[github-actions]

Auto-fallback from the one-click release workflow: direct push to master was blocked by branch protection. Merging this PR publishes via the standard release.yml flow.

Changes since v0.1.201

• feat: wire google lane into snapshot render ([Bug] Snapshot rendering missing google lane: collapse overlay for google sessions still anchored to client bytes; id-ch | [Bug] 快照渲染缺 google 车道:google 会话的折叠覆盖仍锚定客户端字节,id-churn 家族(#3752/#3751)在该车道原样存在(#3759 跟进) #3764)
• build: stop emitting sourcemaps into dist (Remove sourcemaps during bundling to reduce plugin size. | 打包去掉sourcemap 减少插件体积 #3777)
• fix: egress-shed gate keys on the WS transport tag, not any managed transport ([Bug] Jimp is active, but image accumulation still triggers Responses WS 1009; duplicate request size not reduced (v0.1. | [Bug] Jimp 已生效，图片累积仍触发 Responses WS 1009；重复请求未缩小（v0.1.200） #3701 Finding 1)
• test: port-zone autoRestartOnUpdate + occupied target port -> orphan process leak: new instance silently hops to the next port, old instance never exits #3044 leg accepts the [Bug?] Under high load, health check timeouts appear to be misinterpreted as "proxy unavailable": large session requests | [Bug?] 高负载下 health 探活超时疑似被当成「代理不可用」：大会话请求被本地拒发（0.1.199 / plugin native） #3654 twin-adopt exit code 3
• style(kernel): keep openai.ts import block contiguous
• docs(kernel): image fingerprint comments hash the base64 payload, not decoded bytes
• snapshot-render harness: settle on the store tail view instead of a fixed 60ms nap
• Image message ids re-mint whenever an earlier image leaves the visible tail (membership-based ClusterCounter suffix) — prefix-cache breaks per fold, feeds #3748 storage growth #3751: payload-addressed image message ids (media fingerprint seed slot)
• docs: snapshot-render doctrine across AGENTS/MESSAGE-IDENTITY/SESSION-IDENTITY (Snapshot rendering: proxy-lane outbound is now rendered authoritatively server-side; existing sessions are immune to ker | 快照渲染:proxy 车道出站从服务端正史渲染,存量会话对内核升级免疫(#3752 根治) #3756)
• kernel: bump acp-kernel to 0.0.135 for the snapshot-render test seam (kernel-guard, [PR] Snapshot rendering: proxy lane outbound prefix rendered from server-side canonical history; existing sessions immune to | 快照渲染:proxy 车道出站前缀从服务端正史渲染,存量会话对内核升级免疫(#3756) #3759)
• fix: drop unused SnapshotRenderDeps export (unused-exports gate, [PR] Snapshot rendering: proxy lane outbound prefix rendered from server-side canonical history; existing sessions immune to | 快照渲染:proxy 车道出站前缀从服务端正史渲染,存量会话对内核升级免疫(#3756) #3759)
• snapshot render: server-authoritative outbound prefix for proxy lanes (Snapshot rendering: proxy-lane outbound is now rendered authoritatively server-side; existing sessions are immune to ker | 快照渲染:proxy 车道出站从服务端正史渲染,存量会话对内核升级免疫(#3752 根治) #3756)
• test: cover an advisory-adopted side request on the 标题生成等侧请求污染主会话状态:prepareAnthropic 缺 isTitleGen 守卫、快照无条件覆盖、survivedCount 每请求空转 #388 lane ([PR] fix: fork prefix gate is advisory — adopt drifted fork children with a classified warn (#3741) #3757 review)
• test: pin [PR] fix: classify FORK_PREFIX_CONFLICT rejections instead of failing blind (#3740) #3743 classifier branches orphaned by the adoption inversion ([PR] fix: fork prefix gate is advisory — adopt drifted fork children with a classified warn (#3741) #3757 review)
• docs: sync fork-adopt parity-contract comment with the advisory gate ([PR] fix: fork prefix gate is advisory — adopt drifted fork children with a classified warn (#3741) #3757 review)
• fix: fork prefix gate is advisory — adopt drifted fork children with a classified warn (This run failed with 409: "first child request does not match its inherited ordered prefix" | 本轮运行失败409 "first child request does not match its inherited ordered prefix" #3741)
• test: bound teardown 'close' await in two long Windows-flaking suites (Windows CI: Parent test hangs on long test files, child lines batch-cancelled (zero assertion failures, multiple branche | Windows CI：长测试文件父测试挂起、子行被整批取消（零断言失败，多分支中招） #3731)
• test: pin next-turn convergence the clamp restores (fix: overflow-arm stored above the headroom-reserved judging limit → unbounded CONTEXT_WINDOW_EXCEEDED loop #3738)
• fix: acp_delegate_wait wait-timeout no longer reads as task failure (This run failed with 409: "first child request does not match its inherited ordered prefix" | 本轮运行失败409 "first child request does not match its inherited ordered prefix" #3741)
