@huai_xin/dsh-workbuddy 0.1.11 → 0.1.13

package/README.md

@@ -306,15 +306,6 @@ node test/live/policy-probe.mjs       # 两版白名单与准入结果对照
 - **思考档位只认服务端声明的 `reasoning.supportedEfforts`**（两个源合并后国内 15/34、国际 21/28）；只声明了默认档的模型不给档位列表 —— 与网页端行为一致。合并后仍无档位的，请求里不带 `reasoning_effort`，由服务端取默认。
 - 实时**繁忙程度**只有大陆版有（`POST /v2/chat/busy`，海外版 404，`regions.js` 的 `hasBusy` 决定那一列是否显示）。它**不是可用性信号**：只对白名单内的活跃模型返回 `idle/normal/busy/congested`，其余一律 `unknown`（面板显示「未知」，不当成故障）。插件侧 50 秒缓存，与官方前端一致。
 - 两版**条目数会变**：2026-10-22 实测大陆版 30 个（`ok` 22 / `dead` 8）、海外版 25 个（`ok` 13 / `denied` 12）。`gpt-5.3-codex` 只在账号名单里，两个 `/v3/config` 都没有它，倍率因此是「未知」。
-- **长上下文下模型会"说话说一半就收尾"（上游行为，非插件问题）**：海外版 `deepseek-v4.1-flash` 在上下文很长时会不调用工具、只用一句话叙述下一步然后 `finish_reason: "stop"`（例如「我需要先读取该文件，才能准确修改。」）。实测（2026-10-22，`node test/live/long-context-probe.mjs`）：
-
-  | 输入规模 | 正确调用工具 | 提前 `stop` |
-  | --- | --- | --- |
-  | 0k（1 条消息） | 2/2 | 0/2 |
-  | 30k 字符（613 条） | 2/2 | 0/2 |
-  | 89k 字符（1825 条） | 1/2 | **1/2** |
-
-  用户侧复现阈值约 **120k 上下文**起明显（其它模型未测）。判据是**换了上下文长度、其它一切不变**即可复现/消失，且流里**一个 tool-call 分片都没有**、`finish_reason` 就是 `stop` —— 所以既不是适配器丢帧（丢帧会看到 `finish_reason: "tool_calls"`），也不是采样（四档温度 16/16 正常）、不是工具 schema（70 个工具 / 单个 / 简化 schema 9/9 正常）。`stop` 是模型自己采到 EOS，属于上游质量问题；缓解手段只有换模型、把会话压短（harness 的压缩），或回一句"继续"。
 
 ## 上游行为与错误映射
 
@@ -359,7 +350,6 @@ node test/live/stream-cut-probe.mjs cn --selftest     # 离线验证"断流"分
 node test/live/stream-cut-probe.mjs intl 10 --long    # 真实压测断流率（免费模型零消耗）
 node test/live/find-message.mjs "关键词" <会话前缀>     # 按文本定位某一步，并打出前后事件
 node test/live/dump-stream.mjs <会话前缀> <seq,…>      # 打印那一步存下来的原始流帧
-node test/live/sampling-probe.mjs intl 4              # 同一提示在各 temperature 下的"提前 stop"率
 ```
 
 ## 许可

package/docs/workbuddy-protocol.md

@@ -394,7 +394,7 @@ data: [DONE]
 | INTL scoped → `500`（openresty/APISIX，HTML 错误页） | 2026-10-21 起持续；`/v3/config` 正常 ✅ |
 | `glm-5.1` 图片能力两个端点矛盾 | scoped 报 `false`、`/v3/config` 报 `true`，实测能看图（📖 参考实现据此写了覆盖白名单） |
 | 目录包含渠道不服务的模型 | 见 `11102`（§6）✅ |
