Horizon Broker 0.2.0 — asynchronous inference for the edge fleet

The Horizon Broker v0.2.0

Asynchronous inference queue and fault-tolerance manager for a heterogeneous
edge fleet.

Stop treating inference as a blocking call.

Why

95% of planned agent workloads are long-horizon: a recursive codebase scan, a
research sweep, a test-generation run. They do not need an answer in
milliseconds. The broker classifies them as Flex tier work, holds them in a
per-model window until the batch is worth executing, and dispatches them when
the hardware is ready — maximising Model Flop Utilization instead of
minimising latency.

The fleet is also fragile. A Raspberry Pi overheats. A solar trailer hits a
low-voltage shutdown. The broker treats hardware death as an expected event:
state is checkpointed to EdgeRecord continuously, and a lost node triggers an
automatic re-dispatch on a different node inside a 30-second budget.

The four sub-systems

Sub-system	Module	Responsibility
Sleep-Time Scheduler	horizon_broker/scheduler/	Priority queue, per-model windowing, batch planning
Software-Defined Reliability	horizon_broker/recovery/	EdgeRecord snapshots, local journal, recovery coordinator
Heterogeneous Routing	horizon_broker/routing/	Arithmetic-intensity model, PD-DAG router, node placement
Throughput over Latency	horizon_broker/scheduler/window.py	Deliberate delay to build large batches

Plus the fault-detection heartbeat (horizon_broker/nodes/) and the API
ingress (horizon_broker/ingress/).

Commands

• horizon-broker serve --config config/broker.yaml — run the broker and its
  ingress API.
• horizon-broker check-config --config config/broker.yaml — validate the
  configuration and exit.

Install

pip install horizon-broker

Architecture and operational detail live in
docs/ARCHITECTURE.md
and docs/RUNBOOK.md.