• fix(openai wire: prompt_cache_key stamped unconditionally — breaks strict-schema upstreams (NVIDIA NIM 400), no config escape #2645): single-quote the pck-banner HTML fragments so WEB_CLIENT parses
• fix: emit fork-prefix drift labels only for their own failed check
• fix: un-export PublicForkPrefixVerdict per unused-exports gate
• fix: classify FORK_PREFIX_CONFLICT rejections instead of failing blind (acp_delegate_wait reports "Failed" on a normal wait timeout while the delegate is still running #3740)
• docs: cite fix: overflow-arm stored above the headroom-reserved judging limit → unbounded CONTEXT_WINDOW_EXCEEDED loop #3738 in the overflow-arm clamp comment (was an unfilled #N+1 placeholder)
• fix(openai wire: prompt_cache_key stamped unconditionally — breaks strict-schema upstreams (NVIDIA NIM 400), no config escape #2645): de-export two internal-only types tripping the unused-exports gate
• fix(openai wire: prompt_cache_key stamped unconditionally — breaks strict-schema upstreams (NVIDIA NIM 400), no config escape #2645): collapse mitm:// routes to their https egress key in normalizeUpstreamOrigin
• fix: clamp the overflow-arm to the headroom-reserved judging limit
• test(hermes-plugin): pin the nested context_lengths cache read ([PR] fix(hermes-plugin): resolve the real context window from the host #3730)
• docs: describe the openai chat-wire prompt_cache_key policy — stamp-by-default + A/B probe-and-remember (上游侧失效导致缓存命中只有80%用的api反代 workbuddy glm5.3 #2218/openai wire: prompt_cache_key stamped unconditionally — breaks strict-schema upstreams (NVIDIA NIM 400), no config escape #2645/[PR] fix(#2645/#2649/#2583): wire-conformant prompt_cache_key on every egress lane + Cowork lane routing #3556)
• docs: sync omp Cowork caveat with launcher routing ([Bug] Anthropic requests leak prompt_cache_key when Cowork transport bypasses native fetch interception #2583/[PR] fix(#2645/#2649/#2583): wire-conformant prompt_cache_key on every egress lane + Cowork lane routing #3556)
• test: regen storage-render goldens for the nudge low-water stamps (The nudge growth reference point, once clamped by a single high-water mark, never recedes — long sessions must wait unti | nudge 增长参考点被单次高水位钳死后永不回落，长会话只能等到 75% OVER-LIMIT 才压缩 #3674)
• kernel: sync package-lock for acp-kernel 0.0.134 (lock-sync gate)
• fix(hermes-plugin): resolve the real context window from the host
• fix(launcher): mirror OMP proxy truthiness in PI_PROXY_ANTHROPIC fill guard
• launcher(omp): route the Cowork anthropic lane through the proxy (PI_PROXY_ANTHROPIC)
• fix([Bug] Anthropic requests leak prompt_cache_key when Cowork transport bypasses native fetch interception #2583/openai wire: prompt_cache_key stamped unconditionally — breaks strict-schema upstreams (NVIDIA NIM 400), no config escape #2645): withhold the pi-lane prompt_cache_key stamp on un-consumable transports
• fix(openai wire: prompt_cache_key stamped unconditionally — breaks strict-schema upstreams (NVIDIA NIM 400), no config escape #2645): probe-and-remember prompt_cache_key at the egress seam
• kernel: nudge growth references become low-water marks (The nudge growth reference point, once clamped by a single high-water mark, never recedes — long sessions must wait unti | nudge 增长参考点被单次高水位钳死后永不回落，长会话只能等到 75% OVER-LIMIT 才压缩 #3674)
• fix: unexport module-internal symbols flagged by the unused-exports gate ([Bug?] Under high load, health check timeouts appear to be misinterpreted as "proxy unavailable": large session requests | [Bug?] 高负载下 health 探活超时疑似被当成「代理不可用」：大会话请求被本地拒发（0.1.199 / plugin native） #3654)
• test: retry overlay-busy-lock fixture teardown rmSync (Windows ENOTEMPTY)
• fix: [PR] fix: WS-lane egress byte budget with image shedding for close-1009 oversize requests (#3701) #3706 review follow-ups — single log sink for loop shed lines, accurate force-mode comment, no raise-budget advice after a proven 1009 (Findings 2/3/4)
• test: retry overlay-busy-lock fixture teardown rmSync (Windows ENOTEMPTY flake)
• fix: WS-lane egress byte budget with image shedding for close-1009 oversize requests ([Bug] Jimp is active, but image accumulation still triggers Responses WS 1009; duplicate request size not reduced (v0.1. | [Bug] Jimp 已生效，图片累积仍触发 Responses WS 1009；重复请求未缩小（v0.1.200） #3701)
