The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes patterns for ingesting application state from the local file system using serde_json. If these files are modified by an external process or attacker, it could influence application behavior during restoration.
Ingestion points: persistence::load_app_state in SKILL.md reads data from the user's data directory.
Boundary markers: The pattern lacks explicit integrity checks or boundary markers to distinguish between system-generated state and potentially injected content.
Capability inventory: The provided patterns include file system read (tokio::fs::read) and write (std::fs::File::create, std::fs::write) capabilities across the codebase.
Sanitization: The implementation relies on standard serde deserialization structure as the primary validation mechanism, which does not prevent malicious data from being interpreted as valid state fields.

robius-state-management