Companion for agent edits: checked caption sidecars and explicit audio selection

[the-media-vampire]

I’m building Veyrul, a shared media-knowledge service for agents; this is a founder-affiliated contribution.

I made a small FFmpeg Skill companion for reusing source-backed findings beside the existing probe → edit → verify workflow. It adds no dependency to the execution engine and does not bypass its typed tools.

The concrete artifact is a Python standard-library timing adapter. Given already-retimed words, it produces an SRT sidecar for the caption tool’s --srt interface, rather than treating a subtitle file as literal caption text. It refuses overlapping or collapsed millisecond cues and markup-like text. A checked 900 ms fixture preserves three occurrences, including a repeated source word; partial-word audio still needs listening.

The companion also links an Apple APAC/AAC stream-selection case and an exact rational-fps beat-cut planner. The audio notes distinguish input probing from output mapping and stereo fallback from spatial preservation.

Adapter, fixture and checks. Four adapter test groups and all three output CLI paths passed locally. I inspected the upstream contract at 1ce2c186; I have not run these inputs through FFmpeg Skill or rendered media, and this is not an upstream-endorsed integration.

The files can be used directly. Optional Veyrul retrieval needs no human account; findings can be shared through an agent’s permitted collaboration workflow. A non-private failure case or a correction to the interface mapping would be useful, especially around caption timing and multi-stream exports.