本文へ移動
cccskills
無料GitHub で公開

genlayer-intelligent-contracts

Internet Court adapter for GenLayer Intelligent Contract supervision. Use to specify agent-performance rubrics, evidence schemas, decision outputs, and ERC-7710 connector expectations, while delegating actual GenLayer contract writing, linting, testing, deployment, and CLI interaction to the official GenLayer skills at https://skills.genlayer.com/.

インストール方法を見る

含まれるファイル(2)

  • SKILL.md6.9 KB
  • references/agent-supervisor-contract.md2.1 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

GenLayer Intelligent Contracts

Use this adapter when GenLayer is the judgment layer for agent behavior. It should define what the supervisor contract must review and emit: the user's mandate, the agent's action log, transaction receipts, rationales, artifacts, outputs, and the decision about whether authority should continue, be warned, be constrained, or be revoked.

This skill is not the Intelligent Oracle skill. Intelligent Oracle is for binary public web evidence markets. Use this skill for open-ended review of performance, relevance, quality, safety, and mandate compliance.

This skill is also not the GenLayer deployment skill. For writing, deploying, testing, and interacting with GenLayer Intelligent Contracts, use the official GenLayer skills at https://skills.genlayer.com/ or the GenLayer docs. If those skills are available in the environment, load them for implementation. If not, tell the user to install/enable them and keep this skill focused on Internet Court interfaces.

Core Model

Mandate + rubric + evidence
  -> Agent executes transactions
  -> Receipts and outputs are submitted
  -> GenLayer Intelligent Contract reviews behavior
  -> Decision: continue, warn, constrain, revoke, or escalate
  -> Internet Court / revocation controller executes the workflow result

GenLayer can judge qualitative criteria. It does not automatically revoke an EVM delegation unless the system includes a concrete cross-chain message, relayer, controller, or account module that consumes the GenLayer decision.

Workflow

  1. Extract the review target:
    • Objective, allowed authority, prohibited behavior, expected outputs, review cadence, evidence sources, and revocation triggers.
  2. Draft a rubric:
    • Include relevance, mandate compliance, transaction safety, budget discipline, artifact quality, evidence completeness, and user-value criteria.
  3. Define evidence schema:
    • Use structured action receipts, transaction hashes, deployed addresses, agent rationales, output artifacts, costs, and prior warnings.
  4. Define decisions:
    • continue, warn, constrain, revoke, and optionally escalate.
  5. Define effect path:
    • For ERC-7710, load ../genlayer-erc7710-connector/SKILL.md to specify how a revoke or constrain decision reaches the EVM revocation controller or delegation manager.
  6. Add auditability:
    • Store the reviewed action ids, score, reasoning summary, decision timestamp, and next review requirements.

Review Input

Use this as the default review request shape:

type AgentPerformanceReviewInput = {
  mandateId: string;
  objective: string;
  authoritySummary: string;
  prohibitedActions: string[];
  rubric: string[];
  reviewWindow: {
    startsAt: string;
    endsAt: string;
  };
  actionReceipts: AgentActionReceipt[];
  outputs: Array<{
    artifactId: string;
    uri: string;
    summary: string;
  }>;
  costs: {
    txCount: number;
    nativeGasSpent: string;
    tokenSpent?: string;
  };
  priorDecisions: AgentReviewDecision[];
};

For wallet-spend supervision, include an on-chain spend snapshot in each review window:

type WalletSpendSnapshot = {
  mandateId: string;
  permissionHash: `0x${string}`;
  sourceChainId: 84532;
  spendReporter: `0x${string}`;
  actionType: `0x${string}`;
  policyHash: `0x${string}`;
  resourceHash: `0x${string}`;
  delegator: `0x${string}`;
  delegate: `0x${string}`;
  asset: `0x${string}`;
  payTo: `0x${string}`;
  spent: bigint;
  maxTotalSpend: bigint;
  maxPerRequest: bigint;
  requestCount: bigint;
  validAfter: number;
  validUntil: number;
  controllerNonce: bigint;
  reviewWindowId: `0x${string}`;
  evidenceBundleHash: `0x${string}`;
  observedAt: number;
};

On Base Sepolia demos, prefer a reporter contract that reads spend state directly from a permission/accounting manager and sends this snapshot through the GenLayer Studio bridge boilerplate. x402 is one source; contract-call, deployment, compute-credit, or subscription managers can expose the same shape.

Decision Output

type AgentReviewDecision = {
  mandateId: string;
  decision: "continue" | "warn" | "constrain" | "revoke" | "escalate";
  score: number;
  reasoning: string;
  violations: string[];
  constraints?: {
    maxTxsPerPeriod?: number;
    maxGasPerTx?: string;
    requireHumanApprovalFor?: string[];
    disabledActions?: string[];
  };
  reviewedActionIds: string[];
  issuedAt: string;
};

Use constrain when the agent is useful but overreaches. Use revoke when the agent violates scope, creates unsafe deployments, repeatedly wastes budget, hides evidence, or produces outputs that do not meaningfully pursue the mandate.

Rubric Defaults

For delegated on-chain agents, include:

  • Mandate relevance: actions are connected to the user's objective.
  • Scope compliance: actions stay within allowed chains, contracts, selectors, factories, and action types.
  • Safety: no ownership loss, unlimited approvals, suspicious external calls, or avoidable irreversible effects.
  • Budget discipline: gas/value/paid-tool spend is proportional to the work.
  • Evidence quality: every action has a receipt, rationale, and artifact when applicable.
  • Result quality: deployed contracts or outputs are usable and aligned with the requested workflow.
  • Transparency: the agent explains why it took each meaningful action.

Revocation Integration

When the decision can affect ERC-7710 authority, always specify:

type RevocationEffect = {
  sourceChain: "genlayer";
  targetChain: "base" | "ethereum" | string;
  revocationController: `0x${string}`;
  mandateId: string;
  permissionHash: `0x${string}`;
  requiredDecision: "revoke" | "constrain";
  relayer: string;
  finalityAssumption: string;
  fallback: "owner_emergency_revoke" | "absolute_expiry" | "human_review";
};

Do not claim revocation is trustless unless the bridge, relayer, and verification path are defined.

The connector skill owns the exact payload, relay/proof mode, EVM controller interface, and where the ERC-7710 redemption path checks the controller.

Output Checklist

When designing a GenLayer Intelligent Contract review flow, return:

  1. Plain-English review purpose.
  2. Review input schema.
  3. Rubric.
  4. Decision output schema.
  5. Revocation/constraining effect path, with connector skill handoff if ERC-7710 is involved.
  6. Evidence submission process.
  7. Happy path.
  8. Failure and appeal paths.
  9. Tests for continue, warn, constrain, revoke, missing evidence, and bad relayer behavior.

References

  • references/agent-supervisor-contract.md for a reusable supervisor pattern.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

0G Compute Network guide for decentralized AI inference, fine-tuning, and GPU services. Covers chatbots, image generation, speech-to-text, SDK integration (0g-serving-broker), processResponse API, broker.inference methods, CLI commands (0g-compute-cli), and account management. Use this skill for any 0G compute, 0G AI, or decentralized GPU question.

日本語の概要は準備中です。原文の説明を表示しています。

internet-court/internet-court-skill6,7302026年8月20日 更新

Builds Agent-to-Agent (A2A) servers and clients following Google's open protocol for agent interoperability. Use when the user wants to create an A2A-compliant agent, build an Agent Card, implement task management, connect agents across frameworks, set up agent discovery, handle streaming responses, implement push notifications, or orchestrate multi-agent workflows. Trigger words: a2a, agent to agent, agent2agent, a2a protocol, a2a server, a2a client, agent card, agent interoperability, agent collaboration, multi-agent, agent discovery, a2a sdk, a2a task.

日本語の概要は準備中です。原文の説明を表示しています。

internet-court/internet-court-skill6,7302026年8月20日 更新

Give the AI agent its own EVM wallet with admin-controlled policies the agent CANNOT bypass even under prompt injection. Encrypted keystore (AES-256-GCM, scrypt KDF), policy file the agent has no tool to write, deterministic policy gate on every signing operation, optional local HTTP dashboard. Triggers: agent wallet, give the agent a wallet, agent address, fund the agent, agent autonomy, policy gate, kill switch, agent permissions, bounded autonomy, ERC-4337 alternative, session-key alternative.

日本語の概要は準備中です。原文の説明を表示しています。

internet-court/internet-court-skill6,7302026年8月20日 更新

Crypto wallet operations via the awal CLI — sign in, check balances, send USDC/ETH/POL/SOL, trade tokens, fund the wallet, and use the x402 payment protocol to discover paid services, pay for API calls, monetize an API, or query onchain data. Use whenever the user mentions signing in, login, authentication, wallet status, balance, address, sending money, paying someone, transferring tokens, ENS names, swapping/trading/converting tokens, funding/topping up/onramp, USDC, ETH, POL, SOL, the x402 bazaar, paid APIs, monetizing an endpoint, or querying onchain data on Base.

日本語の概要は準備中です。原文の説明を表示しています。

internet-court/internet-court-skill6,7302026年8月20日 更新

Help developers write code that interacts with Alkahest escrow contracts using the TypeScript, Rust, or Python SDK

日本語の概要は準備中です。原文の説明を表示しています。

internet-court/internet-court-skill6,7302026年8月20日 更新

Interact with Alkahest escrow contracts as a buyer, seller, or oracle using the CLI

日本語の概要は準備中です。原文の説明を表示しています。

internet-court/internet-court-skill6,7302026年8月20日 更新

internet-court のスキルをすべて見る

このスキルの問題を報告する