The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted project data (source code, test files, and configuration files such as Cargo.toml and pyproject.toml) to perform quality and coverage evaluations. This establishes a surface where malicious instructions could potentially be embedded in code comments or test strings to influence agent behavior.
Ingestion points: Project directories and test suites scanned via find and git commands as defined in SKILL.md and modules/framework-detection.md.
Boundary markers: The workflow includes a citation verification step using citation_verifier.py to ensure findings are grounded in verbatim anchors from the source material, providing a control against hallucinations or prompt-driven deviation.
Capability inventory: The skill utilizes shell-based discovery and execution of testing frameworks including pytest, cargo, go test, and npm.
Sanitization: The skill requires that every reported finding carries a verbatim Anchor confirmed by the citation verifier, or it must be dropped or labeled as UNVERIFIED.
[COMMAND_EXECUTION]: The skill relies on shell commands to inspect the local filesystem and run diagnostic tools.
Evidence: The instructions contain commands for identifying project types (find . -maxdepth 2 ...) and analyzing changes (git diff --name-only). It also recommends the use of language-specific test runners and coverage tools such as pytest-cov, cargo-tarpaulin, and vitest. These operations are consistent with the skill's stated purpose of auditing test suites.

test-review