[PR] remove: the experimental compressAtFoldPoint feature — fold-point relocation (v1 #5471 + unreleased v2 #5949), #6132 by ranxianglei · Pull Request #6150 · ranxianglei/billion-context

ranxianglei · 2026-11-10T15:49:15Z

这是什么

按 owner 指令（2026-11-10，issue #6132 已附完整事故分析）整体移除实验性功能 compress-at-fold-point（compress.compressAtFoldPoint / BILI_COMPRESS_AT_FOLD_POINT，v1 #5471 + 从未发布的 v2 #5949）：它把重放视图中压缩产生的内容挪到折叠点处，9 天内在其目标模型家族（qwen3-27b 系）造成两次不同的真实事故——2026-11-08 的自发重复压缩循环（#5949 的现场报告，v2 的设计动机）和 2026-11-10 的生产会话作者误认（v0.1.213 = v1 语义：同轮「回答 + compress」被重放时，压缩对挪位后回答文本裸挂在 transcript 尾部，27B 模型把自己的最后一条发言当成用户消息，重复回答了自己的答案，错误认知还固化进了折叠块主题）。

移除零迁移成本：开关 opt-in 默认关，关闭时布局 = 移除后布局（字节级一致）；v2 从未进 tag（最新 v0.1.213 仍是 v1，git tag --contains 为空）。曾考虑直接 git revert #5471：17 个文件冲突（v2 在 master 上改了同一批文件），否决，走正向删除。

改动

kernel（0.0.192 → 0.0.193）

删 kernel/src/relocate-compress.ts + 其测试；relocate-compress-carriers 管线节点与注册
Config.compressAtFoldPoint / .compressAtFoldPointClearThinking / .reasoningDroppable 与公开导出
v2 的 already-folded-echo 解析 kind + pointer marker（parse-compress-input.ts）

host / web / 文档 / 站点

env 旋钮 + resolver（knobs.ts）、文件键（config.ts 字段/字段表/解析循环）、compress-settings 级联与短路可达性、relay round-2 镜像、四个 prepare* 安装点、plugin-render S4 镜像、makeFoldPointReasoningDroppable（server.ts）、stream.ts 的 echo ack
旧配置残留：一次性 INERT 弃用警告（warnInertRemovedCompressKeys，含 top-level 残留，各层级全命名），值忽略
web 配置行 + i18n（en/zh）、config-reference yaml、站点文档/i18n、博客文章直接删除（owner 指令，不留勘误）
docs configuration（en/zh）+ lifecycle 台账（§1 两行删除、§5 记录）、CHANGELOG

Config surface
移除键：compress.compressAtFoldPoint、compress.compressAtFoldPointClearThinking（文件）；BILI_COMPRESS_AT_FOLD_POINT、BILI_COMPRESS_AT_FOLD_POINT_CLEAR_THINKING（env）。实验性、默认关、线上仅 ~10 天（v1），v2 未发布。迁移：无需迁移（功能关闭 = 基线行为）；残留键 = 一次性 INERT 警告 + 忽略值。

kernel-change: 删除 relocate-compress.ts、relocate-compress-carriers 节点、Config.compressAtFoldPoint/.compressAtFoldPointClearThinking/.reasoningDroppable 与 relocateCompressCarriers/planRelocateCompressCarriers 导出——整功能移除（owner 裁定，#6132：9 天内两次目标模型家族上的真实事故，v2 未发布，移除零迁移成本；开关关闭时变换本就是字节级 no-op）。kernel 0.0.192 → 0.0.193（两个 package-lock 同步）。

预检（本地）

kernel: tsc --noEmit ✅；kernel 全量测试 1621/1621 ✅
host: tsc --noEmit ✅；check-unused-exports ✅
CI 守卫本地预跑：lifecycle-ledger-guard ✅、config-surface-guard ✅、gen-config-docs check-coverage ✅、check-architecture-docs ✅
定向测试：removed-compress-keys-inert 4/4（含新增 Remove the compressAtFoldPoint feature — it reorders resent history and caused an authorship-misattribution incident in a live session #6132 用例：四层级命名 + 去重 + 值忽略）✅、compress-reject-reasons ✅、verif-subset-guard + 相关 verif/generic/induction 57 ✅、compress 配置解析套件（rules/absorb/kernel-tier-defaults/strip-images/external-summary）78 ✅
e2e-codex 车道与全量套件判定 = 远端 CI（本地 fake 车道需 codex 二进制长跑，留给 CI）

