omnius 1.0.413 → 1.0.414

package/dist/index.js

@@ -41700,28 +41700,26 @@ var init_transcribe_tool = __esm({
         }
         if (effectiveModel !== askedModel)
           model = effectiveModel;
-        const failures = [];
-        let managedPython = "";
-        try {
-          managedPython = await this.ensureManagedWhisperRuntime();
-          process.env.TRANSCRIBE_PYTHON = managedPython;
-          await ensureTranscribeCliPythonModule(managedPython);
-        } catch (err) {
-          failures.push(`managed ASR runtime bootstrap: ${err instanceof Error ? err.message : String(err)}`);
-        }
         if (backend === "managed-whisper") {
           const managed2 = await this.execViaManagedWhisper(filePath, model, start2);
           if (managed2.success)
             return managed2;
-          if (managed2.error)
-            failures.push(managed2.error);
           return {
             success: false,
             output: "",
-            error: `Managed Whisper ASR failed. ${failures.join(" | ").slice(0, 1200)}`,
+            error: `Managed Whisper ASR failed. ${String(managed2.error || managed2.output || "unknown error").slice(0, 1200)}`,
             durationMs: performance.now() - start2
           };
         }
+        const failures = [];
+        let managedPython = "";
+        try {
+          managedPython = await this.ensureManagedWhisperRuntime();
+          process.env.TRANSCRIBE_PYTHON = managedPython;
+          await ensureTranscribeCliPythonModule(managedPython);
+        } catch (err) {
+          failures.push(`managed ASR runtime bootstrap: ${err instanceof Error ? err.message : String(err)}`);
+        }
         const tc = await loadTranscribeCli() ?? await ensureManagedTranscribeCliRuntime();
         if (tc && backend !== "managed-whisper") {
           try {
@@ -603158,6 +603156,16 @@ function isAudioFailureText(value2) {
   if (!text2) return false;
   return /^(?:ASR|VAD|Transcription)\s+(?:failed|unavailable)|No module named ['"]?transcribe_cli|torchcodec|torchaudio version/i.test(text2);
 }
+function sanitizeLiveAudioError(value2) {
+  const text2 = String(value2 ?? "").trim();
+  if (!text2) return "";
+  if (/transcribe-cli module:\s*No module named ['"]?transcribe_cli/i.test(text2) || /No module named ['"]?transcribe_cli/i.test(text2)) {
+    const parts = text2.split(/\s*\|\s*/g).map((part) => part.trim()).filter((part) => part && !/transcribe-cli module:\s*No module named ['"]?transcribe_cli/i.test(part) && !/No module named ['"]?transcribe_cli/i.test(part));
+    const preserved = parts.join(" | ").trim();
+    return preserved || "ASR unavailable: managed Whisper runtime is still bootstrapping or failed; transcribe-cli is not required for live ASR.";
+  }
+  return text2;
+}
 function readPcm16WavActivity(filePath, bins = 36) {
   try {
     const buf = readFileSync88(filePath);
@@ -603272,39 +603280,12 @@ function parseLiveSpeechDecision(value2) {
   };
 }
 function fallbackLiveSpeechDecision(proposal) {
-  if (proposal.kind === "unknown_person") {
-    return {
-      speak: true,
-      text: proposal.defaultText,
-      confidence: 0.72,
-      action: "speak",
-      reason: "unknown person visible and identity clarification is useful"
-    };
-  }
-  if (proposal.kind === "audio_intent") {
-    return {
-      speak: true,
-      text: proposal.defaultText,
-      confidence: 0.68,
-      action: "speak",
-      reason: "audio intake classified speech as addressed to Omnius"
-    };
-  }
-  if (proposal.urgency === "high") {
-    return {
-      speak: true,
-      text: proposal.defaultText,
-      confidence: 0.6,
-      action: "speak",
-      reason: "high-urgency live trigger"
-    };
