omnius 1.0.412 → 1.0.414

package/dist/index.js

@@ -41412,7 +41412,7 @@ ${content}`
 // packages/execution/dist/tools/transcribe-tool.js
 import { existsSync as existsSync38, mkdirSync as mkdirSync22, writeFileSync as writeFileSync21, readFileSync as readFileSync30, unlinkSync as unlinkSync5, readdirSync as readdirSync14 } from "node:fs";
 import { join as join40, basename as basename7, extname as extname4, resolve as resolve21 } from "node:path";
-import { homedir as homedir11 } from "node:os";
+import { homedir as homedir11, tmpdir as tmpdir5 } from "node:os";
 function transcriptionPythonEnv(extra = {}) {
   const env2 = { ...process.env, ...extra };
   applyMediaCudaDeviceFilterToEnv(env2, "asr");
@@ -41639,6 +41639,11 @@ var init_transcribe_tool = __esm({
           diarize: {
             type: "boolean",
             description: "Enable speaker diarization (identify who said what). Default: false"
+          },
+          backend: {
+            type: "string",
+            description: "ASR backend preference. auto tries fast transcribe-cli first; managed-whisper uses Omnius' isolated Whisper venv directly.",
+            enum: ["auto", "managed-whisper", "transcribe-cli"]
           }
         },
         required: ["path"]
@@ -41652,6 +41657,7 @@ var init_transcribe_tool = __esm({
         const filePath = resolve21(this.workingDir, String(args["path"] ?? ""));
         let model = String(args["model"] ?? "base");
         const diarize = Boolean(args["diarize"] ?? false);
+        const backend = String(args["backend"] ?? "auto").toLowerCase();
         if (!existsSync38(filePath)) {
           return {
             success: false,
@@ -41694,6 +41700,17 @@ var init_transcribe_tool = __esm({
         }
         if (effectiveModel !== askedModel)
           model = effectiveModel;
+        if (backend === "managed-whisper") {
+          const managed2 = await this.execViaManagedWhisper(filePath, model, start2);
+          if (managed2.success)
+            return managed2;
+          return {
+            success: false,
+            output: "",
+            error: `Managed Whisper ASR failed. ${String(managed2.error || managed2.output || "unknown error").slice(0, 1200)}`,
+            durationMs: performance.now() - start2
+          };
+        }
         const failures = [];
         let managedPython = "";
         try {
@@ -41704,7 +41721,7 @@ var init_transcribe_tool = __esm({
           failures.push(`managed ASR runtime bootstrap: ${err instanceof Error ? err.message : String(err)}`);
         }
         const tc = await loadTranscribeCli() ?? await ensureManagedTranscribeCliRuntime();
-        if (tc) {
+        if (tc && backend !== "managed-whisper") {
           try {
             const result = await withProcessEnv(transcriptionPythonEnv(), () => tc.transcribe(filePath, {
               model,
@@ -41894,8 +41911,14 @@ var init_transcribe_tool = __esm({
         }
       }
       async execViaManagedWhisper(filePath, model, start2) {
-        const scriptPath2 = join40(this.workingDir, ".omnius", "tmp", `managed-whisper-${Date.now()}.py`);
-        mkdirSync22(join40(this.workingDir, ".omnius", "tmp"), { recursive: true });
+        let scriptDir = join40(this.workingDir, ".omnius", "tmp");
+        try {
+          mkdirSync22(scriptDir, { recursive: true });
+        } catch {
+          scriptDir = join40(tmpdir5(), "omnius-asr");
+          mkdirSync22(scriptDir, { recursive: true });
+        }
+        const scriptPath2 = join40(scriptDir, `managed-whisper-${Date.now()}-${Math.random().toString(36).slice(2, 8)}.py`);
         const script = `
 import json, os, warnings
 warnings.filterwarnings("ignore")
@@ -41904,13 +41927,15 @@ audio_file = ${JSON.stringify(filePath)}
 model_name = ${JSON.stringify(model)}
 try:
     import whisper
-    device = "cpu"
-    try:
-        import torch
-        if torch.cuda.is_available():
-            device = "cuda"
-    except Exception:
-        pass
+    device = os.environ.get("OMNIUS_ASR_DEVICE", "cpu").strip().lower() or "cpu"
+    if device == "auto":
+        device = "cpu"
+        try:
+            import torch
+            if torch.cuda.is_available():
+                device = "cuda"
+        except Exception:
+            pass
     m = whisper.load_model(model_name, device=device)
     result = m.transcribe(audio_file, fp16=(device == "cuda"), condition_on_previous_text=False)
     segments = []
@@ -42491,7 +42516,7 @@ var init_structured_file = __esm({
 import { spawn as spawn7 } from "node:child_process";
 import { writeFile as writeFile9, mkdtemp, rm as rm2, readdir as readdir4, stat as stat4 } from "node:fs/promises";
 import { join as join41 } from "node:path";
-import { tmpdir as tmpdir5 } from "node:os";
+import { tmpdir as tmpdir6 } from "node:os";
 function runProcess(cmd, args, options2) {
   return new Promise((resolve76) => {
     const proc = spawn7(cmd, args, {
@@ -42689,7 +42714,7 @@ ${result.filesCreated.join("\n")}`);
       // Subprocess mode — temp directory + separate process
       // -------------------------------------------------------------------------
       async #runSubprocess(code8, langConfig, timeoutMs, stdin) {
-        const sandboxDir = await mkdtemp(join41(tmpdir5(), "omnius-sandbox-"));
+        const sandboxDir = await mkdtemp(join41(tmpdir6(), "omnius-sandbox-"));
         try {
           const scriptFile = join41(sandboxDir, `_sandbox_script${langConfig.ext}`);
           await writeFile9(scriptFile, code8, "utf-8");
@@ -42716,7 +42741,7 @@ ${result.filesCreated.join("\n")}`);
           bash: "bash:5"
         };
         const image = images[language] ?? "node:22-slim";
-        const sandboxDir = await mkdtemp(join41(tmpdir5(), "omnius-docker-sandbox-"));
+        const sandboxDir = await mkdtemp(join41(tmpdir6(), "omnius-docker-sandbox-"));
         try {
           const scriptFile = `_sandbox_script${langConfig.ext}`;
           await writeFile9(join41(sandboxDir, scriptFile), code8, "utf-8");
@@ -274241,7 +274266,7 @@ New: ${newNarrative.slice(0, 200)}...`,
 import { spawn as spawn9 } from "node:child_process";
 import { createServer as createServer2 } from "node:net";
 import { join as join44 } from "node:path";
-import { tmpdir as tmpdir6 } from "node:os";
+import { tmpdir as tmpdir7 } from "node:os";
 import { randomBytes as randomBytes11 } from "node:crypto";
 import { unlinkSync as unlinkSync6 } from "node:fs";
 var ReplTool;
@@ -274313,7 +274338,7 @@ var init_repl = __esm({
         if (!this.proc || this.proc.killed || this.proc.exitCode !== null) {
           await this.startProcess();
         }
-        const tempFile = join44(tmpdir6(), `omnius-repl-ctx-${randomBytes11(6).toString("hex")}.txt`);
+        const tempFile = join44(tmpdir7(), `omnius-repl-ctx-${randomBytes11(6).toString("hex")}.txt`);
         const { writeFileSync: writeFs, unlinkSync: unlinkFs } = await import("node:fs");
         writeFs(tempFile, content, "utf8");
         const result = await this.executeCode(`with open(${JSON.stringify(tempFile)}, "r") as _f:
@@ -274540,7 +274565,7 @@ print("__OMNIUS_REPL_READY__")
         if (this.ipcServer)
           return;
         const sockId = randomBytes11(8).toString("hex");
-        this.ipcPath = join44(tmpdir6(), `omnius-repl-ipc-${sockId}.sock`);
+        this.ipcPath = join44(tmpdir7(), `omnius-repl-ipc-${sockId}.sock`);
         return new Promise((resolve76, reject) => {
           this.ipcServer = createServer2((conn) => {
             let buffer2 = new Uint8Array(0);
@@ -287719,7 +287744,7 @@ ${parts.join("\n\n")}`,
 // packages/execution/dist/tools/desktop-click.js
 import { readFileSync as readFileSync35 } from "node:fs";
 import { execSync as execSync16 } from "node:child_process";
-import { tmpdir as tmpdir7 } from "node:os";
+import { tmpdir as tmpdir8 } from "node:os";
 import { join as join57 } from "node:path";
 function getImageDimensions2(filePath) {
   try {
@@ -287936,7 +287961,7 @@ Button: ${button}, Type: ${clickType}`,
               durationMs: performance.now() - start2
             };
           }
-          const screenshotPath = join57(tmpdir7(), `omnius-desktop-click-${Date.now()}.png`);
+          const screenshotPath = join57(tmpdir8(), `omnius-desktop-click-${Date.now()}.png`);
           const screenshotBackend = captureDesktopScreenshot(screenshotPath);
           const dims = getImageDimensions2(screenshotPath);
           if (!dims) {
@@ -288103,7 +288128,7 @@ Screenshot: ${screenshotPath}`,
           if (delayMs > 0) {
             await new Promise((r2) => setTimeout(r2, delayMs));
           }
-          const screenshotPath = join57(tmpdir7(), `omnius-desktop-describe-${Date.now()}.png`);
+          const screenshotPath = join57(tmpdir8(), `omnius-desktop-describe-${Date.now()}.png`);
           const screenshotBackend = captureDesktopScreenshot(screenshotPath);
