From 7c5c3d6fd60a0a3d8a3d40d9320c32975b01b215 Mon Sep 17 00:00:00 2001
From: Mario Zechner <badlogicgames@gmail.com>
Date: Sat, 16 May 2026 23:29:22 +0200
Subject: [PATCH] fix(ai): detect litellm context overflow errors

closes #4563
---
 packages/ai/CHANGELOG.md          | 1 +
 packages/ai/src/utils/overflow.ts | 4 +++-
 packages/ai/test/overflow.test.ts | 7 +++++++
 3 files changed, 11 insertions(+), 1 deletion(-)

diff --git a/packages/ai/CHANGELOG.md b/packages/ai/CHANGELOG.md
index 7fa3a37d..f962fcee 100644
--- a/packages/ai/CHANGELOG.md
+++ b/packages/ai/CHANGELOG.md
@@ -15,6 +15,7 @@
 - Fixed OpenAI Responses requests for models that support disabling reasoning to send `reasoning.effort: "none"` when thinking is off.
 - Fixed Inception Mercury 2 tool calling on OpenRouter by marking `off` as unsupported in `thinkingLevelMap`, so the openai-completions provider omits the reasoning param instead of defaulting to `{reasoning:{effort:"none"}}` (which puts Mercury 2 in instant mode, disabling tool calls).
 - Fixed OpenAI Codex SSE retries to honor `retry-after-ms` and `retry-after` headers before falling back to exponential backoff.
+- Fixed context overflow detection for LiteLLM-wrapped OpenAI-compatible errors using `exceeds the model's maximum context length of ... tokens` wording ([#4563](https://github.com/earendil-works/pi/issues/4563)).
 
 ## [0.74.0] - 2026-05-07
 
diff --git a/packages/ai/src/utils/overflow.ts b/packages/ai/src/utils/overflow.ts
index bf879dc4..7eef4a75 100644
--- a/packages/ai/src/utils/overflow.ts
+++ b/packages/ai/src/utils/overflow.ts
@@ -11,6 +11,7 @@ import type { AssistantMessage } from "../types.js";
  * - Anthropic: "prompt is too long: 213462 tokens > 200000 maximum"
  * - Anthropic: "413 {\"error\":{\"type\":\"request_too_large\",\"message\":\"Request exceeds the maximum size\"}}"
  * - OpenAI: "Your input exceeds the context window of this model"
+ * - OpenAI/LiteLLM: "Requested token count exceeds the model's maximum context length of 131072 tokens"
  * - Google: "The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)"
  * - xAI: "This model's maximum prompt length is 131072 but the request contains 537812 tokens"
  * - Groq: "Please reduce the length of the messages or completion"
@@ -34,6 +35,7 @@ const OVERFLOW_PATTERNS = [
 	/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)
 	/input is too long for requested model/i, // Amazon Bedrock
 	/exceeds the context window/i, // OpenAI (Completions & Responses API)
+	/exceeds (?:the )?(?:model'?s )?maximum context length of [\d,]+ tokens?/i, // OpenAI-compatible proxies (LiteLLM)
 	/input token count.*exceeds the maximum/i, // Google (Gemini)
 	/maximum prompt length is \d+/i, // xAI (Grok)
 	/reduce the length of the messages/i, // Groq
@@ -81,7 +83,7 @@ const NON_OVERFLOW_PATTERNS = [
  *
  * **Reliable detection (returns error with detectable message):**
  * - Anthropic: "prompt is too long: X tokens > Y maximum" or "request_too_large"
- * - OpenAI (Completions & Responses): "exceeds the context window"
+ * - OpenAI (Completions & Responses): "exceeds the context window" or "exceeds the model's maximum context length of X tokens"
  * - Google Gemini: "input token count exceeds the maximum"
  * - xAI (Grok): "maximum prompt length is X but request contains Y"
  * - Groq: "reduce the length of the messages"
diff --git a/packages/ai/test/overflow.test.ts b/packages/ai/test/overflow.test.ts
index 1b50b831..ca6fde24 100644
--- a/packages/ai/test/overflow.test.ts
+++ b/packages/ai/test/overflow.test.ts
@@ -42,6 +42,13 @@ describe("isContextOverflow", () => {
 		expect(isContextOverflow(message, 262144)).toBe(true);
 	});
 
+	it("detects LiteLLM-wrapped OpenAI maximum context length errors", () => {
+		const message = createErrorMessage(
+			"Error: 503 litellm.ServiceUnavailableError: litellm.MidStreamFallbackError: litellm.APIConnectionError: APIConnectionError: OpenAIException - Requested token count exceeds the model's maximum context length of 131072 tokens.",
+		);
+		expect(isContextOverflow(message, 131072)).toBe(true);
+	});
+
 	it("does not treat generic non-overflow Ollama errors as overflow", () => {
 		const message = createErrorMessage("500 `model runner crashed unexpectedly`");
 		expect(isContextOverflow(message, 32768)).toBe(false);