@vellumai/assistant 0.11.0 → 0.11.1-staging.2

package/src/providers/model-catalog.ts

@@ -50,6 +50,15 @@ export interface CatalogModel {
    * before dispatching. Implies `supportsThinking`.
    */
   adaptiveThinkingOnly?: boolean;
+  /**
+   * When true, the model predates adaptive thinking and rejects
+   * `thinking: { type: "adaptive" }` (Anthropic 400s such calls). It only
+   * supports the legacy `{ type: "enabled", budget_tokens }` form, which
+   * Vellum never sends. The daemon drops an enabled thinking config for these
+   * models before dispatching to the Anthropic wire. Mutually exclusive with
+   * `adaptiveThinkingOnly`.
+   */
+  adaptiveThinkingUnsupported?: boolean;
   supportsCaching?: boolean;
   supportsVision?: boolean;
   /**
@@ -296,6 +305,7 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         contextWindowTokens: 200000,
         maxOutputTokens: 64000,
         supportsThinking: true,
+        adaptiveThinkingUnsupported: true,
         supportsCaching: true,
         supportsVision: true,
         supportsToolUse: true,
@@ -312,6 +322,7 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         contextWindowTokens: 200000,
         maxOutputTokens: 64000,
         supportsThinking: true,
+        adaptiveThinkingUnsupported: true,
         supportsCaching: true,
         supportsVision: true,
         supportsToolUse: true,
@@ -328,6 +339,7 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         contextWindowTokens: 200000,
         maxOutputTokens: 64000,
         supportsThinking: true,
+        adaptiveThinkingUnsupported: true,
         supportsCaching: true,
         supportsVision: true,
         supportsToolUse: true,
@@ -404,17 +416,17 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         supportsToolUse: true,
         supportsPromptCacheBreakpoints: true,
         pricing: {
-          inputPer1mTokens: 2.5,
-          outputPer1mTokens: 15.0,
-          cacheWritePer1mTokens: 3.125,
-          cacheReadPer1mTokens: 0.25,
+          inputPer1mTokens: 2.0,
+          outputPer1mTokens: 12.0,
+          cacheWritePer1mTokens: 2.5,
+          cacheReadPer1mTokens: 0.2,
           tiers: [
             {
               inputTokenThreshold: OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
-              inputPer1mTokens: 5,
-              outputPer1mTokens: 22.5,
-              cacheWritePer1mTokens: 6.25,
-              cacheReadPer1mTokens: 0.5,
+              inputPer1mTokens: 4,
+              outputPer1mTokens: 18,
+              cacheWritePer1mTokens: 5,
+              cacheReadPer1mTokens: 0.4,
             },
           ],
         },
@@ -432,17 +444,17 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         supportsToolUse: true,
         supportsPromptCacheBreakpoints: true,
         pricing: {
-          inputPer1mTokens: 1.0,
-          outputPer1mTokens: 6.0,
-          cacheWritePer1mTokens: 1.25,
-          cacheReadPer1mTokens: 0.1,
+          inputPer1mTokens: 0.2,
+          outputPer1mTokens: 1.2,
+          cacheWritePer1mTokens: 0.25,
+          cacheReadPer1mTokens: 0.02,
           tiers: [
             {
               inputTokenThreshold: OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
-              inputPer1mTokens: 2,
-              outputPer1mTokens: 9,
-              cacheWritePer1mTokens: 2.5,
-              cacheReadPer1mTokens: 0.2,
+              inputPer1mTokens: 0.4,
+              outputPer1mTokens: 1.8,
+              cacheWritePer1mTokens: 0.5,
+              cacheReadPer1mTokens: 0.04,
             },
           ],
         },
@@ -881,20 +893,10 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
           cacheReadPer1mTokens: 0.26,
         },
       },
-      {
-        id: "accounts/fireworks/models/kimi-k2p5",
-        displayName: "Kimi K2.5",
-        contextWindowTokens: 256000,
-        maxOutputTokens: 32768,
-        supportsThinking: false,
-        supportsCaching: false,
-        supportsVision: false,
-        supportsToolUse: true,
-        pricing: {
-          inputPer1mTokens: 0.6,
-          outputPer1mTokens: 2.5,
-        },
-      },
+      // Kimi K2.5 (accounts/fireworks/models/kimi-k2p5) is intentionally
+      // absent: Fireworks serves it on-demand/dedicated only, so serverless
+      // chat/completions calls 404 ("not found, inaccessible, and/or not
+      // deployed").
       {
         id: "accounts/fireworks/models/minimax-m3",
         displayName: "MiniMax M3",
@@ -964,7 +966,7 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         },
       },
     ],
-    defaultModel: "accounts/fireworks/models/kimi-k2p5",
+    defaultModel: "accounts/fireworks/models/deepseek-v4-flash",
     apiKeyUrl: "https://fireworks.ai/account/api-keys",
     apiKeyPlaceholder: "fw_...",
   },
@@ -1150,6 +1152,7 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         contextWindowTokens: 200000,
         maxOutputTokens: 64000,
         supportsThinking: true,
+        adaptiveThinkingUnsupported: true,
         supportsCaching: true,
         supportsVision: true,
         supportsToolUse: true,
@@ -1166,6 +1169,7 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         contextWindowTokens: 200000,
         maxOutputTokens: 64000,
         supportsThinking: true,