• fix: replay valid JSON for unparseable proxy-tool arguments in the loop re-request ([Bug] Malformed/truncated parameters of the compress tool are echoed back verbatim → upstream 400 → entire round fails; | [Bug] compress 工具参数畸形/截断被原样回传 → 上游 400 → 整轮失败，nudge 每轮重触发形成死锁 #3661)
• chore(kernel): restore missing EOF newline in kernel/package.json
• docs: document DSH inspector fetch-accessor conflict + workaround ([plugin] Guarded globalThis.fetch accessor blocks coexisting fetch observers (dsh experimental-inspector fails to load) #3528)
• fix: web editor accepts JSONC config text (JSONC doc examples vs bare JSON.parse loader — copying the docs silently drops the config | 文档示例是 JSONC（含 // 注释），loader 却用裸 JSON.parse——照文档抄配置会静默失效 #3572, External summary credential feature entirely unusable + fully opaque failures + no retry | 外部摘要（externalSummary）凭据功能整体不可用 + 失败全程黑箱 + 无重试 #3488)
• fix: startup audit must tolerate JSONC config files (JSONC doc examples vs bare JSON.parse loader — copying the docs silently drops the config | 文档示例是 JSONC（含 // 注释），loader 却用裸 JSON.parse——照文档抄配置会静默失效 #3572)
• fix: report real compression count in cold-load diagnostics
• fix: guard maybeRetry self-heal against a superseded respawn flight
• fix: busy-episode log announces once per episode, not once per caller ([Bug?] Under high load, health check timeouts appear to be misinterpreted as "proxy unavailable": large session requests | [Bug?] 高负载下 health 探活超时疑似被当成「代理不可用」：大会话请求被本地拒发（0.1.199 / plugin native） #3654)
• fix: three-state health probe — a busy proxy is no longer misjudged as dead ([Bug?] Under high load, health check timeouts appear to be misinterpreted as "proxy unavailable": large session requests | [Bug?] 高负载下 health 探活超时疑似被当成「代理不可用」：大会话请求被本地拒发（0.1.199 / plugin native） #3654)
• fix: price opaque parts proportional-bytes in both billing modes; stop pricing responses control items (Image errors accumulating in the conversation | 对话中图片错误堆积 #3398)
• fix: price codec-carried opaque media parts in size decisions (Image errors accumulating in the conversation | 对话中图片错误堆积 #3398)
• fix: close config-redaction gaps in enriched diagnostics blob (Compression conflict banner (bili v0.1.198, node v22.23.2, linux-x64) #3611)
• feat: enrich auto-filed diagnostics reports (Compression conflict banner (bili v0.1.198, node v22.23.2, linux-x64) #3611)
• test: no-op re-run of windows shard — pre-existing lane 统一决策点收尾:四处 B 类名字旁路迁入 + nudgeOn 单点(#3482) #3490-class teardown flake, unrelated to this diff
• fix([dsh-native] takeover-gate refusal (threw state) fails open with the FULL uncompressed transcript; a >window local history dies upstream with 400 (maximum context length) #2924): point the oversize-refusal 422 at a host restart, not a retry ([dsh] Cannot intercept dsh requests? Dead scope retransmission + post-update dual instances trigger #2924 oversize 422 r | [dsh] 无法拦截dsh请求? 死 scope 重发 + 更新后双实例引发 #2924 oversize 422 重试风暴（用户全量日志定案，dsh 0.2.1rc1） #3589)
• fix(ci): sync test-import-graph gate entry list with tsup entries (Hook-model clients (claude/zcode) have no mid-session self-heal when the proxy dies — ECONNREFUSED until next session #1821)
• fix(ci): sync unused-export gate entry list with tsup entries (Hook-model clients (claude/zcode) have no mid-session self-heal when the proxy dies — ECONNREFUSED until next session #1821)
• fix(zcode-mcp): redeclare origin mutable on the watchdog state (typecheck)
• fix(zcode-mcp): watchdog recovers dead lane-owned proxies, waits only on user pins (Hook-model clients (claude/zcode) have no mid-session self-heal when the proxy dies — ECONNREFUSED until next session #1821)
• test(claude-mcp): adapt fixture to new-master launcher typings (Hook-model clients (claude/zcode) have no mid-session self-heal when the proxy dies — ECONNREFUSED until next session #1821)
• test(claude-mcp): normalize the fake CLI log across platforms (Windows CI red)
• docs: CONFIGURATION en/zh — claude's MCP face now points at dist/claude-mcp-entry.js (Hook-model clients (claude/zcode) have no mid-session self-heal when the proxy dies — ECONNREFUSED until next session #1821)
• fix(claude-mcp): keep the watchdog counter hot across a failed bring-up (zcode parity)
• feat: claude lane mid-session proxy self-heal via MCP watchdog (Hook-model clients (claude/zcode) have no mid-session self-heal when the proxy dies — ECONNREFUSED until next session #1821)

