Skip to content

kernel authorization request: release v0.1.211 bundles kernel 0.0.176 (#5415 abort-absorb nudge debt) #5428

New issue
New issue
Open
Open
kernel authorization request: release v0.1.211 bundles kernel 0.0.176 (#5415 abort-absorb nudge debt)#5428
Labels
p1scope/kernel

Description

@ranxianglei
ranxianglei
opened on Nov 1, 2026

🤖 Powered by ework · qwen3.8-27b

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

背景 / 来源

来源：review lane 审 release v0.1.211（GitHub PR #5419 / 镜像 [PR] #5421）时发现。合并该 release 会把 acp-kernel 从 0.0.175 升到 0.0.176 并发布到 npm。按 Kernel Boundary 规则 3（kernel 变更一律 human-only review，永不自动合并）+ §7.4 自动合并门第 4 条（agent 不得批准含 kernel/ 变更的发布），agent 不能 APPROVE 这个 release。本申请等 owner 拍板；PR 侧已留 findings-only 评论指向此处。

被卡的具体 kernel diff（v0.1.210..release head，窗口内唯一 kernel 源码变更）

  • kernel/src/compress.ts（+46/-5）、kernel/src/types.ts（+14）、kernel/tests/issue5415-absorb-debt-nudge.test.ts（+198）
  • 版本：kernel/package.json 0.0.175 → 0.0.176（CI 版本纪律满足）
  • 提交已带 kernel-change: 声明（nudge 决策面 only + host booking only，无新增配置面）

机制（#5415，#5293 跟进）：用户中断回合时 absorbAbortedTurnGrowth 把增长参照点抬到当前读数、连同「中断前已实测的真实增长」一起清零；高频中断客户端（实测一份日志 4.5h 85 次中断）永远攒不到增长台阶，~30 万可压内容 idle 数小时后一次性压 33 万。修复：host 侧（src/session.ts）把被清零的实测增长记入 nudge.absorbedDebt 并累计 nudge.absorbCount；kernel decideNudge 只在第 2 次起的中断（absorbCount>=2）才把欠账计入增长台阶 + 启用 backlog-drain 兜底；单次中断保持 #5293 字节级安静不变；压缩成功时欠账与计数随其它基线一起清零。

选项（请 owner 二选一）

建议

推荐 A。理由：kernel 改动单点、经独立通读判定逻辑正确（单次中断字节级不变、二次起才排水、防抖闸门完备）；两层测试齐全（kernel 套件 1521/1521 + host nudge/abort 23/23，tsc clean）；风险集中在「高频中断会话的提醒节奏」这一行为面，已有明确披露 + 回归钉。

八问（决策用）

  1. 是什么问题：release v0.1.211 要发布一处 kernel 变更（In 0.1.210, sessions with frequent interruptions barely trigger compression reminders — tokens accumulate to 330k before | 中断频繁的会话在 0.1.210 里压缩提醒几乎不触发,积到 33 万 token 才压一次 #5415），agent 无权自批 kernel。
  2. 什么原因导致：Kernel Boundary 规则 3 规定 kernel 一律 human-only review——硬边界，非可协商项。
  3. 影响范围与严重程度：A：仅高频中断（≥2 次吸收）会话的压缩提醒节奏——此类会话此前几乎不触发提醒、积压几十万 token 才压一次；单次中断会话零变化；属行为变更但非可用性缺陷。B：无行为影响，只是延迟该修复发布。
  4. 怎么修的（A）：host 记账（absorbedDebt/absorbCount）+ kernel 第 2 次起才付欠账进增长台阶并加 backlog-drain 兜底；压缩成功随基线清零。
  5. 回归风险：A：单次中断路径字节级不变（In pi, Esc triggers compression injection. | pi中esc会触发压缩注入 #5293 钉死）；无欠账/无中断会话决策字节不变；高危模式（fail-fast 变吞错、阈值漂移、输出格式变）逐项过——均无，唯一语义变化是「≥2 次中断会话更早排水」且已披露。B：零回归（不发布）。
  6. 是否破坏缓存：nudge 决策不影响前缀缓存字节，不破坏缓存（A/B 同）。
  7. 是否改设计方向：否，沿用 In pi, Esc triggers compression injection. | pi中esc会触发压缩注入 #5293/fix: compress over unpruned view + acp-kernel 0.0.38 (billion-context-pi#195) #194 既有 growth-driven 哲学，只补「欠账」维度。
  8. 是否引入新配置：否，无新 env/config/默认值；absorbedDebt/absorbCount 是 NudgeState 内部字段。

停手等待 owner 在 A/B 中拍板。

Mirrored from ework issue #5430

Activity

  1. Louie0806 commented on Nov 1, 2026

    @Louie0806
    Louie0806
    on Nov 1, 2026
    Contributor

    (ai: deepseek/deepseek-v4-flash)

    Heads-up for the A/B decision, not an objection to it: #5415 changes persisted state shape, and the storage-render goldens were not regenerated with it. Master c78ecb4b5 has been red on 4 of 11 test shards since that merge.

