@linxiraos/pi-catalog 1.2.0 → 1.3.0

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

@@ -39,8 +39,17 @@ import {
 	type OpenAICompat,
 	type Provider,
 	type ThinkingConfig,
+	type TokenCost,
 } from "../types";
-import { discoveryFetch, isAnthropicOAuthToken, isRecord, toBoolean, toNumber, toPositiveNumber } from "../utils";
+import {
+	discoveryFetch,
+	isAnthropicOAuthToken,
+	isRecord,
+	toBoolean,
+	toNumber,
+	toPositiveNumber,
+	toPositiveNumberOrNull,
+} from "../utils";
 import { ALIBABA_TOKEN_PLAN_BASE_URL, parseAlibabaTokenPlanCredential } from "../wire/alibaba-token-plan";
 import { normalizeCharmHyperBaseUrl } from "../wire/charm-hyper";
 import { CLINEPASS_API_BASE_URL, clinePassClientHeaders } from "../wire/cline-pass";
@@ -550,6 +559,40 @@ function mapWithBundledReference<TApi extends Api>(
 	};
 }
 
+/**
+ * Rates from an OpenRouter-shaped `pricing` object (USD per token, as decimal
+ * strings). Missing, zero, or negative legs (`-1` marks router-priced rows)
+ * map to 0, which the model manager fills from its reference sources.
+ */
+function perTokenPricingCost(pricing: unknown): TokenCost {
+	const rates = isRecord(pricing) ? pricing : {};
+	return {
+		input: toPositiveNumber(rates.prompt, 0) * 1_000_000,
+		output: toPositiveNumber(rates.completion, 0) * 1_000_000,
+		cacheRead: toPositiveNumber(rates.input_cache_read, 0) * 1_000_000,
+		cacheWrite: toPositiveNumber(rates.input_cache_write, 0) * 1_000_000,
+	};
+}
+
+/**
+ * Rates from NanoGPT's detailed roster: prompt and completion are USD per
+ * million tokens, the cache legs per thousand. Legs the row does not price
+ * keep `fallback`.
+ */
+function nanoGptPricingCost(pricing: unknown, fallback: TokenCost): TokenCost {
+	if (!isRecord(pricing) || pricing.unit !== "per_million_tokens") return fallback;
+	const perThousand = (value: unknown, fallbackRate: number): number => {
+		const rate = toPositiveNumber(value, 0);
+		return rate > 0 ? rate * 1_000 : fallbackRate;
+	};
+	return {
+		input: toPositiveNumber(pricing.prompt, fallback.input),
+		output: toPositiveNumber(pricing.completion, fallback.output),
+		cacheRead: perThousand(pricing.cacheReadInputPer1kTokens, fallback.cacheRead),
+		cacheWrite: perThousand(pricing.cacheWriteInputPer1kTokens, fallback.cacheWrite),
+	};
+}
+
 function normalizeAnthropicBaseUrl(baseUrl: string | undefined, fallback: string): string {
 	const value = baseUrl?.trim();
 	if (!value) {
@@ -1265,6 +1308,87 @@ export function cerebrasModelManagerOptions(
 	});
 }
 
