Skip to content

[Bug] 0.1.209 dsh/Windows: an orphaned proxy outlives its dead owner (33 min, ~51% CPU, write lock held) and is never reaped → the replacement drifts 18789→18790 while the caller stalls 826 s with 0 bytes #5449

Open

Description

@ssamurias

Summary

On the dsh lane (Windows), a bili instance whose owner process died was never collected. It outlived its owner for at least 6 minutes (33 min of runtime total), never answered health on any port, wrote no log lines after the stall began, and burned ~51% of one core the whole time. Because it still held the index.db write lock, the instance that came up after the DSH restart had to degrade to legacy-only, lost the preferred port and drifted 18789 → 18790 — while the dsh client stayed pinned to 18789. The end result for the user: every model request through this machine sat at 826 s with 0 bytes written, then died with kind=upstream-timeout — for every provider, because they all share this proxy.

This is the cost side of #5152/#4994: those fixes (correctly) stopped the watchdog from killing live proxies during a synchronous stall. The same conservatism means a genuinely dead owner is no longer collected either — and that orphan then blocks the replacement.

Environment

field value
bili version 0.1.209 (commit 39a0ada)
usage mode plugin native (dsh) — spawned by the host, lane dsh
client DeepSeek Harness desktop (dsh) 0.2.x, Windows 11 x64, node v26.7.0
upstream api.deepseek.com/anthropic/v1/messages (deepseek-official) + a second lane host, both through the same proxy

Note on timestamps: bili.log writes local time (UTC+8), not UTC. All times below are local UTC+8.

Timeline

18:18:55  instance A starts: pid 20724, parent 5268 (the DSH host)
18:38–47  log density collapses 17–71 lines/min → 1–7 lines/min (stall begins)
18:47:28  [dsh-client] manifest registration failed (timeout after 5000ms: http://127.0.0.1:18789/__bili/plugin/manifest) — retrying
18:49:16  the DSH host (pid 5268) restarts → A's owner is gone
18:49:35  instance B starts: pid 24076
18:49:48  [storage] unified sidecar failed to open …\sessions\index.db (Error: database is locked) — staying legacy-only
18:49:51  port 18789: registry twin 0f60d531 (pid 20724) no longer answers as itself — the socket moved on while it
          drained; treating the port as foreign-held instead of dying twin-collision (#4982)
18:49:51  port 18789 busy — retrying on port 18790 (#3801)
18:49:52  [instances] another bili instance is running (pid 20724, http://127.0.0.1:18789) on lane "dsh" — both
          processes will write the same sessions directory; stop one to avoid state pollution (#394)
18:49:52  [zone] lane "dsh" drifted 18789 → 18790 — the preferred port is still occupied by something else;
          anything pinned to 18789 must re-resolve
18:50:42  [session-961d047e…] client disconnected mid-stream after 826869ms (0 bytes written to client)
18:50:43  client aborted mid-stream: kind=upstream-timeout code=20 … hint=idle budget expired (headers/body)
18:52:10  [storage] second-process pid 20724 has exited — stale warning cleared   ← only because I killed it

At 18:52 I killed pid 20724 by hand (Stop-Process -Force). Measured just before that with Get-Process -Id 20724 | % TotalProcessorTime: 1019.65 s of CPU over 1985 s of wall clock ≈ 51% of one core, for a process that was serving nothing.

Part 1 — the owner-exit reaper never fired

/__bili/health on a healthy instance on this machine reports the mechanism exists:

"watchdog":{"armed":true,"parentPid":21680,"watchers":[21680]}

and the log has the matching line for a successful reaping elsewhere (parent-gone (pid 18660) received — flushing sessions…). So in 0.1.209 the path exists and usually works. In this case it did not: parent 5268 was gone at 18:49:16, and A was still alive (and burning CPU) at 18:52:10, with native-attach diagnosing it as instance is BUSY (event loop saturated), not dead (#3654/#4811) — i.e. everything that looked at it concluded "busy, leave it alone".

