@namzu/zen 2.5.0 → 3.0.0

package/src/models.ts

@@ -20,7 +20,7 @@ export interface ZenModel {
 	readonly outputPrice: number
 	readonly supportsToolUse: boolean
 	readonly supportsStreaming: boolean
-	/** True only for documented anonymous Zen models; omission grants no anonymous access. */
+	/** True only for verified direct-anonymous Zen models; zero price does not grant access. */
 	readonly supportsAnonymousAccess?: boolean
 	/** Exact advertised selectable effort levels; empty means no effort selector. */
 	readonly effortLevels?: readonly ReasoningEffort[]
@@ -35,7 +35,7 @@ export interface ZenModel {
  * derived is an edit to src/catalogue/derive.ts. Both survive regeneration; a
  * hand edit here does not.
  *
- * Refreshed: 2026-09-21
+ * Refreshed: 2026-09-28
  *
  * This is the BUNDLED snapshot: what the driver knows with no network. A host
  * can derive a fresher roster at run time from the same sources, by the same
@@ -44,7 +44,8 @@ export interface ZenModel {
  *
  * Routes come from each service's own documentation page, as the pair
  * (endpoint, AI SDK package) that page states per model; both halves must agree
- * or the run stops. Prices come from that page's per-1M-token table at the base
+ * or the run stops. Three Go ids have narrow reference-client exceptions in
+ * src/catalogue/derive.ts. Prices come from the per-1M-token table at the base
  * tier, and a model with no price row there must be named in the page's
  * free-model list and is then zero. Limits, tool support, modalities and effort
  * options come from the matching provider entry in models.dev, where the
@@ -71,9 +72,9 @@ export interface ZenModel {
  * does not expose those input kinds.
  *
  * Promotional free models with documented routes and complete metadata are
- * included at their advertised zero price; the service enforces access limits.
- * Anonymous admission is explicit rather than inferred from price, and comes
- * from the free-model list on the page. OpenCode's own loader uses the public
+ * included at their advertised zero price. Direct anonymous admission is a
+ * separate, live-verified curation in src/catalogue/derive.ts and still requires
+ * an entry in the page's free-model list. OpenCode's own loader uses the public
  * sentinel when credentials are absent:
  * https://github.com/anomalyco/opencode/blob/16747470f976aca3d362ad730bcd3fe82ecc2c9a/packages/opencode/src/provider/provider.ts#L185
  * Unknown IDs have no inferred protocol or limits. MiniMax and Qwen demonstrate
@@ -102,6 +103,32 @@ const ZEN_MODELS = freezeModels([
 		supportsStreaming: true,
 		effortLevels: ['low', 'medium', 'high', 'xhigh', 'max'],
 	},
+	{
+		id: 'gpt-6-sol',
+		name: 'GPT 6 Sol',
+		protocol: 'responses',
+		contextWindow: 1050000,
+		maxOutputTokens: 128000,
+		inputModalities: ['text', 'image', 'document'],
+		inputPrice: 2.0,
+		outputPrice: 10.0,
+		supportsToolUse: true,
+		supportsStreaming: true,
+		effortLevels: ['none', 'low', 'medium', 'high', 'xhigh', 'max'],
+	},
+	{
+		id: 'gpt-6-luna',
+		name: 'GPT 6 Luna',
+		protocol: 'responses',
+		contextWindow: 1050000,
+		maxOutputTokens: 128000,
+		inputModalities: ['text', 'image', 'document'],
+		inputPrice: 0.1,
+		outputPrice: 0.5,
+		supportsToolUse: true,
+		supportsStreaming: true,
