The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses execFileSync to execute various shell commands, including git for repository state checks and the project's own build command (e.g., npm run build) to ensure the site is correctly generated before auditing.
[DYNAMIC_EXECUTION]: The tool identifies the local package manager and dynamically executes the build script defined in the target repository's package.json file. While this is the intended functionality of a build-gate tool, it involves executing code from the subject repository.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and analyzes a wide range of untrusted data from the repository being gated, such as HTML pages, CSS files, and configuration manifests. This content is processed and formatted into a markdown report for the agent, which could potentially contain malicious instructions intended to influence the agent's behavior.
Ingestion points: Repository files (HTML, CSS, JSON, text) and git metadata are read in scripts/lib/context.mjs and scripts/lib/output.mjs.
Boundary markers: Findings are structured into a markdown report; no specialized adversarial-resistant delimiters are used to wrap ingested content.
Capability inventory: The skill can execute shell commands (execFileSync), write local files (writeFileSync), and perform network requests to fetch sitemaps (net.get).
Sanitization: Content is parsed using standard HTML and robots.txt parsers, but raw snippets from scanned files are included as evidence in reports.
[CREDENTIALS_UNSAFE]: The skill includes a 'hygiene' module designed to search for and identify hardcoded secrets, API keys, and tracked environment files (.env) within the target repository. While it mitigates risk by only reporting the first four characters of detected secrets, the process involves accessing and surfacing sensitive information during the audit.

pre-launch