-| 长上下文下"说话说一半就收尾"（`finish_reason: stop`，无工具调用） | 上游/模型行为，**不是插件问题**。实测：同一提示、同一套工具、只改前面的对话长度 —— 0k 2/2 正常、30k 1/2（一次 181 字后 `stop`）、89k 1/2；每次 finish 诊断块都标 `workbuddy-intl/deepseek-v4.1-flash`。判据：流里**没有任何 tool-call 分片**、`finish_reason` 就是 `stop`（若丢帧会是 `tool_calls` 且能看到分片）。已排除：采样（四档温度 16/16 正常）、工具 schema/数量（70 个 / 单个 / 简化 schema 9/9 正常）、提示词校验（海外版无此错误，请求体按原样转发）。复现：`node test/live/long-context-probe.mjs 2`；排查：`node test/live/toolcall-audit.mjs <会话前缀>`、`sampling-probe.mjs`、`tools-probe.mjs`。缓解只有换模型/压缩上下文/回一句"继续" —— 合法的 `stop` 不能当错误重试。 |
+| 长上下文下"说话说一半就收尾"（`finish_reason: stop`，无工具调用） | 上游/模型行为，**不是插件问题**。实测：同一提示、同一套工具、只改前面的对话长度 —— 0k 2/2 正常、30k 1/2（一次 181 字后 `stop`）、89k 1/2；每次 finish 诊断块都标 `workbuddy-intl/deepseek-v4.1-flash`。判据：流里**没有任何 tool-call 分片**、`finish_reason` 就是 `stop`（若丢帧会是 `tool_calls` 且能看到分片）。已排除：采样（四档温度 16/16 正常）、工具 schema/数量（70 个 / 单个 / 简化 schema 9/9 正常）、提示词校验（海外版无此错误，请求体按原样转发）。复现：`node test/live/long-context-probe.mjs 2`；排查：`node test/live/toolcall-audit.mjs <会话前缀>`、`long-context-probe.mjs`、`repro-probe.mjs`。缓解只有换模型/压缩上下文/回一句"继续" —— 合法的 `stop` 不能当错误重试。 |
 
 ### 4.7 模型选择器里每一项元数据的来源
 
@@ -876,7 +876,7 @@ provider 与全部 HTTP 路由都没注册，设置面板只剩一个 `HTTP 404`
 | `prompt_cache_key` | 发（会话级） | **不发** | 反向差异；实测对本问题无影响（去掉后 1/8 vs 2/8），保留以提高前缀缓存命中 ✓ |
 | `stop` | 仅调用方给 | 不发 | 等价 ✓ |
 | `store` / `prompt_cache_retention` | 不发 | 发（配置值，通常未设置） | 等价 ✓ |
-| 请求头 | `Authorization` / `Accept` / `X-Domain` / `X-Product` / `X-Product-Code` / `User-Agent` | 同类内部头（`X-User-Id`/`X-Enterprise-Id`/`X-Tenant-Id`/`X-Department-Info`/`X-Domain`/`User-Agent`），且会剔除内部头再转发 | 我们的头是目录/聊天都验证过可用的最小集合 ✓ |
+| 请求头 | `Authorization` / `Accept` / `X-Domain` / `X-Product` / `X-Product-Code` / `X-Agent-Purpose` / **`X-Agent-Intent: craft`** / `X-Agent-Runtime` / `X-IDE-Name` / `X-IDE-Type` / `X-IDE-Version` / `X-Session-ID` / `X-Model-ID` / `X-Conversation-ID` / **`X-Conversation-Message-ID`** / `X-Conversation-Request-ID` / **`X-Request-ID`** / `User-Agent` | 同类内部头（`X-User-Id`/`X-Enterprise-Id`/`X-Tenant-Id`/`X-Department-Info`/`X-Domain`/`User-Agent`），转发前还会**剔除**内部头 | **网关会校验**：2026-10-23 实测，裸请求缺 `X-Agent-Purpose`/`X-Conversation-*`/`X-IDE-*` → `400 code 11128 request illegal`。已按官方常量对齐（`X-Agent-Intent` 默认 `craft`；`X-Conversation-Message-ID`/`X-Request-ID` 标识"被回答的那条消息"，用最后一条 user 消息的哈希，重试时保持不变） |
 
 **结论：结构上不是"完全一样"，但没有一处会造成"提前收尾"** —— 唯一的实质差异是官方总带采样参数，
 而实测它对这个问题无效。`refusal` 的含义见 §12.2。
