diff --git a/docs/plans/2026-04-19-comfyui-node-pack-implementation.md b/docs/plans/2026-04-19-comfyui-node-pack-implementation.md
new file mode 100644
index 0000000..e837ebd
--- /dev/null
+++ b/docs/plans/2026-04-19-comfyui-node-pack-implementation.md
@@ -0,0 +1,1041 @@
+# ComfyUI-8cut Node Pack Implementation Plan
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** Build a ComfyUI node pack that ports 8-cut's video scan/review/train/export workflow to a browser-based interface, using file paths instead of image tensors.
+
+**Architecture:** 5 nodes (LoadVideo, AudioScan, VideoReview, TrainModel, ExportClips) passing custom types (`VIDEO_PATH`, `SCAN_REGIONS`, `SCAN_MODEL`). The interactive VideoReview node uses ComfyUI's `ExecutionBlocker` for a two-pass flow: first pass sends data to the frontend widget, second pass (after user clicks Continue) passes edited regions downstream. The `8-cut/core/` package is reused unchanged.
+
+**Tech Stack:** ComfyUI (LiteGraph.js frontend, aiohttp server), Python 3.12, HTML5 `<video>`, Canvas API for timeline, existing 8-cut core (torch, librosa, scikit-learn, ffmpeg)
+
+**Design doc:** `docs/plans/2026-04-19-comfyui-node-pack-design.md`
+
+---
+
+### Task 1: Node pack skeleton and video serving
+
+**Files:**
+- Create: `ComfyUI-8cut/__init__.py`
+- Create: `ComfyUI-8cut/nodes/__init__.py`
+- Create: `ComfyUI-8cut/nodes/load_video.py`
+- Create: `ComfyUI-8cut/server_routes.py`
+- Symlink: `ComfyUI-8cut/core/` → `8-cut/core/`
+
+**Step 1: Create directory structure**
+
+```bash
+mkdir -p ComfyUI-8cut/nodes ComfyUI-8cut/data ComfyUI-8cut/models ComfyUI-8cut/web/js
+```
+
+**Step 2: Symlink core package**
+
+```bash
+ln -s /media/p5/8-cut/core ComfyUI-8cut/core
+```
+
+**Step 3: Create `server_routes.py` — video serving API**
+
+```python
+"""Custom API routes for ComfyUI-8cut."""
+
+import os
+import json
+from aiohttp import web
+
+import server as comfy_server
+
+routes = comfy_server.PromptServer.instance.routes
+
+
+@routes.get("/8cut/video")
+async def serve_video(request):
+    """Serve a video file for HTML5 <video> playback."""
+    path = request.rel_url.query.get("path", "")
+    if not path or not os.path.isfile(path):
+        return web.Response(status=404, text="File not found")
+    return web.FileResponse(path=path)
+```
+
+**Step 4: Create `nodes/load_video.py`**
+
+```python
+"""LoadVideo node — validates a video path and passes it downstream."""
+
+import os
+
+
+class LoadVideo:
+    """Load a video file by path for the 8-cut pipeline."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "video_path": ("STRING", {"default": "", "multiline": False}),
+            },
+        }
+
+    RETURN_TYPES = ("VIDEO_PATH", "STRING")
+    RETURN_NAMES = ("video", "filename")
+    FUNCTION = "run"
+    CATEGORY = "8cut"
+
+    def run(self, video_path):
+        if not video_path or not os.path.isfile(video_path):
+            raise ValueError(f"Video not found: {video_path}")
+        filename = os.path.basename(video_path)
+        return (video_path, filename)
+```
+
+**Step 5: Create `nodes/__init__.py`**
+
+```python
+from .load_video import LoadVideo
+
+NODE_CLASS_MAPPINGS = {
+    "8cut_LoadVideo": LoadVideo,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "8cut_LoadVideo": "Load Video (8cut)",
+}
+```
+
+**Step 6: Create top-level `__init__.py`**
+
+```python
+"""ComfyUI-8cut — tensor-free video scanning workflow."""
