Repository navigation
[PR] fix(plugin-render): keep the acp_summary carrier when a fold reaches the window tail - #4749
ranxianglei wants to merge 6 commits into
Conversation
…the window tail (#4732) A fold whose last covered row is the window tail (the ordinary post- compression state) hit pos === -1 and was skipped without pushing its anchor, while its rows were already added to coveredSeqs — the rendered prefix dropped the whole fold with ok:true. Land the anchor at the window end (snapshot-render's default) and flush it after the merge loop, which is where a pos === viewRows.length anchor can only be consumed.
📦 Built Package ArtifactBranch: Latest build — install this exact versionnpm install -g billion-context@0.0.9424-pr.4749.ea6b239This exact version is immutable and reproducible. The rolling Recent builds (commit → version)
Download artifact
tar xzf billion-context-pr4749.tgz
npm install -g packageInstall instructions are refreshed on each push. |
…lugin-render-tail-anchor
…help-merge sweep)
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) 查重: 不重复 —— 本 PR 是 #4732 的修复 PR(一 issue 一 PR,合规)。检索发现 #4746「When block coverage ends exactly at the tail of the transcript, the anchor is lost」与 #4732 同症状同机制(plugin-render 尾部折叠丢 anchor),两者关系我会在审查中一并核实并说明处置。 标签: bug / scope/wiring / p2(渲染管道缺陷;该车道 rollout 0%、暂无用户可触达,故 p2;若 初步结论:PR 分支 first-parent 历史干净——自有提交仅 1 个核心修复( |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) PR #4749 审查结论(八问) 1. 是什么问题:plugin 车道历史渲染端点( 2. 什么原因导致的:锚点定位 3. 影响范围与严重程度:仅 4. 怎么修复的:镜像兄弟实现 5. 是否有回归问题(审过的既有行为逐项):
6. 是否破坏缓存:不涉及缓存失效机制。新载体只在「刚压完」的当轮出现在尾部;下一轮 untilSeq 越过折叠末行后锚自动回到中部位置——与 D 车道既有的折叠重付成本剖面完全一致,无新增缓存面。 7. 是否改了设计方向:否。纯缺陷修复,向既有 D 车道模式收敛,无架构/边界/数据流转向。 8. 是否引入新配置:否。无新 env/config/默认值。 本地定向验证证据(全量套件以 CI 为准):
运维备注:镜像 API 打标签失败( 判定:可合——核心修复正确、最小、与 D 车道同构;行为变更已披露(尾部折叠:什么都不渲染 → 渲染 acp_summary 载体);回归双向钉住;CI 绿后即可由 owner 合并(合并权在人)。唯一前置提醒: |
…help-merge sweep)
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) 审查已完成,补充 CI 状态更新(完整八问结论见上一条回复): CI 红 = 继承性失败,非本 PR 引入。 逐 job 拉取失败日志比对:本分支 4 个红分片的全部失败用例(master bb7c249 上同名同现,无一为本 PR 特有):
处置:ci-fixes 上的 plugin-fork 测试修复( 判定维持:可合——核心修复与回归钉不变,CI 绿后由 owner 合并。 |
…help-merge sweep)
# Conflicts: # src/plugin-render.ts # tests/plugin-render-endpoint.test.ts
(ai: vllm-qwen/qwen3.8-27b) 冲突已解决并推送(attempt 1/3):master
冲突处置(2 个文件,各 2 个冲突块):
本地定向验证(全量以 CI 为准): CI 实况(镜像缓存,sha 八问:
待 owner 拍板:修复本体已在 master(#4725 已合、#4732 已关),本 PR 现在只剩回归测试增量。合入(补上 master 缺失的 #4732 双向回归钉,有价值)还是关闭(视为已被 #4725 吸收),请 owner 定——关闭权在您,我不代关。CI 分片跑完后如有红项,本车道 attempt 2/3 继续处理。 |
(ai: vllm-qwen/qwen3.8-27b) 本轮核实:「dirty」阻塞是过期信号,当前无冲突、无红项,未做任何 push。 live 验证(全部本 run 实测,非引用旧评论):
为何不 push:PR 处于 八问:
仍待 owner 拍板(上轮已报,状态不变):核心修复已经由 PR #4725 合入 master(#4732 已关),本 PR 现仅剩 #4732 回归测试增量,CI 已绿——合入(补回归钉)还是关闭(视为 #4725 已吸收),请定;关闭权在您,我不代关。 |
[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b) 独立复审结论(专职 reviewer,非照抄前序):本 PR 现仅剩回归测试增量,我独立重做了全部验证,判定可合——且建议合入而非关闭(理由见第 4 问)。 独立验证证据(全部本轮实测,非引用旧评论):
标签:bug / scope/wiring / p1(数据丢失类按 2026-10-19 规则不降档;rollout 0% 暂无人可触达,放量前必须已含此修复)。 八问:
待 owner 拍板:核心修复已在 master(#4725 已合、#4732 已关),本 PR 现仅剩 #4732 双向回归测试增量,CI 已绿、合并零冲突。我的独立判定:可合,且建议合入——master 目前有修复但没有防止复发的回归测试,本 PR 恰好补上这份缺失的钉(已实测有牙)。合入还是关闭(视为 #4725 已吸收),关闭权在您,我不代关。 |
[ework] CI fix verified green ✓ |
Closes #4732
What was wrong
POST /__bili/plugin/render(#4491 S4) suppressed a fold's covered rows (coveredSeqs.add) before locating its anchor position. When the fold's last covered row IS the window tail — the ordinary just-compressed state, no client anomaly needed —viewRows.findIndex((r) => r.seq > maxCover)returns -1 and the old codecontinued without pushing the anchor. Result: the whole fold vanished from the rendered prefix (rows suppressed AND noacp_summarycarrier), while the response stayedok:trueand the log line still counted those rows intofolded=N. Silent context loss for the model.The old comment ("the fold extends past the requested window") described a benign scenario that cannot occur here:
blockSeqsonly contains seqs resolvable inside the window, sopos === -1means exactly "fold reaches the window tail".Fix
Mirror the sibling implementation
src/server/snapshot-render.ts(at = ddefault + post-loop flush):viewRows.lengthwhen no row sits pastmaxCover;pos === viewRows.length(without it they would never be emitted).Behavior change disclosure
Rendered prefix shape on the
pluginLaneServerHistorylane (rollout 0% today): a fold reaching the window tail now renders itsacp_summarycarrier at the tail instead of nothing. No other lane or endpoint is touched; the mid-window anchor path is unchanged (pinned by the negative-direction regression test).Verification
untilSeq= full view and = the cut exactly at the fold's last row) failed before the fix withcarrier missing, passes after; negative-direction test (mid-window fold anchors as before) green both sides.npm run typecheckclean,npm run buildclean locally; full suite via CI.