@rei-standard/amsg-server 2.3.1 → 2.4.0-next.0

package/README.md

@@ -1,6 +1,6 @@
 # @rei-standard/amsg-server
 
-`@rei-standard/amsg-server` 是 ReiStandard 主动消息标准的服务端 SDK（v2.0.1），提供 Blob 租户配置、租户 token 鉴权和标准路由处理器。
+`@rei-standard/amsg-server` 是 ReiStandard 主动消息标准的服务端 SDK：Blob 租户配置、`tenantToken` / `cronToken` 鉴权、标准路由处理器。API 规范见 [API 技术规范](https://github.com/Tosd0/ReiStandard/blob/main/standards/active-messaging-api.md)。
 
 ## v2.0.1 变更摘要
 
@@ -9,6 +9,8 @@
 - 业务端点统一使用 `Authorization: Bearer <tenantToken>`
 - `send-notifications` 支持 `cronToken`（Header 或 query token）
 
+2.2+ 的字段增量（`messages` 数组、`splitPattern`、`avatarUrl` 严格校验）在规范的 [§6.1](https://github.com/Tosd0/ReiStandard/blob/main/standards/active-messaging-api.md#61-ai-消息字段约束) / [§6.2](https://github.com/Tosd0/ReiStandard/blob/main/standards/active-messaging-api.md#62-avatarurl-严格校验)；行为已对齐 `amsg-instant`，向后兼容。
+
 ## 安装
 
 ```bash
@@ -52,7 +54,7 @@ const rei = await createReiServer({
 
 ## 关于 `messageType: 'instant'`
 
-> **Note**：新代码的 instant 消息请用 [@rei-standard/amsg-instant](../instant/README.md)，跳过本端点的"建任务 → 处理 → 删任务" DB 来回。本端点的 `instant` 分支为兼容保留，行为不变、不会有运行时警告。
+> **Note**：新代码的 instant 消息请用 [@rei-standard/amsg-instant](https://github.com/Tosd0/ReiStandard/blob/main/packages/rei-standard-amsg/instant/README.md)，跳过本端点的"建任务 → 处理 → 删任务" DB 来回。本端点的 `instant` 分支为兼容保留，行为不变、不会有运行时警告。
 
 ## AI 接口 `apiUrl` 约束
 
@@ -130,17 +132,45 @@ AI 配置消息的提示词可以用两种形态之一，**互斥二选一**：
 ## 运行环境与要求
 
 - Node.js `>=20`
-- 必须安装 `web-push`
-- 必须安装 `@netlify/blobs`
-- 必须安装至少一个数据库驱动（`@neondatabase/serverless` 或 `pg`）
-- 必须配置：
-  - `VAPID_EMAIL`
-  - `NEXT_PUBLIC_VAPID_PUBLIC_KEY`
-  - `VAPID_PRIVATE_KEY`
-  - `TENANT_CONFIG_KEK`
-  - `TENANT_TOKEN_SIGNING_KEY`
-- 可选配置：
-  - `INIT_SECRET`（配置后 `init-tenant` 必须带 `X-Init-Secret`）
+- 必须装：`web-push`、`@netlify/blobs`、以及至少一个数据库驱动（`@neondatabase/serverless` 或 `pg`）
+
+### 环境变量
+
+必填：
+
+- `VAPID_EMAIL` — VAPID 联系邮箱
+- `NEXT_PUBLIC_VAPID_PUBLIC_KEY` / `VAPID_PRIVATE_KEY` — VAPID 公私钥对，[vapidkeys.com](https://vapidkeys.com) 生成
+- `TENANT_CONFIG_KEK` — 加密 Blob 里租户配置（含 db connection、masterKey）的 KEK
+- `TENANT_TOKEN_SIGNING_KEY` — `tenantToken` / `cronToken` 的 HMAC 签名密钥
+
+可选：
+
+- `INIT_SECRET` — 配了之后 `POST /api/v1/init-tenant` 必须带 `X-Init-Secret` 头才能初始化租户
+- `PUBLIC_BASE_URL` — 生产域名（如 `https://your-domain.com`），用来让 `init-tenant` 返回完整 `cronWebhookUrl`
+- `VERCEL_PROTECTION_BYPASS` — Vercel 部署 + Preview Protection 时给 cron 走的 bypass key
+
+`TENANT_CONFIG_KEK` / `TENANT_TOKEN_SIGNING_KEY` / `INIT_SECRET` 推荐：
+
+```bash
+openssl rand -base64 32
+```
+
+### `.env` 模板
+
+```dotenv
+VAPID_EMAIL=youremail@example.com
+NEXT_PUBLIC_VAPID_PUBLIC_KEY=YOUR-PUBLIC-KEY
+VAPID_PRIVATE_KEY=YOUR-PRIVATE-KEY
+TENANT_CONFIG_KEK=YOUR-KEK-SECRET
+TENANT_TOKEN_SIGNING_KEY=YOUR-TOKEN-SIGNING-KEY
+
+# 可选
+INIT_SECRET=YOUR-INIT-SECRET
+PUBLIC_BASE_URL=https://your-domain.com
+VERCEL_PROTECTION_BYPASS=YOUR_BYPASS_KEY
+```
+
+Vercel 部署配置可参考 [`examples/vercel.json.example`](https://github.com/Tosd0/ReiStandard/blob/main/examples/vercel.json.example)。
 
 ## 相关链接（绝对 URL）
 

package/dist/index.cjs

@@ -482,6 +482,10 @@ function isUniqueViolation(error) {
 
 // src/server/lib/message-processor.js
 
+
+
+
+var _amsgshared = require('@rei-standard/amsg-shared');
 var DEFAULT_SPLIT_REGEX = /([。！？!?]+)/;
 function splitOnceByRegex(chunk, regex) {
   const out = chunk.split(regex).reduce((acc, part, i, arr) => {
@@ -502,6 +506,22 @@ function splitMessageIntoSentences(messageContent, splitPattern = null) {
   }
   return chunks.length > 0 ? chunks : [messageContent];
 }
+function readReasoningContent(llmResponse) {
+  if (!llmResponse || typeof llmResponse !== "object") return null;
+  const choices = (
+    /** @type {{ choices?: unknown }} */
+    llmResponse.choices
+  );
+  if (!Array.isArray(choices) || choices.length === 0) return null;
+  const message = (
+    /** @type {{ message?: { reasoning_content?: unknown } }} */
+    _optionalChain([choices, 'access', _2 => _2[0], 'optionalAccess', _3 => _3.message])
+  );
+  const raw = _optionalChain([message, 'optionalAccess', _4 => _4.reasoning_content]);
