homesec 1.2.1__py3-none-any.whl → 1.2.3__py3-none-any.whl

homesec/sources/rtsp/core.py

@@ -0,0 +1,1424 @@
+"""RTSP motion-detecting clip source.
+
+Ported from motion_recorder.py to keep functionality within homesec.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import random
+import subprocess
+import time
+from datetime import datetime, timedelta
+from enum import Enum
+from pathlib import Path
+
+import numpy as np
+import numpy.typing as npt
+from pydantic import BaseModel, Field, model_validator
+
+from homesec.models.clip import Clip
+from homesec.sources.base import ThreadedClipSource
+from homesec.sources.rtsp.clock import Clock, SystemClock
+from homesec.sources.rtsp.frame_pipeline import FfmpegFramePipeline, FramePipeline
+from homesec.sources.rtsp.hardware import HardwareAccelConfig, HardwareAccelDetector
+from homesec.sources.rtsp.motion import MotionDetector
+from homesec.sources.rtsp.recorder import FfmpegRecorder, Recorder
+from homesec.sources.rtsp.utils import (
+    _format_cmd,
+    _is_timeout_option_error,
+    _next_backoff,
+    _redact_rtsp_url,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class RTSPMotionConfig(BaseModel):
+    """Motion detection configuration."""
+
+    model_config = {"extra": "forbid"}
+
+    pixel_threshold: int = Field(
+        default=45,
+        ge=0,
+        description="Pixel intensity delta required to count a pixel as changed.",
+    )
+    min_changed_pct: float = Field(
+        default=1.0,
+        ge=0.0,
+        description="Percent of pixels that must change to trigger motion (idle state).",
+    )
+    recording_sensitivity_factor: float = Field(
+        default=2.0,
+        ge=1.0,
+        description="Factor to reduce the threshold while recording (>=1.0).",
+    )
+    blur_kernel: int = Field(
+        default=5,
+        ge=0,
+        description="Gaussian blur kernel size (odd or zero; even values are normalized).",
+    )
+
+
+class RTSPRecordingConfig(BaseModel):
+    """Recording lifecycle configuration."""
+
+    model_config = {"extra": "forbid"}
+
+    stop_delay: float = Field(
+        default=10.0,
+        ge=0.0,
+        description="Seconds to keep recording after motion stops.",
+    )
+    max_recording_s: float = Field(
+        default=60.0,
+        gt=0.0,
+        description="Maximum seconds per recording before rotating.",
+    )
+
+
+class RTSPStreamConfig(BaseModel):
+    """RTSP/ffmpeg transport configuration."""
+
+    model_config = {"extra": "forbid"}
+
+    connect_timeout_s: float = Field(
+        default=2.0,
+        ge=0.0,
+        description="RTSP connect timeout (seconds) passed to ffmpeg/ffprobe when supported.",
+    )
+    io_timeout_s: float = Field(
+        default=2.0,
+        ge=0.0,
+        description="RTSP I/O timeout (seconds) passed to ffmpeg/ffprobe when supported.",
+    )
+    ffmpeg_flags: list[str] = Field(
+        default_factory=list,
+        description="Additional ffmpeg flags appended to the command.",
+    )
+    disable_hwaccel: bool = Field(
+        default=False,
+        description="Disable hardware-accelerated decoding.",
+    )
+
+
+class RTSPReconnectConfig(BaseModel):
+    """Reconnect and fallback policy."""
+
+    model_config = {"extra": "forbid"}
+
+    max_attempts: int = Field(
+        default=0,
+        ge=0,
+        description="Max reconnect attempts (0 = retry forever).",
+    )
+    backoff_s: float = Field(
+        default=1.0,
+        ge=0.0,
+        description="Base backoff (seconds) between reconnect attempts.",
+    )
+    detect_fallback_attempts: int = Field(
+        default=3,
+        ge=0,
+        description="Failures before falling back from detect stream to main stream.",
+    )
+
+
+class RTSPRuntimeConfig(BaseModel):
+    """Runtime loop configuration."""
+
+    model_config = {"extra": "forbid"}
+
+    frame_timeout_s: float = Field(
+        default=2.0,
+        ge=0.0,
+        description="Seconds without frames before considering the pipeline stalled.",
+    )
+    frame_queue_size: int = Field(
+        default=20,
+        ge=1,
+        description="Frame queue size used by the frame reader thread.",
+    )
+    heartbeat_s: float = Field(
+        default=30.0,
+        ge=0.0,
+        description="Seconds between heartbeat logs.",
+    )
+    debug_motion: bool = Field(
+        default=False,
+        description="Enable verbose motion detection logging.",
+    )
+
+
+class RTSPSourceConfig(BaseModel):
+    """RTSP source configuration."""
+
+    model_config = {"extra": "forbid"}
+
+    camera_name: str | None = Field(
+        default=None,
+        description="Optional human-friendly camera name.",
+    )
+    rtsp_url_env: str | None = Field(
+        default=None,
+        description="Environment variable containing the RTSP URL.",
+    )
+    rtsp_url: str | None = Field(
+        default=None,
+        description="RTSP URL for the main stream.",
+    )
+    detect_rtsp_url_env: str | None = Field(
+        default=None,
+        description="Environment variable containing the detect stream RTSP URL.",
+    )
+    detect_rtsp_url: str | None = Field(
+        default=None,
+        description="RTSP URL for the detect stream.",
+    )
+    output_dir: str = Field(
+        default="./recordings",
+        description="Directory to store recordings and logs.",
+    )
+
+    motion: RTSPMotionConfig = Field(default_factory=RTSPMotionConfig)
+    recording: RTSPRecordingConfig = Field(default_factory=RTSPRecordingConfig)
+    stream: RTSPStreamConfig = Field(default_factory=RTSPStreamConfig)
+    reconnect: RTSPReconnectConfig = Field(default_factory=RTSPReconnectConfig)
+    runtime: RTSPRuntimeConfig = Field(default_factory=RTSPRuntimeConfig)
+
+    @model_validator(mode="after")
+    def _require_rtsp_url(self) -> RTSPSourceConfig:
+        if not (self.rtsp_url or self.rtsp_url_env):
+            raise ValueError("rtsp_url_env or rtsp_url required for RTSP source")
+        return self
+
+
+class RTSPRunState(str, Enum):
+    IDLE = "idle"
+    RECORDING = "recording"
+    STALLED = "stalled"
+    RECONNECTING = "reconnecting"
+
+
+class RTSPSource(ThreadedClipSource):
+    """RTSP clip source with motion detection.
