Skip to content

[PR] fix: cache full compression input as block originals - #6837

Open
ZhaoJun233 wants to merge 6 commits into
ranxianglei:masterfrom
ZhaoJun233:2026-11-15_original-cache
Open

ZhaoJun233 wants to merge 6 commits into
ranxianglei:masterfrom
ZhaoJun233:2026-11-15_original-cache

Conversation

@ZhaoJun233

@ZhaoJun233 ZhaoJun233 commented Nov 15, 2026 •

Copy link
Copy Markdown

⚠️ MODEL-FACING TEXT CHANGE — restored originals retain complete source text and existing message citations.

Summary

Closes #6834.

After emergency truncation, cache block originals from the same complete input used to apply compression so full decompression can restore text omitted from the display view. Preserve the already rendered citation prefix only when its raw message ID and state ref match; the display body never replaces complete source text.

- cache(nonempty displayed messages, else compression input)
+ cache(compression input ?? displayed messages, existing owned citation metadata)

Evidence

  • Before: The raw-with-display emergency regression failed on earlier master 1ee510c; full restoration lost a middle checksum after both live views were cleared. The source-selection line is unchanged in current base e0d63bf. The first PR revision preserved source text but dropped citation prefixes, causing continuity/fork and storage-golden regressions.
  • After: 46 targeted tests pass against base e0d63bf across original caching, decompression, file restoration, continuity/fork, storage restart, goldens and effective receipt ranges. Coverage keeps exact full source text after both live views are cleared, same-message citations, foreign/quoted-tag rejection, empty display, display-only fallback, and both restoration tiers. Local typecheck and build pass. Existing storage goldens are unchanged.
  • Local synthetic evidence; Full remote CI passed on 9241bcd, including all 11 test shards: CI run.

Regression reasoning

Restored content must retain both the complete source and its existing reference so subsequent lookups resolve the same message. Only a leading displayed tag owned by the matching raw ID/state ref is copied, without duplicating an identical existing prefix or stripping raw quoted content. No new ref or token estimate is synthesized. Empty/absent source handling and tier collectors retain their existing contracts; the current-base effective-range receipt fix (#6855) is preserved.

Merge Danger

Door: two-way

Blast Radius: recovery

New caches retain complete originals and may occupy more space. Existing clipped caches are not rebuilt. Storage format, kernel behavior and release version are unchanged. Human review is required before merge.

@ranxianglei ranxianglei changed the title fix: cache full compression input as block originals [PR] fix: cache full compression input as block originals Nov 15, 2026
@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

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

查重: 不重复 — targeted one-line fix for #6834 (the contentSource selection in applyRanges); no other open PR touches that source-selection line.

Review of PR #6837 (head 8ed436ca, base 2e3cfc973, 2 files +54/-7, external author)

What it does. One line in src/stream.ts applyRanges — the block-originals cache source flips from display-first to raw-first:

  • old: ctx.messages.length > 0 ? ctx.messages : (ctx.compressMessages ?? ctx.messages)
  • new: ctx.compressMessages ?? ctx.messages
    Plus a 3-scenario regression test (tests/compress-original-cache.test.ts).

Why it's correct (consistency, not a new idea). The actual fold (applyCompression, src/stream.ts:1316 at base) and the CCR content store (storeCoveredOriginals, src/stream.ts:1470 at base) BOTH already consume ctx.compressMessages ?? ctx.messages (the raw source). Only the blockContents cache — the resolveDecompress fallback for when both live views are evicted — used display-first. After emergency truncation the display view (ctx.messages = processedMessages) has lost middle text while the raw source (ctx.compressMessages = originalMessages) is complete, so the display-first cache was silently incomplete and whole-block full decompression restored truncated text (the lost middle checksum in #6834). The fix makes the cache use the exact source that was folded.

Verification I ran (local, targeted):

  • New regression test: 3/3 pass.
  • Teeth check: reverting the one line makes raw-with-display FAIL (the bug case) while the other two pass — the test genuinely pins the fix, not a tautology.
  • 7 related decompress/cache suites (proxy-decompress-cache, storage-block-content-join, compress-from-store, decompress-empty-range, fix-decompress-reaper, fix-decompress-tofile, issue2204-decompress-to-file-outcome): 45 pass / 0 fail / 5 skip (better-sqlite3 not installed, expected).
  • tsc --noEmit: clean.
  • CI: still in progress on the pushed head (queued/in_progress, no conclusions yet) — the authoritative green is still pending.

