@huai_xin/dsh-workbuddy 0.1.8 → 0.1.10

package/docs/workbuddy-protocol.md

@@ -175,8 +175,8 @@ POST <chatEndpoint>/v2/chat/completions
 | `max_tokens` | 否 | 实测 `128000` 可用 ✅；未知/别名模型按目录 `maxOutputTokens` |
 | `tools` | 否 | OpenAI function 形态，70 个工具 / 60 KB 实测通过 ✅ |
 | `prompt_cache_key` | 否 | 服务端前缀缓存键；缺失时缓存命中恒为 0 📖 |
-| `thinking` | 否 | `{"type":"enabled"}`，deepseek 系使用 📖✅ |
-| `reasoning_effort` | 否 | `low` / `medium` / `high` / `xhigh` / `max`，取值须在该模型支持列表内 ✅ 网关会校验（见下） |
+| `thinking` | 否 | **本插件不再发送**（2026-10-23 实测：带不带它都不影响 ds 是否思考，见 §12.1） |
+| `reasoning_effort` | 否 | `low` / `medium` / `high` / `xhigh` / `max`，取值须在该模型支持列表内 ✅ 网关会校验（见下）；**这是唯一的思考控制**，按目录声明统一取值（调用方选择 → `defaultEffort` → `high` → 首档） |
 | `temperature` / `stop` | 否 | 直通 ✅ |
 
 **`reasoning_effort` 是真实生效的参数**（2026-10-22 实测，海外版 `deepseek-v4.1-flash`，免费模型，
@@ -860,6 +860,39 @@ provider 与全部 HTTP 路由都没注册，设置面板只剩一个 `HTTP 404`
 | 流空闲监控 | `setInterval` + `streamIdleTimeoutMs`（日志 `Stream idle monitor armed`） | **没有**（依赖 harness 自身超时） | 无 |
 | 请求体改写 | 模型不支持工具时 `stripToolDefinitions`（连 `tool_choice` 一起去掉）、图片扁平化、自定义模型 URL 规整 | CN：系统提示词指纹改写（`Illegal API invocation`）+ 换会话标识重试一次 | — |
 
+#### 12.1 请求体逐字段对照（2026-10-23 补）
+
+官方 `providerData` 里还有 `thinking:{type:'disabled'}`（关闭思考时才发；开启时不发该字段）、
+`reasoning_effort`（档位）。逐字段差异与影响判断：
+
+| 字段 | 本插件 | 官方 CLI | 影响 |
+| --- | --- | --- | --- |
+| `model` / `messages` / `stream` / `tools` / `max_tokens` | 发 | 发 | 一致 ✓ |
+| `stream_options.include_usage` | 发（0.1.8 起） | 发 | 已对齐 ✓ |
+| `tool_choice` / `parallel_tool_calls` | 不发 | 发（配置值） | 服务端默认 `auto` / `true`，等价 ✓ |
+| `temperature` / `top_p` / `frequency_penalty` / `presence_penalty` | **不发**（仅调用方给 temperature 时发） | **总是发**（模型配置值，缺省 `DEFAULT_TEMPERATURE`） | **真实差异**，但实测对"提前收尾"无影响（默认/0.2/0.5/0.7/0.3+top_p0.9 各 8 次：2/8、2/8、3/8、3/8，同一水平）→ 暂不跟随，避免擅自改变输出风格 |
+| `reasoning_effort` | 发（目录档位，**所有模型同一套规则**：调用方选择 → `defaultEffort` → `high` → 首档；无档位不发） | 发（同左） | 一致 ✓ |
+| `thinking` | **不发**（2026-10-23 起） | **开启时不发**；关闭时发 `{type:'disabled'}` | 已对齐 ✓。实测（两个部署各 2–3 次）：带不带 `{type:'enabled'}`，`high` 档都是 0 思考字符；`max` 档不带时思考更多（INTL 384 vs 306、CN 476 vs 418）→ 该字段对"思不思考"没有作用 |
+| `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 ✓ |
+
+**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。
+
+#### 12.2 `finish_reason: "refusal"` 是什么
+
+官方把 `refusal` 与 `content_filter` 归为同一类 `stopReason:"refusal"`，它的来源（读官方 bundle）：
+
+1. **输入侧内容安全拦截**：`prepareSensitiveUserPrompt` 判定 `deny` 时直接返回
+   `stopReason:"refusal"`（日志 `Sensitive user prompt blocked before history`）——**请求根本没发出去**；
+2. **模型/厂商主动拒绝**：`isContentPolicyRefusal(...)` 判定的内容策略拒绝；
+3. 本地命令执行失败等"请求没能完成"的收尾，也复用 `refusal`。
+
+所以 `refusal` ≠ 网络问题、≠ 截断，而是**"这次不回答"**。正确反应是**不重试**（重试会得到同样的拒绝），
+把原因告诉用户 —— 本插件现在正是这样（`INVALID_REQUEST` + 明确文案，见 §4.5）。
+
 三条结论：
 
