[ranxianglei]

Closes #3433

What this does

The incremental-turns lane (#3248) is end-to-end Anthropic-wire-only: it ingests with anthropicToCore, re-emits with coreToAnthropic, and forwards the assembled body to the ORIGINAL request path. But two host entry points claimed "anthropic wire only" without enforcing it:

• src/agent/dsh-native.ts's maybeIncrementalNative checked only Array.isArray(messages) + typeof model === "string" — an OpenAI chat-completions body passes both;
• src/agent/pi.ts's maybeIncrementalTurn (shared by pi AND pi-native by design) had the same gap, and pi genuinely serves openai/responses/google provider APIs.

With the rollout flag on, an OpenAI conversation would be converted onto the lane where: assistant.tool_calls are silently dropped, role:"tool" messages are mis-mapped into user text blocks (tool_use_id association lost), and content:null bodies throw mid-apply. omp/opencode already guarded with wireMessagesLookAnthropic; this brings dsh/pi to parity and enforces the contract at the trust boundary too.

Changes

1. Client guards (the issue's minimal fix): wire the existing wireMessagesLookAnthropic into dsh-native (src/agent/dsh-native.ts) and pi (src/agent/pi.ts). Non-anthropic shapes now stay on full resend — exactly what PLUGIN.md documents.
2. Lane boundary enforcement: src/incremental-lane.ts rejects non-anthropic message shapes with 400 (laneMessagesLookAnthropic, a keep-in-sync mirror of the client helper — same pattern as REQUEST_PARAM_KEYS / INCREMENTAL_TURN_MARKER_HEADER). No host can corrupt the lane's canonical state anymore; the empty-messages sticky-system-update case stays legal (deliberate divergence from the client predicate, documented in the docstring).
3. Regression tests:
   • tests/dsh-native.test.ts — OpenAI shape (tool_calls + role:"tool") rides full resend untouched, no marker; anthropic control still converts.
   • tests/issue3433-pi-incremental-guard.test.ts (new) — real proxy with the lane LIVE (unified storage + rollout forced on) so the probe passes and the GUARD is what decides: OpenAI shape not replaced; anthropic control converts (seed, sinceSeq 0).
   • tests/incremental-lane.test.ts — lane 400s role:"tool" and tool_calls bodies with a clear error, applies nothing, still accepts a valid seed afterwards.

Behavior changes (old → new)

Path	Old	New	Impact today
dsh/pi incremental entry, non-anthropic body	converted onto the anthropic-only lane → tool calls dropped / role:"tool" corrupted / content:null throws	stays on full resend (documented contract)	none — rollout pluginIncrementalTurns is at percent 0, no traffic reaches this code
POST /__bili/plugin/turn (+ marker-diverted turns) with non-anthropic message shapes	TypeError or silent corruption inside applyTurn	400 with explicit error	none — correct hosts never send these shapes; a pre-fix host build gets a clean 400 instead of corruption
dsh empty messages array	passed the array check into planIncrementalTurn	rejected client-side (parity with omp/opencode)	dsh system-only updates ride full resend like every other lane's do

No config surface changes. No kernel/ changes. No version bump.

Verification

• npm run typecheck clean (kernel + pi-subagents + bili).
• Targeted local runs: tests/dsh-native.test.ts 55/55, tests/incremental-lane.test.ts 18/18, tests/issue3433-pi-incremental-guard.test.ts 1/1, adjacent tests/pi-omp-fork-adopt.test.ts 22/22. Full-suite verdict = CI on this head.

Out of scope (tracked as a decision request in #3433): making the lane itself format-aware (detectWireFormat dispatch to openaiToCore/coreToOpenai + per-conversation format recording) — the "real fix" that would let OpenAI lanes actually USE incremental turns; a feature/wire change requiring owner sign-off.

[github-actions]

📦 Built Package Artifact

Branch: 2026-10-17_dsh-pi-incremental-anthropic-guard (6f9956c)

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

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf billion-context-pr3440.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

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

查重：不重复。 open+closed 的 issue/PR 全搜了 wireMessagesLookAnthropic 与 incremental-lane 相关项：只与本 PR 的同源 issue #3433 及已合入的 #3292（phase-2 rollout，不同范围）相关；在途的 #3414/#3340/#3257/#3349/#3355 症状与机制均不同（writeNow 竞态测试、ack 归属、prefix fast path、marker-divert 管理面、文档遗留），无重复。

审核方式：diff 全文通读（6 文件 +229/-3，单 commit 6f9956cdc，分支与 master 同步无落后）；对照原 issue #3433 逐层验证修法是否打到根因；本地定向跑测试（dsh-native 55/55、incremental-lane 18/18、issue3433-pi-incremental-guard 1/1、相邻 pi-omp-fork-adopt 22/22——首跑一次失败是本地 pi-subagents 未构建的环境问题，构建后通过，非代码缺陷）；CI 在本 head 全绿（static-gates + 6 个 test shard + docs/config/lock 门禁，e2e 门控 lane 按设计 skip）；与其他在途 PR 无文件重叠。

