"Agent Half-Dead" Cluster Breakthrough Entry Point + 48-Hour Active Surface A–N Classification Master Ledger (2026-10-29 | 『代理半死不活』簇攻坚入口 + 48小时活跃面 A–N 归类总账(2026-10-29~31) #5297

Activity

1. ranxianglei

   added

   p1

   track

   scope/wiring

   on Oct 31, 2026
2. ranxianglei commented on Oct 31, 2026

   

   ranxianglei

   on Oct 31, 2026

   OwnerAuthor

   More actions

   (ai: sglang-gpu0/qwen3.8-27b)

   第一刀:「代理半死不活」簇全拆解(按机制分四子簇)

   总判断:该簇在 10-29~30 已经过一轮集中修复——「杀健康进程/死端口占坑」族大部分已合 master;当前开放核心 4 项,其中 2 项卡在等 owner 拍板,2 项有 PR 在途。另发现一个系统性缺口:「静默直连」有四个独立触发源、没有统一出口。以下 file:line 均对当前 master 源码核实。

   A. 误死/杀健康进程(基本闭环)

   症状	状态
   parent watchdog 误杀健康 proxy 后 spawner 永久放弃(#4994)	closed;修复合入:#5085(kill 前要求连续确认)+ #5153(活性窗口按自身事件循环延迟加宽),10-29 均 merged
   watchdog 用自己事件循环上的旧心跳读数,一次同步卡顿自杀(#5152)	closed;同上两个 PR
   spawn 超时误杀健康 child	closed;#5010(adopt proven-healthy child)merged
   crash-on-startup 只留死端口	closed;#5097(launcher 尾采样)merged
   wedged proxy 永久占口不自杀(#4762)	closed;#4784(self-exit)merged

   残留:#5004(PR,14 评论,twin-collision:端口持有者是静默 twin 时应先于 persistence init 退出)仍 open、评审停滞——本簇最大在途 PR;同族 #4977(恢复竞态→全体直连撞 1M 上限)仍 open。

   B. 静默直连不压缩(核心痛点;主症状已随 v0.1.209 发,残留待确认)

   用户观感 = 「没有任何提示,压缩消失,上游超限 400,随后被迫换血丢上下文」。触发源至少四个、各自独立:

   1. respawn gave up 后不再重试 → dsh-native never reconnects after respawn gave up (healthy proxy listening) → unattributed over-window body refused locally (422) → session unrecoverable | [dsh-native] respawn 放弃后客户端永不重连（代理已健康监听）→ 未归属超窗 body 被本地 422 拒绝、会话不可恢复 #4755 + dsh lane: BUSY-drop → EADDRINUSE detour → adopt-twin race is silently lost ~half the time, and a one-shot bootstrap failure then becomes permanent #4811 的修复已随 v0.1.209 发布(commit 6483bbfc2 / c11d65d9e:give-up 后恢复驱动不再被 toolsReady 残留自我清除);
   2. 崩溃重启窗口 handshake 永不匹配(instance-file/registry 状态目录不一致)→ Bug: Within the crash-restart window, the proxy starts up successfully but the handshake never matches — model traffic s | Bug: 崩溃重启窗口内 proxy 拉起成功但 handshake 永不匹配，模型流量静默直连（不压缩）且不再重试 #5226 主症状同样被 v0.1.209 覆盖,现 waiting-feedback:等报告人在 0.1.209 复测;残留 facet 归 [Bug][Codex] Temporary separator selected as compression endpoint, causing acp_status to silently fail to report ~180K t | [Bug][Codex] 临时分隔符被选作压缩端点，导致 acp_status 静默漏报约 18 万 token 可压缩历史 #5139;
   3. 旧实例占端口 → bootstrap 连续失败 → 长期直连 → [Bug] Plugin bootstrap spawns a new instance while the old one still holds the port and drops proxying (goes direct): la | [Bug] 插件 bootstrap 在旧实例仍占端口时新起实例并放弃代理（goes direct）：大会话随即以未压缩历史撞上游上限（400 Range of input length / 413） #4125(GitHub 状态仍 open;线程内 10-24 评论称修复 PR [PR] merge: tier-2 consolidation train — 15 PRs + repair of #4525's sidScope compile break #4551 已合并并「补关」,与实际状态不符,需核实是否残留复现);
   4. 隧道准入卡 unresolvable → 见 C-1。

