diff --git a/internal/agent/modes/interactive.go b/internal/agent/modes/interactive.go
index d4570ff..b960635 100644
--- a/internal/agent/modes/interactive.go
+++ b/internal/agent/modes/interactive.go
@@ -560,7 +560,7 @@ func (i *Interactive) redraw() {
 
 	cols, _ := i.cfg.Terminal.Size()
 	if i.agent != nil {
-		i.view.Messages = i.agent.Messages()
+		i.view.Messages = filterHiddenTranscriptMessages(i.agent.Messages())
 	} else {
 		i.view.Messages = nil
 	}
@@ -896,6 +896,33 @@ func alignSliceStartToImageBlock(chat []string, start, end int) int {
 	return start
 }
 
+const hiddenOpenAIImageMirrorPrefix = "Tool output included the following image content:"
+
+func filterHiddenTranscriptMessages(msgs []provider.Message) []provider.Message {
+	if len(msgs) == 0 {
+		return nil
+	}
+	out := make([]provider.Message, 0, len(msgs))
+	for _, m := range msgs {
+		if isHiddenTranscriptMessage(m) {
+			continue
+		}
+		out = append(out, m)
+	}
+	return out
+}
+
+func isHiddenTranscriptMessage(m provider.Message) bool {
+	if m.Role != provider.RoleUser || len(m.Content) == 0 {
+		return false
+	}
+	tb, ok := m.Content[0].(provider.TextBlock)
+	if !ok {
+		return false
+	}
+	return strings.TrimSpace(tb.Text) == hiddenOpenAIImageMirrorPrefix
+}
+
 func clipBottomClippedImages(lines []string) []string {
 	if len(lines) == 0 {
 		return lines
diff --git a/internal/core/agent.go b/internal/core/agent.go
index 7c32c48..75d36f0 100644
--- a/internal/core/agent.go
+++ b/internal/core/agent.go
@@ -181,6 +181,18 @@ func (a *Agent) runLoop(ctx context.Context, sink func(AgentEvent)) error {
 			toolMsg, hadError := a.executeTools(ctx, assistantMsg, sink)
 			a.mu.Lock()
 			a.messages = append(a.messages, toolMsg)
+			// OpenAI's chat-completions tool message shape is text-centric.
+			// Vision models reliably consume images when they arrive as user
+			// content, so when a tool result contains images we mirror them
+			// into a synthetic user message immediately after the tool result.
+			// This keeps the transcript self-contained for providers that can
+			// see image blocks in tool messages while making OpenAI vision
+			// models actually receive the image bytes.
+			if a.Client != nil && a.Client.Name() == "openai" {
+				if mirror := mirrorToolImagesAsUser(toolMsg); len(mirror.Content) > 0 {
+					a.messages = append(a.messages, mirror)
+				}
+			}
 			a.mu.Unlock()
 			// If context was cancelled during tool execution, bail out.
 			if err := ctx.Err(); err != nil {
@@ -388,6 +400,35 @@ func (a *Agent) runOneTool(ctx context.Context, tc provider.ToolCallBlock, sink
 // extractText concatenates all TextBlock content in a message. Used
 // by BeforeAssistantMessage so guards see a single string instead of
 // having to walk provider.Content themselves.
+func mirrorToolImagesAsUser(msg provider.Message) provider.Message {
+	var content []provider.Content
+	for _, c := range msg.Content {
+		tr, ok := c.(provider.ToolResultBlock)
+		if !ok {
+			continue
+		}
+		for _, inner := range tr.Content {
+			switch v := inner.(type) {
+			case provider.TextBlock:
+				// Keep short textual context so the model understands why
+				// the images appeared, but don't duplicate giant read
+				// outputs verbatim.
+				if len(v.Text) > 0 && len(v.Text) <= 500 {
+					content = append(content, v)
+				}
+			case provider.ImageBlock:
+				content = append(content, v)
+			}
+		}
+	}
+	if len(content) == 0 {
+		return provider.Message{}
+	}
+	prefix := provider.TextBlock{Text: "Tool output included the following image content:"}
+	content = append([]provider.Content{prefix}, content...)
+	return provider.Message{Role: provider.RoleUser, Content: content, Time: time.Now()}
+}
+
 func extractText(msg provider.Message) string {
 	var out string
 	for _, c := range msg.Content {
diff --git a/internal/provider/openai.go b/internal/provider/openai.go
index 2a5c04d..7938839 100644
--- a/internal/provider/openai.go
+++ b/internal/provider/openai.go
@@ -174,25 +174,16 @@ func (c *openaiClient) buildRequest(req Request) (*oaiRequest, error) {
 			}
 			out.Messages = append(out.Messages, am)
 		case RoleTool:
-			// Each ToolResultBlock becomes its own tool message.
+			// Each ToolResultBlock becomes its own tool message. Preserve
+			// image blocks for vision-capable OpenAI models instead of
+			// flattening the tool output to plain text.
 			for _, b := range msg.Content {
 				if tr, ok := b.(ToolResultBlock); ok {
-					var text strings.Builder
-					for _, inner := range tr.Content {
-						if tb, ok := inner.(TextBlock); ok {
-							if text.Len() > 0 {
-								text.WriteString("\n")
-							}
-							text.WriteString(tb.Text)
-						}
-					}
-					if tr.IsError && text.Len() > 0 {
-						text.WriteString(" [error]")
-					}
+					content := buildOAIToolContent(tr.Content, tr.IsError)
 					out.Messages = append(out.Messages, oaiMessage{
 						Role:       "tool",
 						ToolCallID: tr.CallID,
-						Content:    text.String(),
+						Content:    content,
 					})
 				}
 			}
@@ -234,6 +225,36 @@ func buildOAIUserContent(blocks []Content) interface{} {
 		}
 		return sb.String()
 	}
+	return buildOAIContentBlocks(blocks, false)
+}
+
+func buildOAIToolContent(blocks []Content, isError bool) interface{} {
+	hasImage := false
+	for _, b := range blocks {
+		if _, ok := b.(ImageBlock); ok {
+			hasImage = true
+			break
+		}
+	}
+	if !hasImage {
+		var sb strings.Builder
+		for _, b := range blocks {
+			if tb, ok := b.(TextBlock); ok {
+				if sb.Len() > 0 {
+					sb.WriteString("\n")
+				}
+				sb.WriteString(tb.Text)
+			}
+		}
+		if isError && sb.Len() > 0 {
+			sb.WriteString(" [error]")
+		}
+		return sb.String()
+	}
+	return buildOAIContentBlocks(blocks, isError)
+}
+
+func buildOAIContentBlocks(blocks []Content, isError bool) []interface{} {
 	var arr []interface{}
 	for _, b := range blocks {
 		switch v := b.(type) {
@@ -246,6 +267,9 @@ func buildOAIUserContent(blocks []Content) interface{} {
 			arr = append(arr, img)
 		}
 	}
+	if isError {
+		arr = append(arr, oaiContentText{Type: "text", Text: "[error]"})
+	}
 	return arr
 }