Skip to content

feat: remove the absorb distillation feature (#3293) - #3309

Merged
ranxianglei merged 1 commit into
masterfrom
2026-10-15_remove-absorb
Oct 16, 2026
Merged

ranxianglei merged 1 commit into
masterfrom
2026-10-15_remove-absorb

Conversation

@ranxianglei

Copy link
Copy Markdown
Owner

Removes the absorb distillation feature entirely.

What & why

The absorb tool let a model distill oversized tool results through an injected tool. Its prompt activation was gated on shouldInject alone, but tool injection was additionally gated on !pluginMode. So in every plugin-mode lane (pi/omp/opencode/dsh) with compress.absorb.enabled=true, the model was told to call a tool that never appears on its wire → unknown-tool spin, wasted turns, growing context pressure. Same symptom family as #3177 (which fixed the compress surface); this is the absorb surface across all plugin lanes.

Owner decision (#3293): remove the feature rather than gate it — it was opt-in, default-off, and found to be useless in practice.

Config surface

Removed config fields (all under compress.absorb.*): enabled, toolName, minTokens, maxTokens, levels. No new or renamed fields.

Compat / migration: the config parser is lenient — old configs that still carry compress.absorb.* are silently ignored (no error, no migration step) because the feature no longer exists. Nothing else reads these keys.

Behavior change (flagged per policy)

  • Old: compress.absorb.enabled=true activated an absorb system prompt + kernel [ACP absorb] markers + an injected absorb tool. In practice only proxy mode worked; plugin mode was broken (see above).
  • New: absorb no longer exists anywhere. Oversized tool results are handled solely by the CCR content store (acp_retrieve), which was already the lossless default path.
  • No change to any other compression behavior.

Scope notes

  • Kernel untouched. The in-repo kernel/ (acp-kernel) still exports its absorb API; bili simply stops passing config.absorb into the loop. kernel.json golden unchanged.
  • Dormant emission-drop machinery retained. The low-level m00885 emission-drop path (adapter absorbArmed params, tag-echo-filter tcHold state machine) is kept but unarmable — after this removal no request can set absorbArmed=true (there is no config.absorb). Full excision would be a high-risk streaming-core refactor disproportionate to removing a user toggle; documented here as a deliberate boundary. Its unit tests (absorb-emission.test.ts, loop-adapters.test.ts, adapter-google.test.ts) stay green against the dormant path.
  • The companion rulesActive gate (same missing !pluginMode) is out of scope — the request was to remove absorb only; rules behavior is unchanged.

Tests

  • Deleted tests/absorb.test.ts + tests/absorb-plugin.test.ts; stripped absorb fixtures/assertions/golden-specs from ~15 affected suites.
  • Regenerated wire-contract goldens: bili.json (−22 lines), manifest-optin.json (−68 lines) — pure absorb deletions; kernel.json + manifest-default.json unchanged.
  • Local targeted run: 286/286 pass across wire-contract, store, ccr-v2, plugin-protocol, rules, image-compress, opencode-v2/native, fix-1897, loop-adapters, absorb-emission, adapter-google, issue1881, cache-seam-*, fork-ccr-adoption, search-plan-aware, decompress-tofile suites. Full suite = CI.

Closes #3293

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b)

接单处理 PR #3309(移除 absorb distillation 功能)。

查重:不重复。 与 #3177/#3287(compress 面在 plugin lane 不可达)同症状家族但不同机制——#3177 修的是 compress 工具面,本 PR 按 owner 在 #3293 的决定直接移除 absorb 整个功能;与在途其他 PR 无文件级重叠需协调的冲突(见下)。

当前状态:PR 分支 2026-10-15_remove-absorb 基于 b2868ca,落后 master 159 个提交,与 master 合并有 15 处冲突标记(平台已转发 conflict-worker,但我已接单,直接由我完成 rebase)。diff 本身干净:49 文件,+111/−1154,纯删除向。

正在 rebase 到最新 master → 定向验证 → push,完成后按八问体例回复结论。

@ranxianglei
ranxianglei force-pushed the 2026-10-15_remove-absorb branch from 8cc5e64 to 84e7a25 Compare October 15, 2026 19:02
@github-actions

github-actions Bot commented Oct 15, 2026 •

Copy link
Copy Markdown
Contributor

📦 Built Package Artifact

Branch: 2026-10-15_remove-absorb (8a43cd8)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.3892-pr.3309.8f19865

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-pr3309.tgz
npm install -g package

Install instructions are refreshed on each push.

