The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external media binaries to process video files using the subprocess module. Detailed analysis confirms these calls are used legitimately for the skill's primary purpose. \n
subprocess.run and subprocess.check_output are utilized in scripts/_common.py and scripts/_mediaskills_common.py to invoke ffprobe and ffmpeg for media inspection. \n
scripts/scan_caption_band.py uses subprocess.run to call ffmpeg for frame extraction. \n
scripts/self_test.py and tests/test_forced_narrative_exact.py invoke utility scripts via uv run to verify the build and validation processes. \n- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from video files and OCR output, creating a potential attack surface for indirect prompt injection. \n
Ingestion points: Untrusted data enters the agent context via OCR text extracted from video frames in scripts/scan_caption_band.py and scripts/refine_boundaries.py. \n
Boundary markers: The references/report-contract.md establishes a strict transcription contract, instructing the agent to treat extracted text as literal dialogue and defining clear inclusion/exclusion rules, which acts as a logical boundary. \n
Capability inventory: The skill scripts have the ability to execute shell commands via subprocess for media processing and directory management. \n
Sanitization: Extracted OCR text is passed through the clean_ocr function in scripts/scan_caption_band.py to remove non-printable characters and normalize whitespace before it is presented to the agent.

forced-narrative-exact