Skip to content

release v0.1.193 - #2767

Merged
ranxianglei merged 12 commits into
masterfrom
2026-10-11_release-v0.1.193
Oct 11, 2026
Merged

ranxianglei merged 12 commits into
masterfrom
2026-10-11_release-v0.1.193

Conversation

@github-actions

Copy link
Copy Markdown
Contributor

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 v0.1.191

@github-actions

github-actions Bot commented Oct 11, 2026 •

Copy link
Copy Markdown
Contributor Author

📦 Built Package Artifact

Branch: 2026-10-11_release-v0.1.193 (bf6d7c9)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.3115-pr.2767.c9b8009

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-pr2767.tgz
npm install -g package

Install instructions are refreshed on each push.

@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 收到 release v0.1.193 PR(来自 one-click workflow 的 auto-fallback)。已确认:release commit 085c6bf37 仅改 package.json(0.1.192→0.1.193)+ package-lock.json 版本行,diff 干净;PR mergeable。CI 正在跑(e2e / lane-guard 已过,部分 lane 仍在排队)。现在开始按版本区间回归审计流程通读 v0.1.191 以来的全部变更(约 229 commits),高危域逐行核查后给出八问结论。

@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Release v0.1.193 版本区间回归审计完成(v0.1.191..master,229 commits / 242 files)

审计覆盖:release commit 本身(package.json 0.1.192→0.1.193 + lockfile 4 行,纯版本号变更,干净);高危域逐行精读——src/update.ts 本窗口零改动(自更新器未动)、dsh 域(dsh-native.ts/dsh-compaction-guard.ts/handle.ts/relay.ts)、lockfile/package.json、request-body-budget、storage 新层(driver/knob 默认值)、kernel golden pins、CI workflows(release.yml 等逐行确认纯增量)、测试断言密度抽查;其余文件按目录 stat+提交历史+定向扫描(fold-reconcile、loop/core thinking-family 扫描、persist 存储开关)。本地未跑测试(按纪律交 CI)。

本版本引入回归:未发现。 逐项既有行为结论:

重大行为变更清单(有意变更,旧值→新值):

  1. compress.stripImages/stripImagesKeepRecent:生效的 wire 侧图片剥离 → INERT(忽略+一次性警告);媒体改由默认折叠处理(建议:compress.stripImages 默认开启——否则带图消息不可压缩会形成只升不降的输入地板撑爆上下文 #2607/Archive-then-fold: default folding for archivable media (Closes #2607, Fixes #2609) #2640,pixels 先归档、decompress({imageRef}) 可恢复)。曾设 stripImages:true 的用户升级后图片会留在 wire 上直到被压缩折叠。CONFIGURATION.md 已标 REMOVED,config.ts 有 INERT 警告 ✅已披露
  2. 请求体预算:解码上限 100MB→200MB;新增并发上限(≥2 个大解码请求时第 3 个返回 503 request_body_busy,retry-after:1);重建后出站体 >100MB 现在返回 413(stage:forward),之前是原样转发给上游
  3. dsh durable-state 载体保护:所有实例钉死 → latest-per-identity(superseded 实例可折叠;用户配置任何 durable* pattern 即整体接管策略)— compress 增加按内容标记的硬排除(protectedContentMarkers),保护会话的基础设施注入免于折叠 #2446 有意设计,附 kernel/host 两层测试
  4. passthrough 通道(x-bili-passthrough):dsh native compaction 调用现在被拒绝(字段证据:45 次被 pipeline 拒、4 次从该通道落地)— [Bug] #1001 归因缺口:客户端原生压缩落地后只剩「未确认档」——bili 已记录的 17 次拒绝(dshCompactionRefusals)与 10 次 DIRECT 旁路未被用于归因 #2709 的修复目标本身
  5. compress 工具描述文案变化(JSON escape hints)→ 升级后全体会话一次性的 prompt-cache 失效重缓存
  6. fork protocol v2(兼容 v1)

新配置/依赖:

