The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines procedures for the agent to ingest conversation transcripts and project metadata, which could contain instructions designed to influence agent behavior.
Ingestion points: host.frames.get (retrieves transcript messages), host.artifacts (retrieves filenames and user upload metadata), and host.sessions.list (retrieves session titles).
Boundary markers: Absent. The skill does not provide specific instructions to the agent on how to differentiate between legitimate data and embedded instructions within these sources.
Capability inventory: The skill utilizes repl_execute to invoke the host API, which has capabilities for file path resolution (host.artifactPath), compute management (host.compute), and messaging (host.sendFrameMessage).
Sanitization: Absent. While the documentation mentions "sanitized runtime segment projections," this does not prevent the agent from interpreting the content of messages or metadata as commands.
[DYNAMIC_EXECUTION]: The skill relies on the execution of JavaScript code snippets within a control REPL to interact with the host environment.
Patterns: The instructions provide various code templates for the agent to use with the repl_execute tool, such as await host.capabilities() and await host.lineage.graph(versionId).

self-awareness