visionA/local-tool/server/internal/api/handlers/camera_handler.go

package handlers

import (
	"fmt"
	"io"
	"net/url"
	"os"
	"path/filepath"
	"strconv"
	"strings"
	"time"

	"visiona-local/server/internal/api/ws"
	"visiona-local/server/internal/camera"
	"visiona-local/server/internal/device"
	"visiona-local/server/internal/driver"
	"visiona-local/server/internal/inference"

	"github.com/gin-gonic/gin"
)

type CameraHandler struct {
	cameraMgr    *camera.Manager
	deviceMgr    *device.Manager
	inferenceSvc *inference.Service
	wsHub        *ws.Hub
	streamer     *camera.MJPEGStreamer
	pipeline     *camera.InferencePipeline
	activeSource camera.FrameSource
	sourceType   camera.SourceType

	// Video seek state — preserved across seek operations
	videoPath      string           // original file path or resolved URL
	videoIsURL     bool             // true if source is a URL
	videoFPS       float64          // target FPS
	videoInfo      camera.VideoInfo // duration, total frames
	activeDeviceID string           // device ID for current video session
}

func NewCameraHandler(
	cameraMgr *camera.Manager,
	deviceMgr *device.Manager,
	inferenceSvc *inference.Service,
	wsHub *ws.Hub,
) *CameraHandler {
	streamer := camera.NewMJPEGStreamer()
	go streamer.Run()
	return &CameraHandler{
		cameraMgr:    cameraMgr,
		deviceMgr:    deviceMgr,
		inferenceSvc: inferenceSvc,
		wsHub:        wsHub,
		streamer:     streamer,
	}
}

func (h *CameraHandler) ListCameras(c *gin.Context) {
	cameras := h.cameraMgr.ListCameras()
	c.JSON(200, gin.H{"success": true, "data": gin.H{"cameras": cameras}})
}

func (h *CameraHandler) StartPipeline(c *gin.Context) {
	var req struct {
		CameraID string `json:"cameraId"`
		DeviceID string `json:"deviceId"`
		Width    int    `json:"width"`
		Height   int    `json:"height"`
	}
	if err := c.ShouldBindJSON(&req); err != nil {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "BAD_REQUEST", "message": err.Error()}})
		return
	}
	if req.Width == 0 {
		req.Width = 640
	}
	if req.Height == 0 {
		req.Height = 480
	}

	// Clean up any existing pipeline
	h.stopActivePipeline()

	// Open camera
	if err := h.cameraMgr.Open(0, req.Width, req.Height); err != nil {
		c.JSON(500, gin.H{"success": false, "error": gin.H{"code": "CAMERA_OPEN_FAILED", "message": err.Error()}})
		return
	}

	// Get device driver
	session, err := h.deviceMgr.GetDevice(req.DeviceID)
	if err != nil {
		c.JSON(404, gin.H{"success": false, "error": gin.H{"code": "DEVICE_NOT_FOUND", "message": err.Error()}})
		return
	}

	// Create inference result channel
	resultCh := make(chan *driver.InferenceResult, 10)

	// Forward results to WebSocket, enriching with device ID
	go func() {
		room := "inference:" + req.DeviceID
		for result := range resultCh {
			result.DeviceID = req.DeviceID
			h.wsHub.BroadcastToRoom(room, result)
		}
	}()

	// Start pipeline with camera as source
	h.activeSource = h.cameraMgr
	h.sourceType = camera.SourceCamera
	h.pipeline = camera.NewInferencePipeline(
		h.cameraMgr,
		camera.SourceCamera,
		session.Driver,
		h.streamer.FrameChannel(),
		resultCh,
	)
	h.pipeline.Start()

	streamURL := "/api/camera/stream"
	c.JSON(200, gin.H{
		"success": true,
		"data": gin.H{
			"streamUrl":  streamURL,
			"sourceType": "camera",
		},
	})
}

func (h *CameraHandler) StopPipeline(c *gin.Context) {
	h.stopActivePipeline()
	c.JSON(200, gin.H{"success": true})
}

func (h *CameraHandler) StreamMJPEG(c *gin.Context) {
	h.streamer.ServeHTTP(c.Writer, c.Request)
}