+
+    Uses ffmpeg for frame extraction and recording; detects motion from
+    downscaled grayscale frames and emits clips when recordings finish.
+    """
+
+    def __init__(
+        self,
+        config: RTSPSourceConfig,
+        camera_name: str,
+        *,
+        frame_pipeline: FramePipeline | None = None,
+        recorder: Recorder | None = None,
+        clock: Clock | None = None,
+    ) -> None:
+        """Initialize RTSP source."""
+        super().__init__()
+        self._clock = clock or SystemClock()
+        rtsp_url = config.rtsp_url
+        if config.rtsp_url_env:
+            env_value = os.getenv(config.rtsp_url_env)
+            if env_value:
+                rtsp_url = env_value
+        if not rtsp_url:
+            raise ValueError("rtsp_url_env or rtsp_url required for RTSP source")
+
+        self.rtsp_url = rtsp_url
+
+        detect_rtsp_url = config.detect_rtsp_url
+        if config.detect_rtsp_url_env:
+            env_value = os.getenv(config.detect_rtsp_url_env)
+            if env_value:
+                detect_rtsp_url = env_value
+
+        derived_detect = self._derive_detect_rtsp_url(self.rtsp_url)
+        if detect_rtsp_url:
+            self.detect_rtsp_url = detect_rtsp_url
+            self._detect_rtsp_url_source = "explicit"
+        elif derived_detect:
+            self.detect_rtsp_url = derived_detect
+            self._detect_rtsp_url_source = "derived_subtype=1"
+        else:
+            self.detect_rtsp_url = self.rtsp_url
+            self._detect_rtsp_url_source = "same_as_rtsp_url"
+
+        self.output_dir = Path(config.output_dir)
+        sanitized_name = self._sanitize_camera_name(camera_name)
+        self.camera_name = sanitized_name or camera_name
+
+        self.pixel_threshold = int(config.motion.pixel_threshold)
+        self.min_changed_pct = float(config.motion.min_changed_pct)
+        self.recording_sensitivity_factor = float(config.motion.recording_sensitivity_factor)
+        self.blur_kernel = self._normalize_blur_kernel(config.motion.blur_kernel)
+        self.motion_stop_delay = float(config.recording.stop_delay)
+        self.max_recording_s = float(config.recording.max_recording_s)
+        self.max_reconnect_attempts = int(config.reconnect.max_attempts)
+        self.detect_fallback_attempts = int(config.reconnect.detect_fallback_attempts)
+        self.debug_motion = bool(config.runtime.debug_motion)
+        self.heartbeat_s = float(config.runtime.heartbeat_s)
+        self.frame_timeout_s = float(config.runtime.frame_timeout_s)
+        self.frame_queue_size = int(config.runtime.frame_queue_size)
+        self.reconnect_backoff_s = float(config.reconnect.backoff_s)
+        self.rtsp_connect_timeout_s = float(config.stream.connect_timeout_s)
+        self.rtsp_io_timeout_s = float(config.stream.io_timeout_s)
+        self.ffmpeg_flags = list(config.stream.ffmpeg_flags)
+
+        if config.stream.disable_hwaccel:
+            logger.info("Hardware acceleration manually disabled")
+            self.hwaccel_config = HardwareAccelConfig(hwaccel=None)
+            self._hwaccel_failed = True
+        else:
+            logger.info("Testing hardware acceleration with camera stream...")
+            self.hwaccel_config = HardwareAccelDetector.detect(self.detect_rtsp_url)
+            self._hwaccel_failed = not self.hwaccel_config.is_available
+
+        self.output_dir.mkdir(parents=True, exist_ok=True)
+
+        self.recording_process: subprocess.Popen[bytes] | None = None
+        self.last_motion_time: float | None = None
+        self.recording_start_time: float | None = None
+        self.recording_start_wall: datetime | None = None
+        self.output_file: Path | None = None
+        self._stderr_log: Path | None = None
+        self._recording_id: str | None = None
+        self._recording_restart_backoff_s = 0.0
+        self._recording_restart_backoff_max_s = 10.0
+        self._recording_restart_base_s = 0.5
+        self._recording_next_restart_at: float | None = None
+        self._stall_grace_until: float | None = None
+        self._motion_detector = MotionDetector(
+            pixel_threshold=self.pixel_threshold,
+            min_changed_pct=self.min_changed_pct,
+            blur_kernel=self.blur_kernel,
+            debug=self.debug_motion,
+        )
+
+        self._frame_pipeline: FramePipeline = frame_pipeline or FfmpegFramePipeline(
+            output_dir=self.output_dir,
+            frame_queue_size=self.frame_queue_size,
+            rtsp_connect_timeout_s=self.rtsp_connect_timeout_s,
+            rtsp_io_timeout_s=self.rtsp_io_timeout_s,
+            ffmpeg_flags=self.ffmpeg_flags,
+            hwaccel_config=self.hwaccel_config,
+            hwaccel_failed=self._hwaccel_failed,
+            on_frame=self._touch_heartbeat,
+            clock=self._clock,
+        )
+        self._recorder: Recorder = recorder or FfmpegRecorder(
+            rtsp_url=self.rtsp_url,
+            ffmpeg_flags=self.ffmpeg_flags,
+            rtsp_connect_timeout_s=self.rtsp_connect_timeout_s,
+            rtsp_io_timeout_s=self.rtsp_io_timeout_s,
+            clock=self._clock,
+        )
+        self._run_state = RTSPRunState.IDLE
+        self._motion_rtsp_url = self.detect_rtsp_url
+        self._detect_stream_available = self.detect_rtsp_url != self.rtsp_url
+        self._detect_fallback_active = False
+        self._detect_fallback_deferred = False
+        self._detect_failure_count = 0
+        self._detect_next_probe_at: float | None = None
+        self._detect_probe_interval_s = 25.0
+        self._detect_probe_backoff_s = self._detect_probe_interval_s
+        self._detect_probe_backoff_max_s = 60.0
+        self.reconnect_count = 0
+        self.last_successful_frame = self._clock.now()
+
+        logger.info(
+            "RTSPSource initialized: camera=%s, output_dir=%s",
+            self.camera_name,
+            self.output_dir,
+        )
+
+    def is_healthy(self) -> bool:
+        """Check if source is healthy."""
