From 1b6ddca87ca041e3b02b387d5a321eb77fc39eca Mon Sep 17 00:00:00 2001
From: David Brailovsky <davidbrai@gmail.com>
Date: Mon, 21 Sep 2026 21:48:11 +0200
Subject: [PATCH] fix(ai): omit empty text parts from multimodal user messages

Prevent OpenAI-compatible providers from rejecting image-only prompts.

Fixes #9797
---
 packages/ai/CHANGELOG.md                      |  4 +++
 packages/ai/src/api/openai-completions.ts     | 32 ++++++++++---------
 ...nai-completions-tool-result-images.test.ts | 29 +++++++++++++++++
 3 files changed, 50 insertions(+), 15 deletions(-)

diff --git a/packages/ai/CHANGELOG.md b/packages/ai/CHANGELOG.md
index 0f85f4d26..94240ff85 100644
--- a/packages/ai/CHANGELOG.md
+++ b/packages/ai/CHANGELOG.md
@@ -2,6 +2,10 @@
 
 ## [Unreleased]
 
+### Fixed
+
+- Fixed image-only user messages being rejected by some OpenAI-compatible providers because they included an empty text part ([#9797](https://github.com/earendil-works/pi/issues/9797))
+
 ## [0.87.0] - 2026-09-21
 
 ### Added
diff --git a/packages/ai/src/api/openai-completions.ts b/packages/ai/src/api/openai-completions.ts
index 4ac0b6206..f5a1bc668 100644
--- a/packages/ai/src/api/openai-completions.ts
+++ b/packages/ai/src/api/openai-completions.ts
@@ -1257,21 +1257,23 @@ export function convertMessages(
 					content: sanitizeSurrogates(msg.content),
 				});
 			} else {
-				const content: ChatCompletionContentPart[] = msg.content.map((item): ChatCompletionContentPart => {
-					if (item.type === "text") {
-						return {
-							type: "text",
-							text: sanitizeSurrogates(item.text),
-						} satisfies ChatCompletionContentPartText;
-					} else {
-						return {
-							type: "image_url",
-							image_url: {
-								url: `data:${item.mimeType};base64,${item.data}`,
-							},
-						} satisfies ChatCompletionContentPartImage;
-					}
-				});
+				const content: ChatCompletionContentPart[] = msg.content
+					.filter((item) => item.type !== "text" || item.text.length > 0)
+					.map((item): ChatCompletionContentPart => {
+						if (item.type === "text") {
+							return {
+								type: "text",
+								text: sanitizeSurrogates(item.text),
+							} satisfies ChatCompletionContentPartText;
+						} else {
+							return {
+								type: "image_url",
+								image_url: {
+									url: `data:${item.mimeType};base64,${item.data}`,
+								},
+							} satisfies ChatCompletionContentPartImage;
+						}
+					});
 				if (content.length === 0) continue;
 				params.push({
 					role: "user",
diff --git a/packages/ai/test/openai-completions-tool-result-images.test.ts b/packages/ai/test/openai-completions-tool-result-images.test.ts
index 6e48efe67..4d303865e 100644
--- a/packages/ai/test/openai-completions-tool-result-images.test.ts
+++ b/packages/ai/test/openai-completions-tool-result-images.test.ts
@@ -69,6 +69,35 @@ function buildEmptyToolResult(toolCallId: string, timestamp: number): ToolResult
 }
 
 describe("openai-completions convertMessages", () => {
+	// Regression test for https://github.com/earendil-works/pi/issues/9797
+	it("omits empty text parts from user messages with images", () => {
+		const { compat: _compat, ...baseModel } = getModel("openai", "gpt-4o-mini");
+		const model: Model<"openai-completions"> = {
+			...baseModel,
+			api: "openai-completions",
+			input: ["text", "image"],
+		};
+		const context = normalizeContext({
+			messages: [
+				{
+					role: "user",
+					content: [
+						{ type: "text", text: "" },
+						{ type: "image", data: "ZmFrZQ==", mimeType: "image/png" },
+					],
+					timestamp: Date.now(),
+				},
+			],
+		});
+
+		expect(convertMessages(model, context, compat)).toEqual([
+			{
+				role: "user",
+				content: [{ type: "image_url", image_url: { url: "data:image/png;base64,ZmFrZQ==" } }],
+			},
+		]);
+	});
+
 	it("batches tool-result images after consecutive tool results", () => {
 		const { compat: _compat, ...baseModel } = getModel("openai", "gpt-4o-mini");
 		const model: Model<"openai-completions"> = {