The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted session history and audit logs to extract 'Generalized Rules' that are then stored in the agent's long-term semantic memory. This creates a surface where an attacker providing malicious input into logs could influence the agent's future behavior across all subsequent sessions.
Ingestion points: Processes content from AUDIT_LOG.md and general Session History during the 'Replay' phase.
Boundary markers: No explicit delimiters or boundary markers are defined in the execution steps to separate untrusted log content from the instruction generation process.
Capability inventory: The skill possesses the ability to modify persistent files (memory/semantic/rules.md and PROJECT_CONTEXT.md) and delete local files (temporary scratchpads and tmp/ files).
Sanitization: While the skill's documentation suggests validating extracted rules against governance constraints, the operational steps do not include automated sanitization or filtering of the input data.

episodic-consolidation