The Agent Skills Directory

[SAFE]: The skill implements a rigorous 'Mutating Action Gate' that requires explicit user approval for any semantic changes, including logic modifications, configuration updates, and dependency changes. This workflow is further protected by a 'Preflight' check and phase-based caps on the number of files and lines changed per approval.
[SAFE]: A dedicated 'Trace mode' facilitates read-only evidence gathering (definitions, references, callers). This mode is explicitly restricted from suggesting fixes or implementation details, providing a secure method for analyzing codebase structure without risk of accidental modification.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted artifacts like logs and stack traces (SKILL.md). While boundary markers like evidence IDs (E1, E2) are present, the skill possesses capabilities like file writes and git commands. Risk is mitigated by a mandatory evidence-first reasoning process and a strict human-in-the-loop approval system that sanitizes outputs before execution.

duck-debug