Two candidate mechanisms, neither verifiable from dist/ (I can't read the source on the user's machine):

  1. The widened grace postpones itself. [Bug] Watchdog takes its stale-heartbeat reading on its own event loop, so a synchronous stall kills a live proxy (parent-gone → uncompressed 400) #5152 made effectiveHeartbeatGrace grow by the ticker's own drift, uncapped on purpose (correctly, per [Bug] 0.1.208 dsh/Windows: event loop fully blocked ~19 min after cold-load/eviction storm on 300MB+ sessions; wedged child holds port so all respawns fail (twin-collision code 3) #5138's 19-minute pause). But a proxy that is itself spinning at ~50% of a core has drifting ticks, so each tick can keep widening the window — the reaper postpones itself exactly in the situation where it is most needed. The worker-thread beat does not cover this case: the owner is dead, so there is nothing to attest liveness; only the kill-0 axis can decide, and that axis is consumed on the same loop.
  2. pid reuse on the kill-0 axis. If process.kill(parentPid, 0) is re-evaluated against a recycled pid, a dead owner looks alive forever. (pid 5268 no longer exists on the machine now, so I can't confirm either way.)

Why it matters beyond this one machine: an orphan on the preferred port is worse than no proxy at all. It holds the DB write lock (forcing the replacement into legacy-only), it forces a port drift (below), and it cannot be reaped by anything inside the system.

Suggested direction: resolve parent liveness from the worker (process.kill(parentPid, 0) off-loop, which cannot be starved), and treat "owner beat stopped and parent pid is not alive from the worker's view" as an unconditional reap, independent of effectiveHeartbeatGrace. A regression test that pins the proxy's loop with a spinning microtask chain and kills the owner would cover it — the existing tests only pin the loop with a live owner (#5152) or kill the owner on an idle loop.

Part 2 — the #4982 drift path strands the caller

#4982 stopped the replacement from dying on twin-collision, which is the right call — but the replacement then starts a second instance on 18790 that nobody talks to, while the caller stays pinned to 18789:

  • 7 × [dsh-client] manifest registration failed (timeout after 5000ms: http://127.0.0.1:18789/__bili/plugin/manifest) in 18:40–18:56, and the plugin surface stayed unavailable for the whole window;
  • the model request queued behind the dead port and produced exactly 0 bytes in 826.9 s before the client gave up.

The log line itself says "anything pinned to 18789 must re-resolve" — but nothing re-resolved. Either the drifted instance should publish the new port in a way the dsh client actually reads on failure, or the replacement should reclaim the preferred port once the holder is proven unreachable and its owner dead. As it stands, #5138's ending (dsh gives up, code 3) and this one (silent drift, stranded caller) are both terminal for the user, who only sees "it thinks forever and prints nothing".

Correction: I killed it, and then everything worked

before: GET http://127.0.0.1:18789/__bili/plugin/manifest → timeout after 5000ms (repeated)
after : Stop-Process -Id 20724 -Force
        GET http://127.0.0.1:18790/__bili/plugin/manifest → 200 in 6 ms, model requests flow again

Workaround running on this machine

A 5-minute schtasks job runs a small PowerShell guard that kills only processes matching all three: cmdline contains billion-context + dist/index.js; parent pid is dead; three consecutive health probes unanswered on ports 18789–18793. It never got to fire on pid 20724 because I killed it by hand first — so I can't yet claim it works, but it is the only reason this box is usable between reproductions.

Why this is not a duplicate

中文摘要

Windows + dsh 桌面版,bili 0.1.209。DSH 重启后(18:49:16),上一代的代理进程 pid 20724 没有退出:父进程已死、任何端口都不应答健康探测、33 分钟内烧掉 1019 秒 CPU(≈51% 单核),却仍占着 index.db 的写锁。于是新起的实例 24076 只能 legacy-only、拿不到 18789、漂移到 18790;而 dsh 客户端仍钉在 18789,模型请求 826 秒零字节后以 upstream-timeout 结束——所有 provider 都受影响,因为它们共用这一个代理。手动 Stop-Process 掉 20724 后,/__bili/plugin/manifest 从"5 秒超时"立刻变成 6 毫秒 200,请求恢复正常。

两条建议:① 父进程存活性判定放到 worker_threads 里做(不受事件循环拖累),并且「owner 心跳停了 + 父进程确实不在」应当无条件回收,不要再被 effectiveHeartbeatGrace 的放大所推迟;② 漂移到新端口后,要么让 dsh 客户端在连接失败时重新读 port-zone.json,要么允许新实例在确认旧持有点不可达且父进程已死之后夺回优选端口。

补充:这台机器上现在挂了一个 5 分钟的看门任务(三条同时满足才杀:命令行含 billion-context + 父进程已死 + 3 次健康探测无响应)。本次没机会触发,因为我是手动杀的。