+  if (typeof raw !== "string") return null;
+  const trimmed = raw.trim();
+  return trimmed.length > 0 ? trimmed : null;
+}
 async function processSingleMessage(task, ctx, providedMasterKey) {
   try {
     const masterKey = providedMasterKey || ctx.masterKey;
@@ -511,19 +531,24 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
     const userKey = deriveUserEncryptionKey(task.user_id, masterKey);
     const decryptedPayload = JSON.parse(decryptFromStorage(task.encrypted_payload, userKey));
     let messageContent;
+    let llmResponse = null;
     if (decryptedPayload.messageType === "fixed") {
       messageContent = decryptedPayload.userMessage;
     } else if (decryptedPayload.messageType === "instant") {
       const hasPrompt = !!decryptedPayload.completePrompt || Array.isArray(decryptedPayload.messages) && decryptedPayload.messages.length > 0;
       if (hasPrompt && decryptedPayload.apiUrl && decryptedPayload.apiKey && decryptedPayload.primaryModel) {
-        messageContent = await _callAI(decryptedPayload);
+        const aiResult = await _callAI(decryptedPayload);
+        messageContent = aiResult.content;
+        llmResponse = aiResult.response;
       } else if (decryptedPayload.userMessage) {
         messageContent = decryptedPayload.userMessage;
       } else {
         throw new Error("Invalid instant message: no content source available");
       }
     } else if (decryptedPayload.messageType === "prompted" || decryptedPayload.messageType === "auto") {
-      messageContent = await _callAI(decryptedPayload);
+      const aiResult = await _callAI(decryptedPayload);
+      messageContent = aiResult.content;
+      llmResponse = aiResult.response;
     } else {
       throw new Error("Invalid message configuration: no content source available");
     }
@@ -532,23 +557,48 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
       throw new Error("VAPID configuration missing - push notifications cannot be sent");
     }
     const pushSubscription = decryptedPayload.pushSubscription;
-    for (let i = 0; i < messages.length; i++) {
-      const notificationPayload = {
+    const sessionId = task.id != null ? `sess_task_${task.id}` : `sess_${_crypto.randomUUID.call(void 0, )}`;
+    const source = decryptedPayload.messageType === "instant" ? "instant" : "scheduled";
+    const messageSubtype = decryptedPayload.messageSubtype || "chat";
+    const avatarUrl = decryptedPayload.avatarUrl || null;
+    const metadata = decryptedPayload.metadata || {};
+    const messageIdBase = task.id != null ? `msg_task_${task.id}` : `msg_${_crypto.randomUUID.call(void 0, )}_instant`;
+    const reasoning = readReasoningContent(llmResponse);
+    if (reasoning) {
+      const reasoningPush = _amsgshared.buildReasoningPush.call(void 0, {
+        messageType: decryptedPayload.messageType,
+        source,
+        messageId: `${messageIdBase}_reasoning`,
+        sessionId,
+        reasoningContent: reasoning,
+        timestamp: (/* @__PURE__ */ new Date()).toISOString(),
         title: `\u6765\u81EA ${decryptedPayload.contactName}`,
+        contactName: decryptedPayload.contactName,
+        avatarUrl,
+        messageSubtype,
+        metadata
+      });
+      await ctx.webpush.sendNotification(pushSubscription, JSON.stringify(reasoningPush));
+      await new Promise((resolve) => setTimeout(resolve, 1500));
+    }
+    for (let i = 0; i < messages.length; i++) {
+      const contentPush = _amsgshared.buildContentPush.call(void 0, {
+        messageType: decryptedPayload.messageType,
+        source,
+        messageId: `${messageIdBase}_${i}`,
+        sessionId,
         message: messages[i],
+        timestamp: (/* @__PURE__ */ new Date()).toISOString(),
+        title: `\u6765\u81EA ${decryptedPayload.contactName}`,
         contactName: decryptedPayload.contactName,
-        messageId: `msg_${_crypto.randomUUID.call(void 0, )}_${task.id || "instant"}_${i}`,
+        avatarUrl,
+        messageSubtype,
         messageIndex: i + 1,
         totalMessages: messages.length,
