HoneyBun — AI Orchestration Platform Architecture

Context Hierarchy

Four context layers. Conflict-resolution baked in.

Context files load bottom-up. Project specificity wins. When ambiguous, the more restrictive constraint applies. The pattern most enterprise AI programs build by year three — designed in on day one.

LAYER 1 / GLOBAL

Always-On Instincts

Non-negotiable behavior rules. Verify-first. No claim without source read. Stakes-asymmetry beats speed.

LAYER 2 / ECOSYSTEM

Projects-Wide Standards

Quality bar, security, testing, performance, git workflow, ship-gate, CI/CD discipline.

LAYER 3 / PLATFORM

HoneyBun Platform Rules

Multi-tenant patterns, vertical theme system, worker conventions, deploy rules, operator vs. internal split.

LAYER 4 / PROJECT

Per-Repo Context

Workers, dashboard, PWA, themes, clients — each surface gets its own CLAUDE.md with deploy commands and gotchas.

PROTOCOL

Task Board API · Single Source of Truth

Every task, every agent, every session writes to one board. Status flow, orphan sweep, parallel-session race protection.

MEMORY

MemStack · Cross-Session Knowledge

SQLite + semantic vector search. Sessions, decisions, lessons, project context. Karpathy-style LLM vault for institutional memory.

HoneyBun Runtime · System Architecture

73 Cloudflare Workers. Seven layers. End-to-end request flow.

HoneyBun's production runtime is 73 Cloudflare Workers (4 core + 69 specialist) — distinct from the Claude Code sub-agent tooling used to build it. Every page view travels through seven layers: acquire → client → WordPress theme → core workers → specialist workers → storage → output surfaces. The mu-plugin bridges WordPress and Cloudflare Workers invisibly. The diagram below is the actual production topology.

Acquire

audit-funnel

Vercel lead doorway · engagement scoring

honeybun/audits/

Per-prospect audit pages · CF Pages (honeybun-audits)

lead-magnets/

Local Node scripts → GHL webhooks

go.gethoneybun.com

Sales pages + pitch decks (every .html auto-deploys)

↓Stripe checkout → /webhook/stripe

Client

Browser

Visitor lands on operator site

↓

WordPress

template-*.php

Per-vertical theme templates

mu-plugins/

hb-heartbeat · hb-hero-preview · DCC bridge

wp-app-root/

hb-deployer.php · hb-theme-sync.php

↓GET /render/:clientId/:pageType/:slug

Core Workers

hb-provision

Monolith router · clones WP · assigns subdomain · proxies to extracted services

hb-content

Sonnet research → Opus copy

hb-render

HTML + CSS + JSON-LD schema

hb-agent

SEO audit · intake · content

↓Orchestrated by

Specialist

Orchestration

orchestrator · task-dispatcher · watchdog · ci-receiver

Monitoring

health-monitor · drift-remediation · cert-probe · status-page

SEO / Audit

seo · seo-playbook · rank-tracker · interlink · schema-validator

Funnel / Leads

funnel · leads · checkout · prospect-intel · density-pricing

Platform

verticals · clients · deploy · auth · session · analytics · seotools · og · image · messages · banner · status-page · assets

Dev Tooling (~/Projects/tools/)

claude-orchestrator (Railway queue) · headroom (compression proxy) · ghl-mcp · gsc-write · css-diagnose

↓Reads / Writes

Storage

KV · HB_CLIENTS

Client registry · DCC configs · cache

R2 · hb-assets

Static assets · theme bundles · artifacts

Supabase

Tasks · lessons · drift events · analytics

↓Returns { html, css, schema }

Output

Operator Site

SEO HTML · schema in <head> · injected DOM

Operator + Internal PWAs

Cloudflare Pages · dual app from one codebase

app.gethoneybun.com

Dashboard · Vercel · analytics · workflows

Build Tooling · The AI Operating System I Use to Build Software

HoneyBun was built by 39 sub-agents. Not engineered by humans.

This is the personal AI operating system I built to direct software work — separate from HoneyBun's production runtime. 39 specialist Claude Code sub-agents, 6 binary quality gates (Plan · Code · Security · Test · Build · Business), and 6 executive-persona reviewers route every task through a custom 270-line dispatch protocol. No gate is skippable. A RED verdict from a relevant persona blocks execution. The platform that runs on AI was itself built by AI orchestration.

plan-checker

Goal-backward plan validation. Max 3 loops.

code-reviewer

No CRITICAL/HIGH lands. Fires on every code change.

security-reviewer

Auth, input, secrets, API. Rotates exposed keys.

tdd-guide

Tests-first enforcement. 80% coverage minimum.

planner

Standard + complex implementation plans.

assumptions-analyzer

Hidden assumptions surfaced with evidence.

advisor-researcher

Gray-area decisions. Parallel × N.

research-synthesizer

Merges parallel research into one brief.

architect

System design, multi-system decisions.

debugger

Root-cause analysis. Stack-trace triage.

api-detective

Failed calls, missing data, OAuth breaks.

database-reviewer

SQL, schema, migrations, performance.

build-error-resolver

Build green with minimal diff.

e2e-runner

Critical-flow verification via browser.

refactor-cleaner

Dead code, knip, depcheck, ts-prune.

doc-updater

Codemaps, READMEs, change docs.

exec-ceo · Bezos

Customer obsession. Day 1. Flywheel.

exec-coo · Cook

Operational precision. Single-thread owner.

exec-cfo · Munger

Inversion. Moats. Unit economics.

exec-cto · Vogels

Failure modes. API-first. Scale-under-load.

exec-cmo · Godin

Tribe. Smallest viable market. Purple cow.

exec-caio · Karpathy

Software 2.0. AI theater detection. Evals.

skeptic

Destruction-test. Severity-rated problems.

strategist

Market intent. Differentiation. Brief.

Factory Pattern

Marginal cost of a new vertical ≈ zero.

Templates ≠ tenants. The 13 golden apps are the factory; live operators are forks of the factory line. Each carries its own git SHA and deployment timestamp.

PROSPECT

→

CLONE GOLDEN

→

DEPLOY LIVE

photo-booth · LIVE

realtor · LIVE

photo-booth · LIVE

plumber · golden

medspa · golden

gym · golden

barbershop · golden

nail salon · golden

+ 8 more verticals

tenant #50 → same cost as #6

Operator Data · One Canonical Record

We already know your business.

Prospecting, the pre-deploy intake, in-app onboarding, the brand & schema generators, and the CRM all read and write a single canonical record per operator. Operators don't fill out forms — they confirm. Every field carries provenance, so operator-confirmed data always wins and a machine re-scrape can never overwrite what a human verified.

SOURCE OF TRUTH · PROVENANCE

One record, every surface

Single canonical store — every operator-data surface reads and writes the same record. No second source, no drift between tools.
Field-level provenance — each field is stamped with its source, timestamp, and confidence (human-confirmed vs. machine-derived). This is data provenance, alongside the SHA-256 deploy ledger that governs file integrity.
Confirmed data wins — a low-confidence scrape can never overwrite a field a human confirmed. Conflicting writes are version-checked, never last-writer-wins.

PRE-SEED · ZERO BLANK FORMS

Pre-seeded, not blank

Seeded before first touch — research and public-data signals fill the record before the operator ever logs in.
Confirm, don't type — the intake and onboarding surfaces show what we already know and ask only for what's missing or worth correcting; screens already answered are skipped.
Less typing, faster launch, fewer errors — the operator's first experience is review, not data entry.

GENERATION · OPERATOR-LEVEL FACTORY

One click → brand & schema

Generated from the record — one action produces a full brand identity and the complete structured-data stack for an operator, drawn from the canonical record.
Zero marginal cost, per operator — the same near-zero marginal cost that adds a vertical extends to every operator: no manual rebuild, no re-keying.
One-way sync outward — confirmed business details flow on to the CRM automatically, while login credentials are never touched.

Mobile Cockpit · Two PWAs, One Codebase

The internal app is the AI operations cockpit F500 programs need.

~63,000 lines across 46 screens sharing 37 cross-app modules. Customer-facing PWA and internal AI ops console deploy from one source. No App Store review cycles.

OPERATOR PWA · CUSTOMER-FACING

operator.gethoneybun.com

edit-sections — 8,743 LOC. Self-serve site editing.
signal — 2,894 LOC. Inbound lead intelligence.
onboarding — 2,162 LOC. Activation flow.
conversations · leads · card · dashboard · edit · account
Workbox service worker · web push · install prompt · offline.html

INTERNAL PWA · AI OPS COCKPIT

ops.gethoneybun.com

leads-platform — 3,325 LOC. Funnel ops surface.
pipeline · rank-tracker · errors · intelligence
prospect-research · prospect-qualifier · prospect-demo — AI-assisted sales workflow
seo-perfector · seo-flywheel · seo-health — autonomous improvement loop
aeo-queue · dcc-admin · converter · lead-prefill · verticals · audits · board · assets · inbox · pages
24 internal screens spanning governance, AI workflows, and observability

Reliability Discipline

Silent failure is the cardinal sin.

Every catch block routes through reportFailure(). Auto-remedy attempted first. Triple-channel escalation if it can't self-heal. Named human owner on every alert.

1 · DETECT

reportFailure()

Client-side hook. Severity tagged. Context attached.

2 · CLIENT RETRY

Retry × 2

0s then 5s delay. Toast if still failing.

3 · WORKER

Auto-Remedy

Health check · dedup · auto-fix task created.

4 · VERIFY

Re-check × 2

5s intervals. Confirm fix or escalate.

5 · ESCALATE

Push · Email · SMS

Named human owner. No failure exits silently.

Lifecycle Communications

Every message governed, channel-aware, and editable without a deploy.

Seven automated cadences carry a contact from prospect to retained operator — each with a defined trigger, channel, and sender. Emails ship via Resend; SMS via Simplex (white-labeled GoHighLevel); push via the PWA. The copy is config-driven, not hardcoded.

CADENCE ENGINE

Seven cadences, one fail-safe path

Prospect: lead reinvite (M2–M5), re-engagement drip, GHL nurture.
Onboarding & live: post-checkout welcome, site-live drip (day 0/1/7/14/30), monthly performance report.
Operator: instant new-lead alert (SMS + email + push) with vertical-aware copy — event-shaped for photo-booth, universal for everything else, safe-default universal.

EDITABLE COPY · NO DEPLOY

A store the cadence reads at send time

Override-or-default — every message reads an editable template from the store, falling back to the hardcoded copy if missing or disabled. A blank or broken edit can never blank a live message.
Branded editor behind a scoped key — edit copy in the browser, live within minutes via a KV cache. Toggle a template off to instantly revert.
Mapped & auditable — a live visual map documents every message's trigger, channel, sender, and verbatim copy.

Territory Economics

One operator, one ~550K cluster — priced by the data, not by hand.

The same factory that makes a new vertical nearly free also makes a new market deterministic: territories are computed from census data, and pricing scales with the geography and page count they generate.

CLUSTER MODEL

ZIP-adjacency to a population cap

~550,000 population per operator per vertical, built by ZIP/ZCTA adjacency from an anchor ZIP — cities are emergent from the ZIP set, not entered by hand.
Pure-function engine (polygon-boundary-v1, ACS census) returns the exact ZIP list + population; growth appends ZIPs (extend-cluster) without reshaping shared geography.
Binary ownership — owned or open, per city/ZIP.

DENSITY PRICING

Tiered by market density

Standard < 250K → 1.00× · Growth 250K–1M → 1.25× · Metro > 1M → 1.50×.
Page-count driven — final price scales with cities × services; Founding-Partner pricing stays decoupled from territory size.

Security Posture

Enterprise-grade isolation. Built in, not bolted on.

Data isolation, deploy integrity, bot protection, and transport security are first-class design constraints — not compliance checkboxes added after the fact.

DATA ISOLATION · AUTH

Scoped operator keys + RBAC

Per-operator API keys — each operator's key is scoped to their own data. One operator's credential cannot reach another operator's records.
Formal permissions library — can(), isPrivilegedAdmin(), and role constants gate every sensitive operation. Not ad-hoc if-checks.
Master key separation — platform operations require isPrivilegedAdmin(). Operator keys can't self-elevate.

DEPLOY INTEGRITY

SHA-256 ledger on every file deploy

Pre-write hash comparison — detects concurrent write conflicts before the write lands. Returns 409 on mismatch; no silent overwrites.
Append-only provenance — every deploy records operator, file path, holder, before-hash, after-hash, and timestamp in Supabase.
3-attempt retry on ledger insert before alerting — provenance is never silently lost even under transient DB pressure.

BOT PROTECTION · TRANSPORT

Turnstile + full security header stack

Cloudflare Turnstile on all lead capture forms — invisible managed CAPTCHA, blocks bots before form submit. Lazy-loaded; zero pageload impact.
HSTS with includeSubDomains — HTTPS enforced across the entire domain tree at the browser level.
CSP with explicit allowlist — script-src and connect-src locked to known endpoints. Frame-ancestors blocks clickjacking. base-uri prevents base tag injection.
Full header stack — nosniff, X-Frame-Options, Referrer-Policy, Permissions-Policy (camera/mic/geo/payment all off).

RATE LIMITING · RESILIENCE

KV gates + CF WAF + timeout discipline

KV-backed throttle gates at the application layer — 429 enforcement on all high-sensitivity endpoints before they touch the database.
CF WAF as the upstream volumetric layer — rate limiting, bot score, and threat scoring handled before the request hits the worker.
AbortSignal.timeout() on every external call — 3–8s depending on path criticality. No hanging requests, no cascading failures.
Read-vs-write retry discrimination — reads retry on 429/502/503/504; writes only retry on 503/504 to prevent double-writes on ambiguous 502s.

Operating Cadence

AI-native ops, not AI features bolted on.

AUTONOMOUS LOOP

Cron-Scheduled Agents

Morning briefing on session start — overnight autopilot, failures, stale tasks, today's schedule
Orphan sweep — stale tasks from dead sessions reclaimed automatically
Pre-flight checks — credentials, upstream reachability, no conflicting in-progress work
Health indicators — green / yellow / red per scheduled automation, surfaced at session start until resolved
Bounded autonomy — never archives human-claimed work; only meta-stale alert noise

LEARNING LOOP

Lessons → Hooks

Every failure produces a per-shard lesson at ~/.claude/lessons/
Recurring patterns get promoted from probabilistic rules to deterministic hooks
5+ formal post-incident write-ups in the workers repo (circuit breakers, drift retries, timeout handling)
143 codified institutional lessons (1,281 lines) in honeybun/lessons.md
Verify-before-work protocol — workers check completed_at + existing code state before claiming any task. Prevents parallel-session re-do.

Autonomous Runner · Tiered Execution Pipeline

Task lands. System decides. You approve or ignore.

Every task is classified by risk the moment it hits the board. Low-risk work runs, verifies, and commits without a single human touch. High-risk work builds in an isolated branch, runs an independent verification pass, and surfaces a one-click approve/reject card with the full diff attached. After merge, a health probe watches the live endpoint — two consecutive failures within five minutes trigger an auto-revert.

Tier 0 · Verify-Only

Read-only DONE_WHEN. Single verifier pass. Closes in ~30s. No Researcher, no Fixer, no worktree.

Tier 1 · Reversible Edit

Single-file or small change. Two-pass Researcher → Builder in isolated worktree. Diff guard + sensitive-path check before push.

Tier 2 · Multi-File / Structural

Builder commits in worktree. Independent Verifier runs DONE_WHEN. Task surfaces as approve/reject card with diff + proof attached.

Tier 3 · Blocked

Infra, auth, migrations, secrets, destructive ops. Blocked from autopilot. Requires a human Claude Code session.

Topology B — Tier 2 Pipeline

The four-station mini-line for multi-file structural changes.

1 · RESEARCHER

Read-only analysis

Maps codebase, produces FIX_SPEC + DONE_WHEN. No writes.

2 · BUILDER

Isolated worktree

Implements + commits on autopilot/<taskId>. Never touches live source repo. Diff guard blocks sensitive paths.

3 · VERIFIER

Independent session

Separate read-only Claude session. Runs DONE_WHEN against the worktree. Reports PASS/FAIL with raw output.

4 · HUMAN GATE

One click

Review card: diff + verifier proof. Approve → merge bot pushes. Reject → branch deleted, task reopens.

5 · PROBE

Post-merge watch

DONE_WHEN re-run on live endpoint. Two failures in 5 min → auto-revert pushed to main before anyone notices.

SAFETY PROPERTIES

Limits that can't be bypassed

200-line / 5-file diff cap — exceeded limit escalates to human review, never auto-merges
9 sensitive-path patterns — wrangler*.toml, .env*, migrations/, auth/, billing/, CI workflows, deploy scripts
Worktree isolation — concurrent tasks on the same repo never collide; each runs on its own branch in /tmp/hb-build/<jobId>
Verifier independence — the agent that builds never verifies its own work; a fresh session runs DONE_WHEN in the worktree before the branch is ever proposed for merge
DONE_WHEN enforced at the API — POST /tasks returns 422 if the description is blank or missing a DONE_WHEN: clause. The verifiable success criterion isn't a convention — it's a hard gate. A task that can't be objectively checked can't enter the queue.
Human-created tasks are never exempt — system-generated notifications ([ALERT], [REVIEW:*], etc.) are exempt from the gate by design. The moment a human presses a button and converts an alert into a work item, full validation applies — a button press is a work decision, not a notification.

METRICS

Before and after

Dispatch latency: 900s cron lag → <5s event-driven (task creates/transitions fire dispatcher immediately)
AUDIT task cost: 3-min two-pass → ~30s single verifier pass
Concurrent tasks: source-repo collisions → zero collisions via isolated worktrees
Post-merge visibility: none → health probe every 15 min, auto-revert on 2 failures in 5 min window
Human involvement (Tier 1): trigger → nothing. System runs, verifies, pushes.
Human involvement (Tier 2): trigger → one approve click with full diff + proof
Code quality enforcement: ad-hoc → auto-review pipeline queues a structured code review on every completed task that touched files

AUTO-REVIEW LOOP

Completed work reviews itself

On every task completion — if files_touched is non-empty, a [REVIEW:level] task is automatically queued: review, review+security, or audit (full 6-pass P0–P5)
Findings → Follow-up — CRITICAL or HIGH findings automatically queue a [FOLLOW-UP] fix task listing every issue. The loop continues up to 3 cycles
Repeat-finding detection — same finding hash appearing after a fix attempt escalates to [HUMAN-REVIEW-REQUIRED] instead of looping. The system recognises when it can't self-resolve
Kill switch + daily cap — auto-review:enabled KV key, 20 tasks/day max, depth cap of 3. Every guard has a documented escalation path; none exit silently

Stack

Production infrastructure, not a notebook.

Backend

Cloudflare Workers Supabase Postgres + Vector REST APIs MCP Servers

Frontend

Vite 6 Workbox PWA WordPress (vertical-aware) Elementor

AI

Claude Opus/Sonnet/Haiku OpenAI GPT Gemini Workers AI Custom MCP

Deploy

Vercel Cloudflare Pages Cloudways Railway Wrangler

Proven Outcomes · Live Proof Case · GSC Verified

Feb 2025 → Feb 2026. What the autonomous engine produced.

10,100

Pages Indexed

Autonomous content engine; GSC verified.

291%

YoY Click Growth

240% YoY impressions. 110 ranked cities.

#44 → #16

Avg Search Position

Moved up 28 positions YoY.

18.5%

Revenue Lift

Closed-loop attribution: impression → revenue.

$45k

/yr Displaced Ad Spend

Across 4 SoCal counties on proof-case site.

40%

Appointments AI-Generated

Sammy conv. AI · voice/SMS/chat · 9 functions.

40%

Show Rate

Zero human engagement. Auto qualify · book · confirm.

4 → 1

Headcount Compressed

Analyst + writer + schema architect + dev → one operator.