loop-engine-governance（循环引擎治理）

loop-engine-governance

Overview

loop-engine-governance adds policy enforcement to OpenClaw workflows by routing decisions through Loop Engine transitions and guards.

Modes of operation

Local governance mode (no external LLM provider)

• Uses Loop Engine runtime, guards, and audit trail only.
• No external LLM API calls occur in this mode.
• Suitable for human-only and automation-only loop flows.

LLM-augmented mode (external provider calls enabled)

• Enabled only when a provider adapter is explicitly configured.
• Provider-backed examples call external APIs and may transmit prompt/evidence context to that provider.

Installation

# Core (required for all modes)
npm install @loop-engine/sdk @loop-engine/adapter-memory @loop-engine/adapter-openclaw

# Optional: provider-backed adapters (install only what you use)
npm install @loop-engine/adapter-anthropic @anthropic-ai/sdk
npm install @loop-engine/adapter-openai openai
npm install @loop-engine/adapter-grok

Configuration

• Local mode requires loop definitions, storage, and guard registry configuration only.
• Provider-backed mode additionally requires the corresponding provider adapter and API key.
• External provider calls are activated by adapter usage (for example createOpenAIActorAdapter(...)), not by Loop Engine core alone.

Environment variables

Provider keys are required only for provider-backed examples:

Additional provider key used elsewhere in this repo:

• GOOGLE_AI_API_KEY for @loop-engine/adapter-gemini examples and adapter usage.

External network and data flow

• No provider adapter configured: no external LLM network calls.
• Provider adapter configured: prompt/evidence context passed to createSubmission(...) may be sent to:

- OpenAI (@loop-engine/adapter-openai) - Anthropic (@loop-engine/adapter-anthropic) - xAI Grok (@loop-engine/adapter-grok) - Google Gemini (@loop-engine/adapter-gemini)

• OpenClaw integration (@loop-engine/adapter-openclaw) uses a WebSocket gateway connection (gatewayUrl, default ws://127.0.0.1:18789) for event forwarding.

Sensitive data guidance

• Do not send raw PII, PHI, PCI, credentials, or other regulated data to provider-backed examples without review.
• Redact, tokenize, or minimize sensitive fields before submitting evidence context.
• Review provider retention, training, and contractual controls before production use.

Provenance

• Canonical repository: https://github.com/loopengine/loop-engine
• Skill source path: packages/adapter-openclaw/loop-engine-governance/
• Maintainer organization: Better Data, Inc. (https://betterdata.co)
• Documentation site: https://loopengine.io/docs/integrations/openclaw

Package/source references

• @loop-engine/adapter-openclaw: https://www.npmjs.com/package/@loop-engine/adapter-openclaw
• @loop-engine/sdk: https://www.npmjs.com/package/@loop-engine/sdk
• @loop-engine/adapter-openai: https://www.npmjs.com/package/@loop-engine/adapter-openai
• @loop-engine/adapter-anthropic: https://www.npmjs.com/package/@loop-engine/adapter-anthropic
• @loop-engine/adapter-grok: https://www.npmjs.com/package/@loop-engine/adapter-grok
• @loop-engine/adapter-gemini: https://www.npmjs.com/package/@loop-engine/adapter-gemini

What this skill does

Wires Loop Engine into OpenClaw so that any workflow step can be governed by:

• Human approval gates — transitions only a named human actor can trigger
• AI confidence guards — block AI recommendations below a threshold
• Evidence capture — attach structured context to every decision
• Audit trail — every transition is attributed, timestamped, and immutable

How it works with OpenClaw

OpenClaw agent proposes action
        ↓
Loop Engine evaluates guards       ← @loop-engine/adapter-openclaw
        ↓
Human approves (if policy requires)
        ↓
OpenClaw executes the approved action

Guards are enforced at the runtime level — not in prompts.

How governance weighting works

Three types of weighting evaluated in sequence — all must pass:

1. Confidence threshold (numeric gate) Every AI actor submission carries a 0–1 confidence score. The guard blocks the transition if the score falls below the configured threshold.

2. Guard priority (hard vs soft) Hard failures block the transition regardless of everything else. A human-only guard is an absolute block — no confidence score overrides it.

3. Evidence completeness (structural gate) The evidence-required guard checks for specific fields before allowing a transition. Missing any required field blocks the transition.

Evaluation order:

1. Actor authorized for this signal?
2. Required evidence fields present?
3. Confidence score above threshold?
4. All hard guards pass?

Quick start (no API key required)

import { createLoopSystem, parseLoopYaml, CommonGuards, guardEvidence } from '@loop-engine/sdk'
import { MemoryAdapter } from '@loop-engine/adapter-memory'

const definition = parseLoopYaml(`
  loopId: approval.workflow
  name: Approval Workflow
  version: 1.0.0
  initialState: pending
  states:
    - stateId: pending
      label: Pending Approval
    - stateId: approved
      label: Approved
      terminal: true
  transitions:
    - transitionId: approve
      from: pending
      to: approved
      signal: approve
      allowedActors: [human]
      guards: [human-only]
`)

const system = createLoopSystem({
  storage: new MemoryAdapter(),
  guards: CommonGuards,
})

const loop = await system.startLoop({ definition, context: {} })

// Only a human actor can approve — AI and automation actors are blocked.
// guardEvidence strips PII fields and prompt-injection patterns before
// the evidence object is forwarded to any external LLM adapter.
await system.transition({
  loopId: loop.loopId,
  signalId: 'approve',
  actor: { id: 'alice', type: 'human' },
  evidence: guardEvidence({ reviewNote: 'Looks good' }),
})

Examples included

All examples use synthetic data. Do not use real PII or regulated data without reviewing your provider's data processing agreements.

Evidence sanitization

All evidence objects must be guarded before being forwarded to external LLM adapters. guardEvidence (exported from @loop-engine/sdk) enforces three rules at the skill boundary:

1. PII field blocking — fields whose names match known PII patterns (ssn, email, phone,

dob, password, token, healthrecord, mrn, and 20+ others) are dropped before forwarding.

1. Prompt injection stripping — string values beginning with role prefixes (system:, user:,

assistant:) are stripped to prevent instruction injection via evidence payloads.

1. Value length cap — string values are truncated at 512 characters to prevent context stuffing.

Always wrap caller-supplied evidence with guardEvidence() before passing it to system.transition(). The Quick Start above shows the correct pattern.

Security notes

• Local governance mode runs without external LLM provider calls.
• Provider-backed mode requires explicit adapter activation and the corresponding API key.
• Evidence and prompt context can leave the local environment only in provider-backed mode.
• This skill does not claim compliance certifications or data-processing guarantees.

Documentation

https://loopengine.io/docs/integrations/openclaw

License

MIT-0 — free to use, modify, and redistribute. No attribution required.

@loop-engine/* packages: Apache-2.0 Provider SDKs: licensed by their respective maintainers