           const dims = getImageDimensions2(screenshotPath);
           const imageBuffer = readFileSync35(screenshotPath);
@@ -289307,7 +289332,7 @@ Language: ${language}
 import { existsSync as existsSync50, statSync as statSync21, readFileSync as readFileSync37, unlinkSync as unlinkSync8 } from "node:fs";
 import { resolve as resolve29, basename as basename10, join as join59 } from "node:path";
 import { execSync as execSync18 } from "node:child_process";
-import { tmpdir as tmpdir8 } from "node:os";
+import { tmpdir as tmpdir9 } from "node:os";
 var PdfToTextTool;
 var init_pdf_to_text = __esm({
   "packages/execution/dist/tools/pdf-to-text.js"() {
@@ -289459,7 +289484,7 @@ ${text2}`,
         if (!ocrCheck.available || !tesCheck.available || !gsCheck.available) {
           return null;
         }
-        const tmpPdf = join59(tmpdir8(), `omnius-ocr-${Date.now()}.pdf`);
+        const tmpPdf = join59(tmpdir9(), `omnius-ocr-${Date.now()}.pdf`);
         try {
           const ocrCmd = `ocrmypdf -l ${language} --skip-text ${JSON.stringify(inputPath)} ${JSON.stringify(tmpPdf)}`;
           execSync18(ocrCmd, { stdio: "pipe", timeout: 6e5 });
@@ -289493,7 +289518,7 @@ import { existsSync as existsSync51, mkdirSync as mkdirSync28, statSync as statS
 import { resolve as resolve30, basename as basename11, dirname as dirname16, join as join60 } from "node:path";
 import { execSync as execSync19 } from "node:child_process";
 import { fileURLToPath as fileURLToPath6 } from "node:url";
-import { homedir as homedir13, tmpdir as tmpdir9 } from "node:os";
+import { homedir as homedir13, tmpdir as tmpdir10 } from "node:os";
 function findOcrScript() {
   const thisDir = dirname16(fileURLToPath6(import.meta.url));
   const devPath3 = resolve30(thisDir, "../../scripts/ocr-advanced.py");
@@ -289726,7 +289751,7 @@ var init_ocr_image_advanced = __esm({
           cmdParts.push("--output-dir", JSON.stringify(resolve30(this.workingDir, outputDir2)));
         let debugDir;
         if (debug) {
-          debugDir = join60(tmpdir9(), `omnius-ocr-debug-${Date.now()}`);
+          debugDir = join60(tmpdir10(), `omnius-ocr-debug-${Date.now()}`);
           cmdParts.push("--debug-dir", debugDir);
         }
         try {
@@ -543195,7 +543220,7 @@ var init_process_health = __esm({
 // packages/execution/dist/tools/camera-capture.js
 import { accessSync, constants, readFileSync as readFileSync44, writeFileSync as writeFileSync28, unlinkSync as unlinkSync11, existsSync as existsSync60, mkdirSync as mkdirSync34, readdirSync as readdirSync21, renameSync as renameSync7 } from "node:fs";
 import { join as join75 } from "node:path";
-import { tmpdir as tmpdir10 } from "node:os";
+import { tmpdir as tmpdir11 } from "node:os";
 function normalizeCameraRotationDegrees(value2) {
   if (value2 === void 0 || value2 === null || value2 === "")
     return 0;
@@ -543693,7 +543718,7 @@ ${formatRouterPlan(plan)}`,
         const height = args["height"] || 720;
         const outputPath3 = args["output_path"];
         const rotation = normalizeCameraRotationDegrees(args["rotate_degrees"] ?? args["rotation_degrees"] ?? args["rotate"]);
-        const captureDir = join75(tmpdir10(), "omnius-camera");
+        const captureDir = join75(tmpdir11(), "omnius-camera");
         if (!existsSync60(captureDir))
           mkdirSync34(captureDir, { recursive: true });
         const tempFile = outputPath3 || join75(captureDir, `capture-${Date.now()}.jpg`);
@@ -544075,7 +544100,7 @@ Auto-installed camera dependencies: ${[
         } catch (err) {
           return { success: false, output: "", error: `QooCam OSC error: ${err instanceof Error ? err.message : String(err)}`, durationMs: performance.now() - start2 };
         }
-        const captureDir = join75(tmpdir10(), "omnius-camera");
+        const captureDir = join75(tmpdir11(), "omnius-camera");
         if (!existsSync60(captureDir))
           mkdirSync34(captureDir, { recursive: true });
         const tempFile = outputPath3 || join75(captureDir, `qoocam-${Date.now()}.jpg`);
@@ -544225,7 +544250,7 @@ Saved to: ${outputPath3}`;
 import { execSync as execSync26 } from "node:child_process";
 import { readFileSync as readFileSync45, unlinkSync as unlinkSync12, existsSync as existsSync61, mkdirSync as mkdirSync35, statSync as statSync25 } from "node:fs";
 import { join as join76 } from "node:path";
-import { tmpdir as tmpdir11 } from "node:os";
+import { tmpdir as tmpdir12 } from "node:os";
 var AudioCaptureTool;
 var init_audio_capture = __esm({
   "packages/execution/dist/tools/audio-capture.js"() {
@@ -544331,7 +544356,7 @@ ${devices.join("\n")}`,
         const channels = args["channels"] || 1;
         const format3 = args["format"] || "wav";
         const outputPath3 = args["output_path"];
-        const captureDir = join76(tmpdir11(), "omnius-audio");
+        const captureDir = join76(tmpdir12(), "omnius-audio");
         if (!existsSync61(captureDir))
           mkdirSync35(captureDir, { recursive: true });
         const ext = format3 === "mp3" ? "mp3" : "wav";
@@ -544418,7 +544443,7 @@ Saved to: ${tempFile}`,
       }
       checkLevel(args, start2) {
         const device = args["device"] || "default";
-        const tempFile = join76(tmpdir11(), `omnius-level-${Date.now()}.raw`);
+        const tempFile = join76(tmpdir12(), `omnius-level-${Date.now()}.raw`);
         try {
           execSync26(`arecord -D ${device} -f S16_LE -r 16000 -c 1 -d 1 -t raw -q ${tempFile}`, { timeout: 5e3, stdio: "pipe" });
           if (!existsSync61(tempFile)) {
@@ -544476,7 +544501,7 @@ Saved to: ${tempFile}`,
 import { execFileSync as execFileSync2, execSync as execSync27, spawn as spawn18 } from "node:child_process";
 import { copyFileSync as copyFileSync4, existsSync as existsSync62, statSync as statSync26, writeFileSync as writeFileSync29, mkdirSync as mkdirSync36, readdirSync as readdirSync22, writeSync, rmSync as rmSync7 } from "node:fs";
 import { basename as basename15, dirname as dirname21, extname as extname12, isAbsolute as isAbsolute4, join as join77, resolve as resolve39 } from "node:path";
-import { homedir as homedir17, tmpdir as tmpdir12 } from "node:os";
+import { homedir as homedir17, tmpdir as tmpdir13 } from "node:os";
 function ttsPythonEnv(extra = {}) {
   const env2 = { ...process.env, ...extra };
   applyMediaCudaDeviceFilterToEnv(env2, "tts");
@@ -545357,7 +545382,7 @@ function ensureLuxttsDaemon() {
     }, 12e4);
     const daemon = spawn18(venvPy, [inferScript], {
       stdio: ["pipe", "pipe", "pipe"],
-      cwd: tmpdir12(),
+      cwd: tmpdir13(),
       env: ttsPythonEnv({ LUXTTS_REPO_PATH: repoDir })
     });
     _luxttsDaemon = daemon;
@@ -545414,7 +545439,7 @@ function luxttsSynthesize(text2, cloneRef, outputPath3, speed = 1) {
       return;
     }
     const id2 = `tts-${++_luxttsRequestId}`;
-    const outPath = outputPath3 || join77(tmpdir12(), `omnius-luxtts-${Date.now()}.wav`);
+    const outPath = outputPath3 || join77(tmpdir13(), `omnius-luxtts-${Date.now()}.wav`);
     const timeout2 = setTimeout(() => {
       _luxttsPending.delete(id2);
       reject(new Error("LuxTTS synthesis timeout"));
@@ -545557,7 +545582,7 @@ function ensureMisottsDaemon() {
     }, 12e4);
     const daemon = spawn18(venvPy, [inferScript], {
       stdio: ["pipe", "pipe", "pipe"],
-      cwd: tmpdir12(),
+      cwd: tmpdir13(),
       env: ttsPythonEnv({ MISO_TTS_REPO_PATH: repoDir, NO_TORCH_COMPILE: "1" })
     });
     _misottsDaemon = daemon;
@@ -545614,7 +545639,7 @@ function misottsSynthesize(text2, cloneRef, outputPath3, maxAudioLengthMs = 3e4)
       return;
     }
     const id2 = `tts-${++_misottsRequestId}`;
-    const outPath = outputPath3 || join77(tmpdir12(), `omnius-misotts-${Date.now()}.wav`);
+    const outPath = outputPath3 || join77(tmpdir13(), `omnius-misotts-${Date.now()}.wav`);
     const timeout2 = setTimeout(() => {
       _misottsPending.delete(id2);
       reject(new Error("MisoTTS synthesis timeout"));
@@ -546272,7 +546297,7 @@ ${tried.map((line) => `- ${line}`).join("\n")}`,
         ], {
           stdio: "pipe",
           timeout: 18e4,
-          cwd: tmpdir12()
+          cwd: tmpdir13()
         });
         return `${voice} (${model})`;
       }
@@ -547069,7 +547094,7 @@ ${info}`, durationMs: performance.now() - start2 };
 import { execSync as execSync30 } from "node:child_process";
 import { readFileSync as readFileSync46, unlinkSync as unlinkSync13, existsSync as existsSync63, mkdirSync as mkdirSync37, statSync as statSync27 } from "node:fs";
 import { join as join78 } from "node:path";
-import { tmpdir as tmpdir13 } from "node:os";
+import { tmpdir as tmpdir14 } from "node:os";
 var SdrScanTool;
 var init_sdr_scan = __esm({
   "packages/execution/dist/tools/sdr-scan.js"() {
@@ -547240,7 +547265,7 @@ Tools installed but rtl_test failed — device may be in use by another process.
