The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands for repository management and automated testing. It uses git for branching and commits, gh (GitHub CLI) for pull request creation, and a configurable <test_command> to run the project's test suite. These operations are essential to the skill's purpose as a software engineering tool and are triggered based on a user-reviewed configuration file (loop.run.yaml).
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection attack surface, as it processes untrusted content from the user prompt and the target repository. This is a common characteristic of agentic development tools.
Ingestion points: Untrusted data is ingested when the subagents read the user's coding prompt and existing repository source code in roles/engineer.md and roles/qa.md.
Boundary markers: The skill does not implement specific delimiters or 'ignore' instructions to isolate ingested content from the subagents' primary instructions.
Capability inventory: The skill has broad capabilities, including modifying source and test files and executing arbitrary shell commands via the <test_command>, git, and gh tools.
Sanitization: No explicit sanitization or filtering of ingested file content or prompt text is performed prior to processing.
[SAFE]: The included Python script tools/quality_check.py performs static analysis of code using the standard library ast module. It does not execute the files it analyzes and does not use any external dependencies or network operations.

swe-loop