@oh-my-pi/pi-catalog 17.3.5 → 17.3.8

package/src/provider-models/cache-provider-id.ts

@@ -73,8 +73,10 @@ export function resolveModelCacheProviderId(providerId: string, options: ModelCa
 		case "openrouter":
 			return "openrouter:pseudo-api";
 		case "vllm": {
+			// v2: qwen3.8 rows cached before the reasoning/template-effort upgrade
+			// carry `reasoning: false` and must be refetched.
 			const baseUrl = options.baseUrl ?? getDefaultModelDiscoveryBaseUrl(providerId)!;
-			return `vllm:${Bun.hash(baseUrl).toString(36)}`;
+			return `vllm:models-v2:${Bun.hash(baseUrl).toString(36)}`;
 		}
 		default:
 			return providerId;

package/src/provider-models/descriptors.ts

@@ -474,17 +474,18 @@ export const CATALOG_PROVIDERS = [
 		defaultModel: "openai/gpt-oss-120b",
 		envVars: ["COREWEAVE_API_KEY", "WANDB_API_KEY"],
 		createModelManagerOptions: (config: ModelManagerConfig) => coreWeaveModelManagerOptions(config),
+		dynamicModelsAuthoritative: true,
 		catalogDiscovery: { label: "CoreWeave Serverless Inference" },
 	},
 	{
 		id: "xai",
-		defaultModel: "grok-4.5",
+		defaultModel: "grok-4.6",
 		envVars: ["XAI_API_KEY"],
 		createModelManagerOptions: (config: ModelManagerConfig) => xaiModelManagerOptions(config),
 	},
 	{
 		id: "xai-oauth",
-		defaultModel: "grok-4.5",
+		defaultModel: "grok-4.6",
 		envVars: ["XAI_OAUTH_TOKEN", "XAI_API_KEY"],
 		createModelManagerOptions: (config: ModelManagerConfig) => xaiOAuthModelManagerOptions(config),
 		catalogDiscovery: {

package/src/provider-models/openai-compat.ts

@@ -16,6 +16,7 @@ import {
 	isGrokReasoningEffortCapable,
 	isKimiK3ModelId,
 	isKimiModelId,
+	isQwen38PlusTemplateEffortModelId,
 	isReasoningGlmModelId,
 } from "../identity/family";
 import { resolveModelReference } from "../identity/reference";
@@ -1073,10 +1074,61 @@ export interface GmiCloudModelManagerConfig {
 	fetch?: FetchImpl;
 }
 
+/**
+ * Map a discovered GMI Cloud model to a full spec.
+ *
+ * GMI's `/v1/models` returns only bare `{id}` rows, so discovery defaults carry
+ * no limits, reasoning, or thinking metadata. When a gmi-cloud bundled
+ * reference exists (the seeded default) it supplies GMI's published tariff and
+ * limits directly. Every other id is an open-weight model GMI resells under its
+ * canonical id (`deepseek-ai/…`, `moonshotai/…`, `zai-org/…`, `Qwen/…`), so its
+ * intrinsic capabilities — context window, output limit, reasoning, thinking
+ * ladder — are recovered from any bundled upstream entry via the canonical
+ * reference index. Pricing is deliberately never borrowed across providers:
+ * GMI's per-model tariff is unknown for these ids, so cost stays zeroed rather
+ * than inheriting another provider's rate.
+ */
+function mapGmiCloudModel(
+	entry: OpenAICompatibleModelRecord,
+	defaults: ModelSpec<"openai-completions">,
+	reference: ModelSpec<"openai-completions"> | undefined,
+): ModelSpec<"openai-completions"> {
+	if (reference) {
+		return mapWithBundledReference(entry, defaults, reference);
+	}
+	const canonical = resolveModelReference(defaults.id, getBundledModelReferenceIndex()) as
+		| ModelSpec<"openai-completions">
+		| undefined;
+	if (!canonical) {
+		return { ...defaults, name: toModelName(entry.name, defaults.name) };
+	}
+	const contextWindow = canonical.contextWindow ?? defaults.contextWindow;
+	const maxTokens =
+		canonical.maxTokens != null && contextWindow != null
+			? Math.min(canonical.maxTokens, contextWindow)
+			: (canonical.maxTokens ?? defaults.maxTokens);
+	return {
+		...defaults,
+		name: toModelName(entry.name, canonical.name ?? defaults.name),
+		reasoning: canonical.reasoning,
+		input: canonical.input,
+		...(canonical.thinking && { thinking: canonical.thinking }),
+		contextWindow,
+		maxTokens,
+	};
+}
+
 export function gmiCloudModelManagerOptions(
 	config?: GmiCloudModelManagerConfig,
 ): ModelManagerOptions<"openai-completions"> {
-	return createSimpleOpenAICompletionsOptions("gmi-cloud", GMI_CLOUD_BASE_URL, config);
+	return createOpenAICompatibleModelManagerOptions({
+		api: "openai-completions",
+		providerId: "gmi-cloud",
+		defaultBaseUrl: GMI_CLOUD_BASE_URL,
+		config,
+		requireApiKey: true,
+		mapModel: mapGmiCloudModel,
+	});
 }
 