         if (!await this.ensureSdrTools()) {
           return { success: false, output: "", error: "rtl-sdr tools not installed. Connect an RTL-SDR dongle and retry (will prompt for install).", durationMs: performance.now() - start2 };
         }
-        const captureDir = join78(tmpdir13(), "omnius-sdr");
+        const captureDir = join78(tmpdir14(), "omnius-sdr");
         if (!existsSync63(captureDir))
           mkdirSync37(captureDir, { recursive: true });
         const outFile = join78(captureDir, `scan-${Date.now()}.csv`);
@@ -547324,7 +547349,7 @@ ${output.slice(0, 2e3)}`,
         } catch {
           return { success: false, output: "", error: "rtl_fm not installed. Run: sudo apt install rtl-sdr", durationMs: performance.now() - start2 };
         }
-        const captureDir = join78(tmpdir13(), "omnius-sdr");
+        const captureDir = join78(tmpdir14(), "omnius-sdr");
         if (!existsSync63(captureDir))
           mkdirSync37(captureDir, { recursive: true });
         const outFile = join78(captureDir, `fm-${Date.now()}.wav`);
@@ -547781,7 +547806,7 @@ ${result.trim()}`, durationMs: performance.now() - start2 };
 // packages/execution/dist/tools/audio-analyze.js
 import { existsSync as existsSync65, mkdirSync as mkdirSync38, writeFileSync as writeFileSync30 } from "node:fs";
 import { basename as basename16, isAbsolute as isAbsolute5, join as join79, resolve as resolve40 } from "node:path";
-import { homedir as homedir18, tmpdir as tmpdir14 } from "node:os";
+import { homedir as homedir18, tmpdir as tmpdir15 } from "node:os";
 import { fileURLToPath as fileURLToPath9 } from "node:url";
 function audioAnalysisPythonEnv(extra = {}) {
   const env2 = { ...process.env, ...extra };
@@ -548212,7 +548237,7 @@ except Exception as e:
         const contextDir = join79(homedir18(), ".omnius", "audio-context");
         if (!existsSync65(contextDir))
           mkdirSync38(contextDir, { recursive: true });
-        const audioFile = join79(tmpdir14(), `omnius-listen-${Date.now()}.wav`);
+        const audioFile = join79(tmpdir15(), `omnius-listen-${Date.now()}.wav`);
         try {
           await execFileText("arecord", ["-D", "default", "-f", "S16_LE", "-r", "16000", "-c", "1", "-d", String(Math.min(duration, 60)), "-q", audioFile], {
             timeout: (duration + 10) * 1e3
@@ -548353,7 +548378,7 @@ ${e2.stderr ? String(e2.stderr) : ""}`.trim();
             return file;
         }
         const duration = args["duration"] || 5;
-        const outFile = join79(tmpdir14(), `omnius-analyze-${Date.now()}.wav`);
+        const outFile = join79(tmpdir15(), `omnius-analyze-${Date.now()}.wav`);
         try {
           await execFileText("arecord", ["-D", "default", "-f", "S16_LE", "-r", "16000", "-c", "1", "-d", String(duration), "-q", outFile], {
             timeout: (duration + 10) * 1e3
@@ -548382,7 +548407,7 @@ ${e2.stderr ? String(e2.stderr) : ""}`.trim();
       }
       /** Run a Python script in the venv and parse JSON output */
       async runPythonScript(script, start2, label) {
-        const scriptFile = join79(tmpdir14(), `omnius-audio-${Date.now()}.py`);
+        const scriptFile = join79(tmpdir15(), `omnius-audio-${Date.now()}.py`);
         writeFileSync30(scriptFile, script);
         try {
           const output = await execFileText(VENV_PYTHON, [scriptFile], {
@@ -548418,7 +548443,7 @@ ${output}`, durationMs: performance.now() - start2 };
 import { execSync as execSync33, spawnSync as spawnSync6 } from "node:child_process";
 import { existsSync as existsSync66, readFileSync as readFileSync47, writeFileSync as writeFileSync31, mkdirSync as mkdirSync39 } from "node:fs";
 import { join as join80 } from "node:path";
-import { tmpdir as tmpdir15, homedir as homedir19 } from "node:os";
+import { tmpdir as tmpdir16, homedir as homedir19 } from "node:os";
 var GPS_USB_IDS, GpsLocationTool;
 var init_gps_location = __esm({
   "packages/execution/dist/tools/gps-location.js"() {
@@ -548543,7 +548568,7 @@ var init_gps_location = __esm({
       /** Run a Python GPS script using pyserial+pynmea2 and return JSON result */
       async runGpsPython(script, timeoutMs = 3e4) {
         await this.ensureGpsVenv();
-        const scriptFile = join80(tmpdir15(), `omnius-gps-${Date.now()}.py`);
+        const scriptFile = join80(tmpdir16(), `omnius-gps-${Date.now()}.py`);
         writeFileSync31(scriptFile, script);
         try {
           const output = execSync33(`${this.GPS_PYTHON} ${scriptFile}`, {
@@ -549015,7 +549040,7 @@ Drift: ${drift != null ? drift + "ms" : "unknown"}`,
       // =========================================================================
       async recordTrack(args, start2) {
         const duration = args["duration"] || 60;
-        const outputPath3 = args["output_path"] || join80(tmpdir15(), `omnius-gps-track-${Date.now()}.gpx`);
+        const outputPath3 = args["output_path"] || join80(tmpdir16(), `omnius-gps-track-${Date.now()}.gpx`);
         if (!await this.ensureGpsd(args)) {
           return { success: false, output: "", error: "gpsd required for track recording. Connect a GPS device.", durationMs: performance.now() - start2 };
         }
@@ -549248,7 +549273,7 @@ def _omnius_normalized_features(features):
 import { execFile as execFile6 } from "node:child_process";
 import { existsSync as existsSync67, mkdirSync as mkdirSync40, writeFileSync as writeFileSync32, readFileSync as readFileSync48 } from "node:fs";
 import { join as join81 } from "node:path";
-import { homedir as homedir20, tmpdir as tmpdir16 } from "node:os";
+import { homedir as homedir20, tmpdir as tmpdir17 } from "node:os";
 function visualMemoryPythonEnv(extra = {}) {
   const env2 = { ...process.env, ...extra };
   applyMediaCudaDeviceFilterToEnv(env2, "vision");
@@ -550020,7 +550045,7 @@ ${objects.join("\n") || "  (none taught)"}`,
         }
       }
       async runVisionPython(script, timeoutMs = 6e4) {
-        const scriptFile = join81(tmpdir16(), `omnius-vmem-${Date.now()}.py`);
+        const scriptFile = join81(tmpdir17(), `omnius-vmem-${Date.now()}.py`);
         writeFileSync32(scriptFile, script);
         try {
           const { stdout: output } = await execFileText2(VENV_PY, [scriptFile], {
@@ -550053,7 +550078,7 @@ ${objects.join("\n") || "  (none taught)"}`,
 import { execSync as execSync34 } from "node:child_process";
 import { appendFileSync as appendFileSync5, existsSync as existsSync68, mkdirSync as mkdirSync41, writeFileSync as writeFileSync33, readFileSync as readFileSync49, readdirSync as readdirSync23 } from "node:fs";
 import { join as join82 } from "node:path";
-import { homedir as homedir21, tmpdir as tmpdir17 } from "node:os";
+import { homedir as homedir21, tmpdir as tmpdir18 } from "node:os";
 import { randomUUID as randomUUID15 } from "node:crypto";
 var MM_DIR, MM_INDEX, MultimodalMemoryTool;
 var init_multimodal_memory = __esm({
@@ -550172,7 +550197,7 @@ with torch.no_grad():
     features = _omnius_normalized_features(model.get_image_features(**inputs))
 print(json.dumps(features[0].cpu().numpy().tolist()))
 `;
-                const scriptFile = join82(tmpdir17(), `mm-clip-${Date.now()}.py`);
+                const scriptFile = join82(tmpdir18(), `mm-clip-${Date.now()}.py`);
                 writeFileSync33(scriptFile, clipScript);
                 const clipOutput = execSync34(`${venvPy} ${scriptFile}`, { encoding: "utf8", timeout: 12e4, env: { ...process.env, PYTHONUNBUFFERED: "1" } });
                 const embedding = JSON.parse(clipOutput.trim().split("\n").pop());
@@ -550194,7 +550219,7 @@ faces = app.get(img) if img is not None else []
 result = [{"confidence": float(f.det_score), "age": int(f.age) if hasattr(f, 'age') else None} for f in faces]
 print(json.dumps(result))
 `;
-                const scriptFile = join82(tmpdir17(), `mm-face-${Date.now()}.py`);
+                const scriptFile = join82(tmpdir18(), `mm-face-${Date.now()}.py`);
                 writeFileSync33(scriptFile, faceScript);
                 const faceOutput = execSync34(`${venvPy} ${scriptFile}`, { encoding: "utf8", timeout: 6e4, env: { ...process.env, PYTHONUNBUFFERED: "1" } });
                 const faces = JSON.parse(faceOutput.trim().split("\n").pop());
@@ -550246,7 +550271,7 @@ with open(model.class_map_path().numpy().decode()) as f:
 top=scores.numpy().mean(axis=0).argsort()[-1]
 print(classes[top])
 `;
-                const scriptFile = join82(tmpdir17(), `mm-yamnet-${Date.now()}.py`);
+                const scriptFile = join82(tmpdir18(), `mm-yamnet-${Date.now()}.py`);
                 writeFileSync33(scriptFile, classifyScript);
                 const soundClass = execSync34(`${mlVenvPy} ${scriptFile}`, { encoding: "utf8", timeout: 12e4 }).trim().split("\n").pop();
                 episode.audio.soundClass = soundClass;
@@ -550342,7 +550367,7 @@ if faces:
 else:
     print(json.dumps({"enrolled": False, "reason": "no face detected"}))
 `;
-                const scriptFile = join82(tmpdir17(), `mm-enroll-${Date.now()}.py`);
+                const scriptFile = join82(tmpdir18(), `mm-enroll-${Date.now()}.py`);
                 writeFileSync33(scriptFile, enrollScript);
                 const enrollOutput = execSync34(`${venvPy} ${scriptFile}`, { encoding: "utf8", timeout: 6e4, env: { ...process.env, PYTHONUNBUFFERED: "1" } });
                 const enrollResult = JSON.parse(enrollOutput.trim().split("\n").pop());
@@ -550397,7 +550422,7 @@ with torch.no_grad():
     features = _omnius_normalized_features(model.get_text_features(**inputs))
 print(json.dumps(features[0].cpu().numpy().tolist()))
 `;
-            const scriptFile = join82(tmpdir17(), `mm-clipq-${Date.now()}.py`);
+            const scriptFile = join82(tmpdir18(), `mm-clipq-${Date.now()}.py`);
             writeFileSync33(scriptFile, clipTextScript);
             const output = execSync34(`${venvPy} ${scriptFile}`, { encoding: "utf8", timeout: 6e4, env: { ...process.env, PYTHONUNBUFFERED: "1" } });
             queryClipEmbedding = JSON.parse(output.trim().split("\n").pop());
@@ -550679,7 +550704,7 @@ ${lines.join("\n")}`,
 import { execSync as execSync35, spawnSync as spawnSync7 } from "node:child_process";
 import { existsSync as existsSync69, mkdirSync as mkdirSync42, writeFileSync as writeFileSync34, readFileSync as readFileSync50, unlinkSync as unlinkSync14 } from "node:fs";
 import { dirname as dirname22, join as join83, resolve as resolve41 } from "node:path";
-import { tmpdir as tmpdir18, homedir as homedir22 } from "node:os";
+import { tmpdir as tmpdir19, homedir as homedir22 } from "node:os";
 import { fileURLToPath as fileURLToPath10 } from "node:url";
 function asrPythonEnv(extra = {}) {
   const env2 = { ...process.env, ...extra };
@@ -550794,7 +550819,7 @@ var init_asr_listen = __esm({
       listenAndTranscribe(args, start2) {
         const duration = args["duration"] || 8;
         const device = args["device"] || "default";
-        const captureDir = join83(tmpdir18(), "omnius-asr");
+        const captureDir = join83(tmpdir19(), "omnius-asr");
         if (!existsSync69(captureDir))
           mkdirSync42(captureDir, { recursive: true });
         const audioFile = join83(captureDir, `listen-${Date.now()}.wav`);
@@ -550941,7 +550966,7 @@ except Exception:
 
 print(json.dumps({"ok": False, "error": "No whisper backend available"}))
 `;
-        const scriptFile = join83(tmpdir18(), `omnius-asr-whisper-${Date.now()}.py`);
+        const scriptFile = join83(tmpdir19(), `omnius-asr-whisper-${Date.now()}.py`);
         writeFileSync34(scriptFile, whisperScript);
         const pyPaths = [
           join83(homedir22(), ".omnius", "venv", "bin", "python3"),
@@ -555719,7 +555744,7 @@ var init_visual_trigger = __esm({
 
 // packages/execution/dist/tools/live-media-loop.js
 import { existsSync as existsSync78, mkdirSync as mkdirSync47, readFileSync as readFileSync57, rmSync as rmSync10, writeFileSync as writeFileSync39 } from "node:fs";
-import { homedir as homedir28, tmpdir as tmpdir19 } from "node:os";
+import { homedir as homedir28, tmpdir as tmpdir20 } from "node:os";
 import { basename as basename19, isAbsolute as isAbsolute8, join as join92, resolve as resolve46 } from "node:path";
 function buildYolo26BootstrapPlan(model = DEFAULT_YOLO26_MODEL) {
   return {
@@ -556038,8 +556063,8 @@ async function ensureRuntime(autoBootstrap) {
   }
 }
 async function exportYolo26Model(python, model, options2, yoloeClasses = []) {
-  const cfgPath = join92(tmpdir19(), `omnius-yolo26-export-${Date.now()}.json`);
-  const scriptPath2 = join92(tmpdir19(), `omnius-yolo26-export-${Date.now()}.py`);
+  const cfgPath = join92(tmpdir20(), `omnius-yolo26-export-${Date.now()}.json`);
+  const scriptPath2 = join92(tmpdir20(), `omnius-yolo26-export-${Date.now()}.py`);
   writeFileSync39(cfgPath, JSON.stringify({ model, options: options2, yoloeClasses }, null, 2), "utf8");
   writeFileSync39(scriptPath2, String.raw`
 import json, sys
@@ -556615,8 +556640,8 @@ var init_live_media_loop = __esm({
             return { success: false, output: "", error: runtime.error, durationMs: performance.now() - start2 };
           const outDir = join92(this.workingDir, ".omnius", "live-media", `run-${Date.now().toString(36)}`);
           mkdirSync47(outDir, { recursive: true });
-          const cfgPath = join92(tmpdir19(), `omnius-live-media-${Date.now()}.json`);
-          const scriptPath2 = join92(tmpdir19(), `omnius-live-media-${Date.now()}.py`);
+          const cfgPath = join92(tmpdir20(), `omnius-live-media-${Date.now()}.json`);
+          const scriptPath2 = join92(tmpdir20(), `omnius-live-media-${Date.now()}.py`);
           writeFileSync39(cfgPath, JSON.stringify({
             source_kind: sourceKind(args),
             source: src2,
@@ -593747,9 +593772,9 @@ ${result}`
             try {
               const { writeFileSync: writeFileSync95, readFileSync: readFileSync137, unlinkSync: unlinkSync37 } = await import("node:fs");
               const { join: join186 } = await import("node:path");
-              const { tmpdir: tmpdir25 } = await import("node:os");
-              const tmpIn = join186(tmpdir25(), `omnius_img_in_${Date.now()}.png`);
-              const tmpOut = join186(tmpdir25(), `omnius_img_out_${Date.now()}.jpg`);
+              const { tmpdir: tmpdir26 } = await import("node:os");
+              const tmpIn = join186(tmpdir26(), `omnius_img_in_${Date.now()}.png`);
+              const tmpOut = join186(tmpdir26(), `omnius_img_out_${Date.now()}.jpg`);
               writeFileSync95(tmpIn, buffer2);
               const pyBin = process.platform === "win32" ? "python" : "python3";
               const resizeScript = [
@@ -595284,7 +595309,7 @@ var init_constraint_learner = __esm({
 import { existsSync as existsSync102, statSync as statSync39, openSync, readSync, closeSync, unlinkSync as unlinkSync18, writeFileSync as writeFileSync49 } from "node:fs";
 import { watch as fsWatch } from "node:fs";
 import { join as join113 } from "node:path";
-import { tmpdir as tmpdir20 } from "node:os";
+import { tmpdir as tmpdir21 } from "node:os";
 import { randomBytes as randomBytes21 } from "node:crypto";
 var NexusAgenticBackend;
 var init_nexusBackend = __esm({
@@ -595486,7 +595511,7 @@ ${suffix}` } : m2);
        * Falls back to unary + word-split if streaming setup fails.
        */
       async *chatCompletionStream(request) {
-        const streamFile = join113(tmpdir20(), `nexus-stream-${randomBytes21(6).toString("hex")}.jsonl`);
+        const streamFile = join113(tmpdir21(), `nexus-stream-${randomBytes21(6).toString("hex")}.jsonl`);
         writeFileSync49(streamFile, "", "utf8");
         const effectiveThink = this.effectiveThink(request);
         const daemonArgs = {
@@ -603069,7 +603094,7 @@ function liveDashboardPreviewWidthForSources(sourceCount, termWidth = process.st
 function formatCameraPane(camera, paneWidth, paneHeight, now2) {
   const contentWidth = Math.max(10, paneWidth - 2);
   const age = Math.max(0, Math.round((now2 - camera.updatedAt) / 1e3));
-  const title = `${compactSourceId(camera.source)} ${age}s${cameraPaneBadge(camera)}`;
+  const title = `↺ ${compactSourceId(camera.source)} ${age}s${cameraPaneBadge(camera)} ↻`;
   const topTitle = ` ${title} `;
   const left = Math.max(0, Math.floor((contentWidth - topTitle.length) / 2));
   const right = Math.max(0, contentWidth - topTitle.length - left);
@@ -603080,6 +603105,30 @@ function formatCameraPane(camera, paneWidth, paneHeight, now2) {
   const body = fitDashboardPane(previewLines, contentWidth, paneHeight, fallback);
   return [top, ...body.map((line) => `│${truncateAnsiCell(line, contentWidth)}│`), bottom];
 }
+function liveDashboardRotationActionFromClick(snapshot, lineText, col) {
+  const plain = stripDashboardAnsi(lineText);
+  const idx = Math.max(0, col - 1);
+  const clicked = plain[idx];
+  if (clicked !== "↺" && clicked !== "↻") return null;
+  const direction = clicked === "↺" ? "ccw" : "cw";
+  const cameras = Object.values(snapshot.cameras ?? (snapshot.video ? { [snapshot.video.source]: snapshot.video } : {}));
+  if (cameras.length === 0) return null;
+  const leftPositions = [];
+  for (let i2 = 0; i2 < plain.length; i2++) {
+    if (plain[i2] === "↺") leftPositions.push(i2);
+  }
+  for (let i2 = 0; i2 < leftPositions.length; i2++) {
+    const left = leftPositions[i2];
+    const nextLeft = leftPositions[i2 + 1] ?? plain.length;
+    const right = plain.indexOf("↻", left + 1);
+    if (right < 0 || right >= nextLeft) continue;
+    if (idx !== left && idx !== right) continue;
+    const segment = plain.slice(left, right + 1);
+    const camera = cameras.find((entry) => segment.includes(compactSourceId(entry.source)));
+    if (camera) return { source: camera.source, direction };
+  }
+  return null;
+}
 function pushDashboardWrapped(lines, value2, width) {
   const contentWidth = Math.max(10, width - 2);
   const words = stripDashboardAnsi(value2).replace(/\s+/g, " ").trim().split(" ").filter(Boolean);
@@ -603107,6 +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);
@@ -603221,39 +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) {
@@ -603286,7 +603318,7 @@ async function classifyLiveSpeech(proposal, config) {
             `urgency=${proposal.urgency}`,
             `summary=${proposal.summary}`,
             `evidence=${proposal.evidence}`,
-            `default_utterance=${proposal.defaultText}`
+            proposal.speechGoal ? `speech_goal=${proposal.speechGoal}` : ""
           ].join("\n").slice(0, 1600)
         }
       ]
@@ -603586,7 +603618,7 @@ function formatLiveDashboardFromSnapshot(snapshot, opts = {}) {
     lines.push(`├${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)}%` : "",
@@ -603969,7 +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("");
@@ -604346,12 +604378,15 @@ var init_live_sensors = __esm({
       autoOrientationInFlight = /* @__PURE__ */ new Set();
       lastFeedbackAt = /* @__PURE__ */ new Map();
       lastClipAt = /* @__PURE__ */ new Map();
+      lastInferenceAt = /* @__PURE__ */ new Map();
       nextVideoSourceIndex = 0;
       liveSpeechSink = null;
       liveSpeechEnabled = false;
       lastLiveSpeechAt = 0;
       lastLiveSpeechByKey = /* @__PURE__ */ new Map();
       liveSpeechInFlight = /* @__PURE__ */ new Set();
+      liveSpeechMutedUntil = 0;
+      liveTtsSpeaking = null;
       intakeAgentConfig;
       setStatusSink(sink) {
         this.statusSink = sink ?? null;
@@ -604369,6 +604404,15 @@ var init_live_sensors = __esm({
       setLiveSpeechEnabled(enabled2) {
         this.liveSpeechEnabled = enabled2;
       }
+      setLiveTtsSpeakingSource(source) {
+        this.liveTtsSpeaking = source ?? null;
+      }
+      muteLiveAudioFor(ms) {
+        this.liveSpeechMutedUntil = Math.max(this.liveSpeechMutedUntil, Date.now() + Math.max(0, ms));
+      }
+      isLiveAudioMutedForTts() {
+        return Date.now() < this.liveSpeechMutedUntil || Boolean(this.liveTtsSpeaking?.());
+      }
       setAudioIntakeAgentConfig(config) {
         this.intakeAgentConfig = config && config.backendUrl && config.model ? { backendUrl: config.backendUrl, model: config.model, apiKey: config.apiKey } : void 0;
       }
@@ -604401,6 +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,
@@ -604439,6 +604487,12 @@ var init_live_sensors = __esm({
         const next = current === 0 ? 90 : current === 90 ? 180 : current === 180 ? 270 : 0;
         return this.setCameraRotation(device, next, "manual", "manual cycle from /live menu");
       }
+      rotateCameraRelative(device, direction) {
+        const current = this.getCameraRotation(device);
+        const delta = direction === "cw" ? 90 : 270;
+        const next = (current + delta) % 360;
+        return this.setCameraRotation(device, next, "manual", `manual ${direction === "cw" ? "clockwise" : "counter-clockwise"} dashboard override`);
+      }
       async autoOrientCamera(device = this.config.selectedCamera, options2 = {}) {
         const target = device || this.config.selectedCamera || firstDeviceId(this.devices.video);
         if (!target) return "No camera selected. Use /live camera <device> or select a camera in /live.";
@@ -604594,7 +604648,7 @@ var init_live_sensors = __esm({
             }, 6e4);
           });
         }
-        void this.sampleVideoNow(true, { mode: "all" }).catch((err) => {
+        void this.sampleVideoNow(true, { mode: "all", forceInferenceNow: true }).catch((err) => {
           this.emitFeedback({
             kind: "error",
             title: "Live video sample failed",
@@ -604781,13 +604835,17 @@ var init_live_sensors = __esm({
           return message2;
         }
       }
-      async sampleSingleVideoNow(source, forceInfer, previewWidth) {
+      async sampleSingleVideoNow(source, forceInfer, previewWidth, options2 = {}) {
         const preview = await this.previewCamera(source, { emit: false, previewWidth });
         let inference = "";
         let clipSummary = "";
         const orientation = this.getCameraOrientation(source);
         const sampledAt = Date.now();
-        if (forceInfer && source) {
+        const inferenceThrottleMs = Math.max(0, Number(options2.inferThrottleMs ?? 2500));
+        const lastInferenceAt = this.lastInferenceAt.get(source) ?? 0;
+        const shouldInfer = Boolean(forceInfer && source && (options2.forceInferenceNow || sampledAt - lastInferenceAt >= inferenceThrottleMs));
+        if (shouldInfer) {
+          this.lastInferenceAt.set(source, sampledAt);
           const loop = new LiveMediaLoopTool(this.repoRoot);
           const inferFromCapturedFrame = preview.ok && Boolean(preview.framePath);
           const result = await loop.execute({
@@ -604856,7 +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);
@@ -604873,7 +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);
@@ -604900,6 +604958,15 @@ ${inference}` : "", clipSummary ? `
 CLIP visual memory:
 ${clipSummary}` : ""].filter(Boolean).join("\n");
       }
+      async sampleCameraNow(source, forceInfer = this.config.inferEnabled || this.config.clipEnabled, options2 = {}) {
+        const previewWidth = liveDashboardPreviewWidthForSources(this.activeVideoSources().length);
+        const result = await this.sampleSingleVideoNow(source, forceInfer, previewWidth, {
+          forceInferenceNow: options2.forceInferenceNow,
+          inferThrottleMs: options2.forceInferenceNow ? 0 : 2500
+        });
+        this.emitDashboard();
+        return result;
+      }
       async sampleVideoNow(forceInfer = this.config.inferEnabled, options2 = {}) {
         if (this.videoInFlight) return "Video sampler is already running.";
         this.videoInFlight = true;
@@ -604911,7 +604978,7 @@ ${clipSummary}` : ""].filter(Boolean).join("\n");
           const outputs = [];
           const previewWidth = liveDashboardPreviewWidthForSources(allSources.length);
           for (const source of sources) {
-            outputs.push(await this.sampleSingleVideoNow(source, forceInfer, previewWidth));
+            outputs.push(await this.sampleSingleVideoNow(source, forceInfer, previewWidth, options2));
           }
           this.emitDashboard();
           return outputs.map((output, index) => `## Camera ${sources[index]}
@@ -604927,6 +604994,20 @@ ${output}`).join("\n\n");
           const preferredInput = preferredLiveAudioInputId(this.devices.audioInputs, this.config.selectedAudioInput) || "default";
           const input = this.config.selectedAudioInput && !isAudioOutputMonitorId(this.config.selectedAudioInput) ? this.config.selectedAudioInput : preferredInput;
           if (input !== this.config.selectedAudioInput) this.config.selectedAudioInput = input;
+          if (this.isLiveAudioMutedForTts()) {
+            const now2 = Date.now();
+            this.snapshot.audio = {
+              updatedAt: now2,
+              input,
+              output: this.config.selectedAudioOutput,
+              transcript: "",
+              analysis: "",
+              error: "Live audio capture muted during Omnius TTS playback to prevent echo."
+            };
+            this.persist();
+            this.emitDashboard();
+            return "Live audio capture muted during Omnius TTS playback to prevent echo.";
+          }
           const outputDir2 = liveDir(this.repoRoot);
           ensureLiveDir(this.repoRoot);
           const recordingPath = join124(outputDir2, `audio-${Date.now()}.wav`);
@@ -604963,7 +605044,11 @@ ${output}`).join("\n\n");
           }
           if (this.config.asrEnabled) {
             try {
-              const asr = await new TranscribeFileTool(this.repoRoot).execute({ path: recordingPath, model: "tiny" });
+              const asr = await new TranscribeFileTool(this.repoRoot).execute({
+                path: recordingPath,
+                model: "tiny",
+                backend: "managed-whisper"
+              });
               if (asr.success) {
                 transcript = cleanPreview(typeof asr.llmContent === "string" ? asr.llmContent : "", 1600);
                 asrBackend = extractAsrBackend(asr.output) ?? "whisper";
@@ -605043,7 +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);
@@ -606316,7 +606401,7 @@ var init_listen = __esm({
               let caughtUncaught = null;
               const uncaughtHandler = (err) => {
                 const msg = err?.message || String(err);
-                if (msg.includes("live_worker") || msg.includes("numpy") || msg.includes("Traceback")) {
+                if (msg.includes("live_worker") || msg.includes("numpy") || msg.includes("transcribe_cli") || msg.includes("ModuleNotFoundError") || msg.includes("Traceback")) {
                   caughtUncaught = err;
                 } else {
                   throw err;
@@ -625268,6 +625353,7 @@ var init_status_bar = __esm({
        * repaint, including selection updates and scroll events.
        */
       _dynamicBlocks = /* @__PURE__ */ new Map();
+      _dynamicBlockClickHandlers = /* @__PURE__ */ new Map();
       /** Sentinel marker — used both for scrollback storage and reflow detection. */
       DYNAMIC_BLOCK_MARK_PREFIX = "
DYNBLOCK:";
       DYNAMIC_BLOCK_MARK_SUFFIX = "
";
@@ -625414,9 +625500,13 @@ var init_status_bar = __esm({
         this.invalidateRowCountCache();
         return `${this.DYNAMIC_BLOCK_MARK_PREFIX}${id2}${this.DYNAMIC_BLOCK_MARK_SUFFIX}`;
       }
+      registerDynamicBlockClickHandler(id2, handler) {
+        this._dynamicBlockClickHandlers.set(id2, handler);
+      }
       /** Unregister a dynamic block. Existing sentinels in scrollback become inert (rendered as empty). */
       unregisterDynamicBlock(id2) {
         this._dynamicBlocks.delete(id2);
+        this._dynamicBlockClickHandlers.delete(id2);
         this._dynamicBlockVersion++;
         this.invalidateRowCountCache();
       }
@@ -627119,9 +627209,23 @@ var init_status_bar = __esm({
        * other body rows are left alone so text selection still works. Returns true
        * when the click was consumed.