+
+from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS
+
+# Import server routes (registers API endpoints as side effect)
+from . import server_routes  # noqa: F401
+
+WEB_DIRECTORY = "./web"
+
+__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS", "WEB_DIRECTORY"]
+```
+
+**Step 7: Test manually**
+
+Install the node pack by symlinking into ComfyUI's `custom_nodes/`:
+
+```bash
+ln -s /path/to/ComfyUI-8cut /path/to/ComfyUI/custom_nodes/ComfyUI-8cut
+```
+
+Start ComfyUI, verify:
+- "Load Video (8cut)" appears in the node menu under "8cut" category
+- Node accepts a video path string and outputs VIDEO_PATH + filename
+- `GET /8cut/video?path=/path/to/video.mp4` serves the file
+
+**Step 8: Commit**
+
+```bash
+git add ComfyUI-8cut/
+git commit -m "feat: ComfyUI-8cut node pack skeleton with LoadVideo and video serving"
+```
+
+---
+
+### Task 2: AudioScan node
+
+**Files:**
+- Create: `ComfyUI-8cut/nodes/audio_scan.py`
+- Modify: `ComfyUI-8cut/nodes/__init__.py`
+
+**Step 1: Create `nodes/audio_scan.py`**
+
+```python
+"""AudioScan node — scan a video for target audio events."""
+
+import os
+import sys
+
+# Ensure core package is importable
+_pack_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+if _pack_dir not in sys.path:
+    sys.path.insert(0, _pack_dir)
+
+from core.audio_scan import scan_video, load_classifier, _EMBED_MODELS
+
+import server as comfy_server
+
+
+class AudioScan:
+    """Scan a video using a trained audio classifier."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        models_dir = os.path.join(
+            os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
+            "models",
+        )
+        model_files = []
+        if os.path.isdir(models_dir):
+            model_files = sorted(
+                f for f in os.listdir(models_dir) if f.endswith(".joblib")
+            )
+        return {
+            "required": {
+                "video": ("VIDEO_PATH",),
+                "model_file": (model_files if model_files else ["(none)"],),
+                "threshold": ("FLOAT", {
+                    "default": 0.5, "min": 0.0, "max": 1.0, "step": 0.01,
+                }),
+                "hop": ("FLOAT", {
+                    "default": 2.0, "min": 0.5, "max": 10.0, "step": 0.5,
+                }),
+            },
+            "hidden": {"unique_id": "UNIQUE_ID"},
+        }
+
+    RETURN_TYPES = ("SCAN_REGIONS",)
+    RETURN_NAMES = ("regions",)
+    FUNCTION = "run"
+    CATEGORY = "8cut"
+
+    def run(self, video, model_file, threshold, hop, unique_id=None):
+        if model_file == "(none)":
+            raise ValueError("No model selected. Train a model first.")
+
+        models_dir = os.path.join(
+            os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
+            "models",
+        )
+        model_path = os.path.join(models_dir, model_file)
+        model = load_classifier(model_path)
+
+        # Progress callback
+        prompt_server = comfy_server.PromptServer.instance
+
+        def progress_cb(current, total):
+            if unique_id is not None:
+                prompt_server.send_sync("progress", {
+                    "value": current, "max": total,
+                    "node": unique_id,
+                })
+
+        regions = scan_video(
+            video, model, threshold=threshold, hop=hop,
+            progress_cb=progress_cb,
+        )
+
+        # Convert to list of dicts for SCAN_REGIONS type
+        result = [
+            {"start": s, "end": e, "score": sc, "disabled": False}
+            for s, e, sc in regions
+        ]
+
+        embed_model = model.get("embed_model", "UNKNOWN")
+        return ({"model": embed_model, "regions": result},)
+```
+
+**Step 2: Register in `nodes/__init__.py`**
+
+Add to imports and mappings:
+
+```python
+from .load_video import LoadVideo
+from .audio_scan import AudioScan
+
+NODE_CLASS_MAPPINGS = {
+    "8cut_LoadVideo": LoadVideo,
+    "8cut_AudioScan": AudioScan,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "8cut_LoadVideo": "Load Video (8cut)",
+    "8cut_AudioScan": "Audio Scan (8cut)",
+}
+```
+
+**Step 3: Test manually**
+
+- Connect LoadVideo → AudioScan
+- Copy a trained `.joblib` model to `ComfyUI-8cut/models/`
+- Run the workflow — verify scan completes, progress bar shows, regions output is populated
+
+**Step 4: Commit**
+
+```bash
+git add ComfyUI-8cut/nodes/audio_scan.py ComfyUI-8cut/nodes/__init__.py
+git commit -m "feat: AudioScan node wrapping core.audio_scan.scan_video"
+```
+
+---
+
+### Task 3: VideoReview node — Python side (ExecutionBlocker pattern)
+
+**Files:**
+- Create: `ComfyUI-8cut/nodes/video_review.py`
+- Modify: `ComfyUI-8cut/nodes/__init__.py`
+- Modify: `ComfyUI-8cut/server_routes.py`
+
+**Step 1: Add review routes to `server_routes.py`**
+
+```python
+import asyncio
+
+# In-memory store for review completion signals
+_review_events: dict[str, asyncio.Event] = {}
+_review_results: dict[str, dict] = {}
+
+
+@routes.post("/8cut/review_done/{node_id}")
+async def review_done(request):
+    """Frontend signals that the user finished reviewing."""