Closes #6132

附：CI 唯一红 lane 的修复（#6160）

本 PR 推送后，CI 唯一红的 lane（real pi package-native lane vs deterministic fake upstream，e2e-native-pi 的 #2791 两例 compress must be CALLED exactly once）经排查非本 PR 引入：干净 master 本地与 schedule CI 同样红。根因在测试假上游 tests/e2e/fake-upstream-chat.mjs——它的脚本队列键 conv|sourceText.slice(0,64) 把 ACP 渲染标签里的易变字节算进了键。首个真实 fold 后，产品侧（#5608-D 把引号空间对齐推迟到下一次成功 fold + #5608-C 抬升 time= 高水位）会重排旧 ref（m00001→m00015，refs 永不复用）并丢掉 time=，于是同一条消息在下一次请求里标签字节变了，64 字切片一变键就变，脚本指令队列整段重放 → 多出一第二次（no-op）compress 调用。CI 上同 SHA 时红时绿：对齐在 fold 后的 scheduleSave 若 unified store 行尚未可见则 no-op（绿），行已可见则重排（红）；本地 node-25 行总是可见 → 确定性红。

修复（issue #6160 先行，commit 13015fec4 随本 PR 分支追加）：队列键改为 conv|<source ref>|<stripAcps(body).slice(0,64)>。ref 是消息出现位置的标识（switch/back 两条字节相同的 prompt 仍各得各的队列；ref 被重排时脚本本就作废），stripAcps 用既有辅助函数剥掉易变外壳。一处改动覆盖共享该假上游的全部 6 个 e2e 套件。

本地验证（pi-stable 0.83.6）：两 #2791 例转绿（fold 运行内队列完整走完 probe→compress→acp_status，恰一次 compress）；native-pi/dsh/opencode/delegate-notify/subagent-sessions 全量跑过，余下两处本地失败在未改动假上游上复现完全相同（本地环境 artifact，与本改动无关）。CI 上该 lane 已在新 head 13015fec4 转绿。

…6132) Owner-directed full removal of the experimental compressAtFoldPoint / compressAtFoldPointClearThinking feature (v1 #5471 + unreleased v2 #5949): two distinct live pathologies in 9 days on its own target model family — 2026-11-08 voluntary re-compress loops (v1, #5949 field report) and 2026-11-10 authorship misattribution (v1 in v0.1.213 production: the pair's relocation left the turn's own answer text bare at the transcript tail, and a 27B model read its own last utterance as a user message, m00376 chain). - delete kernel/src/relocate-compress.ts + kernel/tests/relocate-compress.test.ts - drop the relocate-compress-carriers pipeline node (after hide-compress-calls, before view-latch) and its registration - drop Config.compressAtFoldPoint / .compressAtFoldPointClearThinking / .reasoningDroppable and the public relocateCompressCarriers/plan exports - drop the already-folded-echo parse kind + v2 pointer marker (parse-compress-input.ts) and the two echo tests - kernel 0.0.192 -> 0.0.193 (both package-locks) kernel-change: removed relocate-compress.ts, the pipeline node, 3 Config fields and the public exports (full feature removal, owner decision after the 2026-11-10 incident, #6132); with the switch off the transform was a byte-identical no-op, so removal changes nothing for the default layout. kernel 0.0.192 -> 0.0.193. Closes #6132 (kernel side; host/web/docs follow in the next commit of this PR)

Companion to the kernel removal: strip the experimental fold-point relocation feature (v1 #5471 + unreleased v2 #5949) from everything else. - env knobs BILI_COMPRESS_AT_FOLD_POINT(_CLEAR_THINKING) + resolvers (knobs.ts), file keys compress.compressAtFoldPoint(_ClearThinking) (config.ts: CompressSettings fields, COMPRESS_SETTING_FIELDS, parser loop), compress-settings.ts cascade/pass-through/reachability, relay.ts round-2 mirror + drop-predicate install, the four prepare* install sites, plugin-render.ts S4 mirror, makeFoldPointReasoningDroppable (server.ts), the v2 already-folded-echo ack in stream.ts - stale configs: one-shot INERT deprecation warning extended to the two removed keys (every level incl. top-level), values ignored - web config row + i18n strings (en/zh), config-reference yaml, site docs, site i18n strings, blog post + article entries deleted (owner instruction) - docs configuration en/zh + lifecycle ledger (rows removed, §5 note), CHANGELOG entry - tests: delete tests/compress-at-fold-point.test.ts, drop the echo test from compress-reject-reasons.test.ts, extend removed-compress-keys-inert with the removed fold-point keys Closes #6132

github-actions · 2026-11-10T15:53:22Z

📦 Built Package Artifact

Branch: 2026-11-10_remove-compress-at-fold-point (13015fe)