        */
-      handleContentBlockClick(screenRow) {
+      handleContentBlockClick(screenRow, col) {
         const hit = this.hitTestContentBlock(screenRow);
-        if (!hit || !isCollapsibleBlock(hit.id)) return false;
+        if (!hit) return false;
+        const handler = this._dynamicBlockClickHandlers.get(hit.id);
+        if (handler?.({
+          id: hit.id,
+          offset: hit.offset,
+          col,
+          row: screenRow,
+          lineText: hit.lineText
+        })) {
+          if (this.active && !isOverlayActive() && !this._suspendContentLayer) {
+            this.repaintContent();
+          }
+          return true;
+        }
+        if (!isCollapsibleBlock(hit.id)) return false;
         const plain = stripAnsi(hit.lineText);
         if (plain.includes(READ_MORE_LABEL) || plain.includes(READ_LESS_LABEL)) {
           toggleMore(hit.id);
@@ -627145,7 +627249,7 @@ var init_status_bar = __esm({
         if (!this.active) return;
         const w = termCols();
         if (type === "press" && row >= this.scrollRegionTop) {
-          if (this.handleContentBlockClick(row)) return;
+          if (this.handleContentBlockClick(row, col)) return;
         }
         if (type === "press" && this._suggestions.length > 0) {
           if (this.suggestClickAt(row)) return;
@@ -638356,7 +638460,7 @@ var init_audio_waveform = __esm({
 
 // packages/cli/src/tui/neovim-mode.ts
 import { existsSync as existsSync129, unlinkSync as unlinkSync24 } from "node:fs";
-import { tmpdir as tmpdir21 } from "node:os";
+import { tmpdir as tmpdir22 } from "node:os";
 import { join as join143 } from "node:path";
 function isNeovimActive() {
   return _state2 !== null && !_state2.cleanedUp;
@@ -638402,7 +638506,7 @@ async function startNeovimMode(opts) {
     );
   } catch {
   }
-  const socketPath = join143(tmpdir21(), `omnius-nvim-${process.pid}-${Date.now()}.sock`);
+  const socketPath = join143(tmpdir22(), `omnius-nvim-${process.pid}-${Date.now()}.sock`);
   try {
     if (existsSync129(socketPath)) unlinkSync24(socketPath);
   } catch {
@@ -641943,7 +642047,7 @@ import {
   rmSync as rmSync11
 } from "node:fs";
 import { join as join150, dirname as dirname45, resolve as resolve61 } from "node:path";
-import { homedir as homedir50, tmpdir as tmpdir22, platform as platform6 } from "node:os";
+import { homedir as homedir50, tmpdir as tmpdir23, platform as platform6 } from "node:os";
 import {
   spawn as nodeSpawn
 } from "node:child_process";
@@ -643639,6 +643743,9 @@ except Exception as exc:
           await this.sleep(100);
         }
       }
+      isSpeaking() {
+        return this.speaking || this.speakQueue.length > 0 || Boolean(this.currentPlayback);
+      }
       enqueueSpeech(text2, volume, pitchFactor, speedFactor = 1, stereoDelayMs = 0.6, emotion) {
         if (!this.enabled || !this.ready) return;
         text2 = sanitizeForTTS(text2);
@@ -644036,7 +644143,7 @@ except Exception as exc:
           }
           this.onPCMOutput(Buffer.from(int16.buffer), sampleRate);
         }
-        const wavPath = join150(tmpdir22(), `omnius-voice-${Date.now()}.wav`);
+        const wavPath = join150(tmpdir23(), `omnius-voice-${Date.now()}.wav`);
         this.writeStereoWav(stereo.left, stereo.right, sampleRate, wavPath);
         await this.playWav(wavPath);
         try {
@@ -644549,7 +644656,7 @@ except Exception as exc:
         const cleaned = injectExpressionTags ? applySupertonicExpressionTags(baseText, settings.expression, emotion) : baseText;
         if (!cleaned) return null;
         const wavPath = join150(
-          tmpdir22(),
+          tmpdir23(),
           `omnius-supertonic3-${Date.now()}-${Math.random().toString(36).slice(2, 6)}.wav`
         );
         try {
@@ -644620,7 +644727,7 @@ except Exception as exc:
         return new Promise((resolve76, reject) => {
           const proc = nodeSpawn("sh", ["-c", command], {
             stdio: ["ignore", "pipe", "pipe"],
-            cwd: tmpdir22(),
+            cwd: tmpdir23(),
             env: voicePythonEnv()
           });
           let stdout = "";
@@ -644707,7 +644814,7 @@ Error: ${err instanceof Error ? err.message : String(err)}`
         const mlxModelId = model.mlxModelId ?? "mlx-community/Kokoro-82M-bf16";
         const mlxVoice = model.mlxVoice ?? "af_heart";
         const mlxLangCode = model.mlxLangCode ?? "a";
-        const wavPath = join150(tmpdir22(), `omnius-mlx-${Date.now()}.wav`);
+        const wavPath = join150(tmpdir23(), `omnius-mlx-${Date.now()}.wav`);
         const pyScript = [
           "import sys, json",
           "from mlx_audio.tts import generate as tts_gen",
@@ -644788,7 +644895,7 @@ Error: ${err instanceof Error ? err.message : String(err)}`
         const mlxModelId = model.mlxModelId ?? "mlx-community/Kokoro-82M-bf16";
         const mlxVoice = model.mlxVoice ?? "af_heart";
         const mlxLangCode = model.mlxLangCode ?? "a";
-        const wavPath = join150(tmpdir22(), `omnius-mlx-buf-${Date.now()}.wav`);
+        const wavPath = join150(tmpdir23(), `omnius-mlx-buf-${Date.now()}.wav`);
         const pyScript = [
           "import sys, json",
           "from mlx_audio.tts import generate as tts_gen",
@@ -645564,7 +645671,7 @@ if __name__ == '__main__':
           const env2 = voicePythonEnv({ LUXTTS_REPO_PATH: luxttsRepoDir2() });
           const daemon = nodeSpawn(venvPy, [luxttsInferScript2()], {
             stdio: ["pipe", "pipe", "pipe"],
-            cwd: tmpdir22(),
+            cwd: tmpdir23(),
             env: env2
           });
           this._luxttsDaemon = daemon;
@@ -645654,7 +645761,7 @@ if __name__ == '__main__':
         const ready = await this.ensureLuxttsDaemon();
         if (!ready) return null;
         const wavPath = join150(
-          tmpdir22(),
+          tmpdir23(),
           `omnius-luxtts-${Date.now()}-${Math.random().toString(36).slice(2, 6)}.wav`
         );
         try {
@@ -645795,7 +645902,7 @@ if __name__ == '__main__':
         if (!cleaned) return null;
         const ready = await this.ensureLuxttsDaemon();
         if (!ready) return null;
-        const wavPath = join150(tmpdir22(), `omnius-luxtts-buf-${Date.now()}.wav`);
+        const wavPath = join150(tmpdir23(), `omnius-luxtts-buf-${Date.now()}.wav`);
         try {
           await this.luxttsRequest({
             action: "synthesize",
@@ -645929,7 +646036,7 @@ if __name__ == "__main__":
           });
           const daemon = nodeSpawn(venvPy, [misottsInferScript2()], {
             stdio: ["pipe", "pipe", "pipe"],
-            cwd: tmpdir22(),
+            cwd: tmpdir23(),
             env: env2
           });
           this._misottsDaemon = daemon;
@@ -646013,7 +646120,7 @@ if __name__ == "__main__":
         const ready = await this.ensureMisottsDaemon();
         if (!ready) return null;
         const wavPath = join150(
-          tmpdir22(),
+          tmpdir23(),
           `omnius-misotts-${Date.now()}-${Math.random().toString(36).slice(2, 6)}.wav`
         );
         try {
@@ -646163,7 +646270,7 @@ if __name__ == "__main__":
         if (!cleaned) return null;
         const ready = await this.ensureMisottsDaemon();
         if (!ready) return null;
-        const wavPath = join150(tmpdir22(), `omnius-misotts-buf-${Date.now()}.wav`);
+        const wavPath = join150(tmpdir23(), `omnius-misotts-buf-${Date.now()}.wav`);
         try {
           const settings = this.getMisottsSettings();
           const cloneName = this.misottsCloneRef.split("/").pop() ?? "";
@@ -656904,6 +657011,15 @@ function renderLiveDashboard(ctx3, manager) {
       id2,
       (width) => formatLiveDashboardFromSnapshot(manager.getSnapshot(), { width })
     );
+    host.registerDynamicBlockClickHandler?.(id2, ({ col, lineText }) => {
+      const action = liveDashboardRotationActionFromClick(manager.getSnapshot(), lineText, col);
+      if (!action) return false;
+      manager.rotateCameraRelative(action.source, action.direction);
+      void manager.sampleCameraNow(action.source, true, { forceInferenceNow: true }).catch((err) => {
+        renderWarning(`Camera refresh failed after manual rotation: ${err instanceof Error ? err.message : String(err)}`);
+      });
+      return true;
+    });
     host.appendDynamicBlock(id2);
     return;
   }
@@ -656934,7 +657050,7 @@ ${feedback.message}`);
   renderInfo(`[live ${stamp}] ${feedback.title}
 ${feedback.message}`);
 }
-function attachLiveSinks(ctx3, manager, agentEnabled = false) {
+function attachLiveSinks(ctx3, manager, agentEnabled = false, speechEnabledOverride) {
   manager.setStatusSink(ctx3.setLiveMediaStatus);
   manager.setFeedbackSink((feedback) => renderLiveFeedback(ctx3, manager, feedback));
   manager.setAudioIntakeAgentConfig({
@@ -656942,10 +657058,13 @@ function attachLiveSinks(ctx3, manager, agentEnabled = false) {
     model: ctx3.config.model,
     apiKey: ctx3.config.apiKey
   });
-  const speechEnabled = agentEnabled && Boolean(ctx3.voiceSpeak);
+  const speechEnabled = Boolean(speechEnabledOverride ?? agentEnabled) && Boolean(ctx3.voiceSpeak);
