The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions direct the agent to treat the pyproject.toml file as the primary source of truth for project configuration. This creates a potential surface for indirect prompt injection, where a malicious project configuration could override agent behavior or influence its decision-making during development tasks (e.g., by manipulating linting rules, test execution parameters, or build settings). \n
Ingestion points: The agent reads pyproject.toml, uv.lock, and various Python source files to determine project state (SKILL.md, references/uv-projects.md). \n
Boundary markers: The instructions lack explicit warnings to disregard potentially malicious configuration settings or comments embedded in these files. \n
Capability inventory: The skill utilizes powerful tools like uv run, pytest, ruff, and mypy which can execute code or change project structure based on the ingested configuration (SKILL.md, references/ruff.md). \n
Sanitization: No explicit validation or sanitization steps are defined for the project configuration data before it influences the agent's workflow. \n\n- [EXTERNAL_DOWNLOADS]: The skill provides environment setup instructions that include downloading and executing an installation script for the uv tool. \n
Evidence: curl -LsSf https://astral.sh/uv/install.sh | sh (SKILL.md). \n
Analysis: The download targets astral.sh, which is the official and well-known provider for the uv and ruff toolchain. This represents standard setup procedure for these tools.

python