MapReduce Skill

Skill ID: mapreduce Purpose: Fan-out tasks to multiple providers/agents, then consolidate results Category: Orchestration

Overview

The MapReduce skill enables parallel task execution across multiple AI providers or agent instances, followed by intelligent consolidation of results. This produces higher-quality outputs by leveraging diverse model strengths and cross-validating findings.

Architecture

┌─────────────────────────────────────────────────────────────────────────┐
│                      MAIN THREAD (Orchestrator)                          │
│                                                                         │
│  ┌─────────────────────────────────────────────────────────────────┐    │
│  │  PHASE 1: MAP (Parallel Fan-Out)                                │    │
│  │                                                                 │    │
│  │  Task(worker-1) ──→ output-1.md                                │    │
│  │  Task(worker-2) ──→ output-2.md                                │    │
│  │  Task(worker-3) ──→ output-3.md                                │    │
│  │  bash(codex)   ──→ output-codex.md                             │    │
│  │  bash(gemini)  ──→ output-gemini.md                            │    │
│  └─────────────────────────────────────────────────────────────────┘    │
│                              │                                          │
│                              ▼                                          │
│  ┌─────────────────────────────────────────────────────────────────┐    │
│  │  PHASE 2: COLLECT (Timeout-Based)                               │    │
│  │                                                                 │    │
│  │  TaskOutput(worker-1, timeout=120s)                            │    │
│  │  TaskOutput(worker-2, timeout=120s)                            │    │
│  │  TaskOutput(worker-3, timeout=120s)                            │    │
│  │  Verify: output-codex.md, output-gemini.md exist               │    │
│  └─────────────────────────────────────────────────────────────────┘    │
│                              │                                          │
│                              ▼                                          │
│  ┌─────────────────────────────────────────────────────────────────┐    │
│  │  PHASE 3: REDUCE (Consolidation)                                │    │
│  │                                                                 │    │
│  │  Task(reducer) ──→ reads all outputs ──→ consolidated.md       │    │
│  └─────────────────────────────────────────────────────────────────┘    │
└─────────────────────────────────────────────────────────────────────────┘

Key Constraint

Subagents cannot spawn other subagents. All orchestration happens in the main thread. Workers and reducers are subagents that operate on files.

Use Cases

1. Parallel Planning

Fan out planning task to multiple providers with different strategic biases:

Workers:
  - planner-conservative: Low-risk, proven patterns
  - planner-aggressive: Fast-track, modern patterns
  - planner-security: Security-first approach

Reducer: plan-reducer
Output: specs/ROADMAP.md

See: cookbook/parallel-planning.md

2. Multi-Implementation

Generate the same feature with multiple models, pick best:

Workers:
  - impl-claude: Claude's implementation
  - impl-codex: OpenAI's implementation
  - impl-gemini: Gemini's implementation

Reducer: code-reducer
Output: src/feature/implementation.ts

See: cookbook/multi-impl.md

3. Debug Consensus

Get multiple diagnoses of a bug, verify and select best fix:

Workers:
  - debug-claude: Claude's diagnosis
  - debug-codex: Codex's diagnosis
  - debug-gemini: Gemini's diagnosis

Reducer: debug-reducer
Output: Applied fix + documentation

See: cookbook/debug-consensus.md

Available Reducers

Provider Integration

Claude Subagents (via Task tool)

Task(subagent_type="Plan", prompt="...", run_in_background=true)

External CLI Providers (via spawn skill)

# Codex
codex -m gpt-5.1-codex -a full-auto "${PROMPT}" > output.md

# Gemini
gemini -m gemini-3-pro "${PROMPT}" > output.md

# Cursor
cursor-agent --mode print "${PROMPT}" > output.md

# OpenCode
opencode --provider anthropic "${PROMPT}" > output.md

See: skills/spawn/agent/cookbook/ for detailed CLI patterns.

File Conventions

All MapReduce operations follow standard file conventions:

See: reference/file-conventions.md

Scoring Rubrics

Each reducer uses a specific scoring rubric:

• Plans: Completeness, Feasibility, Risk, Clarity, Innovation
• Code: Correctness, Readability, Maintainability, Performance, Security
• Debug: Correctness, Minimality, Safety, Clarity, Root Cause

See: reference/scoring-rubrics.md

Commands

Example: Full MapReduce

# In main thread:

## Step 1: MAP

Launch planners in a single message (enables parallelism):

Task(subagent_type="Plan", prompt="""
  Create implementation plan for: User Authentication
  Write to: specs/plans/planner-conservative.md
  Strategy: Conservative - proven patterns, minimal risk
""", run_in_background=true)

Task(subagent_type="Plan", prompt="""
  Create implementation plan for: User Authentication
  Write to: specs/plans/planner-aggressive.md
  Strategy: Aggressive - fast, modern patterns
""", run_in_background=true)

Bash("codex -m gpt-5.1-codex -a full-auto 'Create auth plan' > specs/plans/planner-codex.md")

## Step 2: COLLECT

TaskOutput(task_id=conservative-id, block=true, timeout=120000)
TaskOutput(task_id=aggressive-id, block=true, timeout=120000)

# Verify codex output exists
Read("specs/plans/planner-codex.md")

## Step 3: REDUCE

Task(subagent_type="ai-dev-kit:orchestration:plan-reducer", prompt="""
  Consolidate plans in specs/plans/*.md
  Output: specs/ROADMAP.md
  Priority: Security over speed
""")

Cookbook

• parallel-planning.md: Multi-provider planning workflows
• multi-impl.md: Code generation with selection
• debug-consensus.md: Multi-diagnosis bug fixing

Reference

• scoring-rubrics.md: Detailed scoring criteria
• file-conventions.md: Output file standards

Related Skills

• spawn: Provider-specific CLI invocation patterns
• multi-agent-orchestration: General multi-agent patterns
• research: Parallel research with synthesis