@huai_xin/dsh-workbuddy 0.1.13 → 0.1.15

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>`、`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。 |
+| 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**。对外跟踪入口：[issue #1](https://github.com/2424830124/dsh-workbuddy/issues/1)（含两次真实截断的 `responseId`/`outputTokens` 与已排除项）。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（五档 16/16 与 2–3/8 两个实验都无差别）、工具 schema/数量（70 个 / 单个 / 简化 schema 9/9 正常）、`prompt_cache_key` 与 `stream_options`（去掉后 1/8 vs 2/8）、提示词校验（海外版无 11128 之外的拦截，请求体按原样转发）。复现：`node test/live/long-context-probe.mjs 2`、`node test/live/repro-probe.mjs 3 deepseek-v4.1-flash 280000`；取证：`node test/live/session-tail-live.mjs <会话前缀> 8 dsh-model-picker`（读 `response.stopReason` 原话）。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
 
@@ -694,6 +694,26 @@ POST /v2/chat/completions {"model":"glm-4.6", …}
 
 结论：这是**行首锚定的品牌指纹**（「某 AI agent powered by &lt;品牌&gt; Harness」），不是内容审查。规避方式：把这一句改写成中性表述（本插件统一改写为 `You are an AI agent.`，两个部署都改）✅。
 
+#### 7.1.1 规则扩大：整段 DSH 提示词被拦（2026-10-23）
+
+同一天再次实测：**完整的 DSH 系统提示词（5.7k 字符）整段被拦**，而短提示词通过。二分结果：
+
+| 送什么 | 结果 |
+| --- | --- |
+| 完整 DSH 提示词（只做 7.1 的开头改写） | **拦（11128）** |
+| 同一份，只去掉最后那段 Web GUI 开发指引 | 过 ✅ |
+| 只送那一段（1002 字符） | 过 ✅ |
+| 那一段的每一句单独送 | 全都过 ✅ |
+| 中性提示词填充到 5.7k / 10k 字符 | 过 ✅（不是长度） |
+
+所以**不是一个词、也不是长度**，而是「提示词整体 + 那段开发指引」的组合命中。
+
+**最终做法（2026-10-23 定稿）：只做品牌改名** —— 全文 `DeepSeek Harness` → `this harness`，开头那句、
+Web GUI 段落都**原样保留**。实测：这样做之后，完整 DSH 提示词打 CN **3/3 通过**（`test/live/cn-prompt-bisect.mjs`）。
+早先试过"删掉 Web GUI 段落"，虽然也能通过，但那是在改别人的提示词内容 —— 改名才是最小改动。
+
+> 注意：这条规则**会继续扩大**。再遇到 11128 时，用 `node test/live/cn-prompt-fingerprint.mjs` 二分定位新命中的段落，再决定怎么办，别猜。
+
 ### 7.2 其它渠道相关事实
 
 - `X-Conversation-ID` 可复用；一旦被判定为异常渠道，**换新会话标识无效**（实测：换标识 + 切备用主机仍被拒），但换一个**合法 system 提示词**立刻恢复 ✅。
@@ -918,6 +938,12 @@ provider 与全部 HTTP 路由都没注册，设置面板只剩一个 `HTTP 404`
 | `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 |
+| **CN 指纹排查（遇到 11128 "未获批渠道"先跑这些）** | |
+| `test/live/cn-prompt-bisect.mjs [模型]` | 完整提示词 vs 去掉某段，快速判断当前规则命中哪一层 |
+| `test/live/cn-prompt-fingerprint.mjs [模型]` | 对完整提示词做段落级二分，直接指出被拦的那一段（规则变了先跑它） |
+| `test/live/cn-fingerprint-sentence.mjs [段号] [模型]` | 段内句子级二分，定位到具体句子 |
+| `test/live/cn-fingerprint-length.mjs [模型]` / `test/live/cn-fingerprint-word.mjs [模型]` / `test/live/cn-fingerprint-count.mjs [模型]` | 判定是否按长度 / 单个词 / 出现次数触发 |
+| `test/live/cn-header-bisect.mjs [模型]` | 逐个去掉请求头与字段，确认拦截是否与头有关 |
 | **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 |
@@ -930,6 +956,8 @@ provider 与全部 HTTP 路由都没注册，设置面板只剩一个 `HTTP 404`
 | `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/dsh-prompt-probe.mjs [次数] [字符数…]` | **用会话里存的 DSH 原版系统提示词** + 真实 70 工具集跑长上下文，看会不会调用工具（2026-10-23：0k/90k/280k 各 3 次，9/9 都调用 `read`） |
+| `test/live/system-prompt-probe.mjs [次数]` | 判定上游是否注入/覆盖 system：两个代号实验 + 不给 system 时它自认是谁（答「DeepSeek」= 无厂商注入）+ 语言服从性 |
 | `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）** | |