+        adaptiveThinkingUnsupported: true,
         supportsCaching: true,
         supportsVision: true,
         supportsToolUse: true,
@@ -1182,6 +1186,7 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         contextWindowTokens: 200000,
         maxOutputTokens: 64000,
         supportsThinking: true,
+        adaptiveThinkingUnsupported: true,
         supportsCaching: true,
         supportsVision: true,
         supportsToolUse: true,
@@ -1201,6 +1206,10 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
       // `cacheCreationInputTokens`). Long-context (>272K input) is 2x input
       // / 1.5x output / 2x cache-read+write for the whole request, per
       // OpenAI's model cards.
+      //
+      // Rates are OpenRouter's own card (https://openrouter.ai/api/v1/models),
+      // which discounts Terra and Luna below OpenAI's direct list price; Sol
+      // matches direct pricing.
       {
         id: "openai/gpt-5.6-sol",
         displayName: "GPT-5.6 Sol",
@@ -1270,17 +1279,17 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         supportsToolUse: true,
         supportsPromptCacheBreakpoints: true,
         pricing: {
-          inputPer1mTokens: 2.5,
-          outputPer1mTokens: 15.0,
-          cacheWritePer1mTokens: 3.125,
-          cacheReadPer1mTokens: 0.25,
+          inputPer1mTokens: 1.0,
+          outputPer1mTokens: 6.0,
+          cacheWritePer1mTokens: 1.25,
+          cacheReadPer1mTokens: 0.1,
           tiers: [
             {
               inputTokenThreshold: OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
-              inputPer1mTokens: 5,
-              outputPer1mTokens: 22.5,
-              cacheWritePer1mTokens: 6.25,
-              cacheReadPer1mTokens: 0.5,
+              inputPer1mTokens: 2,
+              outputPer1mTokens: 9,
+              cacheWritePer1mTokens: 2.5,
+              cacheReadPer1mTokens: 0.2,
             },
           ],
         },
@@ -1298,17 +1307,17 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         supportsToolUse: true,
         supportsPromptCacheBreakpoints: true,
         pricing: {
-          inputPer1mTokens: 2.5,
-          outputPer1mTokens: 15.0,
-          cacheWritePer1mTokens: 3.125,
-          cacheReadPer1mTokens: 0.25,
+          inputPer1mTokens: 1.0,
+          outputPer1mTokens: 6.0,
+          cacheWritePer1mTokens: 1.25,
+          cacheReadPer1mTokens: 0.1,
           tiers: [
             {
               inputTokenThreshold: OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
-              inputPer1mTokens: 5,
-              outputPer1mTokens: 22.5,
-              cacheWritePer1mTokens: 6.25,
-              cacheReadPer1mTokens: 0.5,
+              inputPer1mTokens: 2,
+              outputPer1mTokens: 9,
+              cacheWritePer1mTokens: 2.5,
+              cacheReadPer1mTokens: 0.2,
             },
           ],
         },
@@ -1326,17 +1335,17 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         supportsToolUse: true,
         supportsPromptCacheBreakpoints: true,
         pricing: {
-          inputPer1mTokens: 1.0,
-          outputPer1mTokens: 6.0,
-          cacheWritePer1mTokens: 1.25,
-          cacheReadPer1mTokens: 0.1,
+          inputPer1mTokens: 0.1,
+          outputPer1mTokens: 0.6,
+          cacheWritePer1mTokens: 0.125,
+          cacheReadPer1mTokens: 0.01,
           tiers: [
             {
               inputTokenThreshold: OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
-              inputPer1mTokens: 2,
-              outputPer1mTokens: 9,
-              cacheWritePer1mTokens: 2.5,
-              cacheReadPer1mTokens: 0.2,
+              inputPer1mTokens: 0.2,
+              outputPer1mTokens: 0.9,
+              cacheWritePer1mTokens: 0.25,
+              cacheReadPer1mTokens: 0.02,
             },
           ],
         },
@@ -1354,17 +1363,17 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         supportsToolUse: true,
         supportsPromptCacheBreakpoints: true,
         pricing: {
-          inputPer1mTokens: 1.0,
-          outputPer1mTokens: 6.0,
-          cacheWritePer1mTokens: 1.25,
-          cacheReadPer1mTokens: 0.1,
+          inputPer1mTokens: 0.1,
+          outputPer1mTokens: 0.6,
+          cacheWritePer1mTokens: 0.125,
+          cacheReadPer1mTokens: 0.01,
           tiers: [
             {
               inputTokenThreshold: OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
-              inputPer1mTokens: 2,
-              outputPer1mTokens: 9,
-              cacheWritePer1mTokens: 2.5,
-              cacheReadPer1mTokens: 0.2,
+              inputPer1mTokens: 0.2,
+              outputPer1mTokens: 0.9,
+              cacheWritePer1mTokens: 0.25,
+              cacheReadPer1mTokens: 0.02,
             },
           ],
         },
@@ -1883,6 +1892,7 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
         contextWindowTokens: 200000,
         maxOutputTokens: 64000,
         supportsThinking: true,
+        adaptiveThinkingUnsupported: true,
         supportsCaching: true,
         supportsVision: true,
         supportsToolUse: true,
@@ -2181,6 +2191,16 @@ export function isModelInCatalog(provider: string, modelId: string): boolean {
   return entry?.models.some((m) => m.id === modelId) ?? false;
 }
 
