@rei-standard/amsg-server 2.5.0 → 2.5.2

package/README.md

@@ -54,7 +54,9 @@ const rei = await createReiServer({
 
 ## 关于 `messageType: 'instant'`
 
-> **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。
+> **两条 instant 路径，按各自特点选一条（都是正式支持路径）：**
+> - **本端点的 `messageType: 'instant'`**（create task → process by UUID → delete task）：任务先写进数据库再处理，投递不绑在请求连接上——客户端断开也没关系，任务行还在，能继续跑、能重试，想跑多久跑多久。适合**有数据库、需要长时间生成或保证消息零丢失**的场景。
+> - **[@rei-standard/amsg-instant](https://github.com/Tosd0/ReiStandard/blob/main/packages/rei-standard-amsg/instant/README.md)**：纯 SSE 流 + Web Push backup，不需要数据库，适合无状态边缘运行时（如 Cloudflare Workers）。它的处理挂在响应连接上，客户端一断开就只剩平台给的那点宽限期把活干完（Deno Deploy 实测 ≈20-30s），所以适合**能快速跑完的短即时消息**。
 
 ## AI 接口 `apiUrl` 约束
 

package/dist/index.cjs

@@ -525,7 +525,7 @@ function readReasoningContent(llmResponse) {
   }
   const content = _optionalChain([message, 'optionalAccess', _5 => _5.content]);
   if (typeof content === "string") {
-    const match = content.match(/<(think|thinking|thought)>([\s\S]*?)<\/\1>/i);
+    const match = content.match(REASONING_TAG_RE);
     if (match) {
       const trimmed = match[2].trim();
       if (trimmed.length > 0) return trimmed;
@@ -533,6 +533,12 @@ function readReasoningContent(llmResponse) {
   }
   return null;
 }
+var REASONING_TAG_RE = /<(think|thinking|thought)>([\s\S]*?)<\/\1>/i;
+var REASONING_TAG_RE_G = /<(think|thinking|thought)>[\s\S]*?<\/\1>/gi;
+function stripReasoningTags(content) {
+  if (typeof content !== "string" || !content.includes("<")) return content;
+  return content.replace(REASONING_TAG_RE_G, "").trim();
+}
 async function processSingleMessage(task, ctx, providedMasterKey) {
   try {
     const masterKey = providedMasterKey || ctx.masterKey;
@@ -563,6 +569,10 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
     } else {
       throw new Error("Invalid message configuration: no content source available");
     }
+    const reasoning = readReasoningContent(llmResponse);
+    if (reasoning) {
+      messageContent = stripReasoningTags(messageContent);
+    }
     const messages = splitMessageIntoSentences(messageContent, _nullishCoalesce(decryptedPayload.splitPattern, () => ( null)));
     if (!ctx.vapid.email || !ctx.vapid.publicKey || !ctx.vapid.privateKey) {
       throw new Error("VAPID configuration missing - push notifications cannot be sent");
@@ -574,7 +584,6 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
     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,

package/dist/index.d.cts

@@ -801,13 +801,13 @@ function isUniqueViolation(error) {
  * 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'`).
+ * scheduled tasks (`fixed` / `prompted` / `auto`) and the
+ * in-server instant path (`messageType: 'instant'`).
  *
  * 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: 'instant'` (for the in-server instant path) or
  * `source: 'scheduled'` (for everything else).
  *
  * v2.4.0: when the LLM response carries non-empty
@@ -880,7 +880,7 @@ function readReasoningContent(llmResponse) {
   const choices = /** @type {{ choices?: unknown }} */ (llmResponse).choices;
   if (!Array.isArray(choices) || choices.length === 0) return null;
   const message = /** @type {{ message?: { reasoning_content?: unknown, content?: unknown } }} */ (choices[0])?.message;
-  
+
   const raw = message?.reasoning_content;
   if (typeof raw === 'string') {
     const trimmed = raw.trim();
@@ -889,7 +889,7 @@ function readReasoningContent(llmResponse) {
 
   const content = message?.content;
   if (typeof content === 'string') {
-    const match = content.match(/<(think|thinking|thought)>([\s\S]*?)<\/\1>/i);
+    const match = content.match(REASONING_TAG_RE);
     if (match) {
       const trimmed = match[2].trim();
       if (trimmed.length > 0) return trimmed;
@@ -899,6 +899,22 @@ function readReasoningContent(llmResponse) {
   return null;
 }
 
