aiwg/agentic/code/addons/rlm/examples/rlm-self-refine.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 Self-Refine Example
# Quality-gated iterative refinement loop
# Based on OpenProse example 40-rlm-self-refine
#
# Pattern: evaluator + refiner loop that iterates until quality threshold met
# Use when: output quality matters more than speed (security analysis, research synthesis)

version: "1.0.0"

root_task:
  node_id: "task-refine00"
  depth: 0
  prompt: "Produce a comprehensive security analysis of the authentication module"
  decomposition_strategy: sequential
  merge_strategy: best-of-n

  # Quality gate enforces iterative refinement
  quality_gate:
    min_score: 85
    scoring_criteria: |
      Completeness: all auth flows covered (login, register, reset, token refresh).
      Accuracy: vulnerabilities correctly identified with OWASP references.
      Actionability: each finding has a concrete remediation step.
      Clarity: non-security engineers can understand the report.
    scorer_model: sonnet
    max_iterations: 5
    fallback: return_best

  children:
    - node_id: "task-draft001"
      parent_id: "task-refine00"
      depth: 1
      prompt: "Draft initial security analysis covering all authentication flows"
      preferred_model: sonnet
      context:
        type: filtered
        source: "file:src/auth/**/*.ts"
        filters:
          file_patterns: ["*.ts"]
      status: pending

    - node_id: "task-eval0001"
      parent_id: "task-refine00"
      depth: 1
      prompt: |
        Evaluate the security analysis against these criteria:
        - Are all auth flows covered?
        - Are OWASP references accurate?
        - Does each finding have a remediation step?
        - Is the language clear for non-specialists?
        Score 0-100 and provide specific improvement feedback.
      preferred_model: sonnet
      context:
        type: summary
        source: "parent_result"
      status: pending

    - node_id: "task-refn0001"
      parent_id: "task-refine00"
      depth: 1
      prompt: |
        Refine the security analysis based on evaluator feedback.
        Address each specific issue raised. Do not remove existing
        correct content — only improve weak areas.
      preferred_model: sonnet
      context:
        type: full
        source: "parent_result"
      status: pending

  status: pending

metadata:
  tree_id: "tree-selfref0"
  root_prompt: "Security analysis with quality-gated refinement"
  max_depth: 1
  total_nodes: 4
  execution_mode: logged

# Notes:
# - The quality_gate on the root node drives the loop: evaluate → refine → re-evaluate
# - Scorer uses Sonnet (cost-efficient for rubric evaluation)
# - Drafter uses Sonnet (good enough for initial pass)
# - After 5 iterations without reaching 85, returns best attempt
# - Iteration history is preserved for analysis