The Agent Skills Directory

[DATA_EXFILTRATION]: The skill transmits project-specific content, including ASR (speech-to-text) transcripts, visual analysis descriptions, and researched background context, to external endpoints at https://api.xiaomimimo.com/v1 and its subdomains. This behavior is primarily implemented in scripts/lib.py through the api_call function. While intended for an LLM-based review process, this involves sending user-provided data to a non-whitelisted third-party domain.
[COMMAND_EXECUTION]: The skill performs dynamic code execution using the importlib.util module. Specifically, scripts/agent_brief.py, scripts/agent_text.py, and scripts/narration_lint.py use spec_from_file_location and loader.exec_module to load and execute the local script scripts/deslop_qc.py at runtime. Dynamic execution of code, even from local sources, is a pattern that requires monitoring.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface (Category 8). It ingests data from external files such as vlm_analysis.json and asr_result.json and interpolates this content into prompts for an LLM reviewer in scripts/review_response.py.
Ingestion points: Files vlm_analysis.json and asr_result.json loaded in scripts/review_runner.py.
Boundary markers: Uses Markdown headers like ## Timeline evidence and ## Context-only evidence to delimit external data.
Capability inventory: Network access via api_call and file writing within the work_dir.
Sanitization: Implements regex-based redaction of API keys and data URLs from error logs in scripts/lib.py.

video-script