Independent co-founder forensic rebuild

Your architecture is right. A working system was deleted — and a cron is killing your agents.

Six months, $400+/mo, agents reinstalled three times, every session forgetting the last. This is the complete reconstruction — published on your own command center so you can finally read all of it in one place: what really happened, the day-one transcript of both agents, the brain decision, every idea from your repos, and the company-as-code plan that gets you to “work on this → done.”

Updated 2026-06-09 5 deep read-only agents 34,330 chat msgs · 522 sessions 72 containers · 23 plan docs GitHub-sourced · live data

00 · Start here

The 5-section forensic rebuild suite

Five deep sections. Each is its own page — built from your live VPS + GitHub, nothing stale, nothing deleted.

Forensics deep-dive →

The bombshell in full: the 50 deleted crons, the 6-hour kill-cron, the 72-container map, the GitHub control-plane audit, the 5 prior audits, the 6-joint chain. Tables + charts.

Day-one transcript →

Both agents — Hermes + OpenClaw — from May 8 to now, chunked phase → task → problem → solution. Public-safe, redacted. Your real story, visible.

The brain decision →

GBrain vs LLM Wiki vs Graphify vs Obsidian vs NotebookLM — evaluated with live docs (Context7), and the right combination chosen. Not a default.

Idea bank →

Every idea from your GitHub YAMLs — 4 groups, ~18 departments, the 47-agent roster, tenants, industries, hiring, operating model. Nothing lost.

The rebuild plan →

Company-as-code, Slack-controlled, never-amnesia. The architecture, the loop, the RESTORE→BRAIN→GitHub-ify sequence, the first 10 task packets — grounded in Karpathy, Cherny, Steinberger.

← Full command center →

Back to the live /migration dashboards: plan, tasks, audit, status, slack, odoo.

How the forensic rebuild fits

Read-only agents

5 deep agents sweep the live VPS + GitHub — no mutations, nothing deleted, nothing stale.

Evidence

34,330 chat msgs · 522 sessions · 72 containers · 23 plan docs — published as the 5 sections above.

Rebuild plan

Findings feed the company-as-code plan: RESTORE → BRAIN → GitHub-ify → the first 10 task packets.

01 · The verdict

The verdict, in your system’s own numbers

Every figure below comes from the system’s own committed surfaces.

The model is right — it’s exactly how Anthropic & OpenAI run: Git = truth, agents = workforce, runtime = disposable. So the design was never the bottleneck. 100% of the failure is execution — and that’s the good news: a broken idea takes years to fix; a broken bootstrap takes hours.

1 / 139tasks done. Throughput ≈ 0 — execution, not planning.

0×times the closure loop ever completed (issue→PR→evidence→close).

29 dayscouncil frozen at round 000 on one flag: pat_revoked:false.

522sessions in 30 days — each a cold start, re-finding the same blockers.

6hinterval a cron force-kills every agent — the “agents keep dying” cause.

50 → 1working crons, deleted by the “fresh rebuild.” see forensics →

02 · The architecture

Company-as-code, Slack-controlled

The target operating model — GitHub as single source of truth + brain, runtime disposable.

  SAM ──▶ SLACK  (top command layer — run the whole company: integrations · automation · templates)
          └─ Discord (2nd surface) · Telegram · WhatsApp  (department / comms channels)
                              │
                              ▼
   HERMES — operator · router · always-on        one per tenant: viewport · mlg · bccl · vinay
                              │  files a TASK PACKET  ( = GitHub issue )
                              ▼
  ╔══════════════════════════════════════════════════════════════════════════╗
  ║  GITHUB   =   SINGLE SOURCE OF TRUTH   +   THE BRAIN                       ║
  ║  • viewport-os monorepo (products · services · agents · tenants · corp)    ║
  ║  • BRAIN = git-versioned markdown vault, MCP-native, 24/7 enrichment loop  ║
  ║  • Issues = the one task ledger     • GitHub Actions = the durable loop    ║
  ╚═══════════════┬══════════════════════════════════════════════════════════╝
                  │  dispatch (durable · retried · audited in git)
                  ▼
   OPENCLAW — 47-seat specialist fleet (sandboxed)    runtimes on OpenAI Codex
   Claude (Max) joins via Claude Code CLI + GitHub Actions  (reviewer · architect · auditor)
                  │  deploys ONLY what GitHub declares  (IaC)
                  ▼
   VPS RUNTIME   =   disposable · rebuildable from Git
                  │  live state
                  ▼
   viewport.llc/migration   =   COMMAND CENTER  (this site — live · mobile · public/private)

Full model, the GitHub namespace, models/auth, and the recovery sequence: the rebuild plan →

03 · The loop

The loop you asked for

Your “Build the Loop” image = your 10-step workflow = the one live object the system runs.

CAPTURESlack/Hermes → issue

AUTOMATEroute + packet

EXECUTEOpenClaw seats

VERIFYreviewer + verifier

PUBLISHevidence + PR + page

REMEMBERbrain writeback (.md)

04 · Unblock

The one key that unfreezes everything

One Sam-only action, then one approval — everything else is done for you.

Sam-only, ~1 hour: revoke the over-scoped admin:enterprise token → mint repo+read:org → flip pat_revoked:true in Migration/council/STATE.md → push the council branch. Your own protocol halts every agent until this is done — it’s been pending 29 days. Everything else, I do.

Your one-click approval: neutralize the 6-hour kill-cron + add model fallbacks (stops agents dying).
Then I run the rest: restore the 50 crons, install the brain, GitHub-ify, rebuild this command center — GitHub-first, Context7-verified, nothing deleted.