diff --git a/packages/coding-agent/src/config/model-resolver.ts b/packages/coding-agent/src/config/model-resolver.ts
index f52c70607..e9b84120f 100644
--- a/packages/coding-agent/src/config/model-resolver.ts
+++ b/packages/coding-agent/src/config/model-resolver.ts
@@ -200,15 +200,19 @@ const AMAZON_BEDROCK_PROVIDER = "amazon-bedrock";
 const BEDROCK_INFERENCE_PROFILE_ARN =
 	/^arn:aws(?:-[a-z]+)*:bedrock:[a-z0-9-]+:[0-9]*:(?:application-inference-profile|inference-profile)\/[a-z0-9][a-z0-9._:-]*$/i;
 
-function resolveBedrockInferenceProfileReference(
-	provider: string,
+function hasBedrockInferenceProfileThinkingSuffix(modelId: string): boolean {
+	const { base, level } = splitThinkingSuffix(modelId);
+	return level !== undefined && BEDROCK_INFERENCE_PROFILE_ARN.test(base.trim());
+}
+
+function resolveBedrockInferenceProfileModelId(
 	modelId: string,
 	availableModels: readonly Model<Api>[],
 ): Model<Api> | undefined {
-	if (provider.toLowerCase() !== AMAZON_BEDROCK_PROVIDER) return undefined;
-
 	const requestedId = modelId.trim();
-	if (!BEDROCK_INFERENCE_PROFILE_ARN.test(requestedId)) return undefined;
+	if (hasBedrockInferenceProfileThinkingSuffix(requestedId) || !BEDROCK_INFERENCE_PROFILE_ARN.test(requestedId)) {
+		return undefined;
+	}
 
 	const providerModels = availableModels.filter(model => model.provider.toLowerCase() === AMAZON_BEDROCK_PROVIDER);
 	const defaultModelId = DEFAULT_MODEL_PER_PROVIDER[AMAZON_BEDROCK_PROVIDER];
@@ -223,6 +227,15 @@ function resolveBedrockInferenceProfileReference(
 	};
 }
 
+function resolveBedrockInferenceProfileReference(
+	provider: string,
+	modelId: string,
+	availableModels: readonly Model<Api>[],
+): Model<Api> | undefined {
+	if (provider.toLowerCase() !== AMAZON_BEDROCK_PROVIDER) return undefined;
+	return resolveBedrockInferenceProfileModelId(modelId, availableModels);
+}
+
 const UPSTREAM_ROUTING_SLUG = /^[a-z0-9](?:[a-z0-9-]*[a-z0-9])?$/i;
 
 /**
@@ -502,6 +515,7 @@ function findExactCanonicalModelMatch(
  * The single model-matching engine. Tries, in order:
  * 1. exact `provider/id` reference (variant-alias and OpenRouter routed/date
  *    fallbacks included),
+
  * 2. exact canonical id (coalesces provider variants),
  * 3. exact bare id (preference-ranked),
  * 4. retired effort-tier variant alias (collapsed catalog entries),
@@ -534,6 +548,11 @@ function matchModel(
 		return pickPreferredModel(exactMatches, context);
 	}
 
+	const bedrockInferenceProfile = resolveBedrockInferenceProfileModelId(modelPattern, availableModels);
+	if (bedrockInferenceProfile) {
+		return bedrockInferenceProfile;
+	}
+
 	// Retired effort-tier variant ids (bare, no provider prefix) resolve to
 	// their collapsed logical model; models from the providers whose table
 	// declared the alias win ties. Auto-derived `X-thinking` pairs resolve
diff --git a/packages/coding-agent/test/model-resolver.test.ts b/packages/coding-agent/test/model-resolver.test.ts
index 76426f488..73d2d26b5 100644
--- a/packages/coding-agent/test/model-resolver.test.ts
+++ b/packages/coding-agent/test/model-resolver.test.ts
@@ -883,7 +883,7 @@ describe("resolveCliModel", () => {
 		expect(result.model?.id).toBe("z-ai/glm-4.7-20251222:nitro");
 	});
 
-	test("accepts Bedrock inference profile ARNs as explicit provider model ids", () => {
+	test("accepts Bedrock inference profile ARNs and preserves thinking suffixes", () => {
 		const defaultBedrockModel = buildModel({
 			id: "us.anthropic.claude-opus-4-8",
 			name: "Claude Opus 4.8 (US)",
@@ -898,19 +898,30 @@ describe("resolveCliModel", () => {
 		});
 		const profileArn = "arn:aws:bedrock:us-east-2:1234567890:application-inference-profile/company-opus-48";
 
-		const result = resolveCliModel({
+		const baseResult = resolveCliModel({
 			cliProvider: "amazon-bedrock",
 			cliModel: profileArn,
 			modelRegistry: {
 				getAll: () => [defaultBedrockModel],
 			},
 		});
+		const offResult = resolveCliModel({
+			cliProvider: "amazon-bedrock",
+			cliModel: `${profileArn}:off`,
+			modelRegistry: {
+				getAll: () => [defaultBedrockModel],
+			},
+		});
 
-		expect(result.error).toBeUndefined();
-		expect(result.model?.provider).toBe("amazon-bedrock");
-		expect(result.model?.api).toBe("bedrock-converse-stream");
-		expect(result.model?.id).toBe(profileArn);
-		expect(result.model?.name).toBe("Bedrock inference profile");
+		expect(baseResult.error).toBeUndefined();
+		expect(baseResult.model?.provider).toBe("amazon-bedrock");
+		expect(baseResult.model?.api).toBe("bedrock-converse-stream");
+		expect(baseResult.model?.id).toBe(profileArn);
+		expect(baseResult.model?.name).toBe("Bedrock inference profile");
+		expect(baseResult.thinkingLevel).toBeUndefined();
+		expect(offResult.error).toBeUndefined();
+		expect(offResult.model?.id).toBe(profileArn);
+		expect(offResult.thinkingLevel).toBe("off");
 	});
 
 	test("returns a clear error when there are no models", () => {