ralph-start

/ralph-start — Activate Autonomous Loop

You are activating a Ralph Wiggum autonomous loop. When Claude tries to exit, the Stop hook will re-feed the task prompt, creating a continuous work loop that persists across iterations.

Input

The user provides:

1. Task prompt (required): What to work on autonomously
2. --max-iterations N (optional): Maximum loop iterations (default: 50)
3. --scope DIR (optional): Directory to freeze edits to
4. --reset-strategy compact|reset (optional): Context management strategy (default: compact)
   • compact: Standard — carry forward context with compaction recovery
   • reset: Full context reset each iteration — clean slate with structured handoff artifact. Better for long-running tasks (10+ iterations) where context anxiety degrades quality.

Activation Protocol

Step 1: Parse Arguments

Extract:

• prompt: The task description (everything that isn't a flag)
• max_iterations: From --max-iterations N or default 50
• scope: From --scope DIR or empty
• reset_strategy: From --reset-strategy or default "compact"

Step 2: Create Loop File

Write the loop configuration to ~/.claude/knowledge/.ralph-active:

cat > ~/.claude/knowledge/.ralph-active << 'RALPH'
{
  "prompt": "THE_TASK_PROMPT",
  "max_iterations": 50,
  "iteration": 0,
  "scope": "DIRECTORY_OR_EMPTY",
  "started_at": "ISO_TIMESTAMP",
  "started_by": "session"
}
RALPH

Use python3 with json.dumps for safe serialization (handles quotes, backslashes, newlines):

python3 << 'PYEOF'
import json, datetime, os
data = {
    "prompt": "THE_PROMPT",
    "max_iterations": MAX_ITER,
    "iteration": 0,
    "scope": "SCOPE_OR_EMPTY",
    "reset_strategy": "RESET_STRATEGY",
    "started_at": datetime.datetime.now().isoformat()
}
ralph_path = os.path.join(os.environ.get("HOME", os.path.expanduser("~")), ".claude", "knowledge", ".ralph-active")
with open(ralph_path, "w") as f:
    json.dump(data, f, indent=2)
PYEOF

Replace THE_PROMPT, MAX_ITER, SCOPE_OR_EMPTY, and RESET_STRATEGY with the actual values. Use a heredoc (<< 'PYEOF') so that the prompt text is never interpolated by the shell — only parsed by Python's JSON serializer.