fix(catalog): expose Baseten Kimi K3 thinking levels

packages/catalog/CHANGELOG.md

@@ -2,6 +2,10 @@
 
 ## [Unreleased]
 
+### Fixed
+
+- Fixed Baseten's `moonshotai/Kimi-K3` catalog metadata so its `low`/`high`/`max` thinking levels are available.
+
 ## [17.2.15] - 2026-08-12
 
 ### Fixed

packages/catalog/src/models.json

@@ -13622,7 +13622,7 @@
 			"api": "openai-completions",
 			"provider": "baseten",
 			"baseUrl": "https://inference.baseten.co/v1",
-			"reasoning": false,
+			"reasoning": true,
 			"input": [
 				"text",
 				"image"
@@ -13636,7 +13636,20 @@
 			"contextWindow": 1048576,
 			"maxTokens": 262144,
 			"supportsComputerUse": false,
-			"supportsComputerUseConfig": false
+			"supportsComputerUseConfig": false,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"low",
+					"high",
+					"max"
+				],
+				"defaultLevel": "max",
+				"effortMap": {
+					"max": "max"
+				},
+				"requiresEffort": true
+			}
 		},
 		"nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B": {
 			"id": "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B",

packages/catalog/src/provider-models/openai-compat.ts

@@ -3550,15 +3550,18 @@ export function basetenModelManagerOptions(
 			const features = Array.isArray(raw.supported_features) ? raw.supported_features : [];
 			const modalities = Array.isArray(raw.input_modalities) ? raw.input_modalities : [];
 
-			// Baseten's reasoning router accepts only the high/max
-			// effort tiers for its GLM-5.2 and gpt-oss routes.
-			const isEffortReasoning =
+			// Keep dynamic discovery conservative: Baseten's generic feature
+			// list can mention reasoning for routes whose wire dialect OMP does
+			// not know. The shared model-thinking policy derives the exact
+			// effort ladder after this spec is built.
+			const isKnownBasetenReasoningRoute =
+				isKimiK3ModelId(defaults.id) ||
 				defaults.id === "openai/gpt-oss-120b" ||
+				defaults.id === "deepseek-ai/DeepSeek-V4-Pro" ||
 				defaults.id === "zai-org/GLM-5.2" ||
 				defaults.id === "zai-org/GLM-5.2-Fast";
-			const isBasetenNativeReasoning = isEffortReasoning || defaults.id === "deepseek-ai/DeepSeek-V4-Pro";
 			const reasoning =
-				isBasetenNativeReasoning && (features.includes("reasoning") || features.includes("reasoning_effort"));
+				isKnownBasetenReasoningRoute && (features.includes("reasoning") || features.includes("reasoning_effort"));
 			const supportsTools = features.includes("tools") ? undefined : false;
 			const vision = modalities.includes("image") || (reference?.input.includes("image") ?? false);
 
@@ -3572,14 +3575,7 @@ export function basetenModelManagerOptions(
 
 			const contextWindow = toPositiveNumber(raw.context_length, reference?.contextWindow ?? defaults.contextWindow);
 			const maxTokens = toPositiveNumber(raw.max_completion_tokens, reference?.maxTokens ?? defaults.maxTokens);
-
 			const baseModel = mapWithBundledReference(entry, defaults, reference);
-			const thinking = isEffortReasoning
-				? {
-						mode: "effort" as const,
-						efforts: [Effort.High, Effort.Max],
-					}
-				: undefined;
 
 			return {
 				...baseModel,
@@ -3588,7 +3584,6 @@ export function basetenModelManagerOptions(
 				cost,
 				contextWindow,
 				maxTokens,
-				...(thinking ? { thinking } : {}),
 				...(supportsTools === false ? { supportsTools } : {}),
 			};
 		},

packages/catalog/test/baseten-provider.test.ts

@@ -1,4 +1,5 @@
 import { describe, expect, test } from "bun:test";
+import { buildModel } from "@oh-my-pi/pi-catalog/build";
 import { basetenModelManagerOptions } from "@oh-my-pi/pi-catalog/provider-models/openai-compat";
 import type { FetchImpl } from "@oh-my-pi/pi-catalog/types";
 
@@ -28,6 +29,20 @@ describe("Baseten provider discovery", () => {
 								input_cache_read: "0.00000016",
 							},
 						},
+						{
+							id: "moonshotai/Kimi-K3",
+							object: "model",
+							name: "Kimi K3",
+							context_length: 1048576,
+							max_completion_tokens: 262144,
+							supported_features: ["tools", "json_mode", "structured_outputs", "reasoning_effort"],
+							input_modalities: ["text", "image"],
+							pricing: {
+								prompt: "0.000003",
+								completion: "0.000015",
+								input_cache_read: "0.0000003",
+							},
+						},
 						{
 							id: "deepseek-ai/DeepSeek-V4-Pro",
 							object: "model",
@@ -90,6 +105,15 @@ describe("Baseten provider discovery", () => {
 			},
 		});
 
+		const kimiK3 = models?.find(model => model.id === "moonshotai/Kimi-K3");
+		if (!kimiK3) throw new Error("Baseten Kimi K3 was not discovered");
+		expect(kimiK3.reasoning).toBe(true);
+		expect(buildModel(kimiK3).thinking).toMatchObject({
+			mode: "effort",
+			efforts: ["low", "high", "max"],
+			defaultLevel: "max",
+		});
+
 		const deepseek = models?.find(model => model.id === "deepseek-ai/DeepSeek-V4-Pro");
 		expect(deepseek).toBeDefined();
 		expect(deepseek).toMatchObject({
@@ -110,14 +134,10 @@ describe("Baseten provider discovery", () => {
 
 		const glmFast = models?.find(model => model.id === "zai-org/GLM-5.2-Fast");
 		expect(glmFast).toBeDefined();
-		expect(glmFast).toMatchObject({
-			provider: "baseten",
-			api: "openai-completions",
-			reasoning: true,
-			thinking: {
-				mode: "effort",
-				efforts: ["high", "max"],
-			},
+		if (!glmFast) throw new Error("Baseten GLM-5.2 Fast was not discovered");
+		expect(buildModel(glmFast).thinking).toMatchObject({
+			mode: "effort",
+			efforts: ["high", "max"],
 		});
 	});
 });