@vellumai/assistant 0.7.0 → 0.7.2

package/src/providers/model-catalog.ts

@@ -1,8 +1,16 @@
+export type LongContextMode =
+  | "native-model"
+  | "provider-request-option"
+  | "unsupported";
+
 export interface CatalogModel {
   id: string;
   displayName: string;
   contextWindowTokens?: number;
   maxOutputTokens?: number;
+  defaultContextWindowTokens?: number;
+  longContextPricingThresholdTokens?: number;
+  longContextMode?: LongContextMode;
   supportsThinking?: boolean;
   supportsCaching?: boolean;
   supportsVision?: boolean;
@@ -15,6 +23,31 @@ export interface CatalogModel {
   };
 }
 
+const DEFAULT_CONTEXT_WINDOW_TOKENS = 200000;
+const OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS = 272000;
+
+function catalogModel(model: CatalogModel): CatalogModel {
+  const configuredDefaultContextWindowTokens =
+    model.defaultContextWindowTokens ?? DEFAULT_CONTEXT_WINDOW_TOKENS;
+  const defaultContextWindowTokens =
+    model.contextWindowTokens === undefined
+      ? configuredDefaultContextWindowTokens
+      : Math.min(
+          configuredDefaultContextWindowTokens,
+          model.contextWindowTokens,
+        );
+
+  return {
+    ...model,
+    defaultContextWindowTokens,
+    longContextMode:
+      model.longContextMode ??
+      ((model.contextWindowTokens ?? 0) > DEFAULT_CONTEXT_WINDOW_TOKENS
+        ? "native-model"
+        : "unsupported"),
+  };
+}
+
 export interface ProviderCatalogEntry {
   id: string;
   displayName: string;
@@ -33,8 +66,26 @@ export interface ProviderCatalogEntry {
   };
 }
 
