Skip to content

DEV_TEST_LOOP.md

Latest commit

 

History

History
136 lines (112 loc) · 5.18 KB

DEV_TEST_LOOP.md

File metadata and controls

136 lines (112 loc) · 5.18 KB
Raw

Dev / Test Loop with docker-agent

This doc exists so future sessions (Claude or human) can quickly understand:

  • What’s on this branch
  • How the app runs on docker-agent
  • How to rapidly test local changes

1. Branch and purpose

  • Branch: voice-design-accent-and-queueing
  • Goal: implement Persona Forge / OmniVoice workflows with accent-specific voice design, streaming segments, and queueing while models load.

Key capabilities:

  • Per-sentence OmniVoice segment generation (audition → cherry-pick → stitch → save).
  • Streaming segments into the rack as they complete.
  • Diverse candidates (guidance_scale), real progress/ETA.
  • Async model loading + job queueing (no hard failure while loading).
  • Lazy Base model swap-back (stays unloaded until needed).
  • Persistent segment library and “Save to library” for VoiceDesign.
  • UX: composer-style Script input, Non-Verbals/Examples toolbars, sidebar polish.

2. docker-agent: how the app runs

  • Machine: docker-agent (access via SSH as usual, root@docker-agent, 192.168.10.72).
  • Container: persona-forge
  • Image: persona-forge:voice-design-accent-and-queueing
  • Port: 8318 → 8318 (HTTP)

Runtime mode:

  • Command:
    • scripts/entrypoint.sh gunicorn persona_forge.app:app
      -w 1 -k gthread --threads 4 --timeout 300
      --bind 0.0.0.0:8318 --log-level info
  • Env (important subset):
    • TTS_BACKEND=openvino
    • MODEL_SIZE=1.7B
    • LOW_RAM_MODE=1
    • FRONTEND_ENABLED=1
    • (REF_TEXT set to a fixed reference; confirm if it must change for your use.)

Key mounts:

  • Host: /var/data/autopirate/persona-forge/voices
    • Container: /voices
  • Host: /var/data/autopirate/persona-forge/segments
    • Container: /segments
  • Host: /var/data/autopirate/persona-forge/reference/voice_A.wav
    • Container: /voice/reference.wav
  • Host: /var/data/autopirate/persona-forge/model
    • Container: /root/.cache/huggingface/hub
  • Host: /var/data/autopirate/persona-forge/ov
    • Container: /ov
  • Host: /home/nick/projects/persona-forge/src
    • Container: /app/src

Important:

  • The running container uses /app/src from:
    • /home/nick/projects/persona-forge/src on docker-agent.
  • That directory is git-tracked to the same repo as this project.

3. Dev-test loop (current default)

Use this when developing locally (or in Claude) and testing on docker-agent.

From the local repo (e.g. this CWD):

  1. Make changes:
    • Edit backend/frontend in src/persona_forge and frontend/src as needed.
  2. Commit and push:
    • git add -A
    • git commit -m "descriptive message"
    • git push origin voice-design-accent-and-queueing
  3. On docker-agent:
    • cd /home/nick/projects/persona-forge
    • git pull origin voice-design-accent-and-queueing
  4. Reload backend:
    • The container’s code mount already reflects the new src, but gunicorn has no auto-reload.
    • Choose one:
      • Fast reload: docker exec persona-forge kill -HUP 1
      • Or full restart (safer for bigger changes): docker restart persona-forge
  5. Test:

Notes:

  • If you change Python-only files: HUP or restart is sufficient.
  • If you change Dockerfile, requirements, or system-level deps: rebuild image and recreate container.

4. Rapid dev mode (optional, for heavy iteration)

When iterating quickly (e.g., tuning OmniVoice behavior), swap gunicorn for uvicorn with --reload.

Example (run on docker-agent, or script it):

  • Stop existing container:

    • docker stop persona-forge && docker rm persona-forge

  • Run in dev-reload mode (same mounts/env, uvicorn instead of gunicorn):

    • docker run -d
      --name persona-forge
      --memory=13g
      -p 8318:8318
      -v /var/data/autopirate/persona-forge/voices:/voices
      -v /var/data/autopirate/persona-forge/segments:/segments
      -v /home/nick/projects/persona-forge/src:/app/src
      -v /var/data/autopirate/persona-forge/reference/voice_A.wav:/voice/reference.wav
      -v /var/data/autopirate/persona-forge/model:/root/.cache/huggingface/hub
      -v /var/data/autopirate/persona-forge/ov:/ov
      -e TTS_BACKEND=openvino
      -e MODEL_SIZE=1.7B
      -e LOW_RAM_MODE=1
      -e FRONTEND_ENABLED=1
      -e "REF_TEXT=Welcome to the shop. How can I help you find what you're looking for today?"
      persona-forge:voice-design-accent-and-queueing
      python -m uvicorn persona_forge.app:app
      --host 0.0.0.0 --port 8318 --reload
      --timeout-graceful-shutdown 300

Effect:

  • Any change in /home/nick/projects/persona-forge/src triggers auto-reload inside ~1-2 seconds.
  • Combine with git pull on docker-agent after pushing from local.

When done iterating, revert to normal gunicorn-based container command.

5. How future Claude sessions should use this

If you’re a new Claude session:

  • Assume:
    • We are developing on voice-design-accent-and-queueing.
    • Running instance is docker-agent:persona-forge (8318), using /app/src mounted from git-tracked repo.
  • Default workflow:
    • Implement in this repo.
    • Commit + push.
    • Instruct user to (or assist with):
      • cd /home/nick/projects/persona-forge && git pull
      • docker exec persona-forge kill -HUP 1 (or docker restart persona-forge).
  • Only propose heavier changes (image rebuilds, env changes) when necessary and confirm first.