Latest build — install this exact version

npm install -g billion-context@0.0.9038-pr.6150.4aea2bb

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
`13015fe`	`billion-context@0.0.9038-pr.6150.4aea2bb`	build
`5211ee9`	`billion-context@0.0.9001-pr.6150.0933416`	build
`e2e05d1`	`billion-context@0.0.8965-pr.6150.f9151bd`	build

Download artifact

Download the artifact from the Actions run
Extract the tarball and install:

tar xzf billion-context-pr6150.tgz
npm install -g package

Install instructions are refreshed on each push.

ranxianglei · 2026-11-10T19:59:06Z

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b)

进度汇报（本会话曾因长时间无输出被平台强制重启一次，现已恢复，继续推进）：

查重：不重复——open [PR] 行中只有本 PR（#6150）针对 #6132，无并行第二 PR。

已核实的当前状态（live 验证：head sha 确认未动，CI 快照取自已刷新的镜像缓存）：

PR [PR] remove: the experimental compressAtFoldPoint feature — fold-point relocation (v1 #5471 + unreleased v2 #5949), #6132 #6150 head = e2e05d1b（2 commits：kernel 侧 + host/web/docs 侧），base master，mergeable_state=blocked。
CI：kernel-guard 红，其余 test/static/e2e 车道全部因它被跳过；守卫族（config-surface / lifecycle-ledger / architecture-docs / lock-sync / version-guard / build-artifact / docs-sync 等）全绿。

红因已钉死：guard 对 PR 正文做字面量匹配 kernel-change:（ci.yml:476），而本 PR 正文写的是 markdown 加粗的 **kernel-change**:——星号插在 "change" 与 ":" 之间，子串不命中，误判为未声明。声明内容本身是齐的（commit message 里也是规范的纯文本形式）。另注意 ci.yml 触发器 types 不含 edit，光改正文不会重跑 CI，必须推 commit。

代码审查结论（预览）：两个 commit 的删除面干净彻底（kernel 节点/导出/Config 三字段/v2 echo kind、host 四个安装点、relay round-2 镜像、web/站点/文档/博客、lifecycle 台账、CHANGELOG 全对齐；残留键走一次性 INERT 警告且四层级命名有测试钉住）。发现 2 处移除后遗留的孤儿 import（src/server/prepare-google.ts:10 的 isStrictReasoningEcho、src/plugin-render.ts:14 的 loadConfigFile, isMcpBiliCompress），按「小问题直接修」将在 PR 分支上直接补一个标注 commit 修掉。golden wire-contract 四个 pin 文件均不含被删符号，无需再生成。

接下来：修 2 处孤儿 import + 定向测试 → 把 PR 正文的声明改回纯文本 kernel-change: → push 触发 CI 全量重跑 → CI 出来后给结构化八问最终结论。

…#6132)

ranxianglei · 2026-11-10T21:01:44Z

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b)

审核收尾。本轮完成：① 修掉移除后遗留的 2 处孤儿 import（src/server/prepare-google.ts:10、src/plugin-render.ts:14，均已 grep 验证无其他使用点），commit 5211ee90 已推上 PR 分支；② PR 正文第 23 行的声明由 markdown 加粗 **kernel-change**: 改为纯文本 kernel-change:（guard 做字面量匹配，加粗星号导致误判未声明——这是 e2e05d1 上 kernel-guard 红、其余 test/e2e 车道被连带跳过的唯一原因）；③ push 触发 synchronize，CI 全量矩阵在新 head 5211ee90 上重跑中。标签已齐（bug / p1 / ci:full / scope/kernel，与本 PR diff 独立判定相符），无需改动。查重结论维持：不重复（open [PR] 行中仅本 PR 针对 #6132）。

