app.py

import gradio as gr
import pandas as pd
import os
from pathlib import Path
from datasets import load_dataset, get_dataset_config_names

# HF Dataset sync
try:
    from hf_dataset_sync import init_dataset_sync
    dataset_sync_enabled = init_dataset_sync()
except Exception as e:
    dataset_sync_enabled = False
    print(f"⚠️ Dataset sync disabled: {e}")

# Predefined datasets (configs fetched dynamically)
PREDEFINED_DATASETS = [
    "abraranwar/agibotworld_alpha_rfm",
    "abraranwar/libero_rfm",
    "abraranwar/usc_koch_rewind_rfm",
    "aliangdw/metaworld",
    "anqil/rh20t_rfm",
    "anqil/rh20t_subset_rfm",
    "jesbu1/auto_eval_rfm",
    "jesbu1/egodex_rfm",
    "jesbu1/epic_rfm",
    "jesbu1/fino_net_rfm",
    "jesbu1/failsafe_rfm",
    "jesbu1/hand_paired_rfm",
    "jesbu1/galaxea_rfm",
    "jesbu1/h2r_rfm",
    "jesbu1/humanoid_everyday_rfm",
    "jesbu1/molmoact_rfm",
    "jesbu1/motif_rfm",
    "jesbu1/oxe_rfm",
    "jesbu1/oxe_rfm_eval",
    "jesbu1/ph2d_rfm",
    "jesbu1/racer_rfm",
    "jesbu1/roboarena_0825_rfm",
    "jesbu1/soar_rfm",
    "ykorkmaz/libero_failure_rfm",
    "aliangdw/usc_xarm_policy_ranking",
    "aliangdw/usc_franka_policy_ranking",
    "aliangdw/utd_so101_policy_ranking",
    "aliangdw/utd_so101_human",
]

# Global state
current_trajectories = []
current_idx = 0
evaluations_df = pd.DataFrame(columns=[
    "dataset_repo", "config_name", "trajectory_id", "task", 
    "decision", "issue_type", "notes", "timestamp"
])


def load_evaluations():
    global evaluations_df
    
    # Try loading from shared HF dataset first (if enabled)
    if dataset_sync_enabled:
        try:
            from huggingface_hub import hf_hub_download
            DATASET_REPO = os.getenv("EVAL_DATASET_REPO")
            HF_TOKEN = os.getenv("HF_TOKEN")
            csv_file = hf_hub_download(
                DATASET_REPO, 
                "traj_evaluations.csv", 
                repo_type="dataset", 
                token=HF_TOKEN,
                force_download=True  # Always get fresh data
            )
            evaluations_df = pd.read_csv(csv_file, keep_default_na=False, na_values=[''])
            # Clean up any legacy "nan" strings from old data
            evaluations_df = evaluations_df.replace(['nan', 'NaN', 'None'], '')
            # Debug: check issue_type values
            unique_issues = evaluations_df['issue_type'].unique()
            print(f"📊 Loaded {len(evaluations_df)} evaluations from shared dataset")
            print(f"🔍 Unique issue_type values: {unique_issues}")
            return
        except Exception as e:
            print(f"⚠️ Could not load from shared dataset: {e}")
    
    # Fallback to local CSV
    csv_path = Path("data/evaluations.csv") if os.getenv("SPACE_ID") else Path("evaluations.csv")
    if csv_path.exists():
        evaluations_df = pd.read_csv(csv_path, keep_default_na=False, na_values=[''])
        # Clean up any legacy "nan" strings from old data
        evaluations_df = evaluations_df.replace(['nan', 'NaN', 'None'], '')
        print(f"📊 Loaded {len(evaluations_df)} evaluations from local CSV")


def save_evaluations():
    # Replace NaN with empty strings before saving
    df_to_save = evaluations_df.fillna("")
    if os.getenv("SPACE_ID"):
        os.makedirs("data", exist_ok=True)
        df_to_save.to_csv("data/evaluations.csv", index=False)
    else:
        df_to_save.to_csv("evaluations.csv", index=False)


