[ranxianglei]

Fixes #1208

What changed (test-only, zero src/ changes)

1. tests/logger-rotation.test.ts — settle() now requires a stable non-zero size before returning and throws on timeout instead of silently returning. Previously last initialized to -1, which is exactly what a missing file reads as, so if the stream's async open had not completed within ~60ms (plausible under windows-latest load), settle returned early while streamFd was still undefined. The clobber then raced the in-flight open: isOrphaned() short-circuits to "not orphaned" while an open is in flight, both post-clobber logs reused the pending stream, and the expected single [warn] never fired (warns.length 0 ≠ 1 — the exact PR fix: register watcher pid for every native client attaching to a shared proxy (#1190) #1200 windows-24 failure). With the precondition guaranteed (open completed + line flushed), the clobber/reopen path is deterministic on every OS: libuv opens files with FILE_SHARE_DELETE on Windows, so the clobber succeeds either way, statSync(logPath) then throws ENOENT, orphan detection fires, the reopen fails with ENOTDIR, and exactly one warn is emitted.

2. tests/session-gc.test.ts ("records rawInputTokens per turn and persists it ([Feature] Default state file cleanup strategy #1082)") — drains in-flight debounced saves with await store.flushAll([]) before the flushSync round-trip assertion (same pattern as Flaky: #1101 T7 cleanup races in-flight persist writes — ENOTEMPTY kills the release gate #1194). With debounceMs: 10, a turn-1 timer can fire between turns and start an async writeNow whose payload snapshot predates turn 2's rawInputTokens; its rename can land after the synchronous flushSync, leaving stale bytes on disk → recorded value survives persistence round-trip fails (the master@e8be19a1 windows-24 failure). After flushAll no competing writer remains.

3. tests/e2e-export-handoff.test.ts — same one-line drain before the teardown rmSync: this test drives real proxy traffic (markDirty → scheduleSave, debounceMs 0) over an enabled store and deleted the dir without draining, the identical race class as Flaky: #1101 T7 cleanup races in-flight persist writes — ENOTEMPTY kills the release gate #1194/T7. Hardened proactively so it cannot become the next random red.

Not changed (verified safe)

Audited every other enabled-store test for the same race class (debounced save in flight at teardown rmSync):

• fix-401-storage, persist-tail, cli-export, search-context-cross-session, store.test.ts envelope round-trip: use awaited writeNow or sync-only flushSync — no in-flight writer at teardown.
• issue1086 T3/T8: foreign sessions never create local state → scheduleSave never called.
• fix-config-session: sync eviction→flushSync only.
• proxy-persist, host-usage-postfold: teardowns either drain (flushAll) or cancel before any write can be in flight; the residual fire-and-forget case is covered by the kernel-side unhandled-rejection fix from release v0.1.42 #158.

The third flake reported in #1208 (ubuntu-22 master@8daba7fd) was identified via the CI job log as issue1086-chain-self-exemption.test.ts T7 (ENOTEMPTY rmdir at teardown) — already fixed on master by 3f61cc1 (#1194); the failing run predated that commit. No action needed.

Behavior changes

None in production code — all three edits are inside tests/. No config schema, wire shape, persistence format, or default changes.

Pre-flight

• npm run typecheck ✓
• npm test: 2377 pass / 3 fail — the 3 failures are in tests/claude-native.test.ts (live ps-walk e2e tests) and fail identically on the untouched tree in this sandbox (verified via stash); they are environment-specific here, not regressions. All 3 touched files pass 5/5 standalone repeated runs.
• npm run build ✓

Note: I did not add blanket CI retry-on-flake (issue suggestion 2) — with the root causes removed deterministically, retries would only mask future real regressions.

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-24_flaky-test-hardening (5b58d20)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-1224

Each push to this PR publishes a new version under the pr-1224 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf billion-context-pr1224.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

一句话总结：审核了 PR #1224（#1208 三个 CI flake 的测试加固）——三个根因均在代码层面独立验证成立、对其他 enabled-store 测试的审计结论复核无误、CI 四平台全绿、纯测试改动零生产行为变更，可以合并。

① 原 issue 方向：#1208 建议「逐一定位 + 确定性等待」，本 PR 对三个观察到的 flake 全部照此实施；明确不做建议 2（CI retry-on-flake）——理由成立：根因已确定性消除，retry 只会掩盖未来真实回归，建议 3（标记 known-flaky）随之不需要。第三个 flake（ubuntu-22 master@8daba7fd）确认为 issue1086-chain-self-exemption.test.ts T7 的 ENOTEMPTY rmdir，已由 3f61cc1 (#1194) 修复 —— 已验证该 commit 是本 PR head 的祖先（git merge-base --is-ancestor），无需动作。方向关通过。