+    node_id = request.match_info["node_id"]
+    data = await request.json()
+    _review_results[node_id] = data
+    event = _review_events.get(node_id)
+    if event:
+        event.set()
+    return web.json_response({"ok": True})
+
+
+@routes.get("/8cut/scan_versions")
+async def scan_versions(request):
+    """Return scan version history for a file/profile/model."""
+    from core.db import ProcessedDB
+    db_path = os.path.join(
+        os.path.dirname(os.path.abspath(__file__)), "data", "8cut.db"
+    )
+    db = ProcessedDB(db_path)
+    filename = request.rel_url.query.get("filename", "")
+    profile = request.rel_url.query.get("profile", "default")
+    model = request.rel_url.query.get("model", "")
+    versions = db.get_scan_versions(filename, profile, model)
+    return web.json_response(versions)
+
+
+@routes.post("/8cut/toggle_region")
+async def toggle_region(request):
+    """Toggle disabled state of a scan result row."""
+    from core.db import ProcessedDB
+    db_path = os.path.join(
+        os.path.dirname(os.path.abspath(__file__)), "data", "8cut.db"
+    )
+    db = ProcessedDB(db_path)
+    data = await request.json()
+    db.toggle_scan_result_disabled(data["row_id"], data["disabled"])
+    return web.json_response({"ok": True})
+
+
+@routes.post("/8cut/add_negatives")
+async def add_negatives(request):
+    """Mark timestamps as hard negatives."""
+    from core.db import ProcessedDB
+    db_path = os.path.join(
+        os.path.dirname(os.path.abspath(__file__)), "data", "8cut.db"
+    )
+    db = ProcessedDB(db_path)
+    data = await request.json()
+    db.add_hard_negatives(
+        data["filename"], data["profile"], data["times"],
+        source_path=data.get("source_path", ""),
+        source_model=data.get("source_model", ""),
+    )
+    return web.json_response({"ok": True})
+```
+
+**Step 2: Create `nodes/video_review.py`**
+
+```python
+"""VideoReview node — interactive video review with ExecutionBlocker pattern."""
+
+import json
+import server as comfy_server
+from comfy_execution.graph_utils import ExecutionBlocker
+
+
+class VideoReview:
+    """Interactive video review — pauses execution for user interaction.
+
+    First pass: displays video + scan regions in the widget, blocks downstream.
+    Second pass: after user clicks Continue, passes edited regions downstream.
+    """
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "video": ("VIDEO_PATH",),
+            },
+            "optional": {
+                "regions": ("SCAN_REGIONS",),
+            },
+            "hidden": {
+                "unique_id": "UNIQUE_ID",
+                "confirmed": ("BOOLEAN", {"default": False}),
+                "edited_regions": ("STRING", {"default": ""}),
+            },
+        }
+
+    RETURN_TYPES = ("SCAN_REGIONS",)
+    RETURN_NAMES = ("regions",)
+    OUTPUT_NODE = True
+    FUNCTION = "run"
+    CATEGORY = "8cut"
+
+    def run(self, video, regions=None, unique_id=None,
+            confirmed=False, edited_regions=""):
+        if confirmed and edited_regions:
+            # Second pass — user clicked Continue, pass data downstream
+            return (json.loads(edited_regions),)
+
+        # First pass — send data to frontend widget, block downstream
+        ui_data = {
+            "video_path": video,
+            "regions": regions if regions else {"model": "", "regions": []},
+            "node_id": unique_id or "",
+        }
+        return {
+            "ui": ui_data,
+            "result": (ExecutionBlocker("Waiting for review..."),),
+        }
+```
+
+**Step 3: Register in `nodes/__init__.py`**
+
+Add VideoReview to imports and both mappings dicts.
