Ralph Mode - Autonomous Development Loops

Ralph Mode - Autonomous Development Loops

Ralph Mode implements the Ralph Wiggum technique adapted for OpenClaw: autonomous task completion through continuous iteration with backpressure gates, completion criteria, and structured planning.

When to Use

Use Ralph Mode when: Building features that require multiple iterations and refinement Working on complex projects with acceptance criteria to validate Need automated testing, linting, or typecheck gates Want to track progress across many iterations systematically Prefer autonomous loops over manual turn-by-turn guidance

Three-Phase Workflow

Phase 1: Requirements Definition Document specs in specs/ (one file per topic of concern) Define acceptance criteria (observable, verifiable outcomes) Create implementation plan with prioritized tasks Phase 2: Planning Gap analysis: compare specs against existing code Generate IMPLEMENTATION_PLAN.md with prioritized tasks No implementation during this phase Phase 3: Building (Iterative) Pick one task from plan per iteration Implement, validate, update plan, commit Continue until all tasks complete or criteria met

Backpressure Gates

Reject incomplete work automatically through validation: Programmatic Gates (Always use these): Tests: [test command] - Must pass before committing Typecheck: [typecheck command] - Catch type errors early Lint: [lint command] - Enforce code quality Build: [build command] - Verify integration Subjective Gates (Use for UX, design, quality): LLM-as-judge reviews for tone, aesthetics, usability Binary pass/fail - converges through iteration Only add after programmatic gates work reliably

Context Efficiency

One task per iteration = fresh context each time Spawn sub-agents for exploration, not main context Lean prompts = smart zone (~40-60% utilization) Plans are disposable - regenerate cheap vs. salvage

File Structure

Create this structure for each Ralph Mode project: project-root/ ├── IMPLEMENTATION_PLAN.md # Shared state, updated each iteration ├── AGENTS.md # Build/test/lint commands (~60 lines) ├── specs/ # Requirements (one file per topic) │ ├── topic-a.md │ └── topic-b.md ├── src/ # Application code └── src/lib/ # Shared utilities

IMPLEMENTATION_PLAN.md

• Priority task list - single source of truth. Format:
• # Implementation Plan
• ## In Progress
• [ ] Task name (iteration N)
• - Notes: discoveries, bugs, blockers
• ## Completed
• [x] Task name (iteration N)
• ## Backlog
• [ ] Future task

Topic Scope Test

Can you describe the topic in one sentence without "and"? ✅ "User authentication with JWT and session management" ❌ "Auth, profiles, and billing" → 3 topics

AGENTS.md - Operational Guide

• Succinct guide for running the project. Keep under 60 lines:
• # Project Operations
• ## Build Commands
• npm run dev # Development server
• npm run build # Production build
• ## Validation
• npm run test # All tests
• npm run lint # ESLint
• npm run typecheck # TypeScript
• npm run e2e # E2E tests
• ## Operational Notes
• Tests must pass before committing
• Typecheck failures block commits
• Use existing utilities from src/lib over ad-hoc copies

Hats (Personas)

Specialized roles for different tasks: Hat: Architect (@architect) High-level design, data modeling, API contracts Focus: patterns, scalability, maintainability Hat: Implementer (@implementer) Write code, implement features, fix bugs Focus: correctness, performance, test coverage Hat: Tester (@tester) Test authoring, validation, edge cases Focus: coverage, reliability, reproducibility Hat: Reviewer (@reviewer) Code reviews, PR feedback, quality assessment Focus: style, readability, adherence to specs Usage: "Spawn a sub-agent with @architect hat to design the data model"

Outer Loop (You coordinate)

Your job as main agent: engineer setup, observe, course-correct. Don't allocate work to main context - Spawn sub-agents Let Ralph Ralph - LLM will self-identify, self-correct Use protection - Sandbox is your security boundary Plan is disposable - Regenerate when wrong/stale Move outside the loop - Sit and watch, don't micromanage

Inner Loop (Sub-agent executes)

Each sub-agent iteration: Study - Read plan, specs, relevant code Select - Pick most important uncompleted task Implement - Write code, one task only Validate - Run tests, lint, typecheck (backpressure) Update - Mark task done, note discoveries, commit Exit - Next iteration starts fresh

Stopping Conditions

Loop ends when: ✅ All IMPLEMENTATION_PLAN.md tasks completed ✅ All acceptance criteria met ✅ Tests passing, no blocking issues ⚠️ Max iterations reached (configure limit) 🛑 Manual stop (Ctrl+C)

Completion Criteria

Define success upfront - avoid "seems done" ambiguity.