Activity

  1. ranxianglei commented on Nov 1, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    (ai: vllm-qwen/qwen3.8-27b)

    查重: 不重复。 按失败签名(orphan / parent-gone / reaper / 端口漂移 / legacy-only / re-resolve)检索 open+closed:无同症状+同机制的条目。最近邻是「代理半死不活」簇 #5302(D 家族、wedge 后遗症),但簇内现有成员要么是误杀方向(#4994/#5152:杀活代理),要么是 owner 仍存活方向(#5138/#5139);本例相反——owner 彻底死亡 + 终结器自身失效,外加第二症状(端口漂移钉死调用方),簇内无成员覆盖。处理完会把它挂进 #5302。你对 #5014/#5359/#5324/#3597 的非重复论证成立;#5138 确实机制不同(那边 owner 还活着)。

    标签: bug dsh scope/wiring p1(孤儿代理持 DB 写锁 + 首选端口 → 全机模型流量中断数分钟,可用性缺陷,按口径 p1)。

    Triage 结论(版本优先,已按 tag v0.1.209 的源码核对)

    先说一个关键事实:你的构建(v0.1.209)不包含最近两周的三个看门狗修复——#4762 事件循环自退出看门狗(PR #4784)、#5152 worker 线程心跳 + grace 扩展(PR #5153)、#4994/#5085 杀确认 tick,全部首发于 v0.1.210。因此:

    1. 你的候选机制 ①(effectiveHeartbeatGrace 自扩展推迟收割)在你的构建里不存在——那段代码不在 v0.1.209 中。你这一版的收割逻辑是:in-loop 每 2s tick,kill-0 连续 3 次失败或心跳 >15s 陈旧 → 移除 → 集合空 → 5s → shutdown。两条轴都跑在主事件循环上——循环不让出,tick 就不发生。你观察到的 stall 期日志密度(1–7 行/分钟)说明循环只是偶尔让出,「检测到 owner 死亡 → 决定退出」仅凭让出频率就可能延迟数分钟——与时间线吻合(owner 18:49:16 死,18:52:10 手动杀,正落在这个窗口里)。还有一层:即便 shutdown 已决策,drain 阶段也会在那条 826s 挂死的连接保持打开期间继续持有进程(最多 10 轮)。
    2. 候选机制 ②(pid 复用)在你的构建里理论可能但概率低:v0.1.209 的 dsh spawn 路径已经注册 owner 心跳(spawner pid = 宿主进程自己的 pid,每隔几秒一跳),pid 即使被复用,心跳轴也能判 stale。但两轴同命:都依赖循环让出。

    当前 master(≥v0.1.210)对 Part 1 已基本覆盖,但有一个残余缺口:

    Part 2(端口漂移钉死调用方)在 master 上仍是真实缺口,已定位机制链:

    • 模型通道 URL 是 launcher overlay 烤好的 /bili/<origin> 形式,origin 嵌在字符串里,bili 运行时无法移动它;
    • 替代实例漂移到 18790 后,确实有 reroute 机制(native attach 模式无运行时 respawn:共享代理的属主退出后,attach 方全部模型请求永久硬失败 #1130 replacedOrigins),但触发条件是「请求实际失败过一次 AND 会话已武装 respawn」。你的事故有两条路把这条链打断:① 请求不快速失败——TCP connect 成功(accept 队列还在收)但零响应,挂到宿主侧超时(你的 826s);② attach 模式的会话可能从未武装过 respawn,于是即使失败,死 origin 也进不了 reroute 集,后续请求重复同样的挂死。终态:日志写着「anything pinned to 18789 must re-resolve」,但系统里没有执行者去做这个 re-resolve。
    • 修法有两个方向,都涉及行为/契约选择,我会另开一个 decision-request issue 列选项(A:客户端强制重解析——健康判死后无条件把烤好的 URL 改写到当前活 origin,接受跨实例压缩基底失同步的风险;B:替代实例回收首选端口——确认旧持有者不可达 + owner 死后绑回 18789,调用方不变;C:fail-fast + 显式错误面——不自动恢复,给用户明确的「代理孤儿了,请重启 dsh」而不是无限思考)。我的倾向是 B 为主 + A 兜底,等你拍板。

