diff --git a/src/headless.rs b/src/headless.rs
index a0af3f3..11810c7 100644
--- a/src/headless.rs
+++ b/src/headless.rs
@@ -364,7 +364,7 @@ fn run_once(configs: &[config::AgentConfig]) -> Result<(), String> {
 }
 
 /// Remove directory listing that Claude CLI prepends to output.
-fn strip_dir_listing(text: &str) -> &str {
+pub fn strip_dir_listing(text: &str) -> &str {
     // Pattern: lines of single words (filenames) at the start, possibly starting with "."
     let mut end = 0;
     for line in text.lines() {
diff --git a/src/tui.rs b/src/tui.rs
index 99210e8..d1dad42 100644
--- a/src/tui.rs
+++ b/src/tui.rs
@@ -176,6 +176,11 @@ impl App {
                         self.ai_scroll = u16::MAX;
                     }
                     OutputEvent::StreamEnd => {
+                        // Strip directory listing that Claude CLI prepends
+                        let clean = crate::headless::strip_dir_listing(&self.ai_output);
+                        if clean.len() != self.ai_output.len() {
+                            self.ai_output = clean.to_string();
+                        }
                         write_private(
                             &format!("{STATE_DIR}/ai.txt"),
                             self.ai_output.as_bytes(),
diff --git a/src/voice/mod.rs b/src/voice/mod.rs
index 3ba7792..42568ad 100644
--- a/src/voice/mod.rs
+++ b/src/voice/mod.rs
@@ -63,18 +63,15 @@ impl VoiceSystem {
         if text.trim().is_empty() { return; }
         let audio = match tts::synthesize(&self.config, text) {
             Ok(data) => data,
-            Err(e) => { eprintln!("tts error: {e}"); return; }
+            Err(_) => return,
         };
-        if let Err(e) = tts::play_audio(&audio, &self.config.tts_model) {
-            eprintln!("audio play error: {e}");
-        }
+        let _ = tts::play_audio(&audio, &self.config.tts_model);
     }
 
     pub fn listen(&self) -> Option<String> {
         match stt::recognize(&self.config) {
             Ok(text) if !text.is_empty() => Some(text),
-            Ok(_) => None,
-            Err(e) => { eprintln!("stt error: {e}"); None }
+            _ => None,
         }
     }
 }
diff --git a/src/voice/stt.rs b/src/voice/stt.rs
index d71ee7a..f456373 100644
--- a/src/voice/stt.rs
+++ b/src/voice/stt.rs
@@ -71,7 +71,7 @@ fn record_vad() -> Result<Vec<i16>, String> {
     let mut speech_count: u32 = 0;
     let mut total_frames: u32 = 0;
 
-    eprintln!("  listening...");
+    // Note: no eprintln here — it corrupts TUI alternate screen
 
     loop {
         match rx.recv_timeout(std::time::Duration::from_millis(100)) {
diff --git a/src/voice/tts.rs b/src/voice/tts.rs
index 456341f..507c3f0 100644
--- a/src/voice/tts.rs
+++ b/src/voice/tts.rs
@@ -45,9 +45,16 @@ pub fn synthesize(config: &VoiceConfig, text: &str) -> Result<Vec<u8>, String> {
         return Err(format!("TTS API error {status}: {body}"));
     }
 
-    resp.bytes()
+    let bytes = resp.bytes()
         .map(|b| b.to_vec())
-        .map_err(|e| format!("TTS read error: {e}"))
+        .map_err(|e| format!("TTS read error: {e}"))?;
+
+    // Verify it's actually audio (MP3 starts with ID3 or 0xFF sync)
+    if bytes.len() < 4 || (bytes[0] != 0xFF && &bytes[0..3] != b"ID3") {
+        return Err(format!("TTS returned invalid audio ({} bytes)", bytes.len()));
+    }
+
+    Ok(bytes)
 }
 
 /// Play audio bytes (MP3) using rodio with pitch shift from ai.json.