// UploadImage handles image file upload for single-shot inference.
func (h *CameraHandler) UploadImage(c *gin.Context) {
	h.stopActivePipeline()

	deviceID := c.PostForm("deviceId")
	if deviceID == "" {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "BAD_REQUEST", "message": "deviceId is required"}})
		return
	}

	file, header, err := c.Request.FormFile("file")
	if err != nil {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "BAD_REQUEST", "message": "file is required"}})
		return
	}
	defer file.Close()

	ext := strings.ToLower(filepath.Ext(header.Filename))
	if ext != ".jpg" && ext != ".jpeg" && ext != ".png" {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "BAD_REQUEST", "message": "only JPG/PNG files are supported"}})
		return
	}

	// Save to temp file
	tmpFile, err := os.CreateTemp("", "edge-ai-image-*"+ext)
	if err != nil {
		c.JSON(500, gin.H{"success": false, "error": gin.H{"code": "STORAGE_ERROR", "message": err.Error()}})
		return
	}
	if _, err := io.Copy(tmpFile, file); err != nil {
		tmpFile.Close()
		os.Remove(tmpFile.Name())
		c.JSON(500, gin.H{"success": false, "error": gin.H{"code": "STORAGE_ERROR", "message": err.Error()}})
		return
	}
	tmpFile.Close()

	// Create ImageSource
	imgSource, err := camera.NewImageSource(tmpFile.Name())
	if err != nil {
		os.Remove(tmpFile.Name())
		c.JSON(500, gin.H{"success": false, "error": gin.H{"code": "IMAGE_DECODE_FAILED", "message": err.Error()}})
		return
	}

	// Get device driver
	session, err := h.deviceMgr.GetDevice(deviceID)
	if err != nil {
		imgSource.Close()
		c.JSON(404, gin.H{"success": false, "error": gin.H{"code": "DEVICE_NOT_FOUND", "message": err.Error()}})
		return
	}

	resultCh := make(chan *driver.InferenceResult, 10)

	go func() {
		room := "inference:" + deviceID
		for result := range resultCh {
			result.DeviceID = deviceID
			h.wsHub.BroadcastToRoom(room, result)
		}
	}()

	h.activeSource = imgSource
	h.sourceType = camera.SourceImage
	h.pipeline = camera.NewInferencePipeline(
		imgSource,
		camera.SourceImage,
		session.Driver,
		h.streamer.FrameChannel(),
		resultCh,
	)
	h.pipeline.Start()

	// Clean up result channel after pipeline completes
	go func() {
		<-h.pipeline.Done()
		close(resultCh)
	}()

	w, ht := imgSource.Dimensions()
	streamURL := "/api/camera/stream"
	c.JSON(200, gin.H{
		"success": true,
		"data": gin.H{
			"streamUrl":  streamURL,
			"sourceType": "image",
			"width":      w,
			"height":     ht,
			"filename":   header.Filename,
		},
	})
}

// UploadVideo handles video file upload for frame-by-frame inference.
func (h *CameraHandler) UploadVideo(c *gin.Context) {
	h.stopActivePipeline()

	deviceID := c.PostForm("deviceId")
	if deviceID == "" {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "BAD_REQUEST", "message": "deviceId is required"}})
		return
	}

	file, header, err := c.Request.FormFile("file")
	if err != nil {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "BAD_REQUEST", "message": "file is required"}})
		return
	}
	defer file.Close()

	ext := strings.ToLower(filepath.Ext(header.Filename))
	if ext != ".mp4" && ext != ".avi" && ext != ".mov" {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "BAD_REQUEST", "message": "only MP4/AVI/MOV files are supported"}})
		return
	}

	// Save to temp file
	tmpFile, err := os.CreateTemp("", "edge-ai-video-*"+ext)
	if err != nil {
		c.JSON(500, gin.H{"success": false, "error": gin.H{"code": "STORAGE_ERROR", "message": err.Error()}})
		return
	}
	if _, err := io.Copy(tmpFile, file); err != nil {
		tmpFile.Close()
		os.Remove(tmpFile.Name())
		c.JSON(500, gin.H{"success": false, "error": gin.H{"code": "STORAGE_ERROR", "message": err.Error()}})
		return
	}
	tmpFile.Close()

	// Probe video info (duration, frame count) before starting pipeline
	videoInfo := camera.ProbeVideoInfo(tmpFile.Name(), 15)

