gateway: an image model's picture reaches a Gemini client (#620, tao-bai: antigravity/gemini-3.1-flash-image 返回 parts: []) — through /v1beta generateContent and streamGenerateContent the answer came back with no parts, though candidatesTokenCount (1567) said the image was made. The IR had no event for a picture the model made: codeAssistDecoder (Antigravity, Gemini CLI and Factory's Gemini route) turned functionCall, thought and text parts into events and dropped inlineData, and geminiParts / geminiEncoder had nothing to write it back as. A new KImage event (Name the media type, Text the base64) carries it; the collector keeps it as an Image part, and the Gemini renderer and stream encoder send it as inlineData in its place among the text. A thought image (a draft, thought: true) stays with the reasoning and isn't sent, as before; thought text and signatures are untouched. Chat, Responses and Anthropic clients still don't get the picture (their renderers skip Image parts). TestGeminiImageModelReturnsItsImage (a fake Factory Gemini upstream, generateContent) and TestCodeAssistStreamsImages (Code Assist chunks through the Gemini stream encoder) fail without it with parts: []. Not tried against a real Antigravity account; the image part's thoughtSignature isn't returned to the client, so multi-turn image editing on Gemini 3 image models may still need it. go vet ./internal/..., go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui -count=1 ./... under a temp HOME pass.

internal/gateway/codeassist.go

@@ -524,6 +524,11 @@ func (d *codeAssistDecoder) decode(data string, emit func(Event)) error {
 				if p.Signature != "" {
 					emit(Event{Kind: KSig, Text: p.Signature})
 				}
+			case p.InlineData != nil && p.InlineData.Data != "":
+				// a picture an image model drew (gemini-*-image): its
+				// thought images, drafts on the way, stay with the
+				// reasoning above (#620)
+				emit(Event{Kind: KImage, Name: p.InlineData.MimeType, Text: p.InlineData.Data})
 			case p.Text != "":
 				emit(Event{Kind: KText, Text: p.Text})
 			}

internal/gateway/gemini.go

@@ -403,6 +403,10 @@ func geminiParts(parts []Part) []map[string]any {
 			if p.Text != "" {
 				out = append(out, map[string]any{"text": p.Text, "thought": true})
 			}
+		case Image:
+			if p.Data != "" {
+				out = append(out, map[string]any{"inlineData": map[string]any{"mimeType": p.MediaType, "data": p.Data}})
+			}
 		case ToolCall:
 			id := p.ID
 			if id == "" {
@@ -487,6 +491,9 @@ func (e *geminiEncoder) event(ev Event) {
 	case KThink:
 		e.flushTool()
 		e.chunk(geminiParts([]Part{{Kind: Thinking, Text: ev.Text}}), "", nil)
+	case KImage:
+		e.flushTool()
+		e.chunk(geminiParts([]Part{{Kind: Image, MediaType: ev.Name, Data: ev.Text}}), "", nil)
 	case KToolStart:
 		e.flushTool()
 		e.tool = &Part{Kind: ToolCall, ID: ev.ID, Name: ev.Name}