+
+**Step 4: Test manually**
+
+- Connect LoadVideo → AudioScan → VideoReview
+- Run workflow — verify AudioScan completes, VideoReview shows "Waiting for review..." status (no widget yet)
+- Verify no server freeze — other ComfyUI operations should still work
+
+**Step 5: Commit**
+
+```bash
+git add ComfyUI-8cut/
+git commit -m "feat: VideoReview node with ExecutionBlocker two-pass pattern"
+```
+
+---
+
+### Task 4: VideoReview frontend — video player and static region display
+
+**Files:**
+- Create: `ComfyUI-8cut/web/js/video_review.js`
+
+This is the core frontend work. Start with playback + read-only region display.
+
+**Step 1: Create `web/js/video_review.js` — extension registration and video player**
+
+```javascript
+import { app } from "../../../scripts/app.js";
+import { api } from "../../../scripts/api.js";
+
+function chainCallback(object, property, callback) {
+    if (object[property]) {
+        const original = object[property];
+        object[property] = function () {
+            const r = original.apply(this, arguments);
+            callback.apply(this, arguments);
+            return r;
+        };
+    } else {
+        object[property] = callback;
+    }
+}
+
+app.registerExtension({
+    name: "8cut.VideoReview",
+
+    async beforeRegisterNodeDef(nodeType, nodeData, app) {
+        if (nodeData?.name !== "8cut_VideoReview") return;
+
+        // On node creation — add DOM widget with video + timeline + table
+        chainCallback(nodeType.prototype, "onNodeCreated", function () {
+            const container = document.createElement("div");
+            container.style.cssText = "width:100%;min-height:480px;background:#1a1a1a;display:flex;flex-direction:column;";
+
+            // Video player
+            const video = document.createElement("video");
+            video.controls = true;
+            video.loop = true;
+            video.style.cssText = "width:100%;max-height:300px;background:#000;";
+            container.appendChild(video);
+
+            // Timeline canvas
+            const timeline = document.createElement("canvas");
+            timeline.height = 60;
+            timeline.style.cssText = "width:100%;height:60px;cursor:crosshair;background:#202020;";
+            container.appendChild(timeline);
+
+            // Region table container
+            const tableWrap = document.createElement("div");
+            tableWrap.style.cssText = "flex:1;overflow-y:auto;max-height:200px;";
+            const table = document.createElement("table");
+            table.style.cssText = "width:100%;border-collapse:collapse;color:#ccc;font-size:12px;";
+            table.innerHTML = "<thead><tr><th>Time</th><th>End</th><th>Score</th></tr></thead><tbody></tbody>";
+            tableWrap.appendChild(table);
+            container.appendChild(tableWrap);
+
+            // Continue button
+            const btnRow = document.createElement("div");
+            btnRow.style.cssText = "display:flex;gap:8px;padding:4px;";
+            const btnContinue = document.createElement("button");
+            btnContinue.textContent = "Continue ▶";
+            btnContinue.style.cssText = "padding:6px 16px;background:#2d7d2d;color:#fff;border:none;border-radius:4px;cursor:pointer;";
+            btnRow.appendChild(btnContinue);
+            container.appendChild(btnRow);
+
+            // Store references
+            this._8cut = { video, timeline, table, btnContinue, regions: null, videoPath: "" };
+
+            // Add as DOM widget
+            this.addDOMWidget("video_review", "preview", container, {
+                serialize: false,
+                getMinHeight: () => 480,
+            });
+
+            // Timeline click → seek
+            timeline.addEventListener("click", (e) => {
+                if (!video.duration) return;
+                const rect = timeline.getBoundingClientRect();
+                const frac = (e.clientX - rect.left) / rect.width;
+                video.currentTime = frac * video.duration;
+            });
+
+            // Timeline rendering
+            const renderTimeline = () => {
+                const ctx = timeline.getContext("2d");
+                const w = timeline.width = timeline.clientWidth;
+                const h = timeline.height;
+                ctx.fillStyle = "#202020";
+                ctx.fillRect(0, 0, w, h);
+
+                if (!video.duration) return;
+
+                // Draw regions
+                const data = this._8cut.regions;
+                if (data && data.regions) {
+                    for (const r of data.regions) {
+                        const x1 = (r.start / video.duration) * w;
+                        const x2 = (r.end / video.duration) * w;
