📦 Edicts — Ground Truth for AI Agents

技能文档

Stop your agent from hallucinating facts that matter.

Edicts is a lightweight ground-truth layer that injects verified facts (launch dates, product constraints, compliance rules, naming conventions, embargoes) directly into your agent's system context. Every prompt gets the facts. No retrieval stack needed.

Install

openclaw plugins install openclaw-plugin-edicts

Or from ClawHub:

openclaw plugins install clawhub:openclaw-plugin-edicts

What It Does

• Provides facts in every prompt — edicts appear in system context automatically
• Read-only by default — agents can list and search edicts; write tools are opt-in
• Categorized and typed — each edict has category, confidence level, TTL, and optional expiry
• Token-efficient — compact YAML/JSON storage, configurable token budget and category limits

Tools Provided

Configuration

In your openclaw.json under plugins.entries:

{
  "openclaw-plugin-edicts": {
    "enabled": true,
    "config": {
      "format": "yaml",
      "maxEdicts": 200,
      "tokenBudget": 4000,
      "autoSave": true,
      "includeSystemContext": true
    }
  }
}

Key Config Options

• format: yaml or json — storage format
• maxEdicts: Maximum number of edicts (default: 200)
• tokenBudget: Max tokens for system context injection (default: 4000)
• categories: Restrict to specific categories
• staleThresholdDays: Days before an edict is considered stale
• tools.enabled: Enable/disable runtime tools (default: true)
• tools.names: Whitelist specific tools
• autoSave: Auto-persist changes (default: true)

Edict Structure

- text: "Product v2.0 launches April 15, NOT before."
  category: product
  confidence: verified    # verified | inferred | user
  ttl: event             # ephemeral | event | durable | permanent
  expiresAt: "2026-04-16"
  tags: [launch, dates]

Use Cases

• Launch dates — "v2.0 launches April 15, NOT before"
• Naming rules — "Always call it 'Casper v2.2.0', never 'Kyoto'"
• Compliance — "NEVER share customer data in group chats"
• Migration status — "Database migration complete as of March 1"
• Embargoes — "Do not discuss partnership until press release on April 5"
• Engagement rules — "Do not engage with @trollaccount on X"

Security and Trust Model

You control what appears in context. Edicts are stored in a local file in your workspace. Only content you write to that file appears in prompts. No remote fetch, no external data sources, no network calls.

Write tools are opt-in. By default, only read tools are active. Write tools (edicts_add, edicts_update, edicts_remove) must be explicitly enabled if you want agents to create or modify edicts. Configuration options:

• tools.enabled: false — read-only prompt injection, no agent mutation
• tools.names: [edicts_list, edicts_search] — whitelist read-only tools
• autoSave: false — runtime changes don't persist across sessions
• includeSystemContext: false — tools without prompt injection

Safest config: tools.enabled: false + autoSave: false — your curated file, injected as-is, no runtime changes possible.

Links

• Docs: https://edicts.ai
• npm: https://www.npmjs.com/package/openclaw-plugin-edicts
• Core library: https://www.npmjs.com/package/edicts
• GitHub: https://github.com/make-software/edicts
• License: MIT