@ranxianglei
ranxianglei force-pushed the 2026-10-15_remove-absorb branch from 84e7a25 to 67fda03 Compare October 15, 2026 19:08
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

冲突已解决并 rebase 到最新 master,另修掉一处遗漏(docs-sync CI 红),已推送到 PR 分支:head 现为 67fda03(2026-10-15_remove-absorb,单 commit,51 files +117/−1207)。CI 正在新 head 上跑,绿了我再确认。

做了什么

  1. Rebase 到最新 master:分支落后 159 个 commit。冲突根因不是逻辑矛盾——master 侧 image_full 无法被正确调用:同一张图每轮重复降采样(38 次)、restore 提示注入 38 次,模型只能猜 ref → 每轮一次失败+重发,缓存命中从 75% 掉到 28% #3266(700722db9)删掉了 image pre-compression(image_full 工具族),本 PR 删 absorb,两条纯删除改动撞在同一批 import 行、工具注入数组、config 字段表、golden 常量上。按并集规则解:每个冲突点同时丢掉两个符号族,两边语义都完整保留。涉及 src/{compress-settings,config,server,server/prepare-{anthropic,google,openai,responses},server/relay,server/side-request}.ts、tests/{fix-1897,wire-contract}.test.ts;src/absorb.ts 与 tests/image-compress.test.ts(master 已删)直接删除。
  2. 顺手清理失效注释(仅注释,无代码行为变化):src/config.ts / src/compress-settings.ts 里多处「merged sub-field-wise like absorb」类比改为仍存在的字段(prompts/ccr/reasoning),并修掉一个指向已不存在符号 seenAbsorbDivergenceWarnings 的注释(改指 seenNamedProviderWarnings)。描述 kernel 管线顺序的注释(kernel 未动、确实还有 absorb 节点)保留不动。
  3. 修复 docs-sync CI 红(本 PR 自身遗漏,非冲突引入):PR 把 absorb 行从 CONFIGURATION.md/.zh-CN.md 删了,但 website/config-reference/compress.yaml 里 6 条 absorb 种子条目没删——docs-sync lane 用 tools/gen-config-docs.mjs generate 从种子重新生成文档再 diff,于是 absorb 行被生成回来 → diff 非空 → 红。已删种子并重跑生成器:纯删除 43 行(yaml −40、website/main.js i18n −2、website/docs/index.html −1),website/ 下 absorb 残留为 0。

结构化八问

  1. 是什么问题:PR 分支与最新 master 文本冲突无法合并;且文档种子文件漏删导致 docs-sync 门禁红。
  2. 什么原因导致的:两个独立的「删除型」变更(master 删 image pre-compression、本 PR 删 absorb)触碰同一批共享代码位置(import 行/注入数组/字段表/golden),产生文本重叠冲突;文档侧是「改了生成产物没改生成源」。
  3. 影响范围与严重程度:仅影响本 PR 分支本身;master 不受影响。合并后行为变化即 PR 描述所述(absorb 彻底移除,plugin lane 的 unknown-tool spin 消失);不合并则分支继续不可合。
  4. 怎么修复的:冲突按「两个删除都保留」的并集规则逐点解决;文档侧补删种子条目并走官方生成器重生成(而不是手改生成产物)。
  5. 是否有回归问题:typecheck 干净;本地定向跑了 20 个相关套件(wire-contract/golden、rules、fix-1897、loop-adapters、absorb-emission 休眠路径、adapter-google、ccr/store 系、cache-seam 系等)267/267 全过。kernel 未动(kernel.json golden 未变)、rulesActive 门控未动(PR 明确 out of scope)。全量套件以 CI 为准,跑完我确认。
  6. 是否破坏缓存:不涉及——无缓存/重试/增量游标/去重键变化。
  7. 是否改了设计方向:不涉及——执行的是 owner 在 开了 absorb 之后,插件模式会话被指示调用一个永远不会出现的 absorb 工具(与 #3177 同症状家族,机制不同) #3293 拍板的删除决策,无新增设计。
  8. 是否引入了新的配置:无新增;仅删除 compress.absorb.* 五个字段,旧配置里的残留键被宽容解析器静默忽略(迁移零成本)。

