fix(provider): replay reasoning_content on v4 deepseek turns

Author: TheGreatAxios

src/provider/openai-compatible-adapter.test.ts

@@ -150,20 +150,58 @@ describe("openai-compatible adapter reasoning_content handling", () => {
       .messages;
   }
 
-  test("strips reasoning_content from input messages for DeepSeek models", () => {
-    const assistant = messagesFor("deepseek-v4").find(
+  test("strips reasoning_content from input messages for pre-V4 DeepSeek models", () => {
+    const assistant = messagesFor("deepseek-v3").find(
       (m) => m["role"] === "assistant",
     );
     expect(assistant).toBeDefined();
     expect("reasoning_content" in defined(assistant)).toBe(false);
   });
 
+  test.each([
+    "deepseek-v4",
+    "deepseek-v4-pro",
+    "deepseek-v4-flash",
+    "deepseek-v4-flash-vision-exp",
+  ])("V4 keeps reasoning_content on replayed assistant turns (%s)", (model) => {
+    const assistant = messagesFor(model).find((m) => m["role"] === "assistant");
+    expect(assistant).toBeDefined();
+    expect(assistant?.["reasoning_content"]).toBe("ponder");
+  });
+
   test("keeps reasoning_content for non-DeepSeek models", () => {
     const assistant = messagesFor("kimi-k2").find(
       (m) => m["role"] === "assistant",
     );
     expect(assistant?.["reasoning_content"]).toBe("ponder");
   });
+
+  test("non-V4/non-DeepSeek models are byte-identical on reasoning_content", () => {
+    for (const model of [
+      "gpt-5",
+      "gpt-5.1",
+      "kimi-k2",
+      "grok-4.6",
+      "claude-sonnet-4.5",
+      "glm-5.3",
+      "openai-compatible-generic",
+    ]) {
+      const assistant = messagesFor(model).find(
+        (m) => m["role"] === "assistant",
+      );
+      expect(assistant?.["reasoning_content"]).toBe("ponder");
+    }
+  });
+
+  test("pre-V4 DeepSeek ids still strip reasoning_content (byte-identical)", () => {
+    for (const model of ["deepseek-v3", "deepseek-r1"]) {
+      const assistant = messagesFor(model).find(
+        (m) => m["role"] === "assistant",
+      );
+      expect(assistant).toBeDefined();
+      expect("reasoning_content" in defined(assistant)).toBe(false);
+    }
+  });
 });
 
 describe("openai-compatible adapter V4 effort wiring", () => {

src/provider/openai-compatible-adapter.ts

@@ -52,11 +52,17 @@ export function createOpenAICompatibleAdapter(
     const hasProviderOptions =
       providerOptions !== undefined && Object.keys(providerOptions).length > 0;
     // DeepSeek returns HTTP 400 if `reasoning_content` appears in input messages,
-    // whereas the base adapter emits it for any model with thinking enabled.
+    // whereas the base adapter emits it for any model with thinking enabled. V4
+    // keeps each turn's reasoning on replay, so only pre-V4 DeepSeek strips it.
     const m = model.toLowerCase().trim();
-    const stripReasoning = m.includes("deepseek");
-    needsDeepSeekPatch = stripReasoning;
-    if (!hasProviderOptions && !stripReasoning) return ensureAccept(built);
+    const isDeepSeek = m.includes("deepseek");
+    const stripReasoning = isDeepSeek && !/^deepseek-v4/i.test(m);
+    needsDeepSeekPatch = isDeepSeek;
+    // V4 always runs the settings step below, even with no providerOptions, so
+    // its sampling/stream defaults reach the wire; non-V4 models without
+    // providerOptions short-circuit here and stay byte-identical.
+    if (!hasProviderOptions && !stripReasoning && !/^deepseek-v4/i.test(m))
+      return ensureAccept(built);
 
     const body = JSON.parse(built.body) as Record<string, unknown>;
     if (hasProviderOptions) Object.assign(body, providerOptions);