Skip to content

[PR] fix: project dsh-mnemon memory snapshots to durable/mnemon-memory (#4495 part A) - #4497

Open
ranxianglei wants to merge 1 commit into
masterfrom
2026-10-24_4495-mnemon-durable-classifier
Open

ranxianglei wants to merge 1 commit into
masterfrom
2026-10-24_4495-mnemon-durable-classifier

Conversation

@ranxianglei

@ranxianglei ranxianglei commented Oct 24, 2026 •

Copy link
Copy Markdown
Owner

Related: #4495 (part A — the folding half; the truncate half (B) awaits the owner decision requested in that thread).

What

dsh-mnemon injects its runtime memory snapshot — the user profile (USER.md) + working reference (MEMORY.md), the one piece of context that should be ALWAYS present — as an ordinary user/text message starting with "MNEMON RUNTIME MEMORY SNAPSHOT". It matched none of the four durable classifiers, so once it left the recent zone it was ordinary foldable history. The plugin re-injects only when the rendered text CHANGES (injectedMemoryText equality gate), so a stable profile is injected ONCE and never again: one fold silently killed the session's persona for the rest of the session.

Change

Behavior change (disclosure)

dsh-lane messages matching the mnemon snapshot header: foldable → protected (latest instance pinned per identity; older superseded instances remain foldable). All other lanes and all other messages are untouched. User takeover of the durable policy (any durable pattern in protectedTools/protectedLatestTools) still wins verbatim.

Tests

New tests/issue4495-mnemon-durable.test.ts (4 tests): unit projection (positive / negative / role-gate / lane registry stays dsh-only); e2e fold executed through /__bili/plugin/tool whose range positionally covers BOTH snapshots — newest survives, superseded folds away (the reported bug); unregistered-lane negative control (both fold away); BLOCKED-ref stamping witnessed at assignment time on turn one. Sibling suites green locally: durable-tool-identity, issue2446-dsh-mcp-catalog, issue2766-durable-role-gate (24 tests). Full-suite verdict = CI on this head.

Not included

Part B (emergency truncate honoring protection) — kernel-side, both variants await the owner's narrow/general choice requested in #4495.


Fixes #4495

(孤儿认领:标题本已指向 #4495,补上关闭关键字 —— 第三批清理)

@ranxianglei ranxianglei changed the title fix: project dsh-mnemon memory snapshots to durable/mnemon-memory (#4495 part A) [PR] fix: project dsh-mnemon memory snapshots to durable/mnemon-memory (#4495 part A) Oct 24, 2026
@github-actions

github-actions Bot commented Oct 24, 2026 •

Copy link
Copy Markdown
Contributor

📦 Built Package Artifact

Branch: 2026-10-24_4495-mnemon-durable-classifier (822052a)

Latest build — install this exact version

npm install -g billion-context@0.0.9842-pr.4497.fcff393

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
822052a billion-context@0.0.9842-pr.4497.fcff393 build
02e2340 billion-context@0.0.9732-pr.4497.2f35a83 build
9b00d0a billion-context@0.0.5535-pr.4497.8551bd7 build
104a547 billion-context@0.0.5495-pr.4497.e0fe89d build
05d9764 billion-context@0.0.5484-pr.4497.9d4fd70 build
2c56f42 billion-context@0.0.5455-pr.4497.7c292dd build
d8b2fbf billion-context@0.0.5441-pr.4497.4657c35 build
ef5f141 billion-context@0.0.5423-pr.4497.fd09152 build
543fc9f billion-context@0.0.5420-pr.4497.397ab17 build
6998998 billion-context@0.0.5413-pr.4497.14f3416 build

Download artifact

  1. Download the artifact from the Actions run
  2. Extract the tarball and install:
tar xzf billion-context-pr4497.tgz
npm install -g package

Install instructions are refreshed on each push.

