The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands to compute CRAP scores and identify code changes. Specifically, it uses gh pr diff, git diff, and a bundled Python script, scripts/compute_crap.py. It also references an external script at plugins/engineer/scripts/dae_guard.py for quality gate validation.
[REMOTE_CODE_EXECUTION]: To generate missing coverage data, the skill identifies and executes the project's native test runner (e.g., npm test, pytest, mvn test, go test). This involves executing code or scripts defined within the repository being analyzed. The skill includes a clear safety protocol requiring the agent to ask for user permission before running any such tests.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface as it processes code and diffs from the repository to generate refactor proposals.
Ingestion points: Git diffs and source code files from the local repository (e.g., as determined in SKILL.md step 1).
Boundary markers: The subagent prompt in references/subagent-prompt.md uses placeholders (<...-body>) but lacks explicit delimiters to distinguish code from instructions.
Capability inventory: The main agent has the ability to execute shell commands and write to the filesystem via edit tools.
Sanitization: Code content is interpolated into prompts for subagents without filtering or sanitization. While this surface exists, the risk is mitigated by the specific scope of the task (drafting refactors and tests).

crap-analyzer