The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill includes functionality to self-evolve by generating new skill files at runtime. In Phase 5, it explicitly calls the tool kilo_synthesize_skill to automatically generate new SKILL.md files when a reusable architectural solution is identified. This represents dynamic script generation that alters the agent's future capabilities.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources during its research phases, creating a vulnerability surface for indirect instructions.
Ingestion points: External grounding data is fetched from GitHub repositories and llms.txt files (Phase 2 and 3).
Boundary markers: The instructions do not specify any delimiters or safety markers to differentiate between the agent's instructions and data retrieved from external research.
Capability inventory: The agent has the ability to write to internal memory (kilo_remember_fact), persist reflections (kilo_record_reflection), and generate new skill files (kilo_synthesize_skill).
Sanitization: There is no mention of sanitization or validation logic for the content retrieved from external GitHub repositories before it is synthesized into architectural decisions or new skills.
[COMMAND_EXECUTION]: The skill utilizes browser automation for verification. Phase 4 mentions the use of Playwright for End-to-End (E2E) testing to verify implementation results.
[EXTERNAL_DOWNLOADS]: The skill performs research by fetching data from external services. In Phase 2, the invoke_subagent tool is instructed to retrieve patterns and design samples from GitHub repositories (specifically those with 10k+ stars) and llms.txt documentation.

grounded-research-benchmark