判定

本地验证(typecheck + 267 定向测试)已过,diff 干净(每一行都与「删 absorb + 对齐 master」相关)。待 CI 全绿即可合并——CI 结果出来我再回一条确认;若有任何 lane 红,我会先定向排查再报。

@ranxianglei
ranxianglei force-pushed the 2026-10-15_remove-absorb branch 2 times, most recently from 602e74c to 3890451 Compare October 15, 2026 19:21
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

冲突已解决:分支已 rebase 到最新 master(6ee4a2ac9)并推送(389045120,force-with-lease),GitHub 已确认 mergeable: true(当前 blocked 仅为 CI 未跑完)。

八问

  1. 是什么问题:本 PR(移除 absorb 蒸馏功能)与 master 冲突——master 在 PR 期间合入了 feat!: remove image pre-compression and the image_full tool (#3266) #3295(image_full 无法被正确调用:同一张图每轮重复降采样(38 次)、restore 提示注入 38 次,模型只能猜 ref → 每轮一次失败+重发,缓存命中从 75% 掉到 28% #3266)移除图片预压缩和 image_full 工具,两个移除动作落在同一批文件的同一批行上。
  2. 什么原因导致的:分支基线 b2868ca68 落后 master 159 个提交;image_full 无法被正确调用:同一张图每轮重复降采样(38 次)、restore 提示注入 38 次,模型只能猜 ref → 每轮一次失败+重发,缓存命中从 75% 掉到 28% #3266 删掉了 src/image-compress.ts 及全部 IMAGE_FULL_*/imageCompressionEnabled 引用,与本 PR 删除 absorb 引用共用同一组 import 行、inject 行 → 13 个文件文本冲突。
  3. 影响范围与严重程度:仅本 PR 分支自身。不影响任何用户;不解决则 PR 无法合并。无生产面影响。
  4. 怎么修复的:rebase 到 6ee4a2ac9,统一解法 = 以 master 版本为准,只去掉 absorb 引用,绝不复活 image_full 引用。另修掉原提交遗漏的三类残留(都属于本 PR 声称要完成的事):① src/proxy-tool.ts 仍 import 已被 image_full 无法被正确调用:同一张图每轮重复降采样(38 次)、restore 提示注入 38 次,模型只能猜 ref → 每轮一次失败+重发,缓存命中从 75% 掉到 28% #3266 删除的 ./image-compress.js 并保留 image_full 判定——不修直接构建失败;② 同文件多余的 image 分支;③ 约 9 处描述已删代码的陈旧注释(src/config.ts 4 处 doc、src/compress-settings.ts 3 处、四个 prepare-* 里的 "strip absorb from the loop config" 段落)。
  5. 是否有回归问题:rebase 本身无回归。逐项验证:typecheck 干净;18 个定向套件 270 用例全绿(wire-contract golden 门禁、fix-1897-leaked-tools-side、store、ccr-v2、plugin-protocol、rules、absorb-emission(休眠 emission-drop 机制按设计保留)、loop-adapters、adapter-google、cache-seam-fuzz/probes、search-plan-aware、opencode-native/v2、fork-ccr-adoption、decompress-tofile×2、storage-corpus);全量回归交给 CI。golden 重新生成后与原提交暂存版本逐字节一致,对 master 的 delta 是纯 absorb 删除(bili.json −22 absorbResponses 块、manifest-optin.json −68),kernel.json 与 manifest-default.json 零改动(kernel 未动,符合 Kernel Boundary)。
  6. 是否破坏缓存:不涉及——未触碰缓存/重试/增量游标/去重键任何机制。
  7. 是否改了设计方向:没有。rebase 完整保留原 PR 设计:absorb 功能整体移除 + 休眠 emission-drop 机制保留(不可再武装)+ 旧 compress.absorb.* 配置键静默忽略。
  8. 是否引入了新的配置:没有。