internal/gateway/gemini_image_test.go

@@ -0,0 +1,96 @@
+package gateway
+
+import (
+	"encoding/json"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/yetone/magpie/internal/provider"
+)
+
+// An image model's picture reaches a Gemini client's generateContent: the
+// upstream's inlineData part comes back, and a thought image (a draft)
+// stays out (#620).
+func TestGeminiImageModelReturnsItsImage(t *testing.T) {
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+	t.Setenv("XDG_CONFIG_HOME", filepath.Join(home, ".config"))
+	t.Setenv("XDG_CACHE_HOME", filepath.Join(home, ".cache"))
+	t.Setenv("XDG_DATA_HOME", filepath.Join(home, ".local", "share"))
+
+	f := &fake{t: t, reply: `data: {"candidates":[{"content":{"parts":[` +
+		`{"thought":true,"inlineData":{"mimeType":"image/png","data":"ZHJhZnQ="}},` +
+		`{"text":"Here it is."},` +
+		`{"inlineData":{"mimeType":"image/png","data":"aGVsbG8="},"thoughtSignature":"c2ln"}` +
+		`]},"finishReason":"STOP"}],"usageMetadata":{"promptTokenCount":5,"candidatesTokenCount":1567}}` + "\n\n"}
+	up := httptest.NewServer(f)
+	t.Cleanup(up.Close)
+	t.Cleanup(provider.FactoryBaseForTest(up.URL, up.URL+"/eu"))
+
+	auth, _ := json.Marshal(map[string]any{
+		"accessToken": "tok", "refreshToken": "r",
+		"expiresAt": time.Now().Add(time.Hour).UnixMilli(),
+		"orgId":     "org_D", "activeOrganizationId": "fac_D", "email": "d@example.com",
+	})
+	dir := filepath.Dir(provider.Path())
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	raw, _ := json.Marshal([]map[string]any{{
+		"agent": "factory", "user": "d@example.com", "plan": "pro", "on": true, "auth": json.RawMessage(auth),
+	}})
+	if err := os.WriteFile(filepath.Join(dir, "logins.json"), raw, 0o600); err != nil {
+		t.Fatal(err)
+	}
+
+	const ask = `{"contents":[{"role":"user","parts":[{"text":"draw a cat"}]}]}`
+	check := func(kind, body string) {
+		t.Helper()
+		if !strings.Contains(body, `"inlineData":{"data":"aGVsbG8=","mimeType":"image/png"}`) {
+			t.Errorf("%s: image dropped: %s", kind, body)
+		}
+		if strings.Contains(body, "ZHJhZnQ=") {
+			t.Errorf("%s: thought image sent: %s", kind, body)
+		}
+		if i, j := strings.Index(body, "Here it is."), strings.Index(body, "aGVsbG8="); i < 0 || j < i {
+			t.Errorf("%s: text and image out of order: %s", kind, body)
+		}
+	}
+
+	// Factory's Gemini route decodes with Antigravity's decoder.
+	code, body := post(t, "/v1beta/models/factory/gemini-3.1-pro-preview:generateContent", ask)
+	if code != 200 || strings.HasPrefix(body, "data:") {
+		t.Fatalf("json %d %s", code, body)
+	}
+	if f.path != "/api/llm/g/v1/generate" {
+		t.Fatalf("upstream %s", f.path)
+	}
+	check("generateContent", body)
+}
+
+// Antigravity and Gemini CLI wrap each chunk in {response}; streamed to a
+// Gemini client, the picture goes out as an inlineData chunk.
+func TestCodeAssistStreamsImages(t *testing.T) {
+	rec := httptest.NewRecorder()
+	enc := &geminiEncoder{w: newSSEWriter(rec), model: "antigravity/gemini-3.1-flash-image"}
+	d := &codeAssistDecoder{}
+	for _, line := range []string{
+		`{"response":{"candidates":[{"content":{"role":"model","parts":[{"thought":true,"inlineData":{"mimeType":"image/png","data":"ZHJhZnQ="}}]}}]}}`,
+		`{"response":{"candidates":[{"content":{"role":"model","parts":[{"inlineData":{"mimeType":"image/jpeg","data":"aGk="}}]},"finishReason":"STOP"}],"usageMetadata":{"promptTokenCount":5,"candidatesTokenCount":1567}}}`,
+	} {
+		d.decode(line, enc.event)
+	}
+	enc.finish()
+	body := rec.Body.String()
+	if !strings.Contains(body, `"inlineData":{"data":"aGk=","mimeType":"image/jpeg"}`) {
+		t.Errorf("image dropped: %s", body)
+	}
+	if strings.Contains(body, "ZHJhZnQ=") {
+		t.Errorf("thought image sent: %s", body)
+	}
+}

internal/gateway/ir.go

@@ -159,6 +159,7 @@ const (
 	KUsage                      // Usage
 	KError                      // Text
 	KSearch                     // Text (the query), Hits: a web search run for the model
+	KImage                      // Name (media type), Text (base64): an image the model made
 )
 
 // Event is one thing a streaming reply said.
@@ -308,6 +309,9 @@ func (c *collector) add(ev Event) {
 	case KSearch:
 		c.closeTool()
 		c.res.Parts = append(c.res.Parts, Part{Kind: Search, Text: ev.Text, Hits: ev.Hits})
+	case KImage:
+		c.closeTool()
+		c.res.Parts = append(c.res.Parts, Part{Kind: Image, MediaType: ev.Name, Data: ev.Text})
 	}
 }