+        if not self._thread_is_healthy():
+            return False
+
+        age = self._clock.now() - self.last_successful_frame
+        return age < (self.frame_timeout_s * 3)
+
+    def _touch_heartbeat(self) -> None:
+        self.last_successful_frame = self._clock.now()
+        super()._touch_heartbeat()
+
+    def _stop_timeout(self) -> float:
+        return 10.0
+
+    def _on_start(self) -> None:
+        logger.info("Starting RTSPSource: %s", self.camera_name)
+
+    def _on_stop(self) -> None:
+        logger.info("Stopping RTSPSource...")
+
+    def _on_stopped(self) -> None:
+        logger.info("RTSPSource stopped")
+
+    def _derive_detect_rtsp_url(self, rtsp_url: str) -> str | None:
+        if "subtype=0" in rtsp_url:
+            return rtsp_url.replace("subtype=0", "subtype=1")
+        return None
+
+    def _sanitize_camera_name(self, name: str | None) -> str | None:
+        if not name:
+            return None
+        raw = str(name).strip()
+        if not raw:
+            return None
+        out: list[str] = []
+        for ch in raw:
+            if ch.isalnum() or ch in ("-", "_"):
+                out.append(ch)
+            elif ch.isspace():
+                out.append("_")
+            else:
+                out.append("_")
+        cleaned = "".join(out).strip("_")
+        while "__" in cleaned:
+            cleaned = cleaned.replace("__", "_")
+        return cleaned or None
+
+    def _recording_prefix(self) -> str:
+        if not self.camera_name:
+            return ""
+        return f"{self.camera_name}_"
+
+    def _make_recording_paths(self, timestamp: str) -> tuple[Path, Path]:
+        prefix = self._recording_prefix()
+        output_file = self.output_dir / f"{prefix}motion_{timestamp}.mp4"
+        stderr_log = self.output_dir / f"{prefix}recording_{timestamp}.log"
+        return output_file, stderr_log
+
+    def _set_recording_state(
+        self,
+        *,
+        proc: subprocess.Popen[bytes],
+        output_file: Path,
+        stderr_log: Path,
+        start_mono: float,
+        start_wall: datetime,
+    ) -> None:
+        self.recording_process = proc
+        self.output_file = output_file
+        self._stderr_log = stderr_log
+        self.recording_start_time = start_mono
+        self.recording_start_wall = start_wall
+        self._recording_id = output_file.name
+
+    def _clear_recording_state(
+        self,
+    ) -> tuple[
+        subprocess.Popen[bytes] | None,
+        Path | None,
+        float | None,
+        datetime | None,
+    ]:
+        proc = self.recording_process
+        output_file = self.output_file
+        start_mono = self.recording_start_time
+        start_wall = self.recording_start_wall
+        self.recording_process = None
+        self.output_file = None
+        self._stderr_log = None
+        self.recording_start_time = None
+        self.recording_start_wall = None
+        self._recording_id = None
+        return proc, output_file, start_mono, start_wall
+
+    def _telemetry_common_fields(self) -> dict[str, object]:
+        return {
+            "pixel_threshold": self.pixel_threshold,
+            "min_changed_pct": self.min_changed_pct,
+            "recording_sensitivity_factor": self.recording_sensitivity_factor,
+            "blur_kernel": self.blur_kernel,
+            "stop_delay": self.motion_stop_delay,
+            "max_recording_s": self.max_recording_s,
+            "detect_fallback_attempts": self.detect_fallback_attempts,
+            "hwaccel": self.hwaccel_config.hwaccel,
+            "hwaccel_device": self.hwaccel_config.hwaccel_device,
+        }
+
+    def _event_extra(self, event_type: str, **fields: object) -> dict[str, object]:
+        extra: dict[str, object] = {
+            "kind": "event",
+            "event_type": event_type,
+            "camera_name": self.camera_name,
+            "recording_id": self._recording_id,
+        }
+        extra.update(self._telemetry_common_fields())
+        extra.update(fields)
+        return extra
+
+    def _probe_stream_info(
+        self,
+        rtsp_url: str,
+        *,
+        timeout_s: float = 10.0,
+    ) -> dict[str, object] | None:
+        base_cmd = [
+            "ffprobe",
+            "-v",
+            "error",
+            "-select_streams",
+            "v:0",
+            "-show_entries",
+            "stream=codec_name,width,height,avg_frame_rate",
+            "-of",
+            "json",
+            "-rtsp_transport",
+            "tcp",
+        ]
+        timeout_args: list[str] = []
+        if self.rtsp_connect_timeout_s > 0:
+            timeout_us_connect = str(int(max(0.1, self.rtsp_connect_timeout_s) * 1_000_000))
+            timeout_args.extend(["-stimeout", timeout_us_connect])
+        if self.rtsp_io_timeout_s > 0:
+            timeout_us_io = str(int(max(0.1, self.rtsp_io_timeout_s) * 1_000_000))
+            timeout_args.extend(["-rw_timeout", timeout_us_io])
+
+        attempts: list[tuple[str, list[str]]] = []
+        if timeout_args:
+            attempts.append(("timeouts", base_cmd + timeout_args + [rtsp_url]))
+        attempts.append(("no_timeouts" if timeout_args else "default", base_cmd + [rtsp_url]))
+
+        for label, cmd in attempts:
+            try:
+                result = subprocess.run(
+                    cmd,
+                    capture_output=True,
+                    text=True,
+                    timeout=timeout_s,
+                    check=False,
+                )
+                if result.returncode != 0:
+                    if _is_timeout_option_error(result.stderr):
+                        logger.debug("ffprobe missing timeout options (%s), retrying", label)
+                        continue
+                    logger.debug("ffprobe failed (%s) with exit code %s", label, result.returncode)
+                    break
+
+                data = json.loads(result.stdout)
+                if not data.get("streams"):
+                    continue
+                stream = data["streams"][0]
+                return {
+                    "codec_name": stream.get("codec_name"),
+                    "width": stream.get("width"),
+                    "height": stream.get("height"),
+                    "avg_frame_rate": stream.get("avg_frame_rate"),
+                }
+            except Exception as exc:
+                logger.warning("Failed to probe stream info: %s", exc, exc_info=True)
+                return None
+
+        return None
+
+    def _redact_rtsp_url(self, url: str) -> str:
+        return _redact_rtsp_url(url)
+
+    def _format_cmd(self, cmd: list[str]) -> str:
+        return _format_cmd(cmd)
+
+    def detect_motion(
+        self, frame: npt.NDArray[np.uint8], *, threshold: float | None = None
+    ) -> bool:
+        """Return True if motion detected in frame."""