Behavior change (flagged per policy). Old → new: the blockContents cache is now built from the raw compression input whenever it's available, instead of the display view. New caches can be larger (complete vs. truncated) and are now complete; in the non-truncation case the cached bytes are effectively unchanged (collectBlockContent reads the covered message ids, whose text is identical in both views). Storage format, kernel, wire shape, and the CCR store are all unchanged. Existing clipped caches are NOT rebuilt (acknowledged in the PR; correct scope — the issue explicitly does not rebuild old caches or invent absent source).

Gate. This is a fix: scoped to one module + one test, but it changes a source-selection fallback, which is a "must stay human" element per §7.4 — and the PR itself says "Human review is required before merge." No auto-merge.


结构化八问

  1. 是什么问题:紧急截断后,块原文缓存(blockContents)用的是"显示视图"(已被截断、丢了中间文本),而不是压缩真正消费的"完整源"。整块 full 解压时优先读这个不完整的缓存,导致中间内容(如校验和)无法恢复。
  2. 什么原因导致的:applyRanges 里 contentSource 的选取是"显示视图优先"(ctx.messages 非空就用它),但 applyCompression(src/stream.ts:1316)和 CCR 内容存储(src/stream.ts:1470)都用 ctx.compressMessages ?? ctx.messages(完整源)。三处本应同源,唯独块原文缓存用了显示视图——截断发生时显示视图缺中间文本,缓存就跟着缺。
  3. 影响范围与严重程度:跨客户端的恢复路径。触发条件=发生紧急截断 + 随后整块 full 解压 + 两个 live 视图都已被逐出。后果=该块恢复不完整(中间文本缺失);原始数据仍在客户端历史里,属 bili 侧恢复完整性问题,触发面较窄。issue Emergency-truncated display content is cached as complete compression originals #6834 建议 p1(恢复/数据丢失类),我认同该分级。
  4. 怎么修复的:把块原文缓存的源改成与 applyCompression/CCR 存储完全同源(ctx.compressMessages ?? ctx.messages),让缓存存的就是被折叠的那份完整原文;compressMessages 缺失时回退显示视图(此时确实没有更完整的源)。
  5. 是否有回归问题:逐项过——wire 消息形状(不变)、CCR 内容存储(本就同源)、applyCompression(本就同源)、存储格式(不变)、kernel(不变)。唯一变化是缓存源选取:非截断场景下 collectBlockContent 按覆盖消息 id 取文本,两视图文本一致 → 缓存字节实际不变;截断场景下缓存变完整(即修复本身)。未发现无意回归。
  6. 是否破坏缓存:不破坏前缀/重试/游标/去重。blockContents 只是 decompress 的回退源,格式不变;新缓存可能更大(完整 vs 截断),属预期。旧的不完整缓存不重建(已披露)。
  7. 是否改了设计方向:没有。只是把一处"异类"的源选取对齐到既有契约(fix(loop): run applyCompression over the unpruned view #197 未剪枝 apply 视图 / fix: stateless decompress — pure retrieval, no deactivate, cache retained (#400) #426 无状态解压缓存),没有转向。
  8. 是否引入了新的配置:没有。无新 env/config/默认值。

判定:修复正确、最小、测试有牙(老代码会红),本地定向验证全绿、typecheck 干净;CI 仍在跑(需等绿)。因涉及源选取回退(§7.4 人工档)+ 作者自标"需人工 review",不建议自动合并——待 CI 绿 + owner 人工确认后合并。

@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

🔶 需要决策

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

Diagnosis (verified live this run)

Head 8ed436ca1b is still red on 6 test shards (mirror cache, 0 pending). Master f648217cb8 is fully green (18/18), so the failures are PR-own, not inherited.

