[ranxianglei]

Addresses item ⑨ of the zero-inheritance-cloning family tracker #4266 — reported as a production incident on that issue.

What was wrong

createForkAdopter's capability gate (src/agent/fork-adopt.ts) treated any failed GET /__bili/plugin/manifest probe as terminal: on the first failure it called done.add(child) with no retry budget. A probe fails for two very different reasons:

1. the proxy genuinely lacks fork support (old build) — terminal is correct;
2. the proxy simply isn't reachable yet (host starting up / restarting / port changing) — transient, should retry.

The gate conflated the two. In case 2 the child was permanently retired from adoption for the process lifetime (the capability cache self-heals, but done.add had already pinned the child), then dead-ended downstream on the first-turn context_length_exceeded. The snapshot/fork degraded path immediately below already classified transient failures and retried within maxAttempts; the capability gate above it had no such classification.

Fix

New probeForkCapability() returns {capable, definitive}:

• definitive=true — manifest read OK but fork not advertised (old proxy): terminal on first probe, unchanged (keeps such proxies off the request path / avoids repeated 404 noise);
• definitive=false — manifest unreadable right now (!res.ok / timeout / network error): transient, deferred and retried up to maxAttempts, adopting once the proxy becomes reachable.

Only definitive results stay cached per base; a failed read is never cached, so the next request re-hits the wire and self-heals. manifestForkCapable remains a thin boolean wrapper so existing callers/tests are untouched. All three native lanes (dsh / opencode / pi) share createForkAdopter, so one fix covers all of them.

Behavior change (old → new)

A single transient capability-probe failure used to retire a child's adoption for the process lifetime; it now retries within the existing maxAttempts budget (default 3). Definitively-incapable proxies are unchanged (still terminal on the first probe). No config surface, wire-shape, persistence, message-identity, or kernel change.

Tests

• Rewrote the exploding-network case: it previously asserted "closes the window" on the first failed probe (pinning the bug); it now asserts deferred-until-cap, then closed.
• Added a regression repro: proxy down on request 1 (5xx manifest), up on request 2 → the child adopts (fork POST fires exactly once) instead of being permanently retired.
• 45/45 green across tests/{pi-omp,dsh,opencode}-fork-adopt.test.ts; typecheck clean.

Part of #4266 (item ⑨). Not closing the tracker — sibling items ①–⑧ are tracked there and land via their own PRs (#3024 / #3039 / #3951 / #4223 / #4263 plus pending owner decisions).

[github-actions]

📦 Built Package Artifact

Branch: 2026-10-23_fork-capability-probe-transient (ec954bf)

Latest build — install this exact version

npm install -g [email protected]

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

Recent builds (commit → version)

commit	version	build
ec954bf	[email protected]	build
5d7f4ca	[email protected]	build
4aedc2b	[email protected]	build

Download artifact

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

tar xzf billion-context-pr4345.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

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

查重: 不重复——本 PR 是 #4266 家族 item ⑨ 的修复(我自己开的);①–⑧ 由各自 PR 承载(#3024/#3039/#3951/#4223/#4263),无其他 open issue/PR 覆盖「capability probe 瞬态失败分类」这一机制。

标签: bug / dsh+pi+opencode / scope/wiring / p1(fork 继承被进程级永久放弃,属「会话数据丢失类」,按 2026-10-19 口径一律 p1)——已打上。

审核结论(八问)

