Skip to content

[PR] review: e98ccda9 → 5c2f24a4e delta anchor — full regression (ci:full) - #4104

Closed
ranxianglei wants to merge 1 commit into
masterfrom
2026-10-21_master-review-r2
Closed

ranxianglei wants to merge 1 commit into
masterfrom
2026-10-21_master-review-r2

Conversation

@ranxianglei

Copy link
Copy Markdown
Owner

Empty review anchor (same drill as #4003). Scope: the 76-commit / 61-file delta since the last review anchor e98ccda — merged lanes include #3983 compress-from-store, #3943 resident-eviction, #4094 conversation-recall, #4101 llm-stream attribution, #4090 egress window fuse, #4093 overflow-teaches-calibration, #4057 snapshot sig covered-bind, #4013 mysql2, external #3924/#3950/#3968, plus CI/test stabilization (#4077/#4084/#4087/#4097/#4100/#4017/#4022). Holistic deep-review + bili gate real-traffic run in parallel. Findings will be posted here. (ai: zhipuai-lb/glm-5.3)

@ranxianglei ranxianglei added the ci:full Run the FULL CI matrix (all e2e lanes + 12-leg test matrix) on this PR label Oct 21, 2026
@ranxianglei ranxianglei changed the title review: e98ccda9 → 5c2f24a4e delta anchor — full regression (ci:full) [PR] review: e98ccda9 → 5c2f24a4e delta anchor — full regression (ci:full) Oct 21, 2026
@github-actions

Copy link
Copy Markdown
Contributor

📦 Built Package Artifact

Branch: 2026-10-21_master-review-r2 (af6f615)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.4896-pr.4104.4bf41cc

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-pr4104.tgz
npm install -g package

Install instructions are refreshed on each push.

@ranxianglei

Copy link
Copy Markdown
Owner Author

(ai: zhipuai-lb/glm-5.3)

e98ccda → 5c2f24a 深度评审结论(第二轮,接 #4003)

范围:上次评审锚 e98ccda 之后的 76 commits / 61 files(+6888/−829),含 #3983/#3943/#4094/#4101/#4090/#4093/#4057/#4013 与外部 #3924/#3950/#3968 及 CI 稳定化批次。方法:三路并行深读(会话身份/估算器守卫/快照+UI+测试基建)+ 本地真模型 gate + 本 PR ci:full 全量矩阵。

总判定:无 P0,3 个 P1(已立案 #4109/#4110/#4111),10 个 P2

P1(用户可见,已立案)

  1. The negotiated recall for unowned retries never triggers on the desktop overlay (routed) lane — learn and claim use two | 无归属重试的协商召回在桌面 overlay(routed)车道上一次都不会触发——learn 和 claim 用了两把不同的 key #4109 [PR] dsh: negotiated session recall for unattributed retries (#4082) #4094 协商召回在桌面 overlay(routed)车道完全失效——claim 用上游 URL 当 key、learn 用 overlay URL,永不相等 → 桌面旗舰车道一次都不触发;测试只覆盖 raw 形态(Test Blind Spot Retrospective: Reinforcing Negative Assertion Discipline, Tightening Fake Upstream Strictness, Thin Inst | 测试盲区复盘:补反向断言纪律、假上游严格化、薄安装冒烟、形态矩阵与生命周期覆盖 #3617 缺行)。
  2. Session recall fingerprints collapse to a single value across the entire store for "pure tool-call" messages — causing c | 会话召回指纹对「纯工具调用」消息退化为全库同值——跨会话错认,B 的历史可能被记进 A 的会话 #4110 召回指纹对「纯工具调用」消息(content:null)退化为全库同值 → 低概率但后果严重的跨会话错认(B 的历史记进 A)。
  3. DSH's unattributed over-limit gate prices base64 images as text (bytes/4) — media-heavy requests are falsely rejected wi | dsh 无归属超限门把 base64 图片当文本计价(bytes/4)——媒体重的请求被假拒 422,上游本会正常接受 #4111 dsh 无归属超限门 bytes/4 把 base64 图片当文本计价 → 媒体重请求假拒 422;relay fuse 同问题明确排除媒体,宿主门没有——同一 PR 族两套口径。

P2(修复方向已写入各路报告,择要)

验证证据

完整报告存档:/home/dog/tmp/bili-1012/review-4104/{session-identity,estimator-wire,snapshot-webui-testinfra}-review.md。

建议收合顺序

P1 三条(#4109/#4110/#4111)发版前修;P2 里 #3950 模式(d) 输入侧锚与 fuse tools 字节两条随下个 PR 带上;其余开 issue 排队。

@ranxianglei

Copy link
Copy Markdown
Owner Author

#4105

yanhuojiyusheng pushed a commit to yanhuojiyusheng/billion-context that referenced this pull request Oct 21, 2026
…e full-mode stories

CI preload (legacy) flips the story under them: the two unified-store
recovery tests fail under BILI_STORAGE_UNIFIED=legacy (refused instead of
applied / ERR_SQLITE_ERROR), deterministic on every shard-1 runner
(ubuntu22/1, ubuntu24/1, win22/1 of PR ranxianglei#4104). Same class as ranxianglei#4076/ranxianglei#4077:
storage-family tests pin their mode instead of inheriting the runner default.
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

查重: 不重复 — Floor 2 第二轮深审的窗口止于锚点 5c2f24a(76 commits/61 files);#4105 合并于该锚之后(2026-10-21T05:48Z,merge commit 11ef088),不在其评审范围内。本条为对 #4105 的独立复审——结论全部基于 PR diff 与 master 代码重做,未复用 Floor 2 的任何结论(#4109/#4110/#4111 系 Floor 2 已立案的 P1,与本条无关)。