@@ -916,6 +916,22 @@ provider 与全部 HTTP 路由都没注册，设置面板只剩一个 `HTTP 404`
 | `test/live/picker-contract-check.mjs` | 适配器实际给模型选择器什么（`providerInfo` / `listModels` / `resolveModel`） |
 | `test/live/harness-contract-replay.mjs` | 复刻 harness 校验规则，定位「选择器整组消失」 |
 | `test/live/dead-code-report.mjs` | 死代码审计：未使用的 CSS 类、词典键、导出 |
+| `test/live/installed-package-check.mjs [dir]` | 检查**已安装副本**（真目录、依赖为空、provider 注册、选择器可选项、资源齐全） |
+| `test/live/leak-scan.mjs [ref]` | 发布前必跑：工作区 + git 历史的账户信息/凭据泄露扫描 |
+| **会话取证（"回答为什么中断"）** | |
+| `test/live/session-tail-live.mjs [前缀] [步数] [项目过滤]` | 读最新会话的末尾几步；容残帧（`ZSTD_e_flush`），带 `stopReason`/`outputTokens`/正文 |
+| `test/live/last-finishes.mjs [前缀] [步数] [项目过滤]` | 只看每步的 `upstream.stopReason` 与"stop 且无工具调用"的步，附分布统计 |
+| `test/live/find-message.mjs "关键词" [会话前缀]` | 按正文定位某一步，并打出前后事件 |
+| `test/live/dump-stream.mjs [会话前缀] <seq,…>` | 打印某一步存下来的原始流帧（含 finish 的 `response` 诊断块） |
+| `test/live/toolcall-audit.mjs [会话前缀] [步数]` | 上游宣告了工具调用、但一个分片都没交付的"丢失指纹" |
+| **流与请求结构实验（改协议相关代码前跑）** | |
+| `test/live/stream-cut-probe.mjs [cn\|intl] [次数] [模型] [--selftest]` | 断流分类：`--selftest` 离线自检，长回答模式可压测断流率 |
+| `test/live/repro-probe.mjs [次数] [模型] [填充字符]` | 长上下文 + 真实 70 工具集下"是否调用工具"，每行打印原始 `finish_reason` |
+| `test/live/long-context-probe.mjs [次数] [模型]` | 只改上下文长度，看提前收尾率（可复现的对照实验） |
+| `test/live/context-size-probe.mjs [字符数…]` | 填充上下文实际值多少 token（网关 `inputTokens` 为准） |
+| `test/live/thinking-field-probe.mjs [次数] [模型] [部署]` | 带/不带 `thinking` 字段时思考量对比（档位才是开关） |
+| `test/live/request-structure.mjs` / `test/live/free-models.mjs` | 各厂商请求结构、免费档清单的逐字段试探 |
+| `test/live/verify-policy.mjs` | 实时目录扣掉台账里 `dead` 的判定后，两版的条目数对照 |
 | **契约与 UI 参考（写 UI 插件用）** | |
 | `test/live/slot-inventory.mjs [outFile]` | 列出客户端实际渲染的所有插槽 |
 | `test/live/model-seat-kind-probe.mjs` | 模型席位的 kind 与占用者（决定能不能接管） |

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@huai_xin/dsh-workbuddy",
-  "version": "0.1.11",
+  "version": "0.1.13",
   "description": "WorkBuddy model provider for DeepSeek Harness (international + China deployments)",
   "type": "module",
   "main": "./src/index.js",

package/src/adapter.js

@@ -13,6 +13,7 @@
  *   X-Product-Code: workbuddy, X-Agent-Purpose/Name/Type/Version, UA.
