refactor(viewer): consolidate to standalone HTML generation

Author: abrichr

CLAUDE.md

@@ -382,20 +382,19 @@ az ml workspace sync-keys -n openadapt-ml -g openadapt-agents
 - Filter/search evaluations by epoch or correctness
 
 ### Viewer Code Consolidation
-**Status**: TODO - HIGH PRIORITY
+**Status**: DONE
 
-**Problem**: Viewer code is fragmented across multiple locations:
+**Problem**: Viewer code was fragmented across multiple locations:
 1. `generate_training_dashboard()` - generates unified viewer template
-2. `_enhance_comparison_to_unified_viewer()` - injects checkpoint_script into comparison.html
-3. `comparison.html` from capture - has its own display logic
-
-This causes issues where changes don't propagate correctly.
+2. `_enhance_comparison_to_unified_viewer()` - injected checkpoint_script into comparison.html
+3. `comparison.html` from capture - had its own display logic
 
-**Solution**: Create a single `generate_unified_viewer()` function that:
-- Takes comparison data as input
-- Generates a complete standalone viewer.html
-- Has all display logic in one place
-- Doesn't rely on injecting scripts into existing HTML
+**Solution implemented**:
+- `generate_unified_viewer_from_output_dir()` now always uses `_generate_unified_viewer_from_extracted_data()`
+- This generates a complete standalone viewer.html without script injection
+- `_enhance_comparison_to_unified_viewer()` marked as deprecated
+- All viewer display logic is now in one place (`_generate_unified_viewer_from_extracted_data`)
+- Changes to viewer code now propagate reliably
 
 ### README API Documentation
 **Status**: VERIFIED

openadapt_ml/cloud/local.py

@@ -58,49 +58,14 @@ def _regenerate_viewer_if_possible(output_dir: Path) -> bool:
 
     Returns True if viewer was regenerated, False otherwise.
     """
-    from openadapt_ml.training.trainer import _enhance_comparison_to_unified_viewer
-
-    # Look for base comparison file
-    base_file = output_dir / "comparison.html"
-    if not base_file.exists():
-        # Try to find any comparison HTML
-        comparison_files = list(output_dir.glob("*comparison*.html"))
-        if comparison_files:
-            base_file = comparison_files[0]
-        else:
-            return False
-
-    # Load predictions from checkpoint files
-    predictions_by_checkpoint = {"None": []}
-    for pred_file in output_dir.glob("predictions_*.json"):
-        checkpoint_name = pred_file.stem.replace("predictions_", "")
-        if "epoch" in checkpoint_name:
-            display_name = checkpoint_name.replace("epoch", "Epoch ").replace("_", " ").title()
-        elif checkpoint_name == "preview":
-            display_name = "Preview"
-        else:
-            display_name = checkpoint_name.title()
-
-        try:
-            with open(pred_file) as f:
-                predictions_by_checkpoint[display_name] = json.load(f)
-        except json.JSONDecodeError:
-            pass
-
-    # Get capture info
-    capture_id = "capture"
-    goal = "Complete the recorded workflow"
+    from openadapt_ml.training.trainer import generate_unified_viewer_from_output_dir
 
     try:
-        _enhance_comparison_to_unified_viewer(
-            base_file,
-            predictions_by_checkpoint,
-            output_dir / "viewer.html",
-            capture_id,
-            goal,
-        )
-        print(f"Regenerated viewer: {output_dir / 'viewer.html'}")
-        return True
+        viewer_path = generate_unified_viewer_from_output_dir(output_dir)
+        if viewer_path:
+            print(f"Regenerated viewer: {viewer_path}")
+            return True
+        return False
     except Exception as e:
         print(f"Could not regenerate viewer: {e}")
         return False
@@ -399,7 +364,7 @@ def cmd_viewer(args: argparse.Namespace) -> int:
     """Regenerate viewer from local training output."""
     from openadapt_ml.training.trainer import (
         generate_training_dashboard,
-        _enhance_comparison_to_unified_viewer,
+        generate_unified_viewer_from_output_dir,
         TrainingState,
         TrainingConfig,
     )
@@ -435,66 +400,12 @@ def cmd_viewer(args: argparse.Namespace) -> int:
         (current_dir / "dashboard.html").write_text(dashboard_html)
         print(f"  Regenerated: dashboard.html")
 
-    # Find comparison HTML to enhance
-    # Try epoch-specific files first, then fall back to generic comparison.html
-    comparison_files = list(current_dir.glob("comparison_epoch*.html"))
-    base_file = None
-    if comparison_files:
-        # Use the latest epoch comparison
-        base_file = sorted(comparison_files)[-1]
-    elif (current_dir / "comparison.html").exists():
-        base_file = current_dir / "comparison.html"
-
-    if base_file:
-        print(f"  Using base file: {base_file.name}")
-
-        # Load all prediction files
-        predictions_by_checkpoint = {"None": []}
-        for pred_file in current_dir.glob("predictions_*.json"):
-            checkpoint_name = pred_file.stem.replace("predictions_", "")
-            # Map to display name
-            if "epoch" in checkpoint_name:
-                display_name = checkpoint_name.replace("epoch", "Epoch ").replace("_", " ").title()
-            elif checkpoint_name == "preview":
-                display_name = "Preview"
-            else:
-                display_name = checkpoint_name.title()
-
-            try:
-                with open(pred_file) as f:
-                    data = json.load(f)
-                    # Handle predictions JSON structure
-                    if isinstance(data, dict) and "predictions" in data:
-                        predictions_by_checkpoint[display_name] = data["predictions"]
-                    else:
-                        predictions_by_checkpoint[display_name] = data
-                print(f"  Loaded predictions from {pred_file.name}")
-            except json.JSONDecodeError:
-                print(f"  Warning: Could not parse {pred_file.name}")
-
-        # Get capture info from training log
-        capture_id = "capture"
-        goal = "Complete the recorded workflow"
-        if log_file.exists():
-            try:
-                with open(log_file) as f:
-                    log_data = json.load(f)
-                capture_path = log_data.get("capture_path", "")
-                if capture_path:
-                    capture_id = Path(capture_path).name
-            except (json.JSONDecodeError, KeyError):
-                pass
-
-        _enhance_comparison_to_unified_viewer(
-            base_file,
-            predictions_by_checkpoint,
-            current_dir / "viewer.html",
-            capture_id,
-            goal,
-        )
-        print(f"\nGenerated: {current_dir / 'viewer.html'}")
+    # Generate unified viewer using consolidated function
+    viewer_path = generate_unified_viewer_from_output_dir(current_dir)
+    if viewer_path:
+        print(f"\nGenerated: {viewer_path}")
     else:
-        print("\nNo comparison.html found. Run comparison first or copy from capture directory.")
+        print("\nNo comparison data found. Run comparison first or copy from capture directory.")
 
     if args.open:
         webbrowser.open(str(current_dir / "viewer.html"))

openadapt_ml/training/trainer.py

@@ -1824,43 +1824,12 @@ def generate_unified_viewer_from_output_dir(output_dir: Path) -> Path | None:
         print("No comparison data found, cannot generate unified viewer")
         return None
 
-    # Generate the unified viewer by enhancing an existing comparison file
+    # Generate the unified viewer using standalone HTML template
+    # (Consolidated approach - always use standalone for reliability)
     viewer_path = output_dir / "viewer.html"
 
-    # Find the best existing comparison file to use as base (prefer epoch files)
-    # Check both main directory and archive folder
-    base_html_file = None
-    search_dirs = [output_dir, output_dir / "archive"]
-
-    for search_dir in search_dirs:
-        if not search_dir.exists():
-            continue
-        # Look for epoch files first (reverse sort = latest first)
-        for comp_file in sorted(search_dir.glob("comparison_epoch*.html"), reverse=True):
-            base_html_file = comp_file
-            break
-        if base_html_file:
-            break
-        # Fall back to comparison_preview.html
-        preview_file = search_dir / "comparison_preview.html"
-        if preview_file.exists():
-            base_html_file = preview_file
-            break
-
-    if base_html_file is None:
-        # Fall back to basic viewer
-        _generate_unified_viewer_from_extracted_data(
-            base_data=base_data,
-            predictions_by_checkpoint=predictions_by_checkpoint,
-            output_path=viewer_path,
-            capture_id=capture_id,
-            goal=goal,
-        )
-        return viewer_path
-
-    # Use the existing comparison file as base and enhance it with unified controls
-    _enhance_comparison_to_unified_viewer(
-        base_html_file=base_html_file,
+    _generate_unified_viewer_from_extracted_data(
+        base_data=base_data,
         predictions_by_checkpoint=predictions_by_checkpoint,
         output_path=viewer_path,
         capture_id=capture_id,
@@ -2307,9 +2276,9 @@ def _generate_unified_viewer_from_extracted_data(
         if (thinkMatch) thinking = thinkMatch[1].trim().substring(0, 150);
 
         if (clickSomMatch) {{
-            action = {{ type: 'click', element: `[${clickSomMatch[1]}]` }};
+            action = {{ type: 'click', element: `[${{clickSomMatch[1]}}]` }};
         }} else if (typeSomMatch) {{
-            action = {{ type: 'type', element: `[${typeSomMatch[1]}]`, text: typeSomMatch[2] }};
+            action = {{ type: 'type', element: `[${{typeSomMatch[1]}}]`, text: typeSomMatch[2] }};
         }} else if (typeSimpleMatch) {{
             action = {{ type: 'type', text: typeSimpleMatch[1] }};
         }} else if (clickCoordMatch) {{
@@ -2323,12 +2292,12 @@ def _generate_unified_viewer_from_extracted_data(
         let html = '';
         if (action) {{
             if (action.type === 'click' && action.element) {{
-                html = `<div style="font-weight:600;color:var(--accent);">CLICK(${action.element})</div>`;
+                html = `<div style="font-weight:600;color:var(--accent);">CLICK(${{action.element}})</div>`;
             }} else if (action.type === 'click' && action.x !== undefined) {{
                 html = `<div style="font-weight:600;color:var(--accent);">CLICK(x=${{action.x.toFixed(2)}}, y=${{action.y.toFixed(2)}})</div>`;
             }} else if (action.type === 'type') {{
-                const elem = action.element ? `${action.element}, ` : '';
-                html = `<div style="font-weight:600;color:var(--accent);">TYPE(${elem}"${{action.text}}")</div>`;
+                const elem = action.element ? `${{action.element}}, ` : '';
+                html = `<div style="font-weight:600;color:var(--accent);">TYPE(${{elem}}"${{action.text}}")</div>`;
             }} else if (action.type === 'raw') {{
                 html = `<div style="color:var(--accent);">${{action.text}}</div>`;
             }}
@@ -2637,6 +2606,10 @@ def _enhance_comparison_to_unified_viewer(
 ) -> None:
     """Enhance an existing comparison HTML file into a unified viewer.
 
+    DEPRECATED: This function uses script injection which is fragile.
+    Use _generate_unified_viewer_from_extracted_data() instead for a
+    standalone viewer that doesn't depend on the comparison.html structure.
+
     Takes the nice openadapt-capture viewer and adds:
     - Simplified nav (Training + Viewer only)
     - Checkpoint dropdown to switch between predictions