@memtensor/memos-local-openclaw-plugin 1.0.9-beta.1 → 1.0.9

package/README.md

@@ -2,7 +2,7 @@
 
 [![npm version](https://img.shields.io/npm/v/@memtensor/memos-local-openclaw-plugin)](https://www.npmjs.com/package/@memtensor/memos-local-openclaw-plugin)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://github.com/MemTensor/MemOS/blob/main/LICENSE)
-[![Node.js >= 18](https://img.shields.io/badge/node-%3E%3D18-brightgreen)](https://nodejs.org/)
+[![Node.js >= 22](https://img.shields.io/badge/node-%3E%3D22-brightgreen)](https://nodejs.org/)
 [![GitHub](https://img.shields.io/badge/GitHub-Source-181717?logo=github)](https://github.com/MemTensor/MemOS/tree/main/apps/memos-local-openclaw)
 
 Persistent local conversation memory for [OpenClaw](https://github.com/nicepkg/openclaw) AI Agents. Every conversation is automatically captured, semantically indexed, and instantly recallable — with **task summarization & skill evolution**, **team sharing for memories and skills**, and **multi-agent collaborative memory**.
@@ -618,7 +618,7 @@ openclaw plugins install @memtensor/memos-local-openclaw-plugin
 
 ## Troubleshooting
 
-> 📖 **详细排查指南 / Detailed troubleshooting guide:** [docs/troubleshooting.html](https://memtensor.github.io/MemOS/apps/memos-local-openclaw/docs/troubleshooting.html) — 包含逐步排查流程、日志查看方法、完全重装步骤等。
+> 📖 **详细排查指南 / Detailed troubleshooting guide:** [Troubleshooting](https://memos-claw.openmem.net/docs/troubleshooting.html) — 包含逐步排查流程、日志查看方法、完全重装步骤等。
 >
 > 📦 **better-sqlite3 official troubleshooting:** [better-sqlite3 Troubleshooting](https://github.com/WiseLibs/better-sqlite3/blob/master/docs/troubleshooting.md) — the upstream guide for native module build issues.
 
@@ -640,7 +640,7 @@ openclaw plugins install @memtensor/memos-local-openclaw-plugin
    Search for `memos-local`, `failed to load`, `Error`, `Cannot find module`.
 
 4. **Check environment**
-   - Node version: `node -v` (requires **>= 18**)
+   - Node version: `node -v` (requires **>= 22**)
    - Plugin directory exists: `ls ~/.openclaw/extensions/memos-local-openclaw-plugin/package.json`
    - Dependencies installed: `ls ~/.openclaw/extensions/memos-local-openclaw-plugin/node_modules/@sinclair/typebox`
      If missing: `cd ~/.openclaw/extensions/memos-local-openclaw-plugin && npm install --omit=dev`
@@ -693,7 +693,7 @@ This section is for contributors who want to develop, test, or modify the plugin
 
 ### Prerequisites
 
-- **Node.js >= 18** (`node -v`)
+- **Node.js >= 22** (`node -v`)
 - **npm >= 9** (`npm -v`)
 - **C++ build tools** (for `better-sqlite3` native module):
   - macOS: `xcode-select --install`
@@ -701,7 +701,7 @@ This section is for contributors who want to develop, test, or modify the plugin
   - Windows: usually not needed (prebuilt binaries available for LTS Node.js); if build fails, install [Visual Studio Build Tools](https://visualstudio.microsoft.com/visual-cpp-build-tools/)
 - **OpenClaw CLI** installed and available in PATH (`openclaw --version`)
 
-> **`better-sqlite3` build issues?** This is the most common installation problem on macOS and Linux. If `npm install` fails, first install the C++ build tools above, then run `npm rebuild better-sqlite3`. For detailed platform-specific solutions, see the [official better-sqlite3 troubleshooting guide](https://github.com/WiseLibs/better-sqlite3/blob/master/docs/troubleshooting.md) and our [installation troubleshooting page](https://memtensor.github.io/MemOS/apps/memos-local-openclaw/docs/troubleshooting.html).
+> **`better-sqlite3` build issues?** This is the most common installation problem on macOS and Linux. If `npm install` fails, first install the C++ build tools above, then run `npm rebuild better-sqlite3`. For detailed platform-specific solutions, see the [official better-sqlite3 troubleshooting guide](https://github.com/WiseLibs/better-sqlite3/blob/master/docs/troubleshooting.md) and our [installation troubleshooting page](https://memos-claw.openmem.net/docs/troubleshooting.html).
 
 ### Clone & Setup
 
@@ -761,8 +761,7 @@ apps/memos-local-openclaw/
 | `node_modules/` | npm dependencies | `npm install` |
 | `dist/` | Compiled JavaScript output | `npm run build` |
 | `package-lock.json` | Dependency lock file | `npm install` (auto-generated) |
-| `www/` | Memory Viewer static site (local preview) | Started automatically by the plugin |
-| `docs/` | Documentation HTML pages | Built from source or viewed at the hosted URL |
+| `www/` | Memory Viewer static site & documentation pages | Started automatically by the plugin |
 | `ppt/` | Presentation files (internal use) | Not needed for development |
 | `.env` | Local environment variables | Copy from `.env.example` |
 

package/index.ts

@@ -389,6 +389,8 @@ const memosLocalPlugin = {
     let currentAgentId = "main";
     const getCurrentOwner = () => `agent:${currentAgentId}`;
 
+    const globalRef = globalThis as any;
+    
     // ─── Check allowPromptInjection policy ───
     // When allowPromptInjection=false, the prompt mutation fields (such as prependContext) in the hook return value
     // will be stripped by the framework. Skip auto-recall to avoid unnecessary LLM/embedding calls.
@@ -2382,8 +2384,45 @@ Groups: ${groupNames.length > 0 ? groupNames.join(", ") : "(none)"}`,
 
     let serviceStarted = false;
 
-    const startServiceCore = async () => {
+    function isGatewayStartCommand(): boolean {
+      const args = process.argv.map(a => String(a || "").toLowerCase());
+      const gIdx = args.lastIndexOf("gateway");
+      if (gIdx === -1) return false;
+      const next = args[gIdx + 1];
+      return !next || next.startsWith("-") || next === "start" || next === "restart";
+    }
+
+    const startServiceCore = async (isHostStart = false) => {
+      if (!isGatewayStartCommand()) {
+        api.logger.info("memos-local: not a gateway start command, skipping service startup.");
+        return;
+      }
+
+      if (globalRef.__memosLocalPluginStopPromise) {
+        await globalRef.__memosLocalPluginStopPromise;
+        globalRef.__memosLocalPluginStopPromise = undefined;
+      }
       if (serviceStarted) return;
+      
+      if (!isHostStart) {
+        if (globalRef.__memosLocalPluginActiveService && globalRef.__memosLocalPluginActiveService !== service) {
+          api.logger.info("memos-local: aborting startServiceCore because a newer plugin instance is active.");
+          return;
+        }
+      }
+
+      if (globalRef.__memosLocalPluginActiveService && globalRef.__memosLocalPluginActiveService !== service) {
+        api.logger.info("memos-local: Stopping previous plugin instance due to start of new instance...");
+        const oldService = globalRef.__memosLocalPluginActiveService;
+        globalRef.__memosLocalPluginActiveService = undefined;
+        try {
+          await oldService.stop({ preserveDb: true });
+        } catch (e: any) {
+          api.logger.warn(`memos-local: Error stopping previous instance: ${e}`);
+        }
+      }
+      
+      globalRef.__memosLocalPluginActiveService = service;
       serviceStarted = true;
 
       if (hubServer) {
@@ -2425,33 +2464,45 @@ Groups: ${groupNames.length > 0 ? groupNames.join(", ") : "(none)"}`,
       );
     };
 
-    api.registerService({
+    const service = {
       id: "memos-local-openclaw-plugin",
-      start: async () => { await startServiceCore(); },
-      stop: async () => {
+      start: async () => { await startServiceCore(true); },
+      stop: async (options?: { preserveDb?: boolean }) => {
+        await viewer.stop();
+        if (globalRef.__memosLocalPluginActiveService === service) {
+          globalRef.__memosLocalPluginActiveService = undefined;
+        }
         await worker.flush();
         await telemetry.shutdown();
         await hubServer?.stop();
-        viewer.stop();
-        store.close();
+        
+        // If we are hot-reloading, the new instance is already using the SAME
+        // database file on disk. Closing the sqlite store here might kill the
+        // connection for the new instance or cause locking issues depending on
+        // how the native binding manages handles.
+        if (!options?.preserveDb) {
+          store.close();
+        }
         api.logger.info("memos-local: stopped");
       },
-    });
+    };
+
+    api.registerService(service);
 
     // Fallback: OpenClaw may load this plugin via deferred reload after
     // startPluginServices has already run, so service.start() never fires.
-    // Start on the next tick instead of waiting several seconds; the
-    // serviceStarted guard still prevents duplicate startup if the host calls
-    // service.start() immediately after registration.
-    const SELF_START_DELAY_MS = 0;
-    setTimeout(() => {
-      if (!serviceStarted) {
+    // Start on a delay instead of next tick so the host has time to call
+    // service.start() during normal startup if this is a fresh launch.
+    const SELF_START_DELAY_MS = 2000;
+    const selfStartTimer = setTimeout(() => {
+      if (!serviceStarted && isGatewayStartCommand()) {
         api.logger.info("memos-local: service.start() not called by host, self-starting viewer...");
         startServiceCore().catch((err) => {
           api.logger.warn(`memos-local: self-start failed: ${err}`);
         });
       }
     }, SELF_START_DELAY_MS);
+    if (selfStartTimer.unref) selfStartTimer.unref();
   },
 };
 

package/openclaw.plugin.json

@@ -20,7 +20,7 @@
     }
   },
   "requirements": {
-    "node": ">=18.0.0",
+    "node": ">=22.0.0",
     "openclaw": ">=2026.2.0"
   },
   "setup": {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@memtensor/memos-local-openclaw-plugin",
-  "version": "1.0.9-beta.1",
+  "version": "1.0.9",
   "description": "MemOS Local memory plugin for OpenClaw — full-write, hybrid-recall, progressive retrieval",
   "type": "module",
   "main": "index.ts",
@@ -45,7 +45,7 @@
   ],
   "license": "MIT",
   "engines": {
-    "node": ">=18.0.0 <25.0.0"
+    "node": ">=22.0.0"
   },
   "dependencies": {
     "@huggingface/transformers": "^3.8.0",

package/src/ingest/providers/anthropic.ts

@@ -121,12 +121,12 @@ export async function generateTaskTitleAnthropic(
     headers,
     body: JSON.stringify({
       model,
-      max_tokens: 100,
+      max_tokens: 2000,
       temperature: 0,
       system: TASK_TITLE_PROMPT,
       messages: [{ role: "user", content: text }],
     }),
-    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
   });
 
   if (!resp.ok) {
@@ -189,12 +189,12 @@ export async function judgeNewTopicAnthropic(
     headers,
     body: JSON.stringify({
       model,
-      max_tokens: 10,
+      max_tokens: 2000,
       temperature: 0,
       system: TOPIC_JUDGE_PROMPT,
       messages: [{ role: "user", content: userContent }],
     }),
-    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
   });
 
   if (!resp.ok) {
@@ -223,10 +223,12 @@ RULES:
 
 OUTPUT — JSON only:
 {"relevant":[1,3],"sufficient":true}
-- "relevant": candidate numbers whose content helps answer the query. [] if none can help.
-- "sufficient": true only if the selected memories fully answer the query.`;
+- "relevant": candidate numbers whose content helps answer the query. [] if none can help. Duplicates removed — only unique information.
+- "sufficient": true only if the selected memories fully answer the query.
 
-import type { FilterResult } from "./openai";
+IMPORTANT FOR REASONING MODELS: After your analysis, you MUST output a valid JSON object in this exact format. Do not output any text after the JSON object.`;
+
+import { parseFilterResult, type FilterResult } from "./openai";
 export type { FilterResult } from "./openai";
 
 export async function filterRelevantAnthropic(
@@ -256,12 +258,12 @@ export async function filterRelevantAnthropic(
     headers,
     body: JSON.stringify({
       model,
-      max_tokens: 200,
+      max_tokens: 2000,
       temperature: 0,
       system: FILTER_RELEVANT_PROMPT,
       messages: [{ role: "user", content: `QUERY: ${query}\n\nCANDIDATES:\n${candidateText}` }],
     }),
-    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
   });
 
   if (!resp.ok) {
@@ -275,23 +277,6 @@ export async function filterRelevantAnthropic(
   return parseFilterResult(raw, log);
 }
 
-function parseFilterResult(raw: string, log: Logger): FilterResult {
-  try {
-    const match = raw.match(/\{[\s\S]*\}/);
-    if (match) {
-      const obj = JSON.parse(match[0]);
-      if (obj && Array.isArray(obj.relevant)) {
-        return {
-          relevant: obj.relevant.filter((n: any) => typeof n === "number"),
-          sufficient: obj.sufficient === true,
-        };
-      }
-    }
-  } catch {}
-  log.warn(`filterRelevant: failed to parse LLM output: "${raw}", fallback to all+insufficient`);
-  return { relevant: [], sufficient: false };
-}
-
 export async function summarizeAnthropic(
   text: string,
   cfg: SummarizerConfig,
@@ -311,7 +296,7 @@ export async function summarizeAnthropic(
     headers,
     body: JSON.stringify({
       model,
-      max_tokens: 100,
+      max_tokens: 2000,
       temperature: cfg.temperature ?? 0,
       system: SYSTEM_PROMPT,
       messages: [{ role: "user", content: `[TEXT TO SUMMARIZE]\n${text}\n[/TEXT TO SUMMARIZE]` }],
@@ -358,12 +343,12 @@ export async function judgeDedupAnthropic(
     headers,
     body: JSON.stringify({
       model,
-      max_tokens: 300,
+      max_tokens: 2000,
       temperature: 0,
       system: DEDUP_JUDGE_PROMPT,
       messages: [{ role: "user", content: `NEW MEMORY:\n${newSummary}\n\nEXISTING MEMORIES:\n${candidateText}` }],
     }),
-    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
   });
 
   if (!resp.ok) {

package/src/ingest/providers/bedrock.ts

@@ -126,9 +126,9 @@ export async function generateTaskTitleBedrock(
     body: JSON.stringify({
       system: [{ text: TASK_TITLE_PROMPT }],
       messages: [{ role: "user", content: [{ text }] }],
-      inferenceConfig: { temperature: 0, maxTokens: 100 },
+      inferenceConfig: { temperature: 0, maxTokens: 2000 },
     }),
-    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
   });
 
   if (!resp.ok) {
@@ -195,9 +195,9 @@ export async function judgeNewTopicBedrock(
     body: JSON.stringify({
       system: [{ text: TOPIC_JUDGE_PROMPT }],
       messages: [{ role: "user", content: [{ text: userContent }] }],
-      inferenceConfig: { temperature: 0, maxTokens: 10 },
+      inferenceConfig: { temperature: 0, maxTokens: 2000 },
     }),
-    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
   });
 
   if (!resp.ok) {
@@ -226,10 +226,12 @@ RULES:
 
 OUTPUT — JSON only:
 {"relevant":[1,3],"sufficient":true}
-- "relevant": candidate numbers whose content helps answer the query. [] if none can help.
-- "sufficient": true only if the selected memories fully answer the query.`;
+- "relevant": candidate numbers whose content helps answer the query. [] if none can help. Duplicates removed — only unique information.
+- "sufficient": true only if the selected memories fully answer the query.
 
-import type { FilterResult } from "./openai";
+IMPORTANT FOR REASONING MODELS: After your analysis, you MUST output a valid JSON object in this exact format. Do not output any text after the JSON object.`;
+
+import { parseFilterResult, type FilterResult } from "./openai";
 export type { FilterResult } from "./openai";
 
 export async function filterRelevantBedrock(
@@ -263,9 +265,9 @@ export async function filterRelevantBedrock(
     body: JSON.stringify({
       system: [{ text: FILTER_RELEVANT_PROMPT }],
       messages: [{ role: "user", content: [{ text: `QUERY: ${query}\n\nCANDIDATES:\n${candidateText}` }] }],
-      inferenceConfig: { temperature: 0, maxTokens: 200 },
+      inferenceConfig: { temperature: 0, maxTokens: 2000 },
     }),
-    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
   });
 
   if (!resp.ok) {
@@ -279,23 +281,6 @@ export async function filterRelevantBedrock(
   return parseFilterResult(raw, log);
 }
 
-function parseFilterResult(raw: string, log: Logger): FilterResult {
-  try {
-    const match = raw.match(/\{[\s\S]*\}/);
-    if (match) {
-      const obj = JSON.parse(match[0]);
-      if (obj && Array.isArray(obj.relevant)) {
-        return {
-          relevant: obj.relevant.filter((n: any) => typeof n === "number"),
-          sufficient: obj.sufficient === true,
-        };
-      }
-    }
-  } catch {}
-  log.warn(`filterRelevant: failed to parse LLM output: "${raw}", fallback to all+insufficient`);
-  return { relevant: [], sufficient: false };
-}
-
 export async function summarizeBedrock(
   text: string,
   cfg: SummarizerConfig,
@@ -321,7 +306,7 @@ export async function summarizeBedrock(
       messages: [{ role: "user", content: [{ text: `[TEXT TO SUMMARIZE]\n${text}\n[/TEXT TO SUMMARIZE]` }] }],
       inferenceConfig: {
         temperature: cfg.temperature ?? 0,
-        maxTokens: 100,
+        maxTokens: 2000,
       },
     }),
     signal: AbortSignal.timeout(cfg.timeoutMs ?? 30_000),
@@ -364,9 +349,9 @@ export async function judgeDedupBedrock(
     body: JSON.stringify({
       system: [{ text: DEDUP_JUDGE_PROMPT }],
       messages: [{ role: "user", content: [{ text: `NEW MEMORY:\n${newSummary}\n\nEXISTING MEMORIES:\n${candidateText}` }] }],
-      inferenceConfig: { temperature: 0, maxTokens: 300 },
+      inferenceConfig: { temperature: 0, maxTokens: 2000 },
     }),
-    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
   });
 
   if (!resp.ok) {

package/src/ingest/providers/gemini.ts

@@ -124,9 +124,9 @@ export async function generateTaskTitleGemini(
     body: JSON.stringify({
       systemInstruction: { parts: [{ text: TASK_TITLE_PROMPT }] },
       contents: [{ parts: [{ text }] }],
-      generationConfig: { temperature: 0, maxOutputTokens: 100 },
+      generationConfig: { temperature: 0, maxOutputTokens: 2000 },
     }),
-    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
   });
 
   if (!resp.ok) {
@@ -192,9 +192,9 @@ export async function judgeNewTopicGemini(
     body: JSON.stringify({
       systemInstruction: { parts: [{ text: TOPIC_JUDGE_PROMPT }] },
       contents: [{ parts: [{ text: userContent }] }],
-      generationConfig: { temperature: 0, maxOutputTokens: 10 },
+      generationConfig: { temperature: 0, maxOutputTokens: 2000 },
     }),
-    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
   });
 
   if (!resp.ok) {
@@ -223,10 +223,12 @@ RULES:
 
 OUTPUT — JSON only:
 {"relevant":[1,3],"sufficient":true}
-- "relevant": candidate numbers whose content helps answer the query. [] if none can help.
-- "sufficient": true only if the selected memories fully answer the query.`;
+- "relevant": candidate numbers whose content helps answer the query. [] if none can help. Duplicates removed — only unique information.
+- "sufficient": true only if the selected memories fully answer the query.
 
-import type { FilterResult } from "./openai";
+IMPORTANT FOR REASONING MODELS: After your analysis, you MUST output a valid JSON object in this exact format. Do not output any text after the JSON object.`;
+
+import { parseFilterResult, type FilterResult } from "./openai";
 export type { FilterResult } from "./openai";
 
 export async function filterRelevantGemini(
@@ -259,9 +261,9 @@ export async function filterRelevantGemini(
     body: JSON.stringify({
       systemInstruction: { parts: [{ text: FILTER_RELEVANT_PROMPT }] },
       contents: [{ parts: [{ text: `QUERY: ${query}\n\nCANDIDATES:\n${candidateText}` }] }],
-      generationConfig: { temperature: 0, maxOutputTokens: 200 },
+      generationConfig: { temperature: 0, maxOutputTokens: 2000 },
     }),
-    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
   });
 
   if (!resp.ok) {
@@ -275,23 +277,6 @@ export async function filterRelevantGemini(
   return parseFilterResult(raw, log);
 }
 
-function parseFilterResult(raw: string, log: Logger): FilterResult {
-  try {
-    const match = raw.match(/\{[\s\S]*\}/);
-    if (match) {
-      const obj = JSON.parse(match[0]);
-      if (obj && Array.isArray(obj.relevant)) {
-        return {
-          relevant: obj.relevant.filter((n: any) => typeof n === "number"),
-          sufficient: obj.sufficient === true,
-        };
-      }
-    }
-  } catch {}
-  log.warn(`filterRelevant: failed to parse LLM output: "${raw}", fallback to all+insufficient`);
-  return { relevant: [], sufficient: false };
-}
-
 export async function summarizeGemini(
   text: string,
   cfg: SummarizerConfig,
@@ -314,7 +299,7 @@ export async function summarizeGemini(
     body: JSON.stringify({
       systemInstruction: { parts: [{ text: SYSTEM_PROMPT }] },
       contents: [{ parts: [{ text: `[TEXT TO SUMMARIZE]\n${text}\n[/TEXT TO SUMMARIZE]` }] }],
-      generationConfig: { temperature: cfg.temperature ?? 0, maxOutputTokens: 100 },
+      generationConfig: { temperature: cfg.temperature ?? 0, maxOutputTokens: 2000 },
     }),
     signal: AbortSignal.timeout(cfg.timeoutMs ?? 30_000),
   });
@@ -355,9 +340,9 @@ export async function judgeDedupGemini(
     body: JSON.stringify({
       systemInstruction: { parts: [{ text: DEDUP_JUDGE_PROMPT }] },
       contents: [{ parts: [{ text: `NEW MEMORY:\n${newSummary}\n\nEXISTING MEMORIES:\n${candidateText}` }] }],
-      generationConfig: { temperature: 0, maxOutputTokens: 300 },
+      generationConfig: { temperature: 0, maxOutputTokens: 2000 },
     }),
-    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
   });
 
   if (!resp.ok) {

package/src/ingest/providers/index.ts

@@ -230,8 +230,13 @@ export class Summarizer {
       return ruleFallback(cleaned);
     }
 
-    const accept = (s: string | undefined): s is string =>
-      !!s && s.length > 0 && s.length < cleaned.length;
+    const accept = (s: string | undefined): s is string => {
+      if (!s || s.length === 0) return false;
+      // Allow the LLM result if it's shorter than input, or if the input is quite short itself (<= 20 chars).
+      // Summarizing a very short text (e.g. "hi there") might yield a slightly longer formal summary (e.g. "User said hi").
+      if (cleaned.length <= 20) return true;
+      return s.length < cleaned.length;
+    };
 
     let llmCalled = false;
     try {