The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze user-supplied source code, establishing a surface for indirect prompt injection.
Ingestion points: The tool accepts a user-defined target file, directory, or subsystem as its input (SKILL.md).
Boundary markers: The skill mandates grounding in references/security-rubric.md and enforces a "Scope Gate" that halts execution if no explicit target is provided, limiting the agent's exposure to intended content.
Capability inventory: The skill can suggest and route code changes via the /changeset and /forge tools, can write findings to .agents/scratch/ dossiers, and can spawn research subagents.
Sanitization: Results are synthesized through a strictly defined canonical output contract with assigned security posture levels, providing structure to the final analysis turn.
[DYNAMIC_EXECUTION]: The skill implements a "Subagent Fan-Out" mechanism to handle large multi-package subsystems. It dynamically launches concurrent research subagents seeded with specific coordinates and roles, representing a platform-native dynamic execution pattern for parallel task processing.

inspect-security