@vellumai/assistant 0.12.5-dev.202609282216.bb92e6d → 0.12.5-dev.202609290048.637f7a7

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vellumai/assistant",
-  "version": "0.12.5-dev.202609282216.bb92e6d",
+  "version": "0.12.5-dev.202609290048.637f7a7",
   "license": "MIT",
   "type": "module",
   "exports": {

package/src/__tests__/anthropic-provider.test.ts

@@ -3653,6 +3653,8 @@ describe("AnthropicProvider — deprecated sampling params (temperature / top_p
     "claude-opus-4-7",
     "claude-sonnet-5",
     "anthropic/claude-sonnet-5",
+    "claude-sonnet-5-5",
+    "anthropic/claude-sonnet-5.5",
     "claude-fable-5",
     "claude-fable-5-1",
     "anthropic/claude-fable-5.1",

package/src/__tests__/retry-thinking-tool-choice.test.ts

@@ -144,6 +144,53 @@ describe("retry normalization: thinking + forced tool_choice", () => {
     expect(lastConfig()?.thinking).toEqual({ type: "adaptive" });
   });
 
+  for (const [providerName, model] of [
+    ["anthropic", "claude-sonnet-5-5"],
+    ["anthropic", "claude-opus-5-5"],
+    ["openrouter", "anthropic/claude-sonnet-5.5"],
+    ["vercel-ai-gateway", "anthropic/claude-sonnet-5.5"],
+  ] as const) {
+    test(`downgrades forced tool_choice to auto for ${providerName} ${model}`, async () => {
+      // These models always reason and reject forced tool use, so stripping
+      // `thinking` is not enough: the forced choice itself has to go.
+      setLlmConfig({
+        callSites: {
+          memoryExtraction: { provider: providerName, model },
+        },
+      });
+      const { provider, lastConfig } = makePipeline(providerName);
+      await provider.sendMessage([userMessage], {
+        config: {
+          callSite: "memoryExtraction",
+          tool_choice: { type: "tool", name: "extract_graph_diff" },
+        },
+      });
+      expect(lastConfig()?.tool_choice).toEqual({ type: "auto" });
+    });
+  }
+
+  test("keeps forced tool_choice for an adaptive-only model off the Anthropic wire", async () => {
+    setLlmConfig({
+      callSites: {
+        memoryExtraction: {
+          provider: "fireworks",
+          model: "accounts/fireworks/models/glm-5p3",
+        },
+      },
+    });
+    const { provider, lastConfig } = makePipeline("fireworks");
+    await provider.sendMessage([userMessage], {
+      config: {
+        callSite: "memoryExtraction",
+        tool_choice: { type: "tool", name: "extract_graph_diff" },
+      },
+    });
+    expect(lastConfig()?.tool_choice).toEqual({
+      type: "tool",
+      name: "extract_graph_diff",
+    });
+  });
+
   test("preserves explicit thinking: disabled with forced tool_choice", async () => {
     // Callers like the retriever explicitly set thinking: { type: "disabled" }
     // alongside forced tool_choice. This should pass through unchanged since

package/src/providers/anthropic/client.ts

@@ -1008,7 +1008,7 @@ export class AnthropicProvider implements Provider {
       const isHaiku = effectiveModel.includes("haiku");
       // Effort support is per-model: Haiku and Sonnet 4.5 reject the param (see isEffortSupported).
       const supportsEffort = isEffortSupported(effectiveModel);
-      // opus-4-7 / opus-4-8 / opus-5 / opus-5-5 and sonnet-5 reject `temperature`,
+      // opus-4-7 / opus-4-8 / opus-5 / opus-5-5 and sonnet-5 / sonnet-5-5 reject `temperature`,
       // `top_p`, and `top_k` with a 400 "`temperature`/`top_p` is deprecated
       // for this model" — model-wide, not effort-conditional (verified
       // 2026-06-23). opus-4-6 / sonnet-4-6 / haiku-4-5 still accept them.

package/src/providers/model-catalog.ts

@@ -331,6 +331,24 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
           cacheReadPer1mTokens: 0.5,
         },
       },
+      {
+        id: "claude-sonnet-5-5",
+        displayName: "Claude Sonnet 5.5",
+        contextWindowTokens: 1000000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens: 200000,
+        supportsThinking: true,
+        adaptiveThinkingOnly: true,
+        supportsCaching: true,
+        supportsVision: true,
+        supportsToolUse: true,
+        pricing: {
+          inputPer1mTokens: 2,
+          outputPer1mTokens: 10,
+          cacheWritePer1mTokens: 2.5,
+          cacheReadPer1mTokens: 0.2,
+        },
+      },
       {
         id: "claude-sonnet-5",
         displayName: "Claude Sonnet 5",
@@ -1336,6 +1354,24 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
           cacheReadPer1mTokens: 0.5,
         },
       },
+      {
+        id: "anthropic/claude-sonnet-5.5",
+        displayName: "Claude Sonnet 5.5",
+        contextWindowTokens: 1000000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens: 200000,
+        supportsThinking: true,
+        adaptiveThinkingOnly: true,
+        supportsCaching: true,
+        supportsVision: true,
+        supportsToolUse: true,
+        pricing: {
+          inputPer1mTokens: 2,
+          outputPer1mTokens: 10,
+          cacheWritePer1mTokens: 2.5,
+          cacheReadPer1mTokens: 0.2,
+        },
+      },
       {
         id: "anthropic/claude-sonnet-5",
         displayName: "Claude Sonnet 5",
@@ -2390,6 +2426,24 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
           cacheReadPer1mTokens: 0.5,
         },
       },
+      {
+        id: "anthropic/claude-sonnet-5.5",
+        displayName: "Claude Sonnet 5.5",
+        contextWindowTokens: 1000000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens: 200000,
+        supportsThinking: true,
+        adaptiveThinkingOnly: true,
+        supportsCaching: true,
+        supportsVision: true,
+        supportsToolUse: true,
+        pricing: {
+          inputPer1mTokens: 2,
+          outputPer1mTokens: 10,
+          cacheWritePer1mTokens: 2.5,
+          cacheReadPer1mTokens: 0.2,
+        },
+      },
       {
         id: "anthropic/claude-sonnet-5",
         displayName: "Claude Sonnet 5",

package/src/providers/retry.ts

@@ -928,6 +928,36 @@ function normalizeSendMessageOptions(
     }
   }
 
+  // Adaptive-thinking-only Claude models (Fable, Opus 5.5, Sonnet 5.5) cannot
+  // turn reasoning off, so stripping `thinking` below does not clear the
+  // forced-tool conflict and Anthropic 400s the request. Downgrade the forced
+  // choice to `auto` so the call goes out. The tool is no longer guaranteed:
+  // forced-tool call sites already treat a missing tool call as a failed
+  // result, which is what the 400 gave them on every request.
+  const forcedToolModel =
+    typeof nextConfig.model === "string" ? nextConfig.model : "";
+  const forcedToolChoice = nextConfig.tool_choice as
+    | Record<string, unknown>
+    | undefined;
+  if (
+    forcedToolChoice != null &&
+    (forcedToolChoice.type === "tool" || forcedToolChoice.type === "any") &&
+    isAdaptiveThinkingOnlyModel(forcedToolModel) &&
+    targetsAnthropicWire(providerName, forcedToolModel)
+  ) {
+    log.warn(
+      {
+        providerName,
+        callSite: config.callSite,
+        model: forcedToolModel,
+        droppedToolChoice: forcedToolChoice,
+      },
+      "Downgrading forced `tool_choice` to `auto` because this model always " +
+        "reasons and rejects forced tool use.",
+    );
+    nextConfig.tool_choice = { type: "auto" };
+  }
+
   // Anthropic (and the gateways fronting Anthropic) rejects requests that
   // combine extended thinking with forced tool use (`tool_choice.type` of
   // `"tool"` or `"any"`).  Strip thinking when both are present so the