+        return self._motion_detector.detect(frame, threshold=threshold)
+
+    def check_recording_health(self) -> bool:
+        """Check if recording process is still alive."""
+        if self.recording_process and not self._recorder.is_alive(self.recording_process):
+            proc = self.recording_process
+            output_file = self.output_file
+            exit_code = getattr(proc, "returncode", None)
+            logger.warning("Recording process died unexpectedly (exit code: %s)", exit_code)
+            if output_file:
+                logger.error(
+                    "Recording process died",
+                    extra=self._event_extra(
+                        "recording_process_died",
+                        recording_id=output_file.name,
+                        recording_path=str(output_file),
+                        exit_code=exit_code,
+                        pid=proc.pid,
+                    ),
+                )
+
+            log_file: Path | None = None
+            if self._stderr_log and self._stderr_log.exists():
+                log_file = self._stderr_log
+            elif output_file:
+                stem = output_file.stem
+                if "motion_" in stem:
+                    prefix_part, timestamp = stem.split("motion_", 1)
+                    candidate = output_file.parent / f"{prefix_part}recording_{timestamp}.log"
+                    if candidate.exists():
+                        log_file = candidate
+
+            if log_file:
+                try:
+                    with open(log_file) as f:
+                        error_lines = f.read()
+                        if error_lines:
+                            logger.warning("Recording error log:\n%s", error_lines[-1000:])
+                except Exception as e:
+                    logger.warning("Could not read log file: %s", e, exc_info=True)
+
+            cleared_proc, cleared_output, start_mono, start_wall = self._clear_recording_state()
+            if cleared_proc is not None:
+                self._stop_recording_process(cleared_proc, cleared_output)
+            self._finalize_clip(cleared_output, start_wall, start_mono)
+            return False
+        return True
+
+    def start_recording(self) -> None:
+        """Start ffmpeg recording process with audio."""
+        if self.recording_process:
+            return
+        now = self._clock.now()
+        if not self._recording_backoff_ready(now):
+            return
+
+        self.output_dir.mkdir(parents=True, exist_ok=True)
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        output_file, stderr_log = self._make_recording_paths(timestamp)
+
+        proc = self._recorder.start(output_file, stderr_log)
+        if not proc:
+            self._bump_recording_backoff(now)
+            logger.error(
+                "Recording failed to start",
+                extra=self._event_extra(
+                    "recording_start_error",
+                    recording_id=output_file.name,
+                    recording_path=str(output_file),
+                    stderr_log=str(stderr_log),
+                ),
+            )
+            return
+
+        self._set_recording_state(
+            proc=proc,
+            output_file=output_file,
+            stderr_log=stderr_log,
+            start_mono=self._clock.now(),
+            start_wall=datetime.now(),
+        )
+        self._reset_recording_backoff()
+
+        logger.info("Started recording: %s (PID: %s)", output_file, proc.pid)
+        logger.debug("Recording logs: %s", stderr_log)
+        logger.info(
+            "Recording started",
+            extra=self._event_extra(
+                "recording_start",
+                recording_id=output_file.name,
+                recording_path=str(output_file),
+                stderr_log=str(stderr_log),
+                pid=proc.pid,
+                detect_stream_source=getattr(self, "_detect_rtsp_url_source", None),
+                detect_stream_is_same=(self.detect_rtsp_url == self.rtsp_url),
+            ),
+        )
+
+    def stop_recording(self) -> None:
+        """Stop ffmpeg recording process."""
+        if not self.recording_process:
+            return
+
+        proc, output_file, started_at, started_wall = self._clear_recording_state()
+        if proc is not None:
+            self._stop_recording_process(proc, output_file)
+        self._finalize_clip(output_file, started_wall, started_at)
+
+        if output_file:
+            duration_s = (self._clock.now() - started_at) if started_at else None
+            logger.info(
+                "Recording stopped",
+                extra=self._event_extra(
+                    "recording_stop",
+                    recording_id=output_file.name,
+                    recording_path=str(output_file),
+                    duration_s=duration_s,
+                    last_changed_pct=self._motion_detector.last_changed_pct,
+                    last_changed_pixels=self._motion_detector.last_changed_pixels,
+                ),
+            )
+
+    def _stop_recording_process(
+        self, proc: subprocess.Popen[bytes], output_file: Path | None
+    ) -> None:
+        try:
+            self._recorder.stop(proc, output_file)
+        except Exception:
+            logger.exception("Failed while stopping recording process (PID: %s)", proc.pid)
+
+    def _rotate_recording_if_needed(self) -> None:
+        if self.recording_process is None or self.recording_start_time is None:
+            return
+
+        now = self._clock.now()
+        if (now - self.recording_start_time) < self.max_recording_s:
+            return
+
+        if self.last_motion_time is None:
+            return
+
+        if (now - self.last_motion_time) > self.motion_stop_delay:
+            return
+
+        old_proc = self.recording_process
+        old_output = self.output_file
+        old_started_wall = self.recording_start_wall
+        old_started_mono = self.recording_start_time
+
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        new_output, new_log = self._make_recording_paths(timestamp)