 // ---------------------------------------------------------------------------
@@ -2980,6 +3032,10 @@ export const ALIBABA_TOKEN_PLAN_DISCOVERED_MODEL_LIMITS: Readonly<Record<string,
 		contextWindow: 1_000_000,
 		maxTokens: 384_000,
 	},
+	"deepseek-v4-pro-0813": {
+		contextWindow: 1_000_000,
+		maxTokens: 384_000,
+	},
 	"deepseek-v3.2": {
 		contextWindow: 131_072,
 		maxTokens: 65_536,
@@ -4994,6 +5050,11 @@ export function vllmModelManagerOptions(config?: VllmModelManagerConfig): ModelM
 					return {
 						...model,
 						contextWindow: toPositiveNumber(entry.max_model_len, model.contextWindow),
+						// vLLM's /v1/models reports no reasoning capability. Qwen 3.8+
+						// open weights always think (the template cannot disable it), so
+						// light up the effort dial; buildModel derives the template
+						// ladder from the id + local-backend compat.
+						reasoning: model.reasoning || isQwen38PlusTemplateEffortModelId(model.id),
 					};
 				},
 				fetch: config?.fetch,
@@ -5071,8 +5132,18 @@ export interface GithubCopilotModelManagerConfig {
 
 const COPILOT_ANTHROPIC_MODEL_PATTERN = /^claude-(haiku|sonnet|opus|fable|mythos)-\d/;
 const isCopilotResponsesModelId = (modelId: string): boolean =>
-	modelId === "grok-4.5" || modelId.startsWith("gpt-5") || modelId.startsWith("oswe") || modelId.startsWith("mai-");
-const COPILOT_CACHE_INVALIDATED_MODEL_IDS = ["grok-4.5", "grok-4.5-1m", "mai-code-1-flash-picker"];
+	modelId === "grok-4.5" ||
+	modelId === "grok-4.6" ||
+	modelId.startsWith("gpt-5") ||
+	modelId.startsWith("oswe") ||
+	modelId.startsWith("mai-");
+const COPILOT_CACHE_INVALIDATED_MODEL_IDS = [
+	"grok-4.5",
+	"grok-4.5-1m",
+	"grok-4.6",
+	"grok-4.6-1m",
+	"mai-code-1-flash-picker",
+];
 