	// Create VideoSource
	videoSource, err := camera.NewVideoSource(tmpFile.Name(), 15)
	if err != nil {
		os.Remove(tmpFile.Name())
		c.JSON(500, gin.H{"success": false, "error": gin.H{"code": "VIDEO_DECODE_FAILED", "message": err.Error()}})
		return
	}
	if videoInfo.TotalFrames > 0 {
		videoSource.SetTotalFrames(videoInfo.TotalFrames)
	}

	// Get device driver
	session, err := h.deviceMgr.GetDevice(deviceID)
	if err != nil {
		videoSource.Close()
		c.JSON(404, gin.H{"success": false, "error": gin.H{"code": "DEVICE_NOT_FOUND", "message": err.Error()}})
		return
	}

	resultCh := make(chan *driver.InferenceResult, 10)

	go func() {
		room := "inference:" + deviceID
		for result := range resultCh {
			result.DeviceID = deviceID
			h.wsHub.BroadcastToRoom(room, result)
		}
	}()

	h.activeSource = videoSource
	h.sourceType = camera.SourceVideo
	h.videoPath = tmpFile.Name()
	h.videoIsURL = false
	h.videoFPS = 15
	h.videoInfo = videoInfo
	h.activeDeviceID = deviceID
	h.pipeline = camera.NewInferencePipeline(
		videoSource,
		camera.SourceVideo,
		session.Driver,
		h.streamer.FrameChannel(),
		resultCh,
	)
	h.pipeline.Start()

	// Notify frontend when video playback completes
	go func() {
		<-h.pipeline.Done()
		close(resultCh)
		h.wsHub.BroadcastToRoom("inference:"+deviceID, map[string]interface{}{
			"type":       "pipeline_complete",
			"sourceType": "video",
		})
	}()

	streamURL := "/api/camera/stream"
	c.JSON(200, gin.H{
		"success": true,
		"data": gin.H{
			"streamUrl":       streamURL,
			"sourceType":      "video",
			"filename":        header.Filename,
			"totalFrames":     videoInfo.TotalFrames,
			"durationSeconds": videoInfo.DurationSec,
		},
	})
}

// ytdlpHosts lists hostnames where yt-dlp should be used to resolve the actual
// video stream URL before passing to ffmpeg.
var ytdlpHosts = map[string]bool{
	"youtube.com": true, "www.youtube.com": true, "youtu.be": true, "m.youtube.com": true,
	"vimeo.com": true, "www.vimeo.com": true,
	"dailymotion.com": true, "www.dailymotion.com": true,
	"twitch.tv": true, "www.twitch.tv": true,
	"bilibili.com": true, "www.bilibili.com": true,
	"tiktok.com": true, "www.tiktok.com": true,
	"facebook.com": true, "www.facebook.com": true, "fb.watch": true,
	"instagram.com": true, "www.instagram.com": true,
	"twitter.com": true, "x.com": true,
}

type urlKind int

const (
	urlDirect urlKind = iota // direct video file or RTSP, pass to ffmpeg directly
	urlYTDLP                 // needs yt-dlp to resolve first
	urlBad                   // invalid or unsupported
)

// classifyVideoURL determines how to handle the given URL.
func classifyVideoURL(rawURL string) (urlKind, string) {
	parsed, err := url.Parse(rawURL)
	if err != nil {
		return urlBad, "Invalid URL format"
	}

	scheme := strings.ToLower(parsed.Scheme)
	host := strings.ToLower(parsed.Hostname())

	// RTSP streams — direct to ffmpeg
	if scheme == "rtsp" || scheme == "rtsps" {
		return urlDirect, ""
	}

	// Must be http or https
	if scheme != "http" && scheme != "https" {
		return urlBad, "Unsupported protocol: " + scheme + ". Use http, https, or rtsp."
	}

	// Known video platforms — use yt-dlp
	if ytdlpHosts[host] {
		return urlYTDLP, ""
	}

	// Everything else — pass directly to ffmpeg
	return urlDirect, ""
}

// StartFromURL handles video/stream inference from a URL (HTTP, HTTPS, RTSP).
func (h *CameraHandler) StartFromURL(c *gin.Context) {
	var req struct {
		URL      string `json:"url"`
		DeviceID string `json:"deviceId"`
	}
	if err := c.ShouldBindJSON(&req); err != nil {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "BAD_REQUEST", "message": err.Error()}})
		return
	}
	if req.URL == "" {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "BAD_REQUEST", "message": "url is required"}})
		return
	}
	if req.DeviceID == "" {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "BAD_REQUEST", "message": "deviceId is required"}})
		return
	}