-        messageType: decryptedPayload.messageType,
-        messageSubtype: decryptedPayload.messageSubtype || "chat",
         taskId: task.id || null,
-        timestamp: (/* @__PURE__ */ new Date()).toISOString(),
-        source: decryptedPayload.messageType === "instant" ? "instant" : "scheduled",
-        avatarUrl: decryptedPayload.avatarUrl || null,
-        metadata: decryptedPayload.metadata || {}
-      };
-      await ctx.webpush.sendNotification(pushSubscription, JSON.stringify(notificationPayload));
+        metadata
+      });
+      await ctx.webpush.sendNotification(pushSubscription, JSON.stringify(contentPush));
       if (i < messages.length - 1) {
         await new Promise((resolve) => setTimeout(resolve, 1500));
       }
@@ -643,11 +693,11 @@ async function _callAI(payload) {
     );
   }
   const aiData = await aiResponse.json();
-  const content = _optionalChain([aiData, 'optionalAccess', _2 => _2.choices, 'optionalAccess', _3 => _3[0], 'optionalAccess', _4 => _4.message, 'optionalAccess', _5 => _5.content]);
+  const content = _optionalChain([aiData, 'optionalAccess', _5 => _5.choices, 'optionalAccess', _6 => _6[0], 'optionalAccess', _7 => _7.message, 'optionalAccess', _8 => _8.content]);
   if (typeof content !== "string" || !content.trim()) {
     throw new Error("AI API error: response missing choices[0].message.content");
   }
-  return content.trim();
+  return { response: aiData, content: content.trim() };
 }
 function buildAiRequestBody(payload) {
   const llmMessages = Array.isArray(payload.messages) && payload.messages.length > 0 ? payload.messages : [{ role: "user", content: payload.completePrompt }];
@@ -1326,7 +1376,7 @@ async function resolveBlobStore(namespace) {
   }
   try {
     const blobModule = await Promise.resolve().then(() => _interopRequireWildcard(require("@netlify/blobs")));
-    const getStore = blobModule.getStore || _optionalChain([blobModule, 'access', _6 => _6.default, 'optionalAccess', _7 => _7.getStore]);
+    const getStore = blobModule.getStore || _optionalChain([blobModule, 'access', _9 => _9.default, 'optionalAccess', _10 => _10.getStore]);
     if (typeof getStore === "function") {
       return getStore({ name: namespace });
     }

package/dist/index.d.cts

@@ -1,4 +1,5 @@
 import { createDecipheriv, createHash, randomBytes, createCipheriv, randomUUID, timingSafeEqual, createHmac } from 'crypto';
+import { buildReasoningPush, buildContentPush } from '@rei-standard/amsg-shared';
 
 /**
  * Adapter Factory
@@ -793,11 +794,23 @@ function isUniqueViolation(error) {
 
 /**
  * Message Processor (SDK version)
- * ReiStandard SDK v2.0.1
+ * ReiStandard amsg-server v2.4.0
+ *
+ * Handles single message content generation and Web Push delivery for
+ * scheduled tasks (`fixed` / `prompted` / `auto`) and the legacy
+ * via-server instant path (`messageType: 'instant'`).
  *
- * Handles single message content generation and Web Push delivery.
- * Receives its dependencies (encryption helpers, webpush, VAPID config)
- * via a context object so that it stays free of process.env references.
+ * Push wire shape comes from `@rei-standard/amsg-shared`'s
+ * discriminated union (`AmsgPush`). The SW (`@rei-standard/amsg-sw`)
+ * routes on `messageKind`. Server-driven pushes always carry
+ * `source: 'instant'` (for the legacy in-server instant) or
+ * `source: 'scheduled'` (for everything else).
+ *
+ * v2.4.0: when the LLM response carries non-empty
+ * `choices[0].message.reasoning_content`, the processor now emits a
+ * standalone `ReasoningPush` **before** the `ContentPush` burst.
+ * `messagesSent` in the return value continues to reflect the sentence
+ * count only (reasoning is an auxiliary push, not a sentence).
  */
 
 
@@ -850,6 +863,25 @@ function splitMessageIntoSentences(messageContent, splitPattern = null) {
   return chunks.length > 0 ? chunks : [messageContent];
 }
 
+/**
+ * Read `choices[0].message.reasoning_content` as a non-empty trimmed
+ * string, or null when absent / empty. Mirrors
+ * `amsg-instant/src/message-processor.js#readReasoningContent`.