   系统性缺口(建议先设计、不动代码):上述四个触发各自有自己的日志行和各自的恢复路径,没有统一的「我现在正在直连/未压缩」可观测状态机,用户侧提示也各说各话。建议对齐面板 banner + /__bili/status attribution(#3581 起的既有面)与 host contract v2 讨论(#5241),单开设计 issue 收敛,而不是再打第四根补丁。

   C. Wedge/挂住(两项都卡决策单,拍了板就有代码可合)

   1. Long-lived proxy wedges on local DNS after WSL network/VPN changes — tunnel admission stays unresolvable (502) until process restart #5221 WSL 网络/VPN 切换后本地 DNS 永久 wedge:src/tunnel-guard.ts:105-119 的 dnsResolveHost(node:dns/promises lookup,2s × 3 次跑满即判 unresolvable)→ 此后每个模型请求 502 tunnel_destination_denied,直到重启 bili 进程;实例注册表会重新收养同一个楔死进程,所以只重启 DSH 宿主无效。Phase 1 PR [PR] fix+docs: make the long-lived-proxy stale-DNS wedge diagnosable and document its remedies (#5224) #5263 已交(错误信息补绕行提示 + README en/zh);Phase 2 自动自检重启(复用 src/restart.ts performSelfRestart)等 A1/A2/A3 拍板,同一 PR 内补齐。
   2. [Bug] Stalled upstream before first response byte holds the client silently up to 12 min and burns host retries — first-byte watchdog is warn-only and deduped per origin #5164 上游首字节前 stall → bili 静默持有客户端最长 12 min:UPSTREAM_TIMEOUT_MS = 12*60*1000(src/fetch-util.ts:24,可经 network.upstreamTimeoutMs 覆盖);dsh 宿主 ~300s 总是先放弃,每次重试全额重付 input tokens,日志只有一条每-origin 终身一次的 warn。止血 PR [PR] fix: first-byte stall warn per (origin,session) with cooldown instead of once-per-origin-per-process (#5164) #5260(warn 改每 (origin,session) 60s 冷却 + 带实际秒数)在 CI;决策单 [PR] ci-fixes convergence: port the four remaining master-red lanes (win32 db path, teardown handles, #5155 settle, #4788 timer host) #5262 等拍板:A=fail-fast+可配置阈值(推荐,默认 120s)/ B=仅诊断 / C=维持现状。
   3. 关联未定位件:[Bug][Codex] Temporary separator selected as compression endpoint, causing acp_status to silently fail to report ~180K t | [Bug][Codex] 临时分隔符被选作压缩端点，导致 acp_status 静默漏报约 18 万 token 可压缩历史 #5139 mega 会话事件循环卡死 ~19min、卡住的 child 占端口导致所有重拉失败(B-2 的残留载体)。

   D. 起不来(Windows EPERM 家族:根因已钉死,修复未落地)

   • 根因([Bug][dsh] EPERM on startup data writes: the real switch is the node runtime's NTFS mandatory integrity label (Low IL), not Defender/ACL — reproducible A/B, and it explains the #5096 crash loop #5154,本机 A/B 实测可重放):不是杀软/ACL,是 node.exe 映像自身的 NTFS 强制完整性标签 = Low → 所有派生 node 进程都是 Low IL,Low 写不了 Medium 对象 → ensureRootCA → writeCombinedBundle open 直接 EPERM(errno -4048)→ [Bug][dsh] Proxy crashes on startup (denied write access to its own data file) — web page inaccessible, root cause burie | [Bug][dsh] proxy 启动即崩溃(写自己数据文件被拒)时 Web 页面打不开、原因埋在 %TEMP%:应在端口上显示启动失败原因并提示权限问题 #5096 的 411 次崩溃循环、Power-Loss Restart Protection Fails to Auto-Recover billion-context Startup Failure on DSH — Troubleshooting Report | 断电重启保护无法自动恢复 billion-context 在 DSH 上启动失败 — 排查报告 #5104 的 400+ 次重试全直连。开关就一条命令:icacls "<node.exe>" /setintegritylevel Low|Medium。
   • 修复 PR [PR] fix: rank the node runtime's Low integrity label first in win32 EPERM hints; add non-blocking startup self-check (#5155) #5160(win32 EPERM 提示把 node runtime Low integrity label 排第一 + 非阻塞启动自检)仍 open、落后 master——今天还有用户机器在盲踩这个坑。
   • spawner 60s 预算族:[PR] fix: defer large-db integrity quick_check past listen so big stores fit the 60 s spawn budget (#4609) #4619(大库 quick_check 推迟到 listen 之后)/ [PR] fix(boot): drop redundant synchronous corpus decode on pre-listen path (#4515) #4539(pre-listen 同步 corpus decode)closed 未合入,承接关系见状态板。

