@tuvren/provider-bridge-ai-sdk

Bridges a Vercel AI SDK LanguageModelV3 into a Tuvren TuvrenProvider, so the runtime's runner can call any AI-SDK-backed model through the neutral provider contract (generate / stream).

Secret Isolation — Edge Confinement (ADR-044)

Provider credentials are confined to the integration edge. They are accepted only by the provider bridge (and the underlying AI SDK model) at request time and never propagate onto any observable, persisted, or replayable runtime surface.

Credentials accepted at this edge:

• the API key baked into the AI SDK provider/model instance you pass to the bridge (e.g. via the provider's own constructor or environment), and
• any secret-bearing defaultHeaders supplied to the bridge.

These values authenticate the model request and nothing else. They are never:

• written to kernel records or any durable state,
• placed on canonical stream events,
• placed on TelemetrySpan / TelemetryEvent attributes, or
• serialized into REPL transcripts.

Provider continuity artifacts (ADR-005; opaque continuation tokens) are non-secret by contract and must not carry credential material. The Kernel Boundary, Durable State Boundary, Telemetry & Observability Boundary, canonical event stream, and transcript surfaces are credential-free zones.

This guarantee is verified — not assumed. The secret-isolation conformance check set (KRT-BD004) configures a representative provider key, runs a turn, and uses a shared runner-owned secret-absence helper to recursively scan the persisted records, stream events, telemetry, and transcript for that secret and its common encoded variants.

Usage

import { createOpenAI } from "@ai-sdk/openai";
import { createAiSdkProviderBridge } from "@tuvren/provider-bridge-ai-sdk";

// The API key lives in the AI SDK provider — the integration edge.
const openai = createOpenAI({ apiKey: process.env.OPENAI_API_KEY! });
const provider = createAiSdkProviderBridge({ model: openai("gpt-4o") });

// `provider` is a TuvrenProvider; the key never leaves this edge.