Plan: one current-truth control layer

Stop the split-brain. Build the Green Board.

A practical control registry for Mike’s AI ecosystem: MacHermes, Herman/VPS3, OpenClaw personas, memory spine, Browserbase/proxies, Vercel, Supabase, Bird’s Eye ROI, and DWS assets. One place to see what is real, verified, authenticated, and safe to use.

Read the build plan See registry model

Council verdict converted to plan

The architecture is right. The missing piece is governance.

Mac as brain and VPS as operator is the correct pattern. The failure mode is coordination drift: too many tools remembering, deploying, scraping, or claiming state without one enforced registry.

One source of operational truth

Not one memory system for everything — one board that records what each tool can do right now.

Runtime checks before claims

Every agent must verify current shell/tools/auth before saying it can execute.

Memory governance

Define canonical memory, retrieval memory, archive notes, and task state so agents stop mixing them.

Registry model

The Green Board columns.

This is the exact table Herman should check before executing meaningful work.

Name	Lane	Runs On	Status	Auth	Memory R/W	Can Deploy	Failure Mode	Next Fix
MacHermesPrimary brain	Strategy / orchestration	Mac Studio	Needs heartbeat	Local	Read/write Obsidian + memory spine	Maybe	Unavailable to VPS shell	Heartbeat endpoint + SSH check
HermanVPS3 operator	Outbound / deploy / reports	VPS3 / Telegram runtime	Partial	Per-shell verified	Honcho + session memory	Vercel API when token present	Memory ≠ shell state	Mandatory runtime audit command
OpenClaw roster22+ personas	Specialist roles	Mac / ecosystem	Context known	Not verified here	Unknown from VPS	No from this shell	Persona names mistaken for live agents	Agent registry + command bridge
Memory SpineHoncho/Hindsight/Mem0/Obsidian/Supabase	Context / recall / archive / task state	Hybrid	Needs contract	Mixed	Mixed	No	Conflicting or stale authority	Write policy + conflict rules
Browserbase / ProxiesBrowser lane	Web interaction / anti-bot routes	Cloud	Unverified	Project ID needed	No	No	YouTube/browser blocks	Set project/env + test flow
VercelDeploy surface	Public artifacts	Cloud	Working now	Token supplied	No	Yes	Token exposure / expiry	Rotate token + env source

One thing first

Build the registry before adding more agents.

Freeze expansion. Make the system prove what is alive, authenticated, and allowed to act. Then the 22+ personas become leverage instead of noise.

PHASE 1

Static truth board

Create JSON registry file.
Render board as Vercel page.
Track lane, runtime, auth, last verified, failures.

PHASE 2

Runtime probes

Check commands, paths, tokens, network, deploy.
Write status snapshots.
Stop agents from claiming unverified tools.

PHASE 3

Memory contract

Obsidian = human source of truth.
Supabase = task/control state.
Honcho/Hindsight/Mem0 = recall/context, not unchecked authority.

PHASE 4

Agent routing

Map each persona to lane and command route.
Add last successful task.
Add owner and escalation rules.

Governance rules

Rules that prevent Herman from bullshitting.

Verify current runtime before execution.

If the tool/path/auth is not visible in the current shell, report it as unverified. Do not rely on memory.

Separate memory from authority.

Recalled memory informs decisions; current probes decide what can actually be done.

One task lifecycle.

Create task → assign lane → pull context → execute → verify → write result → notify Mike.

Outbound needs approval gates.

No scraping/outreach expansion unless Mike explicitly reauthorizes it.

Green means verified today.

Status expires. A green badge must have a timestamp, not vibes.