Session Report · Architecture Stress-Test

The 5-Node Empire,
pressure-tested.

You sketched a decentralised home + business brain — five dedicated nodes, five supervisor bots, a Git-driven "synapse stream", and isolated enterprise personas. This session put it under load and came back with one decision.

3
vision entries
fed in
1
component that's an
outright hazard
~$100
path to 80% of
the value

2026-07-18 · repo valter · branch main · → arrows / space to move

What you proposed

Empire, exploded into five specialist nodes

Each node always-on, each running its own micro-supervisor bot. Empire drops to pure orchestration.

EMPIRE orchestrator Bot+DataMac Mini CodeMBP M5 HomeAssistant NetworkPi Terminusmedia/GPU

Plus: multi-identity Telegram/voice personas · smart-grid EV tariff-shifting · a Git-commit "memory delta" stream syncing every node · isolated Amazon + NZ Bank enterprise branches.

How I tested it

Strip the mythology, weigh each part

Named plainly, "Root Brain / Central State Matrix / synapse stream" is just a few small computers, a message bus, and some Telegram bots. Once you see that, the right size is obvious. Four lenses:

1 · Blast radius

What breaks if this ships — and can it hurt you legally, not just technically?

2 · Over-build

Value per node vs the ops surface each new box + bot drags in.

3 · Right tool

Is the mechanism (Git-as-bus) the correct one, or a heavier reinvention?

4 · Cost / ROI

Cheapest path to 80% of the value, and does the tax angle actually hold?

Finding 1 · existential

The enterprise branch leaks the wrong way

Corporate boundary Amazon · NZ Bank IP, code, project context Personal Empire a Git repo YOU control EXFILTRATION "ingestion channel" = the leak
Your guardrail — "no personal memories leak into the enterprise branch" — watches the enterprise from personal contamination. But the real risk runs the other way: employer IP flowing out into infra you own. The "one-way broadcast" ingestion step is the exfiltration. A private repo doesn't fix it: this breaches confidentiality + IP-assignment terms, and for a bank it can be a reportable regulatory breach. You've been bitten once already — adb's public release needed "Amazon/confidential content scrubbed".

Verdict  Don't build it carefully — don't build it. Enterprise context lives entirely inside their approved environment. The personal brain never touches employer work.

Finding 2 · over-engineering

A datacenter for a house of two

~14
agents once 5 supervisors join the ~9 fleet bots — each needs prompts, memory, patching, and a watcher
1
new single point of failure: Empire now watches everything
2 of 5
nodes actually earn their keep (Pi + Home Assistant)

Premise doesn't survive

The Code Node (MBP M5) assumes local AI bogs Empire down. But your coding is Claude Code (cloud), and heavy local inference already has a home: Terminus's GPU (tax-rag runs there at ~66 tok/s). It's a routing problem, not a $4k Mac.

Genuinely good

A Pi for network monitoring + Home Assistant + local voice is cheap, single-purpose, private. Keep it. Just skip the "HTML5 persona-bot per family member" gold-plating for v1.

Finding 3 · wrong bus

Git is a great archive, a terrible message bus

You killed the odd/even flip-flop for a race condition — correct — then replaced it with a distributed Git transaction system. Much bigger hammer than the nail.

Git-commit synapse stream

  • commit→push→webhook→pull = seconds of latency
  • high-frequency deltas bloat the repo forever (git never forgets)
  • any node that writes → merge conflicts, force-push recovery
  • NAT'd home nodes need reachable webhook listeners → tunnels

MQTT — already in your world

  • the native home-automation protocol; HA speaks it out of the box
  • built for real-time fan-out to many nodes
  • no repo bloat, no merge semantics, sub-second
  • Redis pub/sub or plain Telegram also fine — anything but Git

You don't need a bespoke synapse engine to dodge a status-line race — you need a lock, or a real bus.

Don't lose these

The parts worth keeping

first wedge

EV / tariff shifting

Concrete, bounded, pays real dollars — charge in the 8.85¢ super-off-peak solar window. HA does this natively. Build this first.
$0

Terminus

Already your media node — and your GPU inference node. Lean on it harder instead of buying a Mac.
reuse

Empire + fleet

Keep Empire as orchestrator and extend the existing Hermes bots — don't mint five new ones.
Two guardrails on the EV automation: (1) a bot must never auto-enrol in VPP grid events or change your tariff — money + contract, keep human-in-the-loop; (2) "email receipt → auto-detect EV delivery" is fragile — one misparse fires the wrong policy, so keep the trigger human-confirmed even if the execution is automated.

The decision

Build small. Cut the hazard. Prove before you spend.

CallItemWhy
🟢 BuildPi = network + HA + local voice · Terminus for GPU AI · MQTT bus · EV/tariff automation as the first wedgeCheap, concrete, pays for itself, uses what you already own
🟡 MeasureDedicated Mac Mini "brain" node · multi-identity persona botsOnly if Empire load is a proven bottleneck; personas after HA basics work
🔴 CutEnterprise Amazon/NZ-Bank persona branch · Git-as-synapse transport · MacBook Pro M5 code nodeLegal/compliance landmine · wrong tool · premise doesn't survive (cloud coding, GPU on Terminus)

Bottom line: the vision is coherent and exciting, but it's the same over-build reflex that hit the solar doc — and one component is an outright hazard. The 80% version is one Pi, MQTT, reuse Terminus + Empire, EV automation as proof — for ~$100 instead of several thousand.

The trimmed system

What ~$100 actually buys you

MQTT bus · real-time, no repo bloat EMPIRE orchestrator (reuse) Raspberry Pi ~$100 Home Assistant · Pi-hole local voice · EV/tariff Terminus $0 GPU AI · media (reuse) Telegram family interface (existing fleet)

Two of your five nodes for ~$100; two more at $0 by reusing Terminus + Empire. No new Mac, no Git-synapse, no enterprise branch.

Next actions

Six moves, cheapest-first

Today · $0

  1. Write down the hard rule: the personal brain never ingests employer work. Kill the enterprise-persona branch on paper.
  2. One evening: measure Empire's real load — is anything actually the bottleneck, or is it CPU-bound Ollama?

This week · ~$100

  1. Buy one Raspberry Pi; install Home Assistant + Pi-hole.
  2. Route heavy local AI to Terminus's GPU (already deployed there).

Then · the wedge

  1. Build the EV / tariff automation in HA — human-confirmed trigger, no auto-VPP enrolment.

Bus

  1. Stand up MQTT as the node bus (ships with HA). Retire the Git-synapse idea.

Call to action

Buy the $100 Pi this week.
Prove the value before the $4k.

Wire one automation that pays for itself — EV charging in the solar window — and let that success (not the vision doc) justify the next node.

Turn the 🟢 slice into an adb initiative + tickets
Capture this trimmed architecture + the leak red-flag to the wiki
Run it through a formal small-council vote

Say the word and I'll kick off any of the three.

Session facts & caveats

Transparency footer

What this session did

  • Ingested 3 pasted vision entries; ran a 4-lens architecture stress-test; produced a decision + this deck.
  • No repo files were created or edited by this session — the only artifact is this presentation.
  • The pending items in git status (FIFO tickets, rookcast, memories-ai) are pre-existing / unrelated work, not from this session.

Artifact

  • reports/session-report-2026-07-18-2016-network-node-stress-test.html — this deck

Fact sources

git ✓ adb ✓ mtime scan ✓ gstack (thin) transcript (22MB, skipped)

Recap reflects available context + git facts; a long/compacted session may omit earlier steps. Diagrams are inline SVG; motion is CSS/SVG animation (Remotion-inspired, not a Remotion render — this stays one self-contained file). File lists are derived from git + the mtime scan, not memory.

← → or space to navigate · ▶ to autoplay · ◐ theme
1 / 12