fix(channel): strip agent tags from content parts in outbound messages

StripAgentTags was only applied to the merged content string but not to
individual ContentParts. On channels that don't support RichText (e.g.
Telegram), buildChannelMessage joins part texts directly, causing raw
<attachments>/<reactions>/<speech> blocks to appear in the final message.

internal/conversation/flow/assistant_output.go

@@ -30,6 +30,7 @@ func ExtractAssistantOutputs(messages []conversation.ModelMessage) []conversatio
 			continue
 		}
 		content = agent.StripAgentTags(content)
+		parts = stripAgentTagsFromParts(parts)
 		outputs = append(outputs, conversation.AssistantOutput{Content: content, Parts: parts})
 	}
 	return outputs
@@ -87,6 +88,22 @@ func visibleContentText(parts []conversation.ContentPart) string {
 	return strings.TrimSpace(strings.Join(texts, "\n"))
 }
 
+func stripAgentTagsFromParts(parts []conversation.ContentPart) []conversation.ContentPart {
+	if len(parts) == 0 {
+		return nil
+	}
+	result := make([]conversation.ContentPart, 0, len(parts))
+	for _, p := range parts {
+		if strings.TrimSpace(p.Text) != "" {
+			p.Text = agent.StripAgentTags(p.Text)
+		}
+		if p.HasValue() {
+			result = append(result, p)
+		}
+	}
+	return result
+}
+
 func visibleContentPartText(part conversation.ContentPart) string {
 	if strings.TrimSpace(part.Text) != "" {
 		return part.Text