+
+        logger.info(
+            "Max recording length reached (%.1fs), rotating to: %s",
+            self.max_recording_s,
+            new_output,
+        )
+
+        new_proc = self._recorder.start(new_output, new_log)
+        if new_proc:
+            self.last_motion_time = now
+            self._set_recording_state(
+                proc=new_proc,
+                output_file=new_output,
+                stderr_log=new_log,
+                start_mono=now,
+                start_wall=datetime.now(),
+            )
+
+            if old_output:
+                logger.info(
+                    "Recording rotated",
+                    extra=self._event_extra(
+                        "recording_rotate",
+                        recording_id=old_output.name,
+                        recording_path=str(old_output),
+                        new_recording_id=new_output.name,
+                        new_recording_path=str(new_output),
+                    ),
+                )
+
+            self._stop_recording_process(old_proc, old_output)
+            self._finalize_clip(old_output, old_started_wall, old_started_mono)
+            return
+
+        logger.warning("Rotation start failed; stopping current recording and retrying")
+        proc, output_file, started_mono, started_wall = self._clear_recording_state()
+        if proc is not None:
+            self._stop_recording_process(proc, output_file)
+        self._finalize_clip(output_file, started_wall, started_mono)
+        self.start_recording()
+
+    def _start_frame_pipeline(self) -> None:
+        try:
+            self._frame_pipeline.stop()
+        except Exception:
+            logger.exception("Error stopping frame pipeline before start")
+        self._frame_pipeline.start(self._motion_rtsp_url)
+        self._motion_detector.reset()
+
+    def _stop_frame_pipeline(self) -> None:
+        self._frame_pipeline.stop()
+        self._motion_detector.reset()
+
+    def _wait_for_first_frame(self, timeout_s: float) -> bool:
+        return self._frame_pipeline.read_frame(timeout_s) is not None
+
+    def _reconnect_frame_pipeline(self, *, aggressive: bool) -> bool:
+        initial_backoff_s = 0.2 if aggressive else float(self.reconnect_backoff_s)
+        backoff_s = initial_backoff_s
+        backoff_cap_s = 10.0 if aggressive else float(self.reconnect_backoff_s)
+
+        first_attempt = True
+        while not self._stop_event.is_set():
+            self.reconnect_count += 1
+
+            if self.max_reconnect_attempts == 0:
+                logger.warning(
+                    "Reconnect attempt %s (mode=%s max=inf)...",
+                    self.reconnect_count,
+                    "aggressive" if aggressive else "normal",
+                    extra=self._event_extra(
+                        "rtsp_reconnect_attempt",
+                        attempt=self.reconnect_count,
+                        max_attempts=None,
+                        mode="aggressive" if aggressive else "normal",
+                        detect_is_fallback=self._detect_fallback_active,
+                        motion_rtsp_url=self._redact_rtsp_url(self._motion_rtsp_url),
+                    ),
+                )
+            else:
+                logger.warning(
+                    "Reconnect attempt %s/%s (mode=%s)...",
+                    self.reconnect_count,
+                    self.max_reconnect_attempts,
+                    "aggressive" if aggressive else "normal",
+                    extra=self._event_extra(
+                        "rtsp_reconnect_attempt",
+                        attempt=self.reconnect_count,
+                        max_attempts=self.max_reconnect_attempts,
+                        mode="aggressive" if aggressive else "normal",
+                        detect_is_fallback=self._detect_fallback_active,
+                        motion_rtsp_url=self._redact_rtsp_url(self._motion_rtsp_url),
+                    ),
+                )
+
+            sleep_s = self._reconnect_sleep_s(
+                backoff_s=backoff_s,
+                aggressive=aggressive,
+                first_attempt=first_attempt,
+            )
+            first_attempt = False
+
+            if sleep_s > 0:
+                logger.debug("Waiting %.2fs before reconnect...", sleep_s)
+                self._clock.sleep(sleep_s)
+
+            try:
+                self._start_frame_pipeline()
+            except Exception as exc:
+                logger.warning("Failed to restart frame pipeline: %s", exc, exc_info=True)
+                triggered_fallback = self._note_detect_failure(self._clock.now())
+                if self._reconnect_exhausted(aggressive=aggressive):
+                    return False
+                if triggered_fallback:
+                    backoff_s = initial_backoff_s
+                    first_attempt = True
+                    continue
+                backoff_s = self._bump_reconnect_backoff(
+                    backoff_s,
+                    backoff_cap_s,
+                    aggressive=aggressive,
+                )
+                continue
+
+            startup_timeout_s = min(2.0, max(0.5, float(self.frame_timeout_s)))
+            if self._wait_for_first_frame(startup_timeout_s):
+                logger.info(
+                    "Reconnected successfully",
+                    extra=self._event_extra(
+                        "rtsp_reconnect_success",
+                        attempt=self.reconnect_count,
+                        mode="aggressive" if aggressive else "normal",
+                        detect_is_fallback=self._detect_fallback_active,
+                        motion_rtsp_url=self._redact_rtsp_url(self._motion_rtsp_url),
+                    ),
+                )
+                self.reconnect_count = 0
+                self._detect_failure_count = 0
+                self._detect_fallback_deferred = False
+                return True
+
+            logger.warning("Frame pipeline restarted but still no frames; retrying...")