   邻接(双写者)

   • Auto-update on dsh leaves two proxy processes writing one session store: rotation waits mistake a draining old proxy for one that never exited #5167(dsh 自动更新后新旧两进程共写同一份 index.db;根因是排水中实例的存量 keep-alive 连接照常应答,旋转等待把它误判成「没退出」,探测点 src/launcher.ts:4720 rotateStalePredecessor)— 10-30 closed。
   • [PR] fix(storage): a live second writer no longer blocks startup — sharing a local-disk index.db is a supported topology (rev | fix(storage): 活着的第二写者不再阻止启动——共享一个本地盘 index.db 是受支持拓扑（撤回 #5213 的拒绝策略） #5294(今天新开):撤回 [PR] [storage] Degradation Semantics Bundle: Seven accepted fixes consolidated into one branch — a single merge suffices (inc | [storage] 降级语义全家桶：七个已验收修复合成一个分支，合并一次即可（含撤回损坏库自动改名） #5213 的「活着的第二写者阻止启动」拒绝策略——共享本地盘 index.db 定为受支持拓扑。注意:它直接改变整个 D 簇的行为前提(从「多写者=错误要挡」到「共享写=允许」),建议与 Auto-update on dsh leaves two proxy processes writing one session store: rotation waits mistake a draining old proxy for one that never exited #5167 的修复形态放在一起复核一遍。

   建议下手顺序(待讨论,不代拍板)

   1. 先关两张决策单(成本最低、收益最高):[PR] ci-fixes convergence: port the four remaining master-red lanes (win32 db path, teardown handles, #5155 settle, #4788 timer host) #5262 选 A+120s 与否;Long-lived proxy wedges on local DNS after WSL network/VPN changes — tunnel admission stays unresolvable (502) until process restart #5221 Phase 2 的 A1/A2/A3 → 合计 ≤2 个 PR 可落
   2. [PR] fix(#4977): exit twin-collision before persistence init when the port holder is a silent same-build twin #5004 twin-collision 推完评审落地
   3. [PR] fix: rank the node runtime's Low integrity label first in win32 EPERM hints; add non-blocking startup self-check (#5155) #5160 rebase 落后落后 master 后合并(补 Windows EPERM 提示盲区)
   4. [Bug][Codex] Temporary separator selected as compression endpoint, causing acp_status to silently fail to report ~180K t | [Bug][Codex] 临时分隔符被选作压缩端点，导致 acp_status 静默漏报约 18 万 token 可压缩历史 #5139/Bug: Within the crash-restart window, the proxy starts up successfully but the handshake never matches — model traffic s | Bug: 崩溃重启窗口内 proxy 拉起成功但 handshake 永不匹配，模型流量静默直连（不压缩）且不再重试 #5226 残留深挖(等报告人 0.1.209 复测结论)
   5. 「静默直连统一出口」设计 issue(先讨论方向,不先写码)
3. ranxianglei

   changed the title [-]『代理半死不活』簇攻坚入口 + 48小时活跃面 A–N 归类总账(2026-10-29~31)[/-] [+]"Agent Half-Dead" Cluster Breakthrough Entry Point + 48-Hour Active Surface A–N Classification Master Ledger (2026-10-29 | 『代理半死不活』簇攻坚入口 + 48小时活跃面 A–N 归类总账(2026-10-29~31)[/+] on Oct 31, 2026
4. ranxianglei commented on Oct 31, 2026