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

查重:不重复。 本 PR 是 #4495 Part A(折叠半边)的交付;按 mnemon / durable / truncate 检索全部 issue 与 PR(含已关闭),无同机制的其他条目。截断半边(Part B)仍在 #4495 线程里等你拍板,未另立项。

标签:bug、dsh、scope/kernel、p2(dsh 车道折叠判定面的缺陷修复 PR;仅长会话触发,能用但有明显问题,非数据丢失/缓存泄露类)。

审核结论(独立审了这份 diff,未复用 #4495 侧表述)

代码(src/durable-message-guards.ts,+34/-4,单文件):

  • dshMnemonMemoryPath(src/durable-message-guards.ts:169-173)与四个既有分类器同构:durable 身份接管缺口:classifier 无角色边界,模型回声可夺取 durable/* latest 身份 #2766 角色门在前(assistant 回显永不投影)、contentType text、typeof string 守卫、按头行匹配 marker——布局在下面怎么改都拆不掉这个保护;
  • 接在 dsh 分类链末尾做第五个载体(src/durable-message-guards.ts:177-185),链序不变,五个 marker 字符串互不相交、无身份冲突;lane 注册表仍 dsh-only(KDD release v0.1.2 #9 证据白名单纪律保持);
  • 接线点零改动:src/session.ts:676-689 与 src/server/handle.ts:2214-2215 的 withDurableProtection(config, durableMessageToolPaths[lane]) 原样;用户 durable 策略接管语义(classifiesDurable 整身份接管)完整保留;
  • 无 kernel 改动、无配置面改动、无 wire 协议改动。唯一已知家族内权衡:用户在 user 通道逐字引用头行文本会吃到 latest-per-identity 钉住——这与四个既有载体的语义完全一致,真实暴露向量(模型回显)已被角色门挡掉。

测试(tests/issue4495-mnemon-durable.test.ts,4 例):断言有牙——e2e 用真实 server rig 执行真 compress,区间按位置同时盖住两份快照,断言最新一份存活、被取代的那份折掉、区间内填充消息消失;没有这个修复该断言直接失败,是真回归钉。负向对照齐备(散文不含头行 / 仅内部标签 / skill-catalog 区分 / 非 text / 缺 text / assistant 回显 / 未注册 lane 全折)。本地定向验证:本套件 + durable-tool-identity + issue2446 + issue2766 共 28/28 全绿。

CI 红排查(重点):首推 head 上 10 个 test 分片红了 7 个,但每个分片恰好只挂 1 条,且各不相同,集中在存储域两个时序敏感用例——#3250 …journal_size_limit(WAL autocheckpoint 截断边界)与 media CAS stale-row sweep;最坏分片其余 3499/3510 全绿,静态门禁(kernel-guard/config-surface-guard/docs-sync/static-gates 等)与全部 e2e 车道全绿。三点佐证与 PR 无关:① master 自己在 05:53 的 CI(早于本 PR 08:11 的运行)同样红了 5 个 test 分片、同一分片级模式——master 本来就是红的;② 这两个测试文件在本 head 上本地跑 8/8 过;③ diff 完全不碰存储路径。结论:CI 红是 master 既有的环境性 flake,不是本 PR 引入。

