ORC - Multi-Agent Orchestration System

Overview

ORC (Orchestrator) is an autonomous multi-agent orchestration system for software development. It transforms high-level goals into complete, tested implementations through a structured phased approach.

Priority Order

Quality > Capability > Developer Experience > Speed > Cost

Core Workflow

User Goal → Analyze Phase → Plan Phase → User Approval → Execute Phase → Review Phase → Learn Phase

Commands

Analysis Commands

Planning Commands

Execution Commands

Learning & Utility Commands

Command Routing

When the user invokes /orc, parse the subcommand and route accordingly:

1. analyze - Load commands/analyze.md and invoke Analyzer agent
2. plan - Load commands/plan.md and invoke Planner agent (auto-runs analyze if needed)
3. show - Load commands/show.md and display plan/epic/deviations
4. approve - Load commands/approve.md and update plan status
5. run - Load commands/run.md and invoke Orchestrator for execution
6. next - Load commands/next.md and execute single epic
7. stop - Load commands/stop.md and gracefully halt execution
8. resume - Load commands/resume.md and continue from checkpoint
9. retry - Load commands/retry.md and retry blocked story
10. patterns - Load commands/patterns.md and display patterns
11. learn - Load commands/learn.md and invoke Learner agent
12. status - Load commands/status.md and show state
13. clear - Load commands/clear.md and reset state

Runtime Directory

ORC maintains state in .orc/ directory:

.orc/
├── plan/
│   ├── state.json           # Current execution state
│   ├── plan.json            # Master plan metadata
│   ├── codebase_profile.json # Pre-flight codebase analysis
│   ├── learnings.json       # Accumulated patterns
│   ├── embeddings.json      # Vector embeddings
│   ├── deviations.json      # Deviation log
│   ├── scratchpad.json      # Working memory
│   └── epics/
│       ├── E1.json          # Epic definitions
│       └── ...
└── checkpoints/
    ├── E1-created.json      # Checkpoints
    └── ...

Agent System

Core Agents (7)

Specialist Sub-Agents (12)

Spawned on-demand by core agents based on task requirements:

• Architect - System design, patterns, scalability
• Product Designer - UX flows, acceptance criteria
• DevOps Engineer - CI/CD, Docker, K8s, infrastructure
• Full Stack Dev - End-to-end implementation
• Security Engineer - Auth, encryption, OWASP
• Database Engineer - Schema design, migrations
• Frontend Specialist - React/Vue, CSS, accessibility
• Backend Specialist - REST, GraphQL, API design
• QA Engineer - Test strategy, edge cases
• Biz Analyst - Requirements, business logic
• Content Strategist - SEO, messaging, copy
• Design Researcher - UI reference research (frontend stories)

Plan Hierarchy

Plan
└── Epic (self-contained project milestone)
    └── Feature (self-contained capability)
        └── Story (atomic task with acceptance criteria)

Quality Gates

Story Completion

• All acceptance criteria verified with evidence
• All unit tests pass
• Type checking passes (tsc --noEmit)
• No lint errors
• No new security vulnerabilities
• Approach compliance verified

Feature Completion

• All stories completed or explicitly blocked
• Integration tests pass
• Feature-level acceptance criteria verified

Epic Completion

• All features completed
• E2E tests pass (if applicable)
• Final review completed
• Full security scan passes

Deviation Handling

Any deviation from suggested_approach must be:

1. Documented with reasoning
2. Flagged for user review (if security-related)
3. Tracked in deviation history
4. Factored into confidence scoring

Auto-require review for:

• Security-related deviations
• Architecture changes
• High impact deviations

Auto-Fix Loop

On story failure:

1. Retry with previous failure context (up to 3 attempts)
2. Try alternative patterns if available
3. If still failing, mark as BLOCKED
4. Continue with non-dependent stories
5. Notify user of blocked items

Parallel Execution

Independent stories execute in parallel:

• Dependencies analyzed at plan time
• No conflicts possible (enforced by Planner)
• Parallel stories run concurrently