Relay Action Card

Public schemas, examples, skills, adapters, and verification for reviewing consequential agent actions before execution.

Runtime: Node 24 via .nvmrc. Public package compatibility remains Node >=22.14.0 and is checked in a separate CI job.

How The Neura Stack Fits

• Protocol owns shared message shape, versioning, validation, and conformance.
• Relay evaluates proposed actions and returns Decision Receipts plus refs-only evidence.
• Registry supplies optional identity, standing, capability, and continuity context.
• The developer or operator runtime owns downstream execution or restraint.

This repository is the public proof and integration surface. It is not the private Relay runtime and does not issue production authority or execute downstream work.

Start In 30 Seconds

git clone https://github.com/neurarelay/relay-action-card.git
cd relay-action-card
nvm use
npm ci
npm run check

Then run the golden path:

node examples/core/resolve-action-card.mjs

The example reads a synthetic Action Card, demonstrates a governed decision boundary, and returns a Decision Receipt-shaped result without executing the proposed action.

Golden Path

Start with these files:

1. examples/core/action-card.json: proposed action.
2. examples/core/resolve-action-card.mjs: bounded resolver example.
3. examples/core/decision-receipt.example.json: returned decision evidence.
4. schemas/action-card.v0.1.json: machine-readable Action Card schema.
5. schemas/decision-receipt.v0.1.json: machine-readable Decision Receipt schema.

Use the golden path before choosing a specialized route.

Specialized Routes

MCP And Agent Frameworks

Use examples/mcp/README.md for direct MCP, OpenAI, Anthropic, Google ADK, Microsoft Agent Framework, and related runtime examples. These examples demonstrate integration boundaries; they do not establish official provider endorsement or production partnership.

OpenClaw

Use examples/openclaw/QUICKSTART.md for the public preflight adapter, action-receipt kit, near-miss workbench, and synthetic severe-scenario proof. Published-package source and provenance must remain aligned with this repository.

Choose the lane that matches what you want to prove with the OpenClaw-style receipt kit:

• OpenClaw Developer Journey Proof: read docs/openclaw-developer-journey.md, then run npm run openclaw:proof. The source is examples/openclaw/run-developer-journey-proof.mjs; an explicitly authorized live test uses npm run openclaw:proof -- --live. Verify clean installation with npm run verify:openclaw-clean-consumer.
• OpenClaw OS Decision Receipt Surface: read docs/openclaw-os-decision-receipt-surface.md, then run npm run openclaw:workspace-proof and npm run verify:openclaw-workspace-surface. The source is examples/openclaw/run-workspace-decision-surface.mjs; verifier anchors are verify-openclaw-workspace-surface.mjs and openclaw-workspace-surface.test.mjs.
• Near-miss workbench: read docs/openclaw-near-miss-workbench.md, run npm run openclaw:workbench, review docs/assets/openclaw-near-miss-workbench/near-miss-workbench-desktop.png, and inspect artifacts/openclaw-near-miss-workbench/report.html. The report shows what the agent was about to do, what Neura caught, the receipt route, and the developer-owned next step. Verify the route with npm run verify:openclaw-workbench.
• Severe Scenario Proof Pack: read docs/openclaw-severe-scenario-proof-pack.md, run npm run openclaw:severe-proof, verify with npm run verify:openclaw-severe-proof, test with npm run test:openclaw-severe-proof, and inspect artifacts/openclaw-severe-scenario-proof/report.html.
• Severe Preflight Queue: read docs/openclaw-severe-preflight-queue.md, run npm run openclaw:severe-preflight, verify with npm run verify:openclaw-severe-preflight, test with npm run test:openclaw-severe-preflight, and inspect artifacts/openclaw-severe-preflight-queue/transcript.html.

The OpenClaw route keeps its implementation under examples/openclaw/, including near-miss-workbench/, workspace-surface/, and preflight-adapter/; reusable operator material remains under skills/openclaw/. The workspace proof writes artifacts/openclaw-workspace-decision-surface/report.html.

SDK And A2A

Use examples/sdk/README.md and examples/a2a/README.md for protected Relay client examples. Run npm run example:sdk:authority-routing to see hold_for_registry_backed_authority for unverified public demo context. The current SDK is 0.1.3; the route is runtime-owned and has no public API keys. Production access still requires the relevant Relay authentication and product boundary.

A2A Protected Client Proof

The A2A Controlled Client Pack v0.2 demonstrates the A2A Controlled Runtime v1 boundary:

A2A client -> public Agent Card -> protected `/a2a` message/send -> Decision Receipt task

Run public discovery with npm run example:a2a -- --agent-card-only. With controlled access, run RELAY_A2A_ACCESS_TOKEN=... npm run example:a2a -- --json. The route uses examples/a2a, keeps execution developer-owned, and preserves no public A2A token issuance, no public API keys, and no unprotected execution.

Canonical Schemas

The active public schema mirror is:

• action-card.v0.1.json;
• decision-receipt.v0.1.json;
• agent-io-event-envelope.v0.1.json;
• mcp-approval-receipt.v0.1.json;
• recoverable-observer-envelope.v0.1.json;
• structured-result-trust-receipt.v0.1.json.

Neura Protocol is the canonical authority. schemas/manifest.json records the Protocol source path, versions, and expected SHA-256 values for these byte-identical public mirrors.

Shared-contract movement requires an intentional version change and successful local plus cross-repository compatibility verification.

Skills

Reusable skills live under skills/:

• neura-action-card for constructing bounded Action Cards;
• neura-authority-review for reviewing authority and evidence;
• neura-first-receipt for producing a first receipt-shaped result.

The skills preserve the same boundary as the examples: review and evidence before developer-owned execution.

Package Reality

The OpenClaw preflight adapter is sourced from this public repository and has provenance. The SDK and OpenAI Agents packages currently source from the private Relay repository while their metadata points here.

Existing versions must not be mutated or republished. Before any future SDK or OpenAI Agents version, publishable source, repository metadata, Trusted Publishing identity, provenance, and release tags must be aligned prospectively.

OpenClaw submission-readiness evidence remains indexed at docs/openclaw-clawhub-submission-readiness.md. Verify it with npm run verify:openclaw-submission-readiness and npm run verify:openclaw-clawhub-release; the controlled publisher verifier is verify-openclaw-founder-clawhub-publisher.mjs. The current npm package is @neurarelay/openclaw-preflight-adapter@0.1.5; the separately published ClawHub community package remains 0.1.4, alongside the legacy @rpelevin/neura-relay-preflight-adapter@0.1.1. No official OpenClaw or ClawHub listing, approval, endorsement, partnership, or integration claim exists.

After The First Receipt

Run the public proof first, then choose the next route:

• Open a First receipt feedback issue with receipt and trace refs only.
• Request Sandbox MCP access through the signed-in Relay Workspace.
• Request Production/private MCP access only for a concrete governed-action use case.
• Create the production Agent Passport in Neura Registry before production identity validation: https://www.neuraregistry.com/sign-up?next=%2Fbuilder%2Fagents%2Fnew.
• The Official MCP Registry and com.neurarelay/relay-mcp are distribution references only; they do not establish approval, listing, endorsement, partnership, or integration.

Share refs only. Do not paste private payloads, credentials, tokens, or customer data into issues or documentation.

Verification

Primary development and CI:

nvm use
npm ci
npm run check
npm audit --audit-level=high

GitHub runs the Node 24 proof-contract gate and one Node 22.14.0 public-compatibility job. Branch protection requires the deterministic local-contract and verify checks on an up-to-date branch.

Stable proof snapshot and Action Receipt Kit compatibility index:

• Local and receipt modes: npm run openclaw:dry-run and npm run openclaw:receipts.
• Contract gates: npm run verify:openclaw-action-receipt-kit, npm run verify:openclaw-developer-journey, npm run test:openclaw-kit, and npm run test:openclaw-kit:e2e.
• Verifier anchors: verify-openclaw-near-miss-workbench.mjs, verify-openclaw-developer-journey.mjs, and openclaw-action-receipt-kit.yml; historical release notes remain in CHANGELOG.md.
• Fixture routes include openclaw-memory-write and openclaw-data-export. CI now runs the local kit contract; live receipts remain explicitly authorized and separate.
• The preserved Action Receipt Pack index is docs/openclaw-action-receipt-pack.md, with fixtures under examples/openclaw/action-cards, skills under skills/openclaw/neura-action-card, and route identifiers including openclaw-send-message and openclaw-shell-command.

Contributing And Security

• Contribution workflow: CONTRIBUTING.md.
• Security reporting: SECURITY.md.
• Ownership: .github/CODEOWNERS.

Pull requests should remain synthetic, evidence-backed, and free of private customer data, credentials, tokens, or raw private payloads.

Repository Map

schemas/     Public machine-readable contract mirrors
examples/    Golden path plus specialized integration proofs
skills/      Reusable Action Card and authority-review skills
scripts/     Deterministic verification helpers
tests/       Focused contract tests
docs/        Current proof documentation and preserved history

Boundary

This repository demonstrates how to review agent actions and preserve decision evidence. It does not provide public production credentials, grant authority by itself, execute downstream tools, certify third-party systems, or prove customer adoption.

The complete pre-compression README is preserved at docs/history/README-through-2026-07-21.md.