SMTHN.GD Lockd'In Workspace

M0 status: baseline hardening in progress.

This repository currently contains both migration-source code and new canonical product paths.

• Migration sources: jarvis.py, lockdin_mvp/, DeepSeek-V4-Pro/, csm/
• Canonical destinations for new code: apps/web/, apps/desktop/, packages/, docs/

M0 operational commands:

# Python and local service checks
python scripts/dev_doctor.py

# Soft mode for CI environments where local services are intentionally absent
python scripts/dev_doctor.py --allow-warn

# Milestone critique score (example)
python scripts/critique_milestone.py --milestone M0 --scores-json '{"Repository Hygiene":4,"Secret Safety":5,"Developer Experience":4,"Build Reproducibility":4,"Architecture Governance":4,"Migration Clarity":4}'

See docs/M0_EXECUTION.md for ongoing M0 execution and docs/critique/ for milestone critique skills.

Legacy reference: Jarvis (DeepSeek + Sesame CSM) integration

Overview

• This repo integrates DeepSeek-V4-Pro (text generation) with Sesame CSM (voice generation) into a simple assistant called Jarvis.

Quick start (recommended: conda/miniforge on macOS)

1. Install Miniforge and create env

# Install Miniforge (if not installed) then open a new terminal
conda create -n jarvis python=3.11 -y
conda activate jarvis

2. Install PyTorch + torchaudio via conda

conda install -c pytorch -c conda-forge pytorch torchaudio -y

3. Install the rest of the requirements

pip install --upgrade pip setuptools wheel
# remove or skip torch in requirements if conda installed it
pip install -r requirements.txt --no-deps
pip install openai-whisper sounddevice soundfile

4. DeepSeek V4 asset status

./DeepSeek-V4-Pro currently contains the V4 tokenizer, encoding, configuration, and inference source, but not model weights. DeepSeek V4 Pro is a 1.6T-parameter model whose reference runtime is CUDA/model-parallel and requires Torch 2.10+ and Transformers 5+ in a dedicated environment.

5. (Optional) Set Hugging Face token for downloads

export HUGGINGFACE_HUB_TOKEN="hf_xxx"

Run Jarvis (dry-run, fast)

# dry-run skips heavy models and simulates behavior
python jarvis.py --dry-run

Legacy local model command (not currently runnable)

python jarvis.py --ds-ckpt ./DeepSeek-V4-Pro --ds-config ./DeepSeek-V4-Pro/inference/config.json

Do not use this command until compatible checkpoint shards and a dedicated provider runtime are configured. The legacy Jarvis loader is a migration source; production integration belongs behind the M4 model adapter.

Fetching large assets for team

• To avoid committing large binary files to git, use the provided script scripts/fetch_assets.sh.
• Example: scripts/fetch_assets.sh assets will download the Miniforge installer into assets/.
• You can set DEEPSEEK_URL to a tar.gz that contains the DeepSeek-V4-Pro checkpoint tree, and SESAME_PROMPTS_URL to prompts archive.

Example usage (download Miniforge and DeepSeek archive):

# export DEEPSEEK_URL="https://example.com/deepseek.tar.gz"
DEEPSEEK_URL="<url-to-deepseek-tar.gz>" scripts/fetch_assets.sh assets

Team workflow

• Clone repository with submodules:
  • git clone --recurse-submodules <repo>
• Run scripts/fetch_assets.sh (or set the environment variables and run) to fetch large installers and model checkpoints into place.
• This keeps the main git history small and lets team members obtain large assets from an external host.

Flags

• --dry-run: simulate responses without loading model weights (useful for development)
• --enable-wake: enable microphone wake-word listening (requires sounddevice and openai-whisper)
• --wake-word: set the wake word (default jarvis)
• --device: auto|cpu|mps|cuda (auto-detects mps on macOS)
• --once: print greeting and exit

Notes

• On macOS, generated audio is played automatically with afplay (non-dry runs).
• DeepSeek V4 Pro is not a CPU or single-consumer-GPU runtime target. Use a dedicated multi-GPU provider process or a hosted adapter.
• The csm/ checkout is currently unavailable, so Sesame voice runs only in fallback/mock mode.

If you want, I can help automate environment setup or run a full end-to-end test if you provide model checkpoints and confirm you want me to proceed.

S3 upload/download helpers

• This repo now includes scripts/s3_assets.py and shell wrappers scripts/upload_to_s3.sh and scripts/download_from_s3.sh.
• Usage examples:
  • Upload a file or folder:

    scripts/upload_to_s3.sh my-bucket assets/Miniforge3-MacOSX-x86_64.sh models
    # or recursive folder upload:
    python3 scripts/s3_assets.py upload --bucket my-bucket --key-prefix models --recursive DeepSeek-V4-Pro

  • Download a single key:

    scripts/download_from_s3.sh my-bucket models/Miniforge3-MacOSX-x86_64.sh assets

  • Download recursively by prefix:

    scripts/download_from_s3.sh my-bucket models assets prefix

Security: prefer using IAM roles (CI) or environment variables from .env (see .env.example) — never commit real credentials.