The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts fix_hallucinations.py and enrich_uncertain_with_vision.py invoke ffmpeg and whisper-cli via subprocess.run to perform media processing. The implementation utilizes list-based arguments, which prevents shell injection by separating command parameters from the shell environment. The whisper-cli binary path is constructed using the user's home directory (~/whisper.cpp/build/bin/whisper-cli), assuming the existence of the binary in that specific local folder.
[DYNAMIC_EXECUTION]: The test suite scripts/test_fix.py uses dynamic module loading to import the core logic for validation. The script uses importlib.util to load and execute the local fix_hallucinations.py file. This pattern is typical for unit testing and does not involve remote code execution or untrusted data.
[INDIRECT_PROMPT_INJECTION]: The skill processes transcription JSON data and video frames to generate markdown inspection reports, creating an attack surface for indirect prompt injection. Ingestion points: Reads Whisper JSON transcripts and video files (e.g., scripts/fix_hallucinations.py). Boundary markers: The generated README.md report lacks explicit delimiters or instructions for subsequent LLM processing to ignore content embedded in the transcript text. Capability inventory: The skill has the ability to write files and execute system commands (subprocess.run). Sanitization: The skill performs minimal sanitization, specifically escaping markdown pipe characters (|) within transcript content.
[PROMPT_INJECTION]: The skill's documentation instructs the agent to use tool flags that bypass security permissions. Found in SKILL.md in the Visual Context Frame Extractor workflow example: agy --dangerously-skip-permissions. This instruction encourages the agent to operate without permission prompts, potentially reducing user oversight during tool execution.

whisper-corruption-recovery