已做的动作:把 PR 分支 rebase 到当前 master(无冲突,单 commit 重放 → 543fc9f)并推回分支,CI 正在合并树上重跑。

八问

  1. 是什么问题:dsh-mnemon 注入的用户画像快照是普通 user/text 消息,不命中任何 durable 分类器,离开近期区后被普通折叠吞掉;插件只在文本变化时重注、稳定画像只注入一次,折一次 = 当次会话人设永久消失。
  2. 什么原因导致的:durable 分类链是 KDD release v0.1.2 #9 的证据白名单,mnemon 作为第五个宿主载体缺条目——白名单少一行,不是保护策略本身失效。
  3. 影响范围与严重程度:仅 dsh lane + 装了 dsh-mnemon 的长会话;触发后画像当次会话内不可恢复地丢失(静默失败,模型和用户都无感知信号)。p2。
  4. 怎么修复的:给该载体补上身份投影(头行 → durable/mnemon-memory),交给既有保护机制裁决——默认 durable/* latest-per-identity 自动只钉最新一份快照,被取代的旧份照常可折;受保护消息 assignRefs 时拿 BLOCKED_REF,结构上 target 不到。复用而非新造通道。
  5. 是否有回归问题:逐项审过既有行为——四个原有载体的投影(链序保持)、lane 注册表(仍 dsh-only)、用户 durable 策略接管、durable 身份接管缺口:classifier 无角色边界,模型回声可夺取 durable/* latest 身份 #2766 角色门,全部保持。唯一有意变更已在 PR 正文披露:dsh lane 匹配头行的消息从可折变受保护(latest-per-identity)。无阈值/时序/输出格式漂移。
  6. 是否破坏缓存:不涉及——保护只影响折叠资格判定,不改出站字节布局。
  7. 是否改了设计方向:没有——完全沿用「宿主投影虚拟路径、kernel 按 pattern 保护」的既有架构。
  8. 是否引入新配置:无。

判定

代码与测试通过审核;CI 红为 master 既有 flake(证据如上),rebase 后重跑转绿即可合入。Part B(紧急截断不看保护标记)仍等你拍板:B-narrow(kernel 加 mnemon 专属 excludeIds)/ B-general(截断跳过一切当前受保护消息,我推荐)/ 折中(只跳 text 通道 durable/* 身份)——定了我在 #4495 线程开独立 kernel PR。A 本身闭环,不依赖 B,可先行合入。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

CI 重跑结果 + 完整取证(rebase 后 head 543fc9f,run 40079560367):

主 CI 仍红,但 6 个红分片已逐一取日志核实——全部失败都是 master 既有的存储域时序 flake,与本 PR 零关系:

flaky 用例 本轮红分片 跟踪状态
#3250 …journal_size_limit(WAL 截断边界) ubuntu 24-1 / windows 22-2 / ubuntu 22-1 已有 #4489 + 修复 PR #4502 在途;master 自身红 run(40068050438,早于本 PR)错误字节完全一致
media CAS: stale-row sweep ubuntu 24-3 / ubuntu 22-3 / windows 22-1 无既有跟踪 → 已按「同失败域 = 同 issue」并入 #4489 检查清单(含两轮 run 的分片证据),未另立项
#1365 apply() attach mode windows 22-1 同上,并入 #4489

交叉证据:① master 自身红 run 与本轮红分片集合高度重合、not-ok 行全部落在上述三条内,无任何非存储域失败;② 本 PR 新套件 tests/issue4495-mnemon-durable.test.ts 在 CI 上全绿(ubuntu 22-2 分片 ok 1305/1306/1307/1308,该分片 0 失败);③ 两个主要 flaky 文件在本 head 本地 8/8 过。另:rebase 时顺带把分支从 fe5ba59 基线提到了当前 master(02f544da0),无冲突、单 commit 重放。

八问(本轮增量;完整版见上一条回复,各项结论未变)

  1. 是什么问题:不变——dsh-mnemon 画像快照缺 durable 分类器被折叠吞掉。
  2. 什么原因导致的:不变——KDD release v0.1.2 #9 证据白名单缺第五个载体条目。
  3. 影响范围与严重程度:不变——仅 dsh lane + dsh-mnemon 长会话,p2。
  4. 怎么修复的:不变——补身份投影交既有保护机制裁决,复用而非新造。
  5. 是否有回归问题:不变——四个原有载体/lane 注册表/用户接管/durable 身份接管缺口:classifier 无角色边界,模型回声可夺取 durable/* latest 身份 #2766 角色门全部保持;唯一有意变更已在正文披露。新增:CI 重跑证实本 PR 测试在 CI 环境全绿,无新增回归信号。
  6. 是否破坏缓存:不涉及(不变)。
  7. 是否改了设计方向:没有(不变)。
  8. 是否引入新配置:无(不变)。

