The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads the DeepSWE benchmark corpus from github.com/datacurve-ai/deep-swe and installs the datacurve-pier tool via PyPI or its GitHub repository.
[COMMAND_EXECUTION]: The skill executes several CLI commands, including git clone, uv tool install, and pier run, to set up the environment and run the benchmarking agent.
[INDIRECT_PROMPT_INJECTION]: The skill processes external coding tasks which represents a surface for indirect prompt injection.
Ingestion points: Task definitions are ingested from the tasks/ directory within the cloned deep-swe repository (cloned in SKILL.md).
Boundary markers: The skill requires Docker to be running, noting that pier sandboxes each task in a Docker container by default to isolate execution.
Capability inventory: The skill uses git (network/filesystem), uv (tool installation), and pier (subprocess execution) while communicating with the OpenRouter API.
Sanitization: The primary mitigation is the execution isolation provided by the pier tool's built-in Docker sandbox environment.

run-deep-swe