aiwg/agentic/code/addons/rlm/examples/rlm-divide-conquer.yaml

Deployment tool and support utility for AI context. Copies agents, skills, commands, rules, and behaviors into the paths each AI platform reads (Claude Code, Codex, Copilot, Cursor, Warp, OpenClaw, and 6 more) so one source of truth works across 10 platfo

# RLM Divide-and-Conquer Example
# Semantic chunking pipeline for large document analysis
# Based on OpenProse example 41-rlm-divide-conquer
#
# Pattern: chunker → analyzer → synthesizer pipeline
# Use when: input is too large for single-pass analysis

version: "1.0.0"

root_task:
  node_id: "task-dc00000"
  depth: 0
  prompt: "Analyze the entire codebase for API endpoint documentation completeness"
  decomposition_strategy: map-reduce
  merge_strategy: summarize
  chunking_strategy: semantic-boundary
  batch_size: 25

  children:
    # Phase 1: Chunker identifies semantic boundaries
    - node_id: "task-chunk01"
      parent_id: "task-dc00000"
      depth: 1
      prompt: |
        Scan the codebase and identify all API endpoint files.
        Group them by module/domain into 4-8 semantic chunks.
        Each chunk should contain related endpoints (auth, users, billing, etc.).
        Output a JSON manifest of chunks with file lists.
      preferred_model: haiku
      context:
        type: filtered
        source: "retrieved_documents"
        filters:
          file_patterns: ["src/**/*.ts", "src/**/*.js"]
      status: pending

    # Phase 2: Parallel analyzers process each chunk
    - node_id: "task-anlz001"
      parent_id: "task-dc00000"
      depth: 1
      prompt: |
        For each endpoint in this chunk:
        1. Extract the route, method, and handler
        2. Check if JSDoc/TSDoc exists
        3. Check if OpenAPI spec references it
        4. Rate documentation completeness (0-100)
        Output structured findings per endpoint.
      preferred_model: sonnet
      decomposition_strategy: parallel
      context:
        type: slice
        source: "parent_result"
      status: pending

    # Phase 3: Synthesizer merges chunk results
    - node_id: "task-synth01"
      parent_id: "task-dc00000"
      depth: 1
      prompt: |
        Synthesize all chunk analyses into a single documentation
        completeness report. Include:
        - Overall coverage percentage
        - Top 10 undocumented endpoints by importance
        - Module-by-module breakdown
        - Specific recommendations for each gap
      preferred_model: sonnet
      context:
        type: full
        source: "parent_result"
      quality_gate:
        min_score: 85
        scoring_criteria: "Completeness of coverage, accuracy of gap identification, actionability of recommendations"
        scorer_model: sonnet
        max_iterations: 3
        fallback: return_best
      status: pending

  status: pending

metadata:
  tree_id: "tree-divconq0"
  root_prompt: "Codebase API documentation completeness analysis"
  max_depth: 2
  total_nodes: 4
  execution_mode: logged

# Notes:
# - Chunker uses Haiku (simple file grouping task)
# - Analyzers use Sonnet (moderate code analysis)
# - Synthesizer uses Sonnet with quality gate (final deliverable)
# - semantic-boundary chunking respects module boundaries
# - batch_size: 25 means analyzer processes up to 25 endpoints per sub-agent
# - Parallel analyzers run independently per chunk