+ *
+ * @param {unknown} llmResponse
+ * @returns {string | null}
+ */
+function readReasoningContent(llmResponse) {
+  if (!llmResponse || typeof llmResponse !== 'object') return null;
+  const choices = /** @type {{ choices?: unknown }} */ (llmResponse).choices;
+  if (!Array.isArray(choices) || choices.length === 0) return null;
+  const message = /** @type {{ message?: { reasoning_content?: unknown } }} */ (choices[0])?.message;
+  const raw = message?.reasoning_content;
+  if (typeof raw !== 'string') return null;
+  const trimmed = raw.trim();
+  return trimmed.length > 0 ? trimmed : null;
+}
+
 /**
  * @typedef {Object} ProcessorContext
  * @property {Object}  webpush           - The web-push module instance (already VAPID-configured).
@@ -876,14 +908,19 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
     const decryptedPayload = JSON.parse(decryptFromStorage(task.encrypted_payload, userKey));
 
     let messageContent;
+    /** @type {unknown} */
+    let llmResponse = null;
 
     if (decryptedPayload.messageType === 'fixed') {
       messageContent = decryptedPayload.userMessage;
 
     } else if (decryptedPayload.messageType === 'instant') {
-      const hasPrompt = !!decryptedPayload.completePrompt || (Array.isArray(decryptedPayload.messages) && decryptedPayload.messages.length > 0);
+      const hasPrompt = !!decryptedPayload.completePrompt
+        || (Array.isArray(decryptedPayload.messages) && decryptedPayload.messages.length > 0);
       if (hasPrompt && decryptedPayload.apiUrl && decryptedPayload.apiKey && decryptedPayload.primaryModel) {
-        messageContent = await _callAI(decryptedPayload);
+        const aiResult = await _callAI(decryptedPayload);
+        messageContent = aiResult.content;
+        llmResponse = aiResult.response;
       } else if (decryptedPayload.userMessage) {
         messageContent = decryptedPayload.userMessage;
       } else {
@@ -891,7 +928,9 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
       }
 
     } else if (decryptedPayload.messageType === 'prompted' || decryptedPayload.messageType === 'auto') {
-      messageContent = await _callAI(decryptedPayload);
+      const aiResult = await _callAI(decryptedPayload);
+      messageContent = aiResult.content;
+      llmResponse = aiResult.response;
     } else {
       throw new Error('Invalid message configuration: no content source available');
     }
@@ -908,25 +947,68 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
     }
 
     const pushSubscription = decryptedPayload.pushSubscription;
+    // sessionId is shared across the optional ReasoningPush and every
+    // ContentPush from this LLM round. Pin it to the task id when
+    // available (scheduled tasks) so retries reuse the same id;
+    // otherwise mint a UUID.
+    const sessionId = task.id != null
+      ? `sess_task_${task.id}`
+      : `sess_${randomUUID()}`;
+    const source = decryptedPayload.messageType === 'instant' ? 'instant' : 'scheduled';
+    const messageSubtype = decryptedPayload.messageSubtype || 'chat';
+    const avatarUrl = decryptedPayload.avatarUrl || null;
+    const metadata = decryptedPayload.metadata || {};
+
+    // `messageId` format — deterministic when we have a task.id so a
+    // retry produces the same id for the same (task, sentence) pair
+    // (downstream dedupers can key on it). Falls back to a UUID for
+    // the legacy in-server instant path that has no row id.
+    const messageIdBase = task.id != null
+      ? `msg_task_${task.id}`
+      : `msg_${randomUUID()}_instant`;
+
+    // ReasoningPush — auto-emitted before the content burst when the
+    // LLM response carried non-empty reasoning_content. `fixed` and
+    // explicit-userMessage paths produce no LLM response, so this
+    // block is naturally skipped for them (llmResponse stays null).
