The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands for toolchain activation via mise, dependency installation, and health checks from the cloned project. It also runs a local script for telemetry: python3 utilities/skill-builder/scripts/record_skill_feedback.py as part of the post-run protocol.
[REMOTE_CODE_EXECUTION]: The primary purpose of the skill is to clone and execute setup logic from user-provided GitHub repositories. This is an intended function with documented safety constraints to prevent destructive system changes.
[PROMPT_INJECTION]: The skill includes defensive test cases in references/evals.yaml to ensure the agent resists prompts that attempt to bypass safety checks or skip verification steps. It also identifies an indirect prompt injection surface through the processing of untrusted repository content.
[PROMPT_INJECTION]: Indirect surface evaluation: (1) Ingestion points: all project files and instructions within the cloned GitHub repository. (2) Boundary markers: no explicit markers are defined to isolate untrusted repo content from agent instructions. (3) Capability inventory: includes subprocess execution for git, mise, and package managers, along with file system and network access. (4) Sanitization: includes instructions to automatically redact secrets from all artifacts and logs.
[SAFE]: The skill follows security best practices by implementing a 'fail-fast' approach on blockers, redacting secrets by default, and using a task profile that requires manual human sign-off for high-impact actions.

bootstrap