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

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

接单 · 发版类 PR 全窗回归审计(进行中)

查重:不重复。 在途 open PR 中只有本条 #3781 是 v0.1.202 release PR,无并列第二个发版 PR。相邻条目:#3783 是「do-not-merge」的 master CI probe,#3782 是 kernel lock 同步修复(与本窗 kernel bump 链相关,见下)。

PR 形态已核对: head 8578340a5 相对 origin/master(2b40342ef)仅多一个纯版本 bump commit。按发版类 PR 政策,触发 v0.1.201(ff42dcc13)→ master 全窗回归审计,不是只看两行版本 diff——分段通读整窗 diff(--stat 总览 → 高危域[自动更新/dsh/lockfile/env 模板/network-proxy]精读 → 其余按目录分批),逐项过窗口内 69 个 fix/feat。

预审已钉死的一个高危域点(lockfile): master 上 kernel/package.json=0.0.136,但 kernel/package-lock.json=0.0.133(被 #3767 遗留)——真实不一致,但不阻塞本发布:CI 的 lock-sync gate(ci.yml:506)只重算并比对根 package-lock.json(根 lock 里 kernel=0.0.136、pi-subagents=0.2.1,均在同步态);没有任何 gate 检查 kernel 自己的 lock,kernel test 脚本也不走 frozen lockfile。该不一致已由在途 #3782(2 行改动)跟踪,非本 release 引入,无需塞进本 PR。