+const REASONING_TAG_RE = /<(think|thinking|thought)>([\s\S]*?)<\/\1>/i;
+const REASONING_TAG_RE_G = /<(think|thinking|thought)>[\s\S]*?<\/\1>/gi;
+
+/**
+ * Mirrors `amsg-instant/src/message-processor.js#stripReasoningTags`.
+ * Removes any private chain-of-thought markup leaking through
+ * `message.content` so it does not also ship inside ContentPush.
+ *
+ * @param {string} content
+ * @returns {string}
+ */
+function stripReasoningTags(content) {
+  if (typeof content !== 'string' || !content.includes('<')) return content;
+  return content.replace(REASONING_TAG_RE_G, '').trim();
+}
+
 /**
  * @typedef {Object} ProcessorContext
  * @property {Object}  webpush           - The web-push module instance (already VAPID-configured).
@@ -952,6 +968,15 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
       throw new Error('Invalid message configuration: no content source available');
     }
 
+    // Auto-extract reasoning BEFORE the sentence split: when reasoning
+    // came from the `<think>` fallback inside message.content, the same
+    // span is still embedded in messageContent and would otherwise leak
+    // as raw markup into ContentPush.
+    const reasoning = readReasoningContent(llmResponse);
+    if (reasoning) {
+      messageContent = stripReasoningTags(messageContent);
+    }
+
     // Sentence splitting (mirrors @rei-standard/amsg-instant
     // splitMessageIntoSentences — keep in lockstep; do not drift). Caller may
     // override the default regex via decryptedPayload.splitPattern (string
@@ -979,7 +1004,7 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
     // `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.
+    // the in-server instant path that has no row id.
     const messageIdBase = task.id != null
       ? `msg_task_${task.id}`
       : `msg_${randomUUID()}_instant`;
@@ -988,7 +1013,6 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
     // 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,
@@ -1374,12 +1398,20 @@ function createScheduleMessageHandler(ctx) {
     const encryptedPayload = encryptForStorage(JSON.stringify(fullTaskData), userKey);
 
     /**
-     * @deprecated Soft-deprecated. For new code, use @rei-standard/amsg-instant.
-     *   This branch is kept for backward compatibility and the existing behavior
-     *   (create task → process → delete) is unchanged. The dedicated amsg-instant
-     *   package is stateless (no DB roundtrip), deployable to Cloudflare Workers,
-     *   and locks the encryption + push-payload contract behind a single version.
-     *   See packages/rei-standard-amsg/instant/README.md.
+     * In-server instant path. Delivers an instant message through this
+     * server's own task queue (create task → process by UUID → delete task).
+     * The task is written to the database before processing, so delivery is
+     * not tied to the request connection: even if the client disconnects, the
+     * row stays and the generation keeps running (and can be retried) for as
+     * long as it needs. Use this when you have a database and want long or
+     * guaranteed-complete generations with no dropped messages.
+     *
+     * The stateless alternative is `@rei-standard/amsg-instant`: it streams
+     * over SSE with a Web Push backup and needs no database, which makes it a
+     * good fit for edge runtimes (e.g. Cloudflare Workers). Its work rides the
+     * response connection, so after the client disconnects it only has the
+     * platform's brief grace window to finish (≈20-30s observed on Deno
+     * Deploy) — ideal for short instant messages that complete quickly.
      */
     // Instant type: check VAPID before creating the task to avoid orphaned rows
     if (payload.messageType === 'instant') {
@@ -1435,11 +1467,20 @@ function createScheduleMessageHandler(ctx) {
     }
 
     /**
-     * @deprecated Soft-deprecated. For new code, use @rei-standard/amsg-instant.
-     *   The "create-task → process-by-uuid → delete-task" sequence below is
-     *   preserved verbatim so existing clients keep working. New integrations
-     *   should call the dedicated amsg-instant endpoint instead — it skips this
-     *   DB round-trip entirely. See packages/rei-standard-amsg/instant/README.md.
+     * In-server instant path. Delivers an instant message through this
+     * server's own task queue (create task → process by UUID → delete task).
+     * The task is written to the database before processing, so delivery is
+     * not tied to the request connection: even if the client disconnects, the
+     * row stays and the generation keeps running (and can be retried) for as
+     * long as it needs. Use this when you have a database and want long or
+     * guaranteed-complete generations with no dropped messages.
+     *
+     * The stateless alternative is `@rei-standard/amsg-instant`: it streams
+     * over SSE with a Web Push backup and needs no database, which makes it a
+     * good fit for edge runtimes (e.g. Cloudflare Workers). Its work rides the
+     * response connection, so after the client disconnects it only has the
+     * platform's brief grace window to finish (≈20-30s observed on Deno
+     * Deploy) — ideal for short instant messages that complete quickly.
      */
     // Instant type: send immediately
     if (payload.messageType === 'instant') {