 function inferCopilotApi(modelId: string): Api {
 	if (COPILOT_ANTHROPIC_MODEL_PATTERN.test(modelId)) {
@@ -5705,8 +5776,17 @@ const OPENCODE_ZEN_API_RESOLUTION = createOpenCodeApiResolution("https://opencod
 // /zen/go/v1/chat/completions route does not work for this model while
 // /zen/go/v1/responses does (user-verified against the live gateway,
 // 2026-08-08; Flash only — deepseek-v4-pro serves fine on chat completions).
+//
+// muse-spark-1.2 / muse-spark-1.2-contributor are the same inverse case: the
+// Go gateway's /zen/go/v1/models discovery drops the `provider.npm` hint, so
+// without an override they fall through to openai-completions even though the
+// gateway only serves them at /zen/go/v1/responses (@ai-sdk/openai per
+// https://opencode.ai/docs/go/#endpoints). The completions parser then closes
+// the stream with no finish_reason on every tool-call turn (#8957).
 const OPENCODE_GO_API_RESOLUTION = createOpenCodeApiResolution("https://opencode.ai/zen/go", {
 	"deepseek-v4-flash": "openai-responses",
+	"muse-spark-1.2": "openai-responses",
+	"muse-spark-1.2-contributor": "openai-responses",
 	"minimax-m2.7": "openai-completions",
 	"minimax-m3": "openai-completions",
 	"minimax-m3-free": "openai-completions",

package/src/types.ts

@@ -259,6 +259,16 @@ export interface OpenAICompat {
 	 * Non-Qwen templates ignore the flag, so the auto-detection is safe.
 	 */
 	qwenPreserveThinking?: boolean;
+	/**
+	 * Route the requested thinking effort onto the Qwen 3.8+ chat template's
+	 * `reasoning_effort` kwarg (`low`/`medium`/`xhigh`; template default
+	 * `xhigh`). Emitted inside `chat_template_kwargs` for both Qwen dialects
+	 * (plus the top-level field on the `qwen` dialect, which newer llama.cpp
+	 * builds map natively). Without it the qwen dialects only toggle
+	 * `enable_thinking` and the template always thinks at its `xhigh` default.
+	 * Default: auto-detected (Qwen 3.8+ id on a local llama.cpp-style backend).
+	 */
+	qwenTemplateReasoningEffort?: boolean;
 	/** Whether assistant tool-call messages must include non-empty content. Default: false. */
 	requiresAssistantContentForToolCalls?: boolean;
 	/** Whether the provider supports the `tool_choice` parameter. Default: true. */
@@ -604,6 +614,7 @@ export interface ResolvedOpenAISharedCompat {
 	allowsSyntheticReasoningContentForToolCalls: boolean;
 	replayReasoningContent: boolean;
 	qwenPreserveThinking: boolean;
+	qwenTemplateReasoningEffort: boolean;
 	requiresThinkingAsText: boolean;
 	requiresMistralToolIds: boolean;
 	requiresToolResultName: boolean;
@@ -669,6 +680,7 @@ export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat &
 			| "allowsSyntheticReasoningContentForToolCalls"
 			| "replayReasoningContent"
 			| "qwenPreserveThinking"
+			| "qwenTemplateReasoningEffort"
 			| "requiresThinkingAsText"
 			| "requiresMistralToolIds"
 			| "requiresToolResultName"

package/src/variant-collapse.ts

@@ -111,19 +111,20 @@ function thinkingPair(baseId: string, name: string): EffortVariantFamily {
 	};
 }
 
-type DevinTierRoutes = Partial<Record<"off" | "minimal" | "low" | "medium" | "high" | "xhigh" | "max", string>>;
+type TierRoutes = Partial<Record<"off" | "minimal" | "low" | "medium" | "high" | "xhigh" | "max", string>>;
 
 /** Devin families with a `-max` sibling: five wire tiers, `low` floor. */
 const DEVIN_FIVE_TIER_EFFORTS: readonly Effort[] = [Effort.Low, Effort.Medium, Effort.High, Effort.XHigh, Effort.Max];
 /** Pre-5.6 Devin GPT families top out at `-xhigh`: four wire tiers, `low` floor. */
 const DEVIN_FOUR_TIER_EFFORTS: readonly Effort[] = [Effort.Low, Effort.Medium, Effort.High, Effort.XHigh];
 
