Symptom

Five tests hit the 30s Bun global --timeout on Windows CI even though they mock Npm.install — no real package install, no real network.

All five are bare test(...), not it.live / testEffect — so the helper-level scaling proposed in #14 does NOT reach them.

Evidence (last 5 Windows CI runs on dev, 2026-04-16 → 2026-04-17)

Why this is weird

All 5 tests mock Npm.install — no real download, no child npm process. Ubuntu pass times are 11-341ms. Windows pass times are 18-29s.

That's a 100-3000× gap, and it's NOT Windows being slow at npm (npm isn't running). Candidates to investigate:

• Effect runtime setup cost (Instance.provide, Layer construction) on Windows event loop
• fs.mkdir / Filesystem.exists call latency on GitHub Actions Windows runner disk
• Child process spawn overhead from unrelated setup (tool loading, plugin resolution walking node_modules)
• Mock signal gates (await gate, await ready, await waiting) deadlocking or yielding slowly under Windows scheduling
• Module resolution / first-import cost amortized onto the first test in the file

Root cause is unknown — needs local Windows reproduction or instrumented logging.

Treatment plan

Short term — stopgap PR (separate)

Add explicit { timeout: 60_000 } to each of the 5 tests with a // FIXME(#16) comment linking here. This unblocks CI while root cause is investigated.

Deliberately NOT raising the Bun global --timeout — we want future unrelated hangs to still fail fast at 30s.

Long term — this issue

1. Reproduce at least one of the 5 hangs locally on Windows (or add DEBUG-style timing logs and observe on CI).
2. Instrument timing around suspect areas: Instance.provide boundary, Config.get, Config.waitForDependencies, module import tree inside the mocked install callback.
3. Identify what actually consumes 18-29s.
4. Fix the root cause; remove stopgap explicit timeouts.

Related

• Memory rule: "mock 掉 I/O 的测试不算数" — mocked tests shouldn't flake for I/O reasons. This bucket violates that rule on Windows.
• PR ci(windows): DISABLE_FILEWATCHER env (match upstream) #8 added OPENCODE_EXPERIMENTAL_DISABLE_FILEWATCHER=true for Windows; this bucket persists past that fix.
• Bucket A (shell-loop it.live 3s ceiling): chore(ci): remove dead upstream publish.yml workflow #14 — different root cause (real-clock runner slowness), different fix.
• Assertion / race flakes: flaky(windows): compaction abort does not stop promptly during retry backoff #10 flaky(windows): pty output isolation misses initial socket data #11 flaky(windows): nested directory snapshot revert assertion fails #12 flaky(windows): session diff stays empty after tool execution #13 — unrelated; don't conflate.