PolicyGuard OpenClaw Plugin

PolicyGuard = WDK execution safety layer for OpenClaw.
Integration path: OpenClaw Tool -> PolicyGuard -> WDK -> Chain

This plugin enforces a deterministic approval boundary before funds-related actions execute:

• PASS: safe/non-funds intent can proceed
• CHALLENGE: funds intent requires explicit human approval

README Scoring Map (Criterion -> Evidence)

| Scoring Criterion | Delivered Evidence in Repo | How to Verify Quickly | |---|---|---| | Problem relevance (AI agent fund safety) | README + AWARD_PITCH.md problem framing | Read first 2 sections | | Deterministic safety mechanism | src/policy-engine.ts, src/commands.ts | Run /policy ... and confirm CHALLENGE for funds intents | | Approval gating + idempotency | src/commands.ts challenge state transition | Approve same challenge twice and confirm duplicate approve is blocked | | WDK/OpenClaw integration implementation | src/index.ts, src/wdk-adapter.ts, openclaw.plugin.json | Install plugin and invoke policyguard_command | | Engineering quality | tests/*, package.json scripts | Run npm run build && npm test && npm run validate |

Quickstart (aligned to demo user story)

a) Install policyguard-openclaw-plugin

openclaw plugins install policyguard-openclaw-plugin

(Version pin also works: openclaw plugins install [email protected])

b) Create WDK wallet context

Set the seed env key used by the plugin:

export WDK_SEED="<your mnemonic>"

Recommended plugin config:

{
  "plugins": {
    "policyguard-openclaw-plugin": {
      "persistencePath": "./data/pending-challenges.json",
      "wdkSeedEnvKey": "WDK_SEED",
      "chain": "arbitrum",
      "accountIndex": 0,
      "rpcUrl": "https://arb1.arbitrum.io/rpc",
      "swapProtocolLabel": "velora",
      "swapMaxFee": "0.003"
    }
  }
}

c) Let user transfer some ETH to Arbitrum mainnet

Talk directly to OpenClaw in chat:

1. /policy transfer <amount> ETH to <address>
2. Receive CHALLENGE + challengeId
3. /approve <challengeId> <reason>

d) Monitor hot on-chain tokens, analyze, then buy 1U

Demo user story step (same conversation style):

1. Ask OpenClaw to monitor hot tokens
2. Ask OpenClaw to pick the best opportunity
3. Execute buy with a 1U-sized action through policy challenge + approval flow

Delivered architecture

1. Command layer (src/commands.ts)

   • Handles /policy, /approve, /reject
   • Maintains challenge state transitions

2. Policy layer (src/policy-engine.ts)

   • Deterministic PASS/CHALLENGE logic
   • Stable request fingerprinting

3. Intent layer (src/intent.ts)

   • Extracts entities for transfer/swap flows

4. Execution adapter (src/wdk-adapter.ts)

   • Routes approved intents into execution paths
   • Normalizes failures into structured categories

5. Persistence (src/persistence.ts)

   • Durable challenge records for audit and replay analysis

Delivered technical highlights

• Deterministic policy boundary independent from model randomness
• Idempotent challenge lifecycle to prevent duplicate execution
• Transfer + swap intent extraction with canonical parameters
• Runtime safety checks for seed handling (env-only design)
• Structured error taxonomy: RPC | ALLOWANCE | GAS | BALANCE | TIMEOUT | UNKNOWN
• Local execution-first architecture for controlled deployment environments

Security constraints (implemented)

• Seed-like plaintext keys in config are rejected at startup
• Use environment variables for sensitive material only

Local verification

npm install
npm run build
npm test
npm run validate

Hackathon context source

• Rules/context link provided by organizer:
  • https://hcni4f4mdq79.feishu.cn/wiki/LVzIwMpmKixXeHkeQQ0c8sn8nWg