+/** The `maxOutputTokens` declared for a (provider, model) catalog entry, if any. */
+export function catalogMaxOutputTokens(
+  provider: string,
+  modelId: string,
+): number | undefined {
+  return PROVIDER_CATALOG.find((p) => p.id === provider)?.models.find(
+    (m) => m.id === modelId,
+  )?.maxOutputTokens;
+}
+
 /**
  * Model IDs (across all catalog providers) flagged
  * `supportsPromptCacheBreakpoints`. Consumed by the OpenAI Responses
@@ -2236,6 +2256,27 @@ export function isAdaptiveThinkingOnlyModel(modelId: string): boolean {
   );
 }
 
+/**
+ * Whether the given model predates adaptive thinking and rejects
+ * `thinking: { type: "adaptive" }`, driven by the `adaptiveThinkingUnsupported`
+ * capability in the catalog. Matches the model ID across every provider (same
+ * pattern as {@link isAdaptiveThinkingOnlyModel}), and also matches the
+ * undated aliases Anthropic serves for dated catalog IDs (`claude-haiku-4-5`
+ * resolves the same model as `claude-haiku-4-5-20251001`), since profiles and
+ * the CLI accept either form.
+ */
+export function isAdaptiveThinkingUnsupportedModel(modelId: string): boolean {
+  const stripDateSuffix = (id: string): string => id.replace(/-\d{8}$/, "");
+  const normalized = stripDateSuffix(modelId);
+  return PROVIDER_CATALOG.some((p) =>
+    p.models.some(
+      (m) =>
+        m.adaptiveThinkingUnsupported === true &&
+        (m.id === modelId || stripDateSuffix(m.id) === normalized),
+    ),
+  );
+}
+
 /**
  * Whether a model's serving surface accepts OpenAI chat-completions
  * `input_audio` content parts, driven by the `supportsAudioInput` capability

package/src/providers/model-intents.ts

@@ -18,7 +18,7 @@ const PROVIDER_MODEL_INTENTS: Record<string, Record<ModelIntent, string>> = {
   },
   openai: {
     balanced: "gpt-5.4-mini",
-    "latency-optimized": "gpt-5.4-nano",
+    "latency-optimized": "gpt-5.6-luna",
     "quality-optimized": "gpt-5.4",
     "vision-optimized": "gpt-5.4",
   },
@@ -36,7 +36,7 @@ const PROVIDER_MODEL_INTENTS: Record<string, Record<ModelIntent, string>> = {
   },
   fireworks: {
     balanced: "accounts/fireworks/models/minimax-m3",
-    "latency-optimized": "accounts/fireworks/models/kimi-k2p5",
+    "latency-optimized": "accounts/fireworks/models/deepseek-v4-flash",
     "quality-optimized": "accounts/fireworks/models/kimi-k2p6",
     "vision-optimized": "accounts/fireworks/models/kimi-k2p6",
   },

