JubaKitiashvili/orchestrate

/orchestrate — Multi-Agent Pipeline Execution

You are executing the /orchestrate command. Use the pipeline-orchestrator agent to coordinate a full pipeline across multiple ERNE agents.

Usage

/orchestrate "build user profile screen"
/orchestrate "add pull-to-refresh to the feed"
/orchestrate "implement dark mode toggle"

Pipeline Definition

The standard pipeline runs 5 phases with the following agent sequence:

Phase 1: Plan (Sequential)

Agent: architect

• Decompose the task into an architecture plan
• Define file structure, component hierarchy, data flow, and navigation
• Output: Architecture document with numbered implementation tasks
• Gate: Architecture plan produced and confirmed

Phase 2: Implement (Parallel)

Agents: senior-developer + feature-builder (running simultaneously)

• senior-developer: Implements screens, hooks, API integration, state management
• feature-builder: Implements supporting components, utilities, and wiring
• Work is split by the architect's task breakdown — non-overlapping file ownership
• Output: Complete implementation across all specified files
• Gate: All implementation tasks marked complete, code compiles

Phase 3: Test (Sequential)

Agent: tdd-guide

• Write unit tests for new hooks and utilities
• Write component tests for new screens
• Run the test suite and ensure all tests pass
• Output: Test files and passing test results
• Gate: All tests written and passing

Phase 4: Review (Sequential)

Agent: code-reviewer

• Review all new and modified files
• Check for re-render issues, anti-patterns, platform parity, accessibility, security
• Verdict: PASS, FAIL, or RETRY
  • PASS: Proceed to Phase 5
  • FAIL: Send QA FAIL handoff back to Phase 2 agents with required fixes (up to 3 attempts)
  • RETRY: Minor issues that can be auto-fixed, then re-review
• Gate: PASS verdict from code-reviewer

Phase 5: Validate (Sequential)

Agent: performance-profiler

• Measure rendering performance, bundle size impact, and memory usage
• Verify all metrics are within project targets
• Output: Performance report with metrics table
• Gate: All performance metrics within acceptable ranges

Execution Strategy

User: /orchestrate "build user profile screen"

pipeline-orchestrator:
  ├── Phase 1: architect (plan)
  │     └── Output: Architecture document
  ├── Phase 2: senior-developer + feature-builder (parallel)
  │     └── Output: Implemented feature code
  ├── Phase 3: tdd-guide (test)
  │     └── Output: Tests written and passing
  ├── Phase 4: code-reviewer (review)
  │     ├── PASS → proceed
  │     └── FAIL → retry Phase 2 (max 3 attempts)
  └── Phase 5: performance-profiler (validate)
        └── Output: Performance report

Retry & Escalation

• Failed phases retry up to 3 times with accumulated context from prior failures
• If a phase fails 3 times, the pipeline-orchestrator escalates to the user with:
  • What was attempted
  • Why it failed each time
  • Recommended manual intervention
• Review FAIL loops back to implementation, not to planning (unless the architect's plan is identified as the root cause)

Output Format

The pipeline-orchestrator provides a live progress table and a final summary:

## Pipeline: [Task Description]

| Phase | Agent(s) | Status | Duration | Attempts |
|-------|----------|--------|----------|----------|
| Plan | architect | completed | 42s | 1/3 |
| Implement | senior-developer, feature-builder | completed | 3m 15s | 1/3 |
| Test | tdd-guide | completed | 1m 30s | 1/3 |
| Review | code-reviewer | completed (PASS) | 55s | 1/3 |
| Validate | performance-profiler | completed | 40s | 1/3 |

### Total Duration: 6m 42s
### Files Created/Modified: [list]
### Final Status: SUCCESS

Notes

• The pipeline-orchestrator does not write code — it coordinates agents
• Handoffs between agents follow the templates in docs/handoff-templates.md
• Full pipeline documentation is in docs/pipeline.md
• If the task is too small for a full pipeline (e.g., a one-file fix), consider using /code or /tdd instead