bmad-orchestrator（bmad 协调器）

bmad-orchestrator — BMAD Workflow Orchestration with SSD

When to use this skill

• Initializing BMAD in a new project (with or without SSD)
• Running structured TEA cycles within each BMAD phase
• Checking and resuming BMAD/SSD workflow status
• Routing work across Analysis, Planning, Solutioning, and Implementation
• Managing structured handoff and cross-phase traceability between phases

What is SSD (Structured System Design)?

SSD is a meta-framework that embeds TEA cycles within each BMAD phase, transforming phase execution from "produce a document" into a structured loop:

T → Task:     Decompose the phase into concrete tasks with assigned agents
E → Execute:  Run tasks in parallel via multi-agent team execution
A → Architect: Validate outputs for coherence, completeness, and cross-phase traceability

Each BMAD phase becomes an independent TEA cycle. The architect validation step produces a PASS, PASS_WITH_WARNINGS, REVISE, or FAIL verdict. Only after PASS (or PASS_WITH_WARNINGS) does the human review gate (plannotator) open. This ensures automated structural correctness before human review.

SSD vs. Standard BMAD

Installation

npx skills add https://github.com/akillness/oh-my-skills --skill bmad-orchestrator

Notes for Codex Usage

bmad-orchestrator's default execution path is Claude Code. To run the same flow directly in Codex, we recommend operating BMAD stages via a higher-level orchestration path such as omx/ohmg.

Control Model

BMAD phase routing uses the same three-layer abstraction as JEO:

• settings: platform-specific runtime configuration such as Claude hooks, Codex/Gemini instructions, and MCP setup
• rules: phase constraints such as "do not advance before the current phase document is approved" and "do not reopen the same unchanged phase document for review"
• hooks: platform callbacks such as Claude ExitPlanMode, Codex notify, or Gemini AfterAgent

For BMAD phase gates, the intended rule is strict:

• Review the current phase document before moving forward
• If the document hash has not changed since the last terminal review result, do not relaunch plannotator
• Only a revised document resets the gate and permits another review cycle
• With SSD: plannotator gate requires architect_verdict: "PASS" or "PASS_WITH_WARNINGS" in ssd-state.json before opening

Platform Support Status

Standard BMAD Commands

/workflow-init [--ssd]
/workflow-status

Typical flow:

1. Run /workflow-init to bootstrap BMAD config (add --ssd to also initialize SSD state).
2. Move through phases: Analysis → Planning → Solutioning → Implementation.
3. Run /workflow-status any time to inspect current phase and progress.

SSD Commands (TEA Integration)

Use these commands to run structured TEA cycles within each BMAD phase:

SSD Full Flow

/workflow-init --ssd
       |
       v
  ┌─────────────────────────────────────────────────────┐
  │              PHASE N TEA CYCLE                      │
  │                                                     │
  │  /ssd-decompose  (or /ssd-cycle for full auto)      │
  │       |                                             │
  │       v                                             │
  │  [T] Task decomposition                             │
  │       planner/analyst produces task list            │
  │       → ssd-state.json phases[N].tasks updated      │
  │       |                                             │
  │       v                                             │
  │  /ssd-execute                                       │
  │       |                                             │
  │       v                                             │
  │  [E] Multi-agent execution via /team                │
  │       TeamCreate → TaskCreate per subtask           │
  │       Agents: executor, analyst, designer,          │
  │               test-engineer, security-reviewer…     │
  │       → docs/ssd/phase-N/ artifacts produced        │
  │       → Phase document assembled from artifacts     │
  │       |                                             │
  │       v                                             │
  │  /ssd-validate                                      │
  │       |                                             │
  │       v                                             │
  │  [A] Architect validation                           │
  │       fabric -p bmad_ssd_phase_review               │
  │       (fallback: architect agent via TaskCreate)    │
  │       |                                             │
  │       ├── PASS ──────────→ plannotator review       │
  │       │                         |                   │
  │       │                    ┌────┴────┐              │
  │       │                    │ Approve │ Req Changes  │
  │       │                    └────┬────┘     |        │
  │       │                         |     Loop to [E]   │
  │       │                         v     with feedback │
  │       │                    /ssd-advance              │
  │       │                    → phase N+1               │
  │       │                                             │
  │       ├── PASS_WITH_WARNINGS → plannotator (warned) │
  │       │                                             │
  │       ├── REVISE → targeted re-execute specific     │
  │       │             tasks, then re-validate         │
  │       │                                             │
  │       └── FAIL → full re-execute with changes,      │
  │                   then re-validate                  │
  │                   (max 3 cycles, then escalate)     │
  └─────────────────────────────────────────────────────┘