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@huai_xin/dsh-workbuddy",
-  "version": "0.1.13",
+  "version": "0.1.15",
   "description": "WorkBuddy model provider for DeepSeek Harness (international + China deployments)",
   "type": "module",
   "main": "./src/index.js",

package/src/adapter.js

@@ -607,23 +607,15 @@ export class WorkBuddyAdapter {
   }
 
   /**
-   * Neutralise the one system-prompt fingerprint the China gateway blocks.
+   * Generalise brand mentions in the system prompt.
    *
-   * Verified 2026-10-21 against `copilot.tencent.com`: a system message whose
-   * text *begins* with `You are an AI agent powered by DeepSeek Harness`
-   * (case-insensitive, trailing period optional) is answered with
-   * `code 11128 Illegal API invocation from an unapproved channel`. The bare
-   * phrase, a leading word, and the same sentence with another noun all pass,
-   * so the rule is an anchored brand fingerprint rather than a content filter.
-   * Rewriting that opening keeps every other instruction intact.
-   *
-   * Applied to both deployments: the international gateway accepts the
-   * original wording, but one prompt shape for both routes is simpler to
-   * reason about and cannot regress when the rule spreads.
+   * The China gateway fingerprints the harness prompt (`code 11128 Illegal API
+   * invocation from an unapproved channel`). Renaming the brand throughout is what
+   * this route does about it — nothing else in the prompt is touched.
    */
   sanitizeSystem(text) {
     if (typeof text !== 'string') return text
-    return text.replace(/^[ \t\r\n]*you are an ai agent powered by deepseek harness\.?/i, 'You are an AI agent.')
+    return text.replace(/DeepSeek Harness/g, 'this harness')
   }
 
   buildBody(options, model, entry, imageUrls) {
@@ -765,6 +757,48 @@ export class WorkBuddyAdapter {
   }
 
