The Agent Skills Directory

[COMMAND_EXECUTION]: In run.mjs, the runner script constructs a shell command string by interpolating the runName variable (sourced from the --name command-line argument) directly into a string executed via bash -c. This allows for arbitrary command execution on the host system if the name contains shell metacharacters such as backticks or subcommand substitution.
[INDIRECT_PROMPT_INJECTION]: The browsing logic in src/discovery.js and capture.mjs uses an LLM to drive a web browser based on page element snapshots. A malicious webpage can include instructions in visible or hidden text to manipulate the LLM's context, potentially leading to unauthorized browser actions or data exfiltration. While secret redaction is implemented in src/sensitive.js, it does not prevent logic-based instruction overrides.
Ingestion points: src/discovery.js (DOM snapshots via snapshotInteractables).
Boundary markers: Absent in the snapshot data passed to the LLM.
Capability inventory: Full browser control (click, fill, scroll, wait) and video/asset publication via the moda CLI.
Sanitization: Implements heuristic-based secret redaction for form fields in src/sensitive.js but lacks general prompt protection for untrusted web content.
[DYNAMIC_EXECUTION]: The compile.py script and auth.mjs dynamically determine library search paths (sys.path) and execution targets based on the contents of a local .studio file or environment-specific directory scanning. This configuration allows for the loading of arbitrary modules or the execution of binaries from searched or user-defined paths.
[EXTERNAL_DOWNLOADS]: Fetches browser binaries via Playwright and static utility binaries (ffmpeg-static, ffprobe-static) from the official NPM registry during installation and runtime setup.

moda-video-demo