Feature Request: File-Based Prompts for Ralph Wiggum Plugin

Feature Request: File-Based Prompts for Ralph Wiggum Plugin

Summary

Add support for reading complex multi-line prompts from files in the ralph-wiggum plugin, eliminating Bash parsing issues with structured instructions.

Problem

The current ralph-wiggum:ralph-loop implementation accepts prompts only as CLI arguments. This breaks with complex multi-line prompts due to Bash parsing limitations:

Current behavior:

# Simple prompt - Works ✅
/ralph-loop "Test framework parameter H1"

# Complex multi-line prompt - Fails ❌
/ralph-loop "Test 6 hypotheses (H11-H15, H20).

MEMORY CONTEXT:
- Ken Cheng style: MANDATORY (120)
- Concrete 80% minimum (117)

Each iteration:
  1. Generate post
  2. Evaluate with 3 personas
  ..."

# Error: Bash parse error - cannot handle newlines, special chars

Impact:

• Complex Ralph loops require careful escaping (error-prone)
• Multi-line structured instructions fail
• Copy-paste from files breaks formatting
• Limits Ralph's usability for sophisticated tasks

Proposed Solution

Add fallback: If no CLI arguments provided, read from state/ralph_current_prompt.txt:

# In setup-ralph-loop.sh, after parsing CLI arguments:

# If no prompt from args, try to read from file
if [[ -z "$PROMPT" ]]; then
  # Check for saved Ralph Loop files
  if [[ -f "state/ralph_current_prompt.txt" ]]; then
    echo "📄 Reading prompt from state/ralph_current_prompt.txt..." >&2
    PROMPT=$(cat state/ralph_current_prompt.txt)

    # Also read config if available
    if [[ -f "state/ralph_current_config.json" ]]; then
      echo "📄 Reading config from state/ralph_current_config.json..." >&2

      # Extract max_iterations if not already set
      if [[ $MAX_ITERATIONS -eq 0 ]]; then
        MAX_ITERATIONS=$(grep -o '"max_iterations":[[:space:]]*[0-9]*' state/ralph_current_config.json | grep -o '[0-9]*' || echo "0")
      fi

      # Extract completion_promise if not already set
      if [[ "$COMPLETION_PROMISE" == "null" ]]; then
        COMPLETION_PROMISE=$(grep -o '"completion_promise":[[:space:]]*"[^"]*"' state/ralph_current_config.json | sed 's/.*"completion_promise":[[:space:]]*"\([^"]*\)".*/\1/' || echo "null")
      fi

      echo "✅ Config loaded: max_iterations=$MAX_ITERATIONS, completion_promise=$COMPLETION_PROMISE" >&2
    fi
  fi
fi

Location to patch: setup-ralph-loop.sh, after line PROMPT="${PROMPT_PARTS[*]}" (~line 113)

Benefits

[1] Complex Prompts Work Seamlessly

# Create prompt file
echo "Complex multi-line prompt with:
- Memory context
- Structured instructions
- Special characters: \$, \", \`
- JSON snippets
- Code examples" > state/ralph_current_prompt.txt

# Run Ralph (auto-reads file!)
/ralph-loop

[2] Version Control Friendly

• Prompts can be versioned in Git
• Easy to review/edit in text editor
• Reusable templates

[3] Workflow Integration

Skills like ralph-builder can now:

1. Generate complex Ralph prompts programmatically
2. Save to state/ralph_current_prompt.txt
3. Call /ralph-loop (no CLI escaping needed!)

[4] Better UX

# Old way (error-prone):
/ralph-loop "Escape \"quotes\" and \$vars and \`backticks\`..."

# New way (just works):
# 1. Save prompt to file
# 2. /ralph-loop

Use Cases

Framework Optimization

# state/ralph_current_prompt.txt
Goal: Test 6 framework hypotheses (H11-H15, H20)

MEMORY CONTEXT (Phase 1):
- Ken Cheng MANDATORY (120)
- Concrete 80% minimum (117)

Each iteration:
  1. Generate blog post
  2. Spawn 3 parallel Task agents:
     → Evaluate as Alex CMO (0-50)
     → Evaluate as Victor CEO (0-50)
     → Evaluate as Mike Sales (0-50)
  3. Aggregate scores

<promise>FRAMEWORK_OPTIMIZATION_COMPLETE</promise>

# Run
/ralph-loop  # Auto-reads file!

Spec Execution (Multi-Phase)

# Orchestrator spawns Task agents per phase
# Each phase = fresh context (no bloat!)

Goal: Execute Spec 0076 autonomously

Each iteration (1 per phase):
  1. Read plan.md → Find next incomplete phase
  2. Spawn Task agent for Phase X:
     → Fresh context (0k → 40k → done)
  3. Wait for <promise>PHASE_X_COMPLETE</promise>
  4. Next phase

<promise>ALL_PHASES_COMPLETE</promise>

Backward Compatibility

✅ Fully backward compatible:

• CLI arguments still work as before
• File-reading is fallback only (when no CLI args)
• No breaking changes to existing workflows

Test:

# Old way still works
/ralph-loop "Simple prompt"  ✅

# New way (auto-detects file)
/ralph-loop  ✅ (reads from state/ralph_current_prompt.txt if exists)

Implementation Notes

Error Handling

Updated error message when no prompt found:

if [[ -z "$PROMPT" ]]; then
  echo "❌ Error: No prompt provided" >&2
  echo "" >&2
  echo "   Option 1 - Command line:" >&2
  echo "     /ralph-loop Build a REST API for todos" >&2
  echo "" >&2
  echo "   Option 2 - From file (auto-detected):" >&2
  echo "     Create state/ralph_current_prompt.txt with your prompt" >&2
  echo "     Optionally: state/ralph_current_config.json with config" >&2
  echo "     Then just: /ralph-loop (no arguments needed!)" >&2
  exit 1
fi

Config File Format

Optional state/ralph_current_config.json:

{
  "max_iterations": 50,
  "completion_promise": "TASK_COMPLETE",
  "generated_at": "2025-12-30T15:00:00Z",
  "generated_by": "ralph-builder v1.3",
  "mission": "Framework optimization Phase 2"
}

Testing

Tested on:

• WSL2 Ubuntu (Linux 6.6.87.2-microsoft-standard-WSL2)
• Claude Code CLI
• ralph-wiggum plugin versions: 6d3752c000e2 and unknown

Test cases:

1. ✅ Simple CLI prompt (backward compat)
2. ✅ Multi-line file prompt (new feature)
3. ✅ File + config.json (auto-load config)
4. ✅ No prompt at all (clear error message)

Additional Context

This enhancement enables autonomous multi-session execution patterns:

• Ralph orchestrator spawns Task agents per phase/task
• Each agent gets fresh context (no bloat!)
• Memory handoff via JSON files
• Scales to complex multi-phase workflows

Real-world impact:

• Spec execution: 180k tokens (single session) → 5-10k tokens (orchestrator) + fresh agents
• Framework testing: Sequential evaluations → Parallel Task agents (3x faster)
• Quality: No context degradation from bloat

Related Work

Skills leveraging this feature:

• ralph-builder - Framework optimization loops
• spec-executor - Multi-phase Spec execution
• ralph-hypothesis-generator - Generates complex Ralph prompts

References

• Claude Code: https://github.com/anthropics/claude-code
• Ralph Wiggum technique: https://docs.anthropic.com/claude/docs/ralph-wiggum

---

Would love to see this enhancement merged to enable more sophisticated autonomous workflows! 🚀