The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it reads untrusted data (character cards, episode beats, notes) from Busabase storage and incorporates it into agent context.
Ingestion points: Data enters the context through "content/kelly-drama-app/app/js/providers/busabase-provider.js" and "scripts/lib/drama-busabase.mjs" during project state synchronization.
Boundary markers: No delimiters or strict boundary markers are used when interpolating this data into prompts within "content/kelly-drama-app/app/js/drama-model.js".
Capability inventory: The skill can execute local commands via "spawn" and "execFile", perform network requests to external APIs, and read/write files in the local environment.
Sanitization: No input validation or escaping of the ingested data was identified.
[DYNAMIC_EXECUTION]: The script "scripts/read_hyperframe_status.mjs" accepts an arbitrary filesystem path from the project configuration and performs recursive directory scanning. This behavior allows for the discovery and metadata exposure of files outside the skill's intended workspace if the path is manipulated.
[COMMAND_EXECUTION]: The skill's generation and status monitoring logic spawns external subprocesses, such as "ffprobe" and "python3". The arguments passed to these tools are derived from data retrieved from the project database, representing a risk if the data contains malicious payloads.
[DATA_EXFILTRATION]: The skill is designed to send project metadata and prompts to external AI endpoints at "moonrouter.dev" and "ark.ap-southeast.bytepluses.com". While this is intended for image and video generation, users should be aware that their project content is shared with these services.
[EXTERNAL_DOWNLOADS]: The skill fetches AI models from Hugging Face's official repository at runtime using the "mlx_audio" library for local voice generation. This is documented and utilizes a well-known model repository.

kelly-drama