+                        const alpha = 0.3 + r.score * 0.5;
+                        ctx.fillStyle = r.disabled
+                            ? `rgba(100,100,100,${alpha})`
+                            : `rgba(100,200,255,${alpha})`;
+                        ctx.fillRect(x1, 0, x2 - x1, h);
+                    }
+                }
+
+                // Cursor
+                const cx = (video.currentTime / video.duration) * w;
+                ctx.strokeStyle = "#fff";
+                ctx.lineWidth = 2;
+                ctx.beginPath();
+                ctx.moveTo(cx, 0);
+                ctx.lineTo(cx, h);
+                ctx.stroke();
+            };
+
+            video.addEventListener("timeupdate", renderTimeline);
+            new ResizeObserver(renderTimeline).observe(timeline);
+
+            // Continue button handler
+            btnContinue.addEventListener("click", () => {
+                // Re-queue the prompt with confirmed=true
+                const regions = this._8cut.regions || { model: "", regions: [] };
+                // Set widget values for re-execution
+                this.widgets?.forEach(w => {
+                    if (w.name === "confirmed") w.value = true;
+                    if (w.name === "edited_regions") w.value = JSON.stringify(regions);
+                });
+                app.queuePrompt();
+            });
+        });
+
+        // On executed — receive data from Python node
+        chainCallback(nodeType.prototype, "onExecuted", function (message) {
+            if (!this._8cut || !message) return;
+            const { video_path, regions, node_id } = message;
+            this._8cut.videoPath = video_path;
+            this._8cut.regions = regions;
+            this._8cut.nodeId = node_id;
+
+            // Set video source
+            this._8cut.video.src = api.apiURL(
+                "/8cut/video?path=" + encodeURIComponent(video_path)
+            );
+
+            // Populate table
+            const tbody = this._8cut.table.querySelector("tbody");
+            tbody.innerHTML = "";
+            if (regions && regions.regions) {
+                for (const r of regions.regions) {
+                    const tr = document.createElement("tr");
+                    tr.style.cursor = "pointer";
+                    tr.innerHTML = `<td>${formatTime(r.start)}</td><td>${formatTime(r.end)}</td><td>${r.score.toFixed(2)}</td>`;
+                    tr.addEventListener("click", () => {
+                        this._8cut.video.currentTime = r.start;
+                    });
+                    if (r.disabled) tr.style.color = "#666";
+                    tbody.appendChild(tr);
+                }
+            }
+
+            // Reset confirmed state for next run
+            this.widgets?.forEach(w => {
+                if (w.name === "confirmed") w.value = false;
+            });
+        });
+    },
+});
+
+function formatTime(secs) {
+    const m = Math.floor(secs / 60);
+    const s = (secs % 60).toFixed(1);
+    return `${m}:${s.padStart(4, "0")}`;
+}
+```
+
+**Step 2: Test manually**
+
+- Run the LoadVideo → AudioScan → VideoReview workflow
+- Verify: video plays in the node widget, timeline shows colored regions, table lists results
+- Click a table row → video seeks to that time
+- Click timeline → video seeks
+- Click Continue → prompt re-queues (downstream nodes would execute if connected)
+
+**Step 3: Commit**
+
+```bash
+git add ComfyUI-8cut/web/
+git commit -m "feat: VideoReview frontend — video player, timeline, region table"
+```
+
+---
+
+### Task 5: VideoReview interactivity — region editing, negatives, disable
+
+**Files:**
+- Modify: `ComfyUI-8cut/web/js/video_review.js`
+
+**Step 1: Add region disable toggle (D key or double-click)**
+
+Add to the `onNodeCreated` callback, after table population in `onExecuted`:
+
+- Double-click a table row → toggle `r.disabled`, update row color, redraw timeline
+- Store toggle in local state (the `regions` object)
+
+**Step 2: Add negative marking button**
+
+Add "Add Negative" button to `btnRow`. On click:
+- Collect selected table rows
+- Send `POST /8cut/add_negatives` with their timestamps
+- Mark rows red in the table
+- Update local regions state
+
+**Step 3: Add region edge dragging on timeline**
+
+In the timeline canvas:
+- `mousedown` near a region edge (within 5px) → start drag, track which region + which edge
+- `mousemove` → update region start or end, redraw
+- `mouseup` → finalize, send `POST /8cut/resize_region` to persist
+
+**Step 4: Add version history dropdown**
+
+- Add a `<select>` element above the table
+- On node execution, fetch versions via `GET /8cut/scan_versions?filename=...&profile=...&model=...`