1. 是什么问题:fork 采纳的 capability gate 把「暂时够不到 proxy」当成「proxy 不支持 fork」——第一次 GET /__bili/plugin/manifest 失败就 done.add(child) 永久退役该子会话的采纳资格(直到宿主进程重启),后续请求在首轮 context_length_exceeded 上死掉。
2. 什么原因导致的:probe 失败有两种语义被混为一谈——manifest 读到了但没有 fork(老 proxy,确定态)vs manifest 此刻读不到(宿主启动中/重启中/端口漂移,瞬态)。gate 对两者一刀切终局;而它下方的 snapshot/fork 降级路径本来就有这个分类和 maxAttempts 重试预算,gate 却独缺。
3. 影响范围与严重程度:dsh/opencode/pi 三条 native 车道共享 createForkAdopter,一处修复全覆盖。触发条件=子会话首个请求恰好落在 proxy 不可达窗口;后果=p1:子会话在该进程生命周期内永久失去父上下文继承,下游表现为窗口超限报错。
4. 怎么修复的:probe 结果拆成 {capable, definitive}——definitive=true(读到但无 fork)首次即终局,行为不变;definitive=false(读不到)判瞬态,延迟并在既有 maxAttempts(默认 3)预算内重试;只有 definitive 结果进 per-base 缓存,读失败永不缓存,下一个请求自然重探、自愈。manifestForkCapable 保留为 boolean 薄包装,直接调用方与测试零改动。
5. 是否有回归问题:逐项清点过既有行为——definitive-incapable 路径不变(首探终局+缓存);缓存语义不变(仅成功读取入缓存,与旧行为一致);attempt 预算与下方 tryForkAdoption 瞬态重试共用同一 attempts map 与 maybeAdopt 上限,无新预算;single-flight、done 集、google-wire skip、side-shaped body skip 全部未触碰;CHILD_CONFLICT 处理不变。无语义漂移、无 fail-fast 变吞错、无阈值/时序/输出格式变化。测试侧:原「爆炸网络」用例本来就把 bug 钉成了断言,已重写为 deferred-until-cap;新增回归复现(proxy 第 1 次 5xx、第 2 次可达 → fork POST 恰好发一次)。
6. 是否破坏缓存:不涉及——prefix cache/cache ref/重试游标/去重键路径未触碰;capability cache 仅 per-base 记忆化,语义保持。
7. 是否改了设计方向:不涉及——只在 createForkAdopter 内部给重试纪律补了一个分类维度,与下方既有瞬态重试同机制族。
8. 是否引入新配置:不涉及。

行为变更披露(old→new):单次瞬态 capability-probe 失败,旧=子会话采纳资格进程级永久退役 → 新=在 maxAttempts=3 预算内重试;确定不支持 fork 的 proxy 完全不变(仍首探终局)。PR 正文已载明。

本轮动作(已验证执行)

• rebase 到当前 master(PR 原落后 30 个提交,含 storage [PR] storage: durability trio — open-time quick_check + online VACUUM INTO backup + second-process lockfile (#4175) #4177);
• 修掉一个 CI 红:unused-exports gate 抓到本 PR 导出的 ForkCapabilityProbe 跨模块无人使用 → 去掉 export(独立标注 commit d7cf89b84,按 INTERNAL_ONLY 处置);
• 本地定向验证:typecheck clean;tests/{pi-omp,dsh,opencode}-fork-adopt.test.ts 45/45 绿;unused-exports gate clean;
• mysql-contract 红:master 同 lane 通过(fe5ba59/9284a8828),且本 PR 不碰任何 storage 代码 → 判定 runner flake,等新 head CI 复跑确认。

判定:代码与行为审核通过,无拦下项;新 head d7cf89b84 的全量 CI(test 矩阵 + mysql-contract 复跑)转绿后即可合并——合并操作按规矩留给你。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

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

更正上一条的 flake 判断：mysql-contract 不是 runner flake，是确定性缺陷——查证后收回。本 head d7cf89b84 的 CI 红全部继承自 master（aa198fe61）自身的红，与本 PR diff 无关。已核对到根因与在途修复：