  */
 
+import { createHash, randomUUID } from 'node:crypto'
 import { writeFileSync } from 'node:fs'
 import { tmpdir } from 'node:os'
 import { join } from 'node:path'
@@ -321,7 +322,7 @@ export class WorkBuddyAdapter {
   constructor(ctx, region) {
     this.ctx = ctx
     this.region = region
-    this.sessionId = crypto.randomUUID().replace(/-/g, '')
+    this.sessionId = randomUUID().replace(/-/g, '')
     this.catalog = undefined
     this.catalogAt = 0
     /** Where the current catalog came from: 'scoped' | 'config' | 'cache' | 'none'. */
@@ -358,7 +359,7 @@ export class WorkBuddyAdapter {
 
   /** A fresh conversation identity, used when the gateway rejects the old one. */
   rotateIdentity() {
-    this.sessionId = crypto.randomUUID().replace(/-/g, '')
+    this.sessionId = randomUUID().replace(/-/g, '')
   }
 
   /**
@@ -692,7 +693,24 @@ export class WorkBuddyAdapter {
     return chatRoutes(this.region)
   }
 
-  headers(token, route) {
+  /**
+   * Request headers, aligned with the official CodeBuddy CLI.
+   *
+   * The gateway validates them: a request without `X-Agent-Purpose` /
+   * `X-Conversation-*` / `X-IDE-*` is rejected with `400 code 11128 request illegal`
+   * (measured 2026-10-23), so these are not decoration. Three details come from the
+   * official bundle and were previously missing or wrong here:
+   *
+   * - `X-Agent-Intent` defaults to `craft` ("do the work" mode) there; omitting it
+   *   leaves the deployment to choose, which is the difference between an agent that
+   *   acts and one that only narrates.
+   * - `X-Conversation-Message-ID` / `X-Request-ID` identify *the message being
+   *   answered*, so they are derived from that message and stay stable when the same
+   *   step is retried, instead of a fresh random UUID per attempt.
+   * - `X-Agent-Runtime` / `X-Session-ID` / `X-Model-ID` say which runtime and model
+   *   the route is on.
+   */
+  headers(token, route, model, messageId) {
     return {
       Authorization: `Bearer ${token}`,
       Accept: 'text/event-stream',
@@ -701,21 +719,42 @@ export class WorkBuddyAdapter {
       'X-Product': 'SaaS',
       'X-Product-Code': route.productCode,
       'X-Agent-Purpose': 'conversation',
+      'X-Agent-Intent': 'craft',
+      'X-Agent-Runtime': 'local-agent',
       'X-IDE-Name': route.ideName,
       'X-IDE-Type': route.ideName,
       'X-IDE-Version': route.ideVersion,
       'User-Agent': route.userAgent,
+      'X-Session-ID': this.sessionId,
+      'X-Model-ID': model,
       'X-Conversation-ID': this.sessionId,
-      'X-Conversation-Message-ID': crypto.randomUUID(),
-      'X-Conversation-Request-ID': crypto.randomUUID()
+      'X-Conversation-Message-ID': messageId,
+      'X-Conversation-Request-ID': messageId,
+      'X-Request-ID': messageId
     }
   }
 
-  async send(token, body, signal, route) {
+  /**
+   * A stable id for the message being answered.
+   *
+   * The official client sends the message's own id in `X-Conversation-Message-ID` /
+   * `X-Request-ID`; a fresh random UUID per attempt (what this did before) tells the
+   * gateway every retry is a different message. Hashing the last user turn keeps the
+   * id stable across retries of one step and changes when the conversation moves on.
+   */
+  messageIdFor(options) {
+    const last = [...(options.messages ?? [])].reverse().find((message) => message.role === 'user')
+    const text = typeof last?.content === 'string'
+      ? last.content
+      : (Array.isArray(last?.content) ? last.content.map((block) => String(block?.text ?? '')).join('') : '')
+    return createHash('sha1').update(`${this.sessionId}:${text}`).digest('hex').slice(0, 32)
+  }
+
+  async send(token, body, signal, route, model, messageId) {
     try {
       return await fetch(`${route.endpoint}/v2/chat/completions`, {
         method: 'POST',
-        headers: this.headers(token, route),
+        headers: this.headers(token, route, model, messageId),
         body,
         signal
       })
@@ -734,10 +773,13 @@ export class WorkBuddyAdapter {
     }
     const imageUrls = await this.resolveImages(options.messages)
     const payload = () => JSON.stringify(this.buildBody(options, model, entry, imageUrls))
+    // The message being answered, hashed so every retry of the same step carries the
+    // same id (the official client uses the message's own id here).
