Signed decision evidence for Vercel AI SDK.

Install the native Vercel AI SDK integration.

npm install loopgrid-vercel-ai-sdk ai

Version 0.1.0 is public on npm. The integration depends on @cybertechsoft/loopgrid and targets AI SDK 7 on Node.js 22 or newer. npm Trusted Publishing is configured for future releases.

Bind one consequential decision to one AI SDK call.

import { generateText } from 'ai';
import { LoopGridAISDK } from 'loopgrid-vercel-ai-sdk';

const loopgrid = new LoopGridAISDK({
  baseUrl: 'http://127.0.0.1:8000',
  workspaceId: 'default',
});

const { decisionId } = await loopgrid.startDecision({
  decisionType: 'customer_refund',
  agent: { id: 'refund-agent', version: '1.0.0' },
  authority: { acting_for: 'Support', scope: ['refund:create'] },
  model: { provider: 'configured-provider', name: 'configured-model' },
  context: { prompt_version: 'refund-v3' },
  proposedAction: { tool: 'refund', amount: 25, currency: 'USD' },
  policy: {
    policy_id: 'refund-policy',
    version: '3',
    decision: 'auto_allowed',
  },
});

const result = await generateText({
  model,
  prompt,
  tools,
  telemetry: loopgrid.telemetryOptions(decisionId, {
    functionId: 'refund-agent',
  }),
});

await loopgrid.flush(decisionId);
loopgrid.assertHealthy();

// Application-owned fact observed from the downstream system:
await loopgrid.recordOutcome(decisionId, {
  status: 'settled',
  external_id: 'refund_123',
});

The telemetry object is decision-scoped instead of relying on a global mutable “current decision” value, so concurrent AI SDK calls remain isolated.

Use AI SDK Telemetry instead of wrapping generateText().

AI SDK owns approval mechanics. LoopGrid preserves the evidence.

AI SDK requests approval
  → tool has not executed

application records explicit reviewer
  → loopgrid.recordHumanReview(...)
  → human_approved

AI SDK resumes approved tool
  → tool_requested
  → tool_executed

application observes downstream result
  → loopgrid.recordOutcome(...)
  → outcome_observed
  → evidence_complete
  → verify valid:true

LoopGrid does not infer delegated authority from tool availability, infer policy from AI SDK approval configuration, infer reviewer identity from an approval response, or treat a successful tool callback as proof of the final business outcome. Telemetry callbacks are observational rather than a policy-enforcement boundary.

Commit to model/tool content without storing it by default.

AI SDK telemetry inputs and outputs also default to disabled. A dedicated regression test against ai@7.0.122 verifies that context excluded by telemetry allowlists is not exposed on the diagnostics tracing channel. Transport failures are retained and surfaced through flush() and assertHealthy(); applications that require evidence before execution should enforce that requirement in their own control logic.

Validated through the actual AI SDK runtime and a real LoopGrid Core E2E.

The release E2Es used a deterministic AI SDK test model and sandbox action, so no model API key or real-money movement was required. Verification establishes the integrity/provenance of captured evidence; it does not prove that the underlying decision was correct, that every upstream statement was true, or that the system is legally compliant.