-  }
   return {
     speak: false,
     text: "",
     confidence: 0.55,
-    action: "silent",
-    reason: "ambient event did not justify interrupting with speech"
+    action: proposal.urgency === "high" ? "queue_review" : "silent",
+    reason: "live speech arbiter unavailable; refusing canned vocal fallback"
   };
 }
 async function classifyLiveSpeech(proposal, config) {
@@ -603337,7 +603318,7 @@ async function classifyLiveSpeech(proposal, config) {
             `urgency=${proposal.urgency}`,
             `summary=${proposal.summary}`,
             `evidence=${proposal.evidence}`,
-            `default_utterance=${proposal.defaultText}`
+            proposal.speechGoal ? `speech_goal=${proposal.speechGoal}` : ""
           ].join("\n").slice(0, 1600)
         }
       ]
@@ -603637,7 +603618,7 @@ function formatLiveDashboardFromSnapshot(snapshot, opts = {}) {
     lines.push(`├${horizontal}┤`);
     const transcriptFailure = isAudioFailureText(snapshot.audio.transcript) ? snapshot.audio.transcript : "";
     const transcript = transcriptFailure ? "" : snapshot.audio.transcript;
-    const audioError = [snapshot.audio.error, transcriptFailure].filter(Boolean).join(" | ");
+    const audioError = sanitizeLiveAudioError([snapshot.audio.error, transcriptFailure].filter(Boolean).join(" | "));
     const audioBits = [
       `audio input=${snapshot.audio.input || snapshot.config.selectedAudioInput || "none"}`,
       snapshot.audio.activity ? `activity ${snapshot.audio.activity.waveform} rms=${snapshot.audio.activity.rmsDb.toFixed(1)}dB active=${Math.round(snapshot.audio.activity.activeRatio * 100)}%` : "",
@@ -604020,7 +604001,7 @@ function formatLiveSensorContextFromSnapshot(snapshot, opts = {}) {
   if (snapshot.audio) {
     const transcriptFailure = isAudioFailureText(snapshot.audio.transcript) ? snapshot.audio.transcript : "";
     const transcript = transcriptFailure ? "" : snapshot.audio.transcript;
-    const audioError = [snapshot.audio.error, transcriptFailure].filter(Boolean).join(" | ");
+    const audioError = sanitizeLiveAudioError([snapshot.audio.error, transcriptFailure].filter(Boolean).join(" | "));
     const audioEvents = (snapshot.events ?? []).filter((event) => event.kind === "audio_transcript" || event.kind === "audio_sound").sort((a2, b) => b.observedAt - a2.observedAt).slice(0, 8);
     if (audioEvents.length > 0) {
       lines.push("");
@@ -604464,6 +604445,10 @@ var init_live_sensors = __esm({
       setCameraRotation(device, rotation, source, evidence, confidence2) {
         const target = device || this.config.selectedCamera || firstDeviceId(this.devices.video);
         if (!target) return null;
+        const current = this.config.cameraOrientation?.[target];
+        if (source === "auto" && current?.source === "manual") {
+          return current;
+        }
         const orientation = {
           rotation,
           source,
@@ -604929,7 +604914,7 @@ var init_live_sensors = __esm({
               source,
               summary: `${unknownPeople.length} unknown individual(s) observed on live camera ${source}`,
               evidence: unknownPeople.map((entry) => `${entry.cropPath || entry.trackId || entry.source} ${entry.evidence || ""}`).join("; "),
-              defaultText: "I see someone here, but I do not know who. Who is there?",
+              speechGoal: "If appropriate, ask the visible unknown person for the name Omnius should associate with the face crop. Do not claim an identity.",
               urgency: "high",
               observedAt: sampledAt
             }, 9e4);
@@ -604946,7 +604931,7 @@ var init_live_sensors = __esm({
               source,
               summary: `Visual trigger hit in live camera stream ${source}`,