• test 矩阵 6 腿全红：各 shard 死在测试开始前的绊线（错误签名逐字一致）——process-spawning test files missing from LIVE_FILES … tests/storage-durability.test.ts。[PR] storage: durability trio — open-time quick_check + online VACUUM INTO backup + second-process lockfile (#4175) #4177 新增该 spawn 测试时漏登记 scripts/live-files.mjs，tests/claude-native.test.ts linux live-E2E subtests fail/hang in shared multi-agent sandboxes #1248 绊线直接 die。→ 子缺陷 A，跟踪于 CI red on master since #4177: unregistered spawn test kills all test shards; mysql-contract lane fails #4326/Master CI is all red: tests/storage-durability.test.ts in #4177 is not registered in LIVE_FILES, triggering the process- | master CI 全红：#4177 的 tests/storage-durability.test.ts 未登记 LIVE_FILES，触发 #1248 进程生成绊线 #4333；在途修复 [PR] web: Rotate the "Compressed Structure" tree view of large sessions to vertical full size, no longer shrunk to sub-pixel | web: 大会话「压缩结构」树图旋转为竖向全尺寸，不再缩成亚像素 (#4173) #4174（分支内补登记，commit 133cdd247）与 [PR] ci: register tests/storage-durability.test.ts in LIVE_FILES (#4333) #4334（专用小 PR，同一行登记）——两者内容重复，合其一即可清 A，都领会撞车，请留意。
• mysql-contract 红：该 job 没有任何 kernel 构建步骤，而 [PR] storage: durability trio — open-time quick_check + online VACUUM INTO backup + second-process lockfile (#4175) #4177 把存储导入图延伸到 value-import acp-kernel 的 config.ts，kernel/dist 不入库 → 全新检出必然死于 import 解析（ERR_MODULE_NOT_FOUND … acp-kernel/dist/index.js），本地干净安装已复现。→ 子缺陷 B，跟踪于 CI red on master since #4177: unregistered spawn test kills all test shards; mysql-contract lane fails #4326；在途修复 [PR] fix(ci): build acp-kernel in the mysql-contract lane (sub-defect B of #4326) #4330（job 内加一步 npm --prefix kernel run build）。
• 三个修复 PR 均 open/mergeable/blocked（卡在同一批红检查上），等人工合并。

本 PR 自身状态：guards/static-gates/e2e 各车道全绿；typecheck、定向 45/45、unused-exports gate 本地+CI 双绿。代码面没有待办。

后续安排：master 转绿（A+B 任一组合合入）后，我立即把本分支 rebase 到新 master 并确认全量绿，再交合并。当前无需你操作本 PR。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

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

master 已转绿，本分支已按承诺 rebase 并推送（新 head 65f9db373，基于 72926f6fb）：

1. rebase 干净：两个原 commit 无冲突落地，diff 相对 master 仍只含 src/agent/fork-adopt.ts + tests/pi-omp-fork-adopt.test.ts。上游进展核对：子缺陷 A（LIVE_FILES）随 [PR] web: Rotate the "Compressed Structure" tree view of large sessions to vertical full size, no longer shrunk to sub-pixel | web: 大会话「压缩结构」树图旋转为竖向全尺寸，不再缩成亚像素 (#4173) #4174 合入修复（005e75d99+2d6debc18），子缺陷 B（mysql-contract 缺 kernel 构建步）由 master mysql-contract lane red: #4177's store→runtime-status chain pulled acp-kernel runtime imports into the storage co | master mysql-contract 车道红：#4177 的 store→runtime-status 链把 acp-kernel 运行时 import 带进了 storage 契约测试，job 没构建 kernel #4336 修复合入（fd9043780）；[PR] fix(ci): build acp-kernel in the mysql-contract lane (sub-defect B of #4326) #4330/[PR] ci: register tests/storage-durability.test.ts in LIVE_FILES (#4333) #4334 作为重复关闭——与我在场的判断一致。
2. 本地定向验证：fork-adopt 三套件 45/45 绿；unused-exports gate 绿。
3. 顺手小修（独立标注 commit 65f9db373，与本 PR 主题无关）：rebase 后发现 master 自身有一处 gate 红——[PR] fix(ci): run corrupt index.db degrade checks in a child process so windows cleanup can't hit EBUSY (#4338) #4378 合入的 tests/corrupt-index-child.ts 有 4 个名字（ENGPROOF_CHECK_NAMES/BANNER_CHECK_NAMES/CorruptIndexCheck/CorruptIndexVerdict）export 了但仅文件内引用，unused-exports lane 在任何 head 上都会红（含本 PR）。已去掉 4 个 export：先验证过文件外零引用（父测试只用 ENGPROOF_CHECKS/BANNER_CHECKS/spawnCorruptIndexChild），gate clean + 两个消费方测试文件 17/17 绿。若另有在途修复撞车，以先合入者为准、本 commit 可弃。
4. 待办：全量 typecheck + 回归交给 CI（本地不跑全量）。CI 全绿后我确认并在此回帖，即可交合并。