fix(deepseek): backfill v4 assistant reasoning replay

extensions/deepseek/deepseek.live.test.ts

@@ -176,4 +176,59 @@ describeLive("deepseek plugin live", () => {
     });
     expect(extractNonEmptyAssistantText(result.content).length).toBeGreaterThan(0);
   }, 60_000);
+
+  it("accepts V4 thinking replay after a prior plain assistant message", async () => {
+    const context: Context = {
+      messages: [
+        {
+          role: "user",
+          content: "Say hello.",
+          timestamp: Date.now() - 2,
+        },
+        {
+          role: "assistant",
+          api: "openai-completions",
+          provider: "openai",
+          model: "gpt-5.4",
+          content: [{ type: "text", text: "Hello." }],
+          usage: ZERO_USAGE,
+          stopReason: "stop",
+          timestamp: Date.now() - 1,
+        },
+        {
+          role: "user",
+          content: "Reply with exactly: ok",
+          timestamp: Date.now(),
+        },
+      ],
+    };
+    let capturedPayload: Record<string, unknown> | undefined;
+    const streamFn = createDeepSeekV4ThinkingWrapper(streamSimple, "high");
+    expect(streamFn).toBeDefined();
+
+    const stream = streamFn?.(resolveDeepSeekV4LiveModel(), context, {
+      apiKey: DEEPSEEK_KEY,
+      maxTokens: 64,
+      onPayload: (payload) => {
+        capturedPayload = payload as Record<string, unknown>;
+      },
+    });
+    expect(stream).toBeDefined();
+
+    const result = await (await stream!).result();
+    if (result.stopReason === "error") {
+      throw new Error(
+        result.errorMessage || "DeepSeek V4 plain replay returned error with no message",
+      );
+    }
+
+    const messages = capturedPayload?.messages;
+    expect(Array.isArray(messages)).toBe(true);
+    expect((messages as Array<Record<string, unknown>>)[1]).toMatchObject({
+      role: "assistant",
+      content: "Hello.",
+      reasoning_content: "",
+    });
+    expect(extractNonEmptyAssistantText(result.content).length).toBeGreaterThan(0);
+  }, 60_000);
 });

extensions/deepseek/index.test.ts

@@ -309,6 +309,73 @@ describe("deepseek provider plugin", () => {
     });
   });
 
+  it("adds blank reasoning_content for replayed plain assistant messages", async () => {
+    let capturedPayload: Record<string, unknown> | undefined;
+    const model = {
+      provider: "deepseek",
+      id: "deepseek-v4-pro",
+      name: "DeepSeek V4 Pro",
+      api: "openai-completions",
+      baseUrl: "https://api.deepseek.com",
+      reasoning: true,
+      input: ["text"],
+      cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+      contextWindow: 1_000_000,
+      maxTokens: 384_000,
+      compat: {
+        supportsUsageInStreaming: true,
+        supportsReasoningEffort: true,
+        maxTokensField: "max_tokens",
+      },
+    } as Model<"openai-completions">;
+    const context = {
+      messages: [
+        { role: "user", content: "hi", timestamp: 1 },
+        {
+          role: "assistant",
+          api: "openai-completions",
+          provider: "openai",
+          model: "gpt-5.4",
+          content: [{ type: "text", text: "Hello." }],
+          usage: {
+            input: 0,
+            output: 0,
+            cacheRead: 0,
+            cacheWrite: 0,
+            totalTokens: 0,
+            cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+          },
+          stopReason: "stop",
+          timestamp: 2,
+        },
+        { role: "user", content: "next", timestamp: 3 },
+      ],
+    } as Context;
+    const baseStreamFn = (
+      streamModel: Model<"openai-completions">,
+      streamContext: Context,
+      options?: { onPayload?: (payload: unknown, model: unknown) => unknown },
+    ) => {
+      capturedPayload = buildOpenAICompletionsParams(streamModel, streamContext, {
+        reasoning: "high",
+      } as never);
+      options?.onPayload?.(capturedPayload, streamModel);
+      const stream = createAssistantMessageEventStream();
+      queueMicrotask(() => stream.end());
+      return stream;
+    };
+
+    const wrapThinkingHigh = createDeepSeekV4ThinkingWrapper(baseStreamFn as never, "high");
+    expect(wrapThinkingHigh).toBeDefined();
+    await wrapThinkingHigh?.(model, context, {});
+
+    expect((capturedPayload?.messages as Array<Record<string, unknown>>)[1]).toMatchObject({
+      role: "assistant",
+      content: "Hello.",
+      reasoning_content: "",
+    });
+  });
+
   it("strips replayed reasoning_content when DeepSeek V4 thinking is disabled", async () => {
     let capturedPayload: Record<string, unknown> | undefined;
     const model = {

extensions/venice/index.test.ts

@@ -55,6 +55,7 @@ describe("venice provider plugin", () => {
               },
             ],
           },
+          { role: "assistant", content: "done" },
         ],
       };
       (options as { onPayload?: (payload: Record<string, unknown>) => void })?.onPayload?.(payload);
@@ -87,6 +88,11 @@ describe("venice provider plugin", () => {
             ],
             reasoning_content: "",
           },
+          {
+            role: "assistant",
+            content: "done",
+            reasoning_content: "",
+          },
         ],
       },
     ]);

extensions/venice/stream.ts

@@ -18,7 +18,7 @@ function ensureVeniceDeepSeekV4Replay(payload: Record<string, unknown>): void {
       continue;
     }
     const record = message as Record<string, unknown>;
-    if (record.role === "assistant" && Array.isArray(record.tool_calls)) {
+    if (record.role === "assistant") {
       record.reasoning_content ??= "";
     }
   }