+		effortLevels: ['none', 'low', 'medium', 'high', 'xhigh', 'max'],
+	},
 	{
 		id: 'gpt-5.6-sol',
 		name: 'GPT 5.6 Sol',
@@ -323,6 +350,19 @@ const ZEN_MODELS = freezeModels([
 		supportsStreaming: true,
 		effortLevels: ['low', 'medium', 'high', 'xhigh', 'max'],
 	},
+	{
+		id: 'claude-opus-5-5',
+		name: 'Claude Opus 5.5',
+		protocol: 'messages',
+		contextWindow: 1000000,
+		maxOutputTokens: 128000,
+		inputModalities: ['text', 'image', 'document'],
+		inputPrice: 4.0,
+		outputPrice: 20.0,
+		supportsToolUse: true,
+		supportsStreaming: true,
+		effortLevels: ['low', 'medium', 'high', 'xhigh', 'max'],
+	},
 	{
 		id: 'claude-opus-5',
 		name: 'Claude Opus 5',
@@ -596,6 +636,19 @@ const ZEN_MODELS = freezeModels([
 		supportsStreaming: true,
 		effortLevels: ['minimal', 'low', 'medium', 'high', 'xhigh'],
 	},
+	{
+		id: 'qwen3.8-max',
+		name: 'Qwen3.8 Max',
+		protocol: 'chat',
+		contextWindow: 262144,
+		maxOutputTokens: 131072,
+		inputModalities: ['text', 'image'],
+		inputPrice: 2.0,
+		outputPrice: 6.0,
+		supportsToolUse: true,
+		supportsStreaming: true,
+		effortLevels: [],
+	},
 	{
 		id: 'qwen3.8-flash',
 		name: 'Qwen3.8 Flash',
@@ -806,7 +859,6 @@ const ZEN_MODELS = freezeModels([
 	},
 	{
 		id: 'big-pickle',
-		supportsAnonymousAccess: true,
 		name: 'Big Pickle',
 		protocol: 'chat',
 		contextWindow: 200000,
@@ -819,8 +871,34 @@ const ZEN_MODELS = freezeModels([
 		effortLevels: [],
 	},
 	{
-		id: 'mimo-v2.5-free',
+		id: 'space-bunny-free',
 		supportsAnonymousAccess: true,
+		name: 'Space Bunny Free',
+		protocol: 'chat',
+		contextWindow: 1048576,
+		maxOutputTokens: 524288,
+		inputModalities: ['text', 'image'],
+		inputPrice: 0,
+		outputPrice: 0,
+		supportsToolUse: true,
+		supportsStreaming: true,
+		effortLevels: ['low', 'medium', 'high', 'xhigh', 'max'],
+	},
+	{
+		id: 'longcat-2.5-preview-free',
+		name: 'LongCat 2.5 Preview Free',
+		protocol: 'chat',
+		contextWindow: 1000000,
+		maxOutputTokens: 131072,
+		inputModalities: ['text', 'image'],
+		inputPrice: 0,
+		outputPrice: 0,
+		supportsToolUse: true,
+		supportsStreaming: true,
+		effortLevels: [],
+	},
+	{
+		id: 'mimo-v2.5-free',
 		name: 'MiMo-V2.5 Free',
 		protocol: 'chat',
 		contextWindow: 200000,
@@ -834,7 +912,6 @@ const ZEN_MODELS = freezeModels([
 	},
 	{
 		id: 'ling-3.0-flash-fin-free',
-		supportsAnonymousAccess: true,
 		name: 'Ling 3.0 Flash Fin Free',
 		protocol: 'chat',
 		contextWindow: 262144,
@@ -848,7 +925,6 @@ const ZEN_MODELS = freezeModels([
 	},
 	{
 		id: 'nemotron-3-ultra-free',
-		supportsAnonymousAccess: true,
 		name: 'Nemotron 3 Ultra Free',
 		protocol: 'chat',
 		contextWindow: 1000000,
@@ -862,7 +938,6 @@ const ZEN_MODELS = freezeModels([
 	},
 	{
 		id: 'nemotron-3.5-lightning-free',
-		supportsAnonymousAccess: true,
 		name: 'Nemotron 3.5 Lightning Free',
 		protocol: 'chat',
 		contextWindow: 262144,
@@ -876,7 +951,6 @@ const ZEN_MODELS = freezeModels([
 	},
 	{
 		id: 'muse-spark-1.3-contributor-free',
-		supportsAnonymousAccess: true,
 		name: 'Muse Spark 1.3 Contributor Free',
 		protocol: 'responses',
 		contextWindow: 1048576,
@@ -917,6 +991,19 @@ const GO_MODELS = freezeModels([
 		supportsStreaming: true,
 		effortLevels: ['low', 'medium', 'high', 'xhigh'],
 	},
+	{
+		id: 'gpt-6-luna',
+		name: 'GPT 6 Luna',
+		protocol: 'responses',
+		contextWindow: 1050000,
+		maxOutputTokens: 128000,
+		inputModalities: ['text', 'image', 'document'],
+		inputPrice: 0.1,
+		outputPrice: 0.5,
+		supportsToolUse: true,
+		supportsStreaming: true,
+		effortLevels: ['none', 'low', 'medium', 'high', 'xhigh', 'max'],
+	},
 	{
 		id: 'gpt-5.6-luna',
 		name: 'GPT 5.6 Luna',
@@ -969,19 +1056,6 @@ const GO_MODELS = freezeModels([
 		supportsStreaming: true,
 		effortLevels: ['high', 'max'],
 	},
-	{
-		id: 'glm-5.1',
-		name: 'GLM-5.1',
-		protocol: 'chat',
-		contextWindow: 202752,
-		maxOutputTokens: 32768,
-		inputModalities: ['text'],
-		inputPrice: 1.4,
-		outputPrice: 4.4,
-		supportsToolUse: true,
-		supportsStreaming: true,
-		effortLevels: [],
-	},
 	{
 		id: 'kimi-k3',
 		name: 'Kimi K3',
@@ -1034,6 +1108,19 @@ const GO_MODELS = freezeModels([
 		supportsStreaming: true,
 		effortLevels: [],
 	},
+	{
+		id: 'longcat-2.5-preview-free',
+		name: 'LongCat 2.5 Preview Free',
+		protocol: 'chat',
+		contextWindow: 1000000,
+		maxOutputTokens: 131072,
+		inputModalities: ['text', 'image'],
+		inputPrice: 0,
+		outputPrice: 0,
+		supportsToolUse: true,
+		supportsStreaming: true,
+		effortLevels: [],
+	},
 	{
 		id: 'deepseek-v4.1-flash',
 		name: 'DeepSeek V4.1 Flash',
@@ -1063,7 +1150,7 @@ const GO_MODELS = freezeModels([
 	{
 		id: 'deepseek-v4-flash',
 		name: 'DeepSeek V4 Flash',
-		protocol: 'chat',
+		protocol: 'responses',
 		contextWindow: 1000000,
 		maxOutputTokens: 384000,
 		inputModalities: ['text'],
@@ -1115,7 +1202,7 @@ const GO_MODELS = freezeModels([
 	{
 		id: 'minimax-m3',
 		name: 'MiniMax M3',
-		protocol: 'messages',
+		protocol: 'chat',
 		contextWindow: 1000000,
 		maxOutputTokens: 131072,
 		inputModalities: ['text', 'image'],
@@ -1128,7 +1215,7 @@ const GO_MODELS = freezeModels([
 	{
 		id: 'minimax-m2.7',
 		name: 'MiniMax M2.7',
-		protocol: 'messages',
+		protocol: 'chat',
 		contextWindow: 204800,
 		maxOutputTokens: 131072,
 		inputModalities: ['text'],
@@ -1190,19 +1277,6 @@ const GO_MODELS = freezeModels([
 		supportsStreaming: true,
 		effortLevels: ['low', 'medium', 'xhigh'],
 	},
-	{
-		id: 'qwen3.7-max',
-		name: 'Qwen3.7 Max',
-		protocol: 'messages',
-		contextWindow: 1000000,
-		maxOutputTokens: 65536,
-		inputModalities: ['text'],
-		inputPrice: 2.5,
-		outputPrice: 7.5,
-		supportsToolUse: true,
-		supportsStreaming: true,
-		effortLevels: [],
-	},
 	{
 		id: 'qwen3.7-plus',
 		name: 'Qwen3.7 Plus',
@@ -1216,19 +1290,6 @@ const GO_MODELS = freezeModels([
 		supportsStreaming: true,
 		effortLevels: [],
 	},
-	{
-		id: 'qwen3.6-plus',
-		name: 'Qwen3.6 Plus',
-		protocol: 'messages',
-		contextWindow: 1000000,
-		maxOutputTokens: 65536,
-		inputModalities: ['text', 'image'],
-		inputPrice: 0.5,
-		outputPrice: 3.0,
-		supportsToolUse: true,
-		supportsStreaming: true,
-		effortLevels: [],
-	},
 	{
 		id: 'hy4-preview',
 		name: 'Hy4 preview',
@@ -1255,6 +1316,19 @@ const GO_MODELS = freezeModels([
 		supportsStreaming: true,
 		effortLevels: ['none', 'low', 'high'],
 	},
+	{
+		id: 'space-bunny-free',
+		name: 'Space Bunny Free',
+		protocol: 'chat',
+		contextWindow: 1048576,
+		maxOutputTokens: 524288,
+		inputModalities: ['text', 'image'],
+		inputPrice: 0,
+		outputPrice: 0,
+		supportsToolUse: true,
+		supportsStreaming: true,
+		effortLevels: ['low', 'medium', 'high', 'xhigh', 'max'],
+	},
 ])
 
 /**
@@ -1265,6 +1339,7 @@ const GO_MODELS = freezeModels([
 export const ZEN_OMITTED_MODELS: readonly string[] = Object.freeze([
 	'go/deepseek-flash',
 	'go/glm-5',
+	'go/glm-5.1',
 	'go/grok-4.5',
 	'go/hy3-preview',
 	'go/kimi-k2.5',
@@ -1275,6 +1350,8 @@ export const ZEN_OMITTED_MODELS: readonly string[] = Object.freeze([
 	'go/minimax-m2.5',
 	'go/omen-alpha',
 	'go/qwen3.5-plus',
+	'go/qwen3.6-plus',
+	'go/qwen3.7-max',
 	'zen/claude-sonnet-4',
 	'zen/deepseek-v4-flash-free',
 	'zen/glm-5',

package/src/options.ts

@@ -1,19 +1,33 @@
 import { getModelCapabilities as getMessagesCapabilities } from '@ai-sdk/anthropic/internal'
 import type { LanguageModelV3CallOptions, SharedV3ProviderOptions } from '@ai-sdk/provider'
 import { type ChatCompletionParams, ProviderRequestError, type ProviderRoute } from '@namzu/sdk'
-import { type ZenProtocol, type ZenService, findZenModel } from './models.js'
-import { toModelPrompt } from './prompt.js'
+import type { ZenModel, ZenProtocol, ZenService } from './models.js'
+import { isMuseSparkResponsesModel, toModelPrompt } from './prompt.js'
+
+function goRejectsToolChoice(service: ZenService, model: string): boolean {
+	if (service !== 'go') return false
+	if (model.startsWith('mimo-')) return true
+	if (model === 'deepseek-flash') return true
+	return (
+		/^deepseek-v4(?:\.\d+)?-(?:flash|pro)(?:$|-)/.test(model) &&
+		model !== 'deepseek-v4-flash-vision-exp'
+	)
+}
 
 export function createCallOptions(
 	params: ChatCompletionParams,
 	route: ProviderRoute,
 	service: ZenService,
 	protocol: ZenProtocol,
+	known: ZenModel | undefined,
 ): LanguageModelV3CallOptions {
 	const refuse = (detail: string): never => {
-		throw new ProviderRequestError({ providerId: route.providerId, kind: 'bad_request', detail })
+		throw new ProviderRequestError({
+			providerId: route.providerId,
+			kind: 'bad_request',
+			detail,
+		})
 	}
-	const known = findZenModel(service, params.model)
 	let maxOutputTokens = params.maxTokens ?? 4096
 	if (!Number.isSafeInteger(maxOutputTokens) || maxOutputTokens <= 0)
 		refuse('maxTokens must be a positive integer.')
@@ -46,7 +60,9 @@ export function createCallOptions(
 					: {}),
 				...(params.thinking ? { thinking: { type: params.thinking.type } } : {}),
 				...(params.responseFormat?.type === 'json_schema'
-					? { strictJsonSchema: params.responseFormat.json_schema.strict ?? true }
+					? {
+							strictJsonSchema: params.responseFormat.json_schema.strict ?? true,
+						}
 					: {}),
 			}
 			break
@@ -95,7 +111,10 @@ export function createCallOptions(
 				refuse('This responses model does not support reasoning controls.')
 			options.openai = {
 				store: false,
-				include: ['reasoning.encrypted_content'],
+				// Muse Spark's gateway cannot replay caller-bound encrypted reasoning.
+				...(isMuseSparkResponsesModel(params.model, protocol)
+					? {}
+					: { include: ['reasoning.encrypted_content'] }),
 				...(known?.effortLevels?.length ? { forceReasoning: true } : {}),
 				...(params.effort !== undefined ? { reasoningEffort: params.effort } : {}),
 				...(params.thinking?.type === 'disabled' ? { reasoningEffort: 'none' } : {}),
@@ -104,7 +123,9 @@ export function createCallOptions(
 					? { parallelToolCalls: params.parallelToolCalls }
 					: {}),
 				...(params.responseFormat?.type === 'json_schema'
-					? { strictJsonSchema: params.responseFormat.json_schema.strict ?? true }
+					? {
+							strictJsonSchema: params.responseFormat.json_schema.strict ?? true,
+						}
 					: {}),
 			}
 			break
@@ -211,7 +232,16 @@ export function createCallOptions(
 	// Some Responses upstreams accept only auto tool choice (observed on Muse).
 	// A no-tools turn can preserve its meaning by exposing no tools at all.
 	// Keep required/named choices explicit; never silently weaken them to auto.
-	const withoutTools = protocol === 'responses' && params.toolChoice === 'none'
+	const rejectsToolChoice = goRejectsToolChoice(service, params.model)
+	if (
+		rejectsToolChoice &&
+		params.toolChoice !== undefined &&
+		params.toolChoice !== 'auto' &&
+		params.toolChoice !== 'none'
+	)
+		refuse('This Zen Go model does not support required or named tool choice.')
+	const withoutTools =
+		(protocol === 'responses' || rejectsToolChoice) && params.toolChoice === 'none'
 	const enforced = new Set(params.enforceToolInputSchema)
 	return {
 		prompt: toModelPrompt(params, route, service, protocol),
@@ -241,7 +271,7 @@ export function createCallOptions(
 					})),
 				}
 			: {}),
-		...(params.toolChoice && !withoutTools
+		...(params.toolChoice && !withoutTools && !rejectsToolChoice
 			? {
 					toolChoice:
 						typeof params.toolChoice === 'string'

package/src/prompt.ts

@@ -24,6 +24,20 @@ type NativePart = Extract<LanguageModelV3Content, { type: 'text' | 'reasoning' |
 type AssistantContent = Extract<LanguageModelV3Message, { role: 'assistant' }>['content']
 type RichOutput = Extract<LanguageModelV3ToolResultOutput, { type: 'content' }>['value']
 
+/** Muse Spark's Responses gateway cannot round-trip caller-bound encrypted reasoning. */
+export function isMuseSparkResponsesModel(model: string, protocol: ZenProtocol): boolean {
+	return protocol === 'responses' && model.startsWith('muse-spark-')
+}
+
+/** Go's DeepSeek V4 Responses lane requires reasoning before each assistant turn. */
+export function isGoDeepSeekV4ResponsesModel(
+	service: ZenService,
+	model: string,
+	protocol: ZenProtocol,
+): boolean {
+	return service === 'go' && protocol === 'responses' && model === 'deepseek-v4-flash'
+}
+
 class HistoryConversionError extends Error {}
 
 function record(value: unknown): Record<string, unknown> | undefined {
@@ -137,7 +151,14 @@ function prefixDigest(messages: readonly Message[]): string {
 					chainIndex: message.source.chainIndex,
 				},
 			}),
+			// Bookkeeping the kernel stamps, not content the prefix is defined
+			// by: `id` names which durable record a message came from, and two
+			// requests built from the exact same prefix disagree on it whenever
+			// one side is a fresh, never-recorded copy of the other (or a
+			// fork's copy, carrying a different session's id for identical
+			// content).
 			timestamp: undefined,
+			id: undefined,
 		})),
 	)
 }
@@ -275,7 +296,23 @@ function replayContent(
 		state.projectionDigest !== digest(projection(message))
 	)
 		return undefined
-	return content.map((part) => {
+	// Validate the original completed turn before dropping Muse's reasoning items.
+	// Their encrypted content belongs to the gateway's upstream caller and fails
+	// when replayed in a later tool round. Text and tool calls still round-trip.
+	let replayParts = isMuseSparkResponsesModel(params.model, protocol)
+		? content.filter((part) => part.type !== 'reasoning')
+		: content
+	if (isGoDeepSeekV4ResponsesModel(service, params.model, protocol)) {
+		// The native adapter emits each part in source order. Go rejects an
+		// assistant message between a function_call and its output, so group
+		// captured reasoning and text before this turn's tool-call batch.
+		replayParts = [
+			...replayParts.filter((part) => part.type === 'reasoning'),
+			...replayParts.filter((part) => part.type === 'text'),
+			...replayParts.filter((part) => part.type === 'tool-call'),
+		]
+	}
+	return replayParts.map((part) => {
 		const providerOptions =
 			part.providerMetadata === undefined
 				? undefined
@@ -284,7 +321,10 @@ function replayContent(
 			// The compatible adapter reads Gemini call signatures from the google namespace.
 			const thoughtSignature = metadataString(part, 'thoughtSignature')
 			if (protocol === 'chat' && providerOptions && thoughtSignature !== undefined) {
-				providerOptions.google = { ...providerOptions.google, thoughtSignature }
+				providerOptions.google = {
+					...providerOptions.google,
+					thoughtSignature,
+				}
 			}
 			return {
 				type: 'tool-call',
@@ -294,7 +334,11 @@ function replayContent(
 				...(providerOptions && { providerOptions }),
 			}
 		}
-		return { type: part.type, text: part.text, ...(providerOptions && { providerOptions }) }
+		return {
+			type: part.type,
+			text: part.text,
+			...(providerOptions && { providerOptions }),
+		}
 	})
 }
 
@@ -454,7 +498,9 @@ function isRequestOnlyContext(message: Message): boolean {
 	)
 }
 
-const BREAKPOINT = { anthropic: { cacheControl: { type: 'ephemeral' } } } as const
+const BREAKPOINT = {
+	anthropic: { cacheControl: { type: 'ephemeral' } },
+} as const
 
 /**
  * Messages-protocol cache breakpoints, block-level: after the last static

package/src/types.ts

@@ -2,11 +2,11 @@ import type { ZenCatalogue } from './catalogue/catalogue.js'
 import type { ZenProtocol } from './models.js'
 
 export interface ZenConfig {
-	/** Omit for documented free Zen models. The public sentinel also selects anonymous access. */
+	/** Omit for a model admitted to direct anonymous Zen access; `public` selects the same path. */
 	apiKey?: string
 	/** Stable conversation identity, shared by every turn and auxiliary request. */
 	sessionId?: string
-	/** Defaults to free Muse Spark 1.3 anonymously, or glm-5.3-flash with an API key. */
+	/** Defaults to Space Bunny Free anonymously, or glm-5.3-flash with an API key. */
 	model?: string
 	/** Override the service base URL, primarily for a host-owned proxy. */
 	baseURL?: string
@@ -16,11 +16,12 @@ export interface ZenConfig {
 	protocol?: ZenProtocol
 	/**
 	 * A runtime catalogue, from `@namzu/zen/catalogue`, consulted before the
-	 * bundled snapshot for every lookup: routing, anonymous admission, listing,
-	 * context windows and effort levels. A function is called at each lookup, so
-	 * a host can swap in a fresher catalogue for providers it already built;
-	 * returning `undefined` means the bundled snapshot alone. Omitted, the
-	 * provider uses the bundled snapshot and nothing is ever fetched for it.
+	 * bundled snapshot for every lookup: routing, model listing, context windows
+	 * and effort levels. Anonymous admission also requires the driver's curated
+	 * direct-access ID; a catalogue cannot add to that set. A function is called
+	 * at each lookup, so a host can swap in a fresher catalogue for providers it
+	 * already built. Returning `undefined` means the bundled snapshot alone.
+	 * Omitted, the provider uses the bundled snapshot and fetches nothing.
 	 */
 	catalogue?: ZenCatalogue | (() => ZenCatalogue | undefined)
 }