+    const reasoning = readReasoningContent(llmResponse);
+    if (reasoning) {
+      const reasoningPush = buildReasoningPush({
+        messageType: decryptedPayload.messageType,
+        source,
+        messageId: `${messageIdBase}_reasoning`,
+        sessionId,
+        reasoningContent: reasoning,
+        timestamp: new Date().toISOString(),
+        title: `来自 ${decryptedPayload.contactName}`,
+        contactName: decryptedPayload.contactName,
+        avatarUrl,
+        messageSubtype,
+        metadata,
+      });
+      await ctx.webpush.sendNotification(pushSubscription, JSON.stringify(reasoningPush));
+      await new Promise(resolve => setTimeout(resolve, 1500));
+    }
 
     for (let i = 0; i < messages.length; i++) {
-      const notificationPayload = {
-        title: `来自 ${decryptedPayload.contactName}`,
+      const contentPush = buildContentPush({
+        messageType: decryptedPayload.messageType,
+        source,
+        messageId: `${messageIdBase}_${i}`,
+        sessionId,
         message: messages[i],
+        timestamp: new Date().toISOString(),
+        title: `来自 ${decryptedPayload.contactName}`,
         contactName: decryptedPayload.contactName,
-        messageId: `msg_${randomUUID()}_${task.id || 'instant'}_${i}`,
+        avatarUrl,
+        messageSubtype,
         messageIndex: i + 1,
         totalMessages: messages.length,
-        messageType: decryptedPayload.messageType,
-        messageSubtype: decryptedPayload.messageSubtype || 'chat',
         taskId: task.id || null,
-        timestamp: new Date().toISOString(),
-        source: decryptedPayload.messageType === 'instant' ? 'instant' : 'scheduled',
-        avatarUrl: decryptedPayload.avatarUrl || null,
-        metadata: decryptedPayload.metadata || {}
-      };
+        metadata,
+      });
 
-      await ctx.webpush.sendNotification(pushSubscription, JSON.stringify(notificationPayload));
+      await ctx.webpush.sendNotification(pushSubscription, JSON.stringify(contentPush));
 
       if (i < messages.length - 1) {
         await new Promise(resolve => setTimeout(resolve, 1500));
@@ -1030,7 +1112,15 @@ async function processMessagesByUuid(uuid, ctx, maxRetries = 2, userId, provided
 
 /**
  * Call an OpenAI-compatible API.
+ *
+ * Returns the full response object alongside the extracted (trimmed)
+ * `content` string. Callers that only need the text can ignore
+ * `response`; callers that want `reasoning_content` / `tool_calls`
+ * read from `response.choices[0].message`.
+ *
  * @private
+ * @param {Object} payload
+ * @returns {Promise<{ response: unknown, content: string }>}
  */
 async function _callAI(payload) {
   const normalizedApiUrl = normalizeAiApiUrl(payload.apiUrl);
@@ -1067,7 +1157,7 @@ async function _callAI(payload) {
     throw new Error('AI API error: response missing choices[0].message.content');
   }
 
-  return content.trim();
+  return { response: aiData, content: content.trim() };
 }
 
 /**
@@ -1125,20 +1215,6 @@ function buildAiRequestBody(payload) {
  * to avoid an architectural dependency (server should not depend on the
  * stateless worker package).
  *
- * Rules (idempotent — running it twice equals running it once):
- *   - Already ends with `/chat/completions`             → leave as-is.
- *   - Bare host (no path or just `/`)                   → append `/v1/chat/completions`.
- *   - Path ends with a version segment like `/v1`,
- *     `/v2`, … (with or without trailing slash)         → append only
- *     `/chat/completions`; never doubles `/v1` for
- *     callers who already include it.
- *   - Anything else (custom path that doesn't match the
- *     OpenAI shape, e.g. `/v1/messages` for
- *     Anthropic-style proxies)                          → leave as-is. The caller
- *     knows their own routing.
- *
- * Query string is preserved verbatim.
- *
  * @param {string} apiUrl
  * @returns {string}
  */

package/dist/index.d.ts

@@ -1,4 +1,5 @@
 import { createDecipheriv, createHash, randomBytes, createCipheriv, randomUUID, timingSafeEqual, createHmac } from 'crypto';
+import { buildReasoningPush, buildContentPush } from '@rei-standard/amsg-shared';
 
 /**
  * Adapter Factory
@@ -793,11 +794,23 @@ function isUniqueViolation(error) {
 
 /**
  * Message Processor (SDK version)
- * ReiStandard SDK v2.0.1
+ * ReiStandard amsg-server v2.4.0
+ *
+ * Handles single message content generation and Web Push delivery for
+ * scheduled tasks (`fixed` / `prompted` / `auto`) and the legacy
+ * via-server instant path (`messageType: 'instant'`).
  *
- * Handles single message content generation and Web Push delivery.
- * Receives its dependencies (encryption helpers, webpush, VAPID config)
- * via a context object so that it stays free of process.env references.
+ * Push wire shape comes from `@rei-standard/amsg-shared`'s
+ * discriminated union (`AmsgPush`). The SW (`@rei-standard/amsg-sw`)
+ * routes on `messageKind`. Server-driven pushes always carry
+ * `source: 'instant'` (for the legacy in-server instant) or
+ * `source: 'scheduled'` (for everything else).
+ *
+ * v2.4.0: when the LLM response carries non-empty
+ * `choices[0].message.reasoning_content`, the processor now emits a
+ * standalone `ReasoningPush` **before** the `ContentPush` burst.
+ * `messagesSent` in the return value continues to reflect the sentence
+ * count only (reasoning is an auxiliary push, not a sentence).
  */
 
 
@@ -850,6 +863,25 @@ function splitMessageIntoSentences(messageContent, splitPattern = null) {
   return chunks.length > 0 ? chunks : [messageContent];
 }
 
+/**
+ * Read `choices[0].message.reasoning_content` as a non-empty trimmed
+ * string, or null when absent / empty. Mirrors
+ * `amsg-instant/src/message-processor.js#readReasoningContent`.