-/** Single source of truth for all inference provider metadata and models. */
-export const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
+/**
+ * Canonical assistant catalog for inference provider metadata and models.
+ * `meta/llm-provider-catalog.json` mirrors the client-facing subset and is
+ * kept in parity by `llm-catalog-parity.test.ts`; native-client fallbacks
+ * mirror only the startup-critical display/setup/context metadata.
+ *
+ * Model limits verified 2026-04-30 against official provider docs:
+ * - Anthropic model overview and context window docs:
+ *   https://platform.claude.com/docs/en/about-claude/models/overview
+ *   https://platform.claude.com/docs/en/build-with-claude/context-windows
+ * - OpenAI model comparison and model detail docs:
+ *   https://developers.openai.com/api/docs/models/compare
+ *   https://developers.openai.com/api/docs/models
+ * - Google Gemini model docs:
+ *   https://ai.google.dev/gemini-api/docs/models
+ *
+ * contextWindowTokens is the maximum known input context. maxOutputTokens is
+ * the maximum standard synchronous Messages/Responses output limit.
+ */
+const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
   {
     id: "anthropic",
     displayName: "Anthropic",
@@ -52,8 +103,9 @@ export const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
       {
         id: "claude-opus-4-7",
         displayName: "Claude Opus 4.7",
-        contextWindowTokens: 200000,
-        maxOutputTokens: 32000,
+        contextWindowTokens: 1000000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens: 200000,
         supportsThinking: true,
         supportsCaching: true,
         supportsVision: true,
@@ -68,8 +120,9 @@ export const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
       {
         id: "claude-opus-4-6",
         displayName: "Claude Opus 4.6",
-        contextWindowTokens: 200000,
-        maxOutputTokens: 32000,
+        contextWindowTokens: 1000000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens: 200000,
         supportsThinking: true,
         supportsCaching: true,
         supportsVision: true,
@@ -84,8 +137,9 @@ export const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
       {
         id: "claude-sonnet-4-6",
         displayName: "Claude Sonnet 4.6",
-        contextWindowTokens: 200000,
+        contextWindowTokens: 1000000,
         maxOutputTokens: 64000,
+        longContextPricingThresholdTokens: 200000,
         supportsThinking: true,
         supportsCaching: true,
         supportsVision: true,
@@ -101,7 +155,7 @@ export const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         id: "claude-haiku-4-5-20251001",
         displayName: "Claude Haiku 4.5",
         contextWindowTokens: 200000,
-        maxOutputTokens: 16000,
+        maxOutputTokens: 64000,
         supportsThinking: true,
         supportsCaching: true,
         supportsVision: true,
@@ -137,6 +191,8 @@ export const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         displayName: "GPT-5.5",
         contextWindowTokens: 1050000,
         maxOutputTokens: 128000,
+        longContextPricingThresholdTokens:
+          OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
         supportsThinking: true,
         supportsCaching: true,
         supportsVision: true,
@@ -147,11 +203,29 @@ export const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
           cacheReadPer1mTokens: 0.5,
         },
       },
+      {
+        id: "gpt-5.5-pro",
+        displayName: "GPT-5.5 Pro",
+        contextWindowTokens: 1050000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens:
+          OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
+        supportsThinking: true,
+        supportsCaching: true,
+        supportsVision: true,
+        supportsToolUse: true,
+        pricing: {
+          inputPer1mTokens: 30.0,
+          outputPer1mTokens: 180.0,
+        },
+      },
       {
         id: "gpt-5.4",
         displayName: "GPT-5.4",
-        contextWindowTokens: 400000,
+        contextWindowTokens: 1050000,
         maxOutputTokens: 128000,
+        longContextPricingThresholdTokens:
+          OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
         supportsThinking: true,
         supportsCaching: true,
         supportsVision: true,
@@ -187,9 +261,9 @@ export const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         supportsVision: true,
         supportsToolUse: true,
         pricing: {
-          inputPer1mTokens: 0.5,
-          outputPer1mTokens: 3.0,
-          cacheReadPer1mTokens: 0.05,
+          inputPer1mTokens: 0.75,
+          outputPer1mTokens: 4.5,
+          cacheReadPer1mTokens: 0.075,
         },
       },
       {
@@ -232,6 +306,7 @@ export const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         displayName: "Gemini 3.1 Pro Preview",
         contextWindowTokens: 1048576,
         maxOutputTokens: 65536,
+        longContextPricingThresholdTokens: 200000,
         supportsThinking: true,
         supportsCaching: true,
         supportsVision: true,
@@ -247,6 +322,7 @@ export const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         displayName: "Gemini 3.1 Pro Preview (Custom Tools)",
         contextWindowTokens: 1048576,
         maxOutputTokens: 65536,
+        longContextPricingThresholdTokens: 200000,
         supportsThinking: true,
         supportsCaching: true,
         supportsVision: true,
@@ -320,8 +396,9 @@ export const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
       {
         id: "gemini-2.5-pro",
         displayName: "Gemini 2.5 Pro",
-        contextWindowTokens: 2000000,
+        contextWindowTokens: 1048576,
         maxOutputTokens: 65536,
+        longContextPricingThresholdTokens: 200000,
         supportsThinking: true,
         supportsCaching: true,
         supportsVision: true,
@@ -426,8 +503,9 @@ export const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
       {
         id: "anthropic/claude-opus-4.7",
         displayName: "Claude Opus 4.7",
-        contextWindowTokens: 200000,
-        maxOutputTokens: 32000,
+        contextWindowTokens: 1000000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens: 200000,
         supportsThinking: true,
         supportsCaching: true,
         supportsVision: true,
@@ -442,8 +520,9 @@ export const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
       {
         id: "anthropic/claude-opus-4.6",
         displayName: "Claude Opus 4.6",
-        contextWindowTokens: 200000,
-        maxOutputTokens: 32000,
+        contextWindowTokens: 1000000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens: 200000,
         supportsThinking: true,
         supportsCaching: true,
         supportsVision: true,
@@ -458,8 +537,9 @@ export const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
       {
         id: "anthropic/claude-sonnet-4.6",
         displayName: "Claude Sonnet 4.6",
-        contextWindowTokens: 200000,
+        contextWindowTokens: 1000000,
         maxOutputTokens: 64000,
+        longContextPricingThresholdTokens: 200000,
         supportsThinking: true,
         supportsCaching: true,
         supportsVision: true,
@@ -475,7 +555,7 @@ export const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         id: "anthropic/claude-haiku-4.5",
         displayName: "Claude Haiku 4.5",
         contextWindowTokens: 200000,
-        maxOutputTokens: 16000,
+        maxOutputTokens: 64000,
         supportsThinking: true,
         supportsCaching: true,
         supportsVision: true,
@@ -677,6 +757,12 @@ export const PROVIDER_CATALOG: ProviderCatalogEntry[] = [
   },
 ];
 
