[ranxianglei]

Closes #1997

What

Opt-in throughput-adaptive nudge cadence. When enabled, the fixed compress.nudgeGrowthTokens step is replaced by an adaptive step derived from the session's recent per-call input throughput: large arrivals (bulk file/log reads) widen the step so folding does not interrupt the work; small arrivals (quiet interactive turns) narrow it so context stays lean. Clamped to [nudgeGrowthMin, nudgeGrowthMax] and seeded from nudgeGrowthTokens until three samples exist. The signal references growth magnitude and call count (a bounded window of recent requests) — not wall-clock time or cache occupancy (see the paper §6.3 note on the cache-occupancy denominator artifact).

Config surface (new fields — owner-approved pre-implementation)

Three optional fields under compress (three-level cascade global→provider→model, per-field deepest-wins, exactly like every other compress field):

• nudgeAdaptive?: boolean — master switch. Default unset/false → byte-for-byte identical to today.
• nudgeGrowthMin?: number — default 10000. Lower clamp.
• nudgeGrowthMax?: number — default 200000. Upper clamp.

Why the existing surface cannot express this: nudgeGrowthTokens is a single STATIC step (it flattens the kernel's adaptive band to a constant); no field makes the step track observed throughput. Alternatives considered and rejected: re-introducing window-percentage scaling (forbidden by #379/#380), and cache-/occupancy-driven gating (owner rejected — a low-water fold's apparent miss is a denominator effect while the absolute re-computed tokens stay in the tens of thousands; the trade strongly favors compression). Internal tuning constants (window=8, min-samples=3, proportionality=1) are deliberately NOT exposed as config, to avoid surface bloat.

Behavior change (flagged)

• Default OFF → zero behavior change. With nudgeAdaptive unset, applyAdaptiveNudgeStep returns the resolved config untouched — no allocation, no band change.
• When ON → the per-request nudge.growthFloor/growthCap becomes a moving value S instead of the static step. This is the intended feature.

Both compression modes reasoned (§6):

• Proxy mode: the dynamic band is the intended effect (the proxy executes compress).
• Plugin mode: the proxy-side nudge is injected in BOTH modes (0.1.69 回归: plugin 模式 nudge 注入被 !pluginMode 静默禁用 — INJECT 日志照打但从不投递, 自动压缩只剩 preflight 应急 #451), so the adaptive cadence applies there too; its carrier is an ephemeral trailing user message that never enters the agent's persisted history, so there is no wire drift and no prefix-cache impact beyond what static cadence already had (tool injection remains suppressed).
• Prefix cache ([Bug] ACP: ≥2 compressed blocks render as multiple system messages → SGLang (vLLM-compatible) upstreams return 400 #377): only the nudge band changes; head/system bytes are untouched.

Scope

bili-only — no acp-kernel change (verified kernel 0.0.100 resolveAdaptiveGrowth = min(growthCap, max(growthFloor, round(limit×growthRatio))), so pinning floor == cap == S drives the effective step exactly for any window). Files: new src/nudge-adaptive.ts (pure/deterministic step), src/config.ts (type + parser + misplaced-key hints), src/compress-settings.ts (merge picks + applyAdaptiveNudgeStep), src/server.ts (sample rawInputTokens into a bounded per-session ring persisted via session.metadata, then overlay the band at session bind), docs en/zh + paper §6.3, tests tests/nudge-adaptive.test.ts.

Verification

typecheck ✓ · build ✓ · deterministic unit+integration suite (65) ✓ · focused regression over changed-surface suites (store / e2e-compress-cascade / issue453-backstop / fork-ccr-adoption / cache-friendly-plugin / absorb-plugin / issue402 / issue532, 74) ✓. Rebased onto current master (clean, no conflicts). Full regression delegated to CI.

[github-actions]

📦 Built Package Artifact

Branch: wt-1997-ework-daemon (3676e0f)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.2400-pr.2035.db0d75b

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-pr2035.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review 完成 — 可合并(需人工合并:feat + 新配置面,按 §7.4 不满足 auto-merge,留 human)。

查重:无重复——本 PR 即 #1997 的实现(设计已经 owner floor 12「同意开工」预批),字段名/默认值/三级级联与批准稿完全一致。

结构化七问:

1. 是什么问题:Supports dynamic compress.nudgeGrowthTokens | 支持动态compress.nudgeGrowthTokens #1997 需求——nudge 节奏现为静态 nudgeGrowthTokens 步长,批量读文件/日志时折叠打断工作、安静交互时上下文偏肥;需要可 opt-in 的吞吐自适应步长。
2. 什么原因导致的:静态步长无法表达「步长跟随到达节奏」——nudgeGrowthTokens 把内核自适应 band 扁平化为常数,现有配置面无任何字段能让步长跟踪观测吞吐(window-percentage 缩放被 fix: merge all system/developer messages into one leading message (SGLang #377) #379/探索是否可以集成进sglang，以做到透明管理 #380 禁止;缓存占用闸门被 owner 否决为分母假象)。
3. 怎么修复的:纯函数 computeAdaptiveStep 取近 8 次调用 Δtokens(每次调用到达率,clamp≥0、非有限值先滤除)的均值作步长 S,夹到 [nudgeGrowthMin,nudgeGrowthMax],样本 <3 时回退 nudgeGrowthTokens ?? 50000 种子;然后写 nudge.growthFloor = growthCap = S。已对安装的 acp-kernel 0.0.101 dist 验证 resolveAdaptiveGrowth = min(cap, max(floor, round(limit×ratio))) ⇒ floor==cap==S 时任一窗口有效步长恰为 S(PR 正文写 0.0.100,实际 master pin 是 0.0.101,公式一致且本 PR 不动 pin)。bili-only,零内核改动。
4. 是否有回归问题:无。逐项审过的既有行为:① 默认 OFF → applyAdaptiveNudgeStep 返回同一对象引用,config/wire 逐字节不变(有同引用恒等测试钉住);② 前缀缓存([Bug] ACP: ≥2 compressed blocks render as multiple system messages → SGLang (vLLM-compatible) upstreams return 400 #377):仅 band 变,head/system 字节不动;③ plugin 模式:nudge 双模式都注入(0.1.69 回归: plugin 模式 nudge 注入被 !pluginMode 静默禁用 — INJECT 日志照打但从不投递, 自动压缩只剩 preflight 应急 #451),自适应节奏同样生效(carrier 仍是 ephemeral trailing user message,不进 agent 持久历史 → 无漂移);④ preflight 硬兜底(plugin mode: no hard backstop when the agent ignores the compress nudge — context grows past the window and upstream 400s #453/track: plugin-mode proxy-side emergency truncation (last-resort backstop, deferred from #467) #470)依赖 modelContextLimit,与 band 无关;⑤ emergency nudge 基于 usage,与 band 无关;⑥ 内核 state.nudge 引用值来自持久化 state 而非 band 推导,band 按请求变化安全;⑦ 新 metadata 键 nudgeInputHistory(≤8 个数)即使 OFF 也会写入——与 rawInputTokens([Feature] Default state file cleanup strategy #1082)既有的无条件写 metadata 先例一致,有界、加性键,判定可接受,在此明示;⑧ 配置解析对非法类型严格拒绝,与兄弟字段纪律一致。本地定向验证:typecheck ✓、新套件 20 ✓、9 个变更面回归套件共 97 测试 ✓(store / e2e-compress-cascade / issue453-backstop / fork-ccr-adoption / cache-friendly-plugin / absorb-plugin / issue402 / issue532 / issue833-panel-effective-config)。CI 原 head 16/16 全绿;我补的 2 个 docs commit 后正在重跑(docs-only)。
5. 是否破坏缓存:不涉及——前缀缓存锚点(head system 字节)未动;CCR/content-store、更新节流、去重键均未触碰。
6. 是否改了设计方向:无转向——复用既有三级级联 + session.metadata 持久化通道;信号选择(每次调用到达率而非墙钟/缓存占用)正是 owner 裁定。
7. 是否引入了新的配置:是,3 个 opt-in 字段,无需迁移:compress.nudgeAdaptive(bool,unset/false = 现状)、nudgeGrowthMin(默认 10000)、nudgeGrowthMax(默认 200000);内部常量(window=8 / min-samples=3 / proportionality=1)刻意不暴露,防表面膨胀。CONFIGURATION.md 中英条目照旧补齐。

行为变更 flag:默认 OFF → 零行为变化;ON → 每请求 growthFloor/growthCap 变为移动值 S(即功能本身,已在正文与文档披露)。

review 中发现的小问题,已直接修(不止报告):

