The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to execute system commands for its core functionality. In script/build_memory/build_memory.py, it executes pip install to ensure that required dependencies like numpy and openai are available. In script/build_memory/clipping.py and script/build_memory/omni_core.py, it invokes ffmpeg and ffprobe to slice video into segments and probe file durations.
[EXTERNAL_DOWNLOADS]: The build script script/build_memory/build_memory.py checks for missing dependencies and automatically installs them from the public PyPI registry. While the targeted packages (numpy, openai) are well-known, automatic installation at runtime is an unconventional practice for dependency management.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted media content (video and audio) and uses LLMs to generate transcripts and semantic facts. These results are then interpolated into subsequent LLM prompts used for memory induction. An attacker could embed instructions in the video's audio that might influence the agent's reasoning.
Ingestion points: script/build_memory/stages.py extracts text from media which is then used in script/build_memory/prompts.py.
Boundary markers: Present in the form of specific tags like 【MEMORY_BATCH】 and {{DIALOGUE}} in the prompts, though they do not explicitly warn the model to ignore instructions within the data.
Capability inventory: Includes command execution (ffmpeg, pip), file writing (local memory storage), and network operations (DashScope API).
Sanitization: Absent; the skill relies on the underlying LLM's safety guardrails to ignore malicious instructions in the transcript data.

qwen-mm-plugins-omni-memory