	// Classify the URL
	kind, reason := classifyVideoURL(req.URL)
	if kind == urlBad {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "UNSUPPORTED_URL", "message": reason}})
		return
	}

	// For video platforms (YouTube, etc.), resolve actual stream URL via yt-dlp
	videoURL := req.URL
	if kind == urlYTDLP {
		resolved, err := camera.ResolveWithYTDLP(req.URL)
		if err != nil {
			c.JSON(500, gin.H{"success": false, "error": gin.H{"code": "URL_RESOLVE_FAILED", "message": "無法解析影片連結: " + err.Error()}})
			return
		}
		videoURL = resolved
	}

	h.stopActivePipeline()

	// Probe video info (duration, frame count) - may be slow for remote URLs
	videoInfo := camera.ProbeVideoInfo(videoURL, 15)

	// Create VideoSource from URL (ffmpeg supports HTTP/HTTPS/RTSP natively)
	videoSource, err := camera.NewVideoSourceFromURL(videoURL, 15)
	if err != nil {
		c.JSON(500, gin.H{"success": false, "error": gin.H{"code": "URL_OPEN_FAILED", "message": err.Error()}})
		return
	}
	if videoInfo.TotalFrames > 0 {
		videoSource.SetTotalFrames(videoInfo.TotalFrames)
	}

	// Get device driver
	session, err := h.deviceMgr.GetDevice(req.DeviceID)
	if err != nil {
		videoSource.Close()
		c.JSON(404, gin.H{"success": false, "error": gin.H{"code": "DEVICE_NOT_FOUND", "message": err.Error()}})
		return
	}

	resultCh := make(chan *driver.InferenceResult, 10)

	go func() {
		room := "inference:" + req.DeviceID
		for result := range resultCh {
			h.wsHub.BroadcastToRoom(room, result)
		}
	}()

	h.activeSource = videoSource
	h.sourceType = camera.SourceVideo
	h.videoPath = videoURL
	h.videoIsURL = true
	h.videoFPS = 15
	h.videoInfo = videoInfo
	h.activeDeviceID = req.DeviceID
	h.pipeline = camera.NewInferencePipeline(
		videoSource,
		camera.SourceVideo,
		session.Driver,
		h.streamer.FrameChannel(),
		resultCh,
	)
	h.pipeline.Start()

	go func() {
		<-h.pipeline.Done()
		close(resultCh)
		h.wsHub.BroadcastToRoom("inference:"+req.DeviceID, map[string]interface{}{
			"type":       "pipeline_complete",
			"sourceType": "video",
		})
	}()

	streamURL := "/api/camera/stream"
	c.JSON(200, gin.H{
		"success": true,
		"data": gin.H{
			"streamUrl":       streamURL,
			"sourceType":      "video",
			"filename":        req.URL,
			"totalFrames":     videoInfo.TotalFrames,
			"durationSeconds": videoInfo.DurationSec,
		},
	})
}

// UploadBatchImages handles multiple image files for sequential batch inference.
func (h *CameraHandler) UploadBatchImages(c *gin.Context) {
	h.stopActivePipeline()

	deviceID := c.PostForm("deviceId")
	if deviceID == "" {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "BAD_REQUEST", "message": "deviceId is required"}})
		return
	}

	form, err := c.MultipartForm()
	if err != nil {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "BAD_REQUEST", "message": "multipart form required"}})
		return
	}

	files := form.File["files"]
	if len(files) == 0 {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "BAD_REQUEST", "message": "at least one file is required"}})
		return
	}
	if len(files) > 50 {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "BAD_REQUEST", "message": "maximum 50 images per batch"}})
		return
	}