八问：

1. 是什么问题：增量车道端到端是 Anthropic-wire-only（anthropicToCore 进、coreToAnthropic 出），但 dsh 和 pi 两个宿主入口只查了「messages 是数组 + model 是字符串」——OpenAI chat-completions body 原样满足这两条，开启 pluginIncrementalTurns 后 OpenAI 会话会被转进这条车道，assistant.tool_calls 被静默丢弃、role:"tool" 被错映射成 user 文本块。
2. 什么原因导致的：两层。直接原因是接线遗漏——守卫函数 wireMessagesLookAnthropic 已存在且 omp/opencode 已接入，dsh-native/pi 漏接（bundle grep 计数为 0）；结构性原因是车道服务端信任客户端预筛，信任边界自身没有校验。
3. 影响范围与严重程度：今天零用户影响（rollout pluginIncrementalTurns percent 0，流量到不了这段代码）。但它是潜在静默数据损坏——一旦 rollout 就会命中最常见的 wire shape（报告者机器上 openai 262 个会话 vs anthropic 2 个），且损坏发生在 applyTurn 内部（丢 tool call / 错映射），不易从表面察觉。严重度随启用而升级。
4. 怎么修复的：双层契约强制。① 客户端：把现有 wireMessagesLookAnthropic 接进 dsh-native 和 pi 两个入口，非 anthropic shape 留在 full resend（正是 PLUGIN.md 文档化的契约，pi-native 经 createBiliPlugin() 共享同一守卫一次覆盖）；② 服务端：车道加 laneMessagesLookAnthropic（keep-in-sync 镜像，允许空数组以保 sticky-system-update 合法，分歧已在 docstring 注明），非 anthropic shape 在任何状态变更前直接 400。「真正的修复」（车道 format-aware，detectWireFormat 分发）明确划出范围，作为决策请求留在 [bug] dsh-native not wired to wireMessagesLookAnthropic: Enabling pluginIncrementalTurns on the OpenAI lane silently dro | [bug] dsh-native 未接入 wireMessagesLookAnthropic：OpenAI 车道启用 pluginIncrementalTurns 会静默丢弃 tool_calls / role:"tool" #3433 等 owner 拍板——无决策依赖部分先修、需决策部分挂起，拆分正确。
5. 是否有回归问题：逐项清点触碰路径的既有行为——anthropic shape 转换行为不变（两个测试的 control 断言证明转换仍发生）；full-resend 回退语义保持（return undefined = 不替换、不打 marker，与该 handler 其他早退一致）；车道的空 messages + system sticky 更新仍合法；400 发生在 turn lock 和任何状态变更前（测试断言拒绝后有效 seed 正常落 seq 1）；客户端对 400 有现成处理（noteIncrementalRejection 熔断：reset → 下一轮 reseed，连续 strike 才 dead）→ 旧版宿主 build 拿到干净 400 + 自动回退 full resend，不坏会话。方向是 fail-soft → fail-fast，与高危模式（吞错化）相反。PR 披露的两处行为变更均核对无误：① dsh 空 messages 数组原来会传进 planIncrementalTurn，现在留 full resend（退化输入，full resend 恒安全）；② 车道收非 anthropic shape 从 TypeError/静默损坏变显式 400。
6. 是否破坏缓存：不涉及。拒绝先于 seq/状态变更，缓存键、重试策略、增量游标、去重键均无变化。
7. 是否改了设计方向：否。车道架构、数据流、模式判定均未动；本 PR 只是把 PLUGIN.md 文档化的契约从「文档承诺」变成「代码强制」。镜像函数模式与既有 REQUEST_PARAM_KEYS 镜像一致。
8. 是否引入了新的配置：无。无 env/config/CLI 新增，无 version bump，无 kernel/ 改动。

测试有效性：三组测试双向有牙——既断言回归（OpenAI shape 不被替换 / 车道 400 且不落任何状态），也断言 control（anthropic shape 仍转换 / 有效 seed 仍正常落地），去掉守卫或过度拦截都会红；pi 测试用真实 proxy 且强制车道 LIVE（unified storage + rollout=1），确保 probe 通过、由守卫本身裁决，隔离正确。

结论：可合。 代码、diff 整洁度（每行都服务于目的，无空白/锁文件噪音）、测试质量全部过关；CI 在本 head 全绿。无阻塞项。

[ranxianglei]

(ai: zhipuai-lb/glm-5.3)

已并入 #4590。关单收尾;误判请 reopen。