def get_stats():
    total = len(evaluations_df)
    if total == 0:
        return "No labels yet"
    keeps = len(evaluations_df[evaluations_df['decision'] == 'keep'])
    removes = len(evaluations_df[evaluations_df['decision'] == 'remove'])
    reviews = len(evaluations_df[evaluations_df['decision'] == 'review'])
    return f"Total: {total} | ✅ {keeps} | ❌ {removes} | 🔍 {reviews}"


def fetch_configs(dataset_repo):
    """Fetch configs dynamically (fast API call)."""
    if not dataset_repo:
        return gr.update(choices=[], value=""), "", 0, 20
    try:
        configs = get_dataset_config_names(dataset_repo)
        if configs:
            return gr.update(choices=configs, value=configs[0]), "", 0, 20
        return gr.update(choices=["default"], value="default"), "", 0, 20
    except Exception as e:
        print(f"Config fetch error: {e}")
        return gr.update(choices=["default"], value="default"), "", 0, 20


def analyze_dataset_progress(dataset_repo, config_name):
    """Analyze labeling progress for selected dataset and suggest range."""
    if not dataset_repo:
        return "", 0, 20
    
    config = config_name if config_name and config_name != "default" else None
    
    # Filter evaluations for this dataset
    dataset_evals = evaluations_df[
        (evaluations_df['dataset_repo'] == dataset_repo) & 
        (evaluations_df['config_name'] == (config if config else ''))
    ]
    
    # Try to get dataset size
    try:
        # First try to get from dataset info (fast)
        from datasets import get_dataset_infos
        try:
            infos = get_dataset_infos(dataset_repo)
            config_key = config if config else list(infos.keys())[0]
            if config_key in infos and infos[config_key].splits.get('train'):
                dataset_size = infos[config_key].splits['train'].num_examples
            else:
                raise Exception("No info available")
        except:
            # Fallback: iterate through dataset
            ds = load_dataset(dataset_repo, config, split="train", streaming=True)
            dataset_size = 0
            for i, _ in enumerate(ds):
                dataset_size = i + 1
                if i >= 9999:  # Cap at 10,000 for performance
                    dataset_size = f"~{dataset_size}"
                    break
    except Exception as e:
        dataset_size = "Unknown"
    
    if len(dataset_evals) == 0:
        return f"📊 **No trajectories labeled yet for this dataset**\n\n**Dataset size:** {dataset_size} trajectories", 0, 20
    
    # Get labeled trajectory IDs
    labeled_ids = set(dataset_evals['trajectory_id'].unique())
    
    # Count by decision
    keeps = len(dataset_evals[dataset_evals['decision'] == 'keep'])
    removes = len(dataset_evals[dataset_evals['decision'] == 'remove'])
    reviews = len(dataset_evals[dataset_evals['decision'] == 'review'])
    
    # Try to get total dataset size first (fast)
    try:
        from datasets import get_dataset_infos
        try:
            infos = get_dataset_infos(dataset_repo)
            config_key = config if config else list(infos.keys())[0]
            if config_key in infos and infos[config_key].splits.get('train'):
                dataset_size = infos[config_key].splits['train'].num_examples
            else:
                dataset_size = None
        except:
            dataset_size = None
    except:
        dataset_size = None
    
    # Try to estimate gaps and count trajectories
    try:
        ds = load_dataset(dataset_repo, config, split="train", streaming=True)
        