+export const PROVIDER_CATALOG: ProviderCatalogEntry[] =
+  RAW_PROVIDER_CATALOG.map((entry) => ({
+    ...entry,
+    models: entry.models.map(catalogModel),
+  }));
+
 /** Check if a model ID is in the catalog for a given provider. */
 export function isModelInCatalog(provider: string, modelId: string): boolean {
   const entry = PROVIDER_CATALOG.find((p) => p.id === provider);

package/src/providers/openai/chat-completions-provider.ts

@@ -139,6 +139,9 @@ export class OpenAIChatCompletionsProvider implements Provider {
     const maxTokens = configObj?.max_tokens as number | undefined;
     const modelOverride = configObj?.model as string | undefined;
     const effort = configObj?.effort as string | undefined;
+    const usageAttributionHeaders = configObj?.usageAttributionHeaders as
+      | Record<string, string>
+      | undefined;
 
     try {
       const openaiMessages = this.toOpenAIMessages(messages, systemPrompt);
@@ -196,6 +199,9 @@ export class OpenAIChatCompletionsProvider implements Provider {
       try {
         const stream = await this.client.chat.completions.create(params, {
           signal: timeoutSignal,
+          ...(usageAttributionHeaders
+            ? { headers: usageAttributionHeaders }
+            : {}),
         });
 
         for await (const chunk of stream) {

package/src/providers/openai/responses-provider.ts

@@ -130,6 +130,9 @@ export class OpenAIResponsesProvider implements Provider {
     const modelOverride = configObj?.model as string | undefined;
     const effort = configObj?.effort as string | undefined;
     const verbosity = configObj?.verbosity as string | undefined;
+    const usageAttributionHeaders = configObj?.usageAttributionHeaders as
+      | Record<string, string>
+      | undefined;
 
     try {
       const input = this.toResponsesInput(messages);
@@ -222,11 +225,14 @@ export class OpenAIResponsesProvider implements Provider {
         const responsesApi = this.client.responses as unknown as {
           stream(
             p: Record<string, unknown>,
-            o: { signal: AbortSignal },
+            o: { signal: AbortSignal; headers?: Record<string, string> },
           ): AsyncIterable<ResponsesStreamEvent>;
         };
         const stream = responsesApi.stream(params, {
           signal: timeoutSignal,
+          ...(usageAttributionHeaders
+            ? { headers: usageAttributionHeaders }
+            : {}),
         });
 
         for await (const event of stream) {

package/src/providers/provider-send-message.ts

@@ -13,12 +13,13 @@ import type {
   Message,
   Provider,
   ProviderResponse,
+  SendMessageOptions,
+  ToolDefinition,
   ToolUseContent,
 } from "./types.js";
 
 // Re-export the typed context-overflow error so callsites that dispatch on
 // this category do not need to reach into `./types.js` directly.
-export { ContextOverflowError, isContextOverflowError } from "./types.js";
 
 export interface ConfiguredProviderResult {
   provider: Provider;
@@ -33,6 +34,44 @@ export interface ConfiguredProviderResult {
  */
 let lazyInitPromise: Promise<void> | null = null;
 
+export class CallSiteConfiguredProvider implements Provider {
+  public readonly name: string;
+  public readonly tokenEstimationProvider?: string;
+
+  constructor(
+    private readonly inner: Provider,
+    private readonly callSite: LLMCallSite,
+    private readonly overrideProfile?: string,
+  ) {
+    this.name = inner.name;
+    this.tokenEstimationProvider = inner.tokenEstimationProvider;
+  }
+
+  sendMessage(
+    messages: Message[],
+    tools?: ToolDefinition[],
+    systemPrompt?: string,
+    options?: SendMessageOptions,
+  ): Promise<ProviderResponse> {
+    const config = options?.config;
+    if (config?.callSite) {
+      return this.inner.sendMessage(messages, tools, systemPrompt, options);
+    }
+
+    return this.inner.sendMessage(messages, tools, systemPrompt, {
+      ...options,
+      config: {
+        ...config,
+        callSite: this.callSite,
+        ...(config?.overrideProfile === undefined &&
+        this.overrideProfile !== undefined
+          ? { overrideProfile: this.overrideProfile }
+          : {}),
+      },
+    });
+  }
+}
+
 /**
  * Resolve the configured provider with full selection metadata.
  * If providers haven't been initialized yet (e.g. non-daemon code paths),
@@ -77,7 +116,11 @@ export async function resolveConfiguredProvider(
   try {
     const provider = getProvider(inferenceProvider);
     return {
-      provider,
+      provider: new CallSiteConfiguredProvider(
+        provider,
+        callSite,
+        opts.overrideProfile,
+      ),
       configuredProviderName: inferenceProvider,
     };
   } catch {