Programmatic (Measurable)

All tests pass: [test_command] returns 0 Typecheck passes: No TypeScript errors Build succeeds: Production bundle created Coverage threshold: e.g., 80%+

Subjective (LLM-as-Judge)

For quality criteria that resist automation: ## Completion Check - UX Quality Criteria: Navigation is intuitive, primary actions are discoverable Test: User can complete core flow without confusion ## Completion Check - Design Quality Criteria: Visual hierarchy is clear, brand consistency maintained Test: Layout follows established patterns Run LLM-as-judge sub-agent for binary pass/fail.

Next.js Full Stack

specs/ ├── authentication.md ├── database.md └── api-routes.md src/ ├── app/ # App Router ├── components/ # React components ├── lib/ # Utilities (db, auth, helpers) └── types/ # TypeScript types AGENTS.md: Build: npm run dev Test: npm run test Typecheck: npx tsc --noEmit Lint: npm run lint

Python (Scripts/Notebooks/FastAPI)

specs/ ├── data-pipeline.md ├── model-training.md └── api-endpoints.md src/ ├── pipeline.py ├── models/ ├── api/ └── tests/ AGENTS.md: Build: python -m src.main Test: pytest Typecheck: mypy src/ Lint: ruff check src/

GPU Workloads

specs/ ├── model-architecture.md ├── training-data.md └── inference-pipeline.md src/ ├── models/ ├── training/ ├── inference/ └── utils/ AGENTS.md: Train: python train.py Test: pytest tests/ Lint: ruff check src/ GPU Check: nvidia-smi

Quick Start Command

Start a Ralph Mode session: "Start Ralph Mode for my project at ~/projects/my-app. I want to implement user authentication with JWT. I will: Create IMPLEMENTATION_PLAN.md with prioritized tasks Spawn sub-agents for iterative implementation Apply backpressure gates (test, lint, typecheck) Track progress and announce completion

Operational Learnings

• When Ralph patterns emerge, update AGENTS.md:
• ## Discovered Patterns
• When adding API routes, also add to OpenAPI spec
• Use existing db utilities from src/lib/db over direct calls
• Test files must be co-located with implementation

Escape Hatches

When trajectory goes wrong: Ctrl+C - Stop loop immediately Regenerate plan - "Discard IMPLEMENTATION_PLAN.md and re-plan" Reset - "Git reset to last known good state" Scope down - Create smaller scoped plan for specific work

Advanced: LLM-as-Judge Fixture

For subjective criteria (tone, aesthetics, UX): Create src/lib/llm-review.ts: interface ReviewResult { pass: boolean; feedback?: string; } async function createReview(config: { criteria: string; artifact: string; // text or screenshot path }): Promise<ReviewResult>; Sub-agents discover and use this pattern for binary pass/fail checks.

Critical Operational Requirements

Based on empirical usage, enforce these practices to avoid silent failures:

1. Mandatory Progress Logging

• Ralph MUST write to PROGRESS.md after EVERY iteration. This is non-negotiable.
• Create PROGRESS.md in project root at start:
• # Ralph: [Task Name]
• ## Iteration [N] - [Timestamp]
• ### Status
• [ ] In Progress | [ ] Blocked | [ ] Complete
• ### What Was Done
• [Item 1]
• [Item 2]
• ### Blockers
• None | [Description]
• ### Next Step
• [Specific next task from IMPLEMENTATION_PLAN.md]
• ### Files Changed
• `path/to/file.ts` - [brief description]
• Why: External observers (parent agents, crons, humans) can tail one file instead of scanning directories or inferring state from session logs.

2. Session Isolation & Cleanup

Before spawning a new Ralph session: Check for existing Ralph sub-agents via sessions_list Kill or verify completion of previous sessions Do NOT spawn overlapping Ralph sessions on same codebase Anti-pattern: Spawning Ralph v2 while v1 is still running = file conflicts, race conditions, lost work.

3. Explicit Path Verification

Never assume directory structure. At start of each iteration: // Verify current working directory const cwd = process.cwd(); console.log(`Working in: ${cwd}`); // Verify expected paths exist if (!fs.existsSync('./src/app')) { console.error('Expected ./src/app, found:', fs.readdirSync('.')); // Adapt or fail explicitly } Why: Ralph may be spawned from different contexts with different working directories.

4. Completion Signal Protocol

• When done, Ralph MUST:
• Write final PROGRESS.md with "## Status: COMPLETE"
• List all created/modified files
• Exit cleanly (no hanging processes)
• Example completion PROGRESS.md:
• # Ralph: Influencer Detail Page
• ## Status: COMPLETE ✅
• **Finished:** [ISO timestamp]
• ### Final Verification
• [x] TypeScript: Pass
• [x] Tests: Pass
• [x] Build: Pass
• ### Files Created
• `src/app/feature/page.tsx`
• `src/app/api/feature/route.ts`
• ### Testing Instructions
• 1. Run: `npm run dev`
• 2. Visit: `http://localhost:3000/feature`
• 3. Verify: [specific checks]

5. Error Handling Requirements

If Ralph encounters unrecoverable errors: Log to PROGRESS.md with "## Status: BLOCKED" Describe blocker in detail List attempted solutions Exit cleanly (don't hang) Do not silently fail. A Ralph that stops iterating with no progress log is indistinguishable from one still working.

6. Iteration Time Limits

• Set explicit iteration timeouts:
• ## Operational Parameters
• Max iteration time: 10 minutes
• Total session timeout: 60 minutes
• If iteration exceeds limit: Log blocker, exit
• Why: Prevents infinite loops on stuck tasks, allows parent agent to intervene.

Memory Updates

• After each Ralph Mode session, document:
• ## [Date] Ralph Mode Session
• **Project:** [project-name]
• **Duration:** [iterations]
• **Outcome:** success / partial / blocked
• **Learnings:**
• What worked well
• What needs adjustment
• Patterns to add to AGENTS.md

Appendix: Hall of Failures

Common anti-patterns observed: Anti-PatternConsequencePreventionNo progress loggingParent agent cannot determine statusMandatory PROGRESS.mdSilent failureWork lost, time wastedExplicit error loggingOverlapping sessionsFile conflicts, corrupt stateCheck/cleanup before spawnPath assumptionsWrong directory, wrong filesExplicit verificationNo completion signalParent waits indefinitelyClear COMPLETE statusInfinite iterationResource waste, no progressTime limits + blockersComplex initial promptsSub-agent never starts (empty session logs)SIMPLIFY instructions

Problem: Sub-agents spawn but don't execute

Evidence: Empty session logs (2 bytes), no tool calls, 0 tokens used

Root Causes

Instructions too complex - Overwhelms isolated session initialization No clear execution trigger - Agent doesn't know to start Branching logic - "If X do Y, if Z do W" confuses task selection Multiple files mentioned - Can't decide which to start with

Fix: SIMPLIFIED Ralph Task Template

## Task: [ONE specific thing] **File:** exact/path/to/file.ts **What:** Exact description of change **Validate:** Exact command to run **Then:** Update PROGRESS.md and exit ## Rules 1. Do NOT look at other files 2. Do NOT "check first" 3. Make the change, validate, exit

BEFORE (Bad - causes stalls):

• Fix all TypeScript errors across these files:
• lib/db.ts has 2 errors
• lib/proposal-service.ts has 5 errors
• route.ts has errors
• Check which ones to fix first, then...

AFTER (Good - executes):

Fix lib/db.ts line 27: Change: PoolClient to pg.PoolClient Validate: npm run typecheck Exit immediately after

CRITICAL: Single File Rule

Each Ralph iteration gets ONE file. Not "all errors", not "check then decide". ONE file, ONE change, validate, exit.

CRITICAL: Update PROGRESS.md

• MANDATORY: After EVERY iteration, update PROGRESS.md with:
• ## Iteration [N] - [Timestamp]
• ### Status: Complete ✅ | Blocked ⛔ | Failed ❌
• ### What Was Done
• [Specific changes made]
• ### Validation
• [Test/lint/typecheck results]
• ### Next Step
• [What should happen next]
• Why this matters: Cron job reads PROGRESS.md for status updates. If not updated, status appears stale/repetitive.

Debugging Ralph Stalls

If Ralph stalls: Check session logs (should show tool calls within 60s) If empty after spawn → instructions too complex Reduce: ONE file, ONE line number, ONE change Shorter timeout forces smaller tasks (300s not 600s)

Fixing Stale Status Reports

If cron reports same status repeatedly: Check PROGRESS.md was updated by sub-agent If not updated → sub-agent skipped documentation step Update skill: Add "MANDATORY PROGRESS.md update" to prompt Manual fix: Update PROGRESS.md to reflect actual state

Summary

Ralph works when: Single file focus + explicit change + validate + exit Ralph stalls when: Complex decisions + multiple files + conditional logic