判定

代码与测试通过审核,CI 红为 master 既有存储域 flake(证据链见上 + #4489 线程)。建议路径:先合 #4502 消掉最高频签名,等本 PR 一次自然绿 run 后合入;若你认可取证结论也可直接合。合并动作照旧由你执行。Part B(紧急截断不看保护)仍等你在 #4495 拍板 narrow / general(我推荐)/ 折中。

@ranxianglei
ranxianglei force-pushed the 2026-10-24_4495-mnemon-durable-classifier branch from ef5f141 to d8b2fbf Compare October 24, 2026 10:31
ranxianglei pushed a commit that referenced this pull request Oct 24, 2026
…d cannot misfire on cache-restored test legs

The #4360 pretest cache restores kernel/dist byte-exact (marker-verified)
with the SAVED mtimes, while the freshly checked-out sources are newer.
On the merge ref (which carries bff4d6d's mtime guard), every leg whose
shard runs tests/golden-storage-render.test.ts or
tests/storage-engine-proofs.test.ts then fails its first chat-wire drive
with 'node_modules/acp-kernel/dist is older than kernel/src (#3325)' on
perfectly correct bytes — 6/8 shards red on PR #4497 head 2c56f42
(run 40087540317) and on master itself.

A marker-verified HIT proves the bytes are correct for this source hash,
so the timestamp is refreshed to now on that branch only: the guard keeps
its teeth on genuinely stale workspaces (they still miss and rebuild),
and no rebuild is skipped that wasn't already proven correct.
@ranxianglei
ranxianglei force-pushed the 2026-10-24_4495-mnemon-durable-classifier branch from 05d9764 to 104a547 Compare October 24, 2026 13:07
ranxianglei pushed a commit that referenced this pull request Oct 24, 2026
…d cannot misfire on cache-restored test legs

The #4360 pretest cache restores kernel/dist byte-exact (marker-verified)
with the SAVED mtimes, while the freshly checked-out sources are newer.
On the merge ref (which carries bff4d6d's mtime guard), every leg whose
shard runs tests/golden-storage-render.test.ts or
tests/storage-engine-proofs.test.ts then fails its first chat-wire drive
with 'node_modules/acp-kernel/dist is older than kernel/src (#3325)' on
perfectly correct bytes — 6/8 shards red on PR #4497 head 2c56f42
(run 40087540317) and on master itself.

A marker-verified HIT proves the bytes are correct for this source hash,
so the timestamp is refreshed to now on that branch only: the guard keeps
its teeth on genuinely stale workspaces (they still miss and rebuild),
and no rebuild is skipped that wasn't already proven correct.
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

CI 自动修复回合报告(分支已更新到 104a547f1,基于当前 master tip)。三轮红全部查清:没有一轮是这个 PR 的代码引起的——merge-ref 车道继承的是「master 一边动一边红」的三批独立 facet。

本分支已修(带标注 commit,自包含可审)

