mote_fleet

The fleet control plane: the agent that runs on a robot, and the enrollment, map-registry and dispatch server that runs off it. One package for both halves, because the thing that matters most is that they agree about the wire — and the wire is a single module, protocol.py, that both import.

Same arrangement as mote_perception: a torch-free node on the robot, a server off-board, and one shared wire module (depth_wire.py) so the two cannot drift. Here the robot-side half is the ament package and the off-board half is server/ — plain scripts, because the fleet box runs them without installing anything.

The design doc calls the robot-side process mote_agent, and that is what it is called everywhere it is visible: the node is mote_agent, the service is mote-agent.service, the console script is agent.

Interface contracts	control-plane.md — the MQTT topic tree and payload schemas · fleet-api.md — the HTTP routes, dispatch and audit
Operator runbook	docs/fleet/README.md §6–9, and §11 for maps
Deploying the server	server-pipelines.md — the container stack, gated updates, backup/restore
What was measured	m1-verification.md · m3-verification.md · m4-verification.md · ms-verification.md
Design	docs/design/fleet.md — M1, M3, M4 and Ms, and Q1/Q2/Q3/Q4/Q5

On the robot

pixi run enroll -- --server http://fleet-box:8080 --token <token>
pixi run agent
pixi run publish-map            # after save-map: offer the map to the registry

Module
agent.py	the node: presence/health/pose up, one task at a time down
dispatch.py	the single-in-flight rule; ROS-free and MQTT-free, so the awkward cases are plain function calls
protocol.py	the wire contract — topics, payloads, states. Stdlib only
enroll.py	the enroll CLI
facts.py	hardware facts and the fingerprint enrollment is idempotent on
fleet_config.py	$MOTE_HOME/fleet.yaml — where this robot's fleet lives
mapsync.py	the map registry's robot side: pull the canonical revision, publish a candidate. ROS-free
publish.py	the publish-map CLI

The agent is a bridge and a reporter, never in the control loop. Nav2, SLAM and the behaviour tree run locally and keep running with the fleet server unplugged; a dropped link means the agent stops reporting, not that the robot stops. It is also the robot's sole egress — nothing off-box joins the ROS graph — which is what makes the fleet-scale DDS story a non-problem rather than a design (fleet.md Q1/Q3).

Off the robot

pixi run fleet-broker                                      # mosquitto + WebSockets
pixi run -e fleet fleet-server -- --broker-host fleet-box  # API + dashboard
pixi run -e fleet fleetctl -- dispatch mote-01 goto kitchen

Script
server/fleet_server.py	the fleet API: enrollment, roster, dispatch, audit, basemaps, and the UI — stdlib http.server
server/registry.py	the SQLite row store: robots, enrollment tokens, operators, the audit log, transactional id allocation
server/bundle_store.py	the map registry's byte store: candidate revisions, validation on the way in, the atomic flip that publishes one
server/fleetctl.py	operator CLI: tokens, roster, dispatch, audit, watch
server/ui/	the dashboard: index.html, app.mjs, map.mjs (basemap + the Q5 transform, pan/pinch), mqtt.mjs (a subscribe-only MQTT client), layout.mjs (one pane at a time on a phone)
server/mosquitto.conf, broker.sh	the broker, its WebSocket listener, and where its state goes
deploy/	the deployed shape: an image for the API+UI, a compose file that runs it beside the broker, and fleet-deploy.sh (gated update, rollback, backup, restore)

The server imports mote_fleet.protocol and mote_bringup.bundle from the source tree by path (the depth_server.py pattern) and nothing else — no ROS, no framework, no ament. protocol is the wire the robot and the server agree on, and bundle is the bundle format they agree on, so the server validates an uploaded map revision with the same code that wrote it rather than a second implementation that agrees by convention (fleet.md Q4). protocol is stdlib-only; bundle additionally imports PyYAML and Pillow, which the image installs beside paho — reading these files with anything other than the libraries that write them is exactly the second implementation the rule exists to avoid, and both hand-rolled readers it shipped with proved the point. Server state lives in $MOTE_FLEET_HOME (default ~/.mote-fleet), with the site bundles under sites/.

http.server rather than a web framework stays a floor, not an aspiration: a dozen routes, no templating, no ORM, and one fewer dependency to solve on whatever the fleet box turns out to be. M3 was where a framework was expected to earn its keep and it did not — the UI is static files and there is nothing to render server-side. The same goes for the browser: no bundler, no npm, and no vendored MQTT library, because what the read path needs is five packet types of a published wire format and a minified blob nobody can review is a worse dependency than 200 lines that are tested.

Dispatch is mediated by the server, and only by the server. Every write to task/command — from the dashboard or from fleetctl — is a POST that authorizes an operator token and writes an audit row first. The read path is unchanged and goes straight to the broker.

Uploading a map is not publishing it. A revision a robot uploads is a candidate that changes nothing; an operator promotes one, which flips the floor's map symlink and publishes the retained …/current topic agents pull from. Two robots that map one floor leave two candidates, never a merge.

Tests

pixi run test                 # colcon: everything except the broker tests
pixi run -e dev test-fleet    # + the real-broker end-to-end run

Four tiers, so the same files give full coverage wherever they run:

• contract (test_protocol.py, test_fleet_server.py, test_map_registry.py, test_registry.py) — the code, the JSON Schema files and the doc's field tables checked against each other, and every HTTP route over a real socket with an injected publisher, so a payload or status-code change that nobody described fails here rather than in a dashboard later. api_harness.py is the live server the last two share.
• bridge (test_dispatch.py, test_agent.py, test_mapsync.py) — the single-in-flight rule and the full agent against an injected fake MQTT client, so CI covers it on both architectures without a broker; and the robot's map staging and symlink flip against a real fleet server, with no ROS at all.
• browser (test_ui.py → ui_test.mjs, test_fake_robots.py) — the MQTT packet codec and the world→pixel transform under node, against the same .mjs files the browser loads (skips where there is no node); and the wire-only robots the dashboard is checked against, held to protocol.py and to the task layer's grammar, so the fixture can never become a second definition of the wire. browser_check.mjs is the other half — a real headless browser against a running stack, which needs a docker and a chrome, so it is an operator's tool rather than a test (the reasoning, and what wiring it into CI would take, are in docs/fleet/m3-verification.md §2). It needs no stack of your own: pixi run fleet-ui-check builds one — broker, server, basemap, and the wire-only robots of fake_robots.py — on ports nobody else is using, runs the checks and tears it all down. -- --keep leaves it up instead, which is the loop for working on server/ui/.
• end to end (test_e2e_fleet.py, test_e2e_map_registry.py, test_fleet_outage.py) — a real mosquitto, the real fleet server, the enroll CLI, a real paho client, and the actual mote_tasks behaviour tree driving a mock Nav2; including a dispatch that goes out through the API. The second kills the broker under a live agent: the robot finishes its task anyway and the agent reconnects by itself, which is the claim the fleet server's update pipeline is allowed to have downtime on. The third publishes a map, promotes it with fleetctl, and starts a second robot's agent afterwards — so the only thing that can tell it about the map is a retained message handed over on connect. All skip where there is no broker, and share fleet_harness.py.