               evidence: triggerEvents.map((entry) => entry.summary).join("; "),
-              defaultText: "I noticed something important in view. Should I inspect it?",
+              speechGoal: "If immediate clarification is warranted, ask whether Omnius should inspect the visible trigger.",
               urgency: "high",
               observedAt: sampledAt
             }, 45e3);
@@ -605143,7 +605128,7 @@ ${output}`).join("\n\n");
               source: capture.device,
               summary: "Live audio intake classified speech as intended for Omnius",
               evidence: `transcript=${transcript}; confidence=${intake.confidence.toFixed(2)}; reason=${intake.reason}`,
-              defaultText: "I heard you. I am checking the live context now.",
+              speechGoal: "Respond naturally to the addressed utterance using live context; keep it brief.",
               urgency: "high",
               observedAt: this.snapshot.audio.updatedAt
             }, 2e4);
@@ -695869,7 +695854,10 @@ var init_task_manager_singleton = __esm({
 // packages/cli/src/tui/voicechat.ts
 var voicechat_exports = {};
 __export(voicechat_exports, {
-  VoiceChatSession: () => VoiceChatSession
+  VoiceChatSession: () => VoiceChatSession,
+  buildRealtimeVoiceMessages: () => buildRealtimeVoiceMessages,
+  extractVoiceModelReply: () => extractVoiceModelReply,
+  isLikelyAgentSpeechEcho: () => isLikelyAgentSpeechEcho
 });
 import { EventEmitter as EventEmitter12 } from "node:events";
 function clamp0116(x) {
@@ -695966,19 +695954,138 @@ function stripToolJsonLines(text2) {
   });
   return kept.join("\n").trim();
 }
-var VAD_SILENCE_MS, MAX_SEGMENT_MS, MAX_CONTEXT_TURNS, SYSTEM_PROMPT2, MIN_SIGNAL_SCORE, NOISE_ONLY_RE, VoiceChatSession;
+function compactSpeechText(value2, maxLength = 8e3) {
+  const text2 = String(value2 ?? "").replace(/\s+/g, " ").replace(/\s+([,.!?;:])/g, "$1").trim();
+  return text2 ? text2.slice(0, maxLength) : "";
+}
+function comparableSpeechText(value2) {
+  return String(value2 ?? "").toLowerCase().replace(/[^a-z0-9]+/g, " ").trim();
+}
+function comparableTokens(value2) {
+  return comparableSpeechText(value2).split(/\s+/).filter(Boolean);
+}
+function jaccardSimilarity(aTokens, bTokens) {
+  const a2 = new Set(aTokens);
+  const b = new Set(bTokens);
+  if (!a2.size || !b.size) return 0;
+  let overlap = 0;
+  for (const token of a2) if (b.has(token)) overlap++;
+  return overlap / Math.max(1, Math.min(a2.size, b.size));
+}
+function isMostlyContainedInOrder(needleTokens, haystackTokens) {
+  if (!needleTokens.length || !haystackTokens.length) return false;
+  let index = 0;
+  for (const token of haystackTokens) {
+    if (token === needleTokens[index]) index++;
+    if (index >= needleTokens.length) return true;
+  }
+  return false;
+}
+function isLikelyAgentSpeechEcho({
+  heardText,
+  agentText,
+  elapsedMs: elapsedMs2
+}) {
+  const heard = compactSpeechText(heardText, 1200);
+  if (!heard) return { likelyEcho: false, echoSimilarity: 0, reason: "empty" };
+  if (!Number.isFinite(elapsedMs2) || elapsedMs2 > AGENT_ECHO_WINDOW_MS) {
+    return { likelyEcho: false, echoSimilarity: 0, reason: "outside_window" };
+  }
+  const heardTokens = comparableTokens(heard);
+  const agentTokens = comparableTokens(agentText);
+  if (!agentTokens.length) return { likelyEcho: false, echoSimilarity: 0, reason: "no_agent_speech" };
+  const agentComparable = agentTokens.join(" ");
+  const heardComparable = heardTokens.join(" ");
+  const echoSimilarity = jaccardSimilarity(heardTokens, agentTokens);
+  const exactEcho = Boolean(heardComparable && agentComparable && (agentComparable.includes(heardComparable) || heardComparable.includes(agentComparable)));