判定:冲突闭环,可等 CI 绿灯后由人合并(合并权限在人,我不代合)。

@ranxianglei
ranxianglei force-pushed the 2026-10-15_remove-absorb branch from 3890451 to dac6224 Compare October 15, 2026 19:31
The `absorb` tool let a model distill oversized tool results through an injected
tool, but its prompt activation was gated on `shouldInject` alone while tool
injection was additionally gated on `!pluginMode`. In every plugin-mode lane
(pi/omp/opencode/dsh) with compress.absorb.enabled=true the model was therefore
told to call a tool that never appears on the wire — unknown-tool spin, same
symptom family as #3177 (which fixed the compress surface). Owner decision:
remove the feature entirely rather than gate it.

- delete src/absorb.ts; relocate isProxyToolFor to src/proxy-tool.ts (absorb branch dropped)
- drop absorb from config schema/validation/divergence-warn, compress-settings,
  the four prepare-* wire paths, relay, preflight, side-request, stream*, loop/core,
  plugin manifest, server inject; drop ABSORB_TOOL* from compress-tool + opencode-v2
- delete tests/absorb.test.ts + tests/absorb-plugin.test.ts; strip absorb
  fixtures/assertions/golden-specs from affected suites
- regenerate wire-contract goldens (bili.json, manifest-optin.json); kernel.json unchanged
- docs en/zh: remove absorb config sections and cross-references

The low-level m00885 emission-drop machinery (adapter absorbArmed params,
tag-echo-filter tcHold path) is kept dormant/unarmable: no request can arm it
after this removal. Kernel is untouched (human-gated).
Closes #3293
@ranxianglei
ranxianglei force-pushed the 2026-10-15_remove-absorb branch from dac6224 to 8a43cd8 Compare October 15, 2026 19:52
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b)

PR #3309 已 rebase 到当前 master 并完成全部修复,分支现为单提交 8a43cd86b。 过程中发现并处理了三件事:①与平台 conflict-worker 的并行 rebase 收敛(取两版并集);②补上原 PR 漏掉的三处(docs 种子、lane-guard 配对、#1192 套件残留);③确认剩余 CI 红全部来自 master 基线本身,与本 PR 无关。

做了什么(按时间序)

  1. Rebase:分支落后 master 159 个提交。冲突按「master 删 image 预压缩、本 PR 删 absorb → 两处符号都删」的并集规则解决;顺手清理了所有指向已删机制的过期注释(config.ts / compress-settings.ts / prepare-*)。
  2. 并行收敛:conflict-worker 曾推过它自己的 rebase 版 389045120。逐文件比对后取并集——采纳它对 prepare-{anthropic,google,openai,responses}.ts 更彻底的注释清理(我保留了 4 处过期 absorb 注释),保留我的 docs 种子删除、lane-guard 配对测试和 config.ts 注释修正。它的 CI 正好红在 docs-sync + lane-guard 两个 lane,反证这两处缺一不可。
  3. 补漏三处(原 PR 本地只跑了定向子集,未跑全量,以下均为全量 CI 暴露):
    • website/config-reference/compress.yaml 还留着 6 条 absorb 种子 → docs-sync 生成器把它们写回 CONFIGURATION*.md → 红。已删种子并重新生成(−43 行纯删除)。
    • src/server/side-request.ts 被改动但四个 side-lane 套件一个没动 → lane-guard(dsh 回归防护:泳道判定真值表 + 侧请求不变量 + dsh e2e lane + 运行时金丝雀 + lane-guard(5 项) #2170)红。已在 tests/side-lane-invariants.test.ts 补 (I) 用例钉住「absorb 不再是 bili-owned」。
    • tests/issue1192-disabled-tools.test.ts 还有一个「absorb disabled → 解释文案」子测试——absorb 移除后调用它就是真·未知工具(400+允许清单,正确行为),该子测试测的是已删功能,已删(文件头注释同步去掉对已删套件的引用)。
  4. 验证:typecheck 干净;本地定向 20+ 套件 267/267;上述三个修复各自定向复跑全绿。