   manager.setLiveSpeechEnabled(speechEnabled);
+  manager.setLiveTtsSpeakingSource(ctx3.voiceIsSpeaking);
   manager.setLiveSpeechSink(speechEnabled ? (text2) => {
     if (ctx3.voiceIsEnabled && !ctx3.voiceIsEnabled()) return;
+    const estimatedSpeechMs = Math.max(1500, Math.min(15e3, text2.length * 85));
+    manager.muteLiveAudioFor(estimatedSpeechMs + 1e3);
     ctx3.voiceSpeak?.(text2);
   } : void 0);
   if (agentEnabled && ctx3.startBackgroundPrompt) {
@@ -656958,17 +657077,17 @@ function attachLiveSinks(ctx3, manager, agentEnabled = false) {
   }
 }
 async function startLiveVoicechat(ctx3, manager) {
-  const agentEnabled = Boolean(ctx3.startBackgroundPrompt);
-  attachLiveSinks(ctx3, manager, agentEnabled);
+  const speechEnabled = Boolean(ctx3.voiceSpeak);
+  attachLiveSinks(ctx3, manager, false, speechEnabled);
   await ensureLiveInventory(manager);
   if (!ctx3.voiceChatStart) {
-    renderInfo(manager.startRunMode({ agent: agentEnabled, lowLatency: true, speech: false }));
+    renderInfo(manager.startRunMode({ agent: false, lowLatency: true, speech: false }));
     renderWarning("Voice chat is not available in this context; live audio/video feedback remains active.");
     return;
   }
   ctx3.voiceSetMode?.("voicechat");
   if (ctx3.isVoiceChatActive?.()) {
-    renderInfo(manager.startRunMode({ agent: agentEnabled, lowLatency: true, speech: Boolean(ctx3.voiceSpeak) }));
+    renderInfo(manager.startRunMode({ agent: false, lowLatency: true, speech: speechEnabled }));
     renderInfo("Voice chat is already active. Live audio/video feedback remains attached.");
     return;
   }
@@ -656979,7 +657098,7 @@ async function startLiveVoicechat(ctx3, manager) {
   try {
     renderInfo("Starting live voicechat with audio/video context...");
     await ctx3.voiceChatStart();
-    renderInfo(manager.startRunMode({ agent: agentEnabled, lowLatency: true, speech: Boolean(ctx3.voiceSpeak) }));
+    renderInfo(manager.startRunMode({ agent: false, lowLatency: true, speech: speechEnabled }));
     renderInfo("Live voicechat active: low-latency audio/video context is updating while voice conversation continues.");
   } catch (err) {
     renderError(`Live voicechat failed: ${err instanceof Error ? err.message : String(err)}`);
@@ -695342,15 +695461,15 @@ Content-Type: ${mimeForPath(pathOrFileId, field === "photo" ? "image" : "video")
           let filename = "response.wav";
           let mimeType = "audio/wav";
           try {
-            const { tmpdir: tmpdir25 } = await import("node:os");
+            const { tmpdir: tmpdir26 } = await import("node:os");
             const { join: joinPath } = await import("node:path");
             const {
               writeFileSync: writeFs,
               readFileSync: readFs,
               unlinkSync: unlinkFs
             } = await import("node:fs");
-            const tmpWav = joinPath(tmpdir25(), `omnius-tg-voice-${Date.now()}.wav`);
-            const tmpOgg = joinPath(tmpdir25(), `omnius-tg-voice-${Date.now()}.ogg`);
+            const tmpWav = joinPath(tmpdir26(), `omnius-tg-voice-${Date.now()}.wav`);
+            const tmpOgg = joinPath(tmpdir26(), `omnius-tg-voice-${Date.now()}.ogg`);
             writeFs(tmpWav, wavBuffer);
             await execFileText4(
               "ffmpeg",
@@ -695735,7 +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) {
@@ -695832,19 +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.!?]*)+$/i;
     VoiceChatSession = class extends EventEmitter12 {
@@ -695874,6 +696115,8 @@ Rules:
       silenceTimer = null;
       maxSegmentTimer = null;
       lastSignalScore = null;
+      listenPausedForResponse = false;
+      lastAgentSpeech = null;
       // Abort control for inference
       abortController = null;
       // Callbacks
@@ -695949,6 +696192,10 @@ Rules:
           this.context.push({ role: "system", content: this.toolCatalogNote });
         }
         this.turnCount = 0;
+        this.voiceTranscript = [];
+        this.relayTranscript = [];
+        this.lastAgentSpeech = null;
+        this.listenPausedForResponse = false;
         if (this.verbose) this.onStatus("VoiceChat active — LISTENING");
         this._onTranscript = (...args) => {
           let text2;
@@ -696038,6 +696285,12 @@ Rules:
         if (this._state !== "LISTENING" && this._state !== "CAPTURING") {
           return;
         }
+        const echo = this.classifyAgentEcho(text2);
+        if (echo.likelyEcho) {
+          if (this.debugSnr) this.onStatus(`Ignoring likely agent-speaker echo (${echo.echoSimilarity.toFixed(2)}): ${truncateForLog(text2, 48)}`);
+          this.emit("voiceEchoFiltered", echo);
+          return;
+        }
         if (this._state === "LISTENING") {
           this.setState("CAPTURING");
           this.captureBuffer = "";
@@ -696088,6 +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 });
@@ -696098,9 +696352,7 @@ Rules:
           } catch {
           }
         }
-        while (this.context.length > MAX_CONTEXT_TURNS + 1) {
-          this.context.splice(1, 1);
-        }
+        this.trimContext();
         this.think();
       }
       // ---------------------------------------------------------------------------
@@ -696111,18 +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();
@@ -696176,6 +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) {
@@ -696183,18 +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") {
@@ -696203,9 +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;
@@ -696217,14 +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.
@@ -696236,11 +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",
@@ -696257,10 +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` : `${baseUrl2}/v1/chat/completions`;
         const res = await fetch(endpoint, { method: "POST", headers, body: openaiBody, signal });
@@ -700370,7 +700665,7 @@ __export(graphical_sudo_exports, {
 import { spawn as spawn35 } from "node:child_process";
 import { existsSync as existsSync161, mkdirSync as mkdirSync102, writeFileSync as writeFileSync87, chmodSync as chmodSync6 } from "node:fs";
 import { join as join174 } from "node:path";
-import { tmpdir as tmpdir23 } from "node:os";
+import { tmpdir as tmpdir24 } from "node:os";
 function detectSudoHelper() {
   if (process.platform === "win32") return null;
   if (process.platform === "darwin") return "osascript";
@@ -700391,7 +700686,7 @@ function which2(cmd) {
   return null;
 }
 function ensureAskpassShim(helper, description) {
-  const shimDir = join174(tmpdir23(), "omnius-askpass");
+  const shimDir = join174(tmpdir24(), "omnius-askpass");
   mkdirSync102(shimDir, { recursive: true });
   const shim = join174(shimDir, `${helper}.sh`);
   let body;
@@ -722734,11 +723029,11 @@ async function handleRequest(req3, res, ollamaUrl, verbose, runtimeDefaults = {}
           });
           return;
         }
-        const { tmpdir: tmpdir25 } = await import("node:os");
+        const { tmpdir: tmpdir26 } = await import("node:os");
         const { writeFileSync: writeFileSync95, unlinkSync: unlinkSync37 } = await import("node:fs");
         const { join: pjoin } = await import("node:path");
         const tmpPath = pjoin(
-          tmpdir25(),
+          tmpdir26(),
           `omnius-clone-upload-${Date.now()}-${safeName3}`
         );
         writeFileSync95(tmpPath, buf);
@@ -734213,6 +734508,9 @@ The user pasted a clipboard image saved at ${relPath}. Use the OCR, vision analy
     voiceIsEnabled() {
       return voiceEngine.enabled;
     },
+    voiceIsSpeaking() {
+      return voiceEngine.isSpeaking();
+    },
     voiceGetModel() {
       return voiceEngine.modelId ?? "unknown";
     },
@@ -735092,6 +735390,7 @@ Respond concisely and safely. Remember: you are talking to the general public.`;
     liveDynamicBlockHost: {
       registerDynamicBlock: (id2, render2) => statusBar.registerDynamicBlock(id2, render2),
       appendDynamicBlock: (id2) => statusBar.appendDynamicBlock(id2),
+      registerDynamicBlockClickHandler: (id2, handler) => statusBar.registerDynamicBlockClickHandler(id2, handler),
       refreshDynamicBlocks: () => statusBar.refreshDynamicBlocks()
     },
     // Voice call session — standalone cloudflared tunnel for /call
@@ -738774,7 +739073,7 @@ __export(eval_exports, {
   evalCommand: () => evalCommand,
   expectedStatusesForEvalTask: () => expectedStatusesForEvalTask
 });
-import { tmpdir as tmpdir24 } from "node:os";
+import { tmpdir as tmpdir25 } from "node:os";
 import { mkdirSync as mkdirSync110, writeFileSync as writeFileSync94 } from "node:fs";
 import { join as join184 } from "node:path";
 function expectedStatusesForEvalTask(task, live) {
@@ -738918,7 +739217,7 @@ async function evalCommand(opts, config) {
   process.exit(failed > 0 ? 1 : 0);
 }
 function createTempEvalRepo() {
-  const dir = join184(tmpdir24(), `omnius-eval-${Date.now()}`);
+  const dir = join184(tmpdir25(), `omnius-eval-${Date.now()}`);
   mkdirSync110(dir, { recursive: true });
   mkdirSync110(join184(dir, "src"), { recursive: true });
   mkdirSync110(join184(dir, "tests"), { recursive: true });