[ranxianglei]

Fixes #1797

Problem

test (windows-latest, 22/24) intermittently failed on exactly one test — #983 maybeRetry self-heals a base-less register after a failed respawn (tests/dsh-native.test.ts:1010) — in the two reported forms (proxy is down at src/agent/dsh-native.ts:348; timeout waiting for plugin headers stamped), reproducing on pristine master with ubuntu always green.

Root cause (traced + CI-log forensics)

One mechanism, two timing failures — fire-and-forget attach/recovery chains outliving the test that started them:

1. Cross-test clobber (the issue's forms A/B). Several suites call apply() and tear down their mock proxy without letting the attach chain settle (worst case: the dsh.bundle.patch.yml 禁用自动压缩在 web profile 下静默失效 #1772 web-profile test calls apply() four times against one proxy). On loaded runners the in-flight probe loses the race with proxy.close() → probe DEAD → 30ms evidence grace → the fallback branch evaluates _spawnForTest ?? bootstrap with no stub installed → a REAL bootstrap() child. Its late tail then lands mid-window of a later test and unconditionally clobbers the shared module-global register.base/toolsReady. On CI the children exit instantly code 1 (ci.yml runs npm test BEFORE npm run build, so no dist/ exists) — accounting for every proxy bootstrap failed … (code 1) line in the windows logs.
2. Fixed-sleep race (a third form, same family). The self-heal test assumed the whole initial chain (dead-port probe + 30ms grace + failed spawn) settles within a fixed 50ms sleep. If the probe is slower, the first headersFor takes the registration branch against the still-set dead preset, fetchManifest fails, and parks retryAt at +10s — beyond the 5s wait cap — with nothing re-triggering during the wait → timeout waiting for self-healed tool registration. Reproduced deterministically locally by injecting ≥30ms delay into the :1 probe.

Fix (at the decision point, not wider sleeps)

• src/agent/dsh-native.ts: track in-flight attach/recovery flights (pendingChains + trackChain at the two state.ready = start() sites in apply() and the maybeRetry respawn branch); expose _settleNativeForTest() awaiting them plus one macrotask flush (so .then tails have run). Additive, behavior-neutral outside tests.
• tests/dsh-native.test.ts:
  • afterEach(() => _settleNativeForTest()) — universal backstop: any escaped chain settles at a safe inter-test boundary where the next test's _resetRegisterForTest neutralizes it.
  • command-behavior suites (所有的客户端支持acp-cache命令,在native模式和bili启动器模式下 #1146/[Bug] 插件模式下 /acp 显示的是「最近活跃会话」的状态面板，而非当前会话（/acp-cache 同） #1677 ×4) settle BEFORE closing their mock proxies, removing the fallback-to-real-bootstrap window (also stops orphaned bili children during local dev where dist/ exists).
  • the [dsh] native proxyEnvOrigin() 把临时 origin 冻结进 BILLION_CONTEXT_PROXY，绕过 ensureProxyRunning() 的 instance 验活；工具调用全部 fetch failed，需重启宿主 #983 self-heal test waits for initial-chain settlement event-driven instead of sleeping 50ms.
  • NEW regression test pins the slow-dead-port-probe shape: wraps globalThis.fetch in-test adding 80ms to http://127.0.0.1:1 requests and runs the identical sequence — failed deterministically pre-fix at ≥30ms injected delay.

Behavior change

None user-facing. The product-file diff is additive only (a Set + one exported test hook); wire shapes, both compression modes, defaults, thresholds, and persistence are untouched. Test-side changes only alter when async chains are awaited between tests.

Config surface

N/A — no config fields, env vars, or CLI flags added, renamed, or re-semantics'd.

Verification

• npm run typecheck, npm run build — green
• npm test full suite — 3470 tests / 3451 pass / 0 fail / 19 skipped (e2e-gated)
• slow-loopback sweep on tests/dsh-native.test.ts (SLOW_MS ∈ {30,50,100,150} + jitter, ×2 each): 8/8 green — pre-fix the suite failed from 30ms up
• standalone repro (exact test sequence, injected :1 probe delay 0/30/60/100/200ms ×2): 10/10 pass — pre-fix failed from 30ms up
• duplicate screening: open+closed issues/PRs searched for flaky/windows/maybeRetry/dsh-native variants — no prior report

Follow-up (disclosed in the #1797 thread)

The investigation exposed a latent product-side race of the same family — verifyAttachAndRecover's fallback landed.then/onGiveUp clobber register.base/toolsReady unconditionally, so a stale older chain can destroy a newer healthy base (reachable in prod only if a profile reload re-applies while a dead-preset bootstrap is still in flight; self-heals via maybeRetry but with a direct/uncompressed window). Suggested fix: generation counter — only the newest chain may clobber. Left to owner whether to file it separately (source chain already one level deep).

