---

name: setup description: "Install, verify, configure, and launch the local speech-to-text server." metadata: disco-role: operating disable-model-invocation: true license: GPL 3.0

Setup

Use this sub-skill when the user needs to prepare a local checkout for first launch or recovery: create the environment, install runtime dependencies, verify ffmpeg/ffprobe, place or download models, choose CPU vs CUDA, interpret set.ini, and start the server safely.

Use for

• fresh installs and environment repair
• dependency installation and import failures
• ffmpeg/ffprobe availability checks
• model directory placement and first-run downloads
• CPU vs CUDA startup selection
• launch wrappers and startup troubleshooting

Do not use for

• browser upload flow
• /api or /v1/audio/transcriptions request details
• OpenAI-compatible client examples
• post-launch transcription behavior

For those topics, use transcription. For shared launch failures and broader runtime issues, use the root troubleshooting guide.

What this sub-skill owns

• Python version selection and virtual environment setup
• runtime dependency installation
• ffmpeg / ffprobe checks
• models/ layout and model download expectations
• set.ini startup keys and device selection
• safe launch helpers and startup diagnostics

Bundled files

• references/install-and-launch.md
• references/configuration.md
• references/troubleshooting.md
• ../../scripts/check-runtime.py
• scripts/check-cuda.py
• scripts/launch-server.py

Working order

1. Read references/install-and-launch.md for the normal install and launch path.
2. Read references/configuration.md when the user asks about set.ini, model selection, or CPU/CUDA toggles.
3. Run ../../scripts/check-runtime.py --repo-root <checkout> before launch.
4. Run scripts/check-cuda.py only when CUDA is expected or needs diagnosis.
5. Use scripts/launch-server.py --repo-root <checkout> to start the app from the current checkout.
6. If startup fails, use references/troubleshooting.md, then hand off to transcription only after the server is reachable.

Preserved runtime facts

• start.py launches a Flask + gevent server, starts a background update check and transcription worker, and opens the browser automatically.
• The GUI worker path uses WhisperModel(device=sets.get('devtype'), download_root=<project>/models) and does not pass temperature or cuda_com_type into WhisperModel.
• The API path applies temperature, while devtype still selects CPU vs CUDA.
• set.ini exposes cuda_com_type, but the current observed runtime path does not consume it during model creation.