hivetree — a three-tier HiveMind test mesh

A disposable multi-node HiveMind deployment for exercising the parts of the protocol a single hub can not reach: multi-hop routing, cross-subtree delivery, and relay-loss resilience.

            m0   root master + echo agent      host :6700
           /  \
         r1    r2   relays, upstream -> m0      host :6701 / :6702
        /  \   /  \
   sat-1a 1b 2a 2b   leaf probe clients (not containers)

Each node runs one container: a local OVOS messagebus plus hivemind-core. The root also runs a small echo agent that answers an utterance with echo <text>, so a message that travels up the tree produces a speak that routes back — no skill model load, the tree forms in seconds. Relays carry an empty bus; their downstream traffic escalates to the root.

Upstream links are pre-wired: each relay holds a fixed client credential on the root (see .env), so docker compose up forms the tree with no manual copy step. Every client is granted the full routing message-type whitelist (hivemind-core is deny-by-default).

Run

cp .env.example .env               # throwaway test credentials for the tree
docker compose --env-file .env build
docker compose --env-file .env up -d
docker compose logs -f r1 r2        # look for "Upstream master: m0:5678" + a Noise session

Probe

From the docker host, against the published ports:

~/hivemind/.venv/bin/python -u probes/tree_probe.py all

Tests: escalate (leaf → root agent → back, two hops), propagate (fan-out across subtrees), intercom (targeted cross-subtree delivery), query (participation binding: the answer returns only to the asking leaf, and a forged is_response from another subtree is dropped).

Relay-loss check (the emit path must keep serving other peers when one relay dies):

docker stop hivetree-r1     # then re-run a probe on an r2 leaf; it must not stall

Tear down

docker compose down -v

The image installs the latest HiveMind alphas plus the pinned hivemind-core wheel in wheels/, so the mesh runs the exact build under test.