        # Sample to find gaps in labeled trajectories
        checked_count = 0
        unlabeled_ranges = []
        current_unlabeled_start = None
        
        for i, sample in enumerate(ds):
            traj_id = sample.get("id")
            if traj_id not in labeled_ids:
                if current_unlabeled_start is None:
                    current_unlabeled_start = i
            else:
                if current_unlabeled_start is not None:
                    unlabeled_ranges.append((current_unlabeled_start, i-1))
                    current_unlabeled_start = None
            
            checked_count = i + 1
            if dataset_size is None:
                dataset_size = i + 1
            if checked_count >= 1000:  # Check first 1000 for gaps
                if dataset_size is None or dataset_size == checked_count:
                    dataset_size = f"~{checked_count}"
                break
        
        # Add final range if still unlabeled
        if current_unlabeled_start is not None:
            unlabeled_ranges.append((current_unlabeled_start, checked_count-1))
        
        # Find best range to suggest
        if unlabeled_ranges:
            # Suggest first significant gap (at least 10 trajectories)
            for start, end in unlabeled_ranges:
                if end - start >= 10:
                    suggested_start = start
                    suggested_end = min(start+20, end)
                    break
            else:
                # No big gaps, suggest first unlabeled range
                start, end = unlabeled_ranges[0]
                suggested_start = start
                suggested_end = min(start+20, end)
        else:
            # All checked trajectories are labeled, suggest next batch
            suggested_start = checked_count
            suggested_end = checked_count+20
        
        analysis = f"""📊 **Dataset Progress: {dataset_repo}** {'(' + config + ')' if config else ''}

**Labeled:** {len(labeled_ids)} trajectories
- ✅ Keep: {keeps} ({keeps/len(labeled_ids)*100:.1f}%)
- ❌ Remove: {removes} ({removes/len(labeled_ids)*100:.1f}%)
- 🔍 Review: {reviews} ({reviews/len(labeled_ids)*100:.1f}%)

**Dataset size:** {dataset_size} trajectories (checked: {checked_count})
"""
        
        if unlabeled_ranges[:3]:  # Show first 3 gaps
            gaps = ", ".join([f"{s}-{e}" for s, e in unlabeled_ranges[:3]])
            analysis += f"\n🎯 **Unlabeled gaps:** {gaps}"
        
        return analysis, suggested_start, suggested_end
        
    except Exception as e:
        # If can't analyze structure, suggest continuing after last labeled
        suggested_start = len(labeled_ids)
        suggested_end = len(labeled_ids) + 20
        
        # Still try to get dataset size from info API
        try:
            from datasets import get_dataset_infos
            infos = get_dataset_infos(dataset_repo)
            config_key = config if config else list(infos.keys())[0]
            if config_key in infos and infos[config_key].splits.get('train'):
                ds_size = infos[config_key].splits['train'].num_examples
                size_info = f"**Dataset size:** {ds_size} trajectories\n\n"
            else:
                size_info = ""
        except:
            size_info = ""
        
        return f"""📊 **Dataset Progress: {dataset_repo}**

**Labeled:** {len(labeled_ids)} trajectories
- ✅ Keep: {keeps} ({keeps/len(labeled_ids)*100:.1f}%)
- ❌ Remove: {removes} ({removes/len(labeled_ids)*100:.1f}%)
- 🔍 Review: {reviews} ({reviews/len(labeled_ids)*100:.1f}%)

{size_info}⚠️ Could not analyze dataset structure: {str(e)[:50]}
""", suggested_start, suggested_end


def get_video_url(dataset_repo, video_path):
    """Get direct HuggingFace URL for video (no download needed)."""
    return f"https://huggingface.co/datasets/{dataset_repo}/resolve/main/{video_path}"


def load_trajectories(dataset_repo, config_name, start_idx, end_idx, traj_id):
    """Load trajectories by range or specific ID."""
    global current_trajectories, current_idx
    
