The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the child_process module in open-browser.mjs and cdp-sandbox.mjs to launch Chrome instances and manage the sandboxed execution environment.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: sourcemap-resolver.mjs and api-replay.mjs perform network requests to external domains to fetch source maps or replay API calls, which are necessary for debugging live web applications. The skill also depends on a sibling skill, octocode-scraping, for certain data processing tasks.
[DYNAMIC_EXECUTION]: The skill utilizes dynamic import() calls in cdp-runner.mjs to load task scripts and frequently injects JavaScript into the target browser's execution context via the Chrome DevTools Protocol (CDP).
[INDIRECT_PROMPT_INJECTION]: As the skill processes content from live, untrusted websites, it creates a surface for indirect prompt injection attacks.
Ingestion points: actionability-diagnostics.mjs, affiliates-stealth-probe.mjs, dom-operations-check.mjs, and page-snapshot.mjs extract content and structural data from web pages.
Boundary markers: Absent; the agent receives structured JSON or text representations of the extracted content.
Capability inventory: The skill can perform clicks, fill forms, and navigate the browser (dom-operations-check.mjs, open-browser.mjs).
Sanitization: The skill provides automated redaction of sensitive patterns in HAR files (har-redact.mjs) and limits the length of extracted text strings.
[DATA_EXPOSURE_AND_EXFILTRATION]: Tools such as cookie-bridge.mjs and har-redact.mjs are designed to handle sensitive session data like cookies and authorization tokens; these tools include built-in redaction logic and require explicit user flags (e.g., --i-understand-secrets) to operate.

octocode-chrome-devtools