+    const messageId = this.messageIdFor(options)
     const routes = this.routes()
     let route = routes[0]
 
-    let response = await this.send(token, payload(), options.signal, route)
+    let response = await this.send(token, payload(), options.signal, route, model, messageId)
     let rejection = ''
     if (!response.ok) {
       rejection = await response.text().catch(() => '')
@@ -750,12 +792,12 @@ export class WorkBuddyAdapter {
         dumpRejectedRequest(this.region, model, route, payload(), rejection)
         this.rotateIdentity()
         this.ctx.logger.warn(`[workbuddy] ${this.region.id}: 网关判为未获批渠道，已换会话标识重试一次`)
-        response = await this.send(token, payload(), options.signal, route)
+        response = await this.send(token, payload(), options.signal, route, model, messageId)
         rejection = response.ok ? '' : await response.text().catch(() => '')
         if (!response.ok && routes[1] !== undefined && isUnapprovedChannel(rejection)) {
           this.ctx.logger.warn(`[workbuddy] ${this.region.id}: 主通道仍被拒，改走 ${routes[1].endpoint}`)
           route = routes[1]
-          response = await this.send(token, payload(), options.signal, route)
+          response = await this.send(token, payload(), options.signal, route, model, messageId)
           rejection = response.ok ? '' : await response.text().catch(() => '')
         }
       } else if (response.status === 401 || response.status === 403) {
@@ -763,7 +805,7 @@ export class WorkBuddyAdapter {
         if (rotated.refreshed) {
           token = await readToken(this.ctx, this.region)
           if (token.length > 0) {
-            response = await this.send(token, payload(), options.signal, route)
+            response = await this.send(token, payload(), options.signal, route, model, messageId)
             rejection = response.ok ? '' : await response.text().catch(() => '')
           }
         }
@@ -975,6 +1017,23 @@ export class WorkBuddyAdapter {
     }
 
     const argsTruncated = [...toolCalls.values()].some((block) => isTruncatedArguments(block.text))
+    // A tool-call fragment whose name never arrived cannot be announced, so it never
+    // reaches `toolOrder` — and when upstream then ends with `stop`, the step reads as
+    // "the model chose to stop" while a call was in fact in flight. That is a silent
+    // drop, so it is reported instead: the harness keeps an honest failure, and the
+    // log names what happened.
+    const orphanCalls = [...toolCalls.values()].filter((block) => !hasUsableToolName(block.name) && block.text.trim().length > 0)
+    if (orphanCalls.length > 0) {
+      this.ctx.logger.warn(`[workbuddy] ${this.region.id}: ${options.model} 收到 ${orphanCalls.length} 个工具调用分片但始终没有工具名（finish_reason=${finishReason ?? '(无)'}），按传输错误上报，可自动重试`)
+      yield finishChunk({
+        kind: 'error',
+        failure: {
+          message: `${this.region.provider}: 上游发送了工具调用但没有给出工具名（finish_reason=${finishReason ?? '(无)'}；可重试的传输错误）`,
+          code: 'TRANSPORT'
+        }
+      })
+      return
+    }
     // A stream that stops without the `[DONE]` sentinel was cut, not finished:
     // upstream closed the connection mid-answer. Reporting that as a normal stop
     // ends the turn silently — the model looks like it simply chose to stop — and

package/src/catalog.js

@@ -224,7 +224,7 @@ export function entryFromMeta(id, meta) {
  * normally. So this is a preference order and a "officially sanctioned" marker,
  * never a filter.
  */
-export function whitelistedIds(agents) {
+function whitelistedIds(agents) {
   if (!Array.isArray(agents)) return []
   for (const agentName of ['cli', 'craft']) {
     const agent = agents.find((entry) => typeof entry === 'object' && entry !== null && entry.name === agentName)