结构化八问

是什么问题：实验功能 compress-at-fold-point（把重放视图中压缩产生的内容挪到折叠点处）在其目标模型家族（qwen3-27b 系）9 天内造成两次真实事故——2026-11-08 自发重复压缩循环、2026-11-10 生产会话作者误认（同轮「回答+compress」重放时，挪位后裸挂的回答文本被模型当成用户消息，重复回答了自己的答案并固化进折叠块主题）；owner 裁定整体移除。
什么原因导致的：挪位变换改变了重放视图中消息的呈现位置与归属边界，弱于说话人归因的 27B 系模型把「自己的最后一条发言」误判为用户输入；v2 方案（pointer marker + already-folded-echo 解析）在同一模型家族上仍触发自发压缩循环，且从未发布——继续修补的成本高于移除。
影响范围与严重程度：仅显式开启开关的用户受影响（v1 线上约 10 天暴露面）；默认用户零变化（开关关闭时管线节点直接 fast-path 返回，布局字节级一致）。严重度 p1：作者误认属核心语义破坏（会话内容被错误认知固化），故按 p1 处置。
怎么修复的：正向整删而非 git revert（revert 有 17 文件冲突，因 v2 在 master 改过同一批文件）。kernel：删 relocate-compress.ts + relocate-compress-carriers 节点 + Config 三字段 + 公开导出 + v2 echo kind/pointer marker，版本 0.0.192→0.0.193（两个 package-lock 同步）。host/web/文档：env 旋钮、文件键、compress-settings 级联、relay round-2 镜像、四个 prepare* 安装点、plugin-render S4 镜像、stream echo ack 全部移除；web 配置行/i18n、config-reference yaml、站点文档/i18n、博客文章（owner 指令直接删除）、docs configuration/lifecycle 台账、CHANGELOG 全对齐。残留键走一次性 INERT 弃用警告（四层级全命名、按签名去重、值忽略），并有新测试用例钉住。
是否有回归问题：逐项审过——默认用户零变更（快路径保证字节级一致）；显式 opt-in 用户回到基线布局 = 有意行为变更（见下披露）；含 v2 marker 文本的字符串从 echo kind 落入通用解析失败分类（v2 从未进 tag，生产不可达）；resolveRequestConfig 短路条件恢复特性前形态，净输出不变；golden wire-contract 四个 pin 文件均不含被删符号，无需再生成；孤儿 import 两处已补修。未发现无意回归。
是否破坏缓存：不涉及——不触碰前缀缓存/折叠游标/去重键/重试策略；关闭态本就是字节级 no-op，头部 system 消息保持字节稳定。
是否改了设计方向：是但属「移除实验面」而非转向：两种压缩模式、快照渲染、水位车道的既有方向不变；Kernel Boundary 合规（kernel-change 声明在位且已修正格式、版本已 bump、human-only 评审档适用）。
是否引入新配置：无新增；移除 4 个键——文件 compress.compressAtFoldPoint、compress.compressAtFoldPointClearThinking，env BILI_COMPRESS_AT_FOLD_POINT、BILI_COMPRESS_AT_FOLD_POINT_CLEAR_THINKING。迁移：无需（关闭=基线行为）；残留键=一次性 INERT 警告+值忽略。

行为变更披露（旧→新）：① 显式开启用户的折叠点挪位布局 → 基线布局；② 上述 4 个配置键/env → 从生效变为忽略+一次性警告。原因：目标模型家族两次真实事故，owner 裁定移除（#6132）。

判定：代码侧已就绪，待 5211ee90 上 CI 矩阵转绿即可合并；合并动作按规则留给人工（kernel 变更恒为人工评审档，agent 不合并不发布）。CI 出结果后我会跟进汇报。

…+ tag-stripped body (#6160) The scripted queue key was conv|sourceText.slice(0,64) over the RAW last-user text, whose ACP render tag is volatile by design: the post-fold ledger alignment (#5608-D defers restampRefsToLedger to the next successful fold) renumbers old refs (m00001 -> m00015, refs-never-reused) and raises the #5608-C time-attribute high-water mark, so the very next request re-renders the same message with time= dropped. The 64-char key slice contains the tag, so one scripted turn split into two queues the instant a fold landed and the directive queue replayed from scratch — the #2791 tests' deterministic "compress must be CALLED exactly once (got 2)" on local node-25 and the intermittent red on CI (same SHA red and green across runs; the alignment no-ops when the unified store has no visible rows yet at the post-fold scheduleSave). The key is now conv|<source ref>|<stripAcps(source).slice(0,64)>: the ref names the message occurrence (stable for the turn; a remint voids the script anyway) and keeps byte-identical prompts distinct (switch/back both send 请调用probe_tool); stripAcps (existing helper) drops the volatile wrapper. One site (answerFor) covers all six suites sharing the fake. Verified locally (pi-stable 0.83.6): both #2791 tests green — the fold run's queue continues through the fold (probe -> compress -> acp_status, one call); full pass over native-pi/dsh/opencode/delegate-notify/subagent-sessions (the two remaining local failures reproduce identically on the unmodified fake: pre-existing local-environment artifacts). Closes #6160

github-actions · 2026-11-11T02:25:17Z

📦 Merged Build Published

This merge to master was published as 0.0.1171-master.33104b1 (rolling master tag).

npm install -g billion-context@0.0.1171-master.33104b1  # this exact build
npm install -g billion-context@master  # always the newest merged build

npm view billion-context@master prints the current version at any time.

