The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the local project environment and external tools, creating a potential surface for instructions embedded in data to influence agent behavior.
Ingestion points: The skill ingests outputs from project-specific commands (such as npm test, lint, and build) and reads external report files (e.g., roadtest, harden).
Boundary markers: The instructions specify that outputs should be quoted ("outputs quoted, never recalled"), providing a basic boundary, though it lacks explicit instructions to ignore potentially malicious embedded content.
Capability inventory: The skill is designed to execute shell commands for testing, building, and auditing, and it issues release verdicts based on the processed evidence.
Sanitization: Beyond quoting output, there are no specific mechanisms described for sanitizing or filtering instructions from external tool outputs.
[COMMAND_EXECUTION]: The skill executes various development and auditing commands as part of its primary functionality.
Evidence: The workflow requires running the project's own commands for testing, linting, type-checking, and security auditing (e.g., npm test, git commands, and audit tools). These actions are central to the skill's purpose as a release gate.

cleared