+// ---------------------------------------------------------------------------
+// 3.5 SambaNova
+// ---------------------------------------------------------------------------
+
+/** SambaNova discovery inputs: `apiKey` enables the live roster; `baseUrl` defaults to SambaNova Cloud. */
+export interface SambaNovaModelManagerConfig {
+	apiKey?: string;
+	baseUrl?: string;
+	fetch?: FetchImpl;
+}
+
+/**
+ * One SambaNova per-token USD rate (a decimal string) as a per-million rate.
+ * `absent` is the value for a missing field; a present field that is not a
+ * nonnegative number yields `undefined`.
+ */
+function sambanovaRate(value: unknown, absent: number | undefined): number | undefined {
+	if (value === undefined || value === null) return absent;
+	const parsed = toNumber(value);
+	return parsed !== undefined && parsed >= 0 ? Math.round(parsed * 1e12) / 1e6 : undefined;
+}
+
+/**
+ * Map SambaNova's `pricing` block. Prompt and completion rates are required;
+ * cache rates default to 0 when absent. An unreadable card stays unpriced
+ * (`unknown`) instead of reading as free, and an all-zero card is a declared
+ * free tier.
+ */
+function sambanovaPricing(
+	pricing: unknown,
+	fallback: TokenCost,
+): Pick<ModelSpec<"openai-completions">, "cost" | "pricingStatus"> {
+	const rates = isRecord(pricing) ? pricing : {};
+	const input = sambanovaRate(rates.prompt, undefined);
+	const output = sambanovaRate(rates.completion, undefined);
+	const cacheRead = sambanovaRate(rates.input_cache_read, 0);
+	const cacheWrite = sambanovaRate(rates.input_cache_write, 0);
+	if (input === undefined || output === undefined || cacheRead === undefined || cacheWrite === undefined) {
+		return { cost: fallback, pricingStatus: "unknown" };
+	}
+	const cost = { input, output, cacheRead, cacheWrite };
+	return input + output + cacheRead + cacheWrite === 0 ? { cost, pricingStatus: "free" } : { cost };
+}
+
+/**
+ * SambaNova Cloud's model manager. `/v1/models` requires a key, so a keyless
+ * manager serves only the bundled seed; with a key the live roster replaces
+ * it. Rows carry pricing and limits but no capability metadata: reasoning,
+ * thinking wire format, and image input come from `providers/sambanova.kdl`.
+ */
+export function sambanovaModelManagerOptions(
+	config?: SambaNovaModelManagerConfig,
+): ModelManagerOptions<"openai-completions"> {
+	const apiKey = config?.apiKey;
+	const baseUrl = config?.baseUrl ?? "https://api.sambanova.ai/v1";
+	return {
+		providerId: "sambanova",
+		dynamicModelsAuthoritative: true,
+		// The live `pricing` block is the tariff: an explicit `0` must not revert
+		// to the bundled seed's rate.
+		dynamicCostAuthoritative: true,
+		...(apiKey && {
+			fetchDynamicModels: () =>
+				fetchOpenAICompatibleModels({
+					api: "openai-completions",
+					provider: "sambanova",
+					baseUrl,
+					apiKey,
+					filterModel: (_entry, model) => !isExcludedModel("sambanova", model.id),
+					mapModel: (entry, defaults) => ({
+						...defaults,
+						...sambanovaPricing(entry.pricing, defaults.cost),
+						contextWindow: toPositiveNumber(entry.context_length, defaults.contextWindow),
+						maxTokens: toPositiveNumber(entry.max_completion_tokens, defaults.maxTokens),
+					}),
+					fetch: config?.fetch,
+				}),
+		}),
+	};
+}
+
 // ---------------------------------------------------------------------------
 // 4. Hugging Face
 // ---------------------------------------------------------------------------
@@ -2296,74 +2420,6 @@ export function zhipuCodingPlanModelManagerOptions(
 // 7.5 Fireworks
 // ---------------------------------------------------------------------------
 
