The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill performs runtime monkey-patching of the funasr library's internal export function.
Evidence: In scripts/server.py, the function patch_funasr_onnx_export assigns a local function compat_onnx_export to export_utils._onnx to override the library's default behavior for ONNX exports. This dynamic modification of third-party code at runtime is a notable security risk.
[COMMAND_EXECUTION]: The skill extensively uses shell commands for environment detection, service management, and setup.
Evidence: scripts/init_env.py and scripts/check_env.py execute system commands such as python3, curl, ls, ps, and grep to verify the host environment. scripts/init_env.py specifically runs the user's login shell (shell -l -c "echo $PATH") to capture the full environment PATH.
Evidence: scripts/auto_transcribe.py uses subprocess.Popen to launch the local transcription server.
[EXTERNAL_DOWNLOADS]: The skill downloads large model weights and installs third-party dependencies from external registries.
Evidence: scripts/setup.py and scripts/server.py utilize the modelscope library (snapshot_download) to fetch ASR, VAD, and punctuation models from Alibaba's ModelScope platform.
Evidence: scripts/setup.py executes pip install to install dependencies listed in assets/requirements.txt, which include torch, funasr, and fastapi.
[INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface where instructions spoken within the audio or video files could influence the AI agent's subsequent actions during the summary phase.
Ingestion points: Audio and video files are processed via the file_path parameter in scripts/server.py and scripts/auto_transcribe.py.
Boundary markers: The prompt template in scripts/summary.py (create_summary_prompt) includes the transcribed text after a simple header ("以下是完整文本："), providing minimal isolation between the potentially adversarial data and the system instructions.
Capability inventory: The skill possesses capabilities for file system operations (archiving) and process execution (launching servers).
Sanitization: There is no evidence of sanitization or filtering applied to the transcribed text before it is interpolated into the agent's summary prompt.

funasr-transcribe