The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill utilizes a codebase scanner (scan-codebase.mjs) that ingests content from a target repository, which could potentially contain malicious instructions or deceptive code patterns.
Ingestion points: The scanner reads all source files within the user-provided $TARGET_REPO directory.
Boundary markers: Instructions in SKILL.md advise the agent to treat scanner results as 'leads, not conclusions,' providing a logical boundary for the agent's interpretation.
Capability inventory: The skill possesses significant capabilities including Bash for script execution and Write/Edit for modifying the target repository.
Sanitization: The scanner script provides sanitization by explicitly ignoring files related to credentials, secrets, and environment variables (.env). It also returns metadata rather than raw file snippets to minimize direct processing of untrusted content.
[COMMAND_EXECUTION]: The skill instructions direct the agent to execute multiple Node.js scripts and package management commands.
Evidence: The workflow involves running node scripts/scan-codebase.mjs, pnpm install, and node scripts/validate-stagehand.mjs to perform its primary functions.
[EXTERNAL_DOWNLOADS]: The skill interacts with external registries and vendor-owned web resources.
Evidence: pnpm install downloads dependencies from the NPM registry. The test:e2e:owned script in package.json references https://browserbase.github.io/stagehand-eval-sites/sites/webmcp-test/, which is a vendor-owned test environment. Although automated scanners may flag this test site due to mock forms, it is a legitimate resource within the vendor's ecosystem.
[DYNAMIC_EXECUTION]: The validation component employs dynamic code injection within a controlled browser session.
Evidence: scripts/validate-stagehand.mjs uses the page.addInitScript() functionality to execute local JavaScript files (like webmcp.init.js) inside the browser context to verify tool registration and invocation.

add-webmcp