    行动计划

    1. 立即开修:Part 1 残余缺口(worker 侧父进程存活探测)+ 回归测试,push PR 引用本 issue;
    2. Part 2 开 decision-request issue(p1,选项 + 代价 + 推荐),挂 dsh desktop: newer-than-tested build warning never fires because version lookup cannot see packages inside app.asar | DSH 桌面版:「高于已测版本」告警永不触发(版本探测看不到 app.asar 内的宿主包) #5302;
    3. 进展在本线程更新。

    另外提醒:你的机器跑的是 v0.1.209——升级到 ≥v0.1.210 会直接消除 Part 1 的大部分症状(#4762 自退出 + 心跳确认);你的 schtasks 兜底脚本可保留作纵深防御,互不冲突。

  2. ranxianglei commented on Nov 6, 2026

    @ranxianglei
    Owner

    (ai: vllm-qwen/qwen3.8-27b)

    Facet checklist update from the OpenCode V2 window of #5823 (same family: Windows wedged proxy + failed recovery convergence):

  3. ranxianglei commented on Nov 6, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    (ai: vllm-qwen/qwen3.8-27b)

    收到清单,两点确认:

    1. F3 排序:我手上这条([Bug] 0.1.209 dsh/Windows: an orphaned proxy outlives its dead owner (33 min, ~51% CPU, write lock held) and is never reaped → the replacement drifts 18789→18790 while the caller stalls 826 s with 0 bytes #5449 Part 1)的 diff 严格加性——只在 worker 的 tick 逻辑里新增一条「父进程存活探测」轴(init/beat 消息各加字段、server.ts 调用点传 BILI_PARENT_PID + watcher 集大小),不碰 worker 早死/unarm 路径、不碰 /__bili/health 的 watchdog.armed 面、不碰 spawn-failure 降级。与 F3 的调查方向无冲突;若 F3 先动 event-loop-watchdog.ts,我的分支小面积 rebase 即可。
    2. Part 2 决策单定位:会开,但明确挂在 [Bug][Windows][OpenCode V2] Native proxy fails startup with SQLITE_BUSY; BUSY recovery reattaches to an unresponsive instance #5823 F4 旁边并说清机制分界——F4 = 「端口已释放但替代实例没拿到」(疑似 stale registry/zone);我这侧 = 「漂移发生后钉在旧 origin 的调用方无人重解析」(烤好的 /bili/<origin> URL + attach 会话未武装 respawn → 终态挂死)。两者叠加才是完整收敛:F4 修掉减少不必要的漂移,我这侧修掉消除漂移后调用方的永久失锚。另注:本事件里替代实例在孤儿还活着时就已 spawn(B 18:49:35,A 活到 18:52:10),所以即使 Part 1 快速收割,这个竞态下漂移仍会发生、调用方仍会失锚——两条都要修。

    当前状态:Part 1 实现已完成(worker 侧 off-loop 探测 + 3-tick 确认 + softWarn wedge 门 + watcher-count 守卫),正在本地装依赖跑定向测试,随后 push PR。

  4. ranxianglei commented on Nov 6, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    (ai: vllm-qwen/qwen3.8-27b)

    **Part 1(孤儿代理不被收割)已修复:PR https://github.com/ranxianglei/billion-context/pull/5833**(head 826251354,Closes #5449)。Part 2(端口漂移后钉死调用方)在 master 上确认同样存在,属行为取舍,已立决策单 **https://github.com/ranxianglei/billion-context/issues/5834**(p1)。

    结构化八问(针对 PR #5833):