继续分段通读中,完成后按结构化八问给合并判定。

[github-actions]

📦 Built Package Artifact

Branch: 2026-10-19_release-v0.1.202 (e75e010)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.4505-pr.3781.f885e08

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-pr3781.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)

发版类 PR · v0.1.201→release-head 全窗回归审计 — 结论：可合，无回归

查重：不重复（仅 #3781 是 v0.1.202 发版 PR；相邻 #3783=勿合的 CI 探针、#3782=kernel lock 同步修复）。
PR 形态：head 8578340a5 = release base 2b40342ef(#3770) + 恰好一个纯版本号提交 → 标准发版 PR，触发全窗审计（非两行版本 diff 审阅）。
审计窗口（已更正）：权威发布内容 = v0.1.201(ff42dcc) .. 8578340a5，110 文件。审计期间共享 clone 的 origin/master 被 #3774(snapshot-render-ci-flake) 从 2b40342ef 推进到 5afa77fd9；已核对 #3774 仅改 snapshot-render.ts(+纯可观测 metadata 标记，degrade 仍原样返回 client view、绝不碰出站字节)+其测试(+#3771 静默回退盲区守卫)，纯增量、不改出站行为、不改变任何回归结论，且不在本发布内（随下一版本）。逐文件核过的源/测试文件全部包含在发布窗口内。通读方式：--stat 总览 → 高危域逐行精读 → 其余按目录分批，未整吞 diff。

高危域逐项点名核查

• 自动更新/自更新：src/update.ts 本窗未改（自更新核心不动）；install-lane(plugin-install)、rollout、knobs 逐行核——均 sound；advisory/rollback 通道未变。
• dsh：dsh-native( dsh attach mode test intermittently fails on Windows CI: the async recovery flow reverts the registered origin back to i | dsh attach 模式测试在 Windows CI 偶发红：异步恢复流程把注册 origin 改回旧值（测试隔离竞态） #3616 自愈 supersession 守卫)、fork-adopt(仅注释更新，parity 契约不变)已核；dsh-ledger(dsh 车道自动更新改全:自愈同步三本账(manifest/lockfile/副本)+ 存量漂移显式修复,自动更新后必须还能用 #3496 convergeDshLedger)未改。
• 版本钉死与 lockfile：root package-lock.json 与 vendored 包同步(kernel 0.0.136 / pi-subagents 0.2.1)→ lock-sync 门通过；kernel/package-lock.json 停 0.0.133([PR] Payload-addressed image message IDs: seed adds media fingerprint slot (#3751) | 图片消息 id 按载荷寻址:种子增加媒体指纹槽(#3751) #3767 遗留)——不阻塞(无任何 CI 门检查该文件、kernel 测试无 frozen lockfile)，已在途 [PR] kernel: sync package-lock.json to 0.0.136 (lock left at 0.0.133 by the bump chain) #3782 跟踪，非本窗引入，不并入本 PR。
• systemd/env 模板：launcher env 填充(PI_PROXY_ANTHROPIC 仅在用户未设时补，truthiness 镜像 OMP)已核；窗内无 systemd 单元改动。
• 网络/代理路径：egress-shed / fetch-transport / loop-core / relay / pck-probe / responses-ws / liveness / native-intercept / launcher 逐行读毕，均 sound。

