fix(coding-agent): parse :thinking suffix from custom model IDs in fallback path

Fixes #5552

packages/coding-agent/src/core/model-resolver.ts

@@ -340,9 +340,10 @@ export interface ResolveCliModelResult {
 export function resolveCliModel(options: {
 	cliProvider?: string;
 	cliModel?: string;
+	cliThinking?: string;
 	modelRegistry: ModelRegistry;
 }): ResolveCliModelResult {
-	const { cliProvider, cliModel, modelRegistry } = options;
+	const { cliProvider, cliModel, cliThinking, modelRegistry } = options;
 
 	if (!cliModel) {
 		return { model: undefined, warning: undefined, error: undefined };
@@ -451,12 +452,28 @@ export function resolveCliModel(options: {
 	}
 
 	if (provider) {
-		const fallbackModel = buildFallbackModel(provider, pattern, availableModels);
+		// Parse thinking level suffix from the pattern before building the fallback model,
+		// but only when --thinking is not explicitly provided.
+		// e.g. "zai-org/GLM-5.1-FP8:high" → modelId="zai-org/GLM-5.1-FP8", fallbackThinking="high"
+		let fallbackPattern = pattern;
+		let fallbackThinking: ThinkingLevel | undefined;
+		if (!cliThinking) {
+			const lastColon = pattern.lastIndexOf(":");
+			if (lastColon !== -1) {
+				const suffix = pattern.substring(lastColon + 1);
+				if (isValidThinkingLevel(suffix)) {
+					fallbackPattern = pattern.substring(0, lastColon);
+					fallbackThinking = suffix;
+				}
+			}
+		}
+
+		const fallbackModel = buildFallbackModel(provider, fallbackPattern, availableModels);
 		if (fallbackModel) {
 			const fallbackWarning = warning
-				? `${warning} Model "${pattern}" not found for provider "${provider}". Using custom model id.`
-				: `Model "${pattern}" not found for provider "${provider}". Using custom model id.`;
-			return { model: fallbackModel, thinkingLevel: undefined, warning: fallbackWarning, error: undefined };
+				? `${warning} Model "${fallbackPattern}" not found for provider "${provider}". Using custom model id.`
+				: `Model "${fallbackPattern}" not found for provider "${provider}". Using custom model id.`;
+			return { model: fallbackModel, thinkingLevel: fallbackThinking, warning: fallbackWarning, error: undefined };
 		}
 	}
 

packages/coding-agent/src/main.ts

@@ -360,6 +360,7 @@ function buildSessionOptions(
 		const resolved = resolveCliModel({
 			cliProvider: parsed.provider,
 			cliModel: parsed.model,
+			cliThinking: parsed.thinking,
 			modelRegistry,
 		});
 		if (resolved.warning) {