	// Save all files to temp
	filePaths := make([]string, 0, len(files))
	filenames := make([]string, 0, len(files))
	for _, fh := range files {
		ext := strings.ToLower(filepath.Ext(fh.Filename))
		if ext != ".jpg" && ext != ".jpeg" && ext != ".png" {
			for _, fp := range filePaths {
				os.Remove(fp)
			}
			c.JSON(400, gin.H{"success": false, "error": gin.H{
				"code":    "BAD_REQUEST",
				"message": fmt.Sprintf("unsupported file: %s (only JPG/PNG)", fh.Filename),
			}})
			return
		}

		f, openErr := fh.Open()
		if openErr != nil {
			for _, fp := range filePaths {
				os.Remove(fp)
			}
			c.JSON(500, gin.H{"success": false, "error": gin.H{"code": "STORAGE_ERROR", "message": openErr.Error()}})
			return
		}

		tmpFile, tmpErr := os.CreateTemp("", "edge-ai-batch-*"+ext)
		if tmpErr != nil {
			f.Close()
			for _, fp := range filePaths {
				os.Remove(fp)
			}
			c.JSON(500, gin.H{"success": false, "error": gin.H{"code": "STORAGE_ERROR", "message": tmpErr.Error()}})
			return
		}
		io.Copy(tmpFile, f)
		tmpFile.Close()
		f.Close()

		filePaths = append(filePaths, tmpFile.Name())
		filenames = append(filenames, fh.Filename)
	}

	// Create MultiImageSource
	batchSource, err := camera.NewMultiImageSource(filePaths, filenames)
	if err != nil {
		for _, fp := range filePaths {
			os.Remove(fp)
		}
		c.JSON(500, gin.H{"success": false, "error": gin.H{"code": "IMAGE_DECODE_FAILED", "message": err.Error()}})
		return
	}

	// Get device driver
	session, err := h.deviceMgr.GetDevice(deviceID)
	if err != nil {
		batchSource.Close()
		c.JSON(404, gin.H{"success": false, "error": gin.H{"code": "DEVICE_NOT_FOUND", "message": err.Error()}})
		return
	}

	batchID := fmt.Sprintf("batch-%d", time.Now().UnixNano())
	resultCh := make(chan *driver.InferenceResult, 10)

	go func() {
		room := "inference:" + deviceID
		for result := range resultCh {
			result.DeviceID = deviceID
			h.wsHub.BroadcastToRoom(room, result)
		}
	}()

	h.activeSource = batchSource
	h.sourceType = camera.SourceBatchImage
	h.pipeline = camera.NewInferencePipeline(
		batchSource,
		camera.SourceBatchImage,
		session.Driver,
		h.streamer.FrameChannel(),
		resultCh,
	)
	h.pipeline.Start()

	// Notify frontend when batch completes
	go func() {
		<-h.pipeline.Done()
		close(resultCh)
		h.wsHub.BroadcastToRoom("inference:"+deviceID, map[string]interface{}{
			"type":       "pipeline_complete",
			"sourceType": "batch_image",
			"batchId":    batchID,
		})
	}()

	// Build image list for response
	imageList := make([]gin.H, len(batchSource.Images()))
	for i, entry := range batchSource.Images() {
		imageList[i] = gin.H{
			"index":    i,
			"filename": entry.Filename,
			"width":    entry.Width,
			"height":   entry.Height,
		}
	}

	streamURL := "/api/camera/stream"
	c.JSON(200, gin.H{
		"success": true,
		"data": gin.H{
			"streamUrl":   streamURL,
			"sourceType":  "batch_image",
			"batchId":     batchID,
			"totalImages": len(files),
			"images":      imageList,
		},
	})
}

// GetBatchImageFrame serves a specific image from the active batch by index.
func (h *CameraHandler) GetBatchImageFrame(c *gin.Context) {
	if h.sourceType != camera.SourceBatchImage || h.activeSource == nil {
		c.JSON(404, gin.H{"success": false, "error": gin.H{"code": "NO_BATCH", "message": "no batch image source active"}})
		return
	}
	indexStr := c.Param("index")
	index, err := strconv.Atoi(indexStr)
	if err != nil || index < 0 {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "BAD_REQUEST", "message": "invalid index"}})
		return
	}
	mis, ok := h.activeSource.(*camera.MultiImageSource)
	if !ok {
		c.JSON(500, gin.H{"success": false, "error": gin.H{"code": "INTERNAL_ERROR", "message": "source type mismatch"}})
		return
	}
	jpegData, err := mis.GetImageByIndex(index)
	if err != nil {
		c.JSON(404, gin.H{"success": false, "error": gin.H{"code": "NOT_FOUND", "message": err.Error()}})
		return
	}
	c.Data(200, "image/jpeg", jpegData)
}