+            self._note_detect_failure(self._clock.now())
+            try:
+                self._stop_frame_pipeline()
+            except Exception as exc:
+                logger.warning(
+                    "Failed to stop frame pipeline after reconnect: %s",
+                    exc,
+                    exc_info=True,
+                )
+            if self._reconnect_exhausted(aggressive=aggressive):
+                return False
+            backoff_s = self._bump_reconnect_backoff(
+                backoff_s,
+                backoff_cap_s,
+                aggressive=aggressive,
+            )
+
+        return False
+
+    def _reconnect_sleep_s(
+        self,
+        *,
+        backoff_s: float,
+        aggressive: bool,
+        first_attempt: bool,
+    ) -> float:
+        if first_attempt and aggressive:
+            return 0.0
+        jitter = random.uniform(0.0, min(0.25, backoff_s * 0.25))
+        return backoff_s + jitter
+
+    def _bump_reconnect_backoff(
+        self,
+        backoff_s: float,
+        backoff_cap_s: float,
+        *,
+        aggressive: bool,
+    ) -> float:
+        if not aggressive:
+            return backoff_s
+        return _next_backoff(backoff_s, backoff_cap_s)
+
+    def _recording_backoff_ready(self, now: float) -> bool:
+        next_retry = self._recording_next_restart_at
+        if next_retry is None or now >= next_retry:
+            return True
+        remaining = next_retry - now
+        logger.debug("Recording restart backoff active (%.2fs remaining)", remaining)
+        return False
+
+    def _bump_recording_backoff(self, now: float) -> None:
+        if self._recording_restart_backoff_s <= 0:
+            self._recording_restart_backoff_s = self._recording_restart_base_s
+        else:
+            self._recording_restart_backoff_s = _next_backoff(
+                self._recording_restart_backoff_s,
+                self._recording_restart_backoff_max_s,
+            )
+        self._recording_next_restart_at = now + self._recording_restart_backoff_s
+
+    def _reset_recording_backoff(self) -> None:
+        self._recording_restart_backoff_s = 0.0
+        self._recording_next_restart_at = None
+
+    def _reconnect_exhausted(self, *, aggressive: bool) -> bool:
+        if self.max_reconnect_attempts <= 0:
+            return False
+        if self.reconnect_count < self.max_reconnect_attempts:
+            return False
+        logger.error(
+            "Max reconnect attempts reached. Camera may be offline.",
+            extra=self._event_extra(
+                "rtsp_reconnect_failed",
+                attempt=self.reconnect_count,
+                max_attempts=self.max_reconnect_attempts,
+                mode="aggressive" if aggressive else "normal",
+                detect_is_fallback=self._detect_fallback_active,
+                motion_rtsp_url=self._redact_rtsp_url(self._motion_rtsp_url),
+            ),
+        )
+        return True
+
+    def _note_detect_failure(self, now: float) -> bool:
+        if not self._detect_stream_available:
+            return False
+        if self._motion_rtsp_url != self.detect_rtsp_url:
+            return False
+        if self._detect_fallback_active:
+            return False
+        if self.detect_fallback_attempts <= 0:
+            return False
+
+        self._detect_failure_count = min(
+            self._detect_failure_count + 1,
+            self.detect_fallback_attempts,
+        )
+        if self._detect_failure_count < self.detect_fallback_attempts:
+            return False
+
+        if self.recording_process is not None:
+            if not self._detect_fallback_deferred:
+                logger.warning(
+                    "Detect fallback deferred while recording",
+                    extra=self._event_extra(
+                        "detect_fallback_deferred",
+                        detect_stream_source=getattr(self, "_detect_rtsp_url_source", None),
+                        detect_stream_is_same=False,
+                    ),
+                )
+                self._detect_fallback_deferred = True
+            return False
+
+        self._activate_detect_fallback(now)
+        return True
+
+    def _activate_detect_fallback(self, now: float) -> None:
+        self._detect_fallback_active = True
+        self._detect_fallback_deferred = False
+        self._detect_failure_count = 0
+        self._motion_rtsp_url = self.rtsp_url
+        self._detect_next_probe_at = now + self._detect_probe_interval_s
+        self._detect_probe_backoff_s = self._detect_probe_interval_s
+        logger.warning(
+            "Detect stream failed; falling back to main RTSP stream",
+            extra=self._event_extra(
+                "detect_fallback_enabled",
+                detect_stream_source=getattr(self, "_detect_rtsp_url_source", None),
+                detect_stream_is_same=False,
+            ),
+        )
+
+    def _schedule_detect_probe(self, now: float) -> None:
+        self._detect_next_probe_at = now + self._detect_probe_backoff_s
+        self._detect_probe_backoff_s = _next_backoff(
+            self._detect_probe_backoff_s,
+            self._detect_probe_backoff_max_s,
+        )
+
+    def _maybe_recover_detect_stream(self, now: float) -> bool:
+        if not self._detect_fallback_active:
+            return False
+        if not self._detect_stream_available:
+            return False
+        if self._detect_next_probe_at is not None and now < self._detect_next_probe_at:
+            return False
+
+        info = self._probe_stream_info(self.detect_rtsp_url, timeout_s=3.0)
+        if not info or not info.get("width") or not info.get("height"):
+            self._schedule_detect_probe(now)
+            return False
+
+        logger.info(
+            "Detect stream recovered; attempting switch back",
+            extra=self._event_extra(
+                "detect_fallback_recovering",
+                detect_stream_source=getattr(self, "_detect_rtsp_url_source", None),
+                detect_stream_is_same=False,
+            ),
+        )
+
+        self._motion_rtsp_url = self.detect_rtsp_url
+        try:
+            self._start_frame_pipeline()
+            startup_timeout_s = min(2.0, max(0.5, float(self.frame_timeout_s)))
+            if self._wait_for_first_frame(startup_timeout_s):
+                self._detect_fallback_active = False
+                self._detect_next_probe_at = None
+                self._detect_probe_backoff_s = self._detect_probe_interval_s
+                logger.info(
+                    "Detect stream restored; using detect stream again",
+                    extra=self._event_extra(
+                        "detect_fallback_recovered",
+                        detect_stream_source=getattr(self, "_detect_rtsp_url_source", None),
+                        detect_stream_is_same=False,
+                    ),
+                )
+                return True
+        except Exception as exc:
+            logger.warning("Detect stream switch failed: %s", exc, exc_info=True)
+
+        self._motion_rtsp_url = self.rtsp_url
+        self._detect_fallback_active = True
+        self._schedule_detect_probe(now)
+        try:
+            self._start_frame_pipeline()
+        except Exception as exc:
+            logger.warning(
+                "Failed to restart fallback pipeline after detect switch: %s",
+                exc,
+                exc_info=True,
+            )
+        return False
+
+    def _recording_threshold(self) -> float:
+        if self.recording_sensitivity_factor <= 0:
+            return self.min_changed_pct
+        return max(0.0, self.min_changed_pct / self.recording_sensitivity_factor)
+
+    def _set_run_state(self, state: RTSPRunState) -> None:
+        if self._run_state == state:
+            return
+        logger.debug("RTSP state change: %s -> %s", self._run_state, state)
+        self._run_state = state
+
+    def _ensure_recording(self, now: float) -> None:
+        if self.recording_process:
+            if not self.check_recording_health():
+                logger.warning(
+                    "Recording died, attempting restart",
+                    extra=self._event_extra(
+                        "recording_restart_attempt",
+                        reason="health_check",
+                    ),
+                )
+                if self.last_motion_time is not None and (
+                    (now - self.last_motion_time) <= self.motion_stop_delay
+                ):
+                    self.start_recording()
+                    if self.recording_process:
+                        logger.info(
+                            "Recording restarted",
+                            extra=self._event_extra(
+                                "recording_restart",
+                                reason="health_check",
+                            ),
+                        )
+                else:
+                    self.last_motion_time = None
+            return
+
+        if (
+            self.last_motion_time is not None
+            and (now - self.last_motion_time) <= self.motion_stop_delay
+        ):
+            logger.warning(
+                "Recording missing while motion is recent; restarting",
+                extra=self._event_extra(
+                    "recording_restart_attempt",
+                    reason="missing",
+                ),
+            )
+            self.start_recording()
+            if self.recording_process:
+                logger.info(
+                    "Recording restarted",
+                    extra=self._event_extra(
+                        "recording_restart",
+                        reason="missing",
+                    ),
+                )
+        elif (
+            self.last_motion_time is not None
+            and (now - self.last_motion_time) > self.motion_stop_delay
+        ):
+            self.last_motion_time = None
+
+    def cleanup(self) -> None:
+        """Clean up resources."""