…6150 fold-point removal) Master independently bumped the kernel to 0.0.193 for the compressAtFoldPoint removal; this PR's overlap atomic-reject changes now sit on top of that, so the version must move again (max+1, same precedent as #5609 sync). Both lockfiles regenerated; diff vs master is only the kernel version lines.

… master fold-point removal (#6132/#6150), auto-merged clean (#5385) Master removed the compressAtFoldPoint feature; that removal does not overlap the outbound-prefix hook resolution, so this merge is conflict-free. The observeOutboundPrefix hook still runs on the final sent array (healedMessages).

New master content: #6132/#6150 remove the compressAtFoldPoint fold-point relocation feature (kernel + host + web + docs). Conflict resolved in src/server/prepare-openai.ts: - drop the fold-point thinking-clear block (its helper makeFoldPointReasoningDroppable no longer exists after #6132) - keep the PR's #5664 render-snapshot lane gate verbatim - RenderCaps.droppable pinned to false: with fold-point clear-thinking gone, reasoning blocks are never dropped at folds, so the render plane must account for them (previously mirrored the removed config flag compressAtFoldPointClearThinking)

… (clean auto-merge, incl. ranxianglei#6132/ranxianglei#6150 compressAtFoldPoint removal)

… not the closed unrelated #6150 #6150 is the closed 'remove compressAtFoldPoint' PR — the comment refs read as pointing at a retired feature. Fill the '#6xxx' placeholders with the issue the fix actually closes.

ework-agent added 2 commits November 10, 2026 23:23

ranxianglei changed the title ~~remove: the experimental compressAtFoldPoint feature — fold-point relocation (v1 #5471 + unreleased v2 #5949), #6132~~ [PR] remove: the experimental compressAtFoldPoint feature — fold-point relocation (v1 #5471 + unreleased v2 #5949), #6132 Nov 10, 2026

ranxianglei added bug Something isn't working p1 scope/kernel labels Nov 10, 2026

ranxianglei added the ci:full Run the FULL CI matrix (all e2e lanes + 12-leg test matrix) on this PR label Nov 10, 2026

ranxianglei mentioned this pull request Nov 10, 2026

Remove the compressAtFoldPoint feature — it reorders resent history and caused an authorship-misattribution incident in a live session #6132

Closed

fix(review): drop two orphaned imports left by the fold-point removal (…

5211ee9

…#6132)

ranxianglei mentioned this pull request Nov 11, 2026

e2e-native-pi lane flaky red since 2026-11-10: fake upstream's directive-queue key contains volatile ACP render-tag bytes, so a scripted fold replays the queue ("compress must be CALLED exactly once ... got 2") #6160

Closed

ranxianglei merged commit 33104b1 into master Nov 11, 2026
64 checks passed

ranxianglei pushed a commit to Maziar123/billion-context that referenced this pull request Nov 11, 2026

merge origin/master 33104b1 into fix/event-loop-watchdog-process-exit…

c4b73e3

… (clean auto-merge, incl. ranxianglei#6132/ranxianglei#6150 compressAtFoldPoint removal)

ranxianglei mentioned this pull request Nov 12, 2026

[PR] fix: persist ledger rows before the model call (durability-before-render) #6326

Merged

Navigation Menu

Repository navigation

[PR] remove: the experimental compressAtFoldPoint feature — fold-point relocation (v1 #5471 + unreleased v2 #5949), #6132 - #6150

[PR] remove: the experimental compressAtFoldPoint feature — fold-point relocation (v1 #5471 + unreleased v2 #5949), #6132#6150
ranxianglei merged 4 commits into
masterfrom
2026-11-10_remove-compress-at-fold-point

ranxianglei commented Nov 10, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Nov 10, 2026 •

edited

Loading

Uh oh!

ranxianglei commented Nov 10, 2026

Uh oh!

ranxianglei commented Nov 10, 2026

Uh oh!

Uh oh!

github-actions Bot commented Nov 11, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

ranxianglei commented Nov 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

这是什么

改动

预检（本地）

附：CI 唯一红 lane 的修复（#6160）

Uh oh!

github-actions Bot commented Nov 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

📦 Built Package Artifact

Latest build — install this exact version

Recent builds (commit → version)

Download artifact

Uh oh!

ranxianglei commented Nov 10, 2026

Uh oh!

ranxianglei commented Nov 10, 2026

Uh oh!

Uh oh!

github-actions Bot commented Nov 11, 2026

📦 Merged Build Published

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

ranxianglei commented Nov 10, 2026 •

edited

Loading

github-actions Bot commented Nov 10, 2026 •

edited

Loading