The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The Warden tool analyzes code changes, configuration files (warden.toml), and local skill definitions (.agents/skills/) which can be controlled by contributors to a repository. A malicious actor could inject instructions into these files to influence the agent's behavior during a code review.
Ingestion points: The tool processes the warden.toml configuration, local skill definitions in .agents/skills/, and the source code files being analyzed.
Boundary markers: The documentation does not specify the use of delimiters or 'ignore' instructions when processing content from these files.
Capability inventory: According to creating-skills.md, Warden agents may have access to powerful tools including Bash, Write, Edit, WebFetch, and WebSearch to perform analysis and apply fixes.
Sanitization: No explicit sanitization or validation of the instructions contained within project files is described.
[EXTERNAL_DOWNLOADS]: The warden add --remote and warden sync commands facilitate the fetching and synchronization of skill definitions (markdown instructions) from external GitHub repositories. These instructions are cached locally and subsequently followed by the AI agent during analysis tasks.
[COMMAND_EXECUTION]: The skill instructions describe the use of the warden CLI, which performs file system operations, git commands, and can launch a local web server (via setup-app) to facilitate GitHub App manifest flows.

warden