[ranxianglei]

Addresses the DSH-lane e2e-assertion ask in #2997 §4, and records the triage verdict for the reported post-fold follow-up excess.

What this adds

One hermetic regression test, tests/openai-plugin-postfold-seam.test.ts, on the ONE compression lane #2755/#2741 never covered: DSH-native = plugin mode over /v1/chat/completions (folds arrive via POST /__bili/plugin/tool; rebuild path prepare-openai). It drives a session, fires a multi-disjoint-range fold (production cadence), and asserts the cache-seam invariant §4 asks for:

• the pinned first user message never moves (index stable across every request);
• no summary ever precedes the pinned first user message;
• consecutive steady-state post-fold builds are byte-identical except at the tail — i.e. the provider prefix cache cannot bust early.

Carrier-agnostic: holds whether the seam carries the kernel acp_summary or the client's compress tool pair, so it guards the lane under either host behaviour.

Triage verdict for #2997 (why the test PASSES)

A hermetic sweep over every fold shape I could construct — single, big-contiguous, multi-disjoint-range (up to 9 stacked blocks), and partial folds — shows the bili openai-chat plugin-mode rebuild is deterministic and prefix-stable: the pinned head is byte-identical across consecutive builds and no summary drifts ahead of the pin. The reported 12–17% follow-up hit therefore does not reproduce against the proxy's own output, which points the root cause OUTSIDE bili's rebuild:

• dsh client-side re-serialization (the log's own median is 96.6%, not 100%),
• dsh native compaction-basic, which stays ARMED unless the fix(dsh): profile 层 patch 其实可以关掉 web profile 的原生自动压缩（#1772 结论修正）+ 一个会静默清空整个 preset 的写法坑 #2474 override is applied (the launcher warns about this explicitly),
• provider-side cache keying/TTL/cap on the opencode.ai / deepseek endpoint.
  None of those is modelable in a hermetic proxy unit test; isolating them needs the real dsh binary + real upstream.

Scope / non-goals

• Test-only (test:). No source changes, no config surface touched.
• Does not fix the production symptom (it is outside bili's rebuild); it adds the requested coverage and pins the invariant so any future bili-side regression that reintroduces early-prefix drift fails here.
• The §5 PARTIAL-fold double-seam cost question is a distinct failure domain (commit-granularity tradeoff, needs a design decision) — tracked separately, not bundled here.

Relates to #2997 (addresses §4).

[github-actions]

📦 Built Package Artifact

Branch: 2026-10-13_dsh-lane-postfold-seam-backstop (a0bd762)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.3400-pr.3002.541c1e3

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-pr3002.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review 完成 — 结论:可合。 已拉取分支 pr-3002(单 commit a0bd762,diff = 唯一新增文件 tests/openai-plugin-postfold-seam.test.ts +198 行,纯测试)并做了定向验证。

我做了什么(证据)

• 本地定向运行:node --import tsx --test tests/openai-plugin-postfold-seam.test.ts → pass 1 / fail 0(2.2s,17 requests all head-stable)。
• 变异验证(test 有效性,断言有牙):临时在 src/server/prepare-openai.ts:226 之后的重建结果里给第一条 user 消息注入逐次漂移 nonce(模拟 head drift),测试如预期失败:steady post-fold request[15] diverges from request[14] at item[1] (< 27) — head/summary drift busts the prefix cache —— item[1] 正是 pinned first user message。即该测试能抓住与 [cache] 0.1.196 (d64139f9) 仍复现 post-fold follow-up excess —— #2741/#2755 修复后 2/2 大折叠复发 #2997 同款的回归签名。已还原,工作树干净(git status 空)。
• 源码核对重建路径与 PR 描述一致:src/server/prepare-openai.ts:201(stripKernelSummaries)→ :226(coreToOpenai/systemToUser)→ :234-243(静态 sysParts + injectOpenaiSystem;nudge 走尾部 user 消息保 system byte-stable);plugin fold 入口 POST /__bili/plugin/tool(src/server/admin.ts:459)。
• 覆盖缺口声明成立:现有 seam 类测试(responses-fold-cache-seam、cache-seam-*、abort-rewind-cache-seam)全是 Responses-wire/proxy-mode;host-usage-postfold 直接驱动 loop adapter(unit 级,不经 server 重建);dsh-native.test 只覆盖 install/attach。openai-chat plugin-mode 的 server 级 post-fold seam 此前确实无人钉住。
• CI:head commit 全部 19 个 check-run 绿(test ×4 矩阵 ubuntu/windows × node22/24、static-gates、kernel-guard、lane-guard、lock-sync、version-guard、config-surface-guard、pi-subagents-guard、build-artifact 等)。
• 查重:与既有 seam/backstop 测试及 fix(kernel): fold summaries may never precede the pinned first user message (#2733) #2755/post-fold double prefix-cache penalty: the request AFTER a fold matches only the ~11K head (one-build wire drift after compress) #2741 后续无重复立项。