重大行为变更清单（旧→新+原因，全部有意、已披露、带回归测试）

1. fork 继承前缀门(This run failed with 409: "first child request does not match its inherited ordered prefix" | 本轮运行失败409 "first child request does not match its inherited ordered prefix" #3741)：子会话首请求前缀哈希不匹配 从【永久 409 死锁】→【分类告警后采纳】(publicForkInputCheck 返回 {ok,reason})。分歧子会话首个请求以未折叠态发出=一次性 token 尖峰、永不损坏；保留 side-lane 否决权(publicForkPrefix 仍 false)。原因:原 409 是被拒请求永不计数、无清除/TTL 的死胡同。
2. 代理存活探测([Bug?] Under high load, health check timeouts appear to be misinterpreted as "proxy unavailable": large session requests | [Bug?] 高负载下 health 探活超时疑似被当成「代理不可用」：大会话请求被本地拒发（0.1.199 / plugin native） #3654)：TCP 连上但健康探针超时 从【判死→重生进同一进程(twin 碰撞 autoRestartOnUpdate + occupied target port -> orphan process leak: new instance silently hops to the next port, old instance never exits #3044)】→【判 BUSY(存活但事件循环饱和,如大压缩),持有并抖动复探≤10s;busy 替换实例照路由(请求排队);twin 碰撞退出码 3 改为采纳已在服务的实例而非报错】。真死(ECONNREFUSED)仍走毫秒级快速路径。
3. prompt_cache_key 盖章(openai wire: prompt_cache_key stamped unconditionally — breaks strict-schema upstreams (NVIDIA NIM 400), no config escape #2645/上游侧失效导致缓存命中只有80%用的api反代 workbuddy glm5.3 #2218)：盖章面 从【仅 loopback/admin 改写】→【所有目的形态】(修远程池亲和丢失);保护从静态门→运行时出口 A/B:出口 400 携带 key 则去 key 重发一次,确认拒绝即进程内记忆该 origin+model 直到重启(acp_status/web 可见)。
4. 超尺寸本地拒答文案([dsh] Cannot intercept dsh requests? Dead scope retransmission + post-update dual instances trigger #2924 oversize 422 r | [dsh] 无法拦截dsh请求? 死 scope 重发 + 更新后双实例引发 #2924 oversize 422 重试风暴（用户全量日志定案，dsh 0.2.1rc1） #3589)：422 补救话术 从【重试本轮】→【重启一次客户端宿主(会话状态已落盘,无损)】;状态码/x-bili-gate 头不变。
5. JSONC 配置容忍(JSONC doc examples vs bare JSON.parse loader — copying the docs silently drops the config | 文档示例是 JSONC（含 // 注释），loader 却用裸 JSON.parse——照文档抄配置会静默失效 #3572/External summary credential feature entirely unusable + fully opaque failures + no retry | 外部摘要（externalSummary）凭据功能整体不可用 + 失败全程黑箱 + 无重试 #3488)：启动凭据审计 从【严格 JSON(注释配置静默跳过全局 externalSummary 审计)】→【容忍 JSONC(与运行时加载器一致)】。
6. 不透明媒体计费(Image errors accumulating in the conversation | 对话中图片错误堆积 #3398)：未知形状媒体 从【不计价/flat 回退(曾致 side request(max_tokens<=200)绕过 preflight 且不检查载荷体积,携带全量历史直转上游导致 400 #554 小载体 413)】→【按字节 bytes/4 比例计价】;折叠侧栏为其加 [media: …] 占位避免静默丢失。
7. WS 车道字节背压([Bug] Jimp is active, but image accumulation still triggers Responses WS 1009; duplicate request size not reduced (v0.1. | [Bug] Jimp 已生效，图片累积仍触发 Responses WS 1009；重复请求未缩小（v0.1.200） #3701,新特性+新配置)：Responses WS 出站帧超软预算(默认 10MiB,0 关闭)时 从【原样发送(网关 close 1009 零事件丢弃整请求)】→【先从出站副本归档/降分辨率最老的非保护图片,仍超则一次有界缩小重发】;close-1009-零事件 从【计入告警/熔断(可能砖掉健康端点)】→【排除出熔断核算】。HTTP 车道不受影响。
