Fixes

src/api/providers/openrouter.ts

@@ -46,6 +46,7 @@ export class OpenRouterHandler implements ApiHandler {
 					],
 				}
 				// Add cache_control to the last two user messages
+				// (note: this works because we only ever add one user message at a time, but if we added multiple we'd need to mark the user message before the last assistant message)
 				const lastTwoUserMessages = openAiMessages.filter((msg) => msg.role === "user").slice(-2)
 				lastTwoUserMessages.forEach((msg) => {
 					if (typeof msg.content === "string") {

src/api/transform/openai-format.ts

@@ -66,15 +66,16 @@ export function convertToOpenAiMessages(
 				// "Messages following `tool_use` blocks must begin with a matching number of `tool_result` blocks."
 				// Therefore we need to send these images after the tool result messages
 				// NOTE: it's actually okay to have multiple user messages in a row, the model will treat them as a continuation of the same input (this way works better than combining them into one message, since the tool result specifically mentions (see following user message for image)
-				if (toolResultImages.length > 0) {
-					openAiMessages.push({
-						role: "user",
-						content: toolResultImages.map((part) => ({
-							type: "image_url",
-							image_url: { url: `data:${part.source.media_type};base64,${part.source.data}` },
-						})),
-					})
-				}
+				// UPDATE v2.0: we don't use tools anymore, but if we did it's important to note that the openrouter prompt caching mechanism requires one user message at a time, so we would need to add these images to the user content array instead.
+				// if (toolResultImages.length > 0) {
+				// 	openAiMessages.push({
+				// 		role: "user",
+				// 		content: toolResultImages.map((part) => ({
+				// 			type: "image_url",
+				// 			image_url: { url: `data:${part.source.media_type};base64,${part.source.data}` },
+				// 		})),
+				// 	})
+				// }
 
 				// Process non-tool messages
 				if (nonToolMessages.length > 0) {