The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied audio and video files which may contain adversarial spoken instructions designed to influence the subsequent AI summarization phase.
Ingestion points: Media files are read by transcribe.py, server.py, and moss_mlx.py for transcription.
Boundary markers: The skill uses explicit delimiters  and  to wrap generated summaries.
Capability inventory: The skill uses subprocess to call ffmpeg and python3, and performs file writes to .md and .json files.
Sanitization: Summarization prompts in summary.py include instructions to the AI to remain objective and not fabricate information, providing some defense against obedience to audio-based instructions.
[COMMAND_EXECUTION]: The skill frequently uses subprocesses for operational tasks.
scripts/moss_mlx.py and scripts/align_words.py invoke ffmpeg for audio normalization and slicing.
scripts/auto_transcribe.py uses subprocess.Popen to automatically start the local FastAPI server (scripts/server.py) if it is not already running.
scripts/setup.py and scripts/init_env.py execute python3 and curl for environment configuration and dependency management.
[EXTERNAL_DOWNLOADS]: The skill downloads AI model weights at runtime.
scripts/setup.py and scripts/moss_mlx.py fetch models from ModelScope (modelscope.hub) and Hugging Face. These are well-known, established platforms for AI model distribution.
[DYNAMIC_EXECUTION]: In scripts/server.py, the skill implements an ONNX export patching mechanism that modifies funasr.utils.export_utils._onnx at runtime to ensure compatibility with specific environment configurations. This dynamic modification of imported library logic is used for functional compatibility rather than malicious intent.

local-asr