From 772f6a3d5ea95b09bdd0a3f2645a32769b739765 Mon Sep 17 00:00:00 2001
From: Peter Steinberger <steipete@gmail.com>
Date: Wed, 26 Aug 2026 10:24:03 -0700
Subject: [PATCH] fix(deepinfra): preserve manifest reasoning during live
 discovery (#129998)

---
 extensions/deepinfra/provider-models.test.ts | 83 ++++++++++++++++++++
 extensions/deepinfra/provider-models.ts      | 18 ++---
 2 files changed, 91 insertions(+), 10 deletions(-)

diff --git a/extensions/deepinfra/provider-models.test.ts b/extensions/deepinfra/provider-models.test.ts
index 8e50b3252750..9fadbae9353c 100644
--- a/extensions/deepinfra/provider-models.test.ts
+++ b/extensions/deepinfra/provider-models.test.ts
@@ -263,6 +263,89 @@ describe("discoverDeepInfraModels (chat-only shim)", () => {
     });
   });
 
+  it("preserves bundled reasoning and compat while keeping live model facts authoritative", async () => {
+    const rows = [
+      makeAgentModelEntry({
+        id: "deepseek-ai/DeepSeek-V4-Pro",
+        metadata: {
+          context_length: 96000,
+          max_tokens: 4096,
+          pricing: { input_tokens: 4, output_tokens: 8, cache_read_tokens: 0.4 },
+          tags: ["chat"],
+        },
+      }),
+      makeAgentModelEntry({
+        id: "stepfun-ai/Step-3.7-Flash",
+        metadata: {
+          context_length: 192000,
+          max_tokens: 16384,
+          pricing: { input_tokens: 0.2, output_tokens: 1.15 },
+          tags: ["chat", "vlm", "vision"],
+        },
+      }),
+      makeAgentModelEntry({
+        id: "deepseek-ai/DeepSeek-V3.2",
+        metadata: {
+          context_length: 64000,
+          max_tokens: 8192,
+          pricing: { input_tokens: 1, output_tokens: 2 },
+          tags: ["chat", "reasoning"],
+        },
+      }),
+      makeAgentModelEntry({
+        id: "unlisted/no-reasoning",
+        metadata: { context_length: 32000, max_tokens: 2048, pricing: {}, tags: ["chat"] },
+      }),
+      makeAgentModelEntry({
+        id: "unlisted/with-reasoning",
+        metadata: {
+          context_length: 48000,
+          max_tokens: 4096,
+          pricing: {},
+          tags: ["chat", "reasoning_effort"],
+        },
+      }),
+    ];
+    const mockFetch = vi.fn().mockResolvedValue(jsonResponse({ data: rows }));
+    DEEPINFRA_MODEL_CATALOG.push(DEEPINFRA_MODEL_CATALOG[0]!);
+
+    try {
+      await withFetchPathTest(mockFetch, { DEEPINFRA_API_KEY: "sk-test" }, async () => {
+        const models = await discoverDeepInfraModels();
+
+        expect(models.slice(0, rows.length)).toMatchObject([
+          {
+            id: "deepseek-ai/DeepSeek-V4-Pro",
+            reasoning: true,
+            input: ["text"],
+            contextWindow: 96000,
+            maxTokens: 4096,
+            cost: { input: 4, output: 8, cacheRead: 0.4, cacheWrite: 0 },
+            compat: {
+              codeMode: "capable",
+              supportsUsageInStreaming: true,
+              thinkingFormat: "deepseek",
+            },
+          },
+          {
+            id: "stepfun-ai/Step-3.7-Flash",
+            reasoning: true,
+            input: ["text", "image"],
+            contextWindow: 192000,
+            maxTokens: 16384,
+            cost: { input: 0.2, output: 1.15, cacheRead: 0, cacheWrite: 0 },
+          },
+          { id: "deepseek-ai/DeepSeek-V3.2", reasoning: false },
+          { id: "unlisted/no-reasoning", reasoning: false },
+          { id: "unlisted/with-reasoning", reasoning: true },
+        ]);
+        expect(new Set(models.map((model) => model.id)).size).toBe(models.length);
+      });
+    } finally {
+      DEEPINFRA_MODEL_CATALOG.pop();
+    }
+  });
+
   it("skips entries with no metadata or no surface tag, and deduplicates ids", async () => {
     const mockFetch = vi.fn().mockResolvedValue(
       jsonResponse({
diff --git a/extensions/deepinfra/provider-models.ts b/extensions/deepinfra/provider-models.ts
index 20aa29b85b0c..f39cc37df213 100644
--- a/extensions/deepinfra/provider-models.ts
+++ b/extensions/deepinfra/provider-models.ts
@@ -359,13 +359,15 @@ export function buildDeepInfraModelDefinition(model: ModelDefinitionConfig): Mod
 }
 
 function chatSurfaceModelToModelDefinition(model: DeepInfraSurfaceModel): ModelDefinitionConfig {
+  const manifestModel = DEEPINFRA_MODEL_CATALOG.find((entry) => entry.id === model.id);
   const input: Array<"text" | "image"> = model.tags.includes("vlm") ? ["text", "image"] : ["text"];
   const reasoning = model.tags.includes("reasoning") || model.tags.includes("reasoning_effort");
   return buildDeepInfraModelDefinition({
     id: model.id,
     name: model.name,
-    reasoning,
+    reasoning: manifestModel?.reasoning ?? reasoning,
     input,
+    ...(manifestModel?.compat ? { compat: manifestModel.compat } : {}),
     contextWindow: model.contextWindow ?? DEEPINFRA_DEFAULT_CONTEXT_WINDOW,
     maxTokens: model.maxTokens ?? DEEPINFRA_DEFAULT_MAX_TOKENS,
     cost: {
@@ -480,14 +482,10 @@ export async function discoverDeepInfraModels(options?: {
   }
   const liveModels = chatModels.map(chatSurfaceModelToModelDefinition);
   const seen = new Set(liveModels.map((model) => model.id));
-  const manifestModels = DEEPINFRA_MODEL_CATALOG.map(buildDeepInfraModelDefinition).filter(
-    (model) => {
-      if (seen.has(model.id)) {
-        return false;
-      }
-      seen.add(model.id);
-      return true;
-    },
-  );
+  const manifestModels = DEEPINFRA_MODEL_CATALOG.filter((model) => {
+    const unseen = !seen.has(model.id);
+    seen.add(model.id);
+    return unseen;
+  }).map(buildDeepInfraModelDefinition);
   return [...liveModels, ...manifestModels];
 }