@adjudicate/vercel-ai

Reference Vercel AI SDK (v5+) integration for the adjudicate decision kernel. A thin shim over @adjudicate/adapter-core — every load-bearing concern (tool-use loop, defer/confirm orchestration, audit + ledger wiring, REWRITE handling, confirmation-blob hash verification) lives upstream in adapter-core. This package owns one thing: mapping the AI SDK's wire shapes to the provider-neutral contracts.

Layout

| File | Role | |---|---| | src/bridge-vercel.ts | ProviderBridge<VercelMessage[]> against the AI SDK generateText surface | | src/renderer-vercel.ts | System-prompt + tool-name translation tuned for the AI SDK | | src/adapter.ts | createAdjudicatedAgent — wires bridge + renderer into adapter-core | | src/vercel-types.ts | Structural types mirroring the AI SDK shape (no hard dep) | | src/types.ts | Public surface — VercelHistory, options, agent type | | src/index.ts | Barrel — re-exports adapter-core ergonomics |

Quickstart

import { generateText } from "ai";
import { openai } from "@ai-sdk/openai";
import { installPack } from "@adjudicate/core";
import {
  createAdjudicatedAgent,
  createVercelPromptRenderer,
  createInMemoryConfirmationStore,
  createInMemoryDeferStore,
  createMemoryLedger,
} from "@adjudicate/vercel-ai";
import { paymentsPixPack } from "@adjudicate/pack-payments-pix";

const { pack } = installPack(paymentsPixPack);

const agent = createAdjudicatedAgent({
  pack,
  generateText,
  model: openai("gpt-4o"),
  maxTokens: 1024,
  renderer: createVercelPromptRenderer({
    packId: pack.id,
    toolSchemas: [], // the ToolSchema[] the model sees for this pack's intents
  }),
  deferStore: createInMemoryDeferStore(),
  confirmationStore: createInMemoryConfirmationStore(),
  ledger: createMemoryLedger(),
  executor: {
    invokeRead: async (name, input, state) => /* read tool */,
    invokeIntent: async (envelope, state) => /* side-effect */,
  },
});

const turn = await agent.send({
  sessionId: "s-1",
  userMessage: "Refund the iced-coffee charge.",
  state: { /* … */ },
  context: { /* … */ },
});

createAdjudicatedAgent returns the same { send, resume, confirm } surface as the OpenAI and Anthropic adapters. The kernel sees identical IntentEnvelopes either way; the choice of provider doesn't change which Decision outcomes are reachable.

Why no ai / @ai-sdk/* dependency?

The package accepts any callable satisfying VercelGenerateTextFn plus an opaque LanguageModel handle — a minimal structural interface. The official generateText from ai (>=5) satisfies it; mocks for tests do too. Three reasons:

1. No major-version pin on a fast-moving SDK.
2. Mocks remain trivial.
3. Adopters who already vendor a specific AI SDK version don't get a duplicate copy.

If you want the SDK, install it directly: pnpm add ai @ai-sdk/openai.

Unlike a Chat Completions client, the AI SDK exposes a FREE FUNCTION generateText(...) plus a LanguageModel instance rather than a stateful client object. The bridge injects the callable plus a model handle accordingly.

Determinism + replay

Every intent envelope crosses adjudicateAndAudit() from @adjudicate/core/kernel. The Vercel AI adapter cannot bypass it; the kernel still owns:

• The closed 6-valued Decision algebra.
• Replay suppression via the supplied Ledger.
• Canonical-JSON intentHash (excludes createdAt, includes nonce).
• Fail-closed semantics on throwing guards.
• Confirmation-blob tamper detection at resume.

Cross-provider parity is verified by tests/integration-pix.test.ts — the same canned conversation against the same PIX Pack reaches the same six Decision kinds, with the same audit-record counts and no withBasisAudit drift events.

Notable deltas vs the OpenAI adapter

• Tool-call input arrives ALREADY PARSED as a structured object — no JSON.parse / __raw fallback.
• Token usage maps 1:1 ({ inputTokens, outputTokens }) — no prompt_tokens / completion_tokens rename.
• The output cap field is maxOutputTokens (the v5 name).
• The SDK surface is the free function generateText(...) plus a LanguageModel handle, not a stateful client.
• Tool results use role: "tool" messages whose content is an array of tool-result parts.

Status

Shipped against adapter-core v0.6 and AI SDK v5. Surface stable; the structural VercelGenerateTextFn may expand if the AI SDK ships new request-side fields the adapter needs to pass through.