The Agent Skills Directory

[PROMPT_INJECTION]: The skill's primary mechanism creates a significant surface for indirect prompt injection by learning from untrusted user interactions.\n
Ingestion points: User-provided edits to code and text are monitored via PostTool hooks in references/evolution-analysis.md.\n
Boundary markers: The system lacks delimiters or instructions to ignore embedded commands within user edits, making it possible for users to intentionally inject patterns.\n
Capability inventory: The skill uses Write and Edit tools to modify SKILL.md instruction files and create new documentation in the references/ directory.\n
Sanitization: There is no validation or sanitization of the patterns detected in user edits before they are suggested for integration into the skill's base instructions.\n- [COMMAND_EXECUTION]: The skill manages agent instructions at runtime and executes local shell scripts based on user input.\n
Dynamic Instruction Modification: The skill incorporates an 'Auto-Evolution' feature where high-confidence (0.85+) suggestions are applied to modify SKILL.md files automatically, potentially altering the agent's security posture or behavior without direct oversight.\n
Local Script Execution: Subcommands directly invoke shell scripts (evolution-engine.sh and version-manager.sh) located in the project's hidden configuration directory to perform operations on the filesystem.

skill-evolution