cjm-transcription-tui 0.0.1

Transcription-workflow TUI driver: source/folder selection, candidate (capability, model)-instance setup from manifest config schemas, and per-segment side-by-side transcriber comparison to pick a run's lightweight/accuracy pair — then hands off to cjm-transcription-core's headless pipeline. Born on-graph.

cjm-transcription-tui

Transcription-workflow TUI driver over cjm-transcription-core's headless pipeline.

v0 flow (one keyboard-first Textual app, three stages):

1. Sources — pick audio/video files and/or folders (folders expand to every media file under them, recursive + sorted, via the core's expand_sources).
2. Candidates — stand up candidate (capability, MODEL) instances enumerated from the capability manifests' config schemas (e.g. the whisper family, voxtral mini vs small) through the substrate's CR-10 multi-instance loading.
3. Compare — transcribe ONE VAD-cut segment of a chosen source across every candidate side by side, walk segments to sample more, then mark the run's lightweight and accuracy pair and confirm.

On confirm the app exits and hands the equivalent cjm-transcription-core run invocation to the headless CLI (printed first, for reproducibility), so runs stay journaled + manifest-recorded exactly like hand-launched ones.

cjm-transcription-tui --manifests-dir .cjm/manifests [paths...]

Born on-graph: package content is authored as graph nodes and projected to .py (the journal is the source of truth; these files are generated artifacts).