    # Refresh evaluations from shared dataset to get latest labels
    load_evaluations()
    
    if not dataset_repo:
        return (gr.update(visible=True), gr.update(visible=False), 
                None, "Select a dataset",
                gr.update(value=None), gr.update(visible=False), gr.update(value=None), gr.update(value=""), "",
                evaluations_df.tail(10), "⚠️ Select a dataset")
    
    config = config_name if config_name and config_name != "default" else None
    start = int(start_idx) if start_idx else 0
    end = int(end_idx) if end_idx else start + 20
    target_id = traj_id.strip() if traj_id else None
    
    try:
        ds = load_dataset(dataset_repo, config, split="train", streaming=True)
        
        current_trajectories = []
        for i, sample in enumerate(ds):
            # If looking for specific ID
            if target_id:
                if sample.get("id") == target_id:
                    video_path = sample.get("frames")
                    if video_path:
                        sample["video_url"] = get_video_url(dataset_repo, video_path)
                        sample["dataset_repo"] = dataset_repo
                        sample["config_name"] = config
                        current_trajectories.append(sample)
                    break
                continue
            
            # Range-based loading
            if i < start:
                continue
            if i > end:
                break
            
            video_path = sample.get("frames")
            if video_path:
                sample["video_url"] = get_video_url(dataset_repo, video_path)
                sample["dataset_repo"] = dataset_repo
                sample["config_name"] = config
                current_trajectories.append(sample)
        
        current_idx = 0
        
        if not current_trajectories:
            return (gr.update(visible=True), gr.update(visible=False), 
                    None, "❌ No trajectories found",
                    gr.update(value=None), gr.update(visible=False), gr.update(value=None), gr.update(value=""), "",
                    evaluations_df.tail(10), "❌ No trajectories found")
        
        return show_labeling_view()
        
    except Exception as e:
        return (gr.update(visible=True), gr.update(visible=False),
                None, f"❌ {str(e)[:50]}",
                gr.update(value=None), gr.update(visible=False), gr.update(value=None), gr.update(value=""), "",
                evaluations_df.tail(10), f"❌ Error: {str(e)[:50]}")


def get_trajectory_metadata(traj):
    """Extract and format trajectory metadata."""
    metadata = []
    
    # Success/failure status - check multiple possible fields
    if 'quality_label' in traj and traj['quality_label']:
        label = str(traj['quality_label']).lower()
        if 'success' in label:
            metadata.append("✅ Success")
        elif 'fail' in label:
            metadata.append("❌ Failure")
        elif 'suboptimal' in label:
            metadata.append("⚠️ Suboptimal")
        else:
            metadata.append(f"Quality: {traj['quality_label']}")
    elif 'success' in traj:
        success = traj['success']
        if success == True or success == 1 or success == 1.0:
            metadata.append("✅ Success")
        elif success == False or success == 0 or success == 0.0:
            metadata.append("❌ Failure")
        else:
            metadata.append(f"Status: {success}")
    elif 'is_success' in traj:
        if traj['is_success']:
            metadata.append("✅ Success")
        else:
            metadata.append("❌ Failure")
    
    # Suboptimal marker (separate from quality_label)
    if 'suboptimal' in traj and traj['suboptimal']:
        if "⚠️ Suboptimal" not in metadata:  # Don't duplicate
            metadata.append("⚠️ Suboptimal")
    elif 'is_suboptimal' in traj and traj['is_suboptimal']:
        if "⚠️ Suboptimal" not in metadata:
            metadata.append("⚠️ Suboptimal")
    
    # Source (human/robot)
    if 'is_robot' in traj:
        metadata.append("🤖 Robot" if traj['is_robot'] else "👤 Human")
    elif 'source' in traj:
        source = str(traj['source']).lower()
        if 'human' in source:
            metadata.append("👤 Human")
        elif 'robot' in source or 'policy' in source:
            metadata.append("🤖 Robot")
        else:
            metadata.append(f"Source: {traj['source']}")
    
    return " | ".join(metadata) if metadata else ""


def show_labeling_view():
    """Switch to labeling view with first trajectory."""
    traj = current_trajectories[current_idx]
    video_url = traj.get("video_url")
    task = traj.get("task", "No task")
    traj_id = traj.get("id", "Unknown")
    
    progress = f"Progress: {current_idx + 1}/{len(current_trajectories)} | ID: {traj_id[:8]}..."
    
