The Agent Skills Directory

[PERSISTENCE]: The skill instructions direct the agent to modify the user's shell profile (~/.zshrc) to persist environment variable configurations for various browser providers.
[COMMAND_EXECUTION]: Utilizes the Bash tool to execute agent-browser CLI commands and helper scripts for navigation, interaction, and state management within browser runtimes.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function of browser automation creates an attack surface where untrusted web content could attempt to influence agent behavior.
Ingestion points: Data enters the context through agent-browser read, snapshot, errors, and console (as noted in references/trust-boundaries.md).
Boundary markers: The instructions in references/trust-boundaries.md explicitly tell the agent to treat browser-derived text as data and ignore any embedded instructions.
Capability inventory: The skill uses Bash to drive the agent-browser CLI, which can perform network requests and manipulate page state.
Sanitization: Instructions specify that the agent should verify outcomes independently and avoid reproducing secrets from the browser environment.
[DATA_EXFILTRATION]: The skill accesses local sensitive configuration files in ~/.config/ and transmits authentication tokens to external services, including provider APIs and the Coolify management console at app.coolify.io for managing the vendor's browser pool.
[DYNAMIC_EXECUTION]: Provides functionality to execute JavaScript within the browser context using agent-browser eval, facilitating complex interactions with web applications through runtime code execution.

run-agent-browser