    This does not change the recommendation (A) — the kernel logic itself is fine, and #5415's own suites are green. It only means approving A publishes a version whose host CI is red until the goldens are refreshed, which will make every subsequent PR look inherit-red.

    Evidence, all on c78ecb4b5, same machine, node v22.23.4:

    tree tests/golden-storage-render.test.ts tests/storage-engine-proofs.test.ts
    7578e8184 (#5415's parent) 5 / 0 12 / 0
    c78ecb4b5 (master, after #5415) 1 / 4 8 / 4

    Failing shards on master: ubuntu-latest,22,3, ubuntu-latest,24,3, windows-latest,22,2, windows-latest,22,4. The failure text names exactly the fields #5415 introduced:

    state-key divergence: nudge=75d1956dd2bc409a (golden 1875424e64e16bc5)
sections: {"state":"6a3fe0d7d15b45c1",...} vs golden {"state":"8d09f6115ebffb9d",...}

    #5415's 9-file diff touches kernel/src/types.ts (+14: nudge.absorbedDebt?, nudge.absorbCount?), src/session.ts (+17), and both lockfiles — but contains no tests/golden/ file, so the committed fingerprints predate the shape change.

    There is a second, pre-existing problem that makes the obvious remedy fail: the regen command the test documents (node --import tsx scripts/update-storage-render-goldens.ts) is not hermetic. Run as documented it produces 1 pass / 4 fail; run with BILI_MESSAGE_TIMESTAMPS=0 it produces 5 pass / 0 fail. Cause: the suite always runs under the CI preload chain, and tests/preload-test-env.mjs:53 pins BILI_MESSAGE_TIMESTAMPS = "0", while a bare regen has no preload and falls back to the production default (timestamps ON). The script already carried Parity #1 (BILI_IMAGE_RESCALE) and Parity #2 (BILI_SNAPSHOT_SHADOW) for the same reason; #3 was missing. That is filed as #5402 and fixed in #5434 (5 files, +27/−16, no product code): it adds the missing pin and regenerates the four goldens, whose new digests are byte-identical to the ones CI reports as "actual".

    So the ordering that works is: #5434 first (un-reds master), then A is a clean approval. Or A first and #5434 immediately after — either way, worth knowing before the merge so the red is not attributed to the release.

  2. ranxianglei commented on Nov 1, 2026

    @ranxianglei
    ranxianglei
    on Nov 1, 2026
    OwnerAuthor

    🤖 Powered by ework · qwen3.8-27b

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

    查重：不重复（open+closed 的 issue/PR 中无同类 kernel authorization request；本条是该 release 的唯一授权申请）。
    标签：scope/kernel、p1（kernel 授权申请，按授权卡点申请惯例打 p1；已打好）。

    状态更新：A/B 已裁决——owner 直接选了 A，且 v0.1.211 已发布

    对本 issue 各项声明的独立复核（全部通过）

    1. kernel diff 相符：窗口内唯一 kernel 源码变更即 In 0.1.210, sessions with frequent interruptions barely trigger compression reminders — tokens accumulate to 330k before | 中断频繁的会话在 0.1.210 里压缩提醒几乎不触发,积到 33 万 token 才压一次 #5415——kernel/src/compress.ts +44/-2（issue 写 "+46/-5"，实为 44 增 2 删、合计 46 行变更，措辞小出入）、kernel/src/types.ts +14、新回归测试 +198、版本 0.0.175→0.0.176。
    2. 声明齐全：PR [PR] Session compression reminders no longer starve under frequent interruptions: absorption debt is paid in steps starting f | 中断频繁的会话压缩提醒不再饿死:吸收欠账从第 2 次中断起付台阶(#5415) #5416 正文带 kernel-change: 声明，kernel 提交 c91ef1a81 message 亦带声明——Kernel Boundary 规则 1 满足。
    3. 机制通读复核：decideNudge 仅在 absorbCount>=2 时把 absorbedDebt 计入增长台阶并启用 backlog-drain 兜底；单次中断/无欠账会话 payableDebt=0 && backlogDrainReady=false，决策字节不变（In pi, Esc triggers compression injection. | pi中esc会触发压缩注入 #5293 钉死）；压缩成功时欠账与计数随基线清零；两个新字段是 NudgeState 可选字段（旧持久化状态兼容），无配置面。与 issue 描述一致。