    # Auto-apply previous label if exists
    prev_decision = None
    prev_issue = None
    prev_notes = ""
    if traj_id in evaluations_df['trajectory_id'].values:
        prev_row = evaluations_df[evaluations_df['trajectory_id'] == traj_id].iloc[-1]
        prev_decision = prev_row['decision']
        # Ensure prev_issue is None or a valid choice
        valid_choices = ['too_short', 'too_long', 'wrong_description', 'task_already_completed', 'mislabeled_success', 'mislabeled_failure', 'mislabeled_suboptimal', 'other']
        issue_val = prev_row['issue_type']
        prev_issue = issue_val if (issue_val and str(issue_val).strip() and str(issue_val) in valid_choices) else None
        prev_notes = prev_row['notes'] if pd.notna(prev_row['notes']) else ""
        progress += f" (prev: {prev_decision})"
    
    # Get trajectory metadata
    metadata = get_trajectory_metadata(traj)
    
    # Combine progress, metadata, and task
    if metadata:
        task_with_progress = f"{progress}\n{metadata}\n\n{task}"
    else:
        task_with_progress = f"{progress}\n\n{task}"
    
    return (
        gr.update(visible=False),
        gr.update(visible=True),
        video_url,
        task_with_progress,
        gr.update(value=prev_decision),  # decision_radio
        gr.update(visible=(prev_decision == "review")),  # review_options visibility
        gr.update(value=prev_issue),  # review_reason
        gr.update(value=prev_notes if prev_notes else ""),  # notes_input
        "",  # save_status (empty)
        evaluations_df.tail(10),
        f"✅ Loaded {len(current_trajectories)} trajectories"
    )


def show_current():
    if not current_trajectories or current_idx >= len(current_trajectories):
        return (
            None,
            "No data",
            gr.update(value=None),
            gr.update(visible=False),
            gr.update(value=None),
            gr.update(value="")
        )
    
    traj = current_trajectories[current_idx]
    video_url = traj.get("video_url")
    task = traj.get("task", "No task")
    traj_id = traj.get("id", "Unknown")
    
    progress = f"Progress: {current_idx + 1}/{len(current_trajectories)} | ID: {traj_id[:8]}..."
    
    # Auto-apply previous label if exists
    prev_decision = None
    prev_issue = None
    prev_notes = ""
    if traj_id in evaluations_df['trajectory_id'].values:
        prev_row = evaluations_df[evaluations_df['trajectory_id'] == traj_id].iloc[-1]
        prev_decision = prev_row['decision']
        # Ensure prev_issue is None or a valid choice
        valid_choices = ['too_short', 'too_long', 'wrong_description', 'task_already_completed', 'mislabeled_success', 'mislabeled_failure', 'mislabeled_suboptimal', 'other']
        issue_val = prev_row['issue_type']
        prev_issue = issue_val if (issue_val and str(issue_val).strip() and str(issue_val) in valid_choices) else None
        prev_notes = prev_row['notes'] if pd.notna(prev_row['notes']) else ""
        progress += f" (prev: {prev_decision})"
    
    # Get trajectory metadata
    metadata = get_trajectory_metadata(traj)
    
    # Combine progress, metadata, and task
    if metadata:
        task_with_progress = f"{progress}\n{metadata}\n\n{task}"
    else:
        task_with_progress = f"{progress}\n\n{task}"
    
    return (
        video_url, 
        task_with_progress,
        gr.update(value=prev_decision),  # decision_radio
        gr.update(visible=(prev_decision == "review")),  # review_options
        gr.update(value=prev_issue),  # review_reason
        gr.update(value=prev_notes if prev_notes else "")  # notes_input
    )