package/dist/index.d.ts

@@ -801,13 +801,13 @@ function isUniqueViolation(error) {
  * 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'`).
+ * scheduled tasks (`fixed` / `prompted` / `auto`) and the
+ * in-server instant path (`messageType: 'instant'`).
  *
  * 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: 'instant'` (for the in-server instant path) or
  * `source: 'scheduled'` (for everything else).
  *
  * v2.4.0: when the LLM response carries non-empty
@@ -880,7 +880,7 @@ function readReasoningContent(llmResponse) {
   const choices = /** @type {{ choices?: unknown }} */ (llmResponse).choices;
   if (!Array.isArray(choices) || choices.length === 0) return null;
   const message = /** @type {{ message?: { reasoning_content?: unknown, content?: unknown } }} */ (choices[0])?.message;
-  
+
   const raw = message?.reasoning_content;
   if (typeof raw === 'string') {
     const trimmed = raw.trim();
@@ -889,7 +889,7 @@ function readReasoningContent(llmResponse) {
 
   const content = message?.content;
   if (typeof content === 'string') {
-    const match = content.match(/<(think|thinking|thought)>([\s\S]*?)<\/\1>/i);
+    const match = content.match(REASONING_TAG_RE);
     if (match) {
       const trimmed = match[2].trim();
       if (trimmed.length > 0) return trimmed;
@@ -899,6 +899,22 @@ function readReasoningContent(llmResponse) {
   return null;
 }
 
+const REASONING_TAG_RE = /<(think|thinking|thought)>([\s\S]*?)<\/\1>/i;
+const REASONING_TAG_RE_G = /<(think|thinking|thought)>[\s\S]*?<\/\1>/gi;
+
+/**
+ * Mirrors `amsg-instant/src/message-processor.js#stripReasoningTags`.
+ * Removes any private chain-of-thought markup leaking through
+ * `message.content` so it does not also ship inside ContentPush.
+ *
+ * @param {string} content
+ * @returns {string}
+ */
+function stripReasoningTags(content) {
+  if (typeof content !== 'string' || !content.includes('<')) return content;
+  return content.replace(REASONING_TAG_RE_G, '').trim();
+}
+
 /**
  * @typedef {Object} ProcessorContext
  * @property {Object}  webpush           - The web-push module instance (already VAPID-configured).
@@ -952,6 +968,15 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
       throw new Error('Invalid message configuration: no content source available');
     }
 
+    // Auto-extract reasoning BEFORE the sentence split: when reasoning
+    // came from the `<think>` fallback inside message.content, the same
+    // span is still embedded in messageContent and would otherwise leak
+    // as raw markup into ContentPush.
+    const reasoning = readReasoningContent(llmResponse);
+    if (reasoning) {
+      messageContent = stripReasoningTags(messageContent);
+    }
+
     // Sentence splitting (mirrors @rei-standard/amsg-instant
     // splitMessageIntoSentences — keep in lockstep; do not drift). Caller may
     // override the default regex via decryptedPayload.splitPattern (string
@@ -979,7 +1004,7 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
     // `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.
+    // the in-server instant path that has no row id.
     const messageIdBase = task.id != null
       ? `msg_task_${task.id}`
       : `msg_${randomUUID()}_instant`;
@@ -988,7 +1013,6 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
     // 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,
@@ -1374,12 +1398,20 @@ function createScheduleMessageHandler(ctx) {
     const encryptedPayload = encryptForStorage(JSON.stringify(fullTaskData), userKey);
 
     /**
-     * @deprecated Soft-deprecated. For new code, use @rei-standard/amsg-instant.
-     *   This branch is kept for backward compatibility and the existing behavior
-     *   (create task → process → delete) is unchanged. The dedicated amsg-instant
-     *   package is stateless (no DB roundtrip), deployable to Cloudflare Workers,
-     *   and locks the encryption + push-payload contract behind a single version.
-     *   See packages/rei-standard-amsg/instant/README.md.
+     * In-server instant path. Delivers an instant message through this
+     * server's own task queue (create task → process by UUID → delete task).
+     * The task is written to the database before processing, so delivery is
+     * not tied to the request connection: even if the client disconnects, the
+     * row stays and the generation keeps running (and can be retried) for as
+     * long as it needs. Use this when you have a database and want long or
+     * guaranteed-complete generations with no dropped messages.
+     *
+     * The stateless alternative is `@rei-standard/amsg-instant`: it streams
+     * over SSE with a Web Push backup and needs no database, which makes it a
+     * good fit for edge runtimes (e.g. Cloudflare Workers). Its work rides the
+     * response connection, so after the client disconnects it only has the
+     * platform's brief grace window to finish (≈20-30s observed on Deno
+     * Deploy) — ideal for short instant messages that complete quickly.
      */
     // Instant type: check VAPID before creating the task to avoid orphaned rows
     if (payload.messageType === 'instant') {
@@ -1435,11 +1467,20 @@ function createScheduleMessageHandler(ctx) {
     }
 
     /**
-     * @deprecated Soft-deprecated. For new code, use @rei-standard/amsg-instant.
-     *   The "create-task → process-by-uuid → delete-task" sequence below is
-     *   preserved verbatim so existing clients keep working. New integrations
-     *   should call the dedicated amsg-instant endpoint instead — it skips this
-     *   DB round-trip entirely. See packages/rei-standard-amsg/instant/README.md.
+     * In-server instant path. Delivers an instant message through this
+     * server's own task queue (create task → process by UUID → delete task).
+     * The task is written to the database before processing, so delivery is
+     * not tied to the request connection: even if the client disconnects, the
+     * row stays and the generation keeps running (and can be retried) for as
+     * long as it needs. Use this when you have a database and want long or
+     * guaranteed-complete generations with no dropped messages.
+     *
+     * The stateless alternative is `@rei-standard/amsg-instant`: it streams
+     * over SSE with a Web Push backup and needs no database, which makes it a
+     * good fit for edge runtimes (e.g. Cloudflare Workers). Its work rides the
+     * response connection, so after the client disconnects it only has the
+     * platform's brief grace window to finish (≈20-30s observed on Deno
+     * Deploy) — ideal for short instant messages that complete quickly.
      */
     // Instant type: send immediately
     if (payload.messageType === 'instant') {

package/dist/index.mjs

@@ -525,7 +525,7 @@ function readReasoningContent(llmResponse) {
   }
   const content = message?.content;
   if (typeof content === "string") {
-    const match = content.match(/<(think|thinking|thought)>([\s\S]*?)<\/\1>/i);
+    const match = content.match(REASONING_TAG_RE);
     if (match) {
       const trimmed = match[2].trim();
       if (trimmed.length > 0) return trimmed;
@@ -533,6 +533,12 @@ function readReasoningContent(llmResponse) {
   }
   return null;
 }
+var REASONING_TAG_RE = /<(think|thinking|thought)>([\s\S]*?)<\/\1>/i;
+var REASONING_TAG_RE_G = /<(think|thinking|thought)>[\s\S]*?<\/\1>/gi;
+function stripReasoningTags(content) {
+  if (typeof content !== "string" || !content.includes("<")) return content;
+  return content.replace(REASONING_TAG_RE_G, "").trim();
+}
 async function processSingleMessage(task, ctx, providedMasterKey) {
   try {
     const masterKey = providedMasterKey || ctx.masterKey;
@@ -563,6 +569,10 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
     } else {
       throw new Error("Invalid message configuration: no content source available");
     }
+    const reasoning = readReasoningContent(llmResponse);
+    if (reasoning) {
+      messageContent = stripReasoningTags(messageContent);
+    }
     const messages = splitMessageIntoSentences(messageContent, decryptedPayload.splitPattern ?? null);
     if (!ctx.vapid.email || !ctx.vapid.publicKey || !ctx.vapid.privateKey) {
       throw new Error("VAPID configuration missing - push notifications cannot be sent");
@@ -574,7 +584,6 @@ async function processSingleMessage(task, ctx, providedMasterKey) {
     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,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@rei-standard/amsg-server",
-  "version": "2.5.0",
+  "version": "2.5.2",
   "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",