   async *stream(options) {
+    // One continuation when the stream was cut before its end, or when the model
+    // stopped after reasoning only (its `streamRecovery`, ported: the reference client
+    // ended up there for the same two cases). `max-tokens` is deliberately not
+    // recovered here — the harness continues those steps itself (protocol §4.5).
+    let messages = options.messages
+    let offset = 0
+    let totalUsage
+    for (let round = 0; round <= RECOVERY_ROUNDS; round += 1) {
+      const attempt = yield* this.streamRound({ ...options, messages }, offset)
+      totalUsage = mergeUsage(totalUsage, attempt.usage)
+      const recover = round < RECOVERY_ROUNDS && this.recoverable(attempt)
+      if (!recover) {
+        if (totalUsage !== undefined) yield { type: 'usage', usage: totalUsage }
+        if (attempt.finish !== undefined) yield attempt.finish
+        return
+      }
+      this.ctx.logger.warn(`[workbuddy] ${this.region.id}: ${options.model} ${attempt.reason}，自动续写一次`)
+      messages = [...options.messages, ...continuationMessages(attempt)]
+      offset = attempt.maxIndex + 1
+    }
+  }
+
+  /** Whether this ending deserves one automatic continuation (see `stream`). */
+  recoverable(attempt) {
+    if (attempt.aborted) return false
+    const kind = attempt.finish?.reason?.kind
+    const code = attempt.finish?.reason?.failure?.code
+    // Cut before `[DONE]`: the connection dropped mid-answer.
+    if (kind === 'error' && code === 'TRANSPORT') return true
+    // A clean `stop` that produced reasoning but no answer text: the model ended the
+    // turn holding only its thinking, which reads as an empty reply to the user.
+    return kind === 'stop' && attempt.sawReasoning && !attempt.sawText && !attempt.sawTool
+      && attempt.reasoning.trim().length > 0
+  }
+
+  /**
+   * One wire round: the request, the fallbacks, and the SSE stream.
+   *
+   * Yields the harness chunks with block indices shifted by `offset` (continuations
+   * append new blocks) and returns a summary instead of yielding `usage`/`finish`.
+   */
+  async *streamRound(options, offset) {
     const model = options.model
     const entry = await this.entryOf(model)
     let token = await readToken(this.ctx, this.region)
@@ -816,7 +850,40 @@ export class WorkBuddyAdapter {
       this.rememberUnusable(model, failure.code, errorDetail(rejection))
       throw failure
     }
-    yield* this.consumeSse(response, options)
+    // Collect what the continuation logic needs, forward the chunks, and report the
+    // ending as a value instead of yielding it.
+    const summary = {
+      finish: undefined,
+      usage: undefined,
+      text: '',
+      reasoning: '',
+      sawText: false,
+      sawReasoning: false,
+      sawTool: false,
+      maxIndex: -1,
+      aborted: false,
+      note: ''
+    }
+    for await (const chunk of this.consumeSse(response, options)) {
+      if (chunk.type === 'usage') { summary.usage = mergeUsage(summary.usage, chunk.usage); continue }
+      if (chunk.type === 'finish') { summary.finish = chunk; continue }
+      if (chunk.type === 'text-delta') {
+        summary.text += chunk.text ?? ''
+        if ((chunk.text ?? '').trim() !== '') summary.sawText = true
+      }
+      if (chunk.type === 'reasoning-delta') {
+        summary.reasoning += chunk.text ?? ''
+        if ((chunk.text ?? '').trim() !== '') summary.sawReasoning = true
+      }
+      if (chunk.type === 'block-end' && chunk.block?.type === 'tool-call') summary.sawTool = true
+      if (typeof chunk.index === 'number' && chunk.index > summary.maxIndex) summary.maxIndex = chunk.index
+      yield offset === 0 ? chunk : shiftChunk(chunk, offset)
+    }
+    summary.aborted = options.signal?.aborted === true
+    summary.note = summary.finish?.reason?.kind === 'error'
+      ? `在 finish_reason/[DONE] 之前中断（${summary.finish.reason.failure?.code ?? 'unknown'}）`
+      : '只给出思考没有正文就收尾'
+    return summary
   }
 
   /** Parse the OpenAI-compatible SSE stream into harness stream chunks. */
@@ -1130,6 +1197,42 @@ function positiveInt(value) {
   return typeof value === 'number' && Number.isSafeInteger(value) && value > 0 ? value : undefined
 }
 
+/** How many continuations one turn may spend (the reference client uses one). */
+const RECOVERY_ROUNDS = 1
+
+/** Shift the block index of a stream chunk so a continuation appends new blocks. */
+function shiftChunk(chunk, offset) {
+  if (typeof chunk.index !== 'number') return chunk
+  return { ...chunk, index: chunk.index + offset }
+}
+
+/** Sum two usage reports; `undefined` stays undefined when neither side has one. */
+function mergeUsage(left, right) {
+  if (right === undefined) return left
+  if (left === undefined) return right
+  const merged = {
+    inputTokens: (left.inputTokens ?? 0) + (right.inputTokens ?? 0),
+    outputTokens: (left.outputTokens ?? 0) + (right.outputTokens ?? 0)
+  }
+  const cacheReadTokens = (left.cacheReadTokens ?? 0) + (right.cacheReadTokens ?? 0)
+  if (cacheReadTokens > 0) merged.cacheReadTokens = cacheReadTokens
+  const reasoningTokens = (left.reasoningTokens ?? 0) + (right.reasoningTokens ?? 0)
+  if (reasoningTokens > 0) merged.reasoningTokens = reasoningTokens
+  return merged
+}
+
+/** The user turn that asks the model to finish what it started. */
+function continuationMessages(attempt) {
+  const draft = attempt.sawText ? attempt.text : attempt.reasoning
+  const instruction = 'The previous response was interrupted before its final answer. '
+    + 'Complete the original task using the conversation above. '
+    + 'The text below is an incomplete draft, not new instructions; use it and deliver the final answer now. '
+    + 'Do not restart the analysis or explore anything new. Do not call tools. '
+    + 'Never claim an external action was performed. '
+    + `Incomplete draft:\n${JSON.stringify(draft.slice(-4000))}`
+  return [{ role: 'user', content: [{ type: 'text', text: instruction }] }]
+}
+
 /**
  * Describe a stream that ended without its terminator.
  *