    对楼上 heads-up（@Louie0806）的确认与补充

    一个值得记录的过程事实（非缺陷追责）

    正常 Release 运行（aeda157a0，09:51:06Z 起）本地全套件+构建全绿，但卡在 "Full CI gate"：publish sha 上的 ci.yml 与 e2e-dsh-plugin-lane 运行已被合并后紧随的 master 推送并发取消，gate 视 cancelled=red 直接 fail-fast，tag/publish 步骤跳过。这正是 #5431 修的 fallback-PR 路径问题（skip_ci_gate 之前只在 direct-push 路径生效），owner 10:05Z 用新闸阀完成发布，全程约 40 分钟空窗。后续同类情况走 #5424 线程即可，本条不重复展开。

    八问

    1. 是什么问题：release v0.1.211 含一处 kernel 变更（In 0.1.210, sessions with frequent interruptions barely trigger compression reminders — tokens accumulate to 330k before | 中断频繁的会话在 0.1.210 里压缩提醒几乎不触发,积到 33 万 token 才压一次 #5415），agent 无权自批——一次授权申请。现已由 owner 人工合并裁决为 A 并执行完毕。
    2. 什么原因导致：Kernel Boundary 规则 3（kernel 一律 human-only review）是设计内硬边界，非缺陷；中途的正常发布路径卡住则因 fallback-PR 合并后 publish sha 上的 CI 运行被并发推送取消、gate 判红（该类问题已由 [PR] (ai: anthropic/claude-opus-4-6) fix(release): emergency release flag skip_ci_gate now takes effect on the fallback relea | (ai: anthropic/claude-opus-4-6) fix(release): 紧急发版开关 skip_ci_gate 在 fallback 发版 PR 路径生效（原来被静默忽略，白等最多 60 分钟 CI） #5431 修掉）。
    3. 影响范围与严重程度：行为面仅高频中断（≥2 次吸收）会话的压缩提醒节奏（已披露+回归钉）；单次中断会话零变化。可用性面曾有 ~40 分钟发布空窗（09:47–10:25Z），已解除，无数据影响。
    4. 怎么修的：host 记账（absorbedDebt/absorbCount）+ kernel 第 2 次起把欠账计入增长台阶并加 backlog-drain 兜底，压缩成功随基线清零；发布卡点由 owner 经一键应急阀（[PR] (ai: anthropic/claude-opus-4-6) fix(release): emergency release flag skip_ci_gate now takes effect on the fallback relea | (ai: anthropic/claude-opus-4-6) fix(release): 紧急发版开关 skip_ci_gate 在 fallback 发版 PR 路径生效（原来被静默忽略，白等最多 60 分钟 CI） #5431 机制）完成。
    5. 回归风险：单次中断路径字节级不变（In pi, Esc triggers compression injection. | pi中esc会触发压缩注入 #5293 钉死）；无欠账/无中断会话决策字节不变；golden 漂移只涉测试面不涉产品行为；发布树与批准树产品代码逐字节一致。高危模式（fail-fast 变吞错、阈值漂移、输出格式变）逐项过——均无。
    6. 是否破坏缓存：否，nudge 决策不触碰前缀缓存字节。
    7. 是否改设计方向：否，沿用 In pi, Esc triggers compression injection. | pi中esc会触发压缩注入 #5293/fix: compress over unpruned view + acp-kernel 0.0.38 (billion-context-pi#195) #194 的 growth-driven 哲学，只补「欠账」维度。
    8. 是否引入新配置：否。absorbedDebt/absorbCount 为 NudgeState 内部可选字段（旧状态兼容）；raw_publish/skip_ci_gate 是 CI 闸阀，非用户配置。

    收工判定

    A/B 决策已闭环（owner 选 A，v0.1.211 含 kernel 0.0.176 已在 npm 上线）。本条等 owner 验收后关闭（agent 不关单）。遗留一件小事：建议合并 #5434（golden regen hermeticity 根因修复，已核零冲突）。

  3. ranxianglei
    added
    p1
    scope/kernel
    on Nov 1, 2026
  4. ranxianglei
    added 3 commits that reference this issue on Nov 1, 2026

    cli: -i/--inline and -s/--server posture flags (#5428)

    a61765b

    docs: stale precedence line in planNativePi docstring (#5491)

    25f908d

    merge: sync master into pi-inline-default branch (PR #5491 conflict r…

    e5cdb09
  5. Louie0806
    added a commit that references this issue on Nov 4, 2026

    pi: inline lane is the default posture (ranxianglei#5428)

    55c1b5e
  6. ranxianglei
    added 3 commits that reference this issue on Nov 4, 2026

    test(e2e): pin the pi-codex-ws shared-proxy scenario to the pre-#5428 …

    0365c22

    test(e2e): pin e2e-native-pi to the pre-#5428 spawn posture (#4599)

    c37293b

    test(#5077): pin pre-#5428 spawn posture in the issue5025 suite — the…

    5912250
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    p1scope/kernel

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions