Merge branch 'main' into feat/kimi-code

Author: navedmerchant

packages/types/src/__tests__/lite-llm.test.ts

@@ -12,6 +12,7 @@ describe("LiteLLM preserveReasoning model detection", () => {
 	})
 
 	it("matches provider-prefixed routed model names by their final segment", () => {
+		expect(isLiteLLMPreserveReasoningModel("kimi-k3")).toBe(true)
 		expect(isLiteLLMPreserveReasoningModel("deepseek/deepseek-reasoner")).toBe(true)
 		expect(isLiteLLMPreserveReasoningModel("bedrock/moonshot.kimi-k2-thinking")).toBe(true)
 		expect(isLiteLLMPreserveReasoningModel("fireworks_ai/accounts/fireworks/models/kimi-k2p7-code")).toBe(true)

packages/types/src/__tests__/moonshot.test.ts

@@ -0,0 +1,72 @@
+import type { ModelInfo } from "../model.js"
+import { MOONSHOT_DEFAULT_TEMPERATURE, moonshotDefaultModelId, moonshotModels } from "../providers/moonshot.js"
+
+const modelEntries: [string, ModelInfo][] = Object.entries(moonshotModels)
+const modelInfos: ModelInfo[] = Object.values(moonshotModels)
+
+describe("moonshot registry", () => {
+	describe("moonshotModels registry invariants", () => {
+		it("every entry has a positive maxTokens and contextWindow", () => {
+			for (const [id, info] of modelEntries) {
+				expect(info.maxTokens).toBeGreaterThan(0)
+				expect(info.contextWindow).toBeGreaterThan(0)
+				// Sanity: max output must not exceed the context window.
+				expect(info.maxTokens).toBeLessThanOrEqual(info.contextWindow)
+				void id
+			}
+		})
+
+		it("every entry declares supportsImages and supportsPromptCache", () => {
+			for (const info of modelInfos) {
+				expect(typeof info.supportsImages).toBe("boolean")
+				expect(typeof info.supportsPromptCache).toBe("boolean")
+			}
+		})
+
+		it("models with an array supportsReasoningEffort expose a non-empty allow-list", () => {
+			for (const info of modelInfos) {
+				if (Array.isArray(info.supportsReasoningEffort)) {
+					expect(info.supportsReasoningEffort.length).toBeGreaterThan(0)
+				}
+			}
+		})
+
+		it("every entry declares a reasoningEffort that is covered by its allow-list", () => {
+			for (const info of modelInfos) {
+				if (Array.isArray(info.supportsReasoningEffort) && info.reasoningEffort !== undefined) {
+					expect(info.supportsReasoningEffort).toContain(info.reasoningEffort)
+				}
+			}
+		})
+	})
+
+	describe("kimi-k3", () => {
+		it("exposes always-on reasoning with effort allow-list and reasoning preservation", () => {
+			const info = moonshotModels["kimi-k3"]
+			expect(info).toBeDefined()
+			expect(info.maxTokens).toBe(131_072)
+			expect(info.contextWindow).toBe(1_048_576)
+			expect(info.supportsImages).toBe(true)
+			expect(info.supportsPromptCache).toBe(true)
+			expect(info.supportsReasoningEffort).toEqual(["low", "high", "max"])
+			expect(info.reasoningEffort).toBe("max")
+			expect(info.preserveReasoning).toBe(true)
+			expect(info.defaultTemperature).toBe(1.0)
+			expect(info.inputPrice).toBe(3.0)
+			expect(info.outputPrice).toBe(15.0)
+			expect(info.cacheWritesPrice).toBe(0)
+			expect(info.cacheReadsPrice).toBe(0.3)
+		})
+	})
+
+	describe("defaults", () => {
+		it("the default model id is a curated registry entry", () => {
+			expect(moonshotDefaultModelId).toBe("kimi-k2-0905-preview")
+			expect(moonshotModels[moonshotDefaultModelId]).toBeDefined()
+		})
+
+		it("exposes a deterministic default temperature", () => {
+			expect(MOONSHOT_DEFAULT_TEMPERATURE).toBe(0.6)
+		})
+	})
+})

packages/types/src/__tests__/opencode-go.test.ts

@@ -21,6 +21,7 @@ describe("opencode-go registry", () => {
 		"glm-5",
 		"glm-5.1",
 		"glm-5.2",
+		"kimi-k3",
 		"kimi-k2.5",
 		"kimi-k2.6",
 		"mimo-v2.5",
@@ -60,6 +61,23 @@ describe("opencode-go registry", () => {
 		it("returns undefined for an unknown model ID", () => {
 			expect(getOpencodeGoModelInfo("not-a-real-go-model")).toBeUndefined()
 		})
+
+		it("kimi-k3 exposes always-on reasoning with effort allow-list and reasoning preservation", () => {
+			const info = getOpencodeGoModelInfo("kimi-k3")
+			expect(info).toBeDefined()
+			expect(info?.maxTokens).toBe(131_072)
+			expect(info?.contextWindow).toBe(1_048_576)
+			expect(info?.supportsReasoningEffort).toEqual(["low", "high", "max"])
+			expect(info?.reasoningEffort).toBe("max")
+			expect(info?.preserveReasoning).toBe(true)
+			expect(info?.defaultTemperature).toBe(1.0)
+			expect(info?.supportsPromptCache).toBe(true)
+			expect(info?.supportsMaxTokens).toBe(true)
+			expect(info?.supportsImages).toBe(false)
+			expect(info?.inputPrice).toBe(3.0)
+			expect(info?.outputPrice).toBe(15.0)
+			expect(info?.cacheReadsPrice).toBe(0.3)
+		})
 	})
 
 	describe("OPENCODE_GO_ANTHROPIC_FORMAT_MODELS", () => {

packages/types/src/providers/lite-llm.ts

@@ -48,6 +48,9 @@ export const LITELLM_PRESERVE_REASONING_MODEL_IDS = [
 	"moonshot.kimi-k2-thinking",
 	"kimi-k2p7-code",
 
+	// moonshot.ts, opencode-go.ts
+	"kimi-k3",
+
 	// zai.ts
 	"glm-4.7",
 	"glm-5",

packages/types/src/providers/moonshot.ts

@@ -6,6 +6,21 @@ export type MoonshotModelId = keyof typeof moonshotModels
 export const moonshotDefaultModelId: MoonshotModelId = "kimi-k2-0905-preview"
 
 export const moonshotModels = {
+	"kimi-k3": {
+		maxTokens: 131_072, // Default max_completion_tokens (configurable up to 1,048,576)
+		contextWindow: 1_048_576, // 1M tokens
+		supportsImages: true, // Native visual understanding (text, image, video)
+		supportsPromptCache: true, // Automatic context caching
+		supportsReasoningEffort: ["low", "high", "max"], // Always reasons; default "max"
+		reasoningEffort: "max",
+		preserveReasoning: true,
+		inputPrice: 3.0, // $3.00 per million tokens (cache miss)
+		outputPrice: 15.0, // $15.00 per million tokens
+		cacheWritesPrice: 0, // $0 per million tokens (cache miss)
+		cacheReadsPrice: 0.3, // $0.30 per million tokens (cache hit)
+		defaultTemperature: 1.0, // temperature is fixed at 1.0
+		description: `Kimi K3 is Kimi's most capable flagship model with 2.8 trillion parameters, native visual understanding, and a 1M-token context window, designed for long-horizon coding, knowledge work, and deep reasoning. Thinking is always enabled with configurable reasoning effort (low/high/max, default max).`,
+	},
 	"kimi-k2-0711-preview": {
 		maxTokens: 32_000,
 		contextWindow: 131_072,

packages/types/src/providers/opencode-go.ts

@@ -104,6 +104,23 @@ export const opencodeGoModels: Record<string, ModelInfo> = {
 	},
 
 	// --- Moonshot Kimi ---
+	"kimi-k3": {
+		maxTokens: 131_072, // Default max_completion_tokens (configurable up to 1,048,576)
+		contextWindow: 1_048_576,
+		supportsImages: false,
+		supportsPromptCache: true,
+		supportsMaxTokens: true,
+		supportsReasoningEffort: ["low", "high", "max"], // Always reasons; default "max"
+		reasoningEffort: "max",
+		preserveReasoning: true,
+		defaultTemperature: 1.0,
+		// Go pricing matches Moonshot direct ($3 in / $0.30 cache / $15 out per 1M tokens).
+		inputPrice: 3.0,
+		outputPrice: 15.0,
+		cacheReadsPrice: 0.3,
+		description:
+			"Kimi K3 is Moonshot AI's flagship model with 2.8 trillion parameters, a 1M context window, and always-on reasoning with configurable effort (low/high/max). Available via the Opencode Go plan.",
+	},
 	"kimi-k2.5": {
 		maxTokens: 16_384,
 		contextWindow: 262_144,