结构化八问

1. 是什么问题:[cache] 0.1.196 (d64139f9) 仍复现 post-fold follow-up excess —— #2741/#2755 修复后 2/2 大折叠复发 #2997 报告 DSH-native lane(openai-chat wire 上的 plugin mode)折叠后首个请求缓存命中仅 11.7%~17.3%(应接近全前缀命中),§4 要求给这条 lane 补 e2e 断言。本 PR 补上该断言(proxy 侧不变量形式),并记录 triage 判定。
2. 什么原因导致的:triage 判定为不在 bili 重建内——对全部可构造 fold 形态(single/big-contiguous/multi-disjoint/partial,最多 9 块堆叠)的 hermetic sweep 显示 openai-chat plugin-mode 重建是确定性且前缀稳定的,proxy 自身输出复现不出 12–17% 的 follow-up 缺口;嫌疑指向 proxy 输出之外:dsh 客户端侧重序列化(其日志中位数本就只有 96.6%)、dsh native compaction-basic 未应用 fix(dsh): profile 层 patch 其实可以关掉 web profile 的原生自动压缩（#1772 结论修正）+ 一个会静默清空整个 preset 的写法坑 #2474 override 仍 armed、provider 侧(opencode.ai/deepseek)cache keying/TTL/cap。我的独立核对(重建代码路径 + 定向运行 + 变异复现)支持该判断的合理性;生产症状本身需要真 dsh 二进制 + 真 upstream,hermetic 测试无法建模,PR 已如实声明此边界。
3. 影响范围与严重程度:纯测试新增,零运行时影响;填的是覆盖缺口(DSH-native/openai-chat plugin lane 此前无 post-fold seam 回归守卫)。对用户无可见变化。
4. 怎么修复的:一个 hermetic 回归测试——真实 server + mock upstream,x-bili-plugin header 驱动 plugin mode,经 POST /__bili/plugin/tool 打 multi-disjoint-range fold(生产节奏),随后钉三条不变量:pinned 首条 user 消息索引全程稳定、任何 summary 不得先于 pin、连续 steady-state post-fold 构建除尾部外逐字节前缀一致。Carrier-agnostic(acp_summary 或 client tool pair 承载都守得住)。端口用 listen(0)、XDG env 重定向到 tmpdir,均符合仓库既有纪律(fix: emit /acp panel as persistent custom message and strip it from the proxy model context (#359) #360)。
5. 是否有回归问题:无。审过的既有行为:diff 不触碰任何源文件 ⇒ 不存在被改动的既有行为(语义漂移/fail-fast 变吞错/阈值时序/输出格式漂移各高危模式逐项过,均不适用)。CI 全绿。
6. 是否破坏缓存:不涉及——未触碰任何 cache/retry/增量游标/去重键逻辑。
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。

两个非阻塞观察(记录在案,不拦合并)

• 断言比较的是解析后的 .messages items 而非原始 body 字节(tools/system 字段不在比较内)。该 lane 的动态内容都在 messages 里,且 system prompt 按设计本来就 byte-stable(:228-233 注释),所以可接受;若未来想更严,可扩展到整 body 比较。
• 断言 (3)(summary topic 顺序稳定)在逻辑上被 (2) 的前缀一致蕴含,保留的价值是失败时诊断更清晰——合理。

边界确认:§5 的 PARTIAL-fold double-seam 成本问题确属另一失败域(commit-granularity 取舍,需设计决策),PR 正确地没有捆绑进来,继续在 #2997 下跟踪即可。

判定:可合。 测试有牙(变异验证通过)、范围干净、CI 全绿、triage 判定证据链完整。

[github-actions]

📦 Merged Build Published

This merge to master was published as 0.0.391-master.026997b (rolling master tag).

npm install -g billion-context@0.0.391-master.026997b  # 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.