clip-stream
Pass
Audited by Gen Agent Trust Hub on Nov 21, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill extensively uses the
child_process.spawnmodule to invoke external binaries such asffmpeg,ffprobe,yt-dlp, andstreamlinkfor video analysis, downloading, and cutting. - Evidence: The
lib/process.mjsfile providesrunandrunCapturewrappers used byscripts/cut-clips.mjs,scripts/find-moments.mjs, andscripts/watch-live.mjsto execute tasks on the host system. - [EXTERNAL_DOWNLOADS]: The skill downloads content from external URLs using
yt-dlpandstreamlink. Additionally, it recommends usingnpx hyperframesfor transcription, which triggers a download and execution of code from the npm registry. - Evidence:
SKILL.mdandscripts/watch-live.mjscontain instructions and logic for recording remote video streams. - [INDIRECT_PROMPT_INJECTION]: The skill has an ingestion surface for untrusted user-generated content from Twitch chat and chat logs, which is used to score highlights.
- Ingestion points:
chat.json(read infind-moments.mjs) and live IRC chat fromirc.chat.twitch.tv(read inwatch-live.mjs). - Boundary markers: Not explicitly defined for the chat data, but the content is processed via algorithmic token matching rather than being directly interpolated into agent prompts.
- Capability inventory: The skill can write files, execute subprocesses, and perform network operations.
- Sanitization: The
chatSignalfunction inlib/moments.mjsmitigates risk by tokenizing input and matching against a static whitelist of hype-related tokens (e.g., 'lol', 'pog'), which prevents the interpretation of chat text as instructions. - [DYNAMIC_EXECUTION]: The
watch-live.mjsscript manages its complex workflow by dynamically spawningnodeprocesses to run other scripts provided with the skill. - Evidence:
scripts/watch-live.mjsspawnsnode find-moments.mjsandnode cut-clips.mjsas subprocesses.
Audit Metadata