   

   ranxianglei

   on Oct 31, 2026

   OwnerAuthor

   More actions

   🤖 Powered by ework · qwen3.8-27b

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

   查重: 不重复 —— 本单是 owner 开的簇攻坚入口;按 "half-dead / cluster / A–N ledger / track" 检索 open+closed,无同范围跟踪条目,D 簇各成员(#4994/#5152/#5164/#5221/#5222/#5139/#5154/#4125…)均为独立工作单,本单只做入口+顺序。

   标签: bug + p1 + track + scope/wiring(可用性缺陷簇、p1 密度最高、入口型跟踪、主战场在进程生命周期接线)——已按 OP 的标签说明校正:去掉误挂的 dsh(成员跨 dsh/pi/codex 多车道,总账规则跨车道不打)、p3→p1(跟最重成员对齐)。

   接单后对 GitHub 实况核了一遍(截至 10-31 ~08:00Z),floor 1 快照之后已有 5 处状态变化:

   项	快照时	现在(GitHub 核实)
   #5004 twin-collision(攻击序 2「推完评审落地」)	open、评审停滞	已 merged(10-31 06:07Z)→ 攻击序第 2 项完成
   #5263 WSL DNS Phase 1(诊断+文档)	已交待 CI	已 merged(10-31 07:19Z);且 07:57Z 刚出现 #5326 设计单(真·不重启自愈:fresh-resolver over admission AND egress)+ #5327 PR(wedge 告警上面板,即 Phase 2 的可观测半)——另一会话正在推进,我不抢线
   #5294 第二写者放行启动(邻接双写者)	今天新开	已 merged(10-31 03:36Z);#5213 降级语义 bundle 也在 00:17Z 合入 → 邻接双写者线整体闭环,「共享本地盘 index.db = 受支持拓扑」已是 master 现状
   #5226 handshake 永不匹配	waiting-feedback	已 closed
   #5231 失败压缩回执(同窗口 p1,非本簇但同批)	—	已 merged(05:26Z),follow-up 测试钉 #5310 也已在 07:59Z 合入

   仍开放的决策单(等 owner 拍板,我不代拍):

   1. 首字节 stall fail-fast(止血 PR [PR] fix: first-byte stall warn per (origin,session) with cooldown instead of once-per-origin-per-process (#5164) #5260 CI 已 30 项 0 失败/仅剩 4 pending):A/B/C + 默认值,推荐 A+A-1(默认 120s),拍板后我在 [PR] fix: first-byte stall warn per (origin,session) with cooldown instead of once-per-origin-per-process (#5164) #5260 之上叠加第二个 PR;
   2. WSL DNS Phase 2:原 A1/A2/A3(自检重启)已被 [PR] no-engine refusal: one line — requires Node.js 22+ #5326 的设计方向(fresh-resolver 自愈,不重启)部分重构——建议拍板时把两张单并着看,[PR] no-engine refusal: one line — requires Node.js 22+ #5326 若成立则重启方案降级为兜底。

   本会话开工的非决策依赖项(按「无决策依赖先修」纪律,不等上面两项):

   • [PR] fix: rank the node runtime's Low integrity label first in win32 EPERM hints; add non-blocking startup self-check (#5155) #5160 Windows EPERM 提示排序+启动自检:仍 open 且落后 master——rebase 到当前 master、推上去、看 CI(今天还有用户机器在盲踩这个坑);
   • [Bug] Plugin bootstrap spawns a new instance while the old one still holds the port and drops proxying (goes direct): la | [Bug] 插件 bootstrap 在旧实例仍占端口时新起实例并放弃代理（goes direct）：大会话随即以未压缩历史撞上游上限（400 Range of input length / 413） #4125 状态矛盾核实:快照称线程内 10-24 评论说修复 PR [PR] merge: tier-2 consolidation train — 15 PRs + repair of #4525's sidScope compile break #4551 已合并「补关」但 issue 仍 open;origin 上有 fix/4125-upward-adoption 分支——核实是残留复现还是漏开的 PR,该关线索给关线索、该开 PR 开 PR;
   • 「静默直连统一出口」设计单:先出站查重,干净才立(带来源标记),只讨论方向不写码。

   进展会持续在本线程更新。