+  const orderedEcho = heardTokens.length >= 3 && isMostlyContainedInOrder(heardTokens, agentTokens);
+  const likelyEcho = echoSimilarity >= AGENT_ECHO_SIMILARITY || exactEcho || orderedEcho;
+  return {
+    likelyEcho,
+    echoSimilarity,
+    reason: likelyEcho ? "speaker_echo" : "pass"
+  };
+}
+function clipVoiceReply(text2, maxLength = MAX_VOICE_REPLY_CHARS) {
+  const cleaned = String(text2 || "").replace(/\s+/g, " ").replace(/^[`"']+|[`"']+$/g, "").trim();
+  return cleaned ? cleaned.slice(0, maxLength) : "";
+}
+function tryParseJsonObject(value2) {
+  const text2 = value2.trim();
+  if (!text2.startsWith("{") || !text2.endsWith("}")) return null;
+  try {
+    const parsed = JSON.parse(text2);
+    return parsed && typeof parsed === "object" && !Array.isArray(parsed) ? parsed : null;
+  } catch {
+    return null;
+  }
+}
+function extractVoiceModelReply(value2) {
+  const cleaned = String(value2 ?? "").replace(/^\s*```(?:json)?/i, "").replace(/```\s*$/i, "").trim();
+  const start2 = cleaned.indexOf("{");
+  const end = cleaned.lastIndexOf("}");
+  const embedded = tryParseJsonObject(cleaned) || (start2 >= 0 && end > start2 ? tryParseJsonObject(cleaned.slice(start2, end + 1)) : null);
+  if (embedded && typeof embedded["text"] === "string") {
+    const actionRaw = String(embedded["action"] ?? "speak").toLowerCase();
+    const action = actionRaw === "hangup" ? "hangup" : actionRaw === "silent" || actionRaw === "none" ? "silent" : "speak";
+    return {
+      action,
+      text: clipVoiceReply(embedded["text"]),
+      structured: true,
+      internalReason: typeof embedded["internal_reason"] === "string" ? clipVoiceReply(embedded["internal_reason"], 500) : typeof embedded["internalReason"] === "string" ? clipVoiceReply(embedded["internalReason"], 500) : null
+    };
+  }
+  const text2 = cleaned.replace(/^\s*```(?:[a-z]+)?\s*\n?/i, "").replace(/\n?```\s*$/i, "").replace(/^(?:agent|assistant|ai|voice|speaker|reply)\s*:\s*/i, "").trim();
+  return {
+    action: text2 ? "speak" : "silent",
+    text: clipVoiceReply(text2),
+    structured: false,
+    internalReason: null
+  };
+}
+function isContextSnapshotTurn(turn) {
+  return turn.role === "system" && turn.content.startsWith(CONTEXT_SNAPSHOT_PREFIX);
+}
+function isPinnedSystemTurn(turn) {
+  return turn.role === "system" && (turn.content === SYSTEM_PROMPT2 || turn.content.startsWith("Available tools (emit one-line JSON:"));
+}
+function buildRealtimeVoiceMessages(turns, maxMessages = MAX_REALTIME_MODEL_MESSAGES) {
+  const pinned = [];
+  const dynamic = [];
+  let latestSnapshot = null;
+  for (const turn of turns) {
+    if (isPinnedSystemTurn(turn) && !pinned.some((p2) => p2.content === turn.content)) {
+      pinned.push(turn);
+    } else if (isContextSnapshotTurn(turn)) {
+      latestSnapshot = turn;
+    } else {
+      dynamic.push(turn);
+    }
+  }
+  if (latestSnapshot) dynamic.push(latestSnapshot);
+  const available = Math.max(0, maxMessages - pinned.length);
+  return [...pinned, ...dynamic.slice(-available)];
+}
+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({
   "packages/cli/src/tui/voicechat.ts"() {
     "use strict";
     VAD_SILENCE_MS = 3e3;
     MAX_SEGMENT_MS = 6500;
+    SUMMARY_INJECTION_INTERVAL = 4;
     MAX_CONTEXT_TURNS = 20;
+    MAX_REALTIME_MODEL_MESSAGES = 16;
+    CONTEXT_SNAPSHOT_PREFIX = "Context snapshot (read-only):";
+    MAX_VOICE_REPLY_CHARS = 700;
+    AGENT_ECHO_WINDOW_MS = 12e3;
+    AGENT_ECHO_SIMILARITY = 0.72;
     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:
 - 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.`;
+- Prefer tools for factual queries; otherwise, answer directly with a short reply.
+- Internal notes must never be spoken. If you need an internal reason, return compact JSON: {"action":"speak","text":"exact words to speak aloud","internal_reason":"brief nonspoken reason"}.`;
     MIN_SIGNAL_SCORE = 0.4;
     NOISE_ONLY_RE = /^(?:[.·…\s,;:!?\-–—_()\[\]{}"'`]+|(?:uh|um|erm|hmm|mm+|uhh+|umm+)[\s.!?]*)+$/i;
     VoiceChatSession = class extends EventEmitter12 {
@@ -696008,6 +696115,8 @@ Rules:
       silenceTimer = null;
       maxSegmentTimer = null;
       lastSignalScore = null;
+      listenPausedForResponse = false;
+      lastAgentSpeech = null;
       // Abort control for inference
       abortController = null;
       // Callbacks
@@ -696083,6 +696192,10 @@ Rules:
           this.context.push({ role: "system", content: this.toolCatalogNote });
         }
         this.turnCount = 0;
+        this.voiceTranscript = [];
+        this.relayTranscript = [];
+        this.lastAgentSpeech = null;
+        this.listenPausedForResponse = false;
         if (this.verbose) this.onStatus("VoiceChat active — LISTENING");
         this._onTranscript = (...args) => {
           let text2;
@@ -696172,6 +696285,12 @@ Rules:
         if (this._state !== "LISTENING" && this._state !== "CAPTURING") {
           return;
         }
+        const echo = this.classifyAgentEcho(text2);
+        if (echo.likelyEcho) {
+          if (this.debugSnr) this.onStatus(`Ignoring likely agent-speaker echo (${echo.echoSimilarity.toFixed(2)}): ${truncateForLog(text2, 48)}`);
+          this.emit("voiceEchoFiltered", echo);
+          return;
+        }
         if (this._state === "LISTENING") {
           this.setState("CAPTURING");
           this.captureBuffer = "";
@@ -696222,6 +696341,7 @@ Rules:
           return;
         }
         this.setState("TRANSCRIBING");
+        this.pauseListenForResponse();
         this.onUserSpeech(text2);
         this.voiceTranscript.push({ role: "user", content: text2, ts: Date.now() });
         this.context.push({ role: "user", content: text2 });
@@ -696232,9 +696352,7 @@ Rules:
           } catch {
           }
         }
-        while (this.context.length > MAX_CONTEXT_TURNS + 1) {
-          this.context.splice(1, 1);
-        }
+        this.trimContext();
         this.think();
       }
       // ---------------------------------------------------------------------------
@@ -696245,18 +696363,20 @@ Rules:
         this.setState("THINKING");
         if (this.verbose) this.onStatus("Thinking...");
         this.abortController = new AbortController();
+        const lastUser = this.latestUserText();
         try {
           if (this.toolRelay?.contextSnapshot) {
             try {
               const snap = await Promise.resolve(this.toolRelay.contextSnapshot());
               if (snap && snap.trim()) {
+                this.dropTransientSystemTurns();
                 this.context.push({ role: "system", content: `Context snapshot (read-only):
 ${snap.trim()}` });
+                this.trimContext();
               }
             } catch {
             }
           }
-          const lastUser = [...this.context].reverse().find((m2) => m2.role === "user")?.content || "";
           let preAnswered = false;
           if (this.heuristicsEnabled && this.toolRelay && lastUser) {
             const lower = lastUser.toLowerCase();
@@ -696310,6 +696430,7 @@ ${out}` });
             }
             this.context.push({ role: "system", content: `Tool ${name10} result (authoritative):
 ${toolOutput}` });
+            this.trimContext();
           }
           if (!this.active) return;
           if (this.heuristicsEnabled && this.toolRelay && /\b(can't|cannot)\b/i.test(response) && this.toolCatalogNote) {
@@ -696317,18 +696438,25 @@ ${toolOutput}` });
             response = await this.streamOllamaInference(this.abortController.signal);
           }
           if (response.trim()) {
-            const finalSpoken = stripToolJsonLines(response.trim());
+            const reply = extractVoiceModelReply(stripToolJsonLines(response.trim()));
+            const finalSpoken = reply.text;
+            if (!finalSpoken || reply.action === "silent") {
+              return;
+            }
+            if (typeof this.voice.waitUntilIdle === "function") {
+              try {
+                await this.voice.waitUntilIdle();
+              } catch {
+              }
+            }
             this.context.push({ role: "assistant", content: finalSpoken });
+            this.trimContext();
             this.setState("SPEAKING");
             this.onAgentSpeech(finalSpoken);
