omnius 1.0.419 → 1.0.420

package/dist/index.js

@@ -696639,9 +696639,23 @@ function buildRealtimeVoiceMessages(turns, maxMessages = MAX_REALTIME_MODEL_MESS
       dynamic.push(turn);
     }
   }
-  if (latestSnapshot) dynamic.push(latestSnapshot);
+  if (latestSnapshot) {
+    let lastUserIdx = -1;
+    for (let i2 = dynamic.length - 1; i2 >= 0; i2--) {
+      if (dynamic[i2].role === "user") {
+        lastUserIdx = i2;
+        break;
+      }
+    }
+    if (lastUserIdx >= 0) dynamic.splice(lastUserIdx, 0, latestSnapshot);
+    else dynamic.push(latestSnapshot);
+  }
   const available = Math.max(0, maxMessages - pinned.length);
-  return [...pinned, ...dynamic.slice(-available)];
+  let windowed = dynamic.slice(-available);
+  if (latestSnapshot && available > 0 && !windowed.includes(latestSnapshot)) {
+    windowed = [latestSnapshot, ...windowed.slice(-(available - 1))];
+  }
+  return [...pinned, ...windowed];
 }
 var VAD_SILENCE_MS, MAX_SEGMENT_MS, SUMMARY_INJECTION_INTERVAL, MAX_CONTEXT_TURNS, MAX_REALTIME_MODEL_MESSAGES, CONTEXT_SNAPSHOT_PREFIX, MAX_VOICE_REPLY_CHARS, AGENT_ECHO_WINDOW_MS, AGENT_ECHO_SIMILARITY, SYSTEM_PROMPT2, MIN_SIGNAL_SCORE, NOISE_ONLY_RE, VoiceChatSession;
 var init_voicechat = __esm({
@@ -696659,7 +696673,9 @@ var init_voicechat = __esm({
     SYSTEM_PROMPT2 = `You are a voice assistant having a live spoken conversation. Keep responses extremely brief — 1-2 sentences max. You're speaking aloud, not writing. Be conversational, direct, and helpful. Don't use markdown or formatting — just natural speech.
 
 Rules:
-- Live perception: each turn you may receive a read-only "Context snapshot" system message containing what the camera currently sees (objects, people, classifications, recent visual events) and what is currently heard (sound scene, recent sounds, transcripts). Treat it as your own live sight and hearing. When the user asks what you see, who is there, or what you hear, answer directly from that snapshot. If the snapshot is missing or stale, say you can't see/hear clearly right now — do not invent observations.
+- ALWAYS answer the user's most recent message directly. The live context snapshot is background perception — never the topic. Do not describe cameras, feeds, or surroundings unless the user's latest message asks about them.
+- Never repeat your previous answer. If you already described the scene, don't describe it again unless explicitly asked again — respond to what the user just said.
+- Live perception: each turn you may receive a read-only "Context snapshot" system message containing what the camera currently sees (objects, people, classifications, recent visual events) and what is currently heard (sound scene, recent sounds, transcripts). Treat it as your own live sight and hearing. When the user asks what you SEE, answer from the camera fields; when the user asks what you HEAR, answer from the audio/sound fields. If the snapshot is missing or stale, say you can't see/hear clearly right now — do not invent observations.
 - Never invent environment facts (cwd, OS, specs, repo state). If you need a precise fact from the main agent, request a tool by outputting on a single line EXACTLY one JSON object: {"tool": string, "args": object} and nothing else. Then wait for the tool result before answering.
 - You may also request to relay a user task to the main agent by emitting {"tool":"voice_to_main","args":{"message":"...","start":true}}.
 - Prefer tools for factual queries; otherwise, answer directly with a short reply.
@@ -697015,6 +697031,19 @@ ${toolOutput}` });
             this.context.push({ role: "system", content: `You have tools. Use them. ${this.toolCatalogNote}` });
             response = await this.streamOllamaInference(this.abortController.signal);
           }
+          if (this.lastAgentSpeech?.text && response.trim()) {
+            const prevTokens = comparableTokens(this.lastAgentSpeech.text);
+            const nextTokens = comparableTokens(stripToolJsonLines(response));
+            const similarity3 = jaccardSimilarity(prevTokens, nextTokens);
+            if (similarity3 >= 0.8 && nextTokens.length >= 8) {
+              this.context.push({
+                role: "system",
+                content: `Your draft reply repeats your previous answer. Do NOT describe the scene or cameras again. Respond directly and briefly to the user's latest message: "${truncateForLog(lastUser ?? "", 200)}"`
+              });
+              const retry = await this.streamOllamaInference(this.abortController.signal);
+              if (retry.trim()) response = retry;
+            }
+          }
           if (response.trim()) {
             const reply = extractVoiceModelReply(stripToolJsonLines(response.trim()));
             const finalSpoken = reply.text;

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.419",
+  "version": "1.0.420",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.419",
+      "version": "1.0.420",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.419",
+  "version": "1.0.420",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",