The Agent Skills Directory

[SAFE]: The skill implements a robust 'Engineering-Health Engine' that prioritizes repository integrity. It mandates starting from a 'green' (passing) state, uses the project's own scripts to ensure consistency with CI/CD, and applies changes in small, verifiable waves.
[SAFE]: Strong safety constraints are established in the instructions, such as 'Never auto-commit' and 'Never silence a check to make it pass'. These prevent the agent from making unreviewed changes or bypassing security/quality controls.
[COMMAND_EXECUTION]: The skill executes standard ecosystem tools (such as eslint, tsc, ruff, and mypy) to identify technical debt. This is the intended primary purpose of the skill and is performed using project-local configurations.
[PROMPT_INJECTION]: The skill includes instructions for the agent to adopt a 'smart caveman' persona and use specific emojis. These are benign stylistic constraints and do not attempt to bypass safety filters or override core agent behavior.
[INDIRECT_PROMPT_INJECTION]: As the skill processes repository source code and tool outputs, it possesses an indirect prompt injection surface.
Ingestion points: Source code files and linter/type-checker error logs (SKILL.md).
Boundary markers: The 'TRIAGE (gate)' step requires the agent to present findings in a table and receive explicit user approval before proceeding (health-engine.md).
Capability inventory: The skill can execute shell commands for linting/testing and perform file writes to apply fixes.
Sanitization: While explicit sanitization of tool output is not mentioned, the 'Gate before fixing' and 'Never auto-commit' rules provide a critical human-in-the-loop verification step that mitigates the risk of malicious instructions in codebase comments.

absolute-debt