The Agent Skills Directory

Command Execution: The skill performs local command execution using tools like uv run, python -c, and grep to verify that documentation claims (such as CLI flags and code symbols) match the source of truth. It also utilizes git and gh for repository management. These operations are limited to the local development environment and are essential for the audit process.
Indirect Prompt Injection Surface: The skill ingests untrusted data from documentation files in docs/src/content/docs/ and packages/apm-guide/.apm/skills/apm-usage/. While the subagents are provided with strict task-oriented instructions, the processing of this content creates a surface for indirect prompt injection. The skill manages this risk through the use of discrete boundary instructions in subagent prompts and strict validation of return data against a JSON schema.
Controlled Repository Modification: The skill is designed to apply surgical fixes to documentation files. To maintain security, it enforces a single-writer interlock where only the orchestrator agent is authorized to perform git write operations (commit and push), and all changes are directed to a working branch for maintainer review.

docs-corpus-audit