packages/coding-agent/test/model-resolver.test.ts

@@ -370,6 +370,127 @@ describe("resolveCliModel", () => {
 		expect(result.model?.provider).toBe("openrouter");
 		expect(result.model?.id).toBe("qwen/qwen3-coder:exacto");
 	});
+
+	describe("custom model fallback with :thinking suffix (#5552)", () => {
+		// Models for a provider that has registered models but the specific model ID
+		// is not in the registry (triggers buildFallbackModel path).
+		const neuralwattModel: Model<"anthropic-messages"> = {
+			id: "some-base-model",
+			name: "Some Base Model",
+			api: "anthropic-messages",
+			provider: "neuralwatt",
+			baseUrl: "https://api.neuralwatt.com",
+			reasoning: false,
+			input: ["text"],
+			cost: { input: 1, output: 2, cacheRead: 0.1, cacheWrite: 1 },
+			contextWindow: 128000,
+			maxTokens: 8192,
+		};
+
+		const modelsWithNeuralwatt = [...allModels, neuralwattModel];
+
+		test("strips :thinking suffix from custom model id in fallback path", () => {
+			const registry = {
+				getAll: () => modelsWithNeuralwatt,
+			} as unknown as Parameters<typeof resolveCliModel>[0]["modelRegistry"];
+
+			const result = resolveCliModel({
+				cliModel: "neuralwatt/zai-org/GLM-5.1-FP8:high",
+				modelRegistry: registry,
+			});
+
+			expect(result.error).toBeUndefined();
+			expect(result.model?.provider).toBe("neuralwatt");
+			// The :high suffix must NOT leak into the model id sent to the API
+			expect(result.model?.id).toBe("zai-org/GLM-5.1-FP8");
+			expect(result.thinkingLevel).toBe("high");
+		});
+
+		test("custom model without thinking suffix works normally in fallback path", () => {
+			const registry = {
+				getAll: () => modelsWithNeuralwatt,
+			} as unknown as Parameters<typeof resolveCliModel>[0]["modelRegistry"];
+
+			const result = resolveCliModel({
+				cliModel: "neuralwatt/zai-org/GLM-5.1-FP8",
+				modelRegistry: registry,
+			});
+
+			expect(result.error).toBeUndefined();
+			expect(result.model?.provider).toBe("neuralwatt");
+			expect(result.model?.id).toBe("zai-org/GLM-5.1-FP8");
+			expect(result.thinkingLevel).toBeUndefined();
+		});
+
+		test("all valid thinking levels work in fallback path", () => {
+			const registry = {
+				getAll: () => modelsWithNeuralwatt,
+			} as unknown as Parameters<typeof resolveCliModel>[0]["modelRegistry"];
+
+			for (const level of ["off", "minimal", "low", "medium", "high", "xhigh"]) {
+				const result = resolveCliModel({
+					cliModel: `neuralwatt/zai-org/GLM-5.1-FP8:${level}`,
+					modelRegistry: registry,
+				});
+
+				expect(result.error).toBeUndefined();
+				expect(result.model?.id).toBe("zai-org/GLM-5.1-FP8");
+				expect(result.thinkingLevel).toBe(level);
+			}
+		});
+
+		test("invalid thinking suffix on custom model is treated as part of model id", () => {
+			const registry = {
+				getAll: () => modelsWithNeuralwatt,
+			} as unknown as Parameters<typeof resolveCliModel>[0]["modelRegistry"];
+
+			const result = resolveCliModel({
+				cliModel: "neuralwatt/zai-org/GLM-5.1-FP8:banana",
+				modelRegistry: registry,
+			});
+
+			expect(result.error).toBeUndefined();
+			expect(result.model?.provider).toBe("neuralwatt");
+			// Invalid suffix stays in the id (it's not a thinking level)
+			expect(result.model?.id).toBe("zai-org/GLM-5.1-FP8:banana");
+			expect(result.thinkingLevel).toBeUndefined();
+		});
+
+		test("explicit --provider with custom model:thinking strips suffix correctly", () => {
+			const registry = {
+				getAll: () => modelsWithNeuralwatt,
+			} as unknown as Parameters<typeof resolveCliModel>[0]["modelRegistry"];
+
+			const result = resolveCliModel({
+				cliProvider: "neuralwatt",
+				cliModel: "zai-org/GLM-5.1-FP8:high",
+				modelRegistry: registry,
+			});
+
+			expect(result.error).toBeUndefined();
+			expect(result.model?.provider).toBe("neuralwatt");
+			expect(result.model?.id).toBe("zai-org/GLM-5.1-FP8");
+			expect(result.thinkingLevel).toBe("high");
+		});
+
+		test("with explicit --thinking, :suffix is kept as part of model id", () => {
+			const registry = {
+				getAll: () => modelsWithNeuralwatt,
+			} as unknown as Parameters<typeof resolveCliModel>[0]["modelRegistry"];
+
+			const result = resolveCliModel({
+				cliModel: "neuralwatt/zai-org/GLM-5.1-FP8:high",
+				cliThinking: "medium",
+				modelRegistry: registry,
+			});
+
+			expect(result.error).toBeUndefined();
+			expect(result.model?.provider).toBe("neuralwatt");
+			// :high is kept as part of the model id since --thinking was explicit
+			expect(result.model?.id).toBe("zai-org/GLM-5.1-FP8:high");
+			expect(result.thinkingLevel).toBeUndefined();
+		});
+	});
 });
 
 describe("default model selection", () => {