[PR] feat: plugin lanes accept host declaration of exposed bili tools (#3906) by ranxianglei · Pull Request #3907 · ranxianglei/billion-context

ranxianglei · 2026-10-20T03:51:24Z

Step 1 of the #3884 Step3+ structural direction — owner-approved in the #3884 thread (2026-10-20). Branch is stacked on top of #3885 (the code_run fix it refactors); merging this PR pulls in #3885's two commits as ancestors — either merge order works, but #3885-first is the natural sequence.

What

Plugin lanes accept an optional inbound header x-bili-plugin-exposed-tools — the host's per-request declaration of which bili tools it actually exposed for this request/preset (comma-separated names; present-but-empty = "exposes none"). When present it is ground truth for the reachability decision: it outranks wire-shape inference entirely, profile default included. When absent, the legacy inference (top-level compress / CODE_MODE_OMNIBUS_TOOL_NAMES / profile default) applies bit-for-bit — owner constraint: a request without the header must not change behavior.

Behavior changes (disclosed)

Path	Old	New
plugin lane, header present	n/a	declaration decides: contains `compress` (or top-level `compress`) → surface on; else off — even against a recognized omnibus name
plugin lane, header absent	name-inference (three rungs)	unchanged, bit-for-bit
proxy lane	structurally on	unchanged (a stray header cannot turn it off)
outbound bodies	—	untouched
header egress	—	joins the existing x-bili-plugin* family behavior (that family is forwarded to upstream today; the pre-existing egress gap is tracked separately in #3905)

Why

#3600 and #3884 were the same judgment chain failing twice because the host renamed its code-mode omnibus tool (run_code → code_run). Name-matching can never be complete — every rename is a silent incident until someone reports it. Letting the host declare what it registered makes future renames a non-event. This is the first concrete step of the #3417/#3482 Step3+ ground-truth direction ("reachability truth = host manifest registration state"); host-side adoption (e.g. dsh code-mode presets stamping the header) follows separately and is not blocked by this PR.

Tests

tests/lane-decisions-b-class.test.ts: precedence rows — declaration rescues an UNRECOGNIZED omnibus shape; declaration beats a recognized-name inference (reverse direction); present-empty suppresses; absent keeps legacy both directions; top-level compress stays ground truth vs contradicting declaration; proxy immune; declaration outranks profile default (pi) — plus parser rows (absent/empty/trim/lowercase/array form).
tests/lane-decisions-matrix.test.ts: +2 e2e rows × 4 wires through the real server chain (declared-rescue → surface on; declared-suppress despite recognized omnibus → surface off).
Local targeted green: typecheck clean, b-class 7/7, matrix 74/74. Full suite = remote CI.

Config surface

None — new optional INBOUND header, not a config knob. No new env vars.

github-actions · 2026-10-20T04:38:59Z

📦 Built Package Artifact

Branch: 2026-10-20_exposed-tools-declaration (719bcf2)

Option A — Install this exact build (recommended)

npm install -g [email protected]

This exact version is immutable and reproducible. The rolling pr tag tracks the newest test build across all PRs.

Option B — Download artifact

Download the artifact from the Actions run
Extract the tarball and install:

tar xzf billion-context-pr3907.tgz
npm install -g package

Install instructions are refreshed on each push.

…le (#3884) #3603's single-literal gate (RUN_CODE_TOOL_NAME = "run_code") went stale when DSH 0.2.x moved Code Mode into the dsh-codemode plugin, whose transport is named `code_run` (the reserved name `run_code` belongs to dsh-tools, so the plugin registers its own). On every DSH 0.2.x code-mode session the reachability conjunct failed and the compress surface stayed off forever: nudge ARMED-SUPPRESSED at 42% usage with pendingT1 10x over threshold, zero INJECT, fully silent (#3884). - collect the known omnibus shapes in CODE_MODE_OMNIBUS_TOOL_NAMES (run_code / code_run), provenance-commented per shape; both detection sites match the set - lane-decisions reachability (the reported miss) and ptcNoteMode (proxy-lane exempt note, same-mechanism second site) - matrix: dsh code_run row -> surface on + EMERGENCY twin; existing precision rows (get_weather / acp_status) stay the negative direction - ptc-runcode-tools: code_run -> exempt note on all four wires; bash-only rows pin the negative direction (#3617 bidirectional discipline) Targeted green: typecheck clean; lane-decisions-matrix 66/66; ptc-runcode-tools 16/16; e2e-ptc-runcode-shape 8/8. Full suite = CI.

