Intent-Code Divergence
Medium
- Confidence
- 95% confidence
- Finding
- The skill claims it only produces a strategy, but its workflow instructs the agent to run `ffprobe` on user-provided files. This is a real capability expansion from planning into local command execution, which can expose filesystem metadata, create unsafe trust boundaries, and cause downstream agents to treat this skill as safe-to-run when it is not purely declarative.
