Ralph Wiggum Multi-Model Persistence Loop

---

LIBRARY-FIRST PROTOCOL (MANDATORY)

Before writing ANY code, you MUST check:

Step 1: Library Catalog

• Location: .claude/library/catalog.json
• If match >70%: REUSE or ADAPT

Step 2: Patterns Guide

• Location: .claude/docs/inventories/LIBRARY-PATTERNS-GUIDE.md
• If pattern exists: FOLLOW documented approach

Step 3: Existing Projects

• Location: D:\Projects\*
• If found: EXTRACT and adapt

Decision Matrix

Match	Action
Library >90%	REUSE directly
Library 70-90%	ADAPT minimally
Pattern exists	FOLLOW pattern
In project	EXTRACT
No match	BUILD (add to library after)

---

Purpose

Extend the Ralph Wiggum persistence pattern with intelligent model routing:

• Gemini for research phases (search, megacontext, media)
• Codex for autonomous iteration (yolo, full-auto, sandbox)
• Claude for complex reasoning
• LLM Council for critical decisions

Unique Capability

What This Adds:

• Automatic model selection based on task phase
• Best-of-breed capabilities per iteration
• Fire-and-forget with optimal tool selection
• Multi-model consensus for critical decisions

When to Use

Perfect For:

• Complex tasks requiring multiple model strengths
• Overnight autonomous development
• Tasks mixing research + implementation + testing
• Critical decisions needing consensus
• Large codebase refactoring with validation

Don't Use When:

• Simple single-model tasks
• Time-critical (model switching adds latency)
• Need human oversight at each step

How It Works

ITERATION N:
    |
    +---> Detect Task Phase
    |         |
    |         +---> Research? --> Gemini (search/megacontext)
    |         +---> Media?    --> Gemini (imagen/veo)
    |         +---> Iterate?  --> Codex (yolo/full-auto)
    |         +---> Decide?   --> LLM Council (consensus)
    |         +---> Reason?   --> Claude (agents)
    |
    +---> Execute with Optimal Model
    |
    +---> Check Completion Promise
    |         |
    |         +---> Found? --> EXIT SUCCESS
    |         +---> Not found? --> CONTINUE
    |
    +---> ITERATION N+1 (until max)

Usage

Basic Multi-Model Loop

/ralph-multimodel "Build REST API, research best practices, implement, test, fix failures until all pass"

With Codex Sandbox Mode

CODEX_MODE=sandbox /ralph-multimodel "Experiment with auth refactoring, verify tests"

With LLM Council for Decisions

USE_COUNCIL=true /ralph-multimodel "Design authentication architecture with consensus"

Overnight Task

MAX_ITERATIONS=100 /ralph-multimodel "Complete feature X with documentation, Output <promise>DONE</promise> when finished"

Command Pattern

bash scripts/multi-model/ralph-multimodel.sh "<task>" "<loop_id>"

# Environment options:
MAX_ITERATIONS=30        # Maximum loop iterations
COMPLETION_PROMISE=DONE  # Text signaling completion
USE_COUNCIL=false        # Use LLM Council for decisions
CODEX_MODE=full-auto     # Codex mode: yolo, full-auto, sandbox

Phase Detection & Routing

Phase Detected	Keywords	Model Used
Research	"search", "latest", "documentation", "best practices"	Gemini
Megacontext	"entire codebase", "all files", "architecture overview"	Gemini --all-files
Media	"diagram", "mockup", "image", "video"	Gemini (Imagen/Veo)
Autonomous	"fix tests", "debug", "iterate", "prototype"	Codex (yolo/full-auto)
Decision	"decide", "choose", "architecture decision"	LLM Council
Reasoning	Default	Claude

Memory Integration

Results stored per iteration:

• Gemini: multi-model/gemini/yolo/ralph-{iteration}
• Codex: multi-model/codex/yolo/ralph-{iteration}
• Council: multi-model/council/decisions/ralph-{iteration}

State files:

• ~/.claude/ralph-wiggum/loop-state.json
• ~/.claude/ralph-wiggum/loop-history.log

Integration with Meta-Loop

Ralph Multi-Model connects to the recursive improvement system:

META-LOOP INTEGRATION:
    |
    +---> PROPOSE (auditors detect issues)
    |         |
    |         +---> Ralph Multi-Model for implementation
    |
    +---> TEST (frozen eval harness)
    |
    +---> COMPARE (baseline vs candidate)
    |
    +---> COMMIT (if improved)
    |
    +---> MONITOR (7-day window)
    |
    +---> ROLLBACK (if regressed)

Real-World Examples

Example 1: Full-Stack Feature

/ralph-multimodel "Build user dashboard:
1. Research React dashboard best practices (Gemini)
2. Generate UI mockup (Gemini Media)
3. Implement frontend components (Claude)
4. Build backend API (Codex yolo)
5. Write tests (Claude)
6. Fix all failing tests (Codex full-auto)
Output <promise>ALL_TESTS_PASS</promise> when done"

Example 2: Codebase Refactoring

MAX_ITERATIONS=50 CODEX_MODE=sandbox /ralph-multimodel "
Refactor auth system:
1. Analyze entire codebase architecture (Gemini megacontext)
2. Identify all auth touchpoints
3. Implement new JWT pattern in sandbox (Codex)
4. Run tests and fix failures
Output <promise>REFACTOR_COMPLETE</promise> when all tests pass"

Example 3: Architecture Decision

USE_COUNCIL=true /ralph-multimodel "
Decide database strategy:
1. Research PostgreSQL vs MongoDB for our use case (Gemini)
2. Get multi-model consensus on approach (Council)
3. Document decision
Output <promise>DECISION_MADE</promise>"

Success Indicators

• Loop completes with COMPLETION_PROMISE found
• Optimal model used for each phase
• Memory contains full iteration history
• State files show successful completion

Troubleshooting

Loop Never Completes

• Check COMPLETION_PROMISE is achievable
• Increase MAX_ITERATIONS
• Verify task includes clear completion criteria

Wrong Model Selected

• Be more explicit in task description
• Use phase keywords (see routing table)

Codex Failures

• Check Codex CLI is installed
• Verify ChatGPT Plus subscription active
• Try different CODEX_MODE

Related Resources

• Ralph Wiggum Loop: skills/orchestration/ralph-loop/SKILL.md
• Multi-Model Scripts: scripts/multi-model/
• Meta-Loop: skills/recursive-improvement/
• Memory Namespace: docs/MEMORY-NAMESPACE-SCHEMA.yaml