…ss surface off (#3884) Observability hardening for the allowlist fragility class: when an armed nudge stays ARMED-SUPPRESSED because the lane surface is off for reachability (not user config), the diag line now carries surface=off(no-tools) or surface=off(top-level=[names]) — the next host omnibus rename shows up in bili.log next to the suppression label instead of requiring user archaeology (#3600 and #3884 were both diagnosed from user-reported dumps). No behavior change beyond log text.

… Step3+)

ranxianglei · 2026-10-20T04:47:50Z

🤖 Powered by ework · qwen3.8-27b

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

结论先行：合并冲突已解除（GitHub 实时 mergeable=true），本次无需我再推送。 处理中发现该分支正被并发实例同时 rebase，我已独立复现并验证了等价解法、并确认远端已正确合到最新 master，因此收手以避免重复劳动和 SHA churn。

八问：

是什么问题：PR [PR] feat: plugin lanes accept host declaration of exposed bili tools (#3906) #3907 的分支落后于 master，产生合并冲突（系统转发 merge_conflict / dirty），需 rebase 到最新 master 解决后才能合入。
什么原因导致的：master 在分叉后前进了多个提交，其中 6 个文件与本 PR 改动重叠；真正的文本冲突只有 CHANGELOG.md 一处——master 在 [Unreleased] 顶部新插了「快照渲染影子对账」块，本 PR 也在同一位置插了「### Added」块，两边抢同一行 → 冲突。其余 5 个重叠文件（src/server.ts、四个 prepare-* 适配器）改动不重叠，git 可自动合并。
影响范围与严重程度：只卡住本 PR 能否合入，不影响任何线上行为或其他用户。属「阻塞合入」，非功能缺陷。
怎么修复的：把 PR 的 3 个提交 rebase 到当前 master；CHANGELOG 冲突按「并集」解——两块都原样保留（master 的影子对账块在前、PR 的 Added 块紧随其后），零丢失、零改写长条目。本地核对：typecheck 通过；PR 触及的三个测试文件（lane-decisions b-class / matrix + ptc-runcode-tools）共 97 例全绿。
是否有回归问题：无。相对真实最新 master，净 diff 仍是原 PR 的那 12 个文件、内容与原 PR 一致；auto-merge 的文件经 typecheck + 四 wire 矩阵测试确认无语义漂移；未触碰任何 master 新增逻辑。
是否破坏缓存：不涉及——纯 git 历史操作 + CHANGELOG 文本，无运行时/缓存路径。
是否改了设计方向：不涉及。
是否引入了新的配置：不涉及（本次是解冲突，非功能改动；PR 本身引入的 x-bili-plugin-exposed-tools header 属原 PR 范围，非本次操作）。

为什么我没有再 push（关键）：我本地开始 rebase 时，远端头还是旧的 3adb986f1；期间该分支已被并发实例 rebase 并强推了多次（我先后看到 1c66d99 → ffc415fe）。我独立完成的 rebase 结果与远端在本 PR 的内容上逐字节等价——唯一差异是并发方 rebase 到了更新的 master（现 tip 687096785，已并入 #3899 文档），因此多带了 #3899 对 README 的 15+15 行改动，那属于 master、不属于本 PR。而我的版本基于更旧的 master（838e975）：若推上去，PR 相对真实 master 会显示成删除 #3899 刚合入的 README 文档（倒退）。GitHub 实时状态现已是 mergeable=true / mergeable_state=blocked（冲突已消除；blocked=等待必需检查/分支保护，属正常待合入态）。故按去重/防重复劳动原则收手，不再强推。

一个观察（仅记录，未处置）：采样到的最近一次 CI 里 unused-exports 显示 failure。但采样时分支头仍在变动（我会话内已移动两次），该 CI 快照是瞬态的、且归属当前分支头的持有者，不在本任务（解冲突）范围内，我不据此立项；留待分支稳定后由对应流程复核。

判定：冲突已解除，无需我再操作；等 CI 全绿后由人合并即可。

The #3884 refactor moved ptcNoteMode off the direct constant onto CODE_MODE_OMNIBUS_TOOL_NAMES, leaving RUN_CODE_TOOL_NAME referenced only inside compress-tool.ts (building that set). The unused-exports gate (#1440 P6) flags it as INTERNAL_ONLY; remove the leading export. No behavior change - nothing outside the module imported it.

ranxianglei · 2026-10-20T05:08:11Z

🤖 Powered by ework · qwen3.8-27b

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

CI auto-fix — resolved the failing unused-exports check directly on the PR branch (commit 719bcf215).

