From bb0f4aa602f332838015ce66bdb0a94d8206dc93 Mon Sep 17 00:00:00 2001
From: Vegard Stikbakke <vegard.stikbakke@gmail.com>
Date: Fri, 18 Sep 2026 10:43:20 +0200
Subject: [PATCH] fix(ai): preserve DeepSeek V4 effort metadata

closes #9485
---
 packages/ai/CHANGELOG.md                |  1 +
 packages/ai/scripts/generate-models.ts  | 16 +++++++++++-----
 packages/ai/test/supports-xhigh.test.ts | 12 ++++++++++++
 3 files changed, 24 insertions(+), 5 deletions(-)

diff --git a/packages/ai/CHANGELOG.md b/packages/ai/CHANGELOG.md
index a3b63e880..03bb51b87 100644
--- a/packages/ai/CHANGELOG.md
+++ b/packages/ai/CHANGELOG.md
@@ -8,6 +8,7 @@
 
 ### Fixed
 
+- Fixed DeepSeek V4.1 thinking levels on OpenRouter and OpenCode Go preserving provider effort metadata ([#9485](https://github.com/earendil-works/pi/issues/9485)).
 - Fixed Vercel AI Gateway replaying unsigned thinking as assistant text ([#9676](https://github.com/earendil-works/pi/issues/9676)).
 - Fixed Google Generative AI and Vertex AI using unsupported thinking levels when reasoning is omitted or when model capabilities differ within a Gemini family ([#9455](https://github.com/earendil-works/pi/issues/9455)).
 - Fixed Anthropic-compatible relays breaking signed thinking replay when they report a different response model, while preserving fallback pricing ([#9188](https://github.com/earendil-works/pi/issues/9188)).
diff --git a/packages/ai/scripts/generate-models.ts b/packages/ai/scripts/generate-models.ts
index b1dfcf10f..13dc43252 100644
--- a/packages/ai/scripts/generate-models.ts
+++ b/packages/ai/scripts/generate-models.ts
@@ -1024,7 +1024,11 @@ function applyThinkingLevelMetadata(model: Model<any>): void {
 	if (model.api === "anthropic-messages" && isAnthropicTemperatureUnsupportedModel(model.id)) {
 		mergeAnthropicMessagesCompat(model, { supportsTemperature: false });
 	}
-	if (model.api === "openai-completions" && model.id.includes("deepseek-v4")) {
+	if (
+		model.api === "openai-completions" &&
+		model.id.includes("deepseek-v4") &&
+		model.thinkingLevelMap === undefined
+	) {
 		mergeThinkingLevelMap(
 			model,
 			model.provider === "openrouter"
@@ -2181,10 +2185,12 @@ async function loadModelsDevData(): Promise<Model<any>[]> {
 					}
 				}
 
-				const thinkingLevelMap =
-					api === "google-generative-ai"
-						? getGoogleThinkingLevelMap(modelId, m.reasoning_options ?? [])
-						: undefined;
+				let thinkingLevelMap: NonNullable<Model<Api>["thinkingLevelMap"]> | undefined;
+				if (api === "google-generative-ai") {
+					thinkingLevelMap = getGoogleThinkingLevelMap(modelId, m.reasoning_options ?? []);
+				} else if (variant.provider === "opencode-go" && modelId === "deepseek-v4.1-flash") {
+					thinkingLevelMap = getEffortThinkingLevelMap(m.reasoning_options ?? []);
+				}
 				models.push({
 					id: modelId,
 					name: m.name || modelId,
diff --git a/packages/ai/test/supports-xhigh.test.ts b/packages/ai/test/supports-xhigh.test.ts
index a1e43c40b..e0ae2629f 100644
--- a/packages/ai/test/supports-xhigh.test.ts
+++ b/packages/ai/test/supports-xhigh.test.ts
@@ -94,6 +94,18 @@ describe("getSupportedThinkingLevels", () => {
 		expect(getSupportedThinkingLevels(model!)).toEqual(["off", "low", "high", "max"]);
 	});
 
+	it("preserves low/high/max metadata for DeepSeek V4.1 Flash on OpenRouter", () => {
+		const model = getModel("openrouter", "deepseek/deepseek-v4.1-flash");
+		expect(model).toBeDefined();
+		expect(getSupportedThinkingLevels(model!)).toEqual(["off", "low", "high", "max"]);
+	});
+
+	it("preserves low/high/max metadata for DeepSeek V4.1 Flash on opencode-go", () => {
+		const model = getModel("opencode-go", "deepseek-v4.1-flash");
+		expect(model).toBeDefined();
+		expect(getSupportedThinkingLevels(model!)).toEqual(["low", "high", "max"]);
+	});
+
 	it("includes only high plus off for OpenCode Go Kimi K2.6", () => {
 		const model = getModel("opencode-go", "kimi-k2.6");
 		expect(model).toBeDefined();