package/src/providers/ollama/client.ts

@@ -24,7 +24,9 @@ function resolveBaseUrl(optionBaseUrl?: string): string {
   for (const candidate of [optionBaseUrl, getOllamaBaseUrlEnv()]) {
     if (typeof candidate === "string") {
       const trimmed = candidate.trim();
-      if (trimmed.length > 0) return trimmed;
+      if (trimmed.length > 0) {
+        return trimmed;
+      }
     }
   }
   return DEFAULT_OLLAMA_BASE_URL;

package/src/providers/openai/__tests__/chat-completions-input-audio.test.ts

@@ -29,7 +29,9 @@ const OK_CHUNKS: MockChunk[] = [
 function makeStream(chunks: MockChunk[]): AsyncIterable<MockChunk> {
   return {
     async *[Symbol.asyncIterator]() {
-      for (const c of chunks) yield c;
+      for (const c of chunks) {
+        yield c;
+      }
     },
   };
 }

package/src/providers/openai/__tests__/chat-completions-provider-endpoint.test.ts

@@ -30,7 +30,9 @@ const OK_CHUNKS: MockChunk[] = [
 function makeStream(chunks: MockChunk[]): AsyncIterable<MockChunk> {
   return {
     async *[Symbol.asyncIterator]() {
-      for (const c of chunks) yield c;
+      for (const c of chunks) {
+        yield c;
+      }
     },
   };
 }

package/src/providers/openai/__tests__/chat-completions-provider-reasoning.test.ts

@@ -38,7 +38,9 @@ type MockChunk = {
 function makeStream(chunks: MockChunk[]): AsyncIterable<MockChunk> {
   return {
     async *[Symbol.asyncIterator]() {
-      for (const c of chunks) yield c;
+      for (const c of chunks) {
+        yield c;
+      }
     },
   };
 }

package/src/providers/openai/api-error-normalization.ts

@@ -94,11 +94,21 @@ export function deriveReason(
     return "insufficient_credits";
   }
 
-  if (status === 401 || status === 403) return "invalid_credentials";
-  if (status === 429) return "rate_limited";
-  if (status === 529 || /overloaded/i.test(haystack)) return "overloaded";
-  if (status !== undefined && status >= 500) return "server_error";
-  if (status !== undefined && status >= 400) return "bad_request";
+  if (status === 401 || status === 403) {
+    return "invalid_credentials";
+  }
+  if (status === 429) {
+    return "rate_limited";
+  }
+  if (status === 529 || /overloaded/i.test(haystack)) {
+    return "overloaded";
+  }
+  if (status !== undefined && status >= 500) {
+    return "server_error";
+  }
+  if (status !== undefined && status >= 400) {
+    return "bad_request";
+  }
   return "unknown";
 }
 
@@ -134,19 +144,25 @@ export const captureRawErrorBodyFetch = async (
   init?: RequestInit,
 ): Promise<Response> => {
   const res = await globalThis.fetch(url, init);
-  if (res.ok) return res;
+  if (res.ok) {
+    return res;
+  }
   // Don't drain bodies the SDK will retry: reading a large or slow upstream
   // error page on every attempt would delay those retries and buffer the whole
   // body. We still capture terminal (non-retryable) errors — that's where the
   // actionable upstream detail lives (unsupported model, invalid key, malformed
   // request, etc.).
-  if (sdkWillRetry(res)) return res;
+  if (sdkWillRetry(res)) {
+    return res;
+  }
   // clone() so reading the body leaves the SDK's own read of `res` intact.
   const body = await res
     .clone()
     .text()
     .catch(() => undefined);
-  if (!body) return res;
+  if (!body) {
+    return res;
+  }
   capturedErrorBodies.set(
     res.headers,
     body.length > MAX_CAPTURED_BODY_CHARS
@@ -165,8 +181,12 @@ export const captureRawErrorBodyFetch = async (
  */
 function sdkWillRetry(res: Response): boolean {
   const shouldRetryHeader = res.headers.get("x-should-retry");
-  if (shouldRetryHeader === "true") return true;
-  if (shouldRetryHeader === "false") return false;
+  if (shouldRetryHeader === "true") {
+    return true;
+  }
+  if (shouldRetryHeader === "false") {
+    return false;
+  }
   return (
     res.status === 408 ||
     res.status === 409 ||
@@ -195,17 +215,29 @@ export function normalizeOpenAIAPIError(
     "Request failed";
 
