open-agents-ai 0.80.0 → 0.81.0

package/dist/index.js

@@ -24663,11 +24663,14 @@ var init_call_agent = __esm({
         const feed = getActivityFeed();
         const systemPrompt = this.buildSystemPrompt();
         const runnerOpts = {
-          maxTurns: this.tier === "admin" ? 15 : 5,
-          maxTokens: 4096,
+          maxTurns: this.tier === "admin" ? 8 : 3,
+          // Keep low for fast voice responses
+          maxTokens: 2048,
+          // Short responses for TTS
           temperature: 0.3,
-          requestTimeoutMs: 3e4,
-          taskTimeoutMs: 12e4,
+          requestTimeoutMs: 2e4,
+          // Fast timeout for voice responsiveness
+          taskTimeoutMs: 6e4,
           modelTier: this.opts.modelTier ?? "large",
           contextWindowSize: this.opts.contextWindowSize,
           personality: this.opts.personality,
@@ -24766,17 +24769,17 @@ var init_call_agent = __esm({
           const historyContext = this.conversationHistory.slice(-10).map((h) => `${h.role === "user" ? "User" : "You"}: ${h.text}`).join("\n");
           const feed = getActivityFeed();
           const activitySummary = feed.getSummary(this.tier === "admin" ? 20 : 10, this.tier === "admin");
+          const wantsAction = /\b(read|open|show|run|execute|check|look at|find|search|grep|edit|write|fix|test|build|deploy|install)\b/i.test(text) && !/\b(how are you|what's up|hello|hi|hey|can you hear|stop|quit|bye)\b/i.test(text);
           const taskPrompt = [
-            `The user said (via voice): "${text}"`,
-            "",
-            "Recent conversation:",
-            historyContext,
+            `User said: "${text}"`,
             "",
-            "Recent agent activity:",
-            activitySummary,
-            "",
-            "Respond conversationally and concisely. Your response will be spoken aloud via TTS.",
-            "Call task_complete with your spoken response as the summary."
+            historyContext ? `Conversation so far:
+${historyContext}
+` : "",
+            `Background activity:
+${activitySummary}
+`,
+            wantsAction ? "The user is requesting an action. You may use tools, then call task_complete with a brief spoken summary of what you did." : "RESPOND IMMEDIATELY \u2014 call task_complete with your spoken reply. Do NOT use any tools. Just answer conversationally in 1-3 sentences."
           ].join("\n");
           const result = await this.runner.run(taskPrompt, `Working directory: ${this.repoRoot}`);
           if (result.summary) {
@@ -24797,18 +24800,24 @@ var init_call_agent = __esm({
       }
       buildSystemPrompt() {
         const base = [
-          "You are a voice assistant for an AI coding agent. Users are speaking to you through a live audio call.",
-          "Keep responses SHORT and conversational \u2014 they will be spoken aloud via text-to-speech.",
-          "Avoid code blocks, markdown formatting, or long technical explanations in your spoken responses.",
-          "If the user asks about what the agent is doing, summarize the recent activity concisely."
+          "You are a voice assistant on a LIVE AUDIO CALL. This is a real-time conversation.",
+          "",
+          "CRITICAL RULES FOR VOICE CALLS:",
+          "1. ALWAYS respond IMMEDIATELY with speech. Do NOT use tools before responding.",
+          "2. Your response goes through text-to-speech \u2014 keep it SHORT (1-3 sentences).",
+          "3. NEVER use code blocks, markdown, or long technical text.",
+          "4. Be conversational and natural, like talking to a colleague.",
+          "5. Call task_complete with your spoken response as the summary.",
+          "6. Only use tools (file_read, grep, shell, etc.) if the user EXPLICITLY asks you to look something up, run a command, or make a change. For normal conversation, NEVER call tools.",
+          "7. If the user asks what's happening, summarize from the activity context below \u2014 do NOT run tools to find out."
         ];
         if (this.opts.emotionContext) {
-          base.push("", "Current emotional context:", this.opts.emotionContext);
+          base.push("", "Mood:", this.opts.emotionContext);
         }
         if (this.tier === "admin") {
-          base.push("", "This is an ADMIN call \u2014 you have full access to all tools and can execute commands, read/write files, search code, etc.", "You can take actions on behalf of the user just like the main agent.");
+          base.push("", "ADMIN call \u2014 you CAN use tools IF the user explicitly requests an action (e.g. 'read that file', 'run the tests').", "But for general chat, status questions, or greetings \u2014 respond immediately WITHOUT tools.");
         } else {
-          base.push("", "This is a PUBLIC call \u2014 you have read-only access. You can read files and search code but cannot modify anything.", "You can answer questions about the project and explain what the agent is working on.", "Be helpful and friendly but do not execute any write operations.");
+          base.push("", "PUBLIC call \u2014 read-only access. Answer questions about the project conversationally.");
         }
         return base.join("\n");
       }
@@ -39452,6 +39461,11 @@ Respond concisely and safely. Remember: you are talking to the general public.`;
       if (voiceSession?.isActive) {
         return voiceSession.tunnelUrl;
       }
+      if (!voiceEngine.enabled || !voiceEngine.ready) {
+        writeContent(() => renderInfo("Auto-enabling voice for call session..."));
+        const voiceMsg = await voiceEngine.toggle();
+        writeContent(() => renderInfo(voiceMsg));
+      }
       if (!adminSessionKey) {
         adminSessionKey = generateSessionKey();
       }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "open-agents-ai",
-  "version": "0.80.0",
+  "version": "0.81.0",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",