@vellumai/assistant 0.7.0 → 0.7.1

package/src/providers/call-site-routing.ts

@@ -19,6 +19,8 @@
  * stable identity.
  */
 
+import { AsyncLocalStorage } from "node:async_hooks";
+
 import { resolveCallSiteConfig } from "../config/llm-resolver.js";
 import { getConfig } from "../config/loader.js";
 import type {
@@ -30,14 +32,28 @@ import type {
 } from "./types.js";
 
 export class CallSiteRoutingProvider implements Provider {
-  public readonly name: string;
   public readonly tokenEstimationProvider?: string;
 
+  // Per-call async context that tracks which provider is currently executing.
+  // Using AsyncLocalStorage instead of a plain instance field means concurrent
+  // sendMessage calls (e.g. the main agent turn and a title-generation call
+  // both in-flight at the same time on the same provider instance) each see
+  // their own value — no clobbering, no premature clear.
+  //
+  // During sendMessage, emitLlmCallStartedIfNeeded reads provider.name on the
+  // first text_delta (before the response completes). The getter below returns
+  // the async-context value so streaming trace events carry the routed
+  // provider's name, not the default's.
+  private readonly _activeProviderContext = new AsyncLocalStorage<string>();
+
+  get name(): string {
+    return this._activeProviderContext.getStore() ?? this.defaultProvider.name;
+  }
+
   constructor(
     private readonly defaultProvider: Provider,
     private readonly getProviderByName: (name: string) => Provider | undefined,
   ) {
-    this.name = defaultProvider.name;
     this.tokenEstimationProvider = defaultProvider.tokenEstimationProvider;
   }
 
@@ -48,7 +64,30 @@ export class CallSiteRoutingProvider implements Provider {
     options?: SendMessageOptions,
   ): Promise<ProviderResponse> {
     const target = this.selectProvider(options);
-    return target.sendMessage(messages, tools, systemPrompt, options);
+    const isRouted = target !== this.defaultProvider;
+
+    const doSend = async (): Promise<ProviderResponse> => {
+      const response = await target.sendMessage(
+        messages,
+        tools,
+        systemPrompt,
+        options,
+      );
+      // Also stamp actualProvider on the response so that handleUsage /
+      // llm_call_finished (which read event.actualProvider, not provider.name)
+      // attribute the call to the right provider.
+      if (isRouted && response.actualProvider == null) {
+        return { ...response, actualProvider: target.name };
+      }
+      return response;
+    };
+
+    // Run inside the async context so that any code reading provider.name
+    // during streaming (e.g. emitLlmCallStartedIfNeeded on text_delta) sees
+    // the routed provider's name for this specific call, not the default.
+    return isRouted
+      ? this._activeProviderContext.run(target.name, doSend)
+      : doSend();
   }
 
   /**

package/src/providers/gemini/client.ts

@@ -30,10 +30,20 @@ const GEMINI_CONTEXT_OVERFLOW_TOKEN_PATTERNS =
 const GEMINI_3_UNSIGNED_TOOL_CALL_THOUGHT_SIGNATURE =
   "context_engineering_is_the_way_to_go";
 
-export function isGemini3Model(model: string): boolean {
+function isGemini3Model(model: string): boolean {
   return model.startsWith("gemini-3") || model.startsWith("models/gemini-3");
 }
 
+function stripGeminiHttpOptions(
+  config: genai.GenerateContentConfig,
+): genai.GenerateContentConfig {
+  const { httpOptions: _httpOptions, ...rest } =
+    config as genai.GenerateContentConfig & {
+      httpOptions?: unknown;
+    };
+  return rest;
+}
+
 /**
  * Detect Gemini's context-overflow signals on an `ApiError`. Gemini surfaces
  * this condition via its "RESOURCE_EXHAUSTED" category. The Gemini SDK's
@@ -162,6 +172,9 @@ export class GeminiProvider implements Provider {
     const configObj = config as Record<string, unknown> | undefined;
     const maxTokens = configObj?.max_tokens as number | undefined;
     const modelOverride = configObj?.model as string | undefined;
+    const usageAttributionHeaders = configObj?.usageAttributionHeaders as
+      | Record<string, string>
+      | undefined;
     const activeModel = modelOverride ?? this.model;
 
     try {
@@ -193,6 +206,9 @@ export class GeminiProvider implements Provider {
       const { signal: timeoutSignal, cleanup: cleanupTimeout } =
         createStreamTimeout(this.streamTimeoutMs, signal);
       geminiConfig.abortSignal = timeoutSignal;
+      if (usageAttributionHeaders) {
+        geminiConfig.httpOptions = { headers: usageAttributionHeaders };
+      }
 
       // Accumulate from streaming chunks
       let fullText = "";
@@ -296,7 +312,7 @@ export class GeminiProvider implements Provider {
       const rawRequest = {
         model: activeModel,
         contents: geminiContents,
-        config: geminiConfig,
+        config: stripGeminiHttpOptions(geminiConfig),
       };
       const rawResponse = {
         model: responseModel,

package/src/providers/managed-proxy/context.ts

@@ -92,8 +92,3 @@ export async function managedFallbackEnabledFor(
   if (!meta?.managed) return false;
   return await hasManagedProxyPrereqs();
 }
-
-/** @internal Test-only: reset the cached managed-proxy-enabled flag. */
-export function _resetManagedProxyEnabledCache(): void {
-  _managedProxyEnabled = false;
-}

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);