    1. 是什么问题:dsh/Windows 下代理事件循环卡死期间 owner 进程死亡,孤儿代理长时间不被收割(事故中 33 分钟、~51% CPU),持续占着首选端口和 index.db 写锁,替代实例被迫漂移到 +1 端口,钉在旧端口的调用方挂死。
    2. 什么原因导致的:两条既有收割路径都依赖主循环让出——in-loop watcher 的 2s tick 只有循环 yield 时才跑;你的构建 v0.1.209 连 off-loop 看门狗都没有([bug] After event loop wedge, zombie still holds port 18789 — parent respawn loop spins on EADDRINUSE indefinitely; manu | [bug] 事件循环卡死后僵尸仍握 18789 端口,parent 重拉全部 EADDRINUSE 死循环,只能手工杀进程恢复(event loop wedge holds listen socket, respawn loop fails with EADDR… #4762/[Bug] Watchdog takes its stale-heartbeat reading on its own event loop, so a synchronous stall kills a live proxy (parent-gone → uncompressed 400) #5152/dsh lane: parent watchdog false-positives kill healthy proxy; spawner then gives up permanently #4994 全部首发于 v0.1.210,首条回复已说明你机制假设①引用的代码不在你的构建里)。即便在 master 上,「卡死 <5min + 稀疏 yield」这个中间区间也没有界:in-loop 路径的 detect→decide 取决于 yield 频率(分钟级),[bug] After event loop wedge, zombie still holds port 18789 — parent respawn loop spins on EADDRINUSE indefinitely; manu | [bug] 事件循环卡死后僵尸仍握 18789 端口,parent 重拉全部 EADDRINUSE 死循环,只能手工杀进程恢复(event loop wedge holds listen socket, respawn loop fails with EADDR… #4762 只覆盖连续卡死 ≥5min。pid 复用不是本案根因(你的构建已有心跳证明),但它与心跳轴同命——都被循环 yield 门控。
    3. 影响范围与严重程度:p1 可用性缺陷。owner 死亡 + 循环卡死稀疏 yield 时,全机模型流量中断数分钟至数十分钟(端口被无响应实例占着、DB 锁被持、替代实例接不到首选端口)。
    4. 怎么修复的:给既有 [bug] After event loop wedge, zombie still holds port 18789 — parent respawn loop spins on EADDRINUSE indefinitely; manu | [bug] 事件循环卡死后僵尸仍握 18789 端口,parent 重拉全部 EADDRINUSE 死循环,只能手工杀进程恢复(event loop wedge holds listen socket, respawn loop fails with EADDR… #4762 worker 线程加第二条独立轴——worker 在自己的线程里对 owner pid 做 kill-0 探测(EPERM=alive,[Bug] Windows: Proxy falsely detects parent-gone and self-terminates on every forwarded request (same live PID triggered | [Bug] Windows:代理每转发一次请求就误判 parent-gone 自杀(同一存活 PID 在 2.5h 内触发 12 次) #3890 纪律),仅当三条件同时成立才收割:连续 3 tick 判死 ∧ 主循环静默 ≥30s ∧ 无其他附加 owner(feat: bili CLI + debug in config + README rewrite #7 last-owner 语义,watcher 计数随每次 beat 更新而非冻结在初值)。循环静默门槛保证永不抢占优雅路径:还在 yield 的循环会先走到自己的 in-loop reaper(~11s),worker 最短 ~15s 才可能触发。pid 复用读为 alive → 可能漏收割但绝不会误杀。
    5. 是否有回归问题:逐项审过既有行为——正常 yield 会话不受影响(三条件合取恒不成立);手动 bili start(无 BILI_PARENT_PID)零变化(轴禁用);多宿主共享代理不早死(last-owner guard + 计数随 beat 不冻结);[bug] After event loop wedge, zombie still holds port 18789 — parent respawn loop spins on EADDRINUSE indefinitely; manu | [bug] 事件循环卡死后僵尸仍握 18789 端口,parent 重拉全部 EADDRINUSE 死循环,只能手工杀进程恢复(event loop wedge holds listen socket, respawn loop fails with EADDR… #4762 原 wedge 退出逻辑不动(trip 消息加 reason 字段是纯增量);drain/shutdown 路径不动。测试:新增 7 个测试钉住完整真值表(含负例:健康循环/活 owner/有其他 watcher/轴禁用均不收割),原 7 个 [bug] After event loop wedge, zombie still holds port 18789 — parent respawn loop spins on EADDRINUSE indefinitely; manu | [bug] 事件循环卡死后僵尸仍握 18789 端口,parent 重拉全部 EADDRINUSE 死循环,只能手工杀进程恢复(event loop wedge holds listen socket, respawn loop fails with EADDR… #4762 测试保持绿;本地 typecheck+build+定向套件全绿。
    6. 是否破坏缓存:不涉及(不触碰压缩/缓存/会话状态任何一面)。
    7. 是否改了设计方向:没有。这是沿既有 [bug] After event loop wedge, zombie still holds port 18789 — parent respawn loop spins on EADDRINUSE indefinitely; manu | [bug] 事件循环卡死后僵尸仍握 18789 端口,parent 重拉全部 EADDRINUSE 死循环,只能手工杀进程恢复(event loop wedge holds listen socket, respawn loop fails with EADDR… #4762 架构(同一 worker 线程、同一 trip/exit 机制)的扩展,不是新机制。
    8. 是否引入新配置:没有(无新 env/config/默认值;BILI_PARENT_PID 是既有字段)。