// stopPipelineForSeek stops the pipeline and ffmpeg process but keeps the video file.
func (h *CameraHandler) stopPipelineForSeek() {
	if h.pipeline != nil {
		h.pipeline.Stop()
		h.pipeline = nil
	}
	if h.activeSource != nil {
		if vs, ok := h.activeSource.(*camera.VideoSource); ok {
			vs.CloseWithoutRemove()
		}
	}
	h.activeSource = nil
}

// stopActivePipeline stops the current pipeline and cleans up resources.
func (h *CameraHandler) stopActivePipeline() {
	if h.pipeline != nil {
		h.pipeline.Stop()
		h.pipeline = nil
	}
	// Only close non-camera sources (camera is managed by cameraMgr)
	if h.activeSource != nil && h.sourceType != camera.SourceCamera {
		h.activeSource.Close()
	}
	if h.sourceType == camera.SourceCamera {
		h.cameraMgr.Close()
	}
	h.activeSource = nil
	h.sourceType = ""
	h.videoPath = ""
	h.videoIsURL = false
	h.activeDeviceID = ""
}

// SeekVideo seeks to a specific position in the current video and restarts inference.
func (h *CameraHandler) SeekVideo(c *gin.Context) {
	var req struct {
		TimeSeconds float64 `json:"timeSeconds"`
	}
	if err := c.ShouldBindJSON(&req); err != nil {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "BAD_REQUEST", "message": err.Error()}})
		return
	}

	if h.videoPath == "" || h.sourceType != camera.SourceVideo {
		c.JSON(400, gin.H{"success": false, "error": gin.H{"code": "NO_VIDEO", "message": "no video is currently playing"}})
		return
	}

	// Clamp seek time
	if req.TimeSeconds < 0 {
		req.TimeSeconds = 0
	}
	if h.videoInfo.DurationSec > 0 && req.TimeSeconds > h.videoInfo.DurationSec {
		req.TimeSeconds = h.videoInfo.DurationSec
	}

	// Stop current pipeline without deleting the video file
	h.stopPipelineForSeek()

	// Create new VideoSource with seek position
	var videoSource *camera.VideoSource
	var err error
	if h.videoIsURL {
		videoSource, err = camera.NewVideoSourceFromURLWithSeek(h.videoPath, h.videoFPS, req.TimeSeconds)
	} else {
		videoSource, err = camera.NewVideoSourceWithSeek(h.videoPath, h.videoFPS, req.TimeSeconds)
	}
	if err != nil {
		c.JSON(500, gin.H{"success": false, "error": gin.H{"code": "SEEK_FAILED", "message": err.Error()}})
		return
	}
	if h.videoInfo.TotalFrames > 0 {
		videoSource.SetTotalFrames(h.videoInfo.TotalFrames)
	}

	// Get device driver
	session, err := h.deviceMgr.GetDevice(h.activeDeviceID)
	if err != nil {
		videoSource.Close()
		c.JSON(404, gin.H{"success": false, "error": gin.H{"code": "DEVICE_NOT_FOUND", "message": err.Error()}})
		return
	}

	// Calculate frame offset from seek position
	frameOffset := int(req.TimeSeconds * h.videoFPS)

	resultCh := make(chan *driver.InferenceResult, 10)
	go func() {
		room := "inference:" + h.activeDeviceID
		for result := range resultCh {
			result.DeviceID = h.activeDeviceID
			h.wsHub.BroadcastToRoom(room, result)
		}
	}()

	h.activeSource = videoSource
	h.pipeline = camera.NewInferencePipelineWithOffset(
		videoSource,
		camera.SourceVideo,
		session.Driver,
		h.streamer.FrameChannel(),
		resultCh,
		frameOffset,
	)
	h.pipeline.Start()

	go func() {
		<-h.pipeline.Done()
		close(resultCh)
		h.wsHub.BroadcastToRoom("inference:"+h.activeDeviceID, map[string]interface{}{
			"type":       "pipeline_complete",
			"sourceType": "video",
		})
	}()

	c.JSON(200, gin.H{
		"success": true,
		"data": gin.H{
			"seekTo":      req.TimeSeconds,
			"frameOffset": frameOffset,
		},
	})
}