From bc8eb74b82c9ce700769ae6181bfaa6d7e0bbd90 Mon Sep 17 00:00:00 2001
From: Mario Zechner <badlogicgames@gmail.com>
Date: Fri, 27 Mar 2026 02:50:03 +0100
Subject: [PATCH] fix(ai): detect Ollama overflow errors closes #2626

---
 packages/ai/CHANGELOG.md           |  4 +++
 packages/ai/src/utils/overflow.ts  |  8 +++---
 packages/ai/test/overflow.test.ts  | 42 ++++++++++++++++++++++++++++++
 packages/coding-agent/CHANGELOG.md |  4 +++
 4 files changed, 55 insertions(+), 3 deletions(-)
 create mode 100644 packages/ai/test/overflow.test.ts

diff --git a/packages/ai/CHANGELOG.md b/packages/ai/CHANGELOG.md
index 2e5548bc..8903d63a 100644
--- a/packages/ai/CHANGELOG.md
+++ b/packages/ai/CHANGELOG.md
@@ -2,6 +2,10 @@
 
 ## [Unreleased]
 
+### Fixed
+
+- Fixed context overflow detection to recognize Ollama error responses like `prompt too long; exceeded max context length ...`, so callers can trigger compaction and retry instead of surfacing the raw overflow error ([#2626](https://github.com/badlogic/pi-mono/issues/2626))
+
 ## [0.63.0] - 2026-03-27
 
 ### Breaking Changes
diff --git a/packages/ai/src/utils/overflow.ts b/packages/ai/src/utils/overflow.ts
index c6f4696e..968d9da8 100644
--- a/packages/ai/src/utils/overflow.ts
+++ b/packages/ai/src/utils/overflow.ts
@@ -22,7 +22,7 @@ import type { AssistantMessage } from "../types.js";
  * - Cerebras: Returns "400/413 status code (no body)" - handled separately below
  * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
  * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow
- * - Ollama: Silently truncates input - not detectable via error message
+ * - Ollama: Some deployments truncate silently, others return errors like "prompt too long; exceeded max context length by X tokens"
  */
 const OVERFLOW_PATTERNS = [
 	/prompt is too long/i, // Anthropic
@@ -39,6 +39,7 @@ const OVERFLOW_PATTERNS = [
 	/exceeded model token limit/i, // Kimi For Coding
 	/too large for model with \d+ maximum context length/i, // Mistral
 	/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text
+	/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error
 	/context[_ ]length[_ ]exceeded/i, // Generic fallback
 	/too many tokens/i, // Generic fallback
 	/token limit exceeded/i, // Generic fallback
@@ -71,8 +72,9 @@ const OVERFLOW_PATTERNS = [
  * **Unreliable detection:**
  * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),
  *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.
- * - Ollama: Silently truncates input without error. Cannot be detected via this function.
- *   The response will have usage.input < expected, but we don't know the expected value.
+ * - Ollama: May truncate input silently for some setups, but may also return explicit
+ *   overflow errors that match the patterns above. Silent truncation still cannot be
+ *   detected here because we do not know the expected token count.
  *
  * ## Custom Providers
  *
diff --git a/packages/ai/test/overflow.test.ts b/packages/ai/test/overflow.test.ts
new file mode 100644
index 00000000..214cb123
--- /dev/null
+++ b/packages/ai/test/overflow.test.ts
@@ -0,0 +1,42 @@
+import { describe, expect, it } from "vitest";
+import type { AssistantMessage } from "../src/types.js";
+import { isContextOverflow } from "../src/utils/overflow.js";
+
+function createErrorMessage(errorMessage: string): AssistantMessage {
+	return {
+		role: "assistant",
+		content: [],
+		api: "openai-completions",
+		provider: "ollama",
+		model: "qwen3.5:35b",
+		usage: {
+			input: 0,
+			output: 0,
+			cacheRead: 0,
+			cacheWrite: 0,
+			totalTokens: 0,
+			cost: {
+				input: 0,
+				output: 0,
+				cacheRead: 0,
+				cacheWrite: 0,
+				total: 0,
+			},
+		},
+		stopReason: "error",
+		errorMessage,
+		timestamp: Date.now(),
+	};
+}
+
+describe("isContextOverflow", () => {
+	it("detects explicit Ollama prompt-too-long errors", () => {
+		const message = createErrorMessage("400 `prompt too long; exceeded max context length by 100918 tokens`");
+		expect(isContextOverflow(message, 32768)).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);
+	});
+});
diff --git a/packages/coding-agent/CHANGELOG.md b/packages/coding-agent/CHANGELOG.md
index 8ed999be..73c1e732 100644
--- a/packages/coding-agent/CHANGELOG.md
+++ b/packages/coding-agent/CHANGELOG.md
@@ -2,6 +2,10 @@
 
 ## [Unreleased]
 
+### Fixed
+
+- Fixed auto-compaction overflow recovery for Ollama models when the backend returns explicit `prompt too long; exceeded max context length ...` errors instead of silently truncating input ([#2626](https://github.com/badlogic/pi-mono/issues/2626))
+
 ## [0.63.0] - 2026-03-27
 
 ### Breaking Changes