  • storage.unified(shadow|full,file)+ BILI_STORAGE_UNIFIED(env),缺省 = legacy(config-surface-guard CI 绿,文档已同步)
  • 移除 compress.stripImages/stripImagesKeepRecent(inert+警告,见上)
  • 新 optionalDependency better-sqlite3@^13.0.3(原生模块;optional 保证安装永不失败,引擎探测 better-sqlite3 → node:sqlite 双降级)

CI 状态:38 个 check runs — 31 success / 5 skipped / 0 failed;仅剩 2 个 windows-latest test 腿 in_progress(其余全绿)。且本窗口给 release 流程新增了「发布 sha 全量 CI 门禁」,合并后 CI 会在 publish 前再全量把关一次。

八问:

  1. 是什么问题:例行发版 PR(one-click 直推被分支保护挡回后的自动 fallback),打包 v0.1.191 以来 ~150 个提交的窗口工作。
  2. 什么原因导致的:不涉及(非缺陷修复型 PR)。
  3. 影响范围与严重程度:全体自更新用户;窗口含存储统一、fork v2、媒体折叠、dsh 守卫加固等架构级工作——但均为各 issue 内已审过的有意变更,本审计未发现跨分支交互引入的意外破坏。
  4. 怎么修复的:不适用;审计方法见上(高危域逐行 + 分批通读 + golden/CI/测试有效性核查)。
  5. 是否有回归问题:无(逐项清单见上;高危模式 fail-fast→吞错、阈值漂移、输出格式变化均排查过:413/503 是新增守卫而非吞错,dsh 检测收紧方向安全)。
  6. 是否破坏缓存:默认用户的会话持久化格式不变;prefix cache 有一次性失效(compress 工具描述文案变化);stripImages 用户的 wire 字节变化属有意变更。
  7. 是否改了设计方向:有——unified storage(SQLite 侧车→可选主存储,Phase 1 fidelity 契约钉在 golden corpus)、durable 载体投影到虚拟工具路径(kernel messagePathOf)、媒体折叠取代 wire 剥离、fork protocol v2。均为窗口内 owner 参与的既定方向。
  8. 是否引入了新的配置:storage.unified/BILI_STORAGE_UNIFIED(默认 legacy);移除 compress.stripImages*;新增 optional dep better-sqlite3。

判定:可合。 等最后两个 Windows test 腿跑完即可(预计无风险;即便有异常,新的发布前全量 CI 门禁也会在 publish 前拦下)。合并请人工执行:#2767

@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 CI 收尾确认:38 个 check runs 全部完成 — 33 success / 5 skipped / 0 failed,最后两个 windows-latest test 腿也已转绿。审计结论不变:本版本引入回归:无,可合。请人工合并:https://github.com/ranxianglei/billion-context/pull/2767(合并后 release.yml 会按新门禁在 publish sha 再跑全量 CI,然后自动发布 v0.1.193)。

@ranxianglei

Copy link
Copy Markdown
Owner

递补了几个进来 注意下 比如 #2763 可以自己扫描下

@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 递补提交扫描完成(增量窗口 eacbb5848..master@d0cebb909,即首审之后新并入的 10 个 PR / 25 文件 / +1743−94,src+docs 全量逐行读完):

