@joshuaswarren/openclaw-engram 9.0.12 → 9.0.14

package/dist/index.js

@@ -325,6 +325,11 @@ function parseConfig(raw) {
     localLlmRetryBackoffMs: typeof cfg.localLlmRetryBackoffMs === "number" ? cfg.localLlmRetryBackoffMs : 400,
     localLlm400TripThreshold: typeof cfg.localLlm400TripThreshold === "number" ? cfg.localLlm400TripThreshold : 5,
     localLlm400CooldownMs: typeof cfg.localLlm400CooldownMs === "number" ? cfg.localLlm400CooldownMs : 12e4,
+    // Local LLM fast tier (v9.1)
+    localLlmFastEnabled: cfg.localLlmFastEnabled === true,
+    localLlmFastModel: typeof cfg.localLlmFastModel === "string" && cfg.localLlmFastModel.length > 0 ? cfg.localLlmFastModel : "",
+    localLlmFastUrl: typeof cfg.localLlmFastUrl === "string" && cfg.localLlmFastUrl.length > 0 ? cfg.localLlmFastUrl : typeof cfg.localLlmUrl === "string" && cfg.localLlmUrl.length > 0 ? cfg.localLlmUrl : "http://localhost:1234/v1",
+    localLlmFastTimeoutMs: typeof cfg.localLlmFastTimeoutMs === "number" ? cfg.localLlmFastTimeoutMs : 15e3,
     // Gateway config (passed from index.ts for fallback AI)
     gatewayConfig: cfg.gatewayConfig,
     // v3.0 namespaces (default off)
@@ -980,6 +985,7 @@ var LocalLlmClient = class _LocalLlmClient {
   consecutive400s = 0;
   cooldownUntilMs = 0;
   modelRegistry;
+  _disableThinking = false;
   static HEALTH_CHECK_INTERVAL_MS = 6e4;
   // 1 minute
   static LMS_CACHE_INTERVAL_MS = 3e4;
@@ -988,6 +994,14 @@ var LocalLlmClient = class _LocalLlmClient {
     this.config = config;
     this.modelRegistry = modelRegistry;
   }
+  /**
+   * Disable thinking/reasoning mode for models that support it (e.g. Qwen 3.5).
+   * When enabled, adds chat_template_kwargs to suppress chain-of-thought,
+   * reducing latency for fast-tier operations.
+   */
+  set disableThinking(value) {
+    this._disableThinking = value;
+  }
   resolveHomeDir() {
     return this.config.localLlmHomeDir || process.env.HOME || os.homedir();
   }
@@ -1427,6 +1441,9 @@ var LocalLlmClient = class _LocalLlmClient {
       if (options.responseFormat?.type === "json_schema") {
         requestBody.response_format = options.responseFormat;
       }
+      if (this._disableThinking) {
+        requestBody.chat_template_kwargs = { enable_thinking: false };
+      }
       const baseUrl = this.config.localLlmUrl.replace("localhost", "127.0.0.1").replace(/\/+$/, "");
       const chatUrl = `${baseUrl}/chat/completions`;
       const requestBodyJson = JSON.stringify(requestBody);
@@ -16388,6 +16405,7 @@ var Orchestrator = class _Orchestrator {
   sessionObserver;
   summarizer;
   localLlm;
+  fastLlm;
   modelRegistry;
   relevance;
   negatives;
@@ -16489,6 +16507,14 @@ var Orchestrator = class _Orchestrator {
     this.policyRuntime = new PolicyRuntimeManager(config.memoryDir, config);
     this.summarizer = new HourlySummarizer(config, config.gatewayConfig, this.modelRegistry, this.transcript);
     this.localLlm = new LocalLlmClient(config, this.modelRegistry);
+    this.fastLlm = config.localLlmFastEnabled ? (() => {
+      const client = new LocalLlmClient(
+        { ...config, localLlmModel: config.localLlmFastModel || config.localLlmModel, localLlmUrl: config.localLlmFastUrl, localLlmTimeoutMs: config.localLlmFastTimeoutMs },
+        this.modelRegistry
+      );
+      client.disableThinking = true;
+      return client;
+    })() : this.localLlm;
     this.extraction = new ExtractionEngine(config, this.localLlm, config.gatewayConfig, this.modelRegistry);
     this.threading = new ThreadingManager(
       path30.join(config.memoryDir, "threads"),
@@ -17892,7 +17918,7 @@ ${tmtNode.summary}`);
             id: r.path,
             snippet: r.snippet || r.path
           })),
-          local: this.localLlm,
+          local: this.fastLlm,
           enabled: true,
           timeoutMs: this.config.rerankTimeoutMs,
           maxCandidates: this.config.rerankMaxCandidates,
@@ -19383,7 +19409,7 @@ _Context: ${topQuestion.context}_`
           try {
             const factsText = entity.facts.slice(0, 10).join("; ");
             const prompt = `Summarize this entity in one sentence. Entity: ${entity.name} (${entity.type}). Facts: ${factsText}`;
-            const response = await this.localLlm.chatCompletion(
+            const response = await this.fastLlm.chatCompletion(
               [
                 { role: "system", content: "Respond with a single concise sentence summarizing the entity. No JSON, just plain text." },
                 { role: "user", content: prompt }
@@ -19485,7 +19511,7 @@ _Context: ${topQuestion.context}_`
           const prompt = `You are a memory archivist. Summarize the following ${level}-level memories into 3\u20135 sentences, preserving key facts, decisions, and preferences.
 
 ${texts.map((t, i) => `[${i + 1}] ${t}`).join("\n\n")}`;
-          const response = await this.localLlm.chatCompletion(
+          const response = await this.fastLlm.chatCompletion(
             [
               { role: "system", content: "Respond with a 3\u20135 sentence narrative summary. No JSON, just plain prose." },
               { role: "user", content: prompt }
@@ -19542,7 +19568,7 @@ ${texts.map((t, i) => `[${i + 1}] ${t}`).join("\n\n")}`;
           "Input candidate:",
           JSON.stringify(baseline)
         ].join("\n");
-        const response = await this.localLlm.chatCompletion(
+        const response = await this.fastLlm.chatCompletion(
           [
             { role: "system", content: "Respond with strict JSON only. No markdown." },
             { role: "user", content: prompt }
@@ -20154,7 +20180,7 @@ ${lines.join("\n\n")}`;
           id: r.path,
           snippet: r.snippet || r.path
         })),
-        local: this.localLlm,
+        local: this.fastLlm,
         enabled: true,
         timeoutMs: this.config.rerankTimeoutMs,
         maxCandidates: this.config.rerankMaxCandidates,
@@ -27726,7 +27752,7 @@ var index_default = {
     });
     initLogger(api.logger, cfg.debug);
     log.info(
-      `initialized (debug=${cfg.debug}, qmdEnabled=${cfg.qmdEnabled}, transcriptEnabled=${cfg.transcriptEnabled}, hourlySummariesEnabled=${cfg.hourlySummariesEnabled}, localLlmEnabled=${cfg.localLlmEnabled})`
+      `initialized (debug=${cfg.debug}, qmdEnabled=${cfg.qmdEnabled}, transcriptEnabled=${cfg.transcriptEnabled}, hourlySummariesEnabled=${cfg.hourlySummariesEnabled}, localLlmEnabled=${cfg.localLlmEnabled}${cfg.localLlmFastEnabled ? `, fastLlm=${cfg.localLlmFastModel || "(primary)"}` : ""})`
     );
     const existing = globalThis.__openclawEngramOrchestrator;
     const orchestrator = existing?.recall ? existing : new Orchestrator(cfg);