This PR's one-liner (src/stream.ts applyRanges: contentSource = ctx.compressMessages ?? ctx.messages) produces two distinct failure families:

  1. storage-render + storage-engine-proofs goldens (8 cases) — only the blockContents section drifts. This is the declared behavior change (PR body: "New caches retain complete originals"). The test prescribes regeneration via scripts/update-storage-render-goldens.ts. Mechanical — but only correct once (2) is resolved.
  2. session-continuity-proof scenarios A/B (6 cases) — decompress must return the folded original m00005. This is a real regression, not a golden: the same one-liner also switches the plugin-tool lane's original cache from the tagged processed view the model saw (mem.processed) to the raw client history (mem.original) which carries no ACP tags — so a full decompress returns originals without citable refs. This is word-for-word the regression already fixed in PR [PR] fix: build block-content original cache from the complete source view (#6834) #6852 (commit 8b054a1fd5, "fix: build plugin-lane block-content cache from the tagged view").

The decision: #6837 is a duplicate of #6852

Both PRs close issue #6834:

#6852's final state is a strict superset of #6837's change — if #6852 merges, #6837's diff is fully subsumed. Making #6837 green requires reproducing #6852's two fix commits (8b054a1fd5 + ad77ed1db6) on top of it, i.e. pure duplication.

Options

Default if undecided: leave #6837 open-but-red and let #6852 proceed to merge; owner closes #6837 as a duplicate. (Closing PRs is human-only — I can't do it myself.)

@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

七问(本次 CI 运行小结)

  1. 是什么问题:PR [PR] fix: cache full compression input as block originals #6837 的 6 个 test shard 红——storage-render / storage-engine-proofs golden 漂移(8 例)+ session-continuity-proof A/B 场景失败(6 例)。
  2. 什么原因导致的:本 PR 的一行改动 contentSource = ctx.compressMessages ?? ctx.messages(src/stream.ts applyRanges)。golden 漂移是已声明的有意变更(缓存源 display view → raw compressMessages);session-continuity-proof 失败是真回归——plugin-tool 车道的 original 缓存从带 ACP tag 的 processed 视图切到了无 tag 的 raw client history,full decompress 返回无引用 ref 的原文。
  3. 影响范围与严重程度:仅本 PR 分支(master f648217cb8 全绿 18/18,非继承)。若合入,plugin 车道的 full decompress 会坏(p1:会话恢复/数据完整性)。
  4. 怎么处理的:未直接修——本 PR 是 [PR] fix: build block-content original cache from the complete source view (#6834) #6852 的重复(两者都 close Emergency-truncated display content is cached as complete compression originals #6834;[PR] fix: build block-content original cache from the complete source view (#6834) #6852 已含 plugin 车道解耦 + golden 重生成,CI 已全绿 15 success/15 skipped/0 failed,live 核实)。让 [PR] fix: cache full compression input as block originals #6837 变绿 = 复制 [PR] fix: build block-content original cache from the complete source view (#6834) #6852 的两个修复 commit(8b054a1fd5 + ad77ed1db6),纯重复。已发「🔶 需要决策」评论(选项 A/B/C + 推荐 A:关 [PR] fix: cache full compression input as block originals #6837、留 [PR] fix: build block-content original cache from the complete source view (#6834) #6852)并打上 needs-decision 标签,停手等 owner。
  5. 是否有回归问题:本 PR 自身引入 1 处真回归(session-continuity-proof,plugin 车道 full decompress 丢引用 ref);未合入,master 无影响。
  6. 是否破坏缓存:本 PR 改变 blockContents 缓存源,新缓存更大(已声明);不破坏既有缓存(旧 clipped 缓存不重建)。
  7. 是否改了设计方向:否——未改架构/边界,仅缓存源选择;但「保留哪个 PR 修 Emergency-truncated display content is cached as complete compression originals #6834」属 owner 决策,已升级。

判定:非机械可修(重复 PR + 真回归),按 owner policy 2026-11-01 升级「🔶 需要决策」+ needs-decision 标签,停手。推荐 A:关 #6837、合 #6852。

