[Fix] GLM models reserve too much output context by default (#160)

* fix: cap default glm output reservation

* test: align zai e2e max_tokens expectations

---------

Co-authored-by: Roomote <roomote@roocode.com>

Author: roomote[bot]

apps/vscode-e2e/src/suite/providers/zai.test.ts

@@ -206,12 +206,12 @@ suite("Z.ai GLM provider", function () {
 
 		assert.ok(completionMessage, "Task should complete with the expected Z.ai GLM response")
 
-		// Verify max_tokens is the model's documented limit (131_072), not the 20%-of-context
-		// heuristic cap (40_000) that guards against inaccurate OpenRouter dynamic metadata.
+		// Verify max_tokens uses the restored default clamp (20% of context window)
+		// unless the user explicitly overrides it via modelMaxTokens.
 		assert.strictEqual(
 			requestCapture.maxTokens,
-			131_072,
-			`max_tokens should be the documented glm-5.1 limit (131_072) but was ${requestCapture.maxTokens}`,
+			40_000,
+			`max_tokens should default to the glm-5.1 clamp (40_000) but was ${requestCapture.maxTokens}`,
 		)
 	})
 
@@ -245,12 +245,12 @@ suite("Z.ai GLM provider", function () {
 
 		assert.ok(completionMessage, "Task should complete with the expected Z.ai GLM-5-Turbo response")
 
-		// Verify max_tokens is the model's documented limit (131_072), not the 20%-of-context
-		// heuristic cap (40_000) that guards against inaccurate OpenRouter dynamic metadata.
+		// Verify max_tokens uses the restored default clamp (20% of context window)
+		// unless the user explicitly overrides it via modelMaxTokens.
 		assert.strictEqual(
 			requestCapture.maxTokens,
-			131_072,
-			`max_tokens should be the documented glm-5-turbo limit (131_072) but was ${requestCapture.maxTokens}`,
+			40_551,
+			`max_tokens should default to the glm-5-turbo clamp (40_551) but was ${requestCapture.maxTokens}`,
 		)
 	})
 })

src/api/providers/__tests__/zai.spec.ts

@@ -475,6 +475,34 @@ describe("ZAiHandler", () => {
 	})
 
 	describe("GLM-4.7 Thinking Mode", () => {
+		it("should cap GLM-5.1 max_tokens to 20% of context window by default", async () => {
+			const handlerWithModel = new ZAiHandler({
+				apiModelId: "glm-5.1",
+				zaiApiKey: "test-zai-api-key",
+				zaiApiLine: "international_coding",
+			})
+
+			mockCreate.mockImplementationOnce(() => {
+				return {
+					[Symbol.asyncIterator]: () => ({
+						async next() {
+							return { done: true }
+						},
+					}),
+				}
+			})
+
+			const messageGenerator = handlerWithModel.createMessage("system prompt", [])
+			await messageGenerator.next()
+
+			expect(mockCreate).toHaveBeenCalledWith(
+				expect.objectContaining({
+					model: "glm-5.1",
+					max_tokens: 40_000,
+				}),
+			)
+		})
+
 		it("should enable thinking by default for GLM-4.7 (default reasoningEffort is medium)", async () => {
 			const handlerWithModel = new ZAiHandler({
 				apiModelId: "glm-4.7",

src/api/providers/zai.ts

@@ -11,7 +11,7 @@ import {
 	zaiApiLineConfigs,
 } from "@roo-code/types"
 
-import { type ApiHandlerOptions, shouldUseReasoningEffort } from "../../shared/api"
+import { type ApiHandlerOptions, getModelMaxOutputTokens, shouldUseReasoningEffort } from "../../shared/api"
 import { convertToZAiFormat } from "../transform/zai-format"
 
 import type { ApiHandlerCreateMessageMetadata } from "../index"
@@ -79,11 +79,15 @@ export class ZAiHandler extends BaseOpenAiCompatibleProvider<string> {
 	) {
 		const { id: model, info } = this.getModel()
 
-		// Use info.maxTokens directly — Z.ai model definitions are hand-curated and accurate.
-		// getModelMaxOutputTokens clamps to 20% of contextWindow (a guard for OpenRouter dynamic
-		// metadata where maxTokens ≈ contextWindow), but ApiHandlerOptions omits apiProvider so
-		// the zai bypass in api.ts never fires. glm-5.1 legitimately supports 128k output.
-		const max_tokens = this.options.modelMaxTokens || (info.maxTokens ?? undefined)
+		const max_tokens =
+			this.options.modelMaxTokens ||
+			(getModelMaxOutputTokens({
+				modelId: model,
+				model: info,
+				settings: this.options,
+				format: "openai",
+			}) ??
+				undefined)
 
 		const temperature = this.options.modelTemperature ?? this.defaultTemperature
 

src/shared/__tests__/api.spec.ts

@@ -274,8 +274,7 @@ describe("getModelMaxOutputTokens", () => {
 		})
 	})
 
-	test("should bypass 20% cap for Z.ai provider and use exact configured max tokens", () => {
-		// glm-5.1: maxTokens=131_072 on a 200k context window (65.5%) — would be capped to 40k without bypass
+	test("should still clamp Z.ai models to 20% of context window by default", () => {
 		const model: ModelInfo = {
 			contextWindow: 200_000,
 			supportsPromptCache: true,
@@ -290,7 +289,7 @@ describe("getModelMaxOutputTokens", () => {
 			format: "openai",
 		})
 
-		expect(result).toBe(131_072)
+		expect(result).toBe(40_000)
 	})
 
 	test("should still clamp non-Z.ai models with high maxTokens to 20% of context window", () => {

src/shared/api.ts

@@ -138,15 +138,8 @@ export const getModelMaxOutputTokens = ({
 		// Check if this is a GPT-5 model (case-insensitive)
 		const isGpt5Model = modelId.toLowerCase().includes("gpt-5")
 
-		// Z.ai models have hand-curated maxTokens values from Z.ai's own documentation
-		// (e.g. glm-5.1 supports 128k output on a 200k context window) — bypass the cap.
-		// Note: ZAiHandler.createStreamWithThinking bypasses this function entirely because
-		// ApiHandlerOptions omits apiProvider; this bypass serves callers that pass full
-		// ProviderSettings (Task.ts context management, model-params.ts).
-		const isZaiProvider = settings?.apiProvider === "zai"
-
 		// GPT-5 models bypass the 20% cap and use their full configured max tokens
-		if (isGpt5Model || isZaiProvider) {
+		if (isGpt5Model) {
 			return model.maxTokens
 		}