openclaw-master-skills/skills/agent-mode-upgrades

🚀 Agentic Loop Upgrade

An enhanced agentic loop for OpenClaw with planning, parallel execution, confidence gates, and semantic error recovery.

✨ Features

Feature	Core Loop	Enhanced Loop
Planning	❌ Reactive	✅ Goal decomposition with step tracking
Execution	Sequential	✅ Parallel (independent tools)
Error Handling	Retry-based	✅ Semantic recovery with alternatives
Confidence	Implicit	✅ Explicit gates for risky actions
Context	Overflow-triggered	✅ Proactive summarization
State	Implicit	✅ Observable FSM with checkpointing

🎯 What It Does

Planning & Reflection

The agent decomposes complex goals into step-by-step plans, tracks progress across turns, and reflects after each action to assess if steps are complete.

Parallel Execution

Independent tools execute concurrently for faster task completion. The orchestrator identifies which tools can run in parallel.

Confidence Gates

Before risky operations (file deletions, external messages, etc.), the system assesses confidence and can pause for approval.

Semantic Error Recovery

When tools fail, the system diagnoses the error type and attempts alternative approaches rather than simple retries.

Observable State Machine

Explicit state tracking enables debugging, dashboards, and checkpointing for resuming interrupted tasks.

📦 Installation

From ClawHub

openclaw skill install agentic-loop-upgrade

Manual Installation

1. Clone/download to your skills directory:

   cd ~/.openclaw/skills
   git clone https://github.com/openclaw/skill-agentic-loop-upgrade agentic-loop-upgrade
   

2. Build the TypeScript:

   cd agentic-loop-upgrade/src
   npm install
   npm run build
   

3. Restart OpenClaw:

   openclaw gateway restart
   

🚀 Quick Start

Enable via Dashboard

1. Open OpenClaw Dashboard → Agent → Mode
2. Click Enhanced Loop card
3. Configure settings (or use defaults)
4. Click Save Configuration

Disable

• Mode tab → Click Core Loop → Save
• Or delete: ~/.openclaw/agents/main/agent/enhanced-loop-config.json

⚙️ Configuration

All settings are available in the Mode dashboard:

Planning & Reflection

• Enable Planning: Generate execution plans before complex tasks
• Reflection After Tools: Assess progress after each tool execution
• Max Plan Steps: Maximum steps in a generated plan (2-15)

Execution

• Parallel Tools: Execute independent tools concurrently
• Max Concurrent: Maximum parallel tool executions (1-10)
• Confidence Gates: Assess confidence before risky actions
• Confidence Threshold: Minimum confidence to proceed (30-95%)

Context Management

• Proactive Management: Summarize and prune before overflow
• Summarize After N Iterations: Trigger summarization interval
• Context Threshold: Context fill level to trigger management

Error Recovery

• Semantic Recovery: Diagnose errors and adapt approach
• Max Recovery Attempts: Maximum alternative attempts (1-5)
• Learn From Errors: Store successful recoveries for future use

State Machine

• Enable State Machine: Track agent state transitions
• State Logging: Log all state transitions
• Metrics Collection: Collect timing metrics per state

Orchestrator Model

Select a cost-effective model for planning/reflection calls (e.g., Claude Sonnet 4.5).

📁 File Structure

~/.openclaw/
├── agents/main/agent/
│   └── enhanced-loop-config.json    # Configuration
├── agent-state/                      # Persistent plan state
│   └── {sessionId}.json
└── checkpoints/                      # Checkpoint files
    └── {sessionId}/
        └── ckpt_*.json

🔧 For Developers

Programmatic Usage

import { createOrchestrator } from "@openclaw/enhanced-loop";

const orchestrator = createOrchestrator({
  sessionId: "session_123",
  planning: { enabled: true, maxPlanSteps: 7 },
  approvalGate: { enabled: true, timeoutMs: 15000 },
  retry: { enabled: true, maxAttempts: 3 },
  context: { enabled: true, thresholdTokens: 80000 },
  checkpoint: { enabled: true },
}, {
  onPlanCreated: (plan) => console.log("Plan:", plan.goal),
  onStepCompleted: (id, result) => console.log("✓", result),
});

await orchestrator.init();

Architecture

See SKILL.md for full technical documentation.

🔒 Security & Trust

This skill wraps the agent runner and appends plan context to the agent's prompt. Both operations are bounded, transparent, and auditable:

Property	Value
Outbound network	LLM provider only (inherited from host)
Telemetry / phone-home	❌ None
Prompt modification	✅ Additive-only (appends status text; never replaces core prompt)
Runner bypass	❌ Never — original runner always called
Credential storage	❌ None
Persistence	Local ~/.openclaw/ only
Enabled by default	❌ No — requires explicit opt-in

Post-install verification:

~/.openclaw/skills/agentic-loop-upgrade/scripts/verify.sh

See SECURITY.md for the full audit document.

⚠️ Notes

• Token overhead: Planning and reflection use additional tokens (configurable via orchestrator model selection)
• Easy rollback: One click to switch back to Core Loop
• Checkpoints: Long tasks can be resumed if interrupted

📚 Documentation

• SKILL.md - Full technical documentation
• SECURITY.md - Security & trust audit document
• INSTRUCTIONS.md - Integration guide for agents
• references/ - Component documentation

🔗 Links

• OpenClaw
• Source Code
• ClawHub
• Documentation
• Discord

📄 License

MIT