CI 现状(head 8a43cd86b,43 条 check 全部出结果)

八问

  1. 是什么问题:按 owner 决定整体移除 absorb 蒸馏功能(插件模式下模型被指示调用永远不会出现的工具,空转烧上下文)。
  2. 什么原因导致的:prompt 激活只 gate 在 shouldInject,工具注入额外 gate 了 !pluginMode,两侧门不一致;opt-in 默认关且实测无用,故移除而非修门。
  3. 影响范围与严重程度:仅显式开了 compress.absorb.enabled=true 的用户(默认关,面很窄);其中插件车道是坏的(空转),代理车道可用。移除后超大工具结果统一走 CCR(acp_retrieve),本来就是无损默认路径。
  4. 怎么修复的:整条 absorb 面从 bili 侧切除(配置解析、提示注入、工具注入、side-lane 归属、文档、golden);kernel 不动(仍导出 absorb API,bili 不再传 config.absorb);低层 emission-drop 机制保留但不可武装(无 config.absorb 即无人能设 absorbArmed=true),完整摘除属流式核心大重构,不成比例,作为刻意边界记录在 PR 描述。
  5. 是否有回归问题:审过全部触碰路径的既有行为——rules/ccr/retrieve 注入、side-lane 判定、wire-contract golden、存储渲染语料均保持;唯一行为变化即 absorb 消失(PR 描述已按政策标注旧→新)。本次 rebase 额外修掉的三处都是「PR 自身不完整」,不是行为回归。剩余 CI 红经逐一核对为 master 基线问题(见上),非本 PR 引入。
  6. 是否破坏缓存:不涉及。prefix-cache 锚点、折叠游标、去重键路径均未触碰。
  7. 是否改了设计方向:不涉及架构转向;只是把一个实验性 opt-in 功能下线,压缩主链路(nudge→fold→CCR)不变。
  8. 是否引入了新的配置:无新增;删除 compress.absorb.{enabled,toolName,minTokens,maxTokens,levels} 五个字段,解析器宽容忽略旧配置中的残键,无需迁移。

判定:本 PR 范围内的工作已全部闭环且验证通过;当前 8 个红分片全部由 master 基线红(#3313/#3320、#3317 接手中)造成。等 #3317/#3320 合入后重跑即可全绿,届时可合并;若 owner 接受「已知基线红」直接合并也可以——本 PR 没有任何自己引入的红。我不代合并。

@ranxianglei
ranxianglei merged commit 9a5ff5c into master Oct 16, 2026
35 of 43 checks passed
@github-actions

Copy link
Copy Markdown
Contributor

📦 Merged Build Published

This merge to master was published as 0.0.475-master.9a5ff5c (rolling master tag).

npm install -g billion-context@0.0.475-master.9a5ff5c  # this exact build
npm install -g billion-context@master  # always the newest merged build

npm view billion-context@master prints the current version at any time.

ranxianglei pushed a commit that referenced this pull request Oct 16, 2026
…3312-#3320 batch)