  1. 存储测试红(run 40081428171,6 片):issue3250-wal-checkpoint custom-cap 与 media-cas stale-row sweep 两测试在 base 上就红(Four storage-layer defects: identity hash contaminated with positional metadata (truncation triggers identity swap) / sw | 存储层四缺陷:身份哈希掺位置元数据(截断即换血)/清扫删除回退历史/渲染接缝图片只比长度/fork 谱系死列 #4390 时代遗留)→ 上游已由 [PR] fix(#4506): release media refs on historian retire + budget the #3250 custom-cap slack (#4390 follow-ups) #4511/CI fully red (on master itself): #3250 WAL cap assertion failing across all platform shards — suspected that this mornin | CI 全红(master 自身):#3250 WAL cap 断言在所有平台分片失败,疑似今晨合入的存储层改动破坏 autocheckpoint 截断行为 #4506 修复 → rebase 到绿 master 解决,本地复跑 8/8 过。
  2. docs-sync 红(run 40087540317):[PR] feat: plugin-lane server-owned history — msgid pinning + transcript render endpoint | 插件车道服务端正史:msgid 钉身份 + 正史渲染端点 #4505 手写了 config 文档行但漏了 seed 条目 → 当时 cherry-pick 了现成修复;上游随后以 docs-sync has been all red on master since the merge of #4505: the rollout.features.pluginLaneServerHistory seed entry i | docs-sync 自 #4505 合并起在 master 全红:rollout.features.pluginLaneServerHistory 种子条目丢失 #4520 落了同一内容,本轮 rebase 自动丢弃了该冗余 commit(现在 diff 更干净)。
  3. kernel-dist mtime 守卫红(run 40087540317,6 片 chat-wire golden):根因 = CI Windows test lane takes 20+ minutes: every leg does a full rebuild of kernel/pi-subagents dist — the sharding is way | CI windows 测试车道 20+ 分钟:每条腿都全量重建 kernel/pi-subagents dist,分片也太粗 #4360 预构建缓存恢复保留归档时的旧 mtime × [PR] test: golden drive refuses to run on a stale kernel dist (#3325 lesson) #3330 新增的 mtime 新鲜度守卫——缓存命中跳过重建后,守卫把「字节正确的缓存恢复」误判为过期,每进程首个 drive 必炸(armed-once 语义解释了为何只有 chat wire 挂)。修复:scripts/ci-pretest-cache.mjs 在 marker 验证过的 HIT 分支对恢复出的 dist 刷新 mtime(不重建、不改守卫语义),+ 回归钉 tests/ci-pretest-cache.test.ts(单测 8/8;并用 master 真守卫函数本地复现 stale→true、刷新→false 完整因果链)。该机制即 CI test shards all red since today 10:31Z: kernel-dist freshness guard rejects cache-restored dist as stale — every merge blocked #4541——已在该 issue 留言交叉引用,本 commit 是现存唯一修复,可直接 cherry-pick。

剩余红(master 侧,不在本 PR diff 内处置)

rebase 后的新 head 再撞两批当前 master tip 仍红的 facet:

  1. sidScope 未定义(src/agent/dsh-native.ts:1287,df3cd929f/Plugin lanes: pin identity to client msgids, render outbound from the server-owned history | 插件车道服务端正史:消息身份钉在客户端 id 上,出站前缀由正史渲染 #4491 引入):static-gates 双 OS 编译错 + 各分片 dsh llm/stream 相关 5 个测试运行时 ReferenceError。已有追踪:master typecheck red after #4525 merge: stale sidScope use in dsh-native.ts (fixed in PR #4522) #4550,修复 PR [PR] fix: ingestSession fetches only block-covered rows for seq bounds (#4512) #4522 在途——不重复劳动,等它合入。
  2. BILI_CODEX_STATE_MODE 未登记 LIFECYCLE.md(b7fcd2d/Bili Codex's startup mechanism has various issues that currently rely on band-aid patches. We need to explore a new mechanism to thoroughly resolve these problems. | bili codex 启动机制带了各种问题需要狗皮膏药打补丁,探索新机制彻底解决这些问题 #2222 引入读取时漏登记):lifecycle-ledger-guard 恒红。此前无追踪,已立 CI lifecycle-ledger gate red on master: BILI_CODEX_STATE_MODE read by src/ but missing from LIFECYCLE.md (#2222 follow-up) #4553(带来源标记;一行级修复,§1/§2 归属待 owner 定夺)。
  3. win22s5 分片约 5 分钟死于 responses-ws 测试中途、无 TAP 汇总无错误输出——形态像 runner 侧 OOM/kill 的环境性抖动,重跑确认。

