Greptile Summary

This PR fixes WSL2 networking issues with Ollama by disabling undici's autoSelectFamily option when running inside WSL2, forcing IPv4 connections and preventing TypeError: fetch failed errors caused by unstable IPv6 paths to Windows-host services. It also improves error diagnostics across the Ollama stack by surfacing the nested error.cause.message in both stream error output and provider discovery logging.\n\n- undici-global-dispatcher.ts: resolveAutoSelectFamily() now calls isWSL2Sync() and returns false (override to IPv4) when the system default would otherwise enable autoSelectFamily on a WSL2 host.\n- ollama-stream.ts: Catch block appends err.cause.message to the user-visible error string and logs it via log.warn, helping diagnose undici-wrapped network failures.\n- models-config.providers.discovery.ts: Discovery failure log similarly appends the nested cause message for improved diagnostics.\n- undici-global-dispatcher.test.ts: Adds a ../wsl.js mock and a new test confirming that autoSelectFamily is set to false under WSL2. The beforeEach does not reset the isWSL2Sync mock implementation after vi.clearAllMocks(), which makes the test suite order-dependent — this should be addressed with an explicit vi.mocked(isWSL2Sync).mockReturnValue(false) in beforeEach.

Confidence Score: 4/5

Safe to merge after fixing the isWSL2Sync mock reset in beforeEach to prevent order-dependent test failures.

The core production changes are sound: the WSL2 detection logic is well-established, the IPv4 override is narrowly scoped and idempotency-safe, and the diagnostics improvements are straightforward. The one actionable issue is a test-reliability gap — vi.clearAllMocks() does not reset mockReturnValue implementations, so the new WSL2 test leaves isWSL2Sync returning true for any subsequently added tests. This is a one-line fix in beforeEach and doesn't affect current test outcomes (the WSL2 test is last in its block), but it warrants resolution before merge to avoid silent future regressions.

src/infra/net/undici-global-dispatcher.test.ts — needs explicit isWSL2Sync mock reset in beforeEach.

This is a comment left during a code review.
Path: src/infra/net/undici-global-dispatcher.test.ts
Line: 78-91

Comment:
**Missing `isWSL2Sync` mock reset in `beforeEach`**

`vi.clearAllMocks()` only resets call history (`mock.calls`, `mock.results`), not implementations set by `mockReturnValue`. The WSL2 test at line 160 sets `vi.mocked(isWSL2Sync).mockReturnValue(true)`, and because `clearAllMocks` doesn't reset that implementation, any test added after the WSL2 case will unexpectedly see `isWSL2Sync` returning `true`. This would silently break, e.g., the "replaces default Agent dispatcher" assertion that expects `autoSelectFamily: true`.

Adding an explicit reset in `beforeEach` keeps the suite order-independent:

```suggestion
  beforeEach(async () => {
    vi.resetModules();
    ({
      DEFAULT_UNDICI_STREAM_TIMEOUT_MS,
      ensureGlobalUndiciEnvProxyDispatcher,
      ensureGlobalUndiciStreamTimeouts,
      resetGlobalUndiciStreamTimeoutsForTests,
    } = await import("./undici-global-dispatcher.js"));
    vi.clearAllMocks();
    resetGlobalUndiciStreamTimeoutsForTests();
    setCurrentDispatcher(new Agent());
    getDefaultAutoSelectFamily.mockReturnValue(undefined);
    vi.mocked(hasEnvHttpProxyConfigured).mockReturnValue(false);
    vi.mocked(isWSL2Sync).mockReturnValue(false);
  });
```

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (1): Last reviewed commit: "fix: WSL2 Ollama networking and provider..." | Re-trigger Greptile}

Review: WSL2 Ollama Diagnostics

Verdict: Approve -- clean, focused, and solves a real pain point for WSL2 users.

Reviewed the full diff across all 4 files. This is exactly the right scope (compared to the earlier #55187 which bundled unrelated changes). Notes below from two independent review passes.

What works well

• Targeted fix: autoSelectFamily: false on WSL2 sidesteps the IPv6 routing mess that breaks Ollama connections. Uses the existing isWSL2Sync() helper, no new dependencies.
• Better error messages: Unwrapping error.cause from undici's generic TypeError: fetch failed surfaces the actual network error (ECONNREFUSED, EHOSTUNREACH, etc.) in both the Ollama stream path and provider autodiscovery. Users will actually know what went wrong now.
• Test coverage: New test case validates the WSL2 dispatcher override behavior.

Suggestions (non-blocking)

1. Comment accuracy on autoSelectFamily: The inline comment says disabling autoSelectFamily "forces IPv4," but technically it just skips the Happy Eyeballs algorithm and uses whichever address DNS returns first. In practice this works because Ollama binds to 127.0.0.1, but the comment is slightly misleading. Consider rewording to something like "disables dual-stack auto-selection to avoid WSL2 IPv6 routing issues."

2. Test mock cleanup: The isWSL2Sync mock is set to return true in the WSL2 test block, but vi.clearAllMocks() in beforeEach clears call counts without resetting mockReturnValue implementations. Consider adding vi.restoreAllMocks() or scoping the mock with vi.spyOn to prevent leaking into subsequent tests.

3. Error-cause unwrapping is untested: The .cause surfacing in ollama-stream.ts and models-config.providers.discovery.ts has no corresponding test cases. A quick test that throws a TypeError wrapping an ECONNREFUSED cause and asserts the unwrapped message appears in the error output would lock this behavior in.

4. Pre-existing: isWSL2Sync() does uncached sync I/O: Reads /proc/version synchronously on every call. This is not introduced by this PR and only runs once during dispatcher init, so not blocking. Worth noting for a future follow-up if the function gets called in hotter paths (it is also used in pi-embedded-runner/run/attempt.ts).

Summary

All four items are minor hardening suggestions. The core fix is correct, minimal, and solves the reported issue in #54498. Good to merge.