fix(api): expose models.dev context windows in /v1/models (#1972)

Integrated into release/v3.7.9

src/app/api/v1/models/catalog.ts

@@ -338,10 +338,6 @@ export async function getUnifiedModelsResponse(
         continue;
       }
 
-      // Get default context length from registry (provider-level default)
-      const registryEntry = REGISTRY[alias] || REGISTRY[canonicalProviderId];
-      const defaultContextLength = registryEntry?.defaultContextLength;
-
       for (const model of providerModels) {
         if (!providerSupportsModel(canonicalProviderId, model.id)) continue;
         const aliasId = `${alias}/${model.id}`;
@@ -349,8 +345,6 @@ export async function getUnifiedModelsResponse(
 
         const visionFields =
           getVisionCapabilityFields(aliasId) || getVisionCapabilityFields(model.id);
-        // Model-level context length overrides provider default
-        const contextLength = model.contextLength || defaultContextLength;
 
         models.push({
           id: aliasId,
@@ -360,7 +354,6 @@ export async function getUnifiedModelsResponse(
           permission: [],
           root: model.id,
           parent: null,
-          ...(contextLength ? { context_length: contextLength } : {}),
           ...(visionFields || {}),
         });
 
@@ -378,7 +371,6 @@ export async function getUnifiedModelsResponse(
             permission: [],
             root: model.id,
             parent: aliasId,
-            ...(contextLength ? { context_length: contextLength } : {}),
             ...(providerVisionFields || {}),
           });
         }
@@ -813,7 +805,24 @@ export async function getUnifiedModelsResponse(
       finalModels = filtered;
     }
 
-    const enrichedModels = finalModels.map((model) => enrichCatalogModelEntry(model));
+    const getDefaultContextFallback = (model: any): number | undefined => {
+      if (typeof model.context_length === "number") return undefined;
+      if (model.owned_by === "combo") return undefined;
+      if (model.type && model.type !== "chat") return undefined;
+
+      const provider = typeof model.owned_by === "string" ? model.owned_by : null;
+      if (!provider) return undefined;
+      const canonicalId = aliasToProviderId[provider] || provider;
+      return REGISTRY[canonicalId]?.defaultContextLength;
+    };
+
+    const enrichedModels = finalModels.map((model) => {
+      const enriched = enrichCatalogModelEntry(model);
+      const fallbackContextLength = getDefaultContextFallback(enriched);
+      return fallbackContextLength
+        ? { ...enriched, context_length: fallbackContextLength }
+        : enriched;
+    });
 
     return Response.json(
       {

src/lib/modelCapabilities.ts

@@ -183,16 +183,21 @@ function resolveVisionCapability(
   modalitiesInput: string[],
   modalitiesOutput: string[]
 ): boolean | null {
-  if (typeof spec?.supportsVision === "boolean") return spec.supportsVision;
-  if (typeof registryModel?.supportsVision === "boolean") return registryModel.supportsVision;
-
   const allModalities = [...modalitiesInput, ...modalitiesOutput].map((entry) =>
     String(entry).toLowerCase()
   );
+
+  if (typeof synced?.attachment === "boolean") {
+    return synced.attachment;
+  }
+
   if (allModalities.some((entry) => entry.includes("image"))) {
     return true;
   }
 
+  if (typeof registryModel?.supportsVision === "boolean") return registryModel.supportsVision;
+  if (typeof spec?.supportsVision === "boolean") return spec.supportsVision;
+
   return null;
 }
 