+ *
+ * @param {unknown} llmResponse
+ * @returns {string | null}
+ */
+function readReasoningContent(llmResponse) {
+  if (!llmResponse || typeof llmResponse !== 'object') return null;
+  const choices = /** @type {{ choices?: unknown }} */ (llmResponse).choices;
+  if (!Array.isArray(choices) || choices.length === 0) return null;
+  const message = /** @type {{ message?: { reasoning_content?: unknown } }} */ (choices[0])?.message;
+  const raw = message?.reasoning_content;
+  if (typeof raw !== 'string') return null;
+  const trimmed = raw.trim();
+  return trimmed.length > 0 ? trimmed : null;
+}
+
 /**
  * @typedef {Object} ProcessorContext
  * @property {Object}  webpush           - The web-push module instance (already VAPID-configured).
@@ -876,14 +908,19 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
     const decryptedPayload = JSON.parse(decryptFromStorage(task.encrypted_payload, userKey));
 
     let messageContent;
+    /** @type {unknown} */
+    let llmResponse = null;
 
     if (decryptedPayload.messageType === 'fixed') {
       messageContent = decryptedPayload.userMessage;
 
     } else if (decryptedPayload.messageType === 'instant') {
-      const hasPrompt = !!decryptedPayload.completePrompt || (Array.isArray(decryptedPayload.messages) && decryptedPayload.messages.length > 0);
+      const hasPrompt = !!decryptedPayload.completePrompt
+        || (Array.isArray(decryptedPayload.messages) && decryptedPayload.messages.length > 0);
       if (hasPrompt && decryptedPayload.apiUrl && decryptedPayload.apiKey && decryptedPayload.primaryModel) {
-        messageContent = await _callAI(decryptedPayload);
+        const aiResult = await _callAI(decryptedPayload);
+        messageContent = aiResult.content;
+        llmResponse = aiResult.response;
       } else if (decryptedPayload.userMessage) {
         messageContent = decryptedPayload.userMessage;
       } else {
@@ -891,7 +928,9 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
       }
 
     } else if (decryptedPayload.messageType === 'prompted' || decryptedPayload.messageType === 'auto') {
-      messageContent = await _callAI(decryptedPayload);
+      const aiResult = await _callAI(decryptedPayload);
+      messageContent = aiResult.content;
+      llmResponse = aiResult.response;
     } else {
       throw new Error('Invalid message configuration: no content source available');
     }
@@ -908,25 +947,68 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
     }
 
     const pushSubscription = decryptedPayload.pushSubscription;
+    // sessionId is shared across the optional ReasoningPush and every
+    // ContentPush from this LLM round. Pin it to the task id when
+    // available (scheduled tasks) so retries reuse the same id;
+    // otherwise mint a UUID.
+    const sessionId = task.id != null
+      ? `sess_task_${task.id}`
+      : `sess_${randomUUID()}`;
+    const source = decryptedPayload.messageType === 'instant' ? 'instant' : 'scheduled';
+    const messageSubtype = decryptedPayload.messageSubtype || 'chat';
+    const avatarUrl = decryptedPayload.avatarUrl || null;
+    const metadata = decryptedPayload.metadata || {};
+
+    // `messageId` format — deterministic when we have a task.id so a
+    // retry produces the same id for the same (task, sentence) pair
+    // (downstream dedupers can key on it). Falls back to a UUID for
+    // the legacy in-server instant path that has no row id.
+    const messageIdBase = task.id != null
+      ? `msg_task_${task.id}`
+      : `msg_${randomUUID()}_instant`;
+
+    // ReasoningPush — auto-emitted before the content burst when the
+    // LLM response carried non-empty reasoning_content. `fixed` and
+    // explicit-userMessage paths produce no LLM response, so this
+    // block is naturally skipped for them (llmResponse stays null).