TEA Per-Phase Mapping

Phase 1: Analysis

Phase 2: Planning

Phase 3: Solutioning

Phase 4: Implementation

Fabric Pattern: bmad_ssd_phase_review

Install this custom pattern for automated architect validation:

mkdir -p ~/.config/fabric/patterns/bmad_ssd_phase_review
cat > ~/.config/fabric/patterns/bmad_ssd_phase_review/system.md << 'EOF'
# IDENTITY AND PURPOSE

You are an expert system architect performing a structured phase review for the BMAD Structured System Design (SSD) framework. Your job is to validate that a phase document is internally coherent, externally consistent with prior phase artifacts, and complete enough to advance to the next phase.

Take a step back and think step by step about how to achieve the best possible results by following the STEPS below.

# STEPS

1. IDENTIFY the current BMAD phase (Analysis, Planning, Solutioning, Implementation) from the input metadata.

2. PARSE the phase document and extract all claims, requirements, design decisions, and deliverables.

3. INTERNAL COHERENCE CHECK:
   - Are there contradictory statements within the document?
   - Are all sections complete (no TODOs, placeholders, or TBDs)?
   - Do quantitative claims have justification?

4. CROSS-PHASE TRACEABILITY CHECK (if prior phase artifacts are provided):
   - Phase 2 (Planning): Does every PRD requirement trace to a product brief goal?
   - Phase 3 (Solutioning): Does the architecture address every PRD functional requirement?
   - Phase 4 (Implementation): Does every story map to an architecture component?
   - Flag any orphaned items.

5. COMPLETENESS CHECK against BMAD level expectations:
   - Level 0-1: Minimal viable coverage
   - Level 2: Full requirement coverage with acceptance criteria
   - Level 3: Comprehensive with integration points and risk analysis
   - Level 4: Enterprise-grade with security, performance, and infrastructure coverage

6. RISK ASSESSMENT:
   - Identify assumptions that could invalidate the phase output
   - Flag unresolved dependencies
   - Note scope creep relative to prior phase boundaries

7. PRODUCE a structured verdict.

# OUTPUT INSTRUCTIONS

- Output valid Markdown only.
- Begin with a `## Verdict` section: one of `PASS`, `PASS_WITH_WARNINGS`, `FAIL`, or `REVISE`.
- Follow with `## Internal Coherence` (findings with line references).
- Follow with `## Cross-Phase Traceability` (coverage matrix if applicable).
- Follow with `## Completeness` (missing sections or underspecified areas).
- Follow with `## Risks` (ranked by severity: critical, high, medium, low).
- Follow with `## Required Changes` (concrete, actionable items if verdict is FAIL or REVISE).
- Do not include warnings, disclaimers, or caveats outside the structured sections.

# INPUT

INPUT:
EOF

Usage in /ssd-validate:

# Pipe current phase doc + prior artifacts for cross-phase validation
{
  echo "--- CURRENT PHASE DOCUMENT ---"
  cat "$CURRENT_DOC"
  echo "--- PRIOR PHASE ARTIFACTS ---"
  for prior in "${PRIOR_DOCS[@]}"; do
    echo "--- $(basename "$prior") ---"
    cat "$prior"
  done
} | fabric -p bmad_ssd_phase_review --stream > docs/ssd/phase-N/architect-review.md

# Phase 3 example: validate architecture against PRD
cat docs/architecture-myapp-2026-03-16.md | \
  fabric -p bmad_ssd_phase_review \
  --context "$(cat docs/prd-myapp-2026-03-16.md)" \
  --stream

If fabric is not installed, /ssd-validate falls back to an architect agent:

