The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently executes shell commands via subprocess.run to call external tools like yt-dlp and ffmpeg. These calls include parameters derived from user-provided video URLs, which could lead to command injection if the input is not strictly validated.
Evidence: subprocess.run(['yt-dlp', '-f', ...], ...) in scripts/anibon/ytdlp.py and scripts/_vision.py.
[DYNAMIC_EXECUTION]: The skill uses dynamic module loading via __import__ which can be used to bypass static analysis tools or load modules from computed paths.
Evidence: file=__import__("sys").stderr in scripts/anibon/markdown.py.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes YouTube transcripts (raw_transcript.json), which are untrusted external sources. The instructions include specific logic for sensitive content detection (e.g., Thai royal/political masking), which an attacker could target with malicious transcript content to influence agent behavior.
Ingestion Points: ytdlp.load_transcript(workspace) in scripts/prepare_video.py.
Boundary Markers: None detected in the chunk processing loop.
Capability Inventory: Shell execution, file writing, and SQLite database interaction.
Sanitization: Basic regex-based cleaning is performed in cleaner.py, but the LLM receives the raw text for summarization.
[EXTERNAL_DOWNLOADS]: The skill downloads game-related data from external APIs to build local reference databases. While these are specific to the skill's functionality, they represent a dependency on third-party infrastructure.
Evidence: Downloads from api.atlasacademy.io in scripts/fetch_fgo_db.py and db.ygoprodeck.com in scripts/fetch_ygo_db.py.
[AUTONOMY_CONCERNS]: The handoff protocol instructs the agent to automatically clear its context and "call itself" to resume work without waiting for user intervention, which reduces human oversight of the process.
Evidence: "Automated Handoff Action" section in references/timestamper-handoff.md.

anibon-timestamper-local