• b7e5f112:README/paper 新增行带多余行首空格(cosmetic,渲染不变);
• abc15864:AGENTS.md KDD#7 仍写「the proxy suppresses tool + nudge injection」——自 0.1.69 回归: plugin 模式 nudge 注入被 !pluginMode 静默禁用 — INJECT 日志照打但从不投递, 自动压缩只剩 preflight 应急 #451 起 nudge 双模式都注入(src/server.ts:3900 注释「Inject in BOTH modes (0.1.69 回归: plugin 模式 nudge 注入被 !pluginMode 静默禁用 — INJECT 日志照打但从不投递, 自动压缩只剩 preflight 应急 #451)」+ TECHNICAL-NOTES.md:180 表格行「Proxy-injected nudge | yes | yes」),这句过时措辞正是 PR 正文 plugin-mode 误述的来源,顺手改对;
• PR 正文 plugin-mode 段已 API PATCH 修正(原句「injection is suppressed regardless of the band」与代码事实不符,改为 0.1.69 回归: plugin 模式 nudge 注入被 !pluginMode 静默禁用 — INJECT 日志照打但从不投递, 自动压缩只剩 preflight 应急 #451 语义的准确表述)。

判定:方向正确(预批设计原样落地)、无回归、测试有牙(边界/clamp/同引用/级联/解析拒绝均有断言)、diff 干净——可合并,请人工合并:#2035

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 CI 补报(windows-22 单测 flake,已重跑转绿)

上一轮回复末尾说"正在重跑",结果如下:

现象:两个 review 小修 commit 推上去后(doc-only),主 CI workflow 仅 test (windows-latest, 22) 一个 job 红,其余(version-guard、ubuntu-22/24、windows-24、全部 e2e/overlay/artifact lane)全绿。失败的是单个测试:

not ok 3158 - Responses WS: a title-persona side request rides the lane transport instead of 502ing
tests/responses-websocket.test.ts — AssertionError: input did not match /side request \(agent=title\)/
suite: 3954 tests / 3939 pass / 1 fail / 14 skipped

判定为环境性 flake,非本 PR 回归,证据链:

1. diff 不可能影响:相对上一个全绿 head(6b6a9b1),新增的只有两个 markdown commit(文档行首空格 + AGENTS.md KDD#7 措辞),JS 行为零变化。
2. 同版本对照:同一 Node 22 测试集在 ubuntu-22 通过,windows-24 也通过 → Windows 特有的时序问题,不是代码路径问题。
3. 断言机制:该测试在代理启动后 ~93ms 就断言日志里出现 side request (agent=title);抓到的日志只有启动行 + 约 30 条 [instances] another bili instance is running(无单实例守卫:多个 bili 代理进程并存共写 sessions 目录,伴随孤儿 ref、有注册表无 transcript、计数自相矛盾的污染簇 #394 多实例警告,来自共享 runner 上并行跑的其它测试进程)——典型的共享 Windows runner 高负载下日志竞态。
4. 本地定向复跑:node --import tsx --test tests/responses-websocket.test.ts Linux 上 36/36 通过(2.7s)。

处置:已触发 failed-job 重跑,整个 run 现已 completed success(head abc1586)。无需为这个失败改任何代码;该 flake 无既有 issue 记录,按机制粒度规则不单独立项(单次瞬态、无复现证据链支持系统性缺陷),若后续在 CI 里再次出现同测试失败,再按"重复出现"立项跟进。

结论不变:可合并(需人工 merge——feat + 新配置面,§7.4 人工通道)。

[ranxianglei]

Related: the default-cadence side of this landed as #2278 (window-scaled default step, 20k..50k band — fixes the #2110 RC2 starvation for sub-1M windows). #2035 remains complementary: it adds the opt-in throughput-adaptive cadence (nudgeAdaptive / nudgeGrowthMin / nudgeGrowthMax) on top of the scaled default, with nudgeGrowthTokens as the seed. No overlap in behavior — happy to rebase if desired.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 收到,已核对并顺手把 #2035 rebase 到当前 master(零冲突)。先一个事实更正,再给交互分析:

① 事实更正:#2278 尚未 landed。API 核实:#2278 fix(kernel): window-scale the default nudge growth step (50k flat -> 20k..50k band) 当前 state=open, merged=false(分支 2026-10-07_nudge-growth-window-scaled,kernel 0.0.104→0.0.105)。master 上 AGENTS.md KDD#8 仍写 flat-50K 契约——#2278 是那个契约的唯一待决变更。所以目前是两个 open PR,不是"已落地+互补",合并顺序需要定一下(见 ③)。

② 机制层交互分析(结论:不冲突,顺序二选一均可):