- keep #3298 changes on master's post-absorb shapes: relay round-2 rebuild
  and all four prepare* paths keep withProjectAttribution; absorb imports/
  sections dropped with #3309
- storage-render goldens: metadata-section drift after the merge (absorb
  fields gone from session metadata) — regenerated via
  scripts/update-storage-render-goldens.ts, wire-body sections unchanged
This was referenced Oct 16, 2026
ranxianglei added a commit that referenced this pull request Oct 16, 2026
Restore the absorb machinery reverted by #3309, as the base for the
media-scoped revival (#3398): conflicts resolved keeping both the restored
absorb wiring and post-#3309 additions (withProjectAttribution #3298,
withAcpNotes #3252). src/proxy-tool.ts is deleted again — its only export
isProxyToolFor moves back to absorb.ts (absorb-aware, pre-#3309 shape).

This revert is intentionally NOT pushed alone; the media-only scope
restriction lands on top in the same PR.
ranxianglei pushed a commit that referenced this pull request Oct 16, 2026
…t parity (#3402)

Host applies the product policy the kernel leaves neutral: resolveAbsorbSettings defaults mediaOnly to true (image-bearing results only), so pure-text large results are no longer marked by default and the old over-compression mode is gone; mediaOnly:false restores legacy full-text behavior. absorb.enabled still defaults false. New config key compress.absorb.mediaOnly (boolean) added to the schema, cleaning loop, and plugin-lane divergence fields; config docs regenerated (EN/ZH/website). Adds a regression test pinning that the plugin-lane absorb prompt gate and the manifest advertisement read the same base config block (#3293: prompt appears iff tool is advertised). CHANGELOG backfills the entry #3309 omitted and records this revival.
ranxianglei added a commit that referenced this pull request Oct 16, 2026
Restore the absorb machinery reverted by #3309, as the base for the
media-scoped revival (#3398): conflicts resolved keeping both the restored
absorb wiring and post-#3309 additions (withProjectAttribution #3298,
withAcpNotes #3252). src/proxy-tool.ts is deleted again — its only export
isProxyToolFor moves back to absorb.ts (absorb-aware, pre-#3309 shape).

This revert is intentionally NOT pushed alone; the media-only scope
restriction lands on top in the same PR.
ranxianglei pushed a commit that referenced this pull request Oct 16, 2026
…t parity (#3402)

Host applies the product policy the kernel leaves neutral: resolveAbsorbSettings defaults mediaOnly to true (image-bearing results only), so pure-text large results are no longer marked by default and the old over-compression mode is gone; mediaOnly:false restores legacy full-text behavior. absorb.enabled still defaults false. New config key compress.absorb.mediaOnly (boolean) added to the schema, cleaning loop, and plugin-lane divergence fields; config docs regenerated (EN/ZH/website). Adds a regression test pinning that the plugin-lane absorb prompt gate and the manifest advertisement read the same base config block (#3293: prompt appears iff tool is advertised). CHANGELOG backfills the entry #3309 omitted and records this revival.
ranxianglei added a commit that referenced this pull request Oct 17, 2026
Restore the absorb machinery reverted by #3309, as the base for the
media-scoped revival (#3398): conflicts resolved keeping both the restored
absorb wiring and post-#3309 additions (withProjectAttribution #3298,
withAcpNotes #3252). src/proxy-tool.ts is deleted again — its only export
isProxyToolFor moves back to absorb.ts (absorb-aware, pre-#3309 shape).

This revert is intentionally NOT pushed alone; the media-only scope
restriction lands on top in the same PR.
ranxianglei pushed a commit that referenced this pull request Oct 17, 2026
…t parity (#3402)