@@ -215,18 +220,22 @@ export function getResolvedModelCapabilities(input: CapabilityInput): ResolvedMo
     ) || "";
 
   const supportsTools =
-    typeof spec?.supportsTools === "boolean"
-      ? spec.supportsTools
-      : typeof registryModel?.toolCalling === "boolean"
-        ? registryModel.toolCalling
-        : (synced?.tool_call ?? null);
+    synced?.tool_call ??
+    (typeof registryModel?.toolCalling === "boolean" ? registryModel.toolCalling : null) ??
+    (typeof spec?.supportsTools === "boolean" ? spec.supportsTools : null);
 
   const supportsThinking =
-    typeof spec?.supportsThinking === "boolean"
-      ? spec.supportsThinking
-      : typeof registryModel?.supportsReasoning === "boolean"
-        ? registryModel.supportsReasoning
-        : (synced?.reasoning ?? null);
+    synced?.reasoning ??
+    (typeof registryModel?.supportsReasoning === "boolean"
+      ? registryModel.supportsReasoning
+      : null) ??
+    (typeof spec?.supportsThinking === "boolean" ? spec.supportsThinking : null);
+
+  const contextWindow =
+    synced?.limit_context ??
+    (typeof registryModel?.contextLength === "number" ? registryModel.contextLength : null) ??
+    spec?.contextWindow ??
+    null;
 
   return {
     provider: resolved.provider,
@@ -247,14 +256,10 @@ export function getResolvedModelCapabilities(input: CapabilityInput): ResolvedMo
     attachment: synced?.attachment ?? null,
     structuredOutput: synced?.structured_output ?? null,
     temperature: synced?.temperature ?? null,
-    contextWindow:
-      spec?.contextWindow ??
-      (typeof registryModel?.contextLength === "number" ? registryModel.contextLength : null) ??
-      synced?.limit_context ??
-      null,
-    maxInputTokens: synced?.limit_input ?? spec?.contextWindow ?? null,
+    contextWindow,
+    maxInputTokens: synced?.limit_input ?? contextWindow,
     maxOutputTokens:
-      spec?.maxOutputTokens ?? synced?.limit_output ?? MODEL_SPECS.__default__.maxOutputTokens,
+      synced?.limit_output ?? spec?.maxOutputTokens ?? MODEL_SPECS.__default__.maxOutputTokens,
     defaultThinkingBudget: spec?.defaultThinkingBudget ?? 0,
     thinkingBudgetCap: spec?.thinkingBudgetCap ?? null,
     thinkingOverhead: spec?.thinkingOverhead ?? null,

src/lib/modelsDevSync.ts

@@ -154,12 +154,16 @@ const MODELS_DEV_PROVIDER_MAP: Record<string, string[]> = {
   openai: ["openai", "cx"], // cx = Codex (uses OpenAI models)
   anthropic: ["anthropic", "cc"], // cc = Claude Code
   google: ["gemini", "gemini-cli"],
+  "google-vertex": ["gemini", "vertex"],
+  "google-vertex-anthropic": ["anthropic", "cc", "vertex"],
   vertex_ai: ["gemini", "vertex"],
   deepseek: ["deepseek", "if"], // if = Qoder (routes through DeepSeek)
   groq: ["groq"],
   xai: ["xai"],
   mistral: ["mistral"],
+  togetherai: ["together", "openrouter"],
   together_ai: ["together", "openrouter"],
+  "fireworks-ai": ["fireworks"],
   fireworks: ["fireworks"],
   cerebras: ["cerebras"],
   cohere: ["cohere"],
@@ -172,11 +176,25 @@ const MODELS_DEV_PROVIDER_MAP: Record<string, string[]> = {
   perplexity: ["pplx", "perplexity"],
   // OAuth / special providers
   bedrock: ["kiro", "kr"], // kr = Kiro (AWS Bedrock)
+  "github-copilot": ["github", "gh"],
+  "github-models": ["github", "gh"],
+  kilo: ["kilocode", "kc", "kilo-gateway"],
+  kilocode: ["kilocode", "kc", "kilo-gateway"],
+  "kimi-for-coding": ["kimi-coding", "kmc", "kimi-coding-apikey", "kmca"],
+  opencode: ["opencode-zen"],
+  "opencode-go": ["opencode-go"],
   // Additional providers that may overlap with OmniRoute
   alibaba: ["ali", "alibaba", "bcp", "alicode", "alicode-intl"],
+  "alibaba-cn": ["ali", "alibaba", "bcp"],
+  "alibaba-coding-plan": ["alicode", "alicode-intl"],
+  "alibaba-coding-plan-cn": ["alicode"],
   zai: ["zai", "glm"], // GLM models via Z.AI
+  "zai-coding-plan": ["zai", "glm"],
+  moonshotai: ["moonshot", "kimi"],
+  "moonshotai-cn": ["moonshot", "kimi"],
   moonshot: ["moonshot", "kimi", "kimi-coding", "kmc", "kmca"],
   minimax: ["minimax", "minimax-cn"],
+  "minimax-cn": ["minimax-cn"],
   longcat: ["lc", "longcat"],
   pollinations: ["pol", "pollinations"],
   puter: ["pu", "puter"],
@@ -184,7 +202,6 @@ const MODELS_DEV_PROVIDER_MAP: Record<string, string[]> = {
   scaleway: ["scw"],
   ollama: ["ollamacloud", "ollama-cloud"],
   blackbox: ["bb", "blackbox"],
-  kilocode: ["kc", "kilocode"],
   cline: ["cl", "cline"],
   cursor: ["cu", "cursor"],
   github: ["gh", "github"],