8. claude/zcode MCP 会话内看门狗(Hook-model clients (claude/zcode) have no mid-session self-heal when the proxy dies — ECONNREFUSED until next session #1821)：车道代理会话中途死掉 从【搁置到下次会话启动(zcode 旧 attached 门把 lane-owned 也拦了)】→【仅显式用户 pin 才只等待;lane-owned 走 ensureProxyRunning 恢复,会话中途自愈】;claude MCP 面 从【无条件 mcp add(可能覆盖同名外部 server)】→【幂等+内容所有权(迁移 dist/mcp.js→dist/claude-mcp-entry.js,外部 server 绝不触碰)】。
9. kernel nudge 引用(The nudge growth reference point, once clamped by a single high-water mark, never recedes — long sessions must wait unti | nudge 增长参考点被单次高水位钳死后永不回落，长会话只能等到 75% OVER-LIMIT 才压缩 #3674,host+kernel 协同)：nudge 引用 从【事件门控漂移重置(亚区间下降会死区)+host 侧 reanchorNudgeOnUsageDrop 兜底】→【低水位单调下行(任意 grade 下一读数自纠正),host 镜像删除】。节奏幅度(flat 50K,KDD#8)不变。
10. kernel 图片 id(Image message ids re-mint whenever an earlier image leaves the visible tail (membership-based ClusterCounter suffix) — prefix-cache breaks per fold, feeds #3748 storage growth #3751)：图片核心消息 id 从【纯文本投影】→【载荷指纹寻址】;非图片 id 字节不变;存量图片会话升级时重铸图片 raw-id(order-centric reconcile fix(dsh): model switch drops all folds — fold-reconcile #2396 detect-only cannot re-anchor rewritten tool-call ids → doom-probe 502 session death #2915 吸收,无复用/重复违规)。
11. snapshot-render(Snapshot rendering: proxy-lane outbound is now rendered authoritatively server-side; existing sessions are immune to ker | 快照渲染:proxy 车道出站从服务端正史渲染,存量会话对内核升级免疫(#3752 根治) #3756,新子系统)：默认 0% 休眠(BILI_ROLLOUT_SNAPSHOT_RENDER=1 才启用)→ 本发布零行为变化。
12. sourcemap(Remove sourcemaps during bundling to reduce plugin size. | 打包去掉sourcemap 减少插件体积 #3777)：npm tarball 从【含 ~109MiB .map】→【不含】(无运行时消费者,仅本地调试);零运行时影响。
13. Cowork pck 盖章 + pi/omp anthropic 车道([Bug] Anthropic requests leak prompt_cache_key when Cowork transport bypasses native fetch interception #2583/[Bug] bili omp 扩展 stampPromptCacheKey 无条件注入顶层 prompt_cache_key —— 严格 schema 上游（opencode zen anthropic 端点）全量 400 #1403)：pck 盖章抑制面 从【仅未路由回退分支】→【所有盖章路径(cert-MITM 车道此前漏抑制→Anthropic 严格 schema 每次委派 400)】;pi/omp Cowork anthropic 流量 从【node:https 直连绕过 bili(未压缩不可见)】→【用户未设代理时经可压缩代理】。
14. overflow-arm 钳制(fix: overflow-arm stored above the headroom-reserved judging limit → unbounded CONTEXT_WINDOW_EXCEEDED loop #3738/acp_delegate_wait reports "Failed" on a normal wait timeout while the delegate is still running #3740)：overflow arm 钳制到 headroom-reserved 判定上限(防 arm 越过可用余量)。
15. hermes 上下文窗解析([PR] fix(hermes-plugin): resolve the real context window from the host #3730)：请求路径窗查询 从【每次问宿主】→【记忆化(命中 1h/未命中 30s TTL),<64K 窗不入共享缓存(防投毒),全程 fail-open】。