八问

  1. 是什么问题:PR 代码无缺陷;红来自 master 移动过程中先后落地的三批独立缺陷被 merge-ref 车道继承。
  2. 什么原因导致的:facets 1-2 = 历史遗留红与文档 seed 漏项(均已被上游修复,facet 2 曾需本分支代垫);facet 3 = 两个近期优化(CI Windows test lane takes 20+ minutes: every leg does a full rebuild of kernel/pi-subagents dist — the sharding is way | CI windows 测试车道 20+ 分钟:每条腿都全量重建 kernel/pi-subagents dist,分片也太粗 #4360 缓存 × [PR] test: golden drive refuses to run on a stale kernel dist (#3325 lesson) #3330 守卫)交互打架,缓存命中路径漏掉时间戳物化;facets 4-5 = 新合入功能漏了符号定义/台账登记。
  3. 影响范围与严重程度:舰队级——4/5 清掉前任何 PR(含发版)无法全绿;用户可见行为不受影响。
  4. 怎么修复的:分支内能闭环的闭环(rebase + mtime 守卫小修 + 回归钉);master 侧的走既有追踪(master typecheck red after #4525 merge: stale sidScope use in dsh-native.ts (fixed in PR #4522) #4550→[PR] fix: ingestSession fetches only block-covered rows for seq bounds (#4512) #4522)或新立(CI lifecycle-ledger gate red on master: BILI_CODEX_STATE_MODE read by src/ but missing from LIFECYCLE.md (#2222 follow-up) #4553),不把别人的缺陷塞进本 PR diff。
  5. 是否有回归问题:无。本分支净变更仅 ci-pretest-cache.mjs 的 HIT 分支 + 一个测试文件;MISS/重建路径、test: rebaseline storage-render goldens after remove-absorb (#3309) #3325 守卫语义(真正过期的工作区仍会 miss 并重建)、golden 内容判定均未动。
  6. 是否破坏缓存:不涉及前缀/会话缓存;pretest dist 缓存行为保持(HIT 依旧零重建,只补时间戳)。
  7. 是否改了设计方向:不涉及。
  8. 是否引入了新的配置:不涉及。

判定:按 auto-fix lane 规则停在此处——剩余红全部在 master 侧且已有归属(#4522、#4553、win22s5 待重跑观察),不属于本 PR 的小修范畴。#4522 与 #4553 落地后对本 PR re-run(或任意一次 push)即可全绿;若需要我可以盯这两个条目并在绿后回报。

@ranxianglei
ranxianglei force-pushed the 2026-10-24_4495-mnemon-durable-classifier branch from 104a547 to 9b00d0a Compare October 24, 2026 15:12
ranxianglei pushed a commit that referenced this pull request Oct 24, 2026
…d cannot misfire on cache-restored test legs

The #4360 pretest cache restores kernel/dist byte-exact (marker-verified)
with the SAVED mtimes, while the freshly checked-out sources are newer.
On the merge ref (which carries bff4d6d's mtime guard), every leg whose
shard runs tests/golden-storage-render.test.ts or
tests/storage-engine-proofs.test.ts then fails its first chat-wire drive
with 'node_modules/acp-kernel/dist is older than kernel/src (#3325)' on
perfectly correct bytes — 6/8 shards red on PR #4497 head 2c56f42
(run 40087540317) and on master itself.

A marker-verified HIT proves the bytes are correct for this source hash,
so the timestamp is refreshed to now on that branch only: the guard keeps
its teeth on genuinely stale workspaces (they still miss and rebuild),
and no rebuild is skipped that wasn't already proven correct.
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

CI 自动修复最终回合(2/2)。分支已更新到 9b00d0a7d(rebase 到最新 master tip 2e7052077)。本轮结论:run 40096789570 的 14 个失败全部是 master 侧 facet,无一是本 PR 代码引起;其中两批已被上游修掉并被本次 rebase 带入,剩下一批已有在途修复。

失败分解(run 40096789570)

Facet 表现 归属
sidScope TS2304 + dsh 运行时 ReferenceError static-gates 双 OS + 各分片 dsh/llm-stream 测试(u24s1 4 例、u22s2 #4215、win22s5 等) df3cd92/#4491 引入;上游已修(de29698),本地 typecheck 在 rebase 后干净
BILI_CODEX_STATE_MODE 未登记 LIFECYCLE.md lifecycle-ledger-guard b7fcd2d/#2222 引入;上游已修(6932ab4/afd6287de);我上一轮立的 #4553 可视为已解决
#4260 epoch confirmation … all-missing(true !== false,tests/epoch-confirm-lifecycle.test.ts:61,foldEpochConfirmPending 未清) u22s2 等分片 仍在 master 红:在 PR base、master tip、本分支 head 三处复现一致,与本 PR diff 无关。已有追踪:#4555(born-red 于 train a5d1994)+ 在途修复 PR #4569(对齐 #3641 契约的测试期望)——查重命中,未重复立项
#4427 publish-redaction 4 子测试(win22s5) windows 分片 rebase 后本地全绿,随 master 移动已消解,无需动作