② 回归审查（既有行为清点）：

• src/ 零改动 → 生产行为零变化：config schema、wire shape、persistence format、默认值均未触碰。行为变更声明：无（按交付规则显式声明）。
• settle() 超时语义从「静默继续」改为「throw」：严格改善的失败语义——文件 3s 内不达标就立刻失败并报明确错误，而不是继续跑出后续断言的误导性报错；不可能产生假绿。唯一理论风险是假红（窗口不足），CI 在两个 windows-latest runner 上实测全绿，说明负载下窗口足够。
• 两处新增 drain（flushAll）只改 teardown/读取前时序，不动任何断言逻辑。
• 无「修历史问题顺带改语义」类高危模式：改动全部限定在 test-local helper 与 teardown 块内。

③ 代码/机制审查（逐条读了相关代码路径，不只读 diff）：

• logger-rotation：旧 flake 机制确认——src/logger.ts:79 if (stream === s && streamFd === undefined) return false; // open still in flight。旧 settle 初值哨兵 -1 恰好等于「文件不存在」的读值，async open 未在 ~60ms 内完成时提前返回、streamFd 仍 undefined → isOrphaned() 短路为 not-orphaned → clobber 不被检测 → 0 个 warn，与 PR fix: register watcher pid for every native client attaching to a shared proxy (#1190) #1200 windows-24 失败签名（warns.length 0≠1）完全吻合。新前置条件（稳定非零大小 = open 完成 + 行已落盘）使后续路径确定：clobber → statSync(logPath) ENOENT → orphan 检测命中 → openStream() 的 mkdirSync(dirname) 因 dir 已被普通文件顶替抛 ENOTDIR → warnReopenFailed 恰好一次（第二次 log 被 reopenWarned flag 抑制）。机制成立。✓
• session-gc round-trip：race class 确认——debounceMs: 10 时 turn-1 timer 可在两轮之间触发、启动一个快照早于 turn-2 rawInputTokens 的异步 writeNow，其 rename 可落在同步 flushSync 之后留下旧字节。await store.flushAll([])（src/persist.ts:636）保证 drain 后无竞争写者。与 Flaky: #1101 T7 cleanup races in-flight persist writes — ENOTEMPTY kills the release gate #1194 同模式。✓
• e2e-export-handoff：new SessionStore({ dir, enabled: true, debounceMs: 0 })（line 38），真实代理流量 markDirty→scheduleSave，finally 原先未 drain 直接删目录——同一 race class，主动加固合理。✓
• 独立复核了全部 enabled-store 测试（不止信 PR 清单）：fix-401-storage / persist-tail / cli-export / search-context-cross-session / store.test.ts 均为 awaited writeNow 或 sync-only flushSync；issue404 / persist-encryption / persist-zstd / persist-rollback-guard / proxy-persist / host-usage-postfold / side-request-isolation 均在 rmSync 前 store.cancelAll()；fix-config-session 的 eviction 路径是 getStore().flushSync(s)（src/session.ts:660）同步执行；absorb.test.ts 用同步 flushSync/loadSync 且无 rmSync；store-marker-gate 不删目录。「Not changed (verified safe)」一节成立。✓

④ 测试有效性：三处都是真测试、断言有牙——warns.length===1 + stderr-only 降级 + 文件不复活；rawInputTokens 持久化 round-trip 相等性；export handoff 标记断言。把此前非确定性的行为钉成了确定性回归防护。说明：原 flake 是 Windows 负载下的时序竞争，Linux 本地无法复现原始失败；证据链 = 机制代码级验证 + 本 head 在两个 windows-latest runner 的全量 CI 实跑。kernel 侧相关缺陷已有 acp-kernel#373 单独跟踪，本 PR 用 drain 规避，跨仓部分不在本 PR 范围。

⑤ diff 整洁度：+11/−2、3 个文件、单个线性 commit 直接坐在当前 master head（aae4de1c）上，每行都与目的相关，注释统一带 #1208 出处。无 version bump、无 lockfile/生成物 churn。干净。

验证证据：

• GitHub API：PR open，mergeable_state: clean，base = 当前 master head。
• CI check-runs @ 5b58d20：8/8 success（test windows-latest 22/24、ubuntu-latest 22/24、version-guard、build-artifact、codex e2e ×2）。
• 本地（PR head）：typecheck ✓、build ✓、全量套件 2377 pass / 3 fail——3 个失败均为 claude-native.test.ts 的 live ps-walk e2e，本沙箱环境特有问题（四个 CI 平台上均通过，非回归）；三个被触碰文件连续 5 轮各 15/15 全过。

结论：可以合并。按惯例 Agent 不合并 PR，请人工操作：#1224