+- On change, fetch that version's results and repopulate table + timeline
+
+**Step 5: Test manually**
+
+- Disable a region → gray in timeline + table
+- Mark negative → red highlight, verify via DB
+- Drag region edge → resizes, persists after Continue
+- Switch version → table + timeline update
+
+**Step 6: Commit**
+
+```bash
+git add ComfyUI-8cut/web/
+git commit -m "feat: VideoReview interactivity — disable, negatives, drag resize, versions"
+```
+
+---
+
+### Task 6: TrainModel node
+
+**Files:**
+- Create: `ComfyUI-8cut/nodes/train_model.py`
+- Modify: `ComfyUI-8cut/nodes/__init__.py`
+- Modify: `ComfyUI-8cut/server_routes.py`
+
+**Step 1: Add data query routes to `server_routes.py`**
+
+```python
+@routes.get("/8cut/profiles")
+async def get_profiles(request):
+    from core.db import ProcessedDB
+    db_path = os.path.join(
+        os.path.dirname(os.path.abspath(__file__)), "data", "8cut.db"
+    )
+    db = ProcessedDB(db_path)
+    return web.json_response(db.get_profiles())
+
+
+@routes.get("/8cut/export_folders")
+async def get_export_folders(request):
+    from core.db import ProcessedDB
+    db_path = os.path.join(
+        os.path.dirname(os.path.abspath(__file__)), "data", "8cut.db"
+    )
+    db = ProcessedDB(db_path)
+    profile = request.rel_url.query.get("profile", "default")
+    return web.json_response(db.get_export_folders(profile))
+
+
+@routes.get("/8cut/models")
+async def list_models(request):
+    models_dir = os.path.join(
+        os.path.dirname(os.path.abspath(__file__)), "models"
+    )
+    if not os.path.isdir(models_dir):
+        return web.json_response([])
+    files = sorted(f for f in os.listdir(models_dir) if f.endswith(".joblib"))
+    return web.json_response(files)
+```
+
+**Step 2: Create `nodes/train_model.py`**
+
+```python
+"""TrainModel node — train an audio classifier from labeled data."""
+
+import os
+import sys
+
+_pack_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+if _pack_dir not in sys.path:
+    sys.path.insert(0, _pack_dir)
+
+from core.audio_scan import train_classifier, _EMBED_MODELS
+from core.db import ProcessedDB
+
+import server as comfy_server
+
+
+class TrainModel:
+    """Train an audio event classifier from exported clips."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        db_path = os.path.join(_pack_dir, "data", "8cut.db")
+        db = ProcessedDB(db_path)
+        profiles = db.get_profiles() or ["default"]
+        embed_models = list(_EMBED_MODELS.keys())
+        # Folders will be populated dynamically via frontend JS
+        return {
+            "required": {
+                "profile": (profiles,),
+                "positive_folder": ("STRING", {"default": ""}),
+                "embed_model": (embed_models, {"default": "EAT_LARGE"}),
+                "use_hard_negatives": ("BOOLEAN", {"default": True}),
+            },
+            "optional": {
+                "negative_folder": ("STRING", {"default": ""}),
+                "video_dir": ("STRING", {"default": ""}),
+            },
+            "hidden": {"unique_id": "UNIQUE_ID"},
+        }
+
+    RETURN_TYPES = ("SCAN_MODEL",)
+    RETURN_NAMES = ("model",)
+    FUNCTION = "run"
+    CATEGORY = "8cut"
+
+    def run(self, profile, positive_folder, embed_model,
+            use_hard_negatives, negative_folder="", video_dir="",
+            unique_id=None):
+        db_path = os.path.join(_pack_dir, "data", "8cut.db")
+        db = ProcessedDB(db_path)
+
+        video_infos = db.get_training_data(
+            profile, positive_folder,
+            negative_folder=negative_folder,
+            fallback_video_dir=video_dir,
+            use_hard_negatives=use_hard_negatives,
+        )
+
+        if not video_infos:
+            raise ValueError(
+                f"No training data found for profile '{profile}', "
+                f"folder '{positive_folder}'"
+            )
+
+        model_name = f"{profile}_{embed_model}"
+        model_path = os.path.join(_pack_dir, "models", f"{model_name}.joblib")
+        os.makedirs(os.path.dirname(model_path), exist_ok=True)
+
+        prompt_server = comfy_server.PromptServer.instance
+
+        def progress_cb(current, total):
+            if unique_id is not None:
+                prompt_server.send_sync("progress", {
+                    "value": current, "max": total,
+                    "node": unique_id,
+                })
+
+        train_classifier(
+            video_infos,
+            model_path=model_path,
+            embed_model=embed_model,
+            progress_cb=progress_cb,
+        )
+
+        return (model_path,)
+```
+
+**Step 3: Register in `nodes/__init__.py`**
+
+Add TrainModel to imports and both mappings dicts.
