@sayknow-cli/ai 0.5.9 → 0.5.10

package/CHANGELOG.md

@@ -2,7 +2,11 @@
 
 ## [Unreleased]
 
-## [0.5.9] - 2026-09-10
+## [0.5.10] - 2026-09-10
+
+### Added
+
+- Enabled `reasoning_effort` for native xAI Grok chat-completions and pinned bundled `xai/grok-4.5` / `xai/grok-4.6` thinking ranges to the official docs (`high` for 4.5, `xhigh` for 4.6). Default xAI model is now `grok-4.6`.
 
 ## [0.5.7] - 2026-09-10
 

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@sayknow-cli/ai",
-	"version": "0.5.9",
+	"version": "0.5.10",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"homepage": "https://sayknow-cli.com",
 	"author": "jaybeyond",
@@ -43,7 +43,7 @@
 	"dependencies": {
 		"@anthropic-ai/sdk": "^0.94.0",
 		"@bufbuild/protobuf": "^2.12.0",
-		"@sayknow-cli/utils": "0.5.9",
+		"@sayknow-cli/utils": "0.5.10",
 		"openai": "^6.36.0",
 		"partial-json": "^0.1.7",
 		"zod": "4.4.3"

package/src/model-thinking.ts

@@ -475,6 +475,35 @@ function applyGeneratedModelPolicy(model: ApiModel<Api>): void {
 	if (model.provider !== "opencode-go" && model.id === "minimax-m3") {
 		model.contextWindow = 512_000;
 	}
+	// xAI Grok 4.5/4.6: official reasoning docs (docs.x.ai) expose
+	// low/medium/high, with xhigh only on grok-4.6+. Catalog snapshots and
+	// openai-compat historically clamped everything to high because
+	// supportsReasoningEffort was false for Grok; pin the documented ranges so
+	// profile suffixes and requireSupportedEffort stay honest across regen.
+	if (model.provider === "xai" && /^grok-4\.6(?:$|[-.])/.test(model.id)) {
+		if (model.thinking) {
+			model.thinking = {
+				...model.thinking,
+				mode: "effort",
+				minLevel: Effort.Low,
+				maxLevel: Effort.XHigh,
+				defaultLevel: Effort.High,
+			};
+			delete model.thinking.levels;
+		}
+	}
+	if (model.provider === "xai" && /^grok-4\.5(?:$|[-.])/.test(model.id)) {
+		if (model.thinking) {
+			model.thinking = {
+				...model.thinking,
+				mode: "effort",
+				minLevel: Effort.Low,
+				maxLevel: Effort.High,
+				defaultLevel: Effort.High,
+			};
+			delete model.thinking.levels;
+		}
+	}
 }
 
 function scrubGeneratedModelName(name: string): string {

package/src/models.json

@@ -90830,8 +90830,9 @@
 			"maxTokens": 500000,
 			"thinking": {
 				"mode": "effort",
-				"minLevel": "minimal",
-				"maxLevel": "high"
+				"minLevel": "low",
+				"maxLevel": "high",
+				"defaultLevel": "high"
 			}
 		},
 		"grok-4.6": {
@@ -90855,8 +90856,9 @@
 			"maxTokens": 500000,
 			"thinking": {
 				"mode": "effort",
-				"minLevel": "minimal",
-				"maxLevel": "high"
+				"minLevel": "low",
+				"maxLevel": "xhigh",
+				"defaultLevel": "high"
 			}
 		},
 		"grok-beta": {

package/src/provider-models/descriptors.ts

@@ -166,7 +166,7 @@ export const PROVIDER_DESCRIPTORS: readonly ProviderDescriptor[] = [
 		config => fuguModelManagerOptions(config),
 		catalog("Sakana Fugu", ["FUGU_API_KEY"]),
 	),
-	descriptor("xai", "grok-4-fast-non-reasoning", config => xaiModelManagerOptions(config)),
+	descriptor("xai", "grok-4.6", config => xaiModelManagerOptions(config)),
 	catalogDescriptor(
 		"deepseek",
 		"deepseek-v4-pro",

package/src/providers/openai-completions-compat.ts

@@ -205,7 +205,10 @@ export function detectOpenAICompat(model: Model<"openai-completions">, resolvedB
 		supportsDeveloperRole: !isNonStandard,
 		sendSessionHeaders: false,
 		supportsMultipleSystemMessages: supportsMultipleSystemMessagesDefault,
-		supportsReasoningEffort: !isGrok && !isZai,
+		// xAI Grok 4.5+ accepts reasoning_effort on chat-completions (and
+		// Responses). Older catalogs left this off; keeping it disabled
+		// silently drops profile effort suffixes like :xhigh for grok-4.6.
+		supportsReasoningEffort: !isZai,
 		reasoningEffortMap,
 		supportsUsageInStreaming: !isCerebras,
 		disableReasoningOnForcedToolChoice: isKimiModel || isAnthropicModel || isOpenCodeGoReasoning,