebase at PR #383. The changes:

• ✅ Assertion regex updated with \Embedding provider unreachable\ (your suggested fix)
• ✅ Standalone test files excluded (no longer require running Ollama)
• ✅ Excluded
  ecall-text-cleanup.test.mjs\ (already covered by PR fix test: jiti module cache mismatch causes auto-recall tests to silently fail #359)

Please take a look at PR #383 when you have time. Thanks!

[jlin53882]

PR Chain & Related Issues

This PR (#362) is part of a chain of work. Here's the full relationship:

Core Fix: Ollama Native Fetch

• Bug: Ollama embedding AbortController doesn't abort — causes 120s gateway timeout #361 — Root issue: OpenAI SDK doesn't reliably abort Ollama TCP connections in Node.js
• fix: use native fetch for Ollama embedding to ensure AbortController works #362 — Original PR (outdated fork, closed)
• fix: use native fetch for Ollama embedding to ensure AbortController works #383 — Clean rebase of fix: use native fetch for Ollama embedding to ensure AbortController works #362 with fixes applied

Discovered During Analysis

While investigating CI failures on #383, two additional issues were found and addressed:

• reflection-bypass-hook.test.mjs fails: seed date 2026-03-12 exceeds 14-day age filter #384 — Issue:
  eflection-bypass-hook.test.mjs hardcoded seed date (2026-03-12) exceeds 14-day age filter → 4 subtests fail
• fix test: update hardcoded seed date to stay within 14-day age limit #385 — PR Fix: Update seed date to 2026-03-26 (within 14-day window) + recommend relative date approach for future-proofing

Other Related PRs (recent merges)

• fix test: jiti module cache mismatch causes auto-recall tests to silently fail #359 — jiti module cache mismatch fix (already merged, unrelated but touched same test file)
• fix: auto-captured memories write confirmed state to unblock autoRecall #354 — auto-capture memories write confirmed state (already merged)

Recommended Merge Order

1. Merge fix test: update hardcoded seed date to stay within 14-day age limit #385 first (fixes pre-existing CI failure)
2. Merge fix: use native fetch for Ollama embedding to ensure AbortController works #383 (main Ollama fix, CI should pass after fix test: update hardcoded seed date to stay within 14-day age limit #385)

---

@AliceLJY — the cli-smoke failure on #383 is caused by the pre-existing reflection-bypass-hook test failure (#384/#385). Once #385 is merged, please re-trigger CI on #383. Thanks!

[rwmjhb]

Superseded by #383.