The Agent Skills Directory

[DATA_EXFILTRATION]: The skill employs tools designed to capture sensitive authentication and session data from a user's browser.
Evidence: The documentation for liberate_map_apis and liberate_probe explicitly instructs the agent to capture 'Auth header patterns (X-*, Authorization, cookies)' and inspect 'window globals, localStorage, cookies' via a remote debugging port.
Evidence: The process requires the user to launch a browser with --remote-debugging-port=9222 and log in to the target platform, exposing the active session to the agent's tools.
[COMMAND_EXECUTION]: The skill requires broad tool access to modify the local environment and execute scripts.
Evidence: The allowed-tools list includes Bash, Read, Write, and Edit, which are used throughout the guide to scaffold new directories in src/adapters/ and modify core registration files like src/mcp-server.ts and src/ui/discover.tsx.
Evidence: Phase 4c instructs the agent to execute code via npx tsx src/cli.ts for manual verification of the new adapter.
[INDIRECT_PROMPT_INJECTION]: The skill involves the automated processing of untrusted external content, creating a potential surface for indirect injection.
Ingestion points: The discover and extract functions fetch HTML and JSON data from arbitrary live URLs provided by the user.
Boundary markers: No explicit boundary markers or instruction-guarding delimiters are specified for the code being generated.
Capability inventory: The generated adapters have access to the extraction pipeline's tools, including browser launching via Playwright (launchBrowser).
Sanitization: While the guide mentions sanitizing PII in test fixtures, it does not specify sanitization or validation routines for data processed during live extraction loops.

adapt