OMNI

OMNI is a neural compressor for source code. It beats tar+xz on ratio for Python codebases by combining a trained sequence model with explicit long-range copy matching and entropy coding, instead of general-purpose byte-level compression.

omni compress my_project/
# -> my_project.satish_andromeda

omni decompress my_project.satish_andromeda
# -> my_project/  (byte-identical to the original)

Install

Real one-line installers (apt install omni, curl ... | sh) aren't live yet. For now:

git clone <this repo>
cd omni
./install.sh

which installs the omni CLI in editable mode via pip.

Usage

omni compress <path> [--model NAME] [--out FILE]     # file or directory
omni decompress <file.satish_*> [--out PATH]
omni info <file.satish_*>                              # header only, no model needed
omni models                                             # installed model generations
omni version

Model management — omni model update fetches the latest generation directly from this repo's GitHub Releases (tagged <generation>-<year>, e.g. andromeda-2026), no separate registry server required:

omni model update                                       # install/refresh the latest generation
omni model update --force                                # re-download even if already installed
omni model register <name> <year> <model.pt> --so <arithmetic_coder.so> [--default]
omni model default <name>

omni model update checks ben-blance/omni by default — override with the OMNI_MODEL_REPO env var (owner/repo) to point at a fork.

omni compress always uses the latest installed generation unless you pass --model. omni decompress always uses whatever generation the archive's header says it needs — see docs/satish-format.md.

Try it

examples/ has a walkthrough against a small sample project.

Public vs. private

This repository is the distribution layer only — the CLI, the SATISH container format, docs, and install scripts. It does not contain the compression engine (tokenizer, model architecture, LZ matcher, arithmetic coder, or trained weights).

              this repo (public)
                     │
          ┌──────────┴──────────┐
          │                     │
    CLI (src/omni/)      SATISH format (documented,
          │                docs/satish-format.md)
          ▼
   src/omni/engine.py  ──seam──▶  OMNI engine (private)
                                        │
                                        ▼
                                 model: Andromeda (2026)

src/omni/engine.py is the only file that talks to the engine, and it does so dynamically (via OMNI_ENGINE_SRC, or a private package once one exists) — nothing else in this package needs to change when the engine moves to its own private repo or ships as a compiled binary.

The SATISH format is public and documented on purpose (see docs/satish-format.md) even though the engine isn't: a .satish_* file's header should always be inspectable, independent of whether you have the model or algorithm that produced it.

License: MIT — see LICENSE. This covers the CLI and SATISH format shell in this repo only; it does not extend to the private compression engine or trained model weights, which are distributed separately (see "Public vs. private" above) and are not covered by this license.