The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill instructions recommend sourcing ~/.zshrc to load the OPENROUTER_API_KEY. This involves executing the user's shell configuration file, which often contains sensitive credentials and environment variables.
[EXTERNAL_DOWNLOADS]: The benchmark requires downloading external components, including cloning the deep-swe repository and installing the pier tool from GitHub or PyPI.
[COMMAND_EXECUTION]: The core functionality involves executing shell commands to install tools (uv, git), set up the environment, and run the benchmark via the pier CLI.
[INDIRECT_PROMPT_INJECTION]: The skill processes coding tasks from the DeepSWE benchmark corpus, creating a vulnerability surface where instructions within the tasks could influence agent behavior.
Ingestion points: Task definitions located in deep-swe/tasks (referenced in SKILL.md).
Boundary markers: No explicit boundary markers are used in the prompt to separate task data from instructions.
Capability inventory: Uses pier and mini-swe-agent to execute code and shell commands, typically within Docker containers.
Sanitization: Employs Docker sandboxing for process isolation.

run-deep-swe