-function devinTierFamily(
-	id: string,
-	name: string,
-	routes: DevinTierRoutes,
-	efforts: readonly Effort[],
-): EffortVariantFamily {
+/**
+ * Build one effort-tier family from a tier→wire-id map: routing keeps only the
+ * listed efforts, `members` dedupes the targets in tier order, and thinking is
+ * `mode: "effort"` (mandatory when the family has no `off` route). Shared by the
+ * Devin and Cursor tables, whose per-effort siblings follow the same shape.
+ */
+function tierFamily(id: string, name: string, routes: TierRoutes, efforts: readonly Effort[]): EffortVariantFamily {
 	const routing: Partial<Record<Effort | "off", string>> = {};
 	if (routes.off) routing.off = routes.off;
 	for (const effort of efforts) {
@@ -177,7 +178,7 @@ function devinTierFamily(
 function devinGpt56Families(variant: "luna" | "sol" | "terra", name: string): readonly EffortVariantFamily[] {
 	const base = `gpt-5-6-${variant}`;
 	return [
-		devinTierFamily(
+		tierFamily(
 			base,
 			name,
 			{
@@ -190,7 +191,7 @@ function devinGpt56Families(variant: "luna" | "sol" | "terra", name: string): re
 			},
 			DEVIN_FIVE_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			`${base}-fast`,
 			`${name} Fast`,
 			{
@@ -395,7 +396,7 @@ export const GEMINI_CLI_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 };
 export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 	families: [
-		devinTierFamily(
+		tierFamily(
 			"claude-opus-5",
 			"Claude Opus 5",
 			{
@@ -407,7 +408,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			DEVIN_FIVE_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			"claude-opus-5-fast",
 			"Claude Opus 5 Fast",
 			{
@@ -419,7 +420,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			DEVIN_FIVE_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			"claude-fable-5",
 			"Claude Fable 5",
 			{
@@ -431,7 +432,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			DEVIN_FIVE_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			"claude-sonnet-5",
 			"Claude Sonnet 5",
 			{
@@ -443,7 +444,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			DEVIN_FIVE_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			"claude-opus-4-7",
 			"Claude Opus 4.7",
 			{
@@ -455,7 +456,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			DEVIN_FIVE_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			"claude-opus-4-7-fast",
 			"Claude Opus 4.7 Fast",
 			{
@@ -467,7 +468,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			DEVIN_FIVE_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			"claude-opus-4-8",
 			"Claude Opus 4.8",
 			{
@@ -479,7 +480,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			DEVIN_FIVE_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			"claude-opus-4-8-fast",
 			"Claude Opus 4.8 Fast",
 			{
@@ -491,7 +492,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			DEVIN_FIVE_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			"gpt-5-2",
 			"GPT-5.2",
 			{
@@ -503,7 +504,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			DEVIN_FOUR_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			"gpt-5-3-codex",
 			"GPT-5.3 Codex",
 			{
@@ -514,7 +515,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			DEVIN_FOUR_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			"gpt-5-3-codex-fast",
 			"GPT-5.3 Codex Fast",
 			{
@@ -525,7 +526,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			DEVIN_FOUR_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			"gpt-5-4",
 			"GPT-5.4",
 			{
@@ -537,7 +538,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			DEVIN_FOUR_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			"gpt-5-4-fast",
 			"GPT-5.4 Fast",
 			{
@@ -549,7 +550,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			DEVIN_FOUR_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			"gpt-5-4-mini",
 			"GPT-5.4 Mini",
 			{
@@ -560,7 +561,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			DEVIN_FOUR_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			"gpt-5-5",
 			"GPT-5.5",
 			{
@@ -572,7 +573,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			DEVIN_FOUR_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			"gpt-5-5-fast",
 			"GPT-5.5 Fast",
 			{
@@ -587,7 +588,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 		...devinGpt56Families("luna", "GPT-5.6 Luna"),
 		...devinGpt56Families("sol", "GPT-5.6 Sol"),
 		...devinGpt56Families("terra", "GPT-5.6 Terra"),
-		devinTierFamily(
+		tierFamily(
 			"kimi-k3",
 			"Kimi K3",
 			{
@@ -597,7 +598,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			[Effort.Low, Effort.High, Effort.Max],
 		),
-		devinTierFamily(
+		tierFamily(
 			"swe-1-7",
 			"SWE-1.7",
 			{
@@ -606,7 +607,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			[Effort.Medium, Effort.Max],
 		),
-		devinTierFamily(
+		tierFamily(
 			"grok-4-5",
 			"Grok 4.5",
 			{
@@ -616,7 +617,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			[Effort.Low, Effort.Medium, Effort.High],
 		),
-		devinTierFamily(
+		tierFamily(
 			"inkling",
 			"Inkling",
 			{
@@ -629,7 +630,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			DEVIN_FIVE_TIER_EFFORTS,
 		),
-		devinTierFamily(
+		tierFamily(
 			"gemini-3-1-pro",
 			"Gemini 3.1 Pro",
 			{
@@ -638,7 +639,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			[Effort.Low, Effort.High],
 		),
-		devinTierFamily(
+		tierFamily(
 			"gemini-3-5-flash",
 			"Gemini 3.5 Flash",
 			{
@@ -649,7 +650,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			[Effort.Minimal, Effort.Low, Effort.Medium, Effort.High],
 		),
-		devinTierFamily(
+		tierFamily(
 			"gemini-3-6-flash",
 			"Gemini 3.6 Flash",
 			{
@@ -660,7 +661,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 			},
 			[Effort.Minimal, Effort.Low, Effort.Medium, Effort.High],
 		),
-		devinTierFamily(
+		tierFamily(
 			"gemini-3-flash",
 			"Gemini 3 Flash",
 			{
@@ -694,7 +695,7 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 		},
 		// GLM-5.2 1M — paid variants that consume weekly quota.  Collapse the
 		// three 1M-context variants into one entry with proper effort routing.
-		devinTierFamily(
+		tierFamily(
 			"glm-5-2-1m",
 			"GLM-5.2 1M",
 			{
@@ -707,11 +708,43 @@ export const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
 	],
 };
 
