fix(ai): detect z.ai context overflow

closes #1937

packages/ai/CHANGELOG.md

@@ -2,6 +2,10 @@
 
 ## [Unreleased]
 
+### Fixed
+
+- Fixed context overflow detection to recognize z.ai `model_context_window_exceeded` errors surfaced through OpenAI-compatible stop reason handling ([#1937](https://github.com/badlogic/pi-mono/issues/1937))
+
 ## [0.57.0] - 2026-03-07
 
 ### Added

packages/ai/src/utils/overflow.ts

@@ -38,6 +38,7 @@ const OVERFLOW_PATTERNS = [
 	/context window exceeds limit/i, // MiniMax
 	/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
 	/context[_ ]length[_ ]exceeded/i, // Generic fallback
 	/too many tokens/i, // Generic fallback
 	/token limit exceeded/i, // Generic fallback

packages/ai/test/context-overflow.test.ts

@@ -384,29 +384,32 @@ describe("Context overflow error handling", () => {
 
 	// =============================================================================
 	// z.ai
-	// Special case: Sometimes accepts overflow silently, sometimes rate limits
-	// Detection via usage.input > contextWindow when successful
+	// Special case: may return explicit overflow error text, may accept overflow silently,
+	// or may rate limit instead
 	// =============================================================================
 
 	describe.skipIf(!process.env.ZAI_API_KEY)("z.ai", () => {
-		it("glm-4.5-flash - should detect overflow via isContextOverflow (silent overflow or rate limit)", async () => {
+		it("glm-4.5-flash - should detect overflow via isContextOverflow when z.ai reports it", async () => {
 			const model = getModel("zai", "glm-4.5-flash");
 			const result = await testContextOverflow(model, process.env.ZAI_API_KEY!);
 			logResult(result);
 
 			// z.ai behavior is inconsistent:
+			// - Sometimes returns explicit overflow error text via non-standard finish_reason handling
 			// - Sometimes accepts overflow and returns successfully with usage.input > contextWindow
 			// - Sometimes returns rate limit error
-			// Either way, isContextOverflow should detect it (via usage check or we skip if rate limited)
-			if (result.stopReason === "stop") {
+			if (result.stopReason === "error") {
+				if (result.errorMessage?.match(/model_context_window_exceeded/i)) {
+					expect(isContextOverflow(result.response, model.contextWindow)).toBe(true);
+				} else {
+					console.log("  z.ai returned non-overflow error (possibly rate limited), skipping overflow detection");
+				}
+			} else if (result.stopReason === "stop") {
 				if (result.hasUsageData && result.usage.input > model.contextWindow) {
 					expect(isContextOverflow(result.response, model.contextWindow)).toBe(true);
 				} else {
 					console.log("  z.ai returned stop without overflow usage data, skipping overflow detection");
 				}
-			} else {
-				// Rate limited or other error - just log and pass
-				console.log("  z.ai returned error (possibly rate limited), skipping overflow detection");
 			}
 		}, 120000);
 	});