The Agent Skills Directory

[SAFE]: The skill implements a multi-agent workflow to analyze codebase architecture. Security is maintained by configuring all spawned subagents (explorers, explainers, and critics) with 'readonly: true', which ensures that the agent can only read the codebase and cannot perform any destructive operations or modifications.
[SAFE]: The prompts for the subagents are well-defined and focus on technical fact-gathering and architectural review. There are no patterns of prompt injection, obfuscation, or unauthorized data access beyond the skill's stated purpose of analyzing the codebase.
[SAFE]: No external dependencies, remote script downloads, or unauthorized network operations were identified. The skill relies on standard codebase exploration tools (Glob, Grep, Read) provided by the agent environment.
[INDIRECT_PROMPT_INJECTION]: The skill processes codebase content, creating a theoretical indirect prompt injection surface. However, this is the core function of the skill. Ingestion points: Codebase files accessed via exploration tools. Boundary markers: The prompts guide subagents toward factual mapping but do not include explicit instructions to ignore instructions found within the code. Capability inventory: Restricted to read-only subagents. Sanitization: None mentioned, relying on model guardrails. This is considered safe due to the restricted execution environment.

how