Blame for AGENTS.md

# AGENTS.md

## HDR background retention

HDR environment maps are retained for benchmark runtime generalization even when
no current task or USD statically references them. Do not prune `.hdr` maps under
`background/` based only on dependency reachability. Preserve EXR environment maps
as well. Keep scene-embedded map paths so DomeLight directory sampling still
works. This exception does not restore excluded robots, external packages, or
unused scene USDs.

## Data collection example retention

Data collection examples and replay presets are release roots alongside
benchmark tasks. Retain their complete candidate asset catalogs, metadata,
USD-attached behavior scripts, canonical interaction/grasp data, and configured
HDR background entry layers. Follow recursive USD, material, texture, and script
references before pruning. An asset need not occur in a sampled benchmark scene
to be required by an example task or its other scene instances.

## Engine pipeline retention

`genie_sim_engine_*` bringup scenes are release roots alongside benchmark tasks
and data collection examples. Preserve these configured entry assets and their
complete recursive USD, material, texture, mesh, and attached-script dependencies:

- `scenes/GDK/demo.usda` for pick-and-place and whole-body-control scenes.
- `scenes/blank/blank_cloth.usda` and `scenes/blank/T_Shirt_fold.usd` for the
  Swiftpicker shirt scene.
- `background/home/fried_noodles/background.usda` for the Wuji chef scenes.
- `robot/G2_crsB_wuji/robot.usda` for the chef evaluation visual robot.

The cloth scene references the retained local gridroom at
`background/common/environments/gridroom_black/gridroom_black.usd`, including its
`Materials/Textures` maps. Keep that relative reference local and portable.
Do not replace a missing configured world with a generated flat scene or prune
these dependencies merely because benchmark tasks do not reference them.

## Runtime assembly cache

Engine-generated assemblies belong in `scenes/<scene_stem>/`, including
`manifest.json`, robot USD/URDF, render layers, and payloads. Keep this complete
cache together; manifest paths and generated USD references must resolve after
moving it. Generated `scene_*` directories are ignored by `scenes/.gitignore`.
Retained authored entry assets such as `scenes/GDK/` and `scenes/blank/` remain
tracked; categorized backgrounds and shared assets stay under `background/`.

## Build, Lint, and Test Commands

### Running Tests
```bash
# No explicit test framework found - the project consists mainly of asset configuration files.
```

### Code Style and Conventions

### Python Code Style
- **Formatting**: Use standard PEP 8 style with 4-space indentation
- **Naming Conventions**:
  - `ITEM` dictionary for object metadata (lowercase with underscores)
  - `find_all_items()` and `extract_item_dict()` for function names (snake_case)
  - `usda_path`, `desc_path`, `combined_dict` for variable names (snake_case)
  - `ASSETS_PATH`, `CHROMA_DB_PATH`, `ASSETS_INDEX` for constants (uppercase_with_underscores)
- **Data Structure Style**:
  - Use dictionary definitions with `ITEM = { ... }` format in `item.py` files
  - Keep dictionaries nested with clear keys for objects, shapes, properties
  - Use list format `["name", "position", "quaternion", "scale", "size", "type"]` consistently
  - Include `description` key as empty dict `{}` when no description present
  
### Python Imports
- Import standard library modules at top of file: `os`, `ast`, `hashlib`
- Import pathlib for path operations: `from pathlib import Path`
- Use absolute imports for core functions
- No external library imports in asset files (only dictionaries)

### File Organization
- Asset files: `objects/{category}/{category}_{id}/{id}/item.py`
- Background files: `background/common/{category}/{category}_{id}/item.py`
- Keep `.usda`, `.usd`, `.json`, and texture files in same directory as corresponding item
- Organize by object category with unique identifiers

### Code Organization Principles
```python
# Dictionary-based configuration files
ITEM = {
    "description": {},                # Optional description dict
    "id": "{object_id}",              # Unique identifier
    "shapes": [                       # List of mesh definitions
        {
            "name": "{shape_name}",    # e.g., "bbox", "origin"
            "position": [x, y, z],     # Float position coordinates
            "quaternion": [w, x, y, z],# Quaternion for rotation
            "scale": [x, y, z],       # Scale multipliers
            "size": [x, y, z],        # Physical dimensions
            "type": "{geometry_type}", # e.g., "cube", "sphere"
        }
    ],
    "size": [],                       # Additional size specification
    "url": "path/to/object.usda"      # Path to USD file
}
```

### Error Handling
- Use try-except for file reading operations
- Print informational messages for warnings and errors
- Skip files with errors instead of raising exceptions
- Provide meaningful error messages with file paths

### Type Hints
- Use standard Python types in function signatures
- Path objects with `Path` type hint
- Dictionary types with `dict` hint
- List types with `list` hint
- Number types with `float` or use `number` from typing

### Documentation
- Document complex functions with docstrings
- Use descriptive comments for special cases
- Add error messages inline for critical operations
- Include WARN and ERROR prefixes for important messages

### Asset Tools (Scripts)

The `tools/` directory retains only `.gitignore` and these two standalone scripts:

- `tools/generate_recorder_usda.py`: generate recorder stages and cameras for assets.
- `tools/generate_object_snapshots_and_video.py`: render snapshots and preview videos
  from recorder stages using an Isaac Sim runtime.

Run them from a prepared runtime with `geniesim_assets` installed. The rendering
script starts Isaac Sim; do not import it for a metadata or syntax-only check.