The Agent Skills Directory

[DATA_EXFILTRATION]: The skill transmits the browser's accessibility state (AX tree), containing the content and structure of the visited webpage, to external API providers to facilitate navigation decisions.
Evidence: The decide function in scripts/bridge.mjs makes POST requests to https://api.typesafe.ai/v1/systemone and https://openrouter.ai/api/alpha/decisions containing browser state data.- [CREDENTIALS_UNSAFE]: The skill accesses API keys stored in environment files on the local filesystem.
Evidence: In scripts/bridge.mjs, the loadConfig function retrieves the path to an envFile from the user's configuration directory (~/.config/jev-codex-computer-use/), and the decide function subsequently reads this file to extract keys for the Authorization header.- [INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection via the accessibility tree data it ingests from external webpages.
Ingestion points: scripts/bridge.mjs retrieves untrusted page content via tab.getAXState().
Boundary markers: The content is packaged into a JSON request body sent to a decision model.
Capability inventory: The skill can execute actions like click, scroll, and pressKey in the browser based on the model's response.
Sanitization: The bridge provides instructions to the decision model to treat page content as untrusted data.- [DYNAMIC_EXECUTION]: The skill encourages dynamic loading of its bridge script during runtime initialization.
Evidence: The integration examples in references/runtime.md suggest using await import(bridgeUrl) to load the bridge.mjs module.

jev-codex-computer-use