+    const reasoning = readReasoningContent(llmResponse);
+    if (reasoning) {
+      const reasoningPush = buildReasoningPush({
+        messageType: decryptedPayload.messageType,
+        source,
+        messageId: `${messageIdBase}_reasoning`,
+        sessionId,
+        reasoningContent: reasoning,
+        timestamp: new Date().toISOString(),
+        title: `来自 ${decryptedPayload.contactName}`,
+        contactName: decryptedPayload.contactName,
+        avatarUrl,
+        messageSubtype,
+        metadata,
+      });
+      await ctx.webpush.sendNotification(pushSubscription, JSON.stringify(reasoningPush));
+      await new Promise(resolve => setTimeout(resolve, 1500));
+    }
 
     for (let i = 0; i < messages.length; i++) {
-      const notificationPayload = {
-        title: `来自 ${decryptedPayload.contactName}`,
+      const contentPush = buildContentPush({
+        messageType: decryptedPayload.messageType,
+        source,
+        messageId: `${messageIdBase}_${i}`,
+        sessionId,
         message: messages[i],
+        timestamp: new Date().toISOString(),
+        title: `来自 ${decryptedPayload.contactName}`,
         contactName: decryptedPayload.contactName,
-        messageId: `msg_${randomUUID()}_${task.id || 'instant'}_${i}`,
+        avatarUrl,
+        messageSubtype,
         messageIndex: i + 1,
         totalMessages: messages.length,
-        messageType: decryptedPayload.messageType,
-        messageSubtype: decryptedPayload.messageSubtype || 'chat',
         taskId: task.id || null,
-        timestamp: new Date().toISOString(),
-        source: decryptedPayload.messageType === 'instant' ? 'instant' : 'scheduled',
-        avatarUrl: decryptedPayload.avatarUrl || null,
-        metadata: decryptedPayload.metadata || {}
-      };
+        metadata,
+      });
 
-      await ctx.webpush.sendNotification(pushSubscription, JSON.stringify(notificationPayload));
+      await ctx.webpush.sendNotification(pushSubscription, JSON.stringify(contentPush));
 
       if (i < messages.length - 1) {
         await new Promise(resolve => setTimeout(resolve, 1500));
@@ -1030,7 +1112,15 @@ async function processMessagesByUuid(uuid, ctx, maxRetries = 2, userId, provided
 
 /**
  * Call an OpenAI-compatible API.
+ *
+ * Returns the full response object alongside the extracted (trimmed)
+ * `content` string. Callers that only need the text can ignore
+ * `response`; callers that want `reasoning_content` / `tool_calls`
+ * read from `response.choices[0].message`.
+ *
  * @private
+ * @param {Object} payload
+ * @returns {Promise<{ response: unknown, content: string }>}
  */
 async function _callAI(payload) {
   const normalizedApiUrl = normalizeAiApiUrl(payload.apiUrl);
@@ -1067,7 +1157,7 @@ async function _callAI(payload) {
     throw new Error('AI API error: response missing choices[0].message.content');
   }
 
-  return content.trim();
+  return { response: aiData, content: content.trim() };
 }
 
 /**
@@ -1125,20 +1215,6 @@ function buildAiRequestBody(payload) {
  * to avoid an architectural dependency (server should not depend on the
  * stateless worker package).
  *
- * Rules (idempotent — running it twice equals running it once):
- *   - Already ends with `/chat/completions`             → leave as-is.
- *   - Bare host (no path or just `/`)                   → append `/v1/chat/completions`.
- *   - Path ends with a version segment like `/v1`,
- *     `/v2`, … (with or without trailing slash)         → append only
- *     `/chat/completions`; never doubles `/v1` for
- *     callers who already include it.
- *   - Anything else (custom path that doesn't match the
- *     OpenAI shape, e.g. `/v1/messages` for
- *     Anthropic-style proxies)                          → leave as-is. The caller
- *     knows their own routing.
- *
- * Query string is preserved verbatim.
- *
  * @param {string} apiUrl
  * @returns {string}
  */

package/dist/index.mjs

@@ -482,6 +482,10 @@ function isUniqueViolation(error) {
 
 // src/server/lib/message-processor.js
 import { randomUUID } from "crypto";
+import {
+  buildContentPush,
+  buildReasoningPush
+} from "@rei-standard/amsg-shared";
 var DEFAULT_SPLIT_REGEX = /([。！？!?]+)/;
 function splitOnceByRegex(chunk, regex) {
   const out = chunk.split(regex).reduce((acc, part, i, arr) => {
@@ -502,6 +506,22 @@ function splitMessageIntoSentences(messageContent, splitPattern = null) {
   }
   return chunks.length > 0 ? chunks : [messageContent];
 }
+function readReasoningContent(llmResponse) {
+  if (!llmResponse || typeof llmResponse !== "object") return null;
+  const choices = (
+    /** @type {{ choices?: unknown }} */
+    llmResponse.choices
+  );
+  if (!Array.isArray(choices) || choices.length === 0) return null;
+  const message = (
+    /** @type {{ message?: { reasoning_content?: unknown } }} */
+    choices[0]?.message
+  );
+  const raw = message?.reasoning_content;
+  if (typeof raw !== "string") return null;
+  const trimmed = raw.trim();
+  return trimmed.length > 0 ? trimmed : null;
+}
 async function processSingleMessage(task, ctx, providedMasterKey) {
   try {
     const masterKey = providedMasterKey || ctx.masterKey;
@@ -511,19 +531,24 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
     const userKey = deriveUserEncryptionKey(task.user_id, masterKey);
     const decryptedPayload = JSON.parse(decryptFromStorage(task.encrypted_payload, userKey));
     let messageContent;