tests/unit/model-capabilities-registry.test.ts

@@ -49,7 +49,7 @@ test.after(() => {
   fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
 });
 
-test("canonical model capability resolver merges models.dev data and keeps static overrides authoritative", () => {
+test("canonical model capability resolver lets exact synced metadata override global specs", () => {
   modelsDevSync.saveModelsDevCapabilities({
     openai: {
       "gpt-4o": buildCapability({
@@ -92,11 +92,11 @@ test("canonical model capability resolver merges models.dev data and keeps stati
   const geminiHigh = modelCapabilities.getResolvedModelCapabilities(
     "antigravity/gemini-3.1-pro-high"
   );
-  assert.equal(geminiHigh.toolCalling, true);
-  assert.equal(geminiHigh.reasoning, true);
-  assert.equal(geminiHigh.supportsThinking, true);
-  assert.equal(geminiHigh.contextWindow, 1048576);
-  assert.equal(geminiHigh.maxOutputTokens, 65535);
+  assert.equal(geminiHigh.toolCalling, false);
+  assert.equal(geminiHigh.reasoning, false);
+  assert.equal(geminiHigh.supportsThinking, false);
+  assert.equal(geminiHigh.contextWindow, 1024);
+  assert.equal(geminiHigh.maxOutputTokens, 9999);
   assert.equal(geminiHigh.defaultThinkingBudget, 24576);
   assert.equal(
     modelCapabilities.capThinkingBudget("antigravity/gemini-3.1-pro-high", 40000),

tests/unit/models-catalog-route.test.ts

@@ -14,6 +14,7 @@ const modelsDb = await import("../../src/lib/db/models.ts");
 const combosDb = await import("../../src/lib/db/combos.ts");
 const settingsDb = await import("../../src/lib/db/settings.ts");
 const apiKeysDb = await import("../../src/lib/db/apiKeys.ts");
+const modelsDevSync = await import("../../src/lib/modelsDevSync.ts");
 const v1ModelsCatalog = await import("../../src/app/api/v1/models/catalog.ts");
 
 async function resetStorage() {
@@ -689,6 +690,99 @@ test("v1 models catalog exposes provider-prefixed custom models, filters by raw
   assert.equal(providerAlias.parent, "cl/demo-custom");
 });
 