-            try {
-              this.listen.pause();
-            } catch {
-            }
+            this.lastAgentSpeech = { text: finalSpoken, startedAt: Date.now() };
             this.voice.speak(finalSpoken);
             this.voiceTranscript.push({ role: "assistant", content: finalSpoken, ts: Date.now() });
-            this.voiceTranscript.push({ role: "assistant", content: response.trim(), ts: Date.now() });
-            if (this.runner) {
+            if (this.runner && this.turnCount % SUMMARY_INJECTION_INTERVAL === 0) {
               this.injectSummary();
             }
             if (typeof this.voice.waitUntilIdle === "function") {
@@ -696337,9 +696465,12 @@ ${toolOutput}` });
               } catch {
               }
             } else {
-              const estimatedMs = Math.max(1500, response.length / 5 * (6e4 / 150));
+              const estimatedMs = Math.max(1500, finalSpoken.length / 5 * (6e4 / 150));
               await new Promise((r2) => setTimeout(r2, estimatedMs));
             }
+            if (reply.action === "hangup") {
+              this.active = false;
+            }
           }
         } catch (err) {
           if (!this.active) return;
@@ -696351,14 +696482,44 @@ ${toolOutput}` });
           this.abortController = null;
         }
         if (this.active) {
-          try {
-            await this.listen.resume();
-          } catch {
-          }
+          await this.resumeListenAfterResponse();
           this.setState("LISTENING");
           if (this.verbose) this.onStatus("LISTENING...");
         }
       }
+      latestUserText() {
+        return [...this.context].reverse().find((m2) => m2.role === "user")?.content || "";
+      }
+      pauseListenForResponse() {
+        if (this.listenPausedForResponse) return;
+        try {
+          this.listen.pause();
+        } catch {
+        }
+        this.listenPausedForResponse = true;
+      }
+      async resumeListenAfterResponse() {
+        if (!this.listenPausedForResponse) return;
+        try {
+          await this.listen.resume();
+        } catch {
+        }
+        this.listenPausedForResponse = false;
+      }
+      classifyAgentEcho(text2) {
+        const agentText = this.lastAgentSpeech?.text || "";
+        const elapsedMs2 = this.lastAgentSpeech ? Date.now() - this.lastAgentSpeech.startedAt : Number.POSITIVE_INFINITY;
+        return isLikelyAgentSpeechEcho({ heardText: text2, agentText, elapsedMs: elapsedMs2 });
+      }
+      dropTransientSystemTurns() {
+        this.context = this.context.filter((turn) => !isContextSnapshotTurn(turn));
+      }
+      trimContext() {
+        this.context = buildRealtimeVoiceMessages(this.context, MAX_CONTEXT_TURNS + 2);
+      }
+      buildRealtimeMessagesForRequest() {
+        return buildRealtimeVoiceMessages(this.context, MAX_REALTIME_MODEL_MESSAGES);
+      }
       /**
        * Stream inference. Tries native Ollama /api/chat first (supports think:false
        * for reasoning models), falls back to OpenAI-compat /v1/chat/completions.
@@ -696370,11 +696531,11 @@ ${toolOutput}` });
         try {
           const nativeBody = JSON.stringify({
             model: this.model,
-            messages: this.context,
+            messages: this.buildRealtimeMessagesForRequest(),
             stream: true,
             think: false,
             // Disable reasoning — voice chat needs fast, direct responses
-            options: { temperature: 0.7, num_predict: 256 }
+            options: { temperature: 0.5, num_predict: 192 }
           });
           const res2 = await fetch(`${baseUrl2}/api/chat`, {
             method: "POST",
@@ -696391,10 +696552,10 @@ ${toolOutput}` });
         }
         const openaiBody = JSON.stringify({
           model: this.model,
-          messages: this.context,
+          messages: this.buildRealtimeMessagesForRequest(),
           stream: true,
-          temperature: 0.7,
-          max_tokens: 1024
+          temperature: 0.5,
+          max_tokens: 384
         });
         const endpoint = baseUrl2.includes("/v1") ? `${baseUrl2}/chat/completions` : `${baseUrl2}/v1/chat/completions`;
         const res = await fetch(endpoint, { method: "POST", headers, body: openaiBody, signal });

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.413",
+  "version": "1.0.414",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.413",
+      "version": "1.0.414",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.413",
+  "version": "1.0.414",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",