ffmpeg performs deterministic media transformation after cognitive stages have proposed grounded source moments. It is application tooling, not a model-accessible effect tool.
How Harmonia uses it
- Cut authorized source media to validated Nimi time ranges.
- Reframe output for vertical social formats.
- Burn captions when the installed build supports the subtitles filter.
- Concatenate selected moments into reels.
- Produce internal clip assets before any external publishing proposal.
- Render deterministic local PNG fixtures in explicit mock development mode.
Commands are constructed as argument arrays without a shell. Inputs, output paths, durations, and supported filters are checked before execution.
Security and authority
ffmpeg reads only authorized local or staged media and writes internal artifacts. Rendering does not approve publishing, choose copy, or prove that an external post exists.
Failure behavior
Missing binaries, invalid media, non-zero exits, absent outputs, and unsupported caption filters remain explicit. Captions may be skipped with a recorded note when the local build lacks the filter; the system does not claim captioned output in that case.
Evidence status
Real local ffmpeg invocations and artifact checks have automated coverage. External publication of a rendered asset requires a separate approved effect and verification receipt.
Relevant files
agent/harmonia_agent/clipper.pyagent/harmonia_agent/generative_media.pyagent/harmonia_agent/stages.pyagent/tests/test_generative_media.py
Last modified on August 28, 2026