+test("v1 models catalog uses synced models.dev limits instead of provider defaults", async () => {
+  await seedConnection("openai", { name: "openai-models-dev" });
+
+  try {
+    modelsDevSync.saveModelsDevCapabilities({
+      openai: {
+        "gpt-5.5": {
+          tool_call: true,
+          reasoning: true,
+          attachment: true,
+          structured_output: true,
+          temperature: true,
+          modalities_input: JSON.stringify(["text", "image"]),
+          modalities_output: JSON.stringify(["text"]),
+          knowledge_cutoff: null,
+          release_date: null,
+          last_updated: null,
+          status: null,
+          family: "gpt-5",
+          open_weights: false,
+          limit_context: 1050000,
+          limit_input: 1050000,
+          limit_output: 128000,
+          interleaved_field: null,
+        },
+      },
+    });
+
+    const response = await v1ModelsCatalog.getUnifiedModelsResponse(
+      new Request("http://localhost/api/v1/models")
+    );
+    const body = (await response.json()) as any;
+    const model = body.data.find((item) => item.id === "openai/gpt-5.5");
+
+    assert.equal(response.status, 200);
+    assert.ok(model);
+    assert.equal(model.context_length, 1050000);
+    assert.equal(model.max_input_tokens, 1050000);
+    assert.equal(model.max_output_tokens, 128000);
+  } finally {
+    modelsDevSync.saveModelsDevCapabilities({});
+  }
+});
+
+test("v1 models catalog lets provider-specific synced limits beat global static specs", async () => {
+  await seedConnection("github", {
+    authType: "oauth",
+    name: "github-copilot-models-dev",
+    apiKey: null,
+    accessToken: "github-access",
+  });
+
+  try {
+    modelsDevSync.saveModelsDevCapabilities({
+      github: {
+        "gpt-5.5": {
+          tool_call: true,
+          reasoning: true,
+          attachment: true,
+          structured_output: true,
+          temperature: true,
+          modalities_input: JSON.stringify(["text", "image"]),
+          modalities_output: JSON.stringify(["text"]),
+          knowledge_cutoff: null,
+          release_date: null,
+          last_updated: null,
+          status: null,
+          family: "gpt-5",
+          open_weights: false,
+          limit_context: 400000,
+          limit_input: 272000,
+          limit_output: 128000,
+          interleaved_field: null,
+        },
+      },
+    });
+
+    const response = await v1ModelsCatalog.getUnifiedModelsResponse(
+      new Request("http://localhost/api/v1/models")
+    );
+    const body = (await response.json()) as any;
+    const model = body.data.find((item) => item.id === "gh/gpt-5.5");
+
+    assert.equal(response.status, 200);
+    assert.ok(model);
+    assert.equal(model.context_length, 400000);
+    assert.equal(model.max_input_tokens, 272000);
+    assert.equal(model.max_output_tokens, 128000);
+  } finally {
+    modelsDevSync.saveModelsDevCapabilities({});
+  }
+});
+
 test("v1 models catalog returns 500 when model compatibility lookup crashes", async () => {
   await seedConnection("openai", { name: "openai-compat-crash" });
 

tests/unit/modelsDevSync.test.ts

@@ -412,6 +412,19 @@ describe("modelsDevSync — mapProviderId", () => {
   it("maps moonshot to the canonical provider plus Kimi aliases", () => {
     assert.deepEqual(mapProviderId("moonshot"), ["moonshot", "kimi", "kimi-coding", "kmc", "kmca"]);
   });
+
+  it("maps current models.dev provider IDs used by OmniRoute-compatible providers", () => {
+    assert.deepEqual(mapProviderId("github-copilot"), ["github", "gh"]);
+    assert.deepEqual(mapProviderId("kilo"), ["kilocode", "kc", "kilo-gateway"]);
+    assert.deepEqual(mapProviderId("kimi-for-coding"), [
+      "kimi-coding",
+      "kmc",
+      "kimi-coding-apikey",
+      "kmca",
+    ]);
+    assert.deepEqual(mapProviderId("fireworks-ai"), ["fireworks"]);
+    assert.deepEqual(mapProviderId("togetherai"), ["together", "openrouter"]);
+  });
 });
 
 describe("modelsDevSync — fetchModelsDev (live API)", () => {