测试有效性

强覆盖——每个重大行为变更都有专属回归测试:issue3654-health-probe-three-state(+576)、issue3701-ws-byte-budget(+478)、snapshot-render(+611)、fix-arm-headroom-clamp(+501)、issue3398-opaque-media-reserve(+323)、issue2645-pck-probe(+273)、claude-mcp-entry(+440)、zcode-mcp-entry(+147)、issue3661-malformed-proxy-args-replay(+178)、issue3572/3488、plugin-fork+fork-side-request-lane、issue2218、web-diagnostics-enrichment、hermes-native、omp-native-identity、issue3573-cold-load。issue1595-nudge-reanchor.test.ts 删除(与 kernel #3674 协同,场景转为 kernel 测试)。golden storage-render/*.json 再生(已给理由);wire-contract/kernel.json 未动(flat-50K nudge 常量/提示/工具表稳定)。断言有牙(如 assertDActive 显式抛错命名 degrade 原因,防静默回退读绿)。

结构化八问

1. 是什么问题:这是 v0.1.202 发版 PR(workflow 直推 master 被分支保护挡下后的标准回退发布路径);任务是发版前全窗回归审查,判定能否安全发布,不是修单一 bug。
2. 什么原因导致:不涉及单一故障根因——是对 ~69 个已合入 fix/feat 的合并视角批量回归审查。
3. 影响范围与严重程度:覆盖所有升级用户经 bili 的请求路径(压缩/计费/代理存活/安装车道/持久化/面板)。本窗无破坏性变更,全部为可用性/正确性修复或默认关闭的新特性。严重度:无回归;若干有意行为变更(见清单),均已披露且带测试。
4. 怎么修复的:n/a(审计任务)。方法:分段通读全窗 diff(高危域逐行精读,未整吞)+逐项过既有行为清单+测试有效性核查+lockfile/CI 门核对。
5. 是否有回归问题:无。逐项核过既有行为:fail-fast 未退化为吞错(fork 门硬拒转采纳是有意可用性调整且保留 side-lane 否决;oversize 拒答状态码不变);阈值/时序/输出格式变化均为有意并已披露(WS 预算默认 10MiB、nudge 低水位、pck 盖章面、422 文案等);缓存/重试策略未静默失效(egress-shed 只动出站副本且确定性;pck 撤 key 为进程内记忆可重启重置)。
6. 是否破坏缓存:否。前缀缓存友好——egress-shed 仅改出站副本、确定性(同输入同输出);pck 出口 A/B 失败才撤 key(进程内、重启重置、acp_status/web 可见);snapshot-render 默认 0% 休眠;kernel 图片 id 载荷寻址的重铸由 order-centric reconcile(fix(dsh): model switch drops all folds — fold-reconcile #2396 detect-only cannot re-anchor rewritten tool-call ids → doom-probe 502 session death #2915)吸收。
7. 是否改了设计方向:新增子系统均为增量且默认关/可选(snapshot-render 0% 休眠、egress-shed WS 字节背压、三态 liveness、MCP 会话内看门狗、pck 出口探测),未转向既有架构边界。唯一方向性调整:fork 继承前缀门由「硬拒绝」转「建议式采纳」(This run failed with 409: "first child request does not match its inherited ordered prefix" | 本轮运行失败409 "first child request does not match its inherited ordered prefix" #3741,可用性优先,保留否决权)——属有意调整而非失控漂移。
8. 是否引入了新的配置:是——network.wsOutboundBytes / BILI_WS_OUTBOUND_BYTES(默认 10MiB,0 关闭 WS 预发送 shed;HTTP 上限 network.outboundBodyBytes 不变),已在 CONFIGURATION.md en+zh 记录(L161/L1790)。另有沿用既有 BILI_ROLLOUT_* canary 约定的 BILI_ROLLOUT_SNAPSHOT_RENDER(默认 0%,休眠)。无其它新 env/config 字段。

观察(非缺陷、不阻塞)

claude/zcode MCP 看门狗用布尔健康探针(2s)而非三态 liveness;因有 3 次失败阈值 + ensureProxyRunning 的 twin-collision 采纳兜底,任何「busy 误读为 dead」都收敛到采纳而非砖机,功能安全。仅为一致性备注,无需处理。

判定

本版本引入回归:无。 全窗(110 文件)逐域核毕,既有行为未被无意破坏,行为变更均有意且带测试,lockfile 根锁同步通过。唯一发现(kernel/package-lock.json 停 0.0.133)已在途 #3782 跟踪、非阻塞、不并入本 PR。可合——merge 由人执行,CI 在 push head 上跑全局把关(本地未跑全量套件,符合测试纪律)。