本回合动作与验证

  • rebase 到 2e7052077(窗口与本 PR 四个代码文件零重叠,干净落位);force-push 104a547f1…9b00d0a7d(lease 校验过远端真实 tip)。
  • 本地:npm run typecheck 干净;feature + CI 修复定向测试 12/12 绿;两个继承红文件按预期复现(证明归因)。
  • 插曲:工作区 node_modules 被并发活动清掉过一次,已 npm ci + kernel/pi-subagents 重建恢复,不影响分支内容。

八问

  1. 是什么问题:PR 自身无缺陷;四轮红全部来自 merge-ref 车道继承 master 侧缺陷(master 当天连续落红又连续被修)。
  2. 什么原因导致的:见上表——每批红的引入提交都已点名;本 PR 相关的唯一改动(durable 分类器 + pretest mtime 刷新)均有回归钉且本地全绿。
  3. 影响范围与严重程度:舰队级 CI 可用性——[PR] test: align epoch-confirm/escalation-latch expectations with the merged #3641 contract (#4555) #4569 合入前任何 PR 都会带一个 epoch-confirm 红分片;用户可见行为不受影响。
  4. 怎么修复的:分支内 = rebase 到最绿 tip(吸收上游已修 facet)+ 保留两个带标注的小修(feature、mtime 守卫);master 侧 = 全部走既有追踪(epoch-confirm latch semantics: train a5d1994d test born-red (all-missing) + regression of 2915 escalation-latch-recovery — owner decision needed #4555/[PR] test: align epoch-confirm/escalation-latch expectations with the merged #3641 contract (#4555) #4569),不塞进本 PR diff。
  5. 是否有回归问题:无。typecheck + 定向测试验证;mtime 刷新只作用于 marker 校验过的缓存 HIT,重建路径与 test: rebaseline storage-render goldens after remove-absorb (#3309) #3325 守卫语义不变。
  6. 是否破坏缓存:不涉及前缀/会话缓存;pretest dist 缓存行为保持(HIT 零重建)。
  7. 是否改了设计方向:不涉及。
  8. 是否引入了新的配置:不涉及。

判定(auto-fix lane 终止):剩余唯一红(epoch-confirm)需要「代码 vs 测试期望」的语义决策,不属于小修范畴,且已有 #4555 + 在途 PR #4569 承接——按 lane 规则停在此处。#4569 合入后对本 PR 任意一次 re-run/push 即可全绿,届时可直接合并。

@ranxianglei
ranxianglei force-pushed the 2026-10-24_4495-mnemon-durable-classifier branch from 9b00d0a to 02e2340 Compare November 14, 2026 01:10
ranxianglei pushed a commit that referenced this pull request Nov 14, 2026
…d cannot misfire on cache-restored test legs

The #4360 pretest cache restores kernel/dist byte-exact (marker-verified)
with the SAVED mtimes, while the freshly checked-out sources are newer.
On the merge ref (which carries bff4d6d's mtime guard), every leg whose
shard runs tests/golden-storage-render.test.ts or
tests/storage-engine-proofs.test.ts then fails its first chat-wire drive
with 'node_modules/acp-kernel/dist is older than kernel/src (#3325)' on
perfectly correct bytes — 6/8 shards red on PR #4497 head 2c56f42
(run 40087540317) and on master itself.