def navigate(direction):
    global current_idx
    if direction == "next":
        current_idx = min(current_idx + 1, len(current_trajectories) - 1)
    else:
        current_idx = max(current_idx - 1, 0)
    return show_current()


def save_label(decision, issue_type="", notes=""):
    """Save label and advance. Updates existing if trajectory already labeled."""
    global evaluations_df, current_idx
    
    if not current_trajectories or current_idx >= len(current_trajectories):
        # Returns: video, task, decision, review_vis, review_reason, notes, status, table
        return show_current() + ("", evaluations_df.tail(10))
    
    traj = current_trajectories[current_idx]
    traj_id = traj.get("id", "")
    
    # Debug logging
    print(f"💾 save_label called:")
    print(f"  decision: {decision}")
    print(f"  issue_type: '{issue_type}' (type: {type(issue_type)}, len: {len(str(issue_type))})")
    print(f"  notes: {notes}")
    print(f"  traj_id: {traj_id[:20]}...")
    
    row_data = {
        "dataset_repo": traj.get("dataset_repo", ""),
        "config_name": traj.get("config_name", ""),
        "trajectory_id": traj_id,
        "task": traj.get("task", ""),
        "decision": decision,
        "issue_type": issue_type,
        "notes": notes,
        "timestamp": pd.Timestamp.now().isoformat()
    }
    
    print(f"  📋 row_data issue_type: '{row_data['issue_type']}'")
    
    # Upsert: update if exists
    existing_mask = evaluations_df['trajectory_id'] == traj_id
    is_update = existing_mask.any()
    if is_update:
        idx = evaluations_df[existing_mask].index[-1]
        print(f"  🔄 Updating existing row at index {idx}")
        for col, val in row_data.items():
            evaluations_df.at[idx, col] = val
        print(f"  ✅ After update, issue_type = '{evaluations_df.at[idx, 'issue_type']}'")
        status_msg = f"✅ Updated: {decision}"
    else:
        print(f"  ➕ Adding new row")
        evaluations_df = pd.concat([evaluations_df, pd.DataFrame([row_data])], ignore_index=True)
        new_idx = evaluations_df.index[-1]
        print(f"  ✅ After add, issue_type = '{evaluations_df.at[new_idx, 'issue_type']}'")
        status_msg = f"✅ Added: {decision}"
    
    save_evaluations()
    print(f"  💾 Saved to CSV")
    
    if dataset_sync_enabled:
        from hf_dataset_sync import append_to_dataset
        append_to_dataset(row_data)
    
    current_idx = min(current_idx + 1, len(current_trajectories) - 1)
    # Returns: video, task, decision, review_vis, review_reason, notes, status, table
    return show_current() + (status_msg, evaluations_df.tail(10))


def save_with_decision(decision, review_reason, notes):
    # Ensure issue is always empty string unless decision is "review" with valid reason
    valid_choices = ['too_short', 'too_long', 'wrong_description', 'task_already_completed', 'mislabeled_success', 'mislabeled_failure', 'mislabeled_suboptimal', 'other']
    
    # Debug logging
    print(f"🔍 save_with_decision called:")
    print(f"  decision: {decision} (type: {type(decision)})")
    print(f"  review_reason: {review_reason} (type: {type(review_reason)})")
    print(f"  notes: {notes}")
    
    issue = ""
    if decision == "review" and review_reason and str(review_reason) in valid_choices:
        issue = str(review_reason)
        print(f"  ✅ Setting issue_type to: {issue}")
    else:
        print(f"  ❌ Issue NOT set. Checks:")
        print(f"     decision == 'review': {decision == 'review'}")
        print(f"     review_reason truthy: {bool(review_reason)}")
        if review_reason:
            print(f"     review_reason in valid_choices: {str(review_reason) in valid_choices}")
    
    return save_label(decision, issue, notes)




def back_to_setup():
    return gr.update(visible=True), gr.update(visible=False)


def update_review_visibility(decision):
    return gr.update(visible=(decision == "review"))


