The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documents and implements patterns for fetching benchmark data from external, well-known, and trusted sources. This includes downloading CSV datasets from OpenAI's public blob storage (openaipublic.blob.core.windows.net), scientific data from HuggingFace (huggingface.co), and cloning benchmark repositories from GitHub. These operations align with the skill's primary purpose of porting existing benchmarks to the Harbor platform.
[REMOTE_CODE_EXECUTION]: The documentation for Pytest-based verification scripts includes a command to install the uv tool by piping a remote shell script to the shell: curl -LsSf https://astral.sh/uv/0.9.7/install.sh | sh. While this pattern typically presents a risk, it is the official installation method for a well-known development tool maintained by Astral.
[INDIRECT_PROMPT_INJECTION]: The adapter creation process inherently involves processing external, potentially untrusted benchmark data and interpolating it into agent instructions (instruction.md). This creates an attack surface where a malicious benchmark instance could influence the agent's behavior.
Ingestion points: Data is loaded in adapter.py and run_adapter.py using pandas, the HuggingFace datasets library, or local file readers.
Boundary markers: The skill uses Markdown templates but lacks explicit boundary markers or system instructions to ignore embedded commands in the benchmark data.
Capability inventory: The generated tasks can execute arbitrary shell commands via test.sh and have network access enabled by default (allow_internet = true in task.toml).
Sanitization: Examples demonstrate escaping shell-unsafe characters for the reference solution, but general benchmark data is inserted into instructions without intensive sanitization.
[DYNAMIC_EXECUTION]: The skill generates executable content at runtime, including shell scripts (test.sh, solve.sh) and Dockerfiles. These scripts are built by substituting benchmark data into pre-defined templates. This dynamic generation is a core feature of the adapter system but contributes to the overall risk profile when processing external data.
[COMMAND_EXECUTION]: Generated verification scripts frequently perform command-line operations such as environment setup using apt-get, numerical evaluation with pytest, and package management with uv.

harbor-adapter-creator