Merge pull request #56 from newCman1/fix/deepseek-vision-image-support

fix(models): add image support for deepseek/deepseek-v4-flash-vision-exp

src/models.ts

@@ -43,6 +43,7 @@ export const MODEL_INPUT_MODALITIES: Readonly<Record<string, readonly CommandCod
   "meta/muse-spark-1.1": ["text", "image"],
   "meta/muse-spark-1.2": ["text", "image"],
   "meta/muse-spark-1.2-contributor": ["text", "image"],
+  "deepseek/deepseek-v4-flash-vision-exp": ["text", "image"],
   "moonshotai/Kimi-K2.5": ["text", "image"],
   "moonshotai/Kimi-K2.6": ["text", "image"],
   "moonshotai/Kimi-K2.7-Code": ["text", "image"],

tests/test-models.ts

@@ -103,11 +103,16 @@ describe("commandCodeModelsFromApiResponse()", () => {
   it("matches command-code@1.32.1 image input capabilities", () => {
     assert.deepEqual(inputModalitiesForModel("gpt-5.6-luna"), ["text", "image"])
     assert.deepEqual(inputModalitiesForModel("meta/muse-spark-1.2"), ["text", "image"])
+    assert.deepEqual(inputModalitiesForModel("deepseek/deepseek-v4-flash-vision-exp"), [
+      "text",
+      "image",
+    ])
     assert.deepEqual(inputModalitiesForModel("deepseek/deepseek-v4-pro"), ["text"])
     assert.deepEqual(inputModalitiesForModel("unknown-new-model"), ["text"])
     assert.equal(modelSupportsImageInput("gpt-5.6-luna"), true)
+    assert.equal(modelSupportsImageInput("deepseek/deepseek-v4-flash-vision-exp"), true)
     assert.equal(modelSupportsImageInput("deepseek/deepseek-v4-pro"), false)
-    assert.equal(Object.keys(MODEL_INPUT_MODALITIES).length, 37)
+    assert.equal(Object.keys(MODEL_INPUT_MODALITIES).length, 38)
   })
 
   it("marks only known reasoning models as reasoning-capable", () => {

tests/test-stream.ts

@@ -173,6 +173,87 @@ describe("streamCommandCode — successful streams", () => {
     )
   })
 
+  it("forwards a tool-result image as a following user image for vision-capable models", async () => {
+    server.mockResponse({
+      type: "success",
+      events: [JSON.stringify({ type: "finish", finishReason: "stop" })],
+    })
+    const { streamCommandCode } = createTestDeps({ apiBase: server.baseUrl() })
+
+    const events = await collectEvents(
+      streamCommandCode(
+        makeModel({ id: "deepseek/deepseek-v4-flash-vision-exp" }),
+        makeContext({
+          messages: [
+            { role: "user", content: "read the image" },
+            {
+              role: "assistant",
+              content: [{ type: "toolCall", id: "c1", name: "read", arguments: {} }],
+            },
+            {
+              role: "toolResult",
+              toolCallId: "c1",
+              toolName: "read",
+              content: [
+                { type: "text", text: "image attached" },
+                { type: "image", data: "aGVsbG8=", mimeType: "image/png" },
+              ],
+            },
+          ],
+        }),
+        { apiKey: "mock-key" },
+      ),
+    )
+
+    // No error: the tool-result image must not be rejected for this model.
+    assert.equal(events.at(-1)?.type, "done")
+
+    const body = server.lastRequestBody()
+    // The tool-result text is forwarded on the tool message at index 2.
+    assert.equal(
+      objectAt(body, ["params", "messages", "2", "content", "0", "output", "value"]),
+      "image attached",
+    )
+    // The tool-result image is forwarded as a following user image message at index 3.
+    assert.equal(objectAt(body, ["params", "messages", "3", "role"]), "user")
+    assert.equal(
+      objectAt(body, ["params", "messages", "3", "content", "0", "image"]),
+      "data:image/png;base64,aGVsbG8=",
+    )
+  })
+
+  it("rejects a tool-result image before network access for text-only models", async () => {
+    const { streamCommandCode } = createTestDeps({ apiBase: server.baseUrl() })
+
+    const events = await collectEvents(
+      streamCommandCode(
+        makeModel({ id: "deepseek/deepseek-v4-pro" }),
+        makeContext({
+          messages: [
+            { role: "user", content: "read the image" },
+            {
+              role: "assistant",
+              content: [{ type: "toolCall", id: "c1", name: "read", arguments: {} }],
+            },
+            {
+              role: "toolResult",
+              toolCallId: "c1",
+              toolName: "read",
+              content: [{ type: "image", data: "aGVsbG8=", mimeType: "image/png" }],
+            },
+          ],
+        }),
+        { apiKey: "mock-key" },
+      ),
+    )
+
+    const lastEvent = events.at(-1)
+    assert.equal(lastEvent?.type, "error")
+    if (lastEvent?.type !== "error") throw new Error("expected error event")
+    assert.match(lastEvent.error.errorMessage ?? "", /does not support image content/i)
+    assert.equal(server.requestCount(), 0)
+  })
+
   it("rejects images before network access for text-only models", async () => {
     const { streamCommandCode } = createTestDeps({ apiBase: server.baseUrl() })