The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input by reading and executing commands defined in the docs/design/stage-specs/S{N}.md specification files. This design allows for instructions embedded in documentation to influence agent behavior and trigger command execution.
Ingestion points: The 'DoD' (Definition of Done) section of docs/design/stage-specs/S{N}.md.
Boundary markers: Absent; the skill is instructed to extract and execute every - [ ] assertion without filtering or validation.
Capability inventory: Execution of npx tools and arbitrary shell commands defined under 'Behavioral Acceptance'.
Sanitization: Absent; there is no validation or escaping of the commands extracted from the markdown specifications.
[COMMAND_EXECUTION]: The skill uses npx to execute development tools such as vitest, typescript (via tsc), and playwright. Additionally, it is designed to execute arbitrary commands or tests specified within the behavioral acceptance section of external markdown files, posing a risk of unauthorized shell execution if the source files are untrusted.

stage-gate