[github-actions]

📦 Built Package Artifact

Branch: 2026-10-01_dsh-native-flake-isolation (2566b36)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.1811-pr.1799.cda6f92

This exact version is immutable and reproducible. The rolling pr tag tracks the newest test build across all PRs.

Option B — Download artifact

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

tar xzf billion-context-pr1799.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 PR #1799 复核完成 —— 代码与 CI 已核对无误;但发现一处原始分析未覆盖的关键协调问题,需你拍板后再合。

---

⚠️ 协调警报(最高优先,先读这个)

本 PR(#1799,源出 #1797)与另外两个同窗口、同源缺陷的 open PR 高度重叠。三者都改 tests/dsh-native.test.ts;其中 #1787 与本 PR 同改 src/agent/dsh-native.ts:

PR	创建时间	来源	性质	改动面
#1787	09-30 16:38	#1783 死锁 + 本族竞态	产品侧:新增 landingOwnsRegister() 所有权守卫——过期的 verifyAttachAndRecover 落地链只在仍「拥有」register 时才写(base undefined 或 === 陈旧 origin);并把 retryAt 退避限定到实际失败的 base(修 windows-22 死锁)	dsh-native.ts + 测试
#1798	09-30 19:03	#1791	纯测试:用既有 _stateToolsReadyForTest() gate settle + #1772 加 hermetic spawn-stub seam + #983c 固定 sleep(50) 换 spawn 计数 waitFor	仅 dsh-native.test.ts
#1799	09-30 19:58	#1797	产品+测试(本 PR):pendingChains+trackChain+_settleNativeForTest(),afterEach 统一兜底,5 个 suite 先 settle 再关 proxy,#983 换事件驱动,新增 80ms 慢 probe 回归测试	dsh-native.ts + 测试

核心判定:#1798 与 #1799 是同一根因(跨测试污染共享 register + #983 固定 sleep)的两种竞争修法,都编辑 dsh-native.test.ts 的同一片 finally 块和 #983 用例、且 settle 调用不同(_settleNativeForTest() vs _stateToolsReadyForTest())——二者不能同时干净合并。两个 session 各自做了查重却都报「无重复」,而实际上彼此都早于对方存在(#1798 19:03 < 本 PR 19:58)。这是「One issue = one scope」+ 热文件竞争的标准情形。

建议处置(请你拍板):

1. fix(dsh): stale verifyAttachAndRecover landings must never clobber the self-healed register (#1783) #1787 独立先行——它才是产品正确性修复(在源头让过期落地不再覆盖新 base,符合 root-cause-not-symptom;同时修掉 release v0.1.176 #1783 死锁)。它与本 PR 文本不相交、语义互补,可先合。
2. 测试侧二选一(fix(tests): settle dsh-native attach chains before teardown (#1791) #1798 / fix: contain dsh-native attach/recovery chains within their own test (#1797) #1799),别都留。 我倾向 fix: contain dsh-native attach/recovery chains within their own test (#1797) #1799:① afterEach 统一兜底比逐 suite 插桩更健壮(漏插桩的 suite 也能被安全边界接住);② 带针对形态 C(慢 probe 越过短窗口)的真回归测试(注入 80ms、行为断言、修复前 ≥30ms 即确定性失败);③ pendingChains 显式跟踪所有在飞链,而 state.toolsReady gate 每次 apply() 重建、只反映最新一条。其产物改动极小(Set + 一个 _ForTest 钩子,行为中性,符合既有约定)。
   • 若你坚持零产物面:取 fix(tests): settle dsh-native attach chains before teardown (#1791) #1798、关 fix: contain dsh-native attach/recovery chains within their own test (#1797) #1799,或把 fix: contain dsh-native attach/recovery chains within their own test (#1797) #1799 的两个增量(afterEach 兜底 + 慢 probe 回归测试)cherry-pick 进 fix(tests): settle dsh-native attach chains before teardown (#1791) #1798。
3. 注意:fix(dsh): stale verifyAttachAndRecover landings must never clobber the self-healed register (#1783) #1787 落地后,跨测试 clobber(形态 A)在生产与测试里都变无害(过期落地只能在仍拥有 register 时写);此时测试侧 settle 的价值收敛为 (a) 防本地 dev(dist 存在时)孤儿 bili 子进程,(b) 修 [dsh] native proxyEnvOrigin() 把临时 origin 冻结进 BILLION_CONTEXT_PROXY，绕过 ensureProxyRunning() 的 instance 验活；工具调用全部 fetch failed，需重启宿主 #983 固定 sleep 竞态(形态 B——fix(dsh): stale verifyAttachAndRecover landings must never clobber the self-healed register (#1783) #1787 不碰)。所以无论选哪个,仍需一个测试侧修复。