package/src/providers/ratelimit.ts

@@ -12,7 +12,13 @@ import type {
 const log = getLogger("rate-limit");
 
 export class RateLimitProvider implements Provider {
-  public readonly name: string;
+  // Delegate name dynamically so that wrapper providers (e.g.
+  // CallSiteRoutingProvider) whose name getter reflects per-call async context
+  // (AsyncLocalStorage) are reached correctly during streaming — rather than
+  // returning a stale snapshot captured at construction time.
+  get name(): string {
+    return this.inner.name;
+  }
 
   get tokenEstimationProvider(): string | undefined {
     return this.inner.tokenEstimationProvider;
@@ -25,7 +31,6 @@ export class RateLimitProvider implements Provider {
     private readonly config: RateLimitConfig,
     sharedRequestTimestamps?: number[],
   ) {
-    this.name = inner.name;
     this.requestTimestamps = sharedRequestTimestamps ?? [];
   }
 

package/src/providers/registry.ts

@@ -14,12 +14,13 @@ import { OpenAIResponsesProvider } from "./openai/client.js";
 import { OpenRouterProvider } from "./openrouter/client.js";
 import { RetryProvider } from "./retry.js";
 import type { Provider } from "./types.js";
+import { UsageTrackingProvider } from "./usage-tracking.js";
 
 const providers = new Map<string, Provider>();
 const routingSources = new Map<string, "user-key" | "managed-proxy">();
 
 function registerProvider(name: string, provider: Provider): void {
-  providers.set(name, provider);
+  providers.set(name, new UsageTrackingProvider(provider));
 }
 
 export function getProvider(name: string): Provider {
@@ -40,14 +41,6 @@ export function getProviderRoutingSource(
   return routingSources.get(name);
 }
 
-/**
- * Return the default model for a given provider name.
- * Falls back to the Anthropic default if the provider name is unknown.
- */
-export function getDefaultModel(providerName: string): string {
-  return getProviderDefaultModel(providerName);
-}
-
 export interface ProvidersConfig {
   services: {
     inference: {
@@ -171,6 +164,10 @@ export async function initializeProviders(
             ? { baseURL: anthropicCreds.baseURL }
             : {}),
         }),
+        {
+          forwardUsageAttributionHeaders:
+            anthropicCreds.source === "managed-proxy",
+        },
       ),
     );
     routingSources.set("anthropic", anthropicCreds.source);
@@ -188,6 +185,10 @@ export async function initializeProviders(
           streamTimeoutMs,
           ...(openaiCreds.baseURL ? { baseURL: openaiCreds.baseURL } : {}),
         }),
+        {
+          forwardUsageAttributionHeaders:
+            openaiCreds.source === "managed-proxy",
+        },
       ),
     );
     routingSources.set("openai", openaiCreds.source);
@@ -206,6 +207,10 @@ export async function initializeProviders(
             ? { managedBaseUrl: geminiCreds.baseURL }
             : {}),
         }),
+        {
+          forwardUsageAttributionHeaders:
+            geminiCreds.source === "managed-proxy",
+        },
       ),
     );
     routingSources.set("gemini", geminiCreds.source);

package/src/providers/retry.ts

@@ -1,5 +1,9 @@
 import { resolveCallSiteConfig } from "../config/llm-resolver.js";
 import { getConfig } from "../config/loader.js";
+import {
+  resolveUsageAttribution,
+  sanitizeUsageMetadataValue,
+} from "../usage/attribution.js";
 import { ProviderError } from "../util/errors.js";
 import { getLogger } from "../util/logger.js";
 import {
@@ -24,6 +28,14 @@ import {
 
 const log = getLogger("retry");
 
