The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes codebase evidence such as definitions, references, callers, and logs. While this exposure to untrusted data constitutes an indirect prompt injection surface, it is inherent to the tool's debugging purpose. The skill significantly mitigates this risk by requiring a strict, multi-phase user approval process—Phase 1 (stubs), Phase 2 (wiring), and Phase 3 (implementation)—before any changes are applied to the filesystem, ensuring a human-in-the-loop validation of all generated content.
[SAFE]: The instructions implement a comprehensive 'Mutating action gate' and 'Approval workflow' that forces the agent to wait for explicit user intent tokens like 'approve' or 'confirm' before performing semantic changes, formatting updates, or command execution. It also defines objective 'review-fatigue' triggers that automatically reduce the number of files modified per batch if the diff size becomes too large for a human to review effectively.

duck-debug