The Agent Skills Directory

[DYNAMIC_EXECUTION]: Evaluates JavaScript code generated by the agent or user using the AsyncFunction constructor in scripts/cli/exec-context.mjs. This allows for flexible automation but enables arbitrary code execution within the Node.js environment.
[COMMAND_EXECUTION]: Spawns system processes using child_process. This includes ffmpeg for video encoding during UI recording (scripts/engine/recording/capture.mjs) and powershell.exe to interact with native Windows file dialogs (scripts/engine/forms/file-dialog.mjs).
[DATA_EXFILTRATION]: Sends text content (captions and labels) to external providers including OpenAI, ElevenLabs, and Microsoft (Edge TTS) for text-to-speech synthesis. This is a primary feature for creating narrated demonstration videos, but involves sending application data to third-party APIs.
[INDIRECT_PROMPT_INJECTION]: The skill represents an indirect injection surface because it retrieves and processes data from a web application which could contain malicious instructions.
Ingestion points: Reads form field values, table contents, and spreadsheet data via scripts/engine/forms/state.mjs and scripts/engine/table/grid.mjs.
Boundary markers: Data is returned to the agent as structured JSON, but no explicit instructions are provided to the agent to treat this data as untrusted or to ignore embedded commands.
Capability inventory: The skill possesses capabilities for reading/writing local files, performing network requests, and executing shell commands.
Sanitization: No specific sanitization or filtering is applied to the text data retrieved from the browser session.
[EXTERNAL_DOWNLOADS]: Fetches the Playwright library and the Chromium browser engine from official registries during the initial setup process, as documented in SKILL.md.
[CREDENTIALS_UNSAFE]: The documentation in SKILL.md provides examples of passing credentials directly via command-line arguments (e.g., --password=secret), which may expose them in process listings.

web-test