+        logger.info("Cleaning up...")
+        try:
+            self.stop_recording()
+        except Exception:
+            logger.exception("Error stopping recording")
+        try:
+            self._stop_frame_pipeline()
+        except Exception:
+            logger.exception("Error stopping frame pipeline")
+
+    def _finalize_clip(
+        self,
+        output_file: Path | None,
+        started_wall: datetime | None,
+        started_mono: float | None,
+    ) -> None:
+        if output_file is None:
+            return
+        try:
+            if not output_file.exists() or output_file.stat().st_size == 0:
+                return
+        except Exception as exc:
+            logger.warning(
+                "Failed to stat output clip %s: %s",
+                output_file,
+                exc,
+                exc_info=True,
+            )
+            return
+
+        end_ts = datetime.now()
+        if started_wall is None:
+            if started_mono is not None:
+                duration_s = time.monotonic() - started_mono
+                start_ts = end_ts - timedelta(seconds=duration_s)
+            else:
+                start_ts = end_ts
+                duration_s = 0.0
+        else:
+            start_ts = started_wall
+            duration_s = (end_ts - start_ts).total_seconds()
+
+        clip = Clip(
+            clip_id=output_file.stem,
+            camera_name=self.camera_name,
+            local_path=output_file,
+            start_ts=start_ts,
+            end_ts=end_ts,
+            duration_s=duration_s,
+            source_backend="rtsp",
+        )
+
+        self._emit_clip(clip)
+
+    @staticmethod
+    def _normalize_blur_kernel(blur_kernel: int) -> int:
+        kernel = int(blur_kernel)
+        if kernel < 0:
+            return 0
+        if kernel % 2 == 0 and kernel != 0:
+            return kernel + 1
+        return kernel
+
+    def _log_startup_info(self) -> None:
+        logger.info("Connecting to camera...")
+        logger.info(
+            "Motion: pixel_threshold=%s min_changed_pct=%.3f%% blur=%s stop_delay=%.1fs max_recording_s=%.1fs",
+            self.pixel_threshold,
+            self.min_changed_pct,
+            self.blur_kernel,
+            self.motion_stop_delay,
+            self.max_recording_s,
+        )
+        logger.info(
+            "Motion thresholds: idle=%.3f%% recording=%.3f%% (recording_sensitivity_factor=%.2f)",
+            self.min_changed_pct,
+            self._recording_threshold(),
+            self.recording_sensitivity_factor,
+        )
+        logger.info("Max reconnect attempts: %s", self.max_reconnect_attempts)
+        if self.max_reconnect_attempts == 0:
+            logger.info("Reconnect policy: retry forever")
+        if self._detect_stream_available:
+            logger.info("Detect fallback attempts: %s", self.detect_fallback_attempts)
+
+        if self.hwaccel_config.is_available:
+            device_info = (
+                f" (device: {self.hwaccel_config.hwaccel_device})"
+                if self.hwaccel_config.hwaccel_device
+                else ""
+            )
+            logger.info("Hardware acceleration: %s%s", self.hwaccel_config.hwaccel, device_info)
+        else:
+            logger.info("Hardware acceleration: disabled (using software decoding)")
+
+        logger.info("Detecting camera resolution...")
+        info = self._probe_stream_info(self.rtsp_url)
+        width = info.get("width") if info else None
+        height = info.get("height") if info else None
+        if isinstance(width, (int, str)) and isinstance(height, (int, str)):
+            native_width = int(width)
+            native_height = int(height)
+        else:
+            logger.warning("Failed to probe stream, using default 1920x1080")
+            native_width, native_height = 1920, 1080
+        logger.info("Camera resolution: %sx%s", native_width, native_height)
+
+        if self.detect_rtsp_url != self.rtsp_url:
+            info = self._probe_stream_info(self.detect_rtsp_url)
+            if info and info.get("width") and info.get("height"):
+                logger.info(
+                    "Motion RTSP stream (%s) available: %s (%sx%s @ %s)",
+                    self._detect_rtsp_url_source,
+                    self._redact_rtsp_url(self.detect_rtsp_url),
+                    info.get("width"),
+                    info.get("height"),
+                    info.get("avg_frame_rate"),
+                )
+            else:
+                if self.detect_fallback_attempts > 0:
+                    logger.warning(
+                        "Motion RTSP stream (%s) did not probe cleanly; falling back to main RTSP stream",
+                        self._detect_rtsp_url_source,
+                    )
+                    self._activate_detect_fallback(self._clock.now())
+                else:
+                    logger.warning(
+                        "Motion RTSP stream (%s) did not probe cleanly; fallback disabled",
+                        self._detect_rtsp_url_source,
+                    )
+
+        logger.info("Motion detection: 320x240@10fps (downscaled for efficiency)")
+
+    def _handle_frame_timeout(self) -> bool:
+        pipe_status = self._frame_pipeline.exit_code()
+        if pipe_status is not None:
+            logger.error(
+                "Frame pipeline exited (code: %s). Check logs: %s/frame_pipeline.log",
+                pipe_status,
+                self.output_dir,
+            )
+        else:
+            logger.warning(
+                "No frames received for %.1fs (stall). Last frame %.1fs ago.",
+                self.frame_timeout_s,
+                self._clock.now() - self.last_successful_frame,
+            )
+
+        aggressive = self.recording_process is None
+        return self._reconnect_frame_pipeline(aggressive=aggressive)
+
+    def _stall_grace_remaining(self, now: float) -> float:
+        if not self.recording_process or self.last_motion_time is None:
+            return 0.0
+        return max(0.0, self.motion_stop_delay - (now - self.last_motion_time))
+
+    def _enter_stalled(self, now: float, remaining: float) -> None:
+        self._stall_grace_until = now + remaining
+        logger.warning(
+            "Frame pipeline stalled; keeping recording alive for %.1fs",
+            remaining,
+        )
+
+    def _handle_stalled_wait(self, now: float) -> bool:
+        if self._stall_grace_until is None or now >= self._stall_grace_until:
+            return False
+        self._set_run_state(RTSPRunState.STALLED)
+        self._clock.sleep(min(0.5, self._stall_grace_until - now))
+        return True
+
+    def _handle_reconnect_needed(self, now: float) -> tuple[bool, bool]:
+        self._set_run_state(RTSPRunState.RECONNECTING)
+        if self._handle_frame_timeout():
+            return True, False
+        remaining = self._stall_grace_remaining(now)
+        if remaining > 0:
+            self._enter_stalled(now, remaining)
+            self._handle_stalled_wait(now)
+            return True, True
+        return False, False
+
+    def _handle_heartbeat(
+        self,
+        now: float,
+        frame_count: int,
+        last_heartbeat: float,
+    ) -> tuple[bool, bool, float]:
+        if now - last_heartbeat <= self.heartbeat_s:
+            return True, False, last_heartbeat
+
+        logger.debug(
+            "[HEARTBEAT] Processed %s frames, recording=%s",
+            frame_count,
+            self.recording_process is not None,
+        )
+        if not self._frame_pipeline.is_running():
+            logger.error(
+                "Frame pipeline died! Exit code: %s",
+                self._frame_pipeline.exit_code(),
+            )
+            ok, stalled = self._handle_reconnect_needed(now)
+            if not ok:
+                return False, False, last_heartbeat
+            return True, stalled, now
+
+        return True, False, now
+
+    def _handle_missing_dimensions(self, now: float) -> bool:
+        logger.warning("Frame pipeline missing dimensions; reconnecting")
+        ok, _ = self._handle_reconnect_needed(now)
+        return ok
+
+    def _handle_missing_frame(self, now: float) -> bool:
+        if self._handle_stalled_wait(now):
+            return True
+        ok, _ = self._handle_reconnect_needed(now)
+        return ok
+
+    def _handle_motion_detected(self, now: float, frame_count: int) -> None:
+        logger.debug(
+            "[MOTION DETECTED at frame %s] changed_pct=%.3f%%",
+            frame_count,
+            self._motion_detector.last_changed_pct,
+        )
+        self.last_motion_time = now
+        if not self.recording_process:
+            logger.debug("-> Starting new recording...")
+            self.start_recording()
+            if not self.recording_process:
+                logger.error("-> Recording failed to start!")
+
+    def _update_recording_state(self, now: float) -> None:
+        if self.recording_process and self.last_motion_time is not None:
+            if now - self.last_motion_time > self.motion_stop_delay:
+                logger.info(
+                    "No motion for %.1fs > stop_delay=%.1fs (last changed_pct=%.3f%%), stopping",
+                    now - self.last_motion_time,
+                    self.motion_stop_delay,
+                    self._motion_detector.last_changed_pct,
+                )
+                self.stop_recording()
+                self.last_motion_time = None
+                self._set_run_state(RTSPRunState.IDLE)
+                return
+
+            self._rotate_recording_if_needed()
+            self._set_run_state(RTSPRunState.RECORDING)
+            return
+
+        self._set_run_state(RTSPRunState.IDLE)
+
+    def _process_frame(
+        self,
+        frame: npt.NDArray[np.uint8],
+        now: float,
+        frame_count: int,
+    ) -> None:
+        threshold = (
+            self._recording_threshold()
+            if self.recording_process is not None
+            else self.min_changed_pct
+        )
+        motion_detected = self.detect_motion(frame, threshold=threshold)
+
+        if motion_detected:
+            self._handle_motion_detected(now, frame_count)
+
+        if self.recording_process or self.last_motion_time is not None:
+            self._ensure_recording(now)
+
+        self._update_recording_state(now)
+
+    # State flow (simplified):
+    # Idle -> Recording (motion)
+    # Recording -> Stalled (frame timeout) -> Reconnecting -> Recording/Idle
+    # Recording -> Idle (no motion for stop_delay)
+    def _run(self) -> None:
+        """Start monitoring camera for motion and recording videos."""
+        self._log_startup_info()
+
+        try:
+            try:
+                self._start_frame_pipeline()
+            except Exception as exc:
+                logger.warning("Initial frame pipeline start failed: %s", exc, exc_info=True)
+                if not self._reconnect_frame_pipeline(aggressive=True):
+                    return
+
+            logger.info("Connected, monitoring for motion...")
+            self.reconnect_count = 0
+            self._set_run_state(RTSPRunState.IDLE)
+
+            frame_count = 0
+            last_heartbeat = self._clock.now()
+
+            while not self._stop_event.is_set():
+                now = self._clock.now()
+                if self._handle_stalled_wait(now):
+                    continue
+
+                ok, stalled, last_heartbeat = self._handle_heartbeat(
+                    now,
+                    frame_count,
+                    last_heartbeat,
+                )
+                if not ok:
+                    break
+                if stalled:
+                    continue
+
+                if self._maybe_recover_detect_stream(now):
+                    self._set_run_state(RTSPRunState.RECONNECTING)
+                    continue
+
+                frame_count += 1
+
+                raw_frame = self._frame_pipeline.read_frame(timeout_s=self.frame_timeout_s)
+                if raw_frame is None:
+                    now = self._clock.now()
+                    if self._handle_missing_frame(now):
+                        continue
+                    break
+
+                self._stall_grace_until = None
+                frame_width = self._frame_pipeline.frame_width
+                frame_height = self._frame_pipeline.frame_height
+                if frame_width is None or frame_height is None:
+                    if not self._handle_missing_dimensions(now):
+                        break
+                    continue
+
+                frame = np.frombuffer(raw_frame, dtype=np.uint8).reshape(
+                    (frame_height, frame_width)
+                )
+                now = self._clock.now()
+                self._process_frame(frame, now, frame_count)
+
+        except Exception as e:
+            logger.exception("Unexpected error: %s", e)
+        finally:
+            self.cleanup()