+    let llmResponse = null;
     if (decryptedPayload.messageType === "fixed") {
       messageContent = decryptedPayload.userMessage;
     } else if (decryptedPayload.messageType === "instant") {
       const hasPrompt = !!decryptedPayload.completePrompt || Array.isArray(decryptedPayload.messages) && decryptedPayload.messages.length > 0;
       if (hasPrompt && decryptedPayload.apiUrl && decryptedPayload.apiKey && decryptedPayload.primaryModel) {
-        messageContent = await _callAI(decryptedPayload);
+        const aiResult = await _callAI(decryptedPayload);
+        messageContent = aiResult.content;
+        llmResponse = aiResult.response;
       } else if (decryptedPayload.userMessage) {
         messageContent = decryptedPayload.userMessage;
       } else {
         throw new Error("Invalid instant message: no content source available");
       }
     } else if (decryptedPayload.messageType === "prompted" || decryptedPayload.messageType === "auto") {
-      messageContent = await _callAI(decryptedPayload);
+      const aiResult = await _callAI(decryptedPayload);
+      messageContent = aiResult.content;
+      llmResponse = aiResult.response;
     } else {
       throw new Error("Invalid message configuration: no content source available");
     }
@@ -532,23 +557,48 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
       throw new Error("VAPID configuration missing - push notifications cannot be sent");
     }
     const pushSubscription = decryptedPayload.pushSubscription;
-    for (let i = 0; i < messages.length; i++) {
-      const notificationPayload = {
+    const sessionId = task.id != null ? `sess_task_${task.id}` : `sess_${randomUUID()}`;
+    const source = decryptedPayload.messageType === "instant" ? "instant" : "scheduled";
+    const messageSubtype = decryptedPayload.messageSubtype || "chat";
+    const avatarUrl = decryptedPayload.avatarUrl || null;
+    const metadata = decryptedPayload.metadata || {};
+    const messageIdBase = task.id != null ? `msg_task_${task.id}` : `msg_${randomUUID()}_instant`;
+    const reasoning = readReasoningContent(llmResponse);
+    if (reasoning) {
+      const reasoningPush = buildReasoningPush({
+        messageType: decryptedPayload.messageType,
+        source,
+        messageId: `${messageIdBase}_reasoning`,
+        sessionId,
+        reasoningContent: reasoning,
+        timestamp: (/* @__PURE__ */ new Date()).toISOString(),
         title: `\u6765\u81EA ${decryptedPayload.contactName}`,
+        contactName: decryptedPayload.contactName,
+        avatarUrl,
+        messageSubtype,
+        metadata
+      });
+      await ctx.webpush.sendNotification(pushSubscription, JSON.stringify(reasoningPush));
+      await new Promise((resolve) => setTimeout(resolve, 1500));
+    }
+    for (let i = 0; i < messages.length; i++) {
+      const contentPush = buildContentPush({
+        messageType: decryptedPayload.messageType,
+        source,
+        messageId: `${messageIdBase}_${i}`,
+        sessionId,
         message: messages[i],
+        timestamp: (/* @__PURE__ */ new Date()).toISOString(),
+        title: `\u6765\u81EA ${decryptedPayload.contactName}`,
         contactName: decryptedPayload.contactName,
-        messageId: `msg_${randomUUID()}_${task.id || "instant"}_${i}`,
+        avatarUrl,
+        messageSubtype,
         messageIndex: i + 1,
         totalMessages: messages.length,
-        messageType: decryptedPayload.messageType,
-        messageSubtype: decryptedPayload.messageSubtype || "chat",
         taskId: task.id || null,
-        timestamp: (/* @__PURE__ */ new Date()).toISOString(),
-        source: decryptedPayload.messageType === "instant" ? "instant" : "scheduled",
-        avatarUrl: decryptedPayload.avatarUrl || null,
-        metadata: decryptedPayload.metadata || {}
-      };
-      await ctx.webpush.sendNotification(pushSubscription, JSON.stringify(notificationPayload));
+        metadata
+      });
+      await ctx.webpush.sendNotification(pushSubscription, JSON.stringify(contentPush));
       if (i < messages.length - 1) {
         await new Promise((resolve) => setTimeout(resolve, 1500));
       }
@@ -647,7 +697,7 @@ async function _callAI(payload) {
   if (typeof content !== "string" || !content.trim()) {
     throw new Error("AI API error: response missing choices[0].message.content");
   }
-  return content.trim();
+  return { response: aiData, content: content.trim() };
 }
 function buildAiRequestBody(payload) {
   const llmMessages = Array.isArray(payload.messages) && payload.messages.length > 0 ? payload.messages : [{ role: "user", content: payload.completePrompt }];

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@rei-standard/amsg-server",
-  "version": "2.3.1",
-  "description": "ReiStandard Active Messaging server SDK with pluggable database adapters",
+  "version": "2.4.0-next.0",
+  "description": "ReiStandard Active Messaging server SDK with pluggable database adapters. Three-axis push schema (messageKind / messageType / messageSubtype) from @rei-standard/amsg-shared. Auto-emits ReasoningPush when the LLM response carries reasoning_content.",
   "repository": {
     "type": "git",
     "url": "https://github.com/Tosd0/ReiStandard",
@@ -33,6 +33,7 @@
     "node": ">=20"
   },
   "dependencies": {
+    "@rei-standard/amsg-shared": "0.1.0-next.0",
     "web-push": "^3.6.7",
     "@netlify/blobs": "^8.1.0"
   },