# Load existing
load_evaluations()

# CSS
css = """
.container { max-width: 1000px; margin: 0 auto; }
.decision-btn { min-height: 50px !important; font-size: 16px !important; }
.task-box { background: #f8f9fa; padding: 12px; border-radius: 6px; border-left: 4px solid #667eea; }
.thin-back-btn button { 
    min-height: 35px !important; 
    font-size: 13px !important;
    margin-bottom: 8px !important;
}
#dataset_analysis {
    background: #f0f9ff;
    padding: 16px;
    border-radius: 8px;
    border-left: 4px solid #3b82f6;
    margin: 12px 0;
}
#save_status, #load_status { 
    font-weight: 600; 
    padding: 8px;
    border-radius: 6px;
    text-align: center;
    margin-top: 8px;
}
#save_status {
    color: #10b981; 
    background: #d1fae5;
}
#load_status {
    color: #667eea;
    background: #e0e7ff;
}
#speed_1x, #speed_2x, #speed_4x {
    border: 2px solid #e5e7eb !important;
    transition: all 0.2s;
}
#speed_1x.speed-active, #speed_2x.speed-active, #speed_4x.speed-active {
    background: #667eea !important;
    color: white !important;
    border-color: #667eea !important;
}
"""

with gr.Blocks(title="Trajectory Reviewer", css=css) as demo:
    
    gr.Markdown("# 🎯 Trajectory Reviewer")
    
    # Stage 1: Setup
    with gr.Column(visible=True) as setup_view:
        gr.Markdown("### Dataset")
        
        with gr.Row():
            dataset_dropdown = gr.Dropdown(
                choices=PREDEFINED_DATASETS,
                value="jesbu1/epic_rfm",
                label="Dataset",
                allow_custom_value=True,
                scale=3
            )
            refresh_btn = gr.Button("🔄", scale=0)
        
        config_dropdown = gr.Dropdown(
            choices=[],
            value="",
            label="Config",
            allow_custom_value=True
        )
        
        dataset_analysis = gr.Markdown("", elem_id="dataset_analysis")
        
        gr.Markdown("### Selection")
        with gr.Row():
            with gr.Column():
                start_idx = gr.Number(label="Start Index", value=0, precision=0)
                end_idx = gr.Number(label="End Index", value=20, precision=0)
            with gr.Column():
                traj_id_input = gr.Textbox(label="Or Specific ID", placeholder="Leave empty for range")
        
        load_btn = gr.Button("🚀 Load & Start", variant="primary", size="lg")
        load_status = gr.Markdown("", elem_id="load_status")
    
    # Stage 2: Labeling
    with gr.Column(visible=False) as labeling_view:
        
        with gr.Row():
            with gr.Column(scale=3):
                back_btn = gr.Button("← Back to Setup", variant="secondary", size="sm", elem_classes=["thin-back-btn"])
                video_player = gr.Video(label="Video", elem_id="traj_video", autoplay=True)
                
                # Speed controls
                gr.Markdown("**Playback Speed**")
                with gr.Row():
                    speed_1x = gr.Button("1x", size="sm", elem_id="speed_1x")
                    speed_2x = gr.Button("2x", size="sm", elem_id="speed_2x")
                    speed_4x = gr.Button("4x", size="sm", elem_id="speed_4x")
            
            with gr.Column(scale=2):
                task_display = gr.Textbox(label="📋 Task", interactive=False, lines=3, elem_classes=["task-box"])
                
                gr.Markdown("### Decision")
                decision_radio = gr.Radio(
                    choices=["keep", "remove", "review"],
                    label="Select",
                    value=None
                )
                
                with gr.Column(visible=False) as review_options:
                    review_reason = gr.Radio(
                        choices=[
                            "too_short", 
                            "too_long", 
                            "wrong_description", 
                            "task_already_completed",
                            "mislabeled_success",
                            "mislabeled_failure", 
                            "mislabeled_suboptimal",
                            "other"
                        ],
                        label="Review Reason",
                        value=None
                    )
                
                notes_input = gr.Textbox(label="Notes", placeholder="Optional...", lines=2)
                save_btn = gr.Button("💾 Save & Next", variant="primary", size="lg", elem_classes=["decision-btn"])
                save_status = gr.Markdown("", elem_id="save_status")
                
                with gr.Row():
                    prev_btn = gr.Button("← Prev", size="sm")
                    next_btn = gr.Button("Next →", size="sm")
        
        gr.Markdown("### Recent Labels")
        evals_table = gr.Dataframe(
            value=evaluations_df.tail(10),
            max_height=150
        )
    
    # Speed control JS
    def set_speed_js(rate, btn_id):
        return f"""
        () => {{
            const setSpeed = () => {{
                const video = document.querySelector('#traj_video video');
                if (video) {{
                    video.playbackRate = {rate};
                    // Highlight active button
                    ['#speed_1x', '#speed_2x', '#speed_4x'].forEach(id => {{
                        const btn = document.querySelector(id);
                        if (btn) btn.classList.remove('speed-active');
                    }});
                    document.querySelector('{btn_id}')?.classList.add('speed-active');
                }}
            }};
            setSpeed();
            // Also set on video load events
            setTimeout(setSpeed, 100);
            setTimeout(setSpeed, 500);
        }}
        """
    
    speed_1x.click(None, None, None, js=set_speed_js(1.0, '#speed_1x'))
    speed_2x.click(None, None, None, js=set_speed_js(2.0, '#speed_2x'))
    speed_4x.click(None, None, None, js=set_speed_js(4.0, '#speed_4x'))
    
    # Set 2x on video load
    video_player.change(
        None, None, None,
        js="() => { setTimeout(() => { const v = document.querySelector('#traj_video video'); if (v) v.playbackRate = 2.0; }, 500); }"
    )
    
    # Events
    dataset_dropdown.change(fetch_configs, [dataset_dropdown], [config_dropdown, dataset_analysis, start_idx, end_idx])
    refresh_btn.click(fetch_configs, [dataset_dropdown], [config_dropdown, dataset_analysis, start_idx, end_idx])
    config_dropdown.change(analyze_dataset_progress, [dataset_dropdown, config_dropdown], [dataset_analysis, start_idx, end_idx])
    
    load_btn.click(
        lambda: "⏳ Loading trajectories...",
        None,
        load_status
    ).then(
        load_trajectories,
        [dataset_dropdown, config_dropdown, start_idx, end_idx, traj_id_input],
        [setup_view, labeling_view, video_player, task_display,
         decision_radio, review_options, review_reason, notes_input, save_status,
         evals_table, load_status]
    )
    
    back_btn.click(back_to_setup, outputs=[setup_view, labeling_view])
    decision_radio.change(update_review_visibility, [decision_radio], [review_options])
    
    save_btn.click(
        save_with_decision,
        [decision_radio, review_reason, notes_input],
        [video_player, task_display, decision_radio, review_options, 
         review_reason, notes_input, save_status, evals_table]
    ).then(
        None, None, save_status,
        js="() => { setTimeout(() => document.querySelector('#save_status').textContent = '', 3000); }"
    )
    
    prev_btn.click(
        lambda: navigate("prev"), 
        outputs=[video_player, task_display, decision_radio, 
                 review_options, review_reason, notes_input]
    )
    next_btn.click(
        lambda: navigate("next"), 
        outputs=[video_player, task_display, decision_radio,
                 review_options, review_reason, notes_input]
    )
    
    # Load configs and analysis on startup
    demo.load(fetch_configs, [dataset_dropdown], [config_dropdown, dataset_analysis, start_idx, end_idx])


if __name__ == "__main__":
    demo.launch()