是什么问题
PR head 的 unused-exports 门禁（#1440 P6）红：src/compress-tool.ts 里 RUN_CODE_TOOL_NAME 被判定为 INTERNAL_ONLY —— 导出了但没有任何模块外部引用它。

什么原因导致的
本 PR 的 #3884 重构把 ptcNoteMode() 从「直接比对 RUN_CODE_TOOL_NAME」改成「查 CODE_MODE_OMNIBUS_TOOL_NAMES 集合」。改完之后，RUN_CODE_TOOL_NAME 只剩文件内部一处用法（第 278 行，用来构建那个集合），但行首的 export 还留着 → 门禁按规则报 INTERNAL_ONLY。

影响范围与严重程度
纯 lint/卫生类门禁失败，卡 CI；对运行时行为零影响。该常量本来就已经只在模块内使用，去掉 export 不改变任何运行期语义。

怎么修复的
把 RUN_CODE_TOOL_NAME 行首的 export 去掉（变回模块内常量）。它仍然被定义、仍在文件内用于构建 CODE_MODE_OMNIBUS_TOOL_NAMES，所以不是死代码，只是不再作为对外导出面暴露。一行改动。

是否有回归问题
无。全仓 grep 确认零处 import { RUN_CODE_TOOL_NAME }；仅有的两处引用都是它自身定义和构建集合的那一处，均在 src/compress-tool.ts 内。行为完全不变。本地已跑同一门禁脚本：✓ unused-export gate: clean（exit 0）。

是否破坏缓存
不涉及——未触碰任何缓存 / 重试 / 增量游标 / 去重键逻辑。

是否改了设计方向
无。保留本 PR 自己的设计：CODE_MODE_OMNIBUS_TOOL_NAMES 依旧是对外导出的探测集（lane-decisions 与 ptcNoteMode 消费它）；只是不再导出一个纯内部构件。

是否引入了新的配置
无新增 env / config / 默认值。

已在 PR 分支推送修复（ffc415fe6..719bcf215），失败的 unused-exports job 本地复跑已通过；等远程 CI 在新 head 上全绿。本轮为自动修复通道第 1 次尝试。

ranxianglei · 2026-10-20T05:22:41Z

🤖 Powered by ework · qwen3.8-27b

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

查重：不重复。 独立检索 open+closed（exposed-tools、宿主声明等关键词）：#3905 = 存量 x-bili-plugin* 头被转发上游的卫生缺口，#3884/#3885 = run_code→code_run 事故与名字收敛修复，本 PR(#3906) = 声明通道机制本身——三者各管各的，无第四个覆盖同一机制的条目。

标签：pr / dsh / scope/wiring / p1 —— 动机与落地客户端是 dsh；改动面是新增入站头的协议边界；p1 因为本 PR 叠在 #3885 之上，合入即同时交付 DSH 0.2.x code-mode 压缩面从「全关」恢复可用的 p1 止血。

