The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to locate 'verify skills' within project-specific directories (such as .agents/skills/ and .claude/skills/) and execute the shell commands defined within those external files.
[INDIRECT_PROMPT_INJECTION]: The skill analyzes untrusted content from commit messages and code diffs to determine how to update verification models, which could be exploited by an attacker to influence the agent's logic through crafted commits.
Ingestion points: Reads git commit messages and file diffs between the sync commit and the current head.
Boundary markers: Absent; the instructions do not specify any delimiters or warnings to ignore embedded instructions within the commit messages or diffs.
Capability inventory: The skill can execute commands found in other skill files and modify local verification files (models, harnesses, property files).
Sanitization: Absent; commit messages and diffs are analyzed directly for meaning without escaping or validation.
[DYNAMIC_EXECUTION]: The skill includes instructions to dynamically resolve and follow instructions from relative filesystem paths (e.g., ../<name>/SKILL.md) based on patterns found in text.

maintain-formal-verification-skill