The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads demonstration assets and pre-trained model weights from several well-known and trusted external services.
Fetches video assets and model weights from Hugging Face (huggingface.co) in scripts/prepare_demo_assets.sh and assets/configs/osmo/setup_model_cache.yaml.
Downloads specific perception model weights from Google Cloud Storage (storage.googleapis.com) and Google Drive (drive.usercontent.google.com) in the model cache setup workflow.
[OBFUSCATION]: A model cache setup configuration file utilizes string concatenation to hide a remote resource identifier.
assets/configs/osmo/setup_model_cache.yaml assembles a Google Drive ID (REID_GDRIVE) using "".join() on two separate string literals. This technique obscures the full download URL from simple string-based scanners.
[INDIRECT_PROMPT_INJECTION]: The skill provides an attack surface for indirect prompt injection as it processes external video data and uses that data to influence automated labeling and augmentation via Vision-Language Models (VLMs).
Ingestion points: Untrusted data enters the workflow via the dataset_url and storage_url parameters which point to external video files.
Boundary markers: Prompts in assets/cookbooks/*/auto_labeling/prompts/event_analysis.md use strict JSON formatting and explicit instructions to separate metadata from event analysis, providing some mitigation against obedience to embedded data instructions.
Capability inventory: The workflow tasks have access to powerful tools including the osmo CLI for data management, ffmpeg for video rendering, and custom Python/Shell execution within containerized workers.
Sanitization: scripts/pre_submit_guard.py performs basic validation of video names to ensure they do not contain path separators, reducing local file system manipulation risks.
[COMMAND_EXECUTION]: The skill executes standard system commands and platform-specific CLI tools as part of its intended orchestration role.
Uses osmo credential set in scripts/preflight_credentials.sh to manage environment-specific secrets securely within the platform.
Spawns ffmpeg subprocesses in scripts/render_side_by_side.sh for generating visualization artifacts.
Executes osmo data list via Python's subprocess.run() in scripts/pre_submit_guard.py to validate dataset availability prior to workflow submission.

physical-ai-video-data-augmentation