 1. **我们和官方在"结束原因"上原本差一张表** —— 官方把 `refusal`/`content_filter` 单独归类、

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@huai_xin/dsh-workbuddy",
-  "version": "0.1.8",
+  "version": "0.1.10",
   "description": "WorkBuddy model provider for DeepSeek Harness (international + China deployments)",
   "type": "module",
   "main": "./src/index.js",

package/src/adapter.js

@@ -660,23 +660,30 @@ export class WorkBuddyAdapter {
     const maxTokens = positiveInt(options.maxTokens) ?? entry?.maxTokens
     if (maxTokens !== undefined) body.max_tokens = maxTokens
 
-    // Thinking is pinned on: there is no switch anywhere in the plugin. Upstream
-    // publishes `reasoning.canDisableThinking` without naming a request parameter
-    // that would act on it (protocol §4.3.1), so the effort level is the only
-    // reasoning control, and the explicit `thinking` field is sent where the
-    // official client is known to send it (deepseek family, §4.5) — the other
-    // families ignore it, whatever the value.
-    const isDeepSeek = model.startsWith('deepseek-')
-    if (isDeepSeek) body.thinking = { type: 'enabled' }
-    // `efforts` is optional: many catalog entries declare no level set, and
-    // reading `Object.keys` off that `undefined` used to throw before the
-    // request was even built.
+    // Thinking is pinned on for every model, with no per-family branch: the effort
+    // level is the only reasoning control, and it is picked from what the catalog
+    // declares rather than from the model's name —
+    //   1. the level the caller chose, when the model declares it;
+    //   2. otherwise the model's own `defaultEffort` from the catalog;
+    //   3. otherwise `high` when declared, else the first declared level;
+    //   4. no levels declared → send nothing and let the deployment decide.
+    //
+    // The `thinking` field is deliberately **not** sent for anyone. Measured
+    // 2026-10-23 on both deployments (`test/live/thinking-field-probe.mjs`): with
+    // `reasoning_effort` present, dropping `thinking:{type:'enabled'}` changes
+    // nothing at `high` (0 reasoning chars either way) and yields slightly more
+    // reasoning at `max` (384 vs 306 on INTL, 476 vs 418 on CN). The official client
+    // only sends that field when it turns thinking *off*, so the enabled form was
+    // never part of its contract.
+    //
+    // `efforts` is optional: many catalog entries declare no level set, and reading
+    // `Object.keys` off that `undefined` used to throw before the request was built.
     const levels = entry?.efforts === undefined ? [] : Object.keys(entry.efforts)
-    if (options.reasoningEffort !== undefined && levels.includes(options.reasoningEffort)) {
-      body.reasoning_effort = options.reasoningEffort
-    } else if (isDeepSeek && levels.length > 0) {
-      body.reasoning_effort = levels.includes('high') ? 'high' : levels[0]
-    }
+    const preferred = levels.includes(options.reasoningEffort)
+      ? options.reasoningEffort
+      : (levels.includes(entry?.defaultEffort) ? entry.defaultEffort : undefined)
+    const effort = preferred ?? (levels.includes('high') ? 'high' : levels[0])
+    if (effort !== undefined) body.reasoning_effort = effort
     return body
   }
 
@@ -968,6 +975,20 @@ export class WorkBuddyAdapter {
     }
 
     const argsTruncated = [...toolCalls.values()].some((block) => isTruncatedArguments(block.text))
+    // A step that ends with prose and no tool call *looks* like the model choosing to
+    // stop, so nothing else records it — and "it stopped mid-answer" arrives as a bug
+    // report with no evidence. These two shapes are the ones seen in practice:
+    // an announcement that expects a next action ("……：" / "………") and a statement of
+    // inability ("我需要先读取…" / "我无法…"). Logging them with the raw reason and the
+    // text turns the next report into a fact.
+    if (streamEnded === true && toolOrder.length === 0 && (finishReason === 'stop' || finishReason === undefined)) {
+      const prose = blocks.find((block) => block.kind === 'text')?.text ?? ''
+      const tail = prose.trim()
+      const suspicious = /[:：…]$/.test(tail) || /(我需要|我没有|我无法|先读取|无法直接|不能直接)/.test(tail)
+      if (suspicious) {
+        this.ctx.logger.warn(`[workbuddy] ${this.region.id}: ${options.model} 疑似提前收尾（finish_reason=${finishReason ?? '(无)'}，输出 ${outputTokens ?? '未知'} tokens，无工具调用）：「${tail.slice(0, 60)}」`)
+      }
+    }
     // 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