The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data from benchmark submissions, including JSONL files and trial transcripts, which introduces a potential surface for indirect prompt injection where malicious content in the submissions could attempt to influence the agent's editorial actions.\n
Ingestion points: submissions/ directory containing results.jsonl and gzipped transcripts fetched from the Shpigford/nurb-benchmarks repository (SKILL.md).\n
Boundary markers: The instructions do not define clear boundaries or specify that the agent should ignore instructions found within the submitted data during parsing or re-grading.\n
Capability inventory: The skill possesses the ability to execute commands via uv and git, write files to the local and cloned repositories, and open pull requests on GitHub.\n
Sanitization: While the skill includes instructions to sanitize artifacts by removing local file paths (e.g., /Users/), it does not implement specific sanitization or filtering for adversarial instructions within the transcripts or metadata.\n- [COMMAND_EXECUTION]: The skill utilizes shell commands to manage the benchmarking environment and execute scripts.\n
Evidence: git clone, git log, uv sync, uv run python, and uv run pytest are used throughout the workflow to update code and run evaluation tools.\n
Context: These commands facilitate the primary purpose of the skill and interact with tools provided in the author's own repository.\n- [EXTERNAL_DOWNLOADS]: The skill fetches code and data from a remote source at runtime.\n
Evidence: Clones the Shpigford/nurb-benchmarks repository from GitHub.\n
Context: The external repository is a vendor resource belonging to the skill's author.\n- [REMOTE_CODE_EXECUTION]: The skill downloads and executes code from a remote repository.\n
Evidence: Clones Shpigford/nurb-benchmarks and runs python modules using uv run.\n
Context: As the repository belongs to the skill author, this represents expected vendor functionality.

leaderboard