The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill executes shell commands and inline bash scripts specified in checkpoints.yaml files via the run-checkpoints.sh script. A safety filter is implemented in scripts/lib/command-allowlist.sh that blocks dangerous patterns such as sudo, rm -rf, and piping network downloads directly to a shell. However, the allowlist permits general-purpose interpreters including php -r, python3, and node, which are necessary for the skill's primary assessment functions.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted project files (source code, READMEs, configuration files) and passes their content to LLM-based domain agents for review.
Ingestion points: Files within the project directory are accessed using Read, Glob, and Grep tools.
Boundary markers: Domain agents are provided with a structured prompt template that explicitly instructs them to only report compliance status and not to perform any code modifications.
Capability inventory: The skill possesses the ability to execute shell commands (via a restricted Bash tool) and use jq and gh CLI tools.
Sanitization: Project data is interpolated into prompts without dedicated sanitization, relying on the agent's task-specific instructions to mitigate risks.
[COMMAND_EXECUTION]: The skill automates the execution of common development and quality tools such as composer, phpunit, phpstan, php-cs-fixer, and rector. These commands are typically resolved relative to the project's vendor/bin directory.
[EXTERNAL_DOWNLOADS]: The follow_uses feature in the checkpoint runner fetches content from remote GitHub workflows using the gh api. This process uses the official GitHub API and the operator's local authentication context to inspect reusable workflows.

automated-assessment