The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill framework processes external inputs such as user requests and project documentation (Ingestion points: references/current-guidance.md). It possesses capabilities to modify files and execute verification commands (Capability inventory: Phase 9 implementation and Phase 11 verification). This attack surface is mitigated by mandatory boundary markers like the 'Alignment brief' (Phase 2) and 'Workspace scope checkpoint' (Phase 7), as well as rigorous sanitization through the 'Side-effect audit' (Phase 12) and 'Doubt pass' (Phase 13) which audit for behavior outside the requested slice including secret leaks and authorization bypasses (Sanitization: Phase 12 and 13).
[COMMAND_EXECUTION]: The workflow utilizes repository-local scripts, such as scripts/workflow.py, and standard build/test tools to verify code changes. These executions are scoped within the development cycle phases to ensure evidence-based verification of implementations.

development-cycle