The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill makes extensive use of runpy.run_path() to execute logic from files located outside the skill package. Specifically, scripts/headless_draft.py loads multiple engine components from an external project root, and scripts/asr_once.py loads an ASR executor (defaulting to ~/scripts/transcribe.py) to retrieve sensitive ACCESS_TOKEN and APP_ID values. While headless_draft.py verifies these files against a PINS dictionary of SHA-256 hashes, the asr_once.py executor path is controlled by environment variables and lacks integrity verification.
[COMMAND_EXECUTION]: The skill invokes several external processes via the subprocess module. scripts/edit_plan.py executes ffmpeg for audio rendering and ffprobe for media metadata extraction. scripts/asr_once.py executes a Python script using sys.executable to perform transcription tasks. While these calls generally avoid shell invocation, they operate on paths derived from external input.
[EXTERNAL_DOWNLOADS]: The skill requires a manual installation of the mcncarl/jianying-headless repository from GitHub. Although this is a vendor-owned resource, the skill relies on this external codebase for its core functionality rather than including the necessary logic locally.
[INDIRECT_PROMPT_INJECTION]: The skill processes complex JSON edit plans and interprets media file metadata. There is an attack surface for indirect injection as the skill lacks explicit boundary markers or sanitization when interpolating user-provided or media-derived text (e.g., subtitles, segment reasons) into the plans used by the editing engine. Mandatory evidence: 1. Ingestion points: scripts/edit_plan.py (JSON plan input), scripts/asr_once.py (media source files). 2. Boundary markers: Absent. 3. Capability inventory: subprocess calls (ffmpeg/python), runpy execution, and significant local file system write access. 4. Sanitization: Limited to numerical range validation for timing and volume.
[DATA_EXPOSURE]: The skill accesses sensitive application data stored in the local Jianying project directory (Movies/JianyingPro/User Data/Projects/com.lveditor.draft). It reads project metadata, draft IDs, and timeline structures, which could potentially contain sensitive user information or project details.
[PERSISTENCE]: The skill provides functionality to 'publish' drafts, which involves registering new project entries in the Jianying application's local database and home screen index. This allows the skill to create persistent artifacts within a third-party application on the host machine.

yichen-jianying-edit