PR 内容 审查结论
#2763 客户端 fetch 补丁补上 raw-URL 嗅探:agentless POST 打到原始 provider URL(takeover gate 拒绝路径)且 body 含 DSH_COMPACTION_MARKER_SNIFF 时改道 proxy passthrough lane,由 #2709 的决定性检查裁决;堵住 4/49 现场报告里 raw URL 的残余漏网 ✅ 方向安全:误匹配只多一次 loopback hop、服务端做决定;不可读 body(stream/FormData)fail-open 保持直发;Request 输入走 clone 探测原对象保持可发送;常量 import 不重推 slice(0,48)
#2764(修 #2756) 退化轮次门从「可见字节 > 0」改为「host 是否收到 ANSWER 文本」:纯空白 / 仅 reasoning 的输出不再算已交付答案,空轮重试得以触发;真实文本仍 byte-for-byte 拒绝重试(不会重发已交付答案) ✅ 四条 wire lane(openai/anthropic/google/responses)全部接入 answerProse 记账,含 #2769 的 dropped-tag 按分支判定与 Gemini 大写 finishReason 归一
#2774(修 #2766) 四个 durable classifier 一律先查 msg.role !== "assistant":模型回声不能顶替真载体抢占 durable/<name> 身份(latest-per-identity 下真载体会被降级折叠、伪造副本成为唯一幸存者) ✅ 安全加固,方向正确(carrier 本来就是 host 注入 user 通道)
#2768/#2771/#2772/#2788 storage review-fix pass:BEGIN IMMEDIATE 顶层事务(多进程 SQLITE_BUSY_SNAPSHOT 防护)、损坏 index.db 降级 legacy-only(单条 warn、不自动删文件)、refs/blocks 行按权威集整替(#2695 zombie 复活窗口关闭)、未用的 conversations 表在 v1 前移除、open 失败关句柄(Windows 可删)、freezeReplica 盖 live sidecar 的 forkContentStore、restoreReplica 拒绝覆盖活会话 id(HTTP 409)+ 物化 sidecar ✅ 全部是数据保护方向;conversations 表移除发生在 v1 落盘前,无迁移负担
#2781 Windows EBUSY 测试修复(rmrf 前关 index.db 句柄) ✅ 测试侧
#2784(经 #2795) cache ledger:事件计数跟随单一主因(else-if 链),report 只叙述本维度赢得分区的行;无 cause 字段的旧行保持历史显示 ✅ 修复一次动作翻转多维度的双重计数(dsh 切模型同时翻 prompt 指纹的场景);token 桶计费规则不变

增量行为变更清单(旧→新):① 空轮重试触发面扩大——纯空白/仅 reasoning 终轮现在会重试一次(旧:可见字节 >0 即抑制重试);② durable 身份不再接受 assistant 角色消息为 carrier;③ cache ledger 面板事件计数从「每维度各记一次」改为「单一主因记一次」(展示层,token 桶不变);④ agentless dsh compaction 调用在 raw URL 上也会被拦截进 passthrough lane 裁决(旧:直发漏网)。均为收紧/加固方向,无放宽。

分支状态:release 分支已再次同步到 e7103a90d——现在完整包含 master(含 #2795/#2784),对 master 的 diff 只剩 version bump 两行(package.json 0.1.192→0.1.193 + lockfile 4 行),干净。

CI(最终 head e7103a90d):12/13 条 lane 已 success(含 e2e codex/opencode/dsh/pi、local registry、image billing、PR build artifact);主 CI gate(typecheck+build+全量单测)仍在跑,完成后我补一条确认。中间 head bf4703ed4 上那条 cancelled 的 CI run 是 concurrency cancel-in-progress 的正常行为(被后续 push 取代,同 sha 的 push 事件 CI 已 success)。

结论:递补增量无回归,与首审结论一致——可合(待主 CI gate 转绿)。合并仍由人执行。

@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 CI 收尾确认(最终 head e7103a90d):主 CI gate(pull_request 事件,14 jobs:version-guard / kernel-guard / pi-subagents-guard / lane-guard / config-surface-guard / lock-sync / docs-sync / docs-coverage / unused-exports + test 矩阵 ubuntu×windows × node 22/24)已 completed / success——至此该 head 上全部 13 条 workflow run 均为 success。

结论不变:递补增量无回归,可合。合并由人执行:#2767

ranxianglei added a commit that referenced this pull request Oct 11, 2026
…regressions)

#2807 made the anonymous/identified identity chains hash the wire-NEUTRAL
projection. That exposed three regressions (all three failing legs on the
v0.1.193 release PR #2767):

1. The neutral projection MERGES a run of consecutive same-role messages
   into one turn (required for cross-wire equality — the anthropic render
   joins the run). A client that APPENDS another assistant message after
   an assistant tail (agent harnesses replaying model replies, fork
   follow-ups) re-renders the stored tail turn, so its item hash changes
   and the exact ladder fails by ONE item → fresh pfa-* mint → the fold
   ledger and the usage anchor are orphaned (spurious fork-of-fork
   children in fork-adoption-wires, 58701-vs-5000 sizing in
   issue2033-anonymous-anchor).