@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

[ework] CI fix verified green ✓

@ZhaoJun233
ZhaoJun233 force-pushed the 2026-11-15_original-cache branch from 8ed436c to 9241bcd Compare November 16, 2026 03:29
@ZhaoJun233
ZhaoJun233 marked this pull request as ready for review November 16, 2026 03:49
@ZhaoJun233

Copy link
Copy Markdown
Author

逐项回应初次评审、回归及重复关系诊断、七问小结和后续 CI 通知:

  1. 旧提交的失败判断成立。 8ed436ca1b 丢失可引用 ref 是真实回归,不能通过更新 golden 或忽略 plugin 路径解决。原先“普通场景缓存字节不变”的说明不完整,因为 raw 与 processed 在引用前缀上确实不同。
  2. 当前修复同时保留完整正文和引用。 当前提交 9241bcd1b1f4fdf515263a043ce71ea64164ab7c 仍以 compressMessages ?? messages 为正文来源;只从同一 raw message ID 的显示消息复制开头已有的引用前缀,并要求前缀 ref 等于 state 中该 raw ID 的 ref。显示正文不会覆盖完整正文;错误归属、正文中间的引用不被采纳,源对象不被修改。实现同时供 full:true 和 full:false 两个 collector 使用。
  3. 没有通过更新 originals golden 掩盖回归。 本 PR 的四协议 storage-render golden 全部保留原值。恢复引用前缀后,原来的 storage-render、storage-engine-proofs 和 session-continuity-proof 断言全部通过。因此,当时的 blockContents 漂移不能一概视为必须接受的预期变更。
  4. 验证覆盖。 定向 46 项通过,包含真实紧急截断、两个 live view 清空后的恢复、完整正文与已有引用同时保留、引用归属反例、模型切换/分叉连续性、存储重载及未改动的 goldens。新增回归测试没有放宽原有连续性断言。
  5. 与 [PR] fix: build block-content original cache from the complete source view (#6834) #6852 的关系。 两份 PR 解决同一个 Emergency-truncated display content is cached as complete compression originals #6834,合入时应择一协调,这一点仍成立。不过当前实现不同:核对 [PR] fix: build block-content original cache from the complete source view (#6834) #6852 的 ad77ed1db6e36599a3b7107e35de5d7420b8e8b4 后,它新增 blockContentSource,由 plugin 传入 processed view,并更新部分 storage goldens;本 PR 保持完整源正文,只补回属于同一消息的已有引用,没有新增 context 字段,也没有更新 goldens。因此,“只能复制 [PR] fix: build block-content original cache from the complete source view (#6834) #6852 的两个修复提交才能变绿”以及对旧一行版本的严格包含判断,不能继续用于当前提交。这里也不据此断言 [PR] fix: build block-content original cache from the complete source view (#6834) #6852 存在未经验证的运行缺陷。
  6. 旧缓存及存储边界。 只保证新建缓存使用当时可用的完整源;没有 raw view 时仍按既有 fallback 工作,不补造不存在的正文或引用。旧的 clipped cache 不重建。存储格式、kernel 和 CCR store 没有改动,新缓存可能因保留完整正文而更大。
  7. 可复核的最终 CI。 当前提交已 rebase 到 e0d63bf8c85d37f3e0073ae71dad6997f01c0876;该提交的完整 CI通过,11/11 个 test 分片成功,类型检查和构建通过,PR 已转为 ready for review。这里明确给出 SHA 和 run,避免把旧 head 的红灯或未注明 SHA 的绿灯通知当作当前结论。
  8. 决策项。 评论中提出的代码回归已修复;剩下的是维护者选择哪份最终实现作为 Emergency-truncated display content is cached as complete compression originals #6834 的合入载体。建议比较两份当前 diff 后择一,不重复合入同一修复;我没有关闭另一份 PR,也没有自动合并。

@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

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

CI 状态(重新核验,mirror cache 2026-11-16T04:03:28Z): 新 head 9241bcd1b1 全绿——17 success / 13 skipped / 0 failed,mergeable_state=blocked(来自 needs-decision 标签 + 人工评审门,不是 CI 红)。这条 auto-fix lane 的红灯已解除,CI 侧的活到此完成。

纠正我上一轮的一个判断: 我此前说「#6837 是 #6852 的严格子集、只能照抄 #6852 两个提交才能变绿」——对当前提交不再成立。我读了新 diff,确认两份现在是不同的实现路径:

#6837(当前 9241bcd1b1) #6852
正文源 两条车道都用完整 raw(compressMessages ?? messages) relay 用 raw、plugin 用 processed/display view
引用 从同一 raw ID 的显示消息开头复制引用前缀(ref 须匹配 byRaw[id]) plugin 车道的 processed view 自带引用
新 ctx 字段 无 新增 blockContentSource
goldens 未动 更新部分 storage goldens

我核对了三点:① goldens 确实未动(提交只改 src/stream.ts + 新增 tests/compress-original-cache.test.ts,与你的声明一致);② session-continuity 回归确实被修(full decompress 现在返回带可引用 ref 的原文,CI 全绿印证);③ 引用归属防护到位(新测试覆盖 owned/foreign/非开头 tag + 源对象不修改,代码有 byRaw 匹配 + startsWith 两道闸)。

决策项更新: 代码回归已修、CI 已绿。剩下的是「#6834 的合入载体选 #6837 还是 #6852」——两份现在都是绿的,区别在实现路径与 #6834「完整正文」的覆盖面:#6837 两条车道都用完整 raw 正文,#6852 的 plugin 车道用 processed view。后者对 plugin 车道的完整正文覆盖是否等价,取决于紧急截断是否作用到 plugin 的 display view——这一点我未验证,留给 owner/reviewer 判断。代价上 #6837 更绕(per-message map + 正则 + 三道闸),#6852 更简(一个字段、无正则)。

我维持 needs-decision 标签,不关闭任何一份 PR(关闭是 human-only)。建议 owner 比较两份当前 diff 后择一,不重复合入同一修复。


八问(本轮 CI 修复):

  1. 是什么问题:PR [PR] fix: cache full compression input as block originals #6837 六个 test 分片红——storage-render / storage-engine-proofs 的 golden 漂移 + session-continuity-proof 的真实回归(full decompress 丢可引用 ref)。
  2. 什么原因:旧 head 8ed436ca1b 一行改动把 block-content 原文缓存的正文源从「非空 display view」换成「raw」,plugin 车道的 raw 无 ACP tag → full decompress 返回无引用原文。
  3. 影响范围与严重程度:plugin 车道 full decompress 折叠后丢可引用 ref(可用性缺陷,p1 级);relay 车道另受 Emergency-truncated display content is cached as complete compression originals #6834 完整正文问题影响。
  4. 怎么修复的:本轮我未改代码——贡献者推了新提交 9241bcd1b1(正文源保持完整 raw,只把同一 raw ID 的显示消息开头引用前缀复制回 raw,ref 须匹配 byRaw),我核验 CI 全绿 + goldens 未动 + 回归被修。
  5. 是否有回归:无新回归——goldens 未动、原有断言(storage-render / storage-engine-proofs / session-continuity)全过、新测试覆盖引用归属反例。
  6. 是否破坏缓存:不破坏。存储格式 / kernel / CCR store 未动;新缓存可能因保留完整正文而更大(贡献者声明,属预期)。
  7. 是否改了设计方向:否。仍是「折叠时缓存原文供 decompress」的既有机制,只是正文源选择更精确。
  8. 是否引入新配置:否。无新 env / config / 默认值。

收工判定: CI 红灯已解除(全绿),回归已修复;「#6834 合入载体择一」是 owner 决策项,needs-decision 标签保留,等 owner 拍板。

ZhaoJun233 and others added 5 commits November 16, 2026 17:32
Backport the wire-exit-matrix slice of upstream 297709e. The 400ms fallback could abort before forwarding, leaving requests=0 and closed=0 on Windows CI. Refs ranxianglei#4599.

