Premium Model Budget Governor View on GitHub

Open source / Public beta 0.4.0rc6

Premium Model
Budget Governor

Keep Astra in the work.
Keep the budget in view.

Preview the workflow. Approve the estimate. Inspect the receipt. A local workbench for deliberate premium-model use.

Explore the workbench Get the beta

Direct Astra by default. Read-only execution. No guaranteed savings.

Inside the workbench

Your task. Your model. Your approval.

Development captures, not a live session.
No model calls or account connection.

Start with Astra. Review the selected evidence and estimated spend before authorizing a read-only task.

Try the planner below

Captures show preview states, not proof of savings. Extra stages can cost more. Reduced skill discovery can omit useful guidance.

LOCAL WORKBENCH01 / Direct preview
Astra-preferred Workbench preview with task inputs and explicit approval, before execution

Example project / Preview only / Scroll capture or open full image

01 / Live demo

What fits your task budget?

Illustrative whole-workflow estimates. No model calls, account connection, or live billing.

Conserve mode
Selected workflow Astra direct
Decision Estimate pending
Why

Compare complete workflows including verification and contingency.

Astra stages planned0
Context actionCompress
Approval gateRequired
Total workflow estimatePending

Includes a 0.5-credit contingency and verification. No cache savings assumed. Quality scores are illustrative, not measured predictions.

02 / How it works

Four moves. One governed loop.

  1. 01

    Measure

    Count host input as well as task context. Missing usage stays unknown.

  2. 02

    Plan

    Compare direct Astra and complete hybrid workflows under one budget.

  3. 03

    Execute

    Reserve a bounded call. Astra can investigate, originate, and solve.

  4. 04

    Reconcile

    Record counters, verify outcomes, then decide whether another call helps.

03 / Origin

One allowance gone.
The next already running low.

During Astra-heavy Codex work, one weekly allowance was consumed in roughly a day. After a reset, the replacement allowance was again down to 16% by the next day, and later reached 11% during this launch.

That experience raised a question about repeated context, oversized evidence, and unnecessary handoffs. The account readings alone could not isolate the cause. The goal became:

How can we use as much of Astra’s capability as possible while keeping workflow burn closer to Sol?

These are the creator’s observed account-capacity readings, not a universal provider benchmark.

Idea, research guidance & product management: Sulabh Dubey

Sulabh Dubey · Originator and product lead
Lived problem, core idea, research guidance, product requirements, priorities, edge cases, approvals, and real-project proof-test direction.
Codex by OpenAI · Research and build collaborator
Research synthesis, architecture, policy engine, CLI, MCP server, Codex plugin, security scanners, evals, documentation, visual design, implementation, testing, and release execution.

Independent open-source project. Not an OpenAI product and not endorsed by OpenAI.

04 / Proof

Choose the workflow, not a fixed winner.

Five task families across four workflows: extra handoffs usually cost more than direct Astra. A smaller skill catalog reduced estimated cost on one repeated visual task, but may omit useful guidance. Codex-graded pilot results do not prove broad quality equivalence or weekly savings.

A newer 12-call Astra context pilot passed all frozen checks on three authored tasks with 18.1% fewer input tokens. Estimated credits fell 26.8%, partly from unequal cache hits. A four-call extension checked read-only tools and one named skill; broad capability preservation remains unproven.

Latest evidence and open gates Earlier experiments
Expanded workflow pilot 39 model calls

27 workflows, including 20 matched task/arm runs. All stages counted; cache use recorded.

Opt-in catalog experiment About 17% lower

Mean estimated credits on one visual task, four calls, no cache hits. Not a guarantee; skill discovery may be reduced.

05 / Included

A measured control layer.

CLI, Python library, MCP server, and a ready Codex plugin.

06 / Install

Make your next task a measured one.

Python 3.10+ · Apache-2.0 · Direct-run workbench

Download the installer and wheel into one folder. Run the preview there; add --yes only after reviewing it.

Initial setup uses a terminal. Optional MCP dependencies require a download. Tasks are read-only; direct Astra is the default and extra stages can cost more. Broad savings and unaided onboarding are not yet proven.

python install_governor.py install --wheel premium_model_budget_governor-0.4.0rc6-py3-none-any.whl
Release notes and downloads Browse released source

Build the evidence with us

Useful feedback beats a perfect demo.

Technical or nontechnical: try the setup, bring a small non-sensitive task, and tell us where it helps or gets in the way. Independent testing is welcome. No endorsement expected.

Report your beta experience

Never share private prompts, project files, credentials, or local session links.