+
+**Step 4: Test manually**
+
+- Copy existing `8cut.db` to `ComfyUI-8cut/data/`
+- Add TrainModel node, select profile + positive folder + embed model
+- Run — verify training completes, `.joblib` saved to `models/`
+- Connect output to AudioScan's model input — verify scan works with the trained model
+
+**Step 5: Commit**
+
+```bash
+git add ComfyUI-8cut/
+git commit -m "feat: TrainModel node wrapping core.audio_scan.train_classifier"
+```
+
+---
+
+### Task 7: ExportClips node
+
+**Files:**
+- Create: `ComfyUI-8cut/nodes/export_clips.py`
+- Modify: `ComfyUI-8cut/nodes/__init__.py`
+
+**Step 1: Create `nodes/export_clips.py`**
+
+```python
+"""ExportClips node — export video clips from scan regions."""
+
+import os
+import sys
+import subprocess
+
+_pack_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+if _pack_dir not in sys.path:
+    sys.path.insert(0, _pack_dir)
+
+from core.ffmpeg import build_ffmpeg_command
+from core.paths import build_export_path, _bin
+
+import server as comfy_server
+
+
+def _build_export_spans(regions, fuse_gap=30.0, spread=3.0, min_dur=8.0):
+    """Merge nearby regions into spans and place clips at spread intervals."""
+    if not regions:
+        return []
+    sorted_r = sorted(regions, key=lambda r: r["start"])
+    spans = []
+    s, e = sorted_r[0]["start"], sorted_r[0]["end"]
+    for r in sorted_r[1:]:
+        if r["start"] - e <= fuse_gap:
+            e = max(e, r["end"])
+        else:
+            spans.append((s, e))
+            s, e = r["start"], r["end"]
+    spans.append((s, e))
+
+    groups = []
+    step = max(spread, 1.0)
+    for s, e in spans:
+        dur = e - s
+        if dur < min_dur:
+            continue
+        positions = []
+        t = s
+        while t + min_dur <= e:
+            positions.append(t)
+            t += step
+        if positions:
+            groups.append(positions)
+    return groups
+
+
+class ExportClips:
+    """Export video clips from reviewed scan regions."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "video": ("VIDEO_PATH",),
+                "regions": ("SCAN_REGIONS",),
+                "output_folder": ("STRING", {"default": "/tmp/8cut_export"}),
+                "spread": ("FLOAT", {
+                    "default": 3.0, "min": 1.0, "max": 30.0, "step": 0.5,
+                }),
+                "clip_count": ("INT", {
+                    "default": 3, "min": 1, "max": 20,
+                }),
+                "fuse_gap": ("FLOAT", {
+                    "default": 30.0, "min": 1.0, "max": 120.0, "step": 1.0,
+                }),
+            },
+            "optional": {
+                "short_side": ("INT", {
+                    "default": 512, "min": 128, "max": 2160, "step": 8,
+                }),
+            },
+            "hidden": {"unique_id": "UNIQUE_ID"},
+        }
+
+    RETURN_TYPES = ("STRING",)
+    RETURN_NAMES = ("export_folder",)
+    OUTPUT_NODE = True
+    FUNCTION = "run"
+    CATEGORY = "8cut"
+
+    def run(self, video, regions, output_folder, spread, clip_count,
+            fuse_gap, short_side=512, unique_id=None):
+        # Filter to enabled regions only
+        active = [r for r in regions.get("regions", []) if not r.get("disabled")]
+        if not active:
+            return {"ui": {"text": "No active regions to export"}, "result": (output_folder,)}
+
+        groups = _build_export_spans(active, fuse_gap=fuse_gap, spread=spread)
+        if not groups:
+            return {"ui": {"text": "No spans long enough to export"}, "result": (output_folder,)}
+
+        os.makedirs(output_folder, exist_ok=True)
+        prompt_server = comfy_server.PromptServer.instance
+        exported = []
