The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to capture and bisect data from external, untrusted websites, including full DOM snapshots, console logs, and network traces. An agent using this skill to debug or audit a page could be vulnerable to malicious instructions embedded within that page's content.
Ingestion points: DOM dumps captured in scripts/snapshot-loop.mjs and console/network activity processed in scripts/bisect-cdp.mjs.
Capability inventory: The skill allows access to Bash, Read, and Grep, and it can perform complex file operations and network queries via the browse CLI.
Boundary markers: There are no explicit delimiters or instructions provided to the agent to treat the captured data as untrusted or to ignore embedded instructions.
Sanitization: The content is captured and presented in raw or structured form (JSONL) without filtering for adversarial prompt patterns.
[EXTERNAL_DOWNLOADS]: The skill's setup instructions suggest installing the browse CLI via npm install -g browse. This is an external dependency required for the skill to interact with the Browserbase platform or local Chrome instances.
[COMMAND_EXECUTION]: The skill relies on several Node.js scripts that use spawn and execFileSync to manage background processes for the CDP firehose and sampling loop. It also executes various browse CLI commands to capture browser state.

browser-trace