标签: dsh、scope/wiring、p2(评审锚单本身非缺陷,不打 bug;所涉 P1 发现已单独立案 #4109/#4110/#4111,不在此重复定级)

#4105 独立复审结论(删除 #726 死路冷却 + #3846 拒绝记录加 10 分钟 TTL)

总判定:方向正确、实现干净,本体可收工;发现 2 处收尾缺口(文档漂移 + 缺惯用 inert-key 警告),已开 follow-up PR #4149 修掉。

审核过程(独立产出)

发现的收尾缺口(follow-up 已修)

  1. 文档漂移:被删的键/env 仍作为活配置面写在 CONFIGURATION.md 与 CONFIGURATION.zh-CN.md(各 6 处)及 LIFECYCLE.md(1 行);
  2. 缺惯用警告:仓库对每个被删配置键都有 load-time inert 警告先例(warnInertRemovedCompressKeys 等 4 例),[PR] remove the #726 dead-end cooldown; expire #3846 local-refusal records after 10m (#4079) #4105 没给这个键补上——设过该键的用户会静默失效。
    → follow-up PR [PR] follow-up #4105: finish dead-end cooldown removal — inert-key warning + docs #4149(inert-key 警告 + 双负控测试 + 三文件文档同步;typecheck ✓、定向 21/21 绿)。

结构化八问

  1. 是什么问题:启动载荷超窗的 dsh 用户被锁成全局不可用——「上下文突然满了报错,新会话也没用」(The context window suddenly hit its limit and threw an error. Starting a new session didn't help either. dsh | 上下文突然满了报错 新会话也没用 dsh #3974 卸载级事件)。
  2. 什么原因导致的:估算尺只从成功响应学习,被本地拦截的请求永远学不到偏差(死锁);叠加两把不自愈的锁([Bug] Codex 额度中断后恢复长会话,preflight SSE 摘要为空并持续 502 #726 按请求哈希键住的 5 分钟冷却 + [Bug][preflight] Local 413 is forgotten when the baseline becomes an estimate, reopening oversized probes #3846 永不过期的拒绝记录),把单路由故障放大成全局停摆体验。
  3. 影响范围与严重程度:p1——凡启动载荷超窗的宿主车道完全不可用;触发条件特定但可达(The context window suddenly hit its limit and threw an error. Starting a new session didn't help either. dsh | 上下文突然满了报错 新会话也没用 dsh #3974 实锤)。本 PR 解除「冻结」半边;「仲裁通道」半边留 Context Window Overflow Issue Summary: Raw Requests Sent Directly to Upstream Rejected + Local Intercept Cooldown Lockin | 上下文超窗问题汇总:请求裸冲上游被拒 + 本地拦截冷却锁死会话(统一分析入口) #4079 待拍板。
  4. 怎么修复的:删除冻结机制本身(疏而非堵)+ 旁证记录加 10 分钟 TTL 衰减;重试契约改为有界 walk 重跑 + retryable:false,确定性装不下场景零上游调用。
  5. 是否有回归问题:无无意回归。两处有意变更(已披露+双向钉死):a) 病态宿主重试成本「冻结后零调用」→「每次重跑有界 walk(2–32 次 summary 调用上限)」;b) fail-fast 文案失去冷却时长句。协议形状/状态码/retryable 语义不变。
  6. 是否破坏缓存:不涉及——出站字节、缓存键、fold 状态均未触碰。
  7. 是否改了设计方向:不涉及——在既有 preflight/拒绝账本架构内收敛;与 [PR] fix(#3900): overflow 400s teach the calibration ring — the estimator learns from rejections, not only successes #4093(失败也校准尺子)互补。
  8. 是否引入了新的配置:无新增;移除 2 项(env BILI_PREFLIGHT_DEAD_END_COOLDOWN_MS + 键 network.preflightDeadEndCooldownMs)。兼容:旧配置仍可加载、键变 inert(follow-up [PR] follow-up #4105: finish dead-end cooldown removal — inert-key warning + docs #4149 补 load-time 警告,失效不再静默)。

收工判定:#4105 本体完成(已合、CI 绿);两处收尾缺口随 follow-up PR #4149 合并即闭环。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

本条转发与 Floor 4 的 #4105 链接为同一评论,独立复审结论已在 Floor 6 给出(不重复贴)。补一条 follow-up PR #4149 的进展:

  • CI 首轮在 docs-sync 门红:该门从 website/config-reference/ YAML seed 重新生成 CONFIGURATION*.md / website 文件后查漂移——上一提交手改了生成物却没动 seed,再生成把已删键复活了。
  • 已修(commit 4bbbe2d):seed(process.yaml 条目 + 两行摘要)移除死键、重新生成全部产物;本地三个守卫全绿(generate i18n ok / check-coverage ok / config-surface-guard ok)。
  • 现状:docs-sync ✅,全矩阵 22 绿/0 红,其余长测试分片照常跑。

无其他变更;CI 全绿后 #4149 即可合入。

@ranxianglei

Copy link
Copy Markdown
Owner Author

关闭:review anchor 使命完成。评审窗口 e98ccda→5c2f24a4e 两端点均已进入 master,Floor 2 第二轮深审结论已出。anchor PR 按流程不合并、评审结束即关。(孤儿 PR 扫描)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

ci:full Run the FULL CI matrix (all e2e lanes + 12-leg test matrix) on this PR dsh p2 scope/wiring

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant