Skip to content

BLUEPRINT-ORBIT-V3.md

Latest commit

 

History

History
205 lines (156 loc) · 9.57 KB

BLUEPRINT-ORBIT-V3.md

File metadata and controls

205 lines (156 loc) · 9.57 KB
Raw

Blueprint — Orbit Agentic v3.0

Date: 2026-05-20 Version: 3.0.0 — Orbit Agentic Build type: Multi-session, multi-agent architecture upgrade

What v3.0 is

Orbit goes from a skill library to a living QA team.

Every agent now has 4 parts — like a real hire:

Part What it means Hard/Easy
Skills What this agent knows and can do Easy — already existed
Process HOW we do it at Orbit — SOPs, guardrails, order of ops Hard — this is the real work
MCP The connectors/systems where work happens Easy — we have them
Brain The intelligence layer — recalls past, learns from every approval Evergrowing — never done

The Brain is the hardest because it's like onboarding a new person. However smart they are, they need to learn YOUR products. Orbit ships a starter brain pack so every new install starts with Orbit-level intelligence, not zero.

Architecture overview

┌──────────────────────────────────────────────────────────────────────┐
│                         OPERATOR                                      │
│  Plain language: "Audit example-plugin 2.3" / "RICE score backlog"         │
└───────────────────────────┬──────────────────────────────────────────┘
                            │
                            ▼
┌──────────────────────────────────────────────────────────────────────┐
│                    CLAUDE.md ROUTER                                   │
│  Intent → Agent routing | Team/Admin tier | Approval gate protocol   │
└───────────────────────────┬──────────────────────────────────────────┘
                            │
                            ▼
┌──────────────────────────────────────────────────────────────────────┐
│                    12 AGENT TEAM                                      │
│  Each agent = Skills + Process + MCP + Brain                         │
│                                                                       │
│  01 QA Lead      02 Security     03 Performance   04 Gutenberg Dev   │
│  05 Elementor    06 Designer     07 Release Mgr   08 Compat Eng      │
│  09 Test Auto    10 PM           11 Compliance    12 SEO & Docs      │
└───────────┬───────────────────────────────────────┬──────────────────┘
            │                                       │
            ▼                                       ▼
┌────────────────────────┐          ┌───────────────────────────────────┐
│   MCP LAYER            │          │   BRAIN LAYER                     │
│   (where work happens) │          │   (brain orbit namespace)│
│                        │          │                                   │
│  brain        │          │  Team key:  READ orbit/*          │
│  GitHub (gh CLI)       │          │  Admin key: WRITE orbit/*         │
│  wp-env / wp-cli       │          │                                   │
│  Claude in Chrome      │          │  orbit/plugins/<name>             │
│  DataForSEO            │          │  orbit/patterns/approved          │
│  Apify (via brain)     │          │  orbit/patterns/revised           │
│  Slack / n8n           │          │  orbit/knowledge/*                │
│  WP sites (publish)    │          │  orbit/team/agents/*              │
└────────────────────────┘          └───────────────────────────────────┘

The 4-Part Agent Model (for every team member)

🎓 Skills

"A new hire knows email marketing."

The discrete capabilities the agent has. What it can analyse, detect, generate, or judge. These map directly to Orbit skills + cherry-picked global skills.

Rule: Skills are TOOLS. The agent invokes them. Skills are not agents.

📋 Process

"The process is: upload draft to FluentCRM, set the segment, preview, then send."

The Orbit-specific SOP. Step-by-step order of operations. This is what makes an agent feel like your team member, not a generic Claude instance.

The hard part: Process has guardrails. Every step that could go wrong has a rule:

  • Critical finding → stop and escalate (don't batch)
  • Never test production (always staging)
  • Always cite file:line (never vague findings)
  • Never re-report fixed issues (check brain)

Process = Skills + Order + Guardrails

🔌 MCP + Connectors

"After knowing the skill and process, you need the system where it runs."

Tier Tools available
Team (read-only) brain read, GitHub read, Claude in Chrome, DataForSEO
Admin (write) brain write, GitHub write, WP sites publish, Slack, n8n, Apify, FluentCRM

Each agent file specifies exactly which connectors it uses and at what tier.

🧠 Brain

"A smart new hire still needs to learn your products. Brain is that learning."

The intelligence layer. Two jobs:

  1. Recall — Before any work: 5 brain searches. What do we know about this plugin? What patterns worked? What failed? Never start from zero.
  2. Ingest — After every approval: save what's new. After every revision: save the redline. Brain compounds. By week 4, the agent rarely asks questions.

Brain is evergrowing. It's never "done". Every approve/revise is a new drawer.

Orbit starter brain: 40 pre-loaded drawers shipped with Orbit. Day-1 intelligence.

Brain — Team vs Admin key model

brain
└── orbit/              ← Orbit namespace (separated from other namespaces)
    ├── plugins/
    │   ├── example-plugin/   ← per-plugin findings history
    │   ├── tpa/
    │   └── example-plugin/
    ├── patterns/
    │   ├── approved/   ← "do this again" — approved by operator
    │   └── revised/    ← "don't do this again" — revised/rejected
    ├── knowledge/      ← evergreen WP + Orbit standards
    │   ├── wp-standards
    │   ├── block-editor
    │   ├── elementor
    │   ├── security
    │   └── performance
    └── team/
        └── agents/     ← per-agent learnings (01-qa-lead, 02-security, ...)

Team key: Can search + read orbit/*. Cannot write. Cannot write to brain. Admin key: Can write to orbit/*. Full ingest. Can add to knowledge/ base.

Always-on agents (future — Phase 2)

5 agents will run 9 AM – 6 PM IST like real employees. Which 5 is operator's choice. The framework already supports this — agent files are written for both:

  • Mode A (current): Claude Code — operator types, agent responds
  • Mode B (future): API-based runner — Autonomous runner dispatches task, agent runs headlessly, reports back to ClickUp/Slack

No change to agent files needed between modes. The 5-step WAKE/ANALYSE/PLAN/EXECUTE/INGEST is mode-agnostic.

MCP library — WordPress dev essentials

See full spec at the docs.

Must-have for any WP dev using Orbit:

  1. brain (orbit namespace) — memory
  2. GitHub via gh CLI — code, PRs, issues
  3. Claude in Chrome — visual browser testing
  4. Context7 — live WP/React/PHP docs at runtime
  5. Apify (via brain) — WP.org scraping, CVE mining
  6. DataForSEO — PageSpeed, SERP, performance data

Orbit team additionally: 7. wp-example-plugin / wp-example-plugin / wp-Orbit — WP publish 8. n8n-mcp — workflow automation 10. fluentcrm — release email drafts

v3.0 release scope

What ships in v3.0

  • CLAUDE.md — Team/Admin tier model + full routing
  • 12 agent files — 4-part structure (Skills/Process/MCP/Brain)
  • brain/orbit-brain-spec.md — Brain architecture
  • brain/starter-brain.md — 40 pre-loaded drawers
  • the docs — MCP audit for WP devs
  • the docs — Team vs Admin key setup guide
  • memory/agent-workflow-pattern.md — Updated 5-step workflow
  • routes/routes.yaml — Full skill source map
  • CHANGELOG.md v3.0 entry
  • README.md update — Agentic section

What's Phase 2 (post-v3.0)

  • Always-on 5 agents (Autonomous runner (Phase 2))
  • ClickUp task queue integration
  • Slack status mirror
  • Per-agent daily digest

Build sequence (this session)

Step 1  CLAUDE.md update (Team/Admin + 4-part framework)        ← this file done
Step 2  All 12 agent files with 4-part structure                ← agents/
Step 3  Brain spec + starter brain                              ← brain/
Step 4  MCP library + Team access docs                          ← docs/
Step 5  CHANGELOG v3.0 entry                                    ← CHANGELOG.md