What it does
Common Ground helps a community workshop host keep a workable plan when the budget, weather or venue changes. A Planner and an independent Feasibility Reviewer work concurrently, share priorities and proposals, and reconsider their work when the brief changes. Code calculates costs in integer cents and checks capacity, step-free access, rain shelter, venue closures and a 10% reserve. The interface and planning notes use English, with fictional prices in EUR. The host reviews the final plan; no bookings or payments are made. The hosted demo is a clearly labelled recording of a real local model run. The README explains how to run the live, editable application with a local model and no paid API.
How the agents run concurrently
The Planner and Feasibility Reviewer are separate @mozaik-ai/core agents running independent runLoop loops against one shared RuntimeState. Both receive the same brief and can issue real inference requests concurrently. The Reviewer shares a planning priority while the Planner develops a proposal; runtime events notify the agents of changes to the constraints, priority, proposal and review. The Reviewer then examines the exact current proposal. Responses from older brief revisions are discarded, and a review must match the current proposal ID. Both agents use actual Qwen3-0.6B-Q8_0 inference through llama.cpp with two parallel decoding slots. The current EUR browser run in evidence/live-browser-run.json records 9 completed real model calls, 5 overlapping Planner/Reviewer call pairs and 3 discarded stale responses after a change during inference. The recorded plan totals €462.00, including the 10% reserve, within a €480.00 budget. The repository includes timestamped runtime events, loop IDs, request/response fingerprints and deterministic tests explicitly distinguished from real model runs. Model-generated wording is marked unverified; code performs the hard constraint checks.