   const out: NormalizedOpenAIAPIError = { message };
-  if (body.detail && body.detail !== message) out.detail = body.detail;
+  if (body.detail && body.detail !== message) {
+    out.detail = body.detail;
+  }
   const code = body.apiErrorCode ?? scalar((error as { code?: unknown }).code);
   const type = body.apiErrorType ?? scalar((error as { type?: unknown }).type);
   const param =
     body.apiErrorParam ?? scalar((error as { param?: unknown }).param);
-  if (code) out.apiErrorCode = code;
-  if (type) out.apiErrorType = type;
-  if (param) out.apiErrorParam = param;
+  if (code) {
+    out.apiErrorCode = code;
+  }
+  if (type) {
+    out.apiErrorType = type;
+  }
+  if (param) {
+    out.apiErrorParam = param;
+  }
   const requestId = readHeader(error.headers);
-  if (requestId) out.requestId = requestId;
-  if (rawBody) out.rawBody = rawBody;
+  if (requestId) {
+    out.requestId = requestId;
+  }
+  if (rawBody) {
+    out.rawBody = rawBody;
+  }
   out.reason = deriveReason(out, error.status);
   return out;
 }
@@ -237,9 +269,13 @@ interface BodyDetails {
 }
 
 function extractBody(body: unknown): BodyDetails {
-  if (typeof body === "string") return { message: trunc(body.trim()) };
+  if (typeof body === "string") {
+    return { message: trunc(body.trim()) };
+  }
   const rec = asRecord(body);
-  if (!rec) return {};
+  if (!rec) {
+    return {};
+  }
 
   // OpenAI/OpenRouter nest under `error`; Django puts `detail` at the top.
   // A plain `error: "string"` is the whole message but may still carry
@@ -277,7 +313,9 @@ function extractBody(body: unknown): BodyDetails {
 
 function parseBody(raw: string | undefined): unknown {
   const trimmed = raw?.trim();
-  if (!trimmed) return undefined;
+  if (!trimmed) {
+    return undefined;
+  }
   try {
     return JSON.parse(trimmed);
   } catch {
@@ -292,8 +330,12 @@ function stripLeadingStatus(
   const trimmed = message.trim();
   // SDK sentinel for an unparseable/empty body — carries no signal, so let the
   // caller fall back to "Request failed" rather than surface SDK phrasing.
-  if (/^\d*\s*status code \(no body\)$/i.test(trimmed)) return "";
-  if (typeof status !== "number") return trimmed;
+  if (/^\d*\s*status code \(no body\)$/i.test(trimmed)) {
+    return "";
+  }
+  if (typeof status !== "number") {
+    return trimmed;
+  }
   return trimmed.replace(new RegExp(`^${status}\\s+`), "").trim() || trimmed;
 }
 
@@ -307,8 +349,12 @@ function withScalar(key: keyof BodyDetails, value: unknown): BodyDetails {
 }
 
 function scalar(value: unknown): string | undefined {
-  if (typeof value === "string" && value.length > 0) return value;
-  if (typeof value === "number" && Number.isFinite(value)) return String(value);
+  if (typeof value === "string" && value.length > 0) {
+    return value;
+  }
+  if (typeof value === "number" && Number.isFinite(value)) {
+    return String(value);
+  }
   return undefined;
 }
 
@@ -330,14 +376,18 @@ function readHeaderValue(
   headers: unknown,
   names: readonly string[],
 ): string | undefined {
-  if (!headers) return undefined;
+  if (!headers) {
+    return undefined;
+  }
   const get = (headers as { get?: unknown }).get;
   const getter = typeof get === "function" ? get.bind(headers) : undefined;
   for (const name of names) {
     const raw = getter
       ? (getter(name) as string | null)
       : (asRecord(headers)?.[name] ?? asRecord(headers)?.[name.toLowerCase()]);
-    if (typeof raw === "string" && raw.length > 0) return raw;
+    if (typeof raw === "string" && raw.length > 0) {
+      return raw;
+    }
   }
   return undefined;
 }