Host applies the product policy the kernel leaves neutral: resolveAbsorbSettings defaults mediaOnly to true (image-bearing results only), so pure-text large results are no longer marked by default and the old over-compression mode is gone; mediaOnly:false restores legacy full-text behavior. absorb.enabled still defaults false. New config key compress.absorb.mediaOnly (boolean) added to the schema, cleaning loop, and plugin-lane divergence fields; config docs regenerated (EN/ZH/website). Adds a regression test pinning that the plugin-lane absorb prompt gate and the manifest advertisement read the same base config block (#3293: prompt appears iff tool is advertised). CHANGELOG backfills the entry #3309 omitted and records this revival.
ranxianglei added a commit that referenced this pull request Oct 17, 2026
Restore the absorb machinery reverted by #3309, as the base for the
media-scoped revival (#3398): conflicts resolved keeping both the restored
absorb wiring and post-#3309 additions (withProjectAttribution #3298,
withAcpNotes #3252). src/proxy-tool.ts is deleted again — its only export
isProxyToolFor moves back to absorb.ts (absorb-aware, pre-#3309 shape).

This revert is intentionally NOT pushed alone; the media-only scope
restriction lands on top in the same PR.
ranxianglei pushed a commit that referenced this pull request Oct 17, 2026
…t parity (#3402)

Host applies the product policy the kernel leaves neutral: resolveAbsorbSettings defaults mediaOnly to true (image-bearing results only), so pure-text large results are no longer marked by default and the old over-compression mode is gone; mediaOnly:false restores legacy full-text behavior. absorb.enabled still defaults false. New config key compress.absorb.mediaOnly (boolean) added to the schema, cleaning loop, and plugin-lane divergence fields; config docs regenerated (EN/ZH/website). Adds a regression test pinning that the plugin-lane absorb prompt gate and the manifest advertisement read the same base config block (#3293: prompt appears iff tool is advertised). CHANGELOG backfills the entry #3309 omitted and records this revival.
ranxianglei pushed a commit that referenced this pull request Oct 17, 2026
…-keys INERT list (#3402)

The #3309 revert revived absorb, but #3344 still listed compress.absorb alongside compress.imageCompression as a removed/inert key: old-config parsing dropped absorb and tests/removed-compress-keys-inert.test.ts asserted it parsed to undefined. Pull absorb out of INERT_REMOVED_COMPRESS_KEYS (imageCompression stays removed), correct the warning text, and update the test so absorb is asserted kept while only imageCompression warns.
ranxianglei added a commit that referenced this pull request Oct 17, 2026
Restore the absorb machinery reverted by #3309, as the base for the
media-scoped revival (#3398): conflicts resolved keeping both the restored
absorb wiring and post-#3309 additions (withProjectAttribution #3298,
withAcpNotes #3252). src/proxy-tool.ts is deleted again — its only export
isProxyToolFor moves back to absorb.ts (absorb-aware, pre-#3309 shape).

This revert is intentionally NOT pushed alone; the media-only scope
restriction lands on top in the same PR.
ranxianglei pushed a commit that referenced this pull request Oct 17, 2026
…t parity (#3402)

Host applies the product policy the kernel leaves neutral: resolveAbsorbSettings defaults mediaOnly to true (image-bearing results only), so pure-text large results are no longer marked by default and the old over-compression mode is gone; mediaOnly:false restores legacy full-text behavior. absorb.enabled still defaults false. New config key compress.absorb.mediaOnly (boolean) added to the schema, cleaning loop, and plugin-lane divergence fields; config docs regenerated (EN/ZH/website). Adds a regression test pinning that the plugin-lane absorb prompt gate and the manifest advertisement read the same base config block (#3293: prompt appears iff tool is advertised). CHANGELOG backfills the entry #3309 omitted and records this revival.
ranxianglei pushed a commit that referenced this pull request Oct 17, 2026
…-keys INERT list (#3402)

The #3309 revert revived absorb, but #3344 still listed compress.absorb alongside compress.imageCompression as a removed/inert key: old-config parsing dropped absorb and tests/removed-compress-keys-inert.test.ts asserted it parsed to undefined. Pull absorb out of INERT_REMOVED_COMPRESS_KEYS (imageCompression stays removed), correct the warning text, and update the test so absorb is asserted kept while only imageCompression warns.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

开了 absorb 之后,插件模式会话被指示调用一个永远不会出现的 absorb 工具(与 #3177 同症状家族,机制不同)

1 participant