---

结构化七问(针对本 PR #1799)

1. 是什么问题:windows matrix 上恒为同一测试(#983 maybeRetry self-heals)间歇红,两形态(proxy is down @ dsh-native.ts:348 / timeout waiting for plugin headers stamped),pristine master 亦复现,常态红掩盖真实回归、卡合并判定。
2. 什么原因导致的:fire-and-forget 的 attach/恢复链活过启动它的测试。(a) 跨测试污染:多 suite 调 apply() 后不等链落定就关 mock proxy,慢 runner 上飞行中 probe 输给 close → fallback 求值 _spawnForTest ?? bootstrap 时无 stub → 真 bootstrap 子进程(CI 上 dist 不存在故 code 1 退出,对应日志全部 proxy bootstrap failed (code 1)),其慢尾无条件覆盖模块级共享 register.base/toolsReady;(b) 固定 sleep 竞态:[dsh] native proxyEnvOrigin() 把临时 origin 冻结进 BILLION_CONTEXT_PROXY，绕过 ensureProxyRunning() 的 instance 验活；工具调用全部 fetch failed，需重启宿主 #983 假设初始链 50ms 内落定,probe 一慢首个 headersFor 走注册分支撞死 preset,retryAt 钉到 +10s > 5s 上限 → 超时。ci.yml 顺序已核实(test L43 先于 build L46),支撑 code-1 取证。
3. 怎么修复的:决策点修而非放宽 sleep——产物侧 pendingChains+trackChain(apply 两处 state.ready=start() + maybeRetry respawn 分支)+ _settleNativeForTest()(等全链 + 一宏任务 flush);测试侧 afterEach 统一兜底、5 suite 先 settle 再关 proxy、[dsh] native proxyEnvOrigin() 把临时 origin 冻结进 BILLION_CONTEXT_PROXY，绕过 ensureProxyRunning() 的 instance 验活；工具调用全部 fetch failed，需重启宿主 #983 换事件驱动、新增 80ms 慢 probe 回归测试。
4. 是否有回归问题:无。产物 diff 纯增量(trackChain 返回同一 promise,state.ready 语义不变;Set 每条链 settle 自删,有界不泄漏)。高危模式逐项过:无线形/双压缩模式变更、无 fail-fast 变吞错、无阈值/时序/输出格式漂移、持久化不动。CI 全绿(含 windows 22/24、e2e、build-artifact、version-guard)。
5. 是否破坏缓存:不涉及——未触碰缓存/重试策略/游标/去重键;register.retryAt 语义原样(settle 只改变测试间 await 时机)。
6. 是否改了设计方向:架构/边界/数据流未转向。但必须并入上面的三 PR 协调决定——单独合本 PR 会与未决的 fix(tests): settle dsh-native attach chains before teardown (#1791) #1798 冲突。
7. 是否引入新配置:无新增 env/config/CLI flag/默认值。

---

判定:本 PR 代码健全、CI 全绿,可审;但因与 #1798 冲突,不应独立合并。请你先定 #1798/#1799 取舍(我倾向 #1799)、确认 #1787 先行;定下后我按结果收口(关冗余 PR 或 cherry-pick 增量)。

[ranxianglei]

解决冲突后重新审查一次

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 冲突已解决,重新审查完成。

处理过程

• 状态核对:fix(dsh): stale verifyAttachAndRecover landings must never clobber the self-healed register (#1783) #1787(产品侧 landingOwnsRegister 守卫)与 fix(tests): settle dsh-native attach chains before teardown (#1791) #1798(纯测试 settle)均已合入 master;fix: contain dsh-native attach/recovery chains within their own test (#1797) #1799 因此 mergeable=False(两者都改 tests/dsh-native.test.ts 同一片行)。
• rebase 到当前 master:src/agent/dsh-native.ts 与 fix(dsh): stale verifyAttachAndRecover landings must never clobber the self-healed register (#1783) #1787 干净自动合并(guard 与 fix: contain dsh-native attach/recovery chains within their own test (#1797) #1799 的 tracking 钩子零重叠、互补);冲突只落在 tests/dsh-native.test.ts(6 处)。
• 按 union-of-intent 收敛:保留 master(fix(tests): settle dsh-native attach chains before teardown (#1791) #1798)的全部既有内容(9 处 _stateToolsReadyForTest() 逐 suite settle、dsh.bundle.patch.yml 禁用自动压缩在 web profile 下静默失效 #1772 hermetic spawn-stub 接缝、两个 release v0.1.176 #1783 慢 probe 回归测试),丢弃 fix: contain dsh-native attach/recovery chains within their own test (#1797) #1799 里已被覆盖的部分(per-suite settle 改动 + 那个与 fix(tests): settle dsh-native attach chains before teardown (#1791) #1798 重叠的 80ms 慢 probe 测试),只留 fix: contain dsh-native attach/recovery chains within their own test (#1797) #1799 真正的独有增量。
• 最终 diff(单提交,基于当前 master e669287d):src/agent/dsh-native.ts +26/-3,tests/dsh-native.test.ts +8/-2。

