The Agent Skills Directory

[SAFE]: The skill implements a robust 'find-and-report' only policy. The documentation and prompt templates explicitly instruct the agents NOT to modify any code or perform fixes, ensuring it operates purely as an analytical tool.
[SAFE]: The skill uses disk-mediated dispatch (as per the referenced shared/dispatch-convention.md) for all subagent communication. This ensures that agent interactions are visible and mediated through the filesystem rather than hidden in-memory channels.
[SAFE]: High levels of transparency are maintained through a mandatory communication requirement. The orchestrator must output status updates and write to a pipeline-status.md file after every agent dispatch and completion, allowing the user to monitor all background activities.
[SAFE]: The skill implements strict user-in-the-loop gates. Phase 2 (Analysis) cannot proceed until the user has reviewed and confirmed the subsystem manifest generated in Phase 1 (Scoping).
[SAFE]: Context management and resource limits are well-defined. Hard caps on prompt content (1500 lines) and agent budgets (~20 agents) are enforced to prevent excessive resource consumption and ensure model performance.
[SAFE]: The skill includes comprehensive 'Blind Spots' and 'Synthesis' phases that specifically hunt for cross-cutting concerns like security, performance, and concurrency that might be missed by individual analysis lenses.

audit