初审状态（完整八问结论随后）：

diff 已通读（4 提交、12 文件 +243/-38）：实现与 Plugin lanes: host declares exposed bili tools per request via x-bili-plugin-exposed-tools (structural fix for #3884/#3600 name-guessing) | 插件车道：宿主按请求声明暴露的 bili 工具（根治 #3884/#3600 认名字脆弱性） #3906 设计逐条对齐——声明在场即 ground truth（top-level compress 仍为调用事实）、缺席走旧推断位级不变、proxy 车道免疫、解析 trim+lowercase。测试按 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 纪律双向断言齐全（b-class 优先权正反行 + 矩阵 2 行×4 wire 过真实 server 链）。
CI：head ffc415fe6 的 unused-exports 红已被作者用 719bcf215（去掉 RUN_CODE_TOOL_NAME 冗余 export，一行，我独立 grep 验证无外部引用）修掉；当前无失败项，剩 3 个 test shard 在跑，绿了我在最终结论里确认。
一个需要 owner 拍板的点（不阻塞审查）：你在 [BUG] 0.1.202 DSH 0.2.x code-mode channel: omnibus is actually named code_run, RUN_CODE_TOOL_NAME hardcoded to `run_ | [BUG] 0.1.202 DSH 0.2.x code-mode 通道：omnibus 实际名为 code_run，RUN_CODE_TOOL_NAME 硬编码 run_code ⇒ 压缩面永久压制（#3603 修复复发） #3884 批这个方案时的约束是「新头不向上游转发」。实测现状是整个 x-bili-plugin* 族今天都在被主 HTTP 车道转发给上游模型端点（含本新头），存量缺口由 bili forwards all x-bili-plugin* host headers (and its own chain markers) to upstream model endpoints — egress strip gap | bili 把全部 x-bili-plugin* 宿主头（及自身链标记）转发给上游模型端点——出站剥离缺口 #3905 跟踪。PR 选择「新头随族同生命周期、随 bili forwards all x-bili-plugin* host headers (and its own chain markers) to upstream model endpoints — egress strip gap | bili 把全部 x-bili-plugin* 宿主头（及自身链标记）转发给上游模型端点——出站剥离缺口 #3905 收敛」并已披露。我的建议：维持现状、归 bili forwards all x-bili-plugin* host headers (and its own chain markers) to upstream model endpoints — egress strip gap | bili 把全部 x-bili-plugin* 宿主头（及自身链标记）转发给上游模型端点——出站剥离缺口 #3905 一并收——因为 bili→bili 多跳场景下链标记需要「对下一跳 bili 留、对真上游剥」的车道感知判断，现在只剥这一个头会在第 2 跳丢掉声明，反而引入新退化；bili forwards all x-bili-plugin* host headers (and its own chain markers) to upstream model endpoints — egress strip gap | bili 把全部 x-bili-plugin* 宿主头（及自身链标记）转发给上游模型端点——出站剥离缺口 #3905 里一次做对。若你希望本 PR 内就先无条件剥掉新头（代价：多跳部署丢声明、退化为旧推断），说一声我就补 commit。
叠加关系核实：[PR] fix: dsh 0.2.x code-mode omnibus code_run reads as compress-reachable (#3884) #3885 两个提交是本分支祖先，两种合并顺序都成立，[PR] fix: dsh 0.2.x code-mode omnibus code_run reads as compress-reachable (#3884) #3885 先行是自然顺序——与 PR 描述一致。

ranxianglei · 2026-10-24T16:22:47Z

(ai: zhipuai-lb/glm-5.3)

被超越:master 已含完整演进版(CODE_MODE_OMNIBUS_TOOL_NAMES/#4344/#4372),空挑证实无独特内容。关单收尾;误判请 reopen。

ranxianglei changed the title ~~feat: plugin lanes accept host declaration of exposed bili tools (#3906)~~ [PR] feat: plugin lanes accept host declaration of exposed bili tools (#3906) Oct 20, 2026

ranxianglei force-pushed the 2026-10-20_exposed-tools-declaration branch from 3adb986 to 1c66d99 Compare October 20, 2026 04:38

awork added 3 commits October 20, 2026 12:39

feat: plugin lanes accept host declaration of exposed bili tools (#3884…

ffc415f

… Step3+)

ranxianglei force-pushed the 2026-10-20_exposed-tools-declaration branch from 1c66d99 to ffc415f Compare October 20, 2026 04:39

ranxianglei added dsh p1 scope/wiring labels Oct 20, 2026

ranxianglei closed this Oct 24, 2026

Navigation Menu

Repository navigation

[PR] feat: plugin lanes accept host declaration of exposed bili tools (#3906) - #3907

[PR] feat: plugin lanes accept host declaration of exposed bili tools (#3906)#3907
ranxianglei wants to merge 4 commits into
masterfrom
2026-10-20_exposed-tools-declaration

ranxianglei commented Oct 20, 2026

Uh oh!

github-actions Bot commented Oct 20, 2026 •

edited

Loading

Uh oh!

ranxianglei commented Oct 20, 2026

Uh oh!

ranxianglei commented Oct 20, 2026

Uh oh!

ranxianglei commented Oct 20, 2026

Uh oh!

ranxianglei commented Oct 24, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

ranxianglei commented Oct 20, 2026

What

Behavior changes (disclosed)

Why

Tests

Config surface

Uh oh!

github-actions Bot commented Oct 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

📦 Built Package Artifact

Option A — Install this exact build (recommended)

Option B — Download artifact

Uh oh!

ranxianglei commented Oct 20, 2026

Uh oh!

ranxianglei commented Oct 20, 2026

Uh oh!

ranxianglei commented Oct 20, 2026

Uh oh!

ranxianglei commented Oct 24, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

github-actions Bot commented Oct 20, 2026 •

edited

Loading