fix(responses): keep images in tool results instead of dropping them

llm/transformer/openai/responses/outbound_convert.go

@@ -310,15 +310,42 @@ func convertToolMessageWithType(msg llm.Message, itemType string) Item {
 		output.Text = msg.Content.Content
 	} else if len(msg.Content.MultipleContent) > 0 {
 		for _, p := range msg.Content.MultipleContent {
-			if p.Type == "text" && p.Text != nil {
-				output.Items = append(output.Items, Item{
-					Type: "input_text",
-					Text: p.Text,
-				})
+			switch p.Type {
+			case "text":
+				if p.Text != nil {
+					output.Items = append(output.Items, Item{
+						Type: "input_text",
+						Text: p.Text,
+					})
+				}
+			case "image_url":
+				// Tool results can carry images (Codex's view_image, MCP screenshot
+				// tools, ...). Skipping them here leaves output empty, which the
+				// fallback below turns into "" — the model then sees a successful
+				// but blank tool result and has no way to tell that it lost data.
+				if p.ImageURL != nil {
+					output.Items = append(output.Items, Item{
+						Type:     "input_image",
+						ImageURL: &p.ImageURL.URL,
+						Detail:   p.ImageURL.Detail,
+					})
+				}
 			}
 		}
 	}
 
+	// Content was present but nothing survived the conversion (audio/video/document
+	// parts, which a tool result in this API cannot express). Name what was dropped
+	// rather than falling through to "": a blank-but-successful tool result is
+	// indistinguishable from a genuinely empty one, so the model reads lost data as
+	// "the tool returned nothing" and answers anyway.
+	if output.Text == nil && len(output.Items) == 0 && len(msg.Content.MultipleContent) > 0 {
+		dropped := lo.Uniq(lo.Map(msg.Content.MultipleContent, func(p llm.MessageContentPart, _ int) string {
+			return p.Type
+		}))
+		output.Text = lo.ToPtr("[axonhub] tool output omitted: unsupported content types: " + strings.Join(dropped, ", "))
+	}
+
 	// Some times the tool result is empty, so we need to add an empty string.
 	if output.Text == nil && len(output.Items) == 0 {
 		output.Text = lo.ToPtr("")

llm/transformer/openai/responses/outbound_convert_test.go

@@ -108,7 +108,7 @@ func TestConvertToolMessage(t *testing.T) {
 			},
 		},
 		{
-			name: "tool message with multiple content - mixed types (only text extracted)",
+			name: "tool message with multiple content - text and image preserved in order",
 			msg: llm.Message{
 				Role:       "tool",
 				ToolCallID: lo.ToPtr("call_789"),
@@ -139,6 +139,10 @@ func TestConvertToolMessage(t *testing.T) {
 						Type: "input_text",
 						Text: lo.ToPtr("Text result"),
 					},
+					{
+						Type:     "input_image",
+						ImageURL: lo.ToPtr("https://example.com/image.jpg"),
+					},
 					{
 						Type: "input_text",
 						Text: lo.ToPtr("More text"),
@@ -176,7 +180,7 @@ func TestConvertToolMessage(t *testing.T) {
 			},
 		},
 		{
-			name: "tool message with multiple content but no text parts",
+			name: "tool message with image and unsupported parts keeps the image",
 			msg: llm.Message{
 				Role:       "tool",
 				ToolCallID: lo.ToPtr("call_no_text"),
@@ -201,6 +205,125 @@ func TestConvertToolMessage(t *testing.T) {
 			expected: Item{
 				Type:   "function_call_output",
 				CallID: "call_no_text",
+				Output: &Input{
+					Items: []Item{
+						{
+							Type:     "input_image",
+							ImageURL: lo.ToPtr("https://example.com/image.jpg"),
+						},
+					},
+				},
+			},
+		},
+		{
+			// Shape produced by Codex's view_image tool: the result is a single
+			// base64 image with no text at all. Dropping it made the model see an
+			// empty-but-successful tool result and silently hallucinate instead.
+			name: "image-only tool result is preserved as input_image",
+			msg: llm.Message{
+				Role:       "tool",
+				ToolCallID: lo.ToPtr("call_view_image"),
+				Content: llm.MessageContent{
+					MultipleContent: []llm.MessageContentPart{
+						{
+							Type: "image_url",
+							ImageURL: &llm.ImageURL{
+								URL:    "data:image/png;base64,iVBORw0KGgo=",
+								Detail: lo.ToPtr("high"),
+							},
+						},
+					},
+				},
+			},
+			expected: Item{
+				Type:   "function_call_output",
+				CallID: "call_view_image",
+				Output: &Input{
+					Items: []Item{
+						{
+							Type:     "input_image",
+							ImageURL: lo.ToPtr("data:image/png;base64,iVBORw0KGgo="),
+							Detail:   lo.ToPtr("high"),
+						},
+					},
+				},
+			},
+		},
+		{
+			name: "tool result with text and image keeps both in order",
+			msg: llm.Message{
+				Role:       "tool",
+				ToolCallID: lo.ToPtr("call_mixed"),
+				Content: llm.MessageContent{
+					MultipleContent: []llm.MessageContentPart{
+						{
+							Type: "text",
+							Text: lo.ToPtr("screenshot attached"),
+						},
+						{
+							Type: "image_url",
+							ImageURL: &llm.ImageURL{
+								URL: "https://example.com/shot.png",
+							},
+						},
+					},
+				},
+			},
+			expected: Item{
+				Type:   "function_call_output",
+				CallID: "call_mixed",
+				Output: &Input{
+					Items: []Item{
+						{
+							Type: "input_text",
+							Text: lo.ToPtr("screenshot attached"),
+						},
+						{
+							Type:     "input_image",
+							ImageURL: lo.ToPtr("https://example.com/shot.png"),
+						},
+					},
+				},
+			},
+		},
+		{
+			// A tool result whose parts cannot be expressed here must not look like
+			// an empty-but-successful result, otherwise the model cannot tell that
+			// data was lost.
+			name: "tool message with only unsupported parts reports what was dropped",
+			msg: llm.Message{
+				Role:       "tool",
+				ToolCallID: lo.ToPtr("call_audio_only"),
+				Content: llm.MessageContent{
+					MultipleContent: []llm.MessageContentPart{
+						{
+							Type: "input_audio",
+							InputAudio: &llm.InputAudio{
+								Data:   "audio-data",
+								Format: "wav",
+							},
+						},
+					},
+				},
+			},
+			expected: Item{
+				Type:   "function_call_output",
+				CallID: "call_audio_only",
+				Output: &Input{
+					Text: lo.ToPtr("[axonhub] tool output omitted: unsupported content types: input_audio"),
+				},
+			},
+		},
+		{
+			name: "tool message with no content at all still uses the empty string",
+			msg: llm.Message{
+				Role:       "tool",
+				ToolCallID: lo.ToPtr("call_truly_empty"),
+				Content:    llm.MessageContent{MultipleContent: []llm.MessageContentPart{}},
+			},
+			expected: Item{
+				Type:   "function_call_output",
+				CallID: "call_truly_empty",
 				Output: &Input{
 					Text: lo.ToPtr(""),
 				},