Chakra Vault

Never redownload a 600GB model because your cache broke.

Chakra Vault is a public-safe toolkit for verifying and safely restoring local LLM model files. The current implementation focuses on local verification, provider-neutral planning, safe writes, and a thin Hugging Face model download workflow exposed through a CLI.

Current Status

Chakra Vault currently includes:

• public-safe package foundation and safety checks
• local SHA-256 verification primitives
• provider-neutral download planning
• safe download executor with .part files, verification, and atomic promotion
• read-only Hugging Face metadata and file-source bridge
• composed model download workflow
• thin chakra-vault model download CLI entrypoint

It does not yet include:

• API server
• TUI or web dashboard
• SQLite ledger
• background job scheduler
• restore-test automation
• monitoring daemon
• Hugging Face upload, push, publish, create-repo, or remote-delete flows
• bundled model weights or private inventories

Quick Start

Install the package for local development:

python3 -m venv .venv
.venv/bin/python -m pip install -e ".[dev]"

Run the test suite and safety checks:

python3 -m pytest
.venv/bin/python -m ruff check scripts tests src
python3 scripts/check_no_large_files.py .
python3 scripts/check_no_hf_upload.py .
python3 scripts/check_no_private_paths.py .

Run the model download CLI:

chakra-vault model download \
  --repo-id Qwen/Qwen2.5-0.5B-Instruct \
  --target-dir /tmp/chakra-vault-model \
  --revision main

See CLI documentation for arguments, summary fields, exit codes, and privacy behavior. See download workflow documentation for the component boundaries behind the command.

Documentation

• Install
• CLI
• Download workflow
• Roadmap
• Release checklist
• Changelog

Safety Rules

• No Hugging Face upload, push, publish, create-repo, or remote-delete flows.
• No model weights, local logs, database files, tokens, or personal paths in Git.
• No generated model inventories or private machine paths.
• Unit tests use temporary directories and fake sources instead of real model downloads.