The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and audit persistent memory files (MEMORY.md, .agents/memory.md) and JSON checkpoints, which represent an ingestion surface for untrusted data from previous sessions or external tools.
Ingestion points: Targets MEMORY.md, .agents/memory.md, and compact-working-state.json files for ingestion and compaction.
Boundary markers: The skill explicitly instructs the agent to use provenance tags and delimiter isolation to prevent untrusted content from becoming axiomatic.
Capability inventory: The skill uses read_file, write_file, and edit_file to perform its memory management functions.
Sanitization: Includes robust sanitization protocols, including stripping prompt injection payloads (e.g., <system>, IGNORE PREVIOUS), performing Shannon entropy checks ($H > 4.5$) for secret detection, and implementing regex-based credential scanning.
[DATA_EXPOSURE]: The skill includes a 'Zero Secret Ingestion' policy that actively prevents the exposure of credentials by scanning memory updates for API keys, tokens, and high-entropy strings before writing to disk.

memory-hygiene