The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from source videos (speech, subtitles, OCR) via the omni_call tool to generate dubbing plans and translations.
Ingestion points: omni_call results in workflows/source-analysis/WORKFLOW.md.
Boundary markers: No specific markers are defined to isolate untrusted video content from agent instructions in the analyzed files.
Capability inventory: File system write operations, network requests to the dubbing service, and media rendering tools.
Sanitization: No evidence of sanitization for the transcript or OCR text before it is used in the authoring pipeline.
[COMMAND_EXECUTION]: The reference server implementation (references/launch_dubbing_server.py) uses subprocess.run to execute external binaries like ffmpeg and demucs for audio manipulation and separation. While these are used for legitimate media processing, they represent a vector for command injection if input paths or parameters were improperly handled.
[DATA_EXFILTRATION]: The skill transmits project audio data (in base64 format) to an external service URL defined by the environment variable QWEN_MM_DUBBING_SERVER_URL. While this is necessary for the skill's purpose (using external GPUs for TTS/VAD/Separation), it facilitates the movement of potentially sensitive user media to a remote endpoint.

qwen-mm-plugins-omni-chatcut-video-translation