+        total_clips = sum(min(len(g), clip_count) for g in groups)
+        done = 0
+
+        for gi, group in enumerate(groups):
+            for ci, start in enumerate(group[:clip_count]):
+                out_name = f"clip_{gi:03d}_{ci:03d}.mp4"
+                out_path = os.path.join(output_folder, out_name)
+                cmd = build_ffmpeg_command(
+                    video, start, out_path,
+                    short_side=short_side,
+                )
+                subprocess.run(cmd, capture_output=True, timeout=120)
+                exported.append(out_path)
+                done += 1
+                if unique_id is not None:
+                    prompt_server.send_sync("progress", {
+                        "value": done, "max": total_clips,
+                        "node": unique_id,
+                    })
+
+        return {
+            "ui": {"text": f"Exported {len(exported)} clips to {output_folder}"},
+            "result": (output_folder,),
+        }
+```
+
+**Step 2: Register in `nodes/__init__.py`**
+
+Add ExportClips to imports and both mappings dicts.
+
+**Step 3: Test manually**
+
+- Full pipeline: LoadVideo → AudioScan → VideoReview → ExportClips
+- Run scan, review (click Continue), verify clips appear in output folder
+- Check ffmpeg output — correct resolution, timestamps
+
+**Step 4: Commit**
+
+```bash
+git add ComfyUI-8cut/
+git commit -m "feat: ExportClips node with region fusion and ffmpeg export"
+```
+
+---
+
+### Task 8: DB initialization and profile management
+
+**Files:**
+- Modify: `ComfyUI-8cut/server_routes.py`
+- Create: `ComfyUI-8cut/utils.py` (shared DB path helper)
+
+**Step 1: Create `utils.py` — shared helpers**
+
+Extract the repeated DB path construction:
+
+```python
+"""Shared utilities for ComfyUI-8cut."""
+
+import os
+
+PACK_DIR = os.path.dirname(os.path.abspath(__file__))
+DATA_DIR = os.path.join(PACK_DIR, "data")
+MODELS_DIR = os.path.join(PACK_DIR, "models")
+DB_PATH = os.path.join(DATA_DIR, "8cut.db")
+
+
+def get_db():
+    """Return a ProcessedDB instance for the node pack's database."""
+    from core.db import ProcessedDB
+    os.makedirs(DATA_DIR, exist_ok=True)
+    return ProcessedDB(DB_PATH)
+```
+
+**Step 2: Refactor all DB usage**
+
+Replace all inline `db_path = os.path.join(...)` / `ProcessedDB(db_path)` calls in `server_routes.py` and node files to use `from utils import get_db, MODELS_DIR`.
+
+**Step 3: Test manually**
+
+- Delete `data/8cut.db`, start ComfyUI — verify DB is auto-created
+- Copy existing DB — verify it works with migration
+
+**Step 4: Commit**
+
+```bash
+git add ComfyUI-8cut/
+git commit -m "refactor: extract shared DB/path helpers into utils.py"
+```
+
+---
+
+### Task 9: End-to-end integration test
+
+**Step 1: Full workflow test**
+
+Execute the complete pipeline manually:
+
+1. Place a video file accessible to the server
+2. Copy a trained model to `models/` (or train one via TrainModel node)
+3. Build workflow: LoadVideo → AudioScan → VideoReview → ExportClips
+4. Run — verify scan completes with progress
+5. In VideoReview: play video, click timeline, review regions
+6. Disable a region, mark a negative
+7. Click Continue
+8. Verify ExportClips produces correct clips (disabled regions excluded)
+
+**Step 2: Training round-trip test**
+
+1. Export some clips via ExportClips (writes to DB)
+2. Run TrainModel with the exported data
+3. Rescan with the new model
+4. Verify improved results
+
+**Step 3: Remote access test**
+
+1. Access ComfyUI from a different machine via browser
+2. Verify video streams and plays smoothly
+3. Verify all interactions work over the network
+
+**Step 4: Push**
+
+```bash
+git push
+```