# In Claude Code session:
ssd validate — run architect review on the current phase document before advancing

SSD State File Schema

Location: .omc/state/ssd-state.json

{
  "version": "1.0.0",
  "project_name": "myapp",
  "project_level": 2,
  "active": true,
  "current_phase": 2,
  "created_at": "2026-03-16T10:00:00Z",
  "updated_at": "2026-03-16T14:30:00Z",
  "phases": {
    "1": {
      "name": "analysis",
      "tea_step": "complete",
      "tasks": {
        "decomposed_at": "2026-03-16T10:05:00Z",
        "items": [
          {
            "id": "p1-t1",
            "name": "Market research",
            "agent": "document-specialist",
            "model": "sonnet",
            "status": "completed",
            "output_path": "docs/ssd/phase-1/market-research.md"
          }
        ]
      },
      "architect_review": {
        "verdict": "PASS",
        "method": "fabric",
        "pattern": "bmad_ssd_phase_review",
        "report_path": "docs/ssd/phase-1/architect-review.md",
        "reviewed_at": "2026-03-16T11:00:00Z"
      },
      "plannotator_review": {
        "status": "approved",
        "reviewed_at": "2026-03-16T11:30:00Z",
        "document_hash": "abc123"
      }
    },
    "2": {
      "name": "planning",
      "tea_step": "execute",
      "tasks": {
        "items": [
          {
            "id": "p2-t1",
            "name": "Functional requirements extraction",
            "agent": "analyst",
            "model": "opus",
            "status": "in_progress"
          }
        ]
      },
      "architect_review": null,
      "plannotator_review": null
    }
  },
  "cross_phase_traceability": {
    "phase_1_to_2": { "validated": true, "coverage": 1.0, "orphaned_items": [] },
    "phase_2_to_3": { "validated": false },
    "phase_3_to_4": { "validated": false }
  }
}

tea_step values: null → decompose → execute → validate → complete

plannotator Integration (Phase Review Gate)

Each BMAD phase produces a key document (PRD, Tech Spec, Architecture). Before transitioning to the next phase, review that document with plannotator.

With SSD enabled, plannotator only opens after /ssd-validate produces architect_verdict: "PASS" or "PASS_WITH_WARNINGS". This ensures automated structural correctness before human review.

Phase Review Pattern

# After /prd → docs/prd-myapp-2026-03-16.md is created
# With SSD: run /ssd-validate first, then phase-gate-review.sh
bash scripts/phase-gate-review.sh docs/prd-myapp-2026-03-16.md "PRD Review: myapp"

# After /architecture → docs/architecture-myapp-2026-03-16.md is created
bash scripts/phase-gate-review.sh docs/architecture-myapp-2026-03-16.md "Architecture Review: myapp"

Or submit the plan directly from your AI session:

# In Claude Code after /prd completes:
planno — review the PRD before we proceed to Phase 3

Phase Gate Flow (with SSD)

/prd completes → docs/prd-myapp.md created
       ↓
 /ssd-validate runs (automated architect review)
       ↓ fabric -p bmad_ssd_phase_review
       ↓
  PASS/PASS_WITH_WARNINGS?
       ↓ yes
 phase-gate-review.sh opens plannotator UI
       ↓
  [Approve]              [Request Changes]
       ↓                        ↓
 Obsidian saved          Agent revises doc
 ssd-state updated       /ssd-execute (targeted)
                         then /ssd-validate again
       ↓
 /ssd-advance → Phase 3

Quick Reference

Integration with /team (OMC)

/ssd-execute maps directly to OMC's Team staged pipeline. When you run /ssd-execute:

1. Reads ssd-state.json for the current phase's task list
2. Calls TeamCreate with team name ssd-phase-{N}-{project}
3. Creates tasks via TaskCreate with agent routing from the TEA per-phase table
4. Monitors via TaskList until all tasks reach terminal state
5. Writes outputs to docs/ssd/phase-{N}/
6. Updates ssd-state.json with execution results

# Trigger SSD execution for current phase (Claude Code)
ssd execute — run phase 2 planning tasks with multi-agent team

Or use directly with Team:

/team 4:analyst "Execute BMAD Phase 2 planning: functional requirements, non-functional requirements, user stories, acceptance criteria based on product brief at docs/product-brief-myapp.md"

Quick Start with SSD

# 1. Initialize BMAD with SSD enabled
/workflow-init --ssd

# 2. Run full TEA cycle for Phase 1 (Analysis)
/ssd-cycle

# 3. Check status after cycle
/ssd-status

# 4. When architect validates PASS, human review opens automatically
# (plannotator UI launches)

# 5. After approval, advance to Phase 2
/ssd-advance

# 6. Repeat for each phase
/ssd-cycle  # Phase 2: Planning
/ssd-cycle  # Phase 3: Solutioning
/ssd-cycle  # Phase 4: Implementation

Obsidian Save Format

Approved phase documents are saved to your Obsidian vault:

---
created: 2026-03-16T10:00:00Z
source: plannotator
tags: [bmad, ssd, phase-2, prd, myapp]
ssd_architect_verdict: PASS
ssd_tea_cycle: 1
---

[[BMAD Plans]]

# PRD: myapp
...

TOON Format Hook

If ~/.claude/hooks/toon-inject.mjs is installed, the skill catalog is automatically injected into every prompt. See [bmad-orchestrator SKILL.md — TOON Format Integration] for details.