diff --git a/c64cast.schema.json b/c64cast.schema.json
index d095306..bea3032 100644
--- a/c64cast.schema.json
+++ b/c64cast.schema.json
@@ -785,7 +785,7 @@
               "number",
               "null"
             ],
-            "description": "Seconds before auto-advance. Unset = scene-type default. Video scenes reject this (they run until the file ends). For launcher this is the idle timeout (reset by player input)."
+            "description": "Seconds before auto-advance; 0 = run forever. Unset = scene-type default (webcam/blank run forever when they're the only scene, else 30s; waveform = song length or 30s; slideshow/generative = 30s). Video scenes reject this (they run until the file ends). For launcher this is the idle timeout (reset by player input)."
           },
           "target_fps": {
             "type": [
diff --git a/c64cast/config.py b/c64cast/config.py
index 55320cb..e85cf37 100644
--- a/c64cast/config.py
+++ b/c64cast/config.py
@@ -795,12 +795,17 @@ class SceneCfg:
         default=None,
         metadata={"help": "Display name (shown in interstitials/logs; ensemble match key)."},
     )
-    # None = scene-type default (30s for webcam/blank, songlengths-or-30s for
-    # waveform/midi). Video scenes reject any value (video-driven).
+    # None = scene-type default: webcam/blank run forever in a single-scene
+    # playlist (else 30s so a rotation still advances), songlengths-or-30s for
+    # waveform/midi, 30s for slideshow/generative. 0 = run forever (any type).
+    # Video scenes reject any value (video-driven).
     duration_s: float | None = field(
         default=None,
         metadata={
-            "help": "Seconds before auto-advance. Unset = scene-type default. "
+            "help": "Seconds before auto-advance; 0 = run forever. Unset = "
+            "scene-type default (webcam/blank run forever when they're the "
+            "only scene, else 30s; waveform = song length or 30s; "
+            "slideshow/generative = 30s). "
             "Video scenes reject this (they run until the file ends). "
             "For launcher this is the idle timeout (reset by player input).",
             "applies_to": (
@@ -2982,6 +2987,11 @@ def validate_scene_cfg(s: SceneCfg, cfg: Config, *, audio_enabled: bool) -> None
             "Remove the field (it would be a silent no-op here)."
         )
 
+    # duration_s = 0 is the "run forever" sentinel; negatives are a typo.
+    # (Video rejects any duration_s below in _validate_video.)
+    if s.duration_s is not None and s.duration_s < 0:
+        raise ValueError(f"duration_s must be >= 0 (0 = run forever), got {s.duration_s!r}")
+
     if s.type == "webcam":
         mode = _display_mode_for_scene(s.display, s, cfg)
     elif s.type == "blank":
@@ -3468,11 +3478,27 @@ def build_scene(
             buffered_player=s.asid_buffered_player,
             name=s.name or "ASID",
         )
-    # Video scenes set their own video-driven duration in __init__
-    # (math.inf) — `validate_scene_cfg` above already rejected any explicit
-    # duration_s on them, so honor that by not overwriting it here.
-    if s.duration_s is not None and s.type != "video":
-        scene.duration_s = s.duration_s
+    # Duration resolution. `scene.duration_s = math.inf` means "run until
+    # stopped" (the scene never auto-advances).
+    #   * explicit duration_s == 0 → the "run forever" sentinel (any type);
+    #   * explicit duration_s  > 0 → honored verbatim;
+    #   * unset (None): webcam/blank default to infinite in a SINGLE-scene
+    #     playlist ("leave the camera running"), but keep the base 30 s in a
+    #     multi-scene playlist so the rotation still advances — an infinite
+    #     live scene never becomes is_done and would wedge the playlist. Every
+    #     other type keeps the default already set above (video's video-driven
+    #     math.inf, waveform's song-length, etc.).
+    # Video scenes set their own math.inf in __init__ and reject explicit
+    # duration_s in _validate_video, so leave them untouched here.
+    if s.type != "video":
+        # A single configured scene stays single-scene: interleave_videos is
+        # skipped for a 1-scene playlist (see scenes_from_config), so the
+        # scene count is the whole story here.
+        single_scene_playlist = len(cfg.scenes) <= 1
+        if s.duration_s is not None:
+            scene.duration_s = math.inf if s.duration_s == 0 else s.duration_s
+        elif s.type in ("webcam", "blank") and single_scene_playlist:
+            scene.duration_s = math.inf
     if s.target_fps is not None:
         scene.target_fps = float(s.target_fps)
     # Per-scene pixel effect (validated frame-bearing in validate_scene_cfg).
@@ -3568,7 +3594,8 @@ def scenes_from_config(
                 api, None, HiresDisplayMode(style="edges"), source, cfg.audio, "Live Hi-Res Edges"
             )
         )
-        base[-1].duration_s = 30.0
+        # The sole scene when nothing is configured — leave it running.
+        base[-1].duration_s = math.inf
 
