The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data in the form of unified diffs and plan drafts. These inputs serve as vectors for potential malicious instructions hidden in code comments or strings intended to influence the agent's analysis or file-writing behavior.
Ingestion points: Data enters the context via the diff, intent_summary, and plan.md inputs across all execution modes.
Boundary markers: The instructions do not define explicit delimiters or 'ignore' directives to prevent the underlying model from following instructions contained within the analyzed content.
Capability inventory: The skill possesses powerful capabilities, including the Edit and Write tools for modifying the local codebase and memory.write for persistent state storage.
Sanitization: There is no documented mechanism for sanitizing or filtering untrusted input content before it is processed by the AI.
[COMMAND_EXECUTION]: In its apply mode, the skill executes shell commands to run scoped tests and type-checks (e.g., pnpm test, tsc, pytest) to verify its changes. While these actions are intended as safety checks, they involve dynamic execution based on file extensions and paths.
[PERSISTENCE]: The skill implements a self-improvement loop via the 'LoreKit' memory system. It persists 'lessons' (judgments and calibration data) across sessions using memory.write, which influences future agent behavior. While this is a core feature, it provides a persistent channel for information to survive between independent task executions.

optimize-approach