diff --git a/recorder.go b/recorder.go
index c8ad720..d5ffd35 100644
--- a/recorder.go
+++ b/recorder.go
@@ -79,8 +79,9 @@ func (rm *RecorderManager) StopAll() {
 //      d. On failure: .part file stays (picked up by cleaner), wait 5s, retry
 //
 // Each segment is exactly 5 minutes. Gaps between segments are ~1-2s
-// (FFmpeg restart overhead). The -t 300 flag provides a hard timeout
-// preventing hung FFmpeg processes.
+// (FFmpeg restart overhead). The -t 300 flag bounds the segment by input
+// duration, while -rw_timeout makes ffmpeg exit (rather than hang) when the
+// RTSP stream stalls, so the retry loop below can reconnect.
 func (rm *RecorderManager) startRecorder(cam CameraConfig) {
 	rtspURL := cam.RTSPMain
 	if rtspURL == "" {
@@ -125,14 +126,16 @@ func (rm *RecorderManager) startRecorder(cam CameraConfig) {
 		partFile := filepath.Join(outDir, "rec_"+timeStr+".part.mp4")
 		finalFile := filepath.Join(outDir, "rec_"+timeStr+".mp4")
 
-		// Build FFmpeg command with hard 5-minute timeout.
+		// Build FFmpeg command: 5-minute segment, with a socket I/O timeout so a
+		// stalled stream exits ffmpeg instead of hanging the recorder loop.
 		cmd := exec.Command("ffmpeg",
 			"-hide_banner", "-loglevel", "error",
 			"-rtsp_transport", "tcp",
+			"-rw_timeout", "10000000", // 10s socket I/O timeout — exit on stalled stream
 			"-i", rtspURL,
 			"-an",              // drop audio (pcm_mulaw not supported in MP4)
 			"-c:v", "copy",     // stream-copy video: zero transcoding, near-zero CPU
-			"-t", "300",        // hard stop after 300 seconds (5 minutes)
+			"-t", "300",        // stop after 300 seconds (5 minutes) of input
 			"-y",               // overwrite output without asking
 			partFile,
 		)