    行为变更披露(旧→新):旧——「卡死+稀疏 yield」下的孤儿代理只能靠 in-loop tick(分钟级)或 #4762 连续卡死 ≥5min 退出被收割;新——额外地,owner 判死 ×3 tick + 循环静默 ≥30s + 无其他 owner 时 ~15–45s 内退出。原因:补上「卡死 <5min + 稀疏 yield」中间区间,把孤儿持有时间从分钟级收敛到秒级。

    Part 2 机制链(master 已核实):dsh 插件车道的模型 URL 在会话启动时把 origin 烤进 /bili/<origin>/...,运行时不可移动;重路由只对 replacedOrigins 里的 origin 生效,而进入该集合需要真实发送失败 AND 已武装的 state.respawn;卡死的 holder 接受 TCP 但永不响应 → 没有快速失败 → 请求挂到宿主自己的 idle 预算耗尽(事故中的 826.9s);attach 模式会话的 respawn 未武装 → 即使硬失败也不重路由 → 永久搁浅。日志写着 "anything pinned to 18789 must re-resolve",但没有角色执行它。选项 A(客户端强制重解析)/B(替代实例在 holder 被证明不可达且 owner 死后回收首选端口)/C(快速失败+明确报错)及推荐(B 主 + A 兜底 + C 作报错底线)见 #5834;与 #5823 F4 的关系(分配侧 vs 寻址侧、两个面都要)也已在该单里写明。

    CI 状态:上一轮唯一红是 unused-exports 静态门(我导出的类型未被外部引用,已按门的规定改为模块内类型并本地验证门绿),核心矩阵当时是被该门下游 skip;新 head 上全套正在重跑,红了我会按 #4599 流程跟进。

    另外重申首条回复的版本结论:你这台机器升级到 ≥v0.1.210 本身就会消掉大部分 Part 1 症状(#4762 覆盖连续卡死、#5152 修心跳误杀);本 PR 补的是剩下的中间区间。

  5. ranxianglei commented on Nov 6, 2026

    @ranxianglei
    Owner

    🤖 Powered by ework · qwen3.8-27b

    (ai: vllm-qwen/qwen3.8-27b)

    更正一条(诚实披露):我上一条回复声称「孤儿持有时间收敛到 ~15–45s」对单 spawner 孤儿(正是本事故形状)不成立——ci-check 在分支上用复现抓到了这个缺陷并已修复:

    • 缺陷:last-owner 守卫采样的是 proxyWatchers.size,而该集合创建时就种入了 spawner pid、且只有 in-loop reaper 会移除它——而 in-loop reaper 恰恰被本轴要兜底的 wedge 饿死。于是单 spawner 孤儿的计数永远冻结在 1,off-loop 轴在事故形状下根本不会触发(复现:dead owner + wedged loop never tripped)。
    • 修复(分支 commit 39c740aad):计数排除被探测的 parentPid 本身——它的死活正是探测对象,且把它移出集合依赖的正是被兜底的那条路径;复用 pid 在该处仍读 alive,不引入误杀面。多宿主安全不变(其他附加 owner 仍在计数里,dead-streak 门独立生效)。
    • 回归钉:新增生产形状行「dead spawner 仍在 SET 中必须触发」+ last-owner 负例升级为同一 Set 派生接线;本地 15/15 绿。
    • 同批还有一处顺手小修(3b01a0802):tripwire 报错文案指向 scripts/run-ci-tests.mjs 但表早已移到 scripts/live-files.mjs(CI flake storm: concurrent fixed-port e2e files (port-zone-live / claude-native / auto-restart) collide on the 18787 zone under --test-concurrency 4 #3182),文案对齐。

    当前 head 39c740aad 全套矩阵正在跑(截至刚才 0 失败)。此前两轮红的分账不变:第一轮 unused-exports(已修)、第二轮全分片 tripwire(新测试文件未登记 LIVE_FILES,已登记并在 #4599 留了结论)。PR 仍是 https://github.com/ranxianglei/billion-context/pull/5833,决策单仍是 https://github.com/ranxianglei/billion-context/issues/5834。

  6. 8 remaining items

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions