The Agent Skills Directory

[SAFE]: The documentation enforces a clear security boundary by distinguishing between administrative Management Tokens and participant-scoped Peer Tokens, explicitly warning developers to keep the former restricted to backend environments and to rotate them if compromised.
[EXTERNAL_DOWNLOADS]: The skill references and retrieves configuration files, such as OpenAPI specifications and protobuf definitions, from official repositories under the fishjam-cloud organization. These are recognized as legitimate vendor resources and do not pose a security risk.
[PROMPT_INJECTION]: The Fishjam platform supports custom metadata for peers and tracks, as described in glossary.md and notifications-taxonomy.md. This represents a surface for indirect prompt injection if the data is subsequently processed by an agent without sanitization.
Ingestion points: Metadata fields for peers (peer.metadata) and tracks (track.metadata) found in SKILL.md, glossary.md, and notifications-taxonomy.md.
Boundary markers: Not present; the documentation defines metadata as an opaque string or object.
Capability inventory: The skill provides documentation on managing room lifecycles and media streaming via SDKs but contains no executable code that processes this metadata.
Sanitization: Not addressed in the platform documentation; users of the SDK are expected to handle data validation independently.

fishjam-platform