The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill performs a version check for the @intentsolutions/audit-harness package by querying https://registry.npmjs.org. This is a standard operation for managing package dependencies and verifying updates from a well-known service.
[COMMAND_EXECUTION]: The skill uses Bash to execute various local build and testing tools (e.g., pytest, pnpm, go, cargo) and invokes its own bundled scripts like crap-score.py, arch-check.sh, and harness-hash.sh to analyze the repository.
[DYNAMIC_EXECUTION]: Within the scripts/crap-score.py utility, the skill uses subprocess.run to call external complexity and coverage binaries (like radon, gocyclo, or c8) depending on the programming language detected in the repository.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the repository being audited, including Gherkin feature files, product documentation, and commit history, creating a potential surface for indirect prompt injection.
Ingestion points: Data is read from the repository via Read, Glob, and Grep tools, specifically targeting files like features/.feature, docs/.md, and git commit logs.
Boundary markers: The skill documentation defines a clear ownership model (Human vs AI) and policy sections in TESTING.md that are hash-pinned to detect unauthorized modifications, although specific technical delimiters for data passed to sub-agents are not explicitly provided.
Capability inventory: The skill has broad execution permissions via Bash (invoking multiple language runtimes), the Task tool for agent orchestration, and the Skill tool for workflow handoffs.
Sanitization: There is no evidence of explicit sanitization or escaping of the ingested repository content before it is processed by the diagnostic sub-agents.

audit-tests