+const USAGE_ATTRIBUTION_HEADER_NAMES = {
+  callSite: "X-Vellum-LLM-Call-Site",
+  inferenceProfile: "X-Vellum-Inference-Profile",
+  inferenceProfileSource: "X-Vellum-Inference-Profile-Source",
+  resolvedProvider: "X-Vellum-Resolved-Provider",
+  resolvedModel: "X-Vellum-Resolved-Model",
+} as const;
+
 /** Providers that support the `effort` config (extended thinking / reasoning). */
 const EFFORT_SUPPORTED_PROVIDERS = new Set([
   "anthropic",
@@ -118,12 +130,18 @@ function isRetryableError(error: unknown): boolean {
 function normalizeSendMessageOptions(
   providerName: string,
   options?: SendMessageOptions,
+  normalizeOptions: { forwardUsageAttributionHeaders?: boolean } = {},
 ): SendMessageOptions | undefined {
   const config = options?.config;
   if (!config) return options;
 
   const nextConfig: Record<string, unknown> = { ...config };
 
+  // Internal metadata must be derived here, not accepted from callers, and it
+  // must never leak into provider JSON request bodies.
+  delete nextConfig.usageAttributionHeaders;
+  delete nextConfig.usageTracking;
+
   // `overrideProfile` is a routing/resolution-time concern (consumed by the
   // resolver below and `CallSiteRoutingProvider`'s provider selection); it is
   // not a wire-format field. Strip unconditionally so it never leaks into
@@ -134,6 +152,10 @@ function normalizeSendMessageOptions(
     const resolved = resolveCallSiteConfig(config.callSite, getConfig().llm, {
       overrideProfile: config.overrideProfile,
     });
+    const attribution = resolveUsageAttribution({
+      callSite: config.callSite,
+      overrideProfile: config.overrideProfile,
+    });
 
     const explicitModel =
       typeof config.model === "string" && config.model.trim().length > 0
@@ -143,6 +165,18 @@ function normalizeSendMessageOptions(
     // Routing key is consumed by the resolver above and must not leak
     // downstream as a wire-format field.
     delete nextConfig.callSite;
+    if (normalizeOptions.forwardUsageAttributionHeaders === true) {
+      const usageAttributionHeaders = buildUsageAttributionHeaders({
+        callSite: attribution.callSite,
+        appliedProfile: attribution.appliedProfile,
+        profileSource: attribution.profileSource,
+        resolvedProvider: attribution.resolvedProvider,
+        resolvedModel: attribution.resolvedModel,
+      });
+      if (Object.keys(usageAttributionHeaders).length > 0) {
+        nextConfig.usageAttributionHeaders = usageAttributionHeaders;
+      }
+    }
 
     // Apply resolved values, letting per-call explicit fields win where set.
     nextConfig.model = explicitModel ?? resolved.model;
@@ -194,12 +228,11 @@ function normalizeSendMessageOptions(
       nextConfig.openrouter = { only: resolved.openrouter.only };
     }
     // `contextWindow` and `provider` are server-side concerns, not provider
-    // request parameters: `contextWindow` is consumed by the agent loop's
-    // overflow recovery and the conversation manager directly from
-    // `config.llm.default.contextWindow.*`; `provider` selection is handled
-    // by `CallSiteRoutingProvider` upstream. Forwarding them as per-call
-    // config leaks unknown fields into provider request bodies — Anthropic
-    // (and other strict-schema clients) reject the request with
+    // request parameters: effective context is resolved per call site/profile
+    // by the agent/conversation path, while `provider` selection is handled by
+    // `CallSiteRoutingProvider` upstream. Forwarding them as per-call config
+    // leaks unknown fields into provider request bodies — Anthropic (and other
+    // strict-schema clients) reject the request with
     // "Extra inputs are not permitted".
   }
 