   Fix: per-source-message tail FRAGMENTS (fragmentHash) are now carried
   on every NeutralTurn, persisted on the chain (ChainEntry.tailFrags,
   snapshot round-trip included) and forwarded through note() from all
   resolve() paths. New rung 1.4 (append-continuation): a stored hv2
   chain whose first K-1 items are byte-equal AND whose stored tail
   fragments are a strict PREFIX of the incoming turn-at-K-1 fragments
   is the same conversation extended in place → match instead of mint.

2. The v1 (pre-#2791 raw-wire) fallback rung could OUT-VOTE a matching
   hv2 chain by raw depth (raw counts MESSAGES, neutral counts TURNS —
   16 vs 13 for the same history) and steal the request to a stale
   pre-upgrade duplicate session. Any user upgrading 0.1.192 → 0.1.193
   hits this seam on their first post-upgrade request. Fix: within the
   ladder, hv2 candidates now take precedence over hv1; hv1 still
   resolves untouched when no neutral chain matches (legacy bridge).

3. Image placeholder text blocks were joined with "" into the neutral
   text, so the USER_IMAGE_PLACEHOLDER line filter could not strip them
   (issue2601 filter-branch replay). Placeholder blocks/parts are now
   dropped during neutral-text collection.

Local: full targeted affinity sweep (27 suites) green; failing trio green.
ranxianglei added a commit that referenced this pull request Oct 11, 2026
…regressions)

#2807 made the anonymous/identified identity chains hash the wire-NEUTRAL
projection. That exposed three regressions (all three failing legs on the
v0.1.193 release PR #2767):

1. The neutral projection MERGES a run of consecutive same-role messages
   into one turn (required for cross-wire equality — the anthropic render
   joins the run). A client that APPENDS another assistant message after
   an assistant tail (agent harnesses replaying model replies, fork
   follow-ups) re-renders the stored tail turn, so its item hash changes
   and the exact ladder fails by ONE item → fresh pfa-* mint → the fold
   ledger and the usage anchor are orphaned (spurious fork-of-fork
   children in fork-adoption-wires, 58701-vs-5000 sizing in
   issue2033-anonymous-anchor).

   Fix: per-source-message tail FRAGMENTS (fragmentHash) are now carried
   on every NeutralTurn, persisted on the chain (ChainEntry.tailFrags,
   snapshot round-trip included) and forwarded through note() from all
   resolve() paths. New rung 1.4 (append-continuation): a stored hv2
   chain whose first K-1 items are byte-equal AND whose stored tail
   fragments are a strict PREFIX of the incoming turn-at-K-1 fragments
   is the same conversation extended in place → match instead of mint.

2. The v1 (pre-#2791 raw-wire) fallback rung could OUT-VOTE a matching
   hv2 chain by raw depth (raw counts MESSAGES, neutral counts TURNS —
   16 vs 13 for the same history) and steal the request to a stale
   pre-upgrade duplicate session. Any user upgrading 0.1.192 → 0.1.193
   hits this seam on their first post-upgrade request. Fix: within the
   ladder, hv2 candidates now take precedence over hv1; hv1 still
   resolves untouched when no neutral chain matches (legacy bridge).

3. Image placeholder text blocks were joined with "" into the neutral
   text, so the USER_IMAGE_PLACEHOLDER line filter could not strip them
   (issue2601 filter-branch replay). Placeholder blocks/parts are now
   dropped during neutral-text collection.

Local: full targeted affinity sweep (27 suites) green; failing trio green.
@ranxianglei
ranxianglei merged commit 68e7675 into master Oct 11, 2026
22 of 38 checks passed
@ranxianglei
ranxianglei deleted the 2026-10-11_release-v0.1.193 branch October 15, 2026 05:33
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant