@pensar/apex 2.6.1 → 2.7.0-canary.21cb83dc

package/src/core/ai/models/claude-capabilities.ts

@@ -1,22 +1,43 @@
-type ClaudeCapabilities = {
+export type ClaudeThinkingEffort = "low" | "medium" | "high" | "xhigh" | "max";
+
+export type ClaudeCapabilities = {
   alwaysOnThinking: boolean;
   bindsThinking?: boolean;
+  defaultEffort: ClaudeThinkingEffort;
 };
 
 const CLAUDE_CAPABILITIES: Record<string, ClaudeCapabilities> = {
-  "claude-sonnet-5-5": { alwaysOnThinking: true, bindsThinking: true },
-  "claude-opus-5-5": { alwaysOnThinking: true, bindsThinking: true },
-  "claude-fable-5-1": { alwaysOnThinking: true, bindsThinking: true },
-  "claude-opus-5": { alwaysOnThinking: false },
-  "claude-sonnet-5": { alwaysOnThinking: false },
-  "claude-fable-5": { alwaysOnThinking: true },
+  "claude-haiku-5-5": {
+    alwaysOnThinking: false,
+    bindsThinking: true,
+    defaultEffort: "medium",
+  },
+  "claude-sonnet-5-5": {
+    alwaysOnThinking: true,
+    bindsThinking: true,
+    defaultEffort: "high",
+  },
+  "claude-opus-5-5": {
+    alwaysOnThinking: true,
+    bindsThinking: true,
+    defaultEffort: "medium",
+  },
+  "claude-fable-5-1": {
+    alwaysOnThinking: true,
+    bindsThinking: true,
+    defaultEffort: "high",
+  },
+  "claude-opus-5": { alwaysOnThinking: false, defaultEffort: "high" },
+  "claude-sonnet-5": { alwaysOnThinking: false, defaultEffort: "high" },
+  "claude-fable-5": { alwaysOnThinking: true, defaultEffort: "high" },
 };
 
 export function getClaudeCapabilities(
   modelId: string,
 ): ClaudeCapabilities | undefined {
   const nativeId = modelId
-    .replace(/^(concentrate:|anthropic\/)/, "")
+    .replace(/^(concentrate:|pensar:)/, "")
+    .replace(/^(?:(?:us|global|eu|au|jp)\.)?anthropic[./]/, "")
     .replace(/(\d)\.(\d)/g, "$1-$2");
   return CLAUDE_CAPABILITIES[nativeId];
 }

package/src/core/ai/models/concentrate.ts

@@ -1,5 +1,15 @@
 import type { ModelInfo } from "../ai";
 
+// Concentrate does not route on max_output_tokens, so these use the lowest
+// ceiling among each slug's upstream providers.
+export const CONCENTRATE_MAX_OUTPUT_TOKENS: Readonly<Record<string, number>> = {
+  "concentrate:kimi-k3": 32_768,
+  "concentrate:qwen3.8-max": 32_768,
+  "concentrate:qwen3.8-27b": 32_768,
+  "concentrate:deepseek-v4-1-flash": 32_768,
+  "concentrate:deepseek-v4-1-flash-1017": 131_072,
+};
+
 // Curated from Concentrate's public model catalog. The local prefix keeps
 // gateway selections distinct from direct-provider models and is not sent.
 export const CONCENTRATE_MODELS: ModelInfo[] = [
@@ -153,6 +163,36 @@ export const CONCENTRATE_MODELS: ModelInfo[] = [
     provider: "concentrate",
     contextLength: 1000000,
   },
+  {
+    id: "concentrate:qwen3.8-max",
+    name: "Qwen3.8 Max",
+    provider: "concentrate",
+    contextLength: 256000,
+  },
+  {
+    id: "concentrate:qwen3.8-27b",
+    name: "Qwen3.8 27B",
+    provider: "concentrate",
+    contextLength: 262144,
+  },
+  {
+    id: "concentrate:deepseek-v4-1-flash-1017",
+    name: "DeepSeek V4.1 Flash 1017",
+    provider: "concentrate",
+    contextLength: 1048576,
+  },
+  {
+    id: "concentrate:deepseek-v4-1-flash",
+    name: "DeepSeek V4.1 Flash",
+    provider: "concentrate",
+    contextLength: 1000000,
+  },
+  {
+    id: "concentrate:deepseek-v4-pro-0813",
+    name: "DeepSeek V4 Pro 0813",
+    provider: "concentrate",
+    contextLength: 1048576,
+  },
   {
     id: "concentrate:deepseek-v3-2",
     name: "DeepSeek V3.2",
@@ -165,6 +205,12 @@ export const CONCENTRATE_MODELS: ModelInfo[] = [
     provider: "concentrate",
     contextLength: 256000,
   },
+  {
+    id: "concentrate:kimi-k3",
+    name: "Kimi K3",
+    provider: "concentrate",
+    contextLength: 1048576,
+  },
   {
     id: "concentrate:kimi-k2-thinking",
     name: "Kimi K2 Thinking",
@@ -177,6 +223,12 @@ export const CONCENTRATE_MODELS: ModelInfo[] = [
     provider: "concentrate",
     contextLength: 204800,
   },
+  {
+    id: "concentrate:mimo-v2.6-pro",
+    name: "MiMo V2.6 Pro",
+    provider: "concentrate",
+    contextLength: 1048576,
+  },
   {
     id: "concentrate:glm-5.3",
     name: "GLM-5.3",

package/src/core/ai/models/index.ts

@@ -12,9 +12,15 @@ import { ANTHROPIC_MODELS } from "./anthropic";
 import { BEDROCK_MODELS } from "./bedrock";
 import { getClaudeCapabilities } from "./claude-capabilities";
 
-export { getClaudeCapabilities } from "./claude-capabilities";
+export {
+  type ClaudeThinkingEffort,
+  getClaudeCapabilities,
+} from "./claude-capabilities";
 
-import { CONCENTRATE_MODELS } from "./concentrate";
+import {
+  CONCENTRATE_MAX_OUTPUT_TOKENS,
+  CONCENTRATE_MODELS,
+} from "./concentrate";
 import { GOOGLE_MODELS } from "./google";
 import { INCEPTION_MODELS } from "./inception";
 // Bedrock Mantle (OpenAI Responses API) — curated manually.
@@ -100,6 +106,8 @@ export function getMaxOutputTokens(
 function lookupOutputBudgetByPattern(modelId: string): number {
   if (modelId === "chat-latest") return 128_000;
   if (getClaudeCapabilities(modelId)) return 128_000;
+  const concentrateBudget = CONCENTRATE_MAX_OUTPUT_TOKENS[modelId];
+  if (concentrateBudget) return concentrateBudget;
   // OpenRouter uses dots in Claude version numbers (anthropic/claude-opus-4.6)
   // while native Anthropic uses dashes (claude-opus-4-6-20250929). Normalize
   // digit.digit sequences to dashes so all Claude patterns match both forms.
@@ -219,14 +227,22 @@ function lookupOutputBudgetByPattern(modelId: string): number {
     return 131_072;
   }
 
+  if (modelId.includes("mimo-v2.6-pro")) {
+    return 131_072;
+  }
+
+  if (modelId.includes("kimi-k3")) {
+    return 131_072;
+  }
+
   // GLM 5 models ship a ~131K (128Ki) max-output window. Matching the family
   // keeps Bedrock and OpenRouter model IDs on the same budget.
   if (modelId.includes("glm-5")) {
     return 131_072;
   }
 
-  // V4.1 Flash's larger output window must precede the legacy DeepSeek fallback.
-  if (modelId.includes("deepseek-v4.1-flash")) {
+  // DeepSeek V4's larger output window must precede the legacy DeepSeek fallback.
+  if (/deepseek-v4[.-]/.test(modelId)) {
     return 384_000;
   }
 
@@ -235,6 +251,11 @@ function lookupOutputBudgetByPattern(modelId: string): number {
     return 8_192;
   }
 
+  // Qwen3.8's first-party 131K output window must precede the Qwen fallback.
+  if (modelId.includes("qwen3.8")) {
+    return 131_072;
+  }
+
   // Qwen3 Coder ships a 16K max-output window over its 128K+ context. Match the
   // whole `qwen` family so future Bedrock/OpenRouter qwen ids don't silently
   // inherit the 4,096 default.

package/src/core/ai/models/models.test.ts

@@ -103,6 +103,11 @@ describe("getMaxOutputTokens", () => {
   });
 
   it("recognizes Claude tier-specific budgets", () => {
+    expect(getMaxOutputTokens("claude-haiku-5-5")).toBe(128_000);
+    expect(getMaxOutputTokens("global.anthropic.claude-haiku-5-5")).toBe(
+      128_000,
+    );
+    expect(getMaxOutputTokens("anthropic/claude-haiku-5.5")).toBe(128_000);
     expect(getMaxOutputTokens("claude-sonnet-4-5-20250929")).toBe(64_000);
     expect(getMaxOutputTokens("claude-opus-4-5-20250101")).toBe(64_000);
     expect(getMaxOutputTokens("claude-haiku-4-5-20251001")).toBe(64_000);
@@ -111,6 +116,24 @@ describe("getMaxOutputTokens", () => {
     expect(getMaxOutputTokens("claude-3-5-haiku-20241022")).toBe(8_192);
   });
 
+  it("registers Claude Haiku 5.5 across Anthropic, Bedrock, and OpenRouter", () => {
+    expect(getModelInfo("claude-haiku-5-5")).toMatchObject({
+      name: "Claude Haiku 5.5",
+      provider: "anthropic",
+      contextLength: 1_000_000,
+    });
+    expect(getModelInfo("global.anthropic.claude-haiku-5-5")).toMatchObject({
+      name: "Claude Haiku 5.5 (Global)",
+      provider: "bedrock",
+      contextLength: 1_000_000,
+    });
+    expect(getModelInfo("anthropic/claude-haiku-5.5")).toMatchObject({
+      name: "Claude Haiku 5.5",
+      provider: "openrouter",
+      contextLength: 1_000_000,
+    });
+  });
+
   it("pins latest-tier Claude (4.6 + 4.7 + 4.8) to 128K output", () => {
     // These ship the 128K window; they MUST not fall through to the
     // generic `claude-opus-4-` / `claude-sonnet-4-` catch-alls (32K/64K)
@@ -288,3 +311,134 @@ describe("OpenAI reasoning effort support", () => {
     );
   });
 });
+
+describe("open-weight gateway models", () => {
+  it.each([
+    {
+      id: "xiaomi/mimo-v2.6-pro",
+      name: "MiMo-V2.6-Pro",
+      provider: "openrouter",
+      contextLength: 1_050_000,
+      maxOutputTokens: 131_072,
+    },
+    {
+      id: "concentrate:mimo-v2.6-pro",
+      name: "MiMo V2.6 Pro",
+      provider: "concentrate",
+      contextLength: 1_048_576,
+      maxOutputTokens: 131_072,
+    },
+    {
+      id: "moonshotai/kimi-k3",
+      name: "Kimi K3",
+      provider: "openrouter",
+      contextLength: 1_048_576,
+      maxOutputTokens: 131_072,
+    },
+    {
+      id: "concentrate:kimi-k3",
+      name: "Kimi K3",
+      provider: "concentrate",
+      contextLength: 1_048_576,
+      maxOutputTokens: 32_768,
+    },
+    {
+      id: "qwen/qwen3.8-max-0902",
+      name: "Qwen3.8 Max (0902)",
+      provider: "openrouter",
+      contextLength: 1_000_000,
+      maxOutputTokens: 131_072,
+    },
+    {
+      id: "qwen/qwen3.8-27b",
+      name: "Qwen3.8 27B",
+      provider: "openrouter",
+      contextLength: 1_000_000,
+      maxOutputTokens: 131_072,
+    },
+    {
+      id: "qwen/qwen3.8-flash",
+      name: "Qwen3.8 Flash",
+      provider: "openrouter",
+      contextLength: 1_000_000,
+      maxOutputTokens: 131_072,
+    },
+    {
+      id: "concentrate:qwen3.8-max",
+      name: "Qwen3.8 Max",
+      provider: "concentrate",
+      contextLength: 256_000,
+      maxOutputTokens: 32_768,
+    },
+    {
+      id: "concentrate:qwen3.8-27b",
+      name: "Qwen3.8 27B",
+      provider: "concentrate",
+      contextLength: 262_144,
+      maxOutputTokens: 32_768,
+    },
+    {
+      id: "deepseek/deepseek-v4.1-flash-1017",
+      name: "DeepSeek V4.1 Flash 1017",
+      provider: "openrouter",
+      contextLength: 1_048_576,
+      maxOutputTokens: 384_000,
+    },
+    {
+      id: "deepseek/deepseek-v4-pro-0813",
+      name: "DeepSeek V4 Pro 0813",
+      provider: "openrouter",
+      contextLength: 1_048_576,
+      maxOutputTokens: 384_000,
+    },
+    {
+      id: "concentrate:deepseek-v4-1-flash",
+      name: "DeepSeek V4.1 Flash",
+      provider: "concentrate",
+      contextLength: 1_000_000,
+      maxOutputTokens: 32_768,
+    },
+    {
+      id: "concentrate:deepseek-v4-1-flash-1017",
+      name: "DeepSeek V4.1 Flash 1017",
+      provider: "concentrate",
+      contextLength: 1_048_576,
+      maxOutputTokens: 131_072,
+    },
+    {
+      id: "concentrate:deepseek-v4-pro-0813",
+      name: "DeepSeek V4 Pro 0813",
+      provider: "concentrate",
+      contextLength: 1_048_576,
+      maxOutputTokens: 384_000,
+    },
+  ])("registers $id with its gateway limits", ({
+    id,
+    maxOutputTokens,
+    ...info
+  }) => {
+    expect(getModelInfo(id)).toMatchObject(info);
+    expect(getMaxOutputTokens(id)).toBe(maxOutputTokens);
+  });
+
+  it.each([
+    "concentrate:mimo-v2.6-pro",
+    "concentrate:kimi-k3",
+    "concentrate:qwen3.8-max",
+    "concentrate:qwen3.8-27b",
+    "concentrate:deepseek-v4-1-flash",
+    "concentrate:deepseek-v4-1-flash-1017",
+    "concentrate:deepseek-v4-pro-0813",
+  ])("offers every effort Concentrate can route for %s", (id) => {
+    expect(modelSupportsOpenAIReasoning(id)).toBe(true);
+    expect(getOpenAIReasoningEfforts(id)).toEqual([
+      "low",
+      "medium",
+      "high",
+      "xhigh",
+      "max",
+    ]);
+    expect(normalizeOpenAIReasoningEffort(id)).toBe("medium");
+    expect(normalizeOpenAIReasoningEffort(id, "max")).toBe("max");
+  });
+});

package/src/core/ai/models/openrouter.ts

@@ -3,6 +3,12 @@
 
 import type { ModelInfo } from "../ai";
 export const OPENROUTER_MODELS: ModelInfo[] = [
+  {
+    id: "anthropic/claude-haiku-5.5",
+    name: "Claude Haiku 5.5",
+    provider: "openrouter",
+    contextLength: 1000000,
+  },
   {
     id: "anthropic/claude-sonnet-5.5",
     name: "Claude Sonnet 5.5",
@@ -177,12 +183,24 @@ export const OPENROUTER_MODELS: ModelInfo[] = [
     provider: "openrouter",
     contextLength: 128000,
   },
+  {
+    id: "deepseek/deepseek-v4.1-flash-1017",
+    name: "DeepSeek V4.1 Flash 1017",
+    provider: "openrouter",
+    contextLength: 1048576,
+  },
   {
     id: "deepseek/deepseek-v4.1-flash",
     name: "DeepSeek V4.1 Flash",
     provider: "openrouter",
     contextLength: 1048576,
   },
+  {
+    id: "deepseek/deepseek-v4-pro-0813",
+    name: "DeepSeek V4 Pro 0813",
+    provider: "openrouter",
+    contextLength: 1048576,
+  },
   {
     id: "deepseek/deepseek-chat",
     name: "DeepSeek Chat",
@@ -217,7 +235,7 @@ export const OPENROUTER_MODELS: ModelInfo[] = [
     id: "moonshotai/kimi-k3",
     name: "Kimi K3",
     provider: "openrouter",
-    contextLength: 1000000,
+    contextLength: 1048576,
   },
   {
     id: "moonshot/kimi-k2-turbo",
@@ -249,6 +267,24 @@ export const OPENROUTER_MODELS: ModelInfo[] = [
     provider: "openrouter",
     contextLength: 1048576,
   },
+  {
+    id: "qwen/qwen3.8-max-0902",
+    name: "Qwen3.8 Max (0902)",
+    provider: "openrouter",
+    contextLength: 1000000,
+  },
+  {
+    id: "qwen/qwen3.8-27b",
+    name: "Qwen3.8 27B",
+    provider: "openrouter",
+    contextLength: 1000000,
+  },
+  {
+    id: "qwen/qwen3.8-flash",
+    name: "Qwen3.8 Flash",
+    provider: "openrouter",
+    contextLength: 1000000,
+  },
   {
     id: "qwen/qwen3-8b",
     name: "Qwen 3 8B Thinking",
@@ -315,6 +351,12 @@ export const OPENROUTER_MODELS: ModelInfo[] = [
     provider: "openrouter",
     contextLength: 256000,
   },
+  {
+    id: "xiaomi/mimo-v2.6-pro",
+    name: "MiMo-V2.6-Pro",
+    provider: "openrouter",
+    contextLength: 1050000,
+  },
   {
     id: "z-ai/glm-4.7-flash",
     name: "GLM 4.7 Flash",

package/src/core/ai/providers/anthropic-models.test.ts

@@ -12,12 +12,14 @@ import { getMaxOutputTokens } from "../models";
 import { getProviderModel } from "../utils";
 
 const boundModels = new Set([
+  "claude-haiku-5-5",
   "claude-sonnet-5-5",
   "claude-opus-5-5",
   "claude-fable-5-1",
 ]);
 
 const models = [
+  { id: "claude-haiku-5-5", required: false },
   { id: "claude-sonnet-5-5", required: true },
   { id: "claude-opus-5-5", required: true },
   { id: "claude-fable-5-1", required: true },
@@ -105,7 +107,10 @@ describe.each(models)("Anthropic $id", ({ id, required }) => {
     expect(body).toMatchObject({
       model: id,
       max_tokens: 128_000,
-      output_config: { effort: "medium", format: { type: "json_schema" } },
+      output_config: {
+        ...(enableThinking || required ? { effort: "medium" } : {}),
+        format: { type: "json_schema" },
+      },
       thinking: { type: enableThinking || required ? "adaptive" : "disabled" },
     });
     expect(body.temperature).toBeUndefined();
@@ -287,3 +292,57 @@ describe.each(models)("Anthropic $id", ({ id, required }) => {
     );
   });
 });
+
+describe("Anthropic Claude Haiku 5.5 effort", () => {
+  it.each([
+    "xhigh",
+    "max",
+  ] as const)("sends the %s effort with adaptive thinking", async (effort) => {
+    const fetchMock = vi.fn(
+      async (_input: RequestInfo | URL, _init?: RequestInit) =>
+        completion("claude-haiku-5-5"),
+    );
+    vi.stubGlobal("fetch", fetchMock);
+
+    await getProviderModel("claude-haiku-5-5", {
+      anthropicAPIKey: "test-key",
+    }).doGenerate({
+      prompt: [{ role: "user", content: [{ type: "text", text: "Check" }] }],
+      providerOptions: buildReasoningProviderOptions("claude-haiku-5-5", {
+        enableThinking: true,
+        thinkingEffort: effort,
+      }),
+    });
+
+    const call = fetchMock.mock.calls[0];
+    if (!call) throw new Error("Expected a Claude request");
+    expect(JSON.parse(String(call[1]?.body))).toMatchObject({
+      thinking: { type: "adaptive" },
+      output_config: { effort },
+    });
+  });
+
+  it("uses the documented medium effort by default", async () => {
+    const fetchMock = vi.fn(
+      async (_input: RequestInfo | URL, _init?: RequestInit) =>
+        completion("claude-haiku-5-5"),
+    );
+    vi.stubGlobal("fetch", fetchMock);
+
+    await getProviderModel("claude-haiku-5-5", {
+      anthropicAPIKey: "test-key",
+    }).doGenerate({
+      prompt: [{ role: "user", content: [{ type: "text", text: "Check" }] }],
+      providerOptions: buildReasoningProviderOptions("claude-haiku-5-5", {
+        enableThinking: true,
+      }),
+    });
+
+    const call = fetchMock.mock.calls[0];
+    if (!call) throw new Error("Expected a Claude request");
+    expect(JSON.parse(String(call[1]?.body))).toMatchObject({
+      thinking: { type: "adaptive" },
+      output_config: { effort: "medium" },
+    });
+  });
+});

package/src/core/ai/providers/bedrock-models.test.ts

@@ -0,0 +1,81 @@
+import { afterEach, describe, expect, it, vi } from "vitest";
+import { buildReasoningProviderOptions } from "../ai";
+import { getProviderModel } from "../utils";
+
+function completion() {
+  return Response.json({
+    output: {
+      message: {
+        role: "assistant",
+        content: [{ text: "ok" }],
+      },
+    },
+    stopReason: "end_turn",
+    usage: {
+      inputTokens: 12,
+      outputTokens: 2,
+      totalTokens: 14,
+    },
+  });
+}
+
+afterEach(() => vi.unstubAllGlobals());
+
+describe("Bedrock Claude Haiku 5.5", () => {
+  it("sends adaptive max effort to the global inference profile", async () => {
+    const fetchMock = vi.fn(
+      async (_input: RequestInfo | URL, _init?: RequestInit) => completion(),
+    );
+    vi.stubGlobal("fetch", fetchMock);
+
+    await getProviderModel("global.anthropic.claude-haiku-5-5", {
+      bedrock: { apiKey: "test-key", region: "us-east-1" },
+    }).doGenerate({
+      prompt: [{ role: "user", content: [{ type: "text", text: "Check" }] }],
+      providerOptions: buildReasoningProviderOptions(
+        "global.anthropic.claude-haiku-5-5",
+        { enableThinking: true, thinkingEffort: "max" },
+      ),
+    });
+
+    const call = fetchMock.mock.calls[0];
+    if (!call) throw new Error("Expected a Bedrock request");
+    expect(String(call[0])).toContain(
+      "/model/global.anthropic.claude-haiku-5-5/converse",
+    );
+    expect(new Headers(call[1]?.headers).get("authorization")).toBe(
+      "Bearer test-key",
+    );
+    expect(JSON.parse(String(call[1]?.body))).toMatchObject({
+      additionalModelRequestFields: {
+        thinking: { type: "adaptive", display: "summarized" },
+        output_config: { effort: "max" },
+      },
+    });
+  });
+
+  it("forwards thinking disabled explicitly", async () => {
+    const fetchMock = vi.fn(
+      async (_input: RequestInfo | URL, _init?: RequestInit) => completion(),
+    );
+    vi.stubGlobal("fetch", fetchMock);
+
+    await getProviderModel("global.anthropic.claude-haiku-5-5", {
+      bedrock: { apiKey: "test-key", region: "us-east-1" },
+    }).doGenerate({
+      prompt: [{ role: "user", content: [{ type: "text", text: "Check" }] }],
+      providerOptions: buildReasoningProviderOptions(
+        "global.anthropic.claude-haiku-5-5",
+        { enableThinking: false },
+      ),
+    });
+
+    const call = fetchMock.mock.calls[0];
+    if (!call) throw new Error("Expected a Bedrock request");
+    expect(JSON.parse(String(call[1]?.body))).toMatchObject({
+      additionalModelRequestFields: {
+        thinking: { type: "disabled" },
+      },
+    });
+  });
+});

package/src/core/ai/providers/concentrate.test.ts

@@ -79,6 +79,37 @@ describe("createConcentrateModel", () => {
     });
   });
 
+  it.each([
+    "mimo-v2.6-pro",
+    "kimi-k3",
+    "qwen3.8-max",
+    "qwen3.8-27b",
+    "deepseek-v4-1-flash",
+    "deepseek-v4-1-flash-1017",
+    "deepseek-v4-pro-0813",
+  ])("forces reasoning on the %s Responses request", async (slug) => {
+    const fetchMock = vi.fn(
+      async (_input: RequestInfo | URL, _init?: RequestInit) =>
+        completedResponse(),
+    );
+    const model = createConcentrateModel(`concentrate:${slug}`, {
+      apiKey: "sk-cn-test",
+      fetch: fetchMock,
+    });
+
+    await model.doGenerate({
+      prompt: [{ role: "user", content: [{ type: "text", text: "Say ok" }] }],
+      providerOptions: { openai: { reasoningEffort: "max" } },
+    } satisfies LanguageModelV3CallOptions);
+
+    const body = JSON.parse(String(fetchMock.mock.calls[0]?.[1]?.body));
+    expect(body).toMatchObject({
+      model: slug,
+      store: false,
+      reasoning: { effort: "max", summary: "auto" },
+    });
+  });
+
   it("fails before constructing a request when the key is missing", () => {
     expect(() =>
       createConcentrateModel("concentrate:glm-5.3", { apiKey: " " }),

package/src/core/ai/providers/concentrate.ts

@@ -9,6 +9,16 @@ import { getClaudeCapabilities } from "../models";
 
 export const CONCENTRATE_BASE_URL = "https://api.concentrate.ai/v1";
 export const CONCENTRATE_GLM_5_3_MODEL_ID = "concentrate:glm-5.3";
+// Concentrate routes each reasoning effort only to upstreams that accept it.
+export const CONCENTRATE_REASONING_MODEL_IDS: ReadonlySet<string> = new Set([
+  "concentrate:mimo-v2.6-pro",
+  "concentrate:kimi-k3",
+  "concentrate:qwen3.8-max",
+  "concentrate:qwen3.8-27b",
+  "concentrate:deepseek-v4-1-flash",
+  "concentrate:deepseek-v4-1-flash-1017",
+  "concentrate:deepseek-v4-pro-0813",
+]);
 
 const RETRYABLE_STATUSES = new Set([424, 429, 500, 503, 504]);
 
@@ -139,6 +149,7 @@ export function createConcentrateModel(
   return withConcentrateDefaults(
     concentrate.responses(upstreamModelId),
     modelId === CONCENTRATE_GLM_5_3_MODEL_ID ||
+      CONCENTRATE_REASONING_MODEL_IDS.has(modelId) ||
       /^gpt-6[.-]/.test(upstreamModelId) ||
       !!getClaudeCapabilities(modelId),
   );