@@ -274,6 +307,55 @@ function normalizeSendMessageOptions(
   };
 }
 
+function buildUsageAttributionHeaders(input: {
+  callSite: string | null;
+  appliedProfile: string | null;
+  profileSource: string;
+  resolvedProvider: string;
+  resolvedModel: string;
+}): Record<string, string> {
+  const headers: Record<string, string> = {};
+  addSanitizedHeader(
+    headers,
+    USAGE_ATTRIBUTION_HEADER_NAMES.callSite,
+    input.callSite,
+  );
+  addSanitizedHeader(
+    headers,
+    USAGE_ATTRIBUTION_HEADER_NAMES.inferenceProfile,
+    input.appliedProfile,
+  );
+  if (input.appliedProfile) {
+    addSanitizedHeader(
+      headers,
+      USAGE_ATTRIBUTION_HEADER_NAMES.inferenceProfileSource,
+      input.profileSource,
+    );
+  }
+  addSanitizedHeader(
+    headers,
+    USAGE_ATTRIBUTION_HEADER_NAMES.resolvedProvider,
+    input.resolvedProvider,
+  );
+  addSanitizedHeader(
+    headers,
+    USAGE_ATTRIBUTION_HEADER_NAMES.resolvedModel,
+    input.resolvedModel,
+  );
+  return headers;
+}
+
+function addSanitizedHeader(
+  headers: Record<string, string>,
+  name: string,
+  value: unknown,
+): void {
+  const sanitized = sanitizeUsageMetadataValue(value);
+  if (sanitized != null) {
+    headers[name] = sanitized;
+  }
+}
+
 /**
  * `RetryProvider` sets `retriesExhausted = true` on the final thrown error
  * when the retry loop burned through all attempts against a retryable error
@@ -289,7 +371,10 @@ export class RetryProvider implements Provider {
     return this.inner.tokenEstimationProvider;
   }
 
-  constructor(private readonly inner: Provider) {
+  constructor(
+    private readonly inner: Provider,
+    private readonly options: { forwardUsageAttributionHeaders?: boolean } = {},
+  ) {
     this.name = inner.name;
   }
 
@@ -302,7 +387,10 @@ export class RetryProvider implements Provider {
     let lastError: unknown;
     let didRetry = false;
 
-    const normalizedOptions = normalizeSendMessageOptions(this.name, options);
+    const normalizedOptions = normalizeSendMessageOptions(this.name, options, {
+      forwardUsageAttributionHeaders:
+        this.options.forwardUsageAttributionHeaders === true,
+    });
 
     for (let attempt = 0; attempt <= DEFAULT_MAX_RETRIES; attempt++) {
       try {

package/src/providers/speech-to-text/provider-catalog.ts

@@ -34,7 +34,7 @@ import type {
  *   Twilio. A `<Stream>` media-stream is opened and the daemon transcribes
  *   audio server-side via the provider's batch API.
  */
-export type TelephonyStrategyKind =
+type TelephonyStrategyKind =
   | "conversation-relay-native"
   | "media-stream-custom";
 
@@ -51,7 +51,7 @@ export type TwilioNativeProvider = "Deepgram" | "Google";
  * ConversationRelay. Only present when `strategyKind` is
  * `"conversation-relay-native"`.
  */
-export interface TwilioNativeMapping {
+interface TwilioNativeMapping {
   /** Twilio-native provider name for the TwiML `transcriptionProvider` attribute. */
   readonly provider: TwilioNativeProvider;
   /**
@@ -69,7 +69,7 @@ export interface TwilioNativeMapping {
  * The telephony routing resolver reads these fields from the catalog
  * instead of maintaining its own hardcoded maps.
  */
-export interface TelephonyRouting {
+interface TelephonyRouting {
   /** Which Twilio call-setup strategy this provider uses. */
   readonly strategyKind: TelephonyStrategyKind;
   /**
@@ -84,10 +84,10 @@ export interface TelephonyRouting {
 // ---------------------------------------------------------------------------
 
 /** How the provider's credentials are configured by the user. */
-export type SttSetupMode = "api-key" | "cli";
+type SttSetupMode = "api-key" | "cli";
 
 /** Guide for obtaining API credentials from a provider. */
-export interface SttCredentialsGuide {
+interface SttCredentialsGuide {
   readonly description: string;
   readonly url: string;
   readonly linkLabel: string;
@@ -100,7 +100,7 @@ export interface SttCredentialsGuide {
 /**
  * Metadata for a single STT provider.
  */
-export interface SttProviderEntry {
+interface SttProviderEntry {
   /** Canonical provider identifier (must match an {@link SttProviderId} variant). */
   readonly id: SttProviderId;
 
@@ -262,8 +262,7 @@ const CATALOG: ReadonlyMap<SttProviderId, SttProviderEntry> = new Map<
       subtitle:
         "High-accuracy speech-to-text powered by OpenAI Whisper. Requires an OpenAI API key.",
       setupMode: "api-key",
-      setupHint:
-        "Enter your OpenAI API key to enable Whisper transcription.",
+      setupHint: "Enter your OpenAI API key to enable Whisper transcription.",
       credentialProvider: "openai",
       supportedBoundaries: new Set<SttBoundaryId>([
         "daemon-batch",