omp packages

omp-bonsai-system-one0.1.0

Local System One judge for omp on frozen Ternary-Bonsai-4B, with an optional model router

senna_lang33 Published 14 days ago Apache-2.0 ~52 downloads/mo

Built for omp

The author tagged this package omp-plugin.

Did it work for you?

Result

One report per person. A new report replaces your old one.

Reports 0

No reports yet.

Readme

omp-bonsai-system-one

A plugin for the omp coding agent that runs omp's judge role on a local System One model: bonsai-4b-system-one, a frozen Ternary-Bonsai-4B served over HTTP. The judge drives omp's per-prompt auto thinking level and its other typed judgments; with this plugin they run offline and at no cost.

An optional router feature also switches the session's model once, from the first prompt's rated difficulty, to your own roles:

Rated difficulty Role
light: a question about the code, or a tiny mechanical edit @smol
standard: an ordinary feature, fix, test, or review within one area @mid, or @default when no mid role is set
complex: subtle design, cross-cutting change, or hard debugging @slow

The plugin assigns no models; a tier whose roles are unset keeps the current model.

Experimental. Not affiliated with TypeSafe, PrismML, or the omp project.

Install

  1. Install and run the local server from bonsai-4b-system-one (Apple Silicon: MLX, about 1.1 GB of weights).

  2. Install the plugin from npm:

    omp plugin install 'omp-bonsai-system-one[router]'     # drop [router] for the judge only

    or from this repository as a marketplace:

    omp plugin marketplace add senna-lang/omp-bonsai-system-one
omp plugin install omp-bonsai-system-one@omp-bonsai-system-one
omp plugin features omp-bonsai-system-one --enable router     # optional

  3. Select the model for the judge role in ~/.omp/agent/config.yml:

    modelRoles:
  judge: bonsai-local/bonsai-4b-system-one
defaultThinkingLevel: auto

Starting the server

Keep python -m system_one_bonsai.serve running yourself, or let the plugin start it on the first prompt of a session by pointing it at your environment:

export BONSAI_SYSTEM_ONE_SERVE_CMD="/path/to/bonsai-4b-system-one/.venv/bin/python -m system_one_bonsai.serve"

The first prompt waits up to 20 s for the server to come up. The started server keeps running after omp exits (about 1.5 GB resident on MLX); its log is ~/.cache/bonsai-system-one/serve.log. BONSAI_SYSTEM_ONE_URL overrides the default address http://127.0.0.1:8765.

If the server is unavailable, the turn continues: auto thinking keeps its previous level and the router keeps the current model.

Accuracy

On 80 coding-agent requests with reference effort levels from a strong model (Claude Opus; not human labels), asked omp's own auto-thinking question through omp's judge path:

Judge Exact effort Within one level Median latency (Apple M2)
bonsai-local/bonsai-4b-system-one 0.66 0.99 0.8 s
omp built-in local/lfm2-1.2b 0.29 0.69 3.0 s
omp built-in local/lfm2.5-230m 0.25 0.56 0.3 s

It tends to rate one level low, mostly high for xhigh requests. The set is small and the reference labels come from a single model.

License

Apache-2.0. See LICENSE.