A marker-verified HIT proves the bytes are correct for this source hash,
so the timestamp is refreshed to now on that branch only: the guard keeps
its teeth on genuinely stale workspaces (they still miss and rebuild),
and no rebuild is skipped that wasn't already proven correct.
@ranxianglei
ranxianglei force-pushed the 2026-10-24_4495-mnemon-durable-classifier branch from 02e2340 to 822052a Compare November 14, 2026 03:20
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b)

冲突已解、PR 已收敛为纯功能改动。本轮处理如下:

做了什么

  1. 冲突核查:git merge-tree 确认 master 自 merge-base 起未触碰本 PR 的任何文件——实际无文本冲突;"dirty" 是分支落后 master 82 个提交造成的落后态,不是真冲突。
  2. rebase 到最新 master:master 期间又前进 2 个提交([PR] fix: watermark fastpath (plugin full-resend watermark fastpath) silently disabled by input release mutation | fix: 水印快路径(plugin full-resend watermark fastpath)被输入释放改动静默禁用 #6640 watermark fastpath),已 rebase 到最新 master tip(e5e1647ba0),无冲突。
  3. 丢弃 CI-fix 提交:原 PR 第二个提交 fix(ci): refresh verified-hit dist mtimes… 测的是 master 的 scripts/ci-pretest-cache.mjs mtime restamp(master 代码,与本 PR 的 mnemon 功能无关)。该测试 flaky:断言 mtimeMs >= before 过紧——Date.now() 取整毫秒,而文件系统 mtime 存 sub-ms 精度,实测 13/30 失败、delta 恒为 -0.001ms(restamp 代码本身工作正常)。且它要修的 mtime-guard 红项已随 rebase 进 master(master 的 restamp 代码已在位)。按「一个 issue 一个 scope」原则将其移除,PR 收敛为纯功能。
  4. 验证:npm run typecheck 干净(exit 0);mnemon 功能测试 4/4 通过。
  5. 推送:force-push 到 PR 分支(822052ac68,纯功能,2 文件:src/durable-message-guards.ts + tests/issue4495-mnemon-durable.test.ts)。

当前状态

  • mergeable: true(无冲突);mergeable_state: blocked(等人工 approve,正常——合并 human-only)。
  • CI 在新 head 上运行中(22 项 queued,0 failed);本地定向验证已过,全量回归以远程 CI 为准。

八问

  1. 是什么问题:PR 与 base 处于 dirty(冲突)态 + 有红项,需 CI 工程师一次处理。
  2. 什么原因导致的:「冲突」实为分支落后 master 82 提交(非真冲突);红项为继承性(notice-registry 存储默认值、mtime guard),master 已各自修复。
  3. 影响范围与严重程度:仅本 PR 的合并态;不影响运行时行为(功能改动本身是 dsh-mnemon 记忆快照的 durable 分类,已单独验证)。
  4. 怎么修复的:rebase 到最新 master 消解落后态;丢弃与功能无关的 flaky CI 测试提交,收敛为纯功能 PR。
  5. 是否有回归问题:无。功能改动只增 durable-message-guards.ts 的 mnemon 分类分支 + 对应测试,不改既有路径;typecheck + 定向测试过。
  6. 是否破坏缓存:不涉及(不触碰缓存/重试/游标逻辑)。
  7. 是否改了设计方向:不涉及(纯分类补充,无架构转向)。
  8. 是否引入了新的配置:不涉及(无新 env/config/默认值)。

遗留说明:master 的 ci-pretest-cache.mjs mtime restamp 目前无可靠回归测试(本 PR 原带的那个是 flaky 的,已移除)。如需钉住该行为,建议单独 PR 补一个非 flaky 测试(断言用容差或「近期」判定,而非 >= before)。

结论:冲突已解、PR 已收敛为纯功能、本地验证过;等 CI 全量绿 + 人工 approve 即可合并。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b)

状态更新(完整八问报告见上一条):

结论不变:冲突已解、PR 收敛为纯功能、本地验证过;等 CI 全量绿 + 人工 approve 即可合并。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[ework] CI fix verified green ✓

1 similar comment
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 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

Labels

bug Something isn't working dsh p2 scope/kernel

Projects

None yet

1 participant