Validated on Windows with Node 22.23.4: the delayed-dispatch scenario reproduces the original assertion before the fix; all four delayed abort cases pass after it. The original wire-exit and compress-original-cache suites pass 12 tests. Typecheck and build pass.
@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

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

冲突已解决并推送,PR 恢复可合并状态。

本轮处理:PR #6837 与 master 的合并冲突(merge state: dirty)。

  1. 拉取 PR 头 6a3531aa4a(贡献者 fork ZhaoJun233/billion-context-contrib 的 2026-11-15_original-cache 分支),把当前 master 6a641f3785(领先 PR 基线 301e1a5459 共 26 个提交)合并进 PR 分支——沿用 PR 既有的 merge 方式(分支里已有 merge commit),不 rebase、不 force-push。
  2. 唯一冲突 tests/issue2929-codex-fold-double-miss.test.ts:PR 分支保留硬编码请求下标锚点(PROXY_TOOL_AT=6/FOLD_AT=7)并加了 nudge 步长固定(NUDGE_GROWTH_TOKENS + minGrowthFloor/minGrowthRatio);master 已改成调度无关方案(fold.state/armIdx/foldIdx,注释明确说明「硬编码请求下标锚点每次 nudge 门 token 调度漂移都会断」)。自动合并的测试主体已是 master 的调度无关版本,PR 的硬编码锚点因此变成死代码 → 取 master 版本(更稳健,且是 master 当前绿灯版本;PR 的脆弱锚点方案已被 master 取代)。
  3. 本地验证:npm ci + npm run typecheck(含 kernel/pi-subagents 构建)通过;定向测试 tests/compress-original-cache.test.ts(Emergency-truncated display content is cached as complete compression originals #6834 修复,4 项)+ tests/issue2929-codex-fold-double-miss.test.ts(冲突解决后的 codex 命中率问题分析 #2929 探针,1 项)全部通过(5/5,探针输出「re-request 与 next turn 在 summary 锚点处字节一致」)。
  4. 推送合并提交 6b70ba6ba2("Merge branch 'master' into 2026-11-15_original-cache",无 [ework:ci-held] 标记——这是最终 head)到 fork 的 PR 分支。
  5. 验证 PR 状态:mergeable: True(冲突已清除),mergeable_state: blocked(因 needs-decision 标签 + 人工评审门禁,非 CI 红或冲突)。

CI 状态: 新 head 6b70ba6ba2 的 CI 已触发,镜像缓存显示 13 个 check 排队中(尚未完成)。按状态新鲜度政策,我不在此断言 CI 绿/红——以远程 CI 在推送 head 上的结果为准。

七问:

  1. 是什么问题:PR [PR] fix: cache full compression input as block originals #6837 与基线分支 master 产生合并冲突(merge state: dirty),阻塞合并。
  2. 什么原因导致的:master 自 PR 基线前进了 26 个提交;其中 tests/issue2929-codex-fold-double-miss.test.ts 在两侧都有改动——PR 分支保留硬编码请求下标锚点并加了 nudge 步长固定,master 改成了调度无关方案,两者在同一区域冲突。
  3. 影响范围与严重程度:仅影响本 PR 的可合并性(dirty 状态阻塞合并);无用户侧影响,无数据/行为影响。
  4. 怎么修复的:把当前 master 合并进 PR 分支(沿用 PR 既有 merge 方式,不 rebase、不 force-push);唯一冲突取 master 的调度无关版本(PR 的硬编码锚点已被 master 取代且变成死代码)。
  5. 是否有回归问题:无。冲突解决取的是 master 当前绿灯版本;本地 typecheck + 定向测试(Emergency-truncated display content is cached as complete compression originals #6834 修复 + codex 命中率问题分析 #2929 探针)全绿;自动合并的其余文件均来自 master(已绿)。
  6. 是否破坏缓存:不涉及(无缓存/重试/游标逻辑改动)。
  7. 是否改了设计方向 / 引入新配置:不涉及(纯冲突解决,无新配置、无架构转向)。

判定: 冲突已解决,PR 恢复可合并状态;CI 在新 head 上运行中,待其绿后由 owner 决定 #6834 的合入载体(#6837 vs #6852,needs-decision 标签保留)。

@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

[ework] CI fix verified green ✓

2 similar comments
@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

[ework] CI fix verified green ✓

@ranxianglei

Copy link
Copy Markdown
Owner

🤖 Powered by ework · qwen3.8-27b

[ework] CI fix verified green ✓

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Emergency-truncated display content is cached as complete compression originals

2 participants