#1799 现在只剩什么(相对 #1798+#1787)
一层全链边界的纵深防御:

• 产物侧:行为中性的 pendingChains Set + trackChain()(返回同一个 promise)+ _settleNativeForTest()(仅测试调用)。
• 测试侧:全局 afterEach(() => { void _settleNativeForTest(); }) —— 每个测试结束后 drain 所有在飞 attach/恢复链。
• 为什么仍有价值:fix(tests): settle dsh-native attach chains before teardown (#1791) #1798 用的是 _stateToolsReadyForTest(),它只 await 最近一次 apply() 的 gate;多 apply 的 suite 可能留下更早一条链在飞。这个 afterEach 补上「按测试边界保证无一链越界」这层 fix(tests): settle dsh-native attach chains before teardown (#1791) #1798 没给的保证,防止未来新增 suite 忘了插桩时复发轮换类污染。

---

结构化七问

1. 是什么问题:不是新 bug,是协调问题——fix(tests): settle dsh-native attach chains before teardown (#1791) #1798/fix(dsh): stale verifyAttachAndRecover landings must never clobber the self-healed register (#1783) #1787 合入后 fix: contain dsh-native attach/recovery chains within their own test (#1797) #1799 与 master 文本冲突(mergeable=False),需把它收敛为「相对已合入工作仍独有」的最小增量再复审。
2. 什么原因导致的:两个并行 session 对同一 flake 各出 PR;owner 先合了 fix(tests): settle dsh-native attach chains before teardown (#1791) #1798(纯测试)+fix(dsh): stale verifyAttachAndRecover landings must never clobber the self-healed register (#1783) #1787(产品守卫)。二者落地后,fix: contain dsh-native attach/recovery chains within their own test (#1797) #1799 的 per-suite settle 与慢 probe 测试已被覆盖,净增量只剩「全链 drain」一处。
3. 怎么修复的:rebase→union-of-intent 解冲突,保留 fix(tests): settle dsh-native attach chains before teardown (#1791) #1798 既有机制、丢弃 fix: contain dsh-native attach/recovery chains within their own test (#1797) #1799 冗余项,仅保留独有增量(全局 afterEach 全链 drain + 行为中性钩子)。
4. 是否有回归问题:无。产物 diff 纯增量(trackChain 返回同一 promise、Set 每条链 settle 自删有界、_settleNativeForTest 仅测试调用);afterEach 对已 settle 的 suite 是 no-op。高危模式逐项过:无线形/双压缩模式变更、无 fail-fast 变吞错、无阈值/时序/输出格式漂移、持久化不动。本地证据:typecheck✅ build✅ dsh-native 41/41×7(含负载循环)✅ 全量套件 3463 pass / 0 fail / 19 skip✅。
5. 是否破坏缓存:不涉及——未触碰缓存/重试策略/游标/去重键;register.retryAt 语义原样。
6. 是否改了设计方向:未转向。方向仍是「attach/恢复链不越过测试边界」;只是从 fix: contain dsh-native attach/recovery chains within their own test (#1797) #1799 的完整方案收敛为「在 fix(tests): settle dsh-native attach chains before teardown (#1791) #1798 逐 suite settle 之上补一层全链兜底」。
7. 是否引入新配置:无新增 env/config/CLI flag/默认值。

---

判定 & 需要你拍板的一点

• 冲突已解决,PR 现 mergeable=True;CI 在新 head 2566b36c 上重跑中(本地全绿,预期 CI 转绿后即可合)。
• 诚实披露:瘦身后 fix: contain dsh-native attach/recovery chains within their own test (#1797) #1799 的净增量已经很薄,而且它没有自带针对该 afterEach 的专属回归测试(其机制依赖 fix(tests): settle dsh-native attach chains before teardown (#1791) #1798 已有的链 settle 测试间接覆盖;要给「晚 spawn 的链也会被 drain」写确定性测试较难)。它的定位是防御性纵深,不是修某个具体复现。
• 所以两条路请你定:
  • (A) 认可这层纵深防御 → 等 CI 绿即可合;
  • (B) 认为 fix(tests): settle dsh-native attach chains before teardown (#1791) #1798+fix(dsh): stale verifyAttachAndRecover landings must never clobber the self-healed register (#1783) #1787 已充分 → 直接关 fix: contain dsh-native attach/recovery chains within their own test (#1797) #1799 视为被取代(我不反对,这样最干净)。

你选哪个我就照做(A 我等 CI,B 我关 PR 并在此说明)。