melt - Task-Agnostic Autonomous Execution

/melt - Task-Agnostic Autonomous Execution

Complete guide to the /melt skill for autonomous task execution with optional Agent Teams planning and completion validation. (/build is a legacy alias.)

Table of Contents

1. Overview
2. When to Use
3. Planning
4. Workflow
5. Completion Checkpoint
6. Comparison with /appfix
7. Skill Fluidity
8. Troubleshooting

Overview

/melt is the universal autonomous execution skill. It provides:

• Optional Agent Teams Planning - Encouraged for complex tasks, not mandated
• 100% Autonomous Operation - No permission prompts, no confirmation requests
• Completion Checkpoint - Deterministic boolean validation before stopping
• Browser Verification - Mandatory testing in real browser
• Strict Linter Policy - Fix ALL errors, including pre-existing ones

Key Features

When to Use

Use /melt

Use /repair Instead

Planning

For complex tasks, spawn an Agent Team to ensure optimal implementation planning. Use TeamCreate to coordinate parallel agents, or launch parallel Task() calls directly.

Recommended Agents (3-5)

Important: Forge reads the core agent prompts from ~/.claude/skills/heavy/SKILL.md at runtime. This ensures prompts stay in sync - when heavy improves, forge automatically benefits.

Why Agent Teams?

Synthesis Output

After agents return, synthesize their insights:

TRADEOFF: [topic]
- First Principles: Delete X because [reason]
- AGI-Pilled: Expand Y because [capability argument]
- Resolution: [chosen approach with rationale]

Workflow

Phase 0: Activation

# The skill creates the state file on activation
mkdir -p .claude && cat > .claude/autonomous-state.json << 'EOF'
{"mode": "melt", "started_at": "2025-01-26T10:00:00Z", "task": "user task"}
EOF

mkdir -p ~/.claude && cat > ~/.claude/autonomous-state.json << 'EOF'
{"mode": "melt", "started_at": "2025-01-26T10:00:00Z", "origin_project": "/path/to/project"}
EOF

Phase 0.5: Planning

Encouraged for complex tasks before making any changes.

1. EnterPlanMode - Switch to planning mode
2. Explore the codebase:
   • Project structure and architecture
   • Recent commits: git log --oneline -15
   • Environment and deployment configs
   • Relevant code patterns for the task
   • Existing tests and validation
3. Spawn Agent Team via TeamCreate or parallel Task() calls (First Principles + AGI-Pilled + task-specific experts, 3-5 recommended)
4. Synthesize tradeoffs and write to plan file
5. ExitPlanMode - Get plan approved

Why this matters: Jumping straight to code leads to broken functionality, inconsistent patterns, and wasted effort.

Phase 1: Execute

1. Make code changes (Edit tool)
2. Run linters, fix ALL errors
3. Commit and push
4. Deploy if needed

Phase 2: Verify (MANDATORY)

CRITICAL: Use Surf CLI first, not Chrome MCP.

1. Run Surf CLI: python3 ~/.claude/hooks/surf-verify.py --urls "https://..."
2. Check .claude/web-smoke/summary.json for pass/fail
3. Only fall back to Chrome MCP if Surf CLI unavailable
4. Check console for errors
5. Verify feature works as expected

Phase 3: Complete

1. Update completion checkpoint
2. Try to stop
3. If blocked: address issues, try again
4. If passed: clean up state files

Completion Checkpoint

Create .claude/completion-checkpoint.json:

{
  "self_report": {
    "is_job_complete": true,
    "code_changes_made": true,
    "linters_pass": true,
    "category": "refactor"
  },
  "reflection": {
    "what_was_done": "Implemented feature X, deployed to staging, verified in browser",
    "what_remains": "none",
    "key_insight": "Reusable lesson for future sessions (>50 chars)",
    "search_terms": ["keyword1", "keyword2"],
    "memory_that_helped": []
  }
}

Required Fields

Comparison with /appfix

/melt is the universal base with optional Agent Teams planning. /appfix adds debugging-specific phases.

Skill Fluidity

Skills are capabilities, not cages. You may use techniques from any skill for sub-problems without switching modes.

Troubleshooting

"Checkpoint validation failed"

Your checkpoint has incomplete fields. Check:

1. is_job_complete - Are you honestly done?
2. linters_pass - Did all linters pass?
3. what_remains - Is it "none"?
4. key_insight - Is it >50 chars?

"Stop hook blocked me"

This is expected when work is incomplete:

• If is_job_complete: false → you're blocked
• Complete the work, update checkpoint, try again

"linters_pass is false"

Fix ALL linter errors:

# JavaScript/TypeScript
npm run lint && npx tsc --noEmit

# Python
ruff check --fix . && pyright

"These errors aren't related to our code" is NOT acceptable.

State Files

Cleanup

Remove state files when done:

rm -f ~/.claude/autonomous-state.json .claude/autonomous-state.json