     if not cfg.playlist.interleave_videos:
         return base
diff --git a/c64cast/scenes.py b/c64cast/scenes.py
index 3c6cbd7..37a0c28 100644
--- a/c64cast/scenes.py
+++ b/c64cast/scenes.py
@@ -283,9 +283,11 @@ def setup(self) -> None:
         fps_str = f"{self.target_fps:.0f}fps" if self.target_fps else "auto-fps"
         overlay_names = [getattr(ov, "name", type(ov).__name__) for ov in self.overlays]
         ov_str = ", ".join(overlay_names) if overlay_names else "no overlays"
-        # VideoScene sets duration_s = math.inf because its lifetime
-        # is video-driven; format that distinctly instead of "inf.0s".
-        dur_str = "video-driven" if math.isinf(self.duration_s) else f"{self.duration_s:.1f}s"
+        # duration_s = math.inf means the scene never auto-advances: a video
+        # (lifetime is video-driven, ends on EOF) or a single-scene
+        # webcam/blank (runs until stopped). Format that distinctly instead
+        # of "inf.0s".
+        dur_str = "unbounded" if math.isinf(self.duration_s) else f"{self.duration_s:.1f}s"
         log.info(
             "scene %r: mode=%s duration=%s %s [%s]", self.name, mode_name, dur_str, fps_str, ov_str
         )
diff --git a/config/c64cast.example.toml b/config/c64cast.example.toml
index d63bc46..21d75a5 100644
--- a/config/c64cast.example.toml
+++ b/config/c64cast.example.toml
@@ -607,8 +607,10 @@ profile_interval = 10.0             # seconds between profiler summary lines
 # Scenes run in declaration order; each plays for `duration_s` seconds,
 # then advances. Common fields:
 #   type        — "webcam" | "video" | "waveform" | "midi"
-#   duration_s  — seconds before advancing. Rejected on `video`
-#                 scenes (they run until the video file ends).
+#   duration_s  — seconds before advancing; 0 = run forever. Unset =
+#                 scene-type default (webcam/blank run forever when they're
+#                 the only scene, else 30s). Rejected on `video` scenes
+#                 (they run until the video file ends).
 #   name        — display name (used by the interstitial). Optional.
 #   target_fps  — per-scene FPS cap. Default: system rate
 #                 (60 NTSC / 50 PAL). Bitmap (hires/mhires) video, webcam,
diff --git a/docs/architecture.md b/docs/architecture.md
index 8e8ffa8..3eb7250 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -125,7 +125,7 @@ Mechanism (`modes.py`): `setup()` zeroes both VIC banks' bitmap+screen, pins ban
 
 ### `scenes.py` — Scene state machine
 
-Playlist calls `setup()` → `process_frame()` * → `teardown()` for each scene, with an interstitial scene (built by an injected `interstitial_factory`) between them. `WebcamScene` is tuned for low latency: each camera frame is pushed straight through, no delay buffer. Audio follows the global `[audio].enabled` flag — when on, the webcam/blank scene picks up the AudioStreamer automatically; a per-scene `audio = false` opts back out (useful for muting one segment in an otherwise audible playlist). When attached it runs uncorrelated to the video — sync between mic and display isn't preserved. Backpressure is handled at the Playlist layer via deadline-based frame-dropping rather than a per-scene queue check — the DMA socket's TCP send buffer absorbs short bursts, and missed frames get dropped at the deadline instead of bursting to catch up. `VideoScene` is the opposite: it uses the audio playback position as the master clock and picks the closest video frame against it, so A/V never drift. Its lifetime is video-driven — `process_frame` returns False once the source reports `finished`, and `__init__` pins `self.duration_s = math.inf` so the base-class duration timer can't truncate playback. `config.validate_scene_cfg` rejects any user-supplied `duration_s` on a video cfg (the field would either be a silent no-op or a truncation footgun); `Scene.setup` formats the inf as `duration=video-driven` in the startup log.
+Playlist calls `setup()` → `process_frame()` * → `teardown()` for each scene, with an interstitial scene (built by an injected `interstitial_factory`) between them. `WebcamScene` is tuned for low latency: each camera frame is pushed straight through, no delay buffer. Its duration default is context-dependent (resolved in `config.build_scene`): webcam/blank with an unset `duration_s` default to `math.inf` ("leave the camera running") **only when they're the sole scene** in a single-scene playlist — in a multi-scene playlist they keep the base 30 s, because an infinite live scene never becomes `is_done` and would wedge the rotation. An explicit `duration_s = 0` is the universal "run forever" sentinel (mapped to `math.inf` for any non-video type); `< 0` is rejected at load. Audio follows the global `[audio].enabled` flag — when on, the webcam/blank scene picks up the AudioStreamer automatically; a per-scene `audio = false` opts back out (useful for muting one segment in an otherwise audible playlist). When attached it runs uncorrelated to the video — sync between mic and display isn't preserved. Backpressure is handled at the Playlist layer via deadline-based frame-dropping rather than a per-scene queue check — the DMA socket's TCP send buffer absorbs short bursts, and missed frames get dropped at the deadline instead of bursting to catch up. `VideoScene` is the opposite: it uses the audio playback position as the master clock and picks the closest video frame against it, so A/V never drift. Its lifetime is video-driven — `process_frame` returns False once the source reports `finished`, and `__init__` pins `self.duration_s = math.inf` so the base-class duration timer can't truncate playback. `config.validate_scene_cfg` rejects any user-supplied `duration_s` on a video cfg (the field would either be a silent no-op or a truncation footgun); `Scene.setup` formats the inf as `duration=video-driven` in the startup log.
 
