subkit0.1.0

subkit turns video and other media files into subtitle and transcript artifacts from a command-line interface, giving creators and developers a repeatable way to produce timed text. Written in Go, it runs as a standalone CLI alongside ffmpeg and treats transcription as a sequence of pipeline stages rather than one opaque operation.

Each stage caches its result independently, so repeating a command with the same input and options can reuse existing artifacts, while a missing final file can be rebuilt from the rendered-output cache. Transcripts, raw provider responses, cues, and rendered files are cached by default; audio remains temporary unless --cache-audio is used, which stores normalized audio for later reuse. Cache keys include the source-media hash, step options, pipeline version, and ffmpeg version for audio extraction, with the default cache location derived from Go’s os.UserCacheDir().

Batch commands accept individual paths or globs and process up to four files concurrently by default. --output-template supports {dir}, {base}, {input_ext}, {kind}, and {format}, while --output-dir handles batch destinations; --out remains an exact single-input path for extract-audio, transcribe, and cues. Multi-file progress uses an interactive Bubble Tea display in a terminal and plain logs when redirected, with auto, tui, plain, or off modes available; failures are summarized after the batch unless --fail-fast stops queued work after the first error.

The prototype uses Deepgram’s nova-2-video model by default, but Portuguese pt-PT input currently requires --model general --language pt-PT because nova-2-video was rejected for that language. Its provider-neutral transcript model preserves structures that map to Deepgram, AssemblyAI, Google, AWS, and Whisper-style timings and segments.