+/** Cursor's Grok tier tokens equal the effort id, so routing is a direct map. */
+const CURSOR_GROK_45_EFFORTS: readonly Effort[] = [Effort.Low, Effort.Medium, Effort.High];
+const CURSOR_GROK_46_EFFORTS: readonly Effort[] = [Effort.Low, Effort.Medium, Effort.High, Effort.XHigh];
+
+/**
+ * Cursor serves Grok 4.5/4.6 as per-effort sibling ids
+ * (`cursor-grok-4.6-low|-medium|-high|-xhigh`) alongside a parallel `-fast`
+ * service-tier lane (`-low-fast`, …). Each lane collapses into its own logical
+ * model — the `-fast` axis is a sibling family, never a second routing
+ * dimension — with effort routing onto the live sibling wire ids.
+ */
+function cursorGrokFamilies(version: "4.5" | "4.6", efforts: readonly Effort[]): readonly EffortVariantFamily[] {
+	const build = (fast: boolean): EffortVariantFamily => {
+		const suffix = fast ? "-fast" : "";
+		const routes: TierRoutes = {};
+		for (const effort of efforts) {
+			routes[effort] = `cursor-grok-${version}-${effort}${suffix}`;
+		}
+		return tierFamily(`cursor-grok-${version}${suffix}`, `Grok ${version}${fast ? " Fast" : ""}`, routes, efforts);
+	};
+	return [build(false), build(true)];
+}
+
+/** `cursor` Grok families: per-effort siblings collapsed per service-tier lane. */
+export const CURSOR_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
+	families: [
+		...cursorGrokFamilies("4.5", CURSOR_GROK_45_EFFORTS),
+		...cursorGrokFamilies("4.6", CURSOR_GROK_46_EFFORTS),
+	],
+};
+
 /** Provider id → hand collapse table. The CCA providers diverge on thinking transport. */
 export const VARIANT_COLLAPSE_TABLES: Readonly<Record<string, VariantCollapseTable>> = {
 	"google-antigravity": ANTIGRAVITY_VARIANT_COLLAPSE_TABLE,
 	"google-gemini-cli": GEMINI_CLI_VARIANT_COLLAPSE_TABLE,
 	devin: DEVIN_VARIANT_COLLAPSE_TABLE,
+	cursor: CURSOR_VARIANT_COLLAPSE_TABLE,
 };
 
 /**