 Each Scene also carries an `overlays: list[Overlay]`. The Playlist runs every overlay's `setup`/`process_frame`/`teardown` around the scene's lifecycle — scene paints first, overlays paint on top (in declaration order).
 
diff --git a/tests/test_config.py b/tests/test_config.py
index f83ae4f..a96561b 100644
--- a/tests/test_config.py
+++ b/tests/test_config.py
@@ -4,6 +4,7 @@
 from __future__ import annotations
 
 import argparse
+import math
 import os
 import tempfile
 import unittest
@@ -464,6 +465,16 @@ def test_valid_blank_scene_passes(self):
         s = cfgmod.SceneCfg(type="blank")
         cfgmod.validate_scene_cfg(s, self._cfg(), audio_enabled=False)
 
+    def test_negative_duration_rejected(self):
+        s = cfgmod.SceneCfg(type="blank", duration_s=-1.0)
+        with self.assertRaisesRegex(ValueError, "duration_s must be >= 0"):
+            cfgmod.validate_scene_cfg(s, self._cfg(), audio_enabled=False)
+
+    def test_zero_duration_allowed(self):
+        # 0 is the "run forever" sentinel, not an error.
+        s = cfgmod.SceneCfg(type="blank", duration_s=0)
+        cfgmod.validate_scene_cfg(s, self._cfg(), audio_enabled=False)
+
     def test_blank_scene_rejects_wrong_display(self):
         s = cfgmod.SceneCfg(type="blank", display="mhires")
         with self.assertRaisesRegex(ValueError, "blank scene must use"):
@@ -939,6 +950,65 @@ def test_pre_emphasis_defaults_to_none_auto(self):
         self.assertIsNone(scene.pre_emphasis)
 
 
+class SceneDurationDefaultTest(unittest.TestCase):
+    """build_scene's duration resolution: webcam/blank default to infinite
+    in a single-scene playlist ("leave the camera running"), keep 30 s in a
+    multi-scene playlist (so the rotation still advances), and treat
+    duration_s = 0 as a universal "run forever" sentinel."""
+
+    def setUp(self):
+        import os
+        import sys
+        from typing import cast
+
+        from c64cast.api import Ultimate64API
+        from c64cast.video import WebcamSource
+
+        sys.path.insert(0, os.path.join(os.path.dirname(__file__)))
+        from _fakes import FakeAPI
+
+        self.api = cast(Ultimate64API, FakeAPI())
+        self.source = cast(WebcamSource, object())
+
+    def _cfg(self, *scenes: cfgmod.SceneCfg) -> cfgmod.Config:
+        cfg = cfgmod.Config()
+        cfg.scenes = list(scenes)
+        return cfg
+
+    def _build(self, cfg: cfgmod.Config, s: cfgmod.SceneCfg):
+        return cfgmod.build_scene(s, cfg, self.api, None, self.source)
+
+    def test_single_scene_webcam_unset_is_infinite(self):
+        s = cfgmod.SceneCfg(type="webcam", display="petscii")
+        scene = self._build(self._cfg(s), s)
+        self.assertTrue(math.isinf(scene.duration_s))
+
+    def test_single_scene_blank_unset_is_infinite(self):
+        s = cfgmod.SceneCfg(type="blank")
+        scene = self._build(self._cfg(s), s)
+        self.assertTrue(math.isinf(scene.duration_s))
+
+    def test_multi_scene_webcam_unset_stays_30s(self):
+        # Two scenes → rotation; an infinite live scene would wedge it, so
+        # the webcam keeps the finite base default and advances.
+        s = cfgmod.SceneCfg(type="webcam", display="petscii")
+        other = cfgmod.SceneCfg(type="blank")
+        scene = self._build(self._cfg(s, other), s)
+        self.assertEqual(scene.duration_s, 30.0)
+
+    def test_zero_is_run_forever_sentinel_even_in_multi_scene(self):
+        # Explicit 0 overrides the finite multi-scene default.
+        s = cfgmod.SceneCfg(type="webcam", display="petscii", duration_s=0)
+        other = cfgmod.SceneCfg(type="blank")
+        scene = self._build(self._cfg(s, other), s)
+        self.assertTrue(math.isinf(scene.duration_s))
+
+    def test_positive_duration_honored(self):
+        s = cfgmod.SceneCfg(type="webcam", display="petscii", duration_s=45.0)
+        scene = self._build(self._cfg(s), s)
+        self.assertEqual(scene.duration_s, 45.0)
+
+
 class FollowerOnlyRotationFilterTest(unittest.TestCase):
     """scenes_from_config skips follower_only scenes — they're available
     for follower-override lookup via cfg.scenes but must never reach the