The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The scripts/setup.sh script downloads the official Frevana setup utility from the author's repository at github.com/FinpeakInc/frevana-cli-releases. As this is a vendor-owned resource used for necessary environment configuration, it is considered a legitimate functional requirement.
[REMOTE_CODE_EXECUTION]: The setup script executes the downloaded remote content via exec bash. This pattern is used to bootstrap the local Frevana daemon and CLI binary from the vendor's official distribution channel.
[COMMAND_EXECUTION]: The scripts/automate.sh script executes the local frevana binary and uses an inline Python script to process JSON payloads for browser automation. The skill interacts with the local file system using mktemp for temporary data handling during tool execution.
[PROMPT_INJECTION]: As a browser automation skill, it possesses an attack surface for indirect prompt injection when processing snapshots of untrusted web content.
Ingestion points: Web content is ingested into the agent context via the frevana_automate tool called in scripts/automate.sh.
Boundary markers: The skill uses an integer-based [ref] system to decouple agent actions from raw HTML/CSS and includes explicit instructions in SKILL.md to stop for human confirmation during sensitive flows.
Capability inventory: The skill can perform click, type, navigate, select, and pressKey operations via the frevana binary.
Sanitization: The skill relies on behavioral constraints (e.g., the --confirm-payment flag requirement) and vendor-level tool safety rather than programmatic input sanitization.

browser-automate