The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands to collect application metrics. This includes running the project's full test suite via bundle exec rspec or bundle exec rails test, and running bundle exec rubycritic. Executing a test suite involves running arbitrary code defined within the audited project.
[EXTERNAL_DOWNLOADS]: The skill may perform external package installations using bundle install to set up simplecov and rubycritic if they are not already present in the project. While these are well-known development tools, the process involves fetching packages from external registries.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze entire codebases, which act as untrusted input. A malicious codebase could contain embedded instructions in comments or strings designed to manipulate the agent's audit report or behavior.
Ingestion points: Reads files from app/, spec/, test/, config/, db/, and lib/ (SKILL.md).
Boundary markers: The skill does not explicitly define delimiters for external code when processing it for analysis.
Capability inventory: Uses Glob, Grep, and Read tools for analysis, shell execution for metrics collection, and Write to generate RAILS_AUDIT_REPORT.md.
Sanitization: There is no explicit sanitization or filtering mentioned for the content read from the codebase before it is used to generate the report.
[DYNAMIC_EXECUTION]: The subagents dynamically modify the project's configuration files at runtime, specifically prepending instrumentation code to test helpers (spec/rails_helper.rb or test/test_helper.rb) and modifying the Gemfile to inject new dependencies.

rails-audit-thoughtbot