Release Notes: QSR Store Agent#
Version 2026.3.0#
Release Date: December 2026
New:
Operator chat with live answers from connected store services, using the local Qwen3 INT4 model (set by
MODEL_ID) served by the OpenVINO™ Model Server on an Intel® GPU.Event-driven recommendations with operator approval for any store event type.
Learning from operator responses, with operator review before learned rules take effect.
Scope guardrail that declines questions unrelated to the store.
Weather service with on-demand simulation and optional live Open-Meteo data.
Store skills for kiosk operations, order accuracy, kitchen safety, digital signage, café operations, table management, and footfall analytics.
Single-command deployment with Docker Compose, using a pre-built image or a local build.
Store data comes only from the connected store services; the built-in simulations, except weather, are removed.
Learning from the live events of any connected service.
Fixed:
make learn ARGS=runon a fresh install creates the learner profile itself.A learned skill can read its own service and every context its rules check.
A service snapshot that is already stale when read now fails with a clock-skew message instead of five retries.
When the model proposes an action without reading the context it requires, the agent reads it and asks the model to reassess, instead of failing after five rounds.
Learned rules waiting for review survive container restarts and rebuilds.
make learntrusts the self-signed certificates of remote services.Learning no longer adds a rule from only two records that the other rules miss.
Operator chat loads the matching store skill before it answers or acts.
Store skills use the current tools and fields of their services, and answer time-window and count questions correctly.
Known Limitations:
Weather data is simulated unless live Open-Meteo weather is configured.
The operator UI has no built-in authentication. Restrict network access in production.
Self-signed certificates on remote services are trusted on first use at start-up. Use properly signed certificates in production.
One operator conversation and one event are processed at a time.
A learned skill lists when to act but cannot forbid an action, so the base skill may still propose one operators usually reject.