monitored-ralph-loop（受监控的拉尔夫循环）

Ralph Loop (Event-Driven)

Enhanced Ralph pattern with event-driven notifications — Codex/Claude calls OpenClaw when it needs attention instead of polling.

Key Concepts

Clean Sessions

Each iteration spawns a fresh agent session with clean context. This is intentional:

• Avoids context window limits
• Each codex exec is a new process with no memory of previous runs
• Memory persists via files: IMPLEMENTATION_PLAN.md, AGENTS.md, git history

File-Based Notification Fallback

If OpenClaw is rate-limited when Codex sends a wake notification:

1. The notification is written to .ralph/pending-notification.txt
2. Wake is attempted (may fail)
3. When OpenClaw recovers, it checks for pending notifications
4. Work is never lost — it's all in git/files

File Structure

project/
├── PROMPT.md                      # Loaded each iteration (mode-specific)
├── AGENTS.md                      # Project context, test commands, learnings
├── IMPLEMENTATION_PLAN.md         # Task list with status
├── specs/                         # Requirements specs
│   ├── overview.md
│   └── <feature>.md
└── .ralph/
    ├── ralph.log                  # Execution log
    ├── pending-notification.txt   # Current pending notification (if any)
    └── last-notification.txt      # Previous notification (for reference)

Notification Format

.ralph/pending-notification.txt:

{
  "timestamp": "2026-02-07T02:30:00+01:00",
  "project": "/home/user/my-project",
  "message": "DONE: All tasks complete.",
  "iteration": 15,
  "max_iterations": 20,
  "cli": "codex",
  "status": "pending"
}

Status values:

• pending — Wake failed or not attempted
• delivered — Wake succeeded

OpenClaw Recovery Procedure

When coming back online after rate limit or downtime, check for pending notifications:

# Find all pending notifications across projects
find ~/projects -name "pending-notification.txt" -path "*/.ralph/*" 2>/dev/null

# Or check a specific project
cat /path/to/project/.ralph/pending-notification.txt

Recovery Actions by Message Prefix

Injecting Responses

To answer a decision/question for the next iteration:

echo "## Human Decisions
- [$(date '+%Y-%m-%d %H:%M')] Q: <question>? A: <answer>" >> AGENTS.md

The next Codex session will read AGENTS.md and see the answer.

Clearing Notifications

After processing a notification, clear it:

mv .ralph/pending-notification.txt .ralph/last-notification.txt

Workflow

1. Collect Requirements

Ask for (if not provided):

• Goal/JTBD: What outcome is needed?
• CLI: codex, claude, opencode, goose
• Mode: PLANNING, BUILDING, or BOTH
• Tech stack: Language, framework, database
• Test command: How to verify correctness
• Max iterations: Default 20

2. Generate Specs

Break the goal into topics of concern → specs/*.md:

# specs/overview.md
## Goal
<one-sentence JTBD>

## Tech Stack
- Language: Python 3.11
- Framework: FastAPI
- Database: SQLite
- Frontend: HTMX + Tailwind

## Success Criteria
- [ ] Criterion 1
- [ ] Criterion 2

3. Generate AGENTS.md

# AGENTS.md

## Project
<brief description>

## Commands
- **Install**: `pip install -e .`
- **Test**: `pytest`
- **Lint**: `ruff check .`
- **Run**: `python -m app`

## Backpressure
Run after each implementation:
1. `ruff check . --fix`
2. `pytest`

## Human Decisions
<!-- Decisions made by humans are recorded here -->

## Learnings
<!-- Agent appends operational notes here -->

4. Generate PROMPT.md (Mode-Specific)

PLANNING Mode

# Ralph PLANNING Loop

## Goal
<JTBD>

## Context
- Read: specs/*.md
- Read: Current codebase structure
- Update: IMPLEMENTATION_PLAN.md

## Rules
1. Do NOT implement code
2. Do NOT commit
3. Analyze gaps between specs and current state
4. Create/update IMPLEMENTATION_PLAN.md with prioritized tasks
5. Each task should be small (< 1 hour of work)
6. If requirements are unclear, list questions

## Notifications
When you need input or finish planning:

openclaw gateway wake --text "PLANNING: <your message>" --mode now

Use prefixes:
- `DECISION:` — Need human input on a choice
- `QUESTION:` — Requirements unclear
- `DONE:` — Planning complete

## Completion
When plan is complete and ready for building, add to IMPLEMENTATION_PLAN.md:

STATUS: PLANNING_COMPLETE

Then notify:

openclaw gateway wake --text "DONE: Planning complete. X tasks identified." --mode now

BUILDING Mode

# Ralph BUILDING Loop

## Goal
<JTBD>

## Context
- Read: specs/*.md, IMPLEMENTATION_PLAN.md, AGENTS.md
- Implement: One task per iteration
- Test: Run backpressure commands from AGENTS.md

## Rules
1. Pick the highest priority incomplete task from IMPLEMENTATION_PLAN.md
2. Investigate relevant code before changing
3. Implement the task
4. Run backpressure commands (lint, test)
5. If tests pass: commit with clear message, mark task done
6. If tests fail: try to fix (max 3 attempts), then notify
7. Update AGENTS.md with any operational learnings
8. Update IMPLEMENTATION_PLAN.md with progress

## Notifications
Call OpenClaw when needed:

openclaw gateway wake --text "<PREFIX>: <message>" --mode now

Prefixes:
- `DECISION:` — Need human input (e.g., "SQLite vs PostgreSQL?")
- `ERROR:` — Tests failing after 3 attempts
- `BLOCKED:` — Missing dependency, credentials, or unclear spec
- `PROGRESS:` — Major milestone complete (optional)
- `DONE:` — All tasks complete

## Completion
When all tasks are done:
1. Add to IMPLEMENTATION_PLAN.md: `STATUS: COMPLETE`
2. Notify:

openclaw gateway wake --text "DONE: All tasks complete. Summary: <what was built>" --mode now

5. Run the Loop

Use the provided scripts/ralph.sh:

# Default: 20 iterations with Codex
./scripts/ralph.sh 20

# With Claude Code
RALPH_CLI=claude ./scripts/ralph.sh 10

# With tests
RALPH_TEST="pytest" ./scripts/ralph.sh

Parallel Execution

For independent tasks, use git worktrees:

# Create worktrees for parallel work
git worktree add /tmp/task-auth main
git worktree add /tmp/task-upload main

# Spawn parallel sessions (each is clean/fresh)
exec pty:true background:true workdir:/tmp/task-auth command:"codex exec --full-auto 'Implement user authentication...'"
exec pty:true background:true workdir:/tmp/task-upload command:"codex exec --full-auto 'Implement image upload...'"

Track sessions:

Each Codex notifies independently. Check .ralph/pending-notification.txt in each worktree.

CLI-Specific Notes

Codex

• Requires git repository
• Each codex exec is a fresh session — no memory between calls
• --full-auto: Auto-approve in workspace (sandboxed)
• --yolo: No sandbox, no approvals (dangerous but fast)
• Default model: gpt-5.2-codex

Claude Code

• --dangerously-skip-permissions: Auto-approve (use in sandbox)
• No git requirement
• Each invocation is fresh

OpenCode

• opencode run "$(cat PROMPT.md)"

Goose

• goose run "$(cat PROMPT.md)"

Safety

⚠️ Auto-approve flags are dangerous. Always:

1. Run in a dedicated directory/branch
2. Use a sandbox (Docker/VM) for untrusted projects
3. Have git reset --hard ready as escape hatch
4. Review commits before pushing

Quick Start

# 1. Create project directory
mkdir my-project && cd my-project && git init

# 2. Copy templates from skill
cp /path/to/ralph-loop/templates/* .
mv PROMPT-PLANNING.md PROMPT.md

# 3. Create specs
mkdir specs
cat > specs/overview.md << 'EOF'
## Goal
Build a web app that...

## Tech Stack
- Python 3.11 + FastAPI
- SQLite
- HTMX + Tailwind

## Features
1. Feature one
2. Feature two
EOF

# 4. Edit PROMPT.md with your goal

# 5. Run the loop
./ralph.sh 20

Example: Antique Catalogue

# specs/overview.md
## Goal
Web app for cataloguing antique items with metadata, images, and categories.

## Tech Stack
- Python 3.11 + FastAPI
- SQLite + SQLAlchemy
- HTMX + Tailwind CSS
- Local file storage for images

## Features
1. CRUD for items (name, description, age, purchase info, dimensions)
2. Image upload (multiple per item)
3. Tags and categories
4. Search and filter
5. Multiple view modes (grid, list, detail)

The agent will:

1. (PLANNING) Break this into 10-15 tasks
2. (BUILDING) Implement each task, one per iteration
3. Commit after each successful implementation
4. Notify on completion or if blocked