• fix(kernel): window-scale the default nudge growth step (50k flat -> 20k..50k band) #2278 只改 kernel 默认 growthFloor 50000→20000(resolveAdaptiveGrowth 公式不变:min(growthCap, max(growthFloor, round(limit×ratio))))。feat(#1997): throughput-adaptive nudge cadence (compress.nudgeAdaptive) #2035 的机制是 pin floor == cap == S,代入公式恒等于 S(min(S, max(S, x)) = S;limit≤0 时返回 growthFloor 也 = S)——与默认值是多少无关,已在 master 的 in-repo kernel(kernel/src/compress.ts:1625)重新逐行验证。两机制正交。
• 唯一接缝是 adaptive 的 seed:<3 个样本前,feat(#1997): throughput-adaptive nudge cadence (compress.nudgeAdaptive) #2035 用 nudgeGrowthTokens ?? 50000(镜像当前 flat-50k 默认)起步。若 fix(kernel): window-scale the default nudge growth step (50k flat -> 20k..50k band) #2278 先合:sub-400k 窗口下 adaptive-ON 前几次调用从 50k 起步、adaptive-OFF 停在 20k——瞬态(≤3 次调用)、随后被吞吐数据接管,自纠正,但起步不一致。若 feat(#1997): throughput-adaptive nudge cadence (compress.nudgeAdaptive) #2035 先合:seed 与当前默认精确相等,零接缝;待 fix(kernel): window-scale the default nudge growth step (50k flat -> 20k..50k band) #2278 落地后再补一个小 follow-up 把 seed 对齐到 window-scaled 默认即可(我可以做,一行常量+注释的事)。
• 文档面:两者都动 CONFIGURATION.md/README 的 nudge 段落,无论谁先合并,后者的文本冲突都是相邻行、trivial 可解。
• 顺带确认 master 这 410 个提交里新增的 nudgeSuppressed(session)(session-self-heal 会话卡死自愈机制：卡死模式检测与代理侧主动补救 #2155,zombie lane 抑制注入)与本 PR 正交:它是 per-session 注入开关,不改 band/threshold;抑制期间 ring 照常采样,恢复时 band 反映的是真实近期吞吐,行为反而更准。

③ 建议合并顺序:#2035 先合(seed 零接缝),#2278 后合时我补 seed 对齐 follow-up;若 owner 更想先落 #2278(修 #2110 starvation 紧迫性更高),也可以——代价只是上述瞬态 seed 不一致,我同样可以先把 seed 对齐 commit 加进 #2035 再等 #2278。请拍板顺序。

④ Rebase 报告(新 head 3676e0f1,原 abc15864 落后 master 410 提交):

• rebase 零文本冲突;diff 保持原样(12 文件 +349/−8,含 review 小修)。
• 语义审计(热点文件 server.ts 这 410 提交里大改过):overlay 点仍在所有 prepare* 路径之前;master 新增的 output-headroom modelContextLimit 收缩发生在 overlay 之后,但对 floor==cap pin 无效(公式保证),动态步长不受影响;storeEffectiveConfig 面板 stamp 行为不变。
• 本地验证(rebased head):typecheck ✓ · build ✓ · nudge-adaptive 20/20 ✓ · 聚焦回归 8 套件 74/74 ✓。全量回归交 CI,结果出来我再补报。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 CI 补报:rebased head 3676e0f1 全绿 ✅

全部 13 条 workflow lane completed success:CI(ubuntu-22/24 + windows-22/24 + version-guard)、Theorem Lane、E2E real codex、e2e-native-dsh / e2e-native-pi / e2e-native-opencode、e2e-dsh-plugin-lane、e2e-pi-codex-ws、e2e-fake-codex、e2e-local-registry、ci-image、PR Build Artifact、CI Windows codex overlay。

当前状态汇总:

• feat(#1997): throughput-adaptive nudge cadence (compress.nudgeAdaptive) #2035 已 rebase 到 master @9f82bbed8,零冲突,diff 保持原样;本地 typecheck/build/定向回归全过,远程 CI 全绿。
• 与 fix(kernel): window-scale the default nudge growth step (50k flat -> 20k..50k band) #2278 的机制交互已分析(正交,唯一接缝是 adaptive seed 的瞬态起步值),待 owner 拍板合并顺序:建议 feat(#1997): throughput-adaptive nudge cadence (compress.nudgeAdaptive) #2035 先合(seed 零接缝);若 fix(kernel): window-scale the default nudge growth step (50k flat -> 20k..50k band) #2278 先合,我补一个 seed 对齐的小 follow-up commit。
• 结论不变:可合并(需人工 merge——feat + 新配置面,§7.4 人工通道)。