-/**
- * Fireworks-published cap for the Kimi K2 family. Fireworks' `/v1/models`
- * envelope generically reports `max_completion_tokens: 65536` for every Kimi
- * deployment, but Kimi K2 (instruct / thinking / turbo) on Fireworks is
- * documented to ship long reasoning traces that should be bounded — capping
- * at 32,768 prevents handing callers a budget the router cannot honor.
- * See https://github.com/can1357/oh-my-pi/issues/1849.
- */
-export const FIREWORKS_KIMI_MAX_TOKENS = 32_768;
-
-/**
- * Fireworks' output ceiling for Kimi K2.7-Code specifically. Its `/v1/models`
- * generic `max_completion_tokens` is 65,536 and Fireworks serves it in full —
- * verified with a single completion emitting 58,971 output tokens and
- * `max_tokens: 200000` accepted without error. Unlike the older K2.5/K2.6
- * family (see {@link FIREWORKS_KIMI_MAX_TOKENS}), K2.7-Code is not clamped to
- * 32,768; that ceiling only truncated it.
- */
-export const FIREWORKS_KIMI_K27_CODE_MAX_TOKENS = 65_536;
-
-/**
- * Returns true for the Kimi K2.5 / K2.6 family served by Fireworks-backed
- * providers (`fireworks` direct, `firepass` router) that share the 32,768
- * `maxTokens` ceiling. Matches both the public catalog id (`kimi-k2.5`,
- * `kimi-k2.6`, `kimi-k2.6-turbo`) and the canonical Fireworks wire id
- * (`accounts/fireworks/{models,routers}/kimi-k2…`).
- *
- * K2.7-Code (incl. `-fast` / `-highspeed`) is deliberately excluded: unlike the
- * earlier K2 family it serves its full context on Fireworks — verified with a
- * single completion emitting 58,971 output tokens and `max_tokens: 200000`
- * accepted without error — so the 32,768 cap would only truncate it. It inherits
- * Fireworks' reported `max_completion_tokens` (65,536) instead.
- */
-export function isFireworksKimiK2ModelId(modelId: string): boolean {
-	const identity = classifyModel("fireworks", modelId, { lenient: true });
-	if (identity.class !== "kimi" || identity.family === "k2.7-code") return false;
-	return identity.family?.startsWith("k2") === true;
-}
-
-/**
- * Clamp the Kimi K2 family's `maxTokens` to {@link FIREWORKS_KIMI_MAX_TOKENS}
- * on Fireworks-backed providers, leaving every other model untouched.
- */
-export function clampFireworksKimiMaxTokens(modelId: string, candidate: number): number;
-export function clampFireworksKimiMaxTokens(modelId: string, candidate: number | null): number | null;
-export function clampFireworksKimiMaxTokens(modelId: string, candidate: number | null): number | null {
-	if (candidate === null) return null;
-	return isFireworksKimiK2ModelId(modelId) ? Math.min(candidate, FIREWORKS_KIMI_MAX_TOKENS) : candidate;
-}
-
-/**
- * Kimi K2.7 Code's documented recommended output budget. Some provider
- * discovery rows report the context-sized `max_completion_tokens` instead.
- */
-export const KIMI_K27_CODE_RECOMMENDED_MAX_TOKENS = 32_768;
-
-export function isKimiK27CodeModelId(modelId: string): boolean {
-	const identity = classifyModel("fireworks", modelId, { lenient: true });
-	return identity.class === "kimi" && identity.family === "k2.7-code";
-}
-
-export function clampKimiK27CodeMaxTokens(modelId: string, candidate: number): number;
-export function clampKimiK27CodeMaxTokens(modelId: string, candidate: number | null): number | null;
-export function clampKimiK27CodeMaxTokens(modelId: string, candidate: number | null): number | null {
-	if (candidate === null) return null;
-	return isKimiK27CodeModelId(modelId) ? Math.min(candidate, KIMI_K27_CODE_RECOMMENDED_MAX_TOKENS) : candidate;
-}
-
 /**
  * Fireworks Fast variants we surface: the Fast table on
  * https://docs.fireworks.ai/serverless/serverless-modes, minus the US-only
@@ -2493,15 +2549,9 @@ function mapFireworksControlPlaneModel(
 	const supportsImage = toBoolean(record.supportsImageInput) === true;
 	const supportsTools = toBoolean(record.supportsTools);
 	const contextWindow = toPositiveNumber(record.contextLength, reference?.contextWindow ?? null);
-	// The control plane reports no max-output budget. Default K2.7-Code to its
-	// verified 65,536 ceiling, the older K2.5/K2.6 family to the clamped 32,768,
-	// everyone else to the discovery fallback, then clamp.
-	const fallbackMaxTokens = isKimiK27CodeModelId(publicModelId)
-		? FIREWORKS_KIMI_K27_CODE_MAX_TOKENS
-		: isFireworksKimiK2ModelId(publicModelId)
-			? FIREWORKS_KIMI_MAX_TOKENS
-			: null;
-	const maxTokens = clampFireworksKimiMaxTokens(publicModelId, reference?.maxTokens ?? fallbackMaxTokens);
+	// The control plane reports no max-output budget; Fireworks' per-model
+	// output caps and fallbacks are rule-owned (providers/fireworks.kdl).
+	const maxTokens = reference?.maxTokens ?? null;
 	const base: ModelSpec<"openai-completions"> = reference ?? {
 		id: publicModelId,
 		name,
@@ -2941,8 +2991,9 @@ function buildClinePassFreeModel(id: string, liveCatalog?: ClinePassLiveCatalog)
 		input: upstream?.input ?? live?.input ?? fallback.input,
 		contextWindow: upstream?.contextWindow ?? live?.contextWindow ?? fallback.contextWindow,
 		maxTokens: upstream?.maxTokens ?? live?.maxTokens ?? fallback.maxTokens,
-		// A non-reasoning reference must not carry the Cline effort ladder.
-		...(reasoning ? {} : { thinking: undefined }),
+		// No reasoning-effort control means no effort ladder: the cline-pass
+		// rules trust only explicit thinking, so a copied ladder would surface.
+		thinking: undefined,
 		compat: { supportsReasoningEffort: false, wireModelIdMode: "raw" },
 	};
 }
@@ -3215,12 +3266,18 @@ export interface OpenCodeModelManagerConfig {
 	fetch?: FetchImpl;
 }
 
+const OPENCODE_ZEN_BASE_PATH = "https://opencode.ai/zen";
+const OPENCODE_ZEN_GOOGLE_BASE_URL = "https://opencode.ai/inference/google/v1beta";
+
 function normalizeOpenCodeBasePath(baseUrl: string | undefined, fallbackBasePath: string): string {
 	const value = normalizeAnthropicBaseUrl(baseUrl, fallbackBasePath);
 	return value.endsWith("/v1") ? value.slice(0, -3) : value;
 }
 
 function openCodeBaseUrlForApi(api: Api, basePath: string): string {
+	if (api === "google-generative-ai" && basePath === OPENCODE_ZEN_BASE_PATH) {
+		return OPENCODE_ZEN_GOOGLE_BASE_URL;
+	}
 	return api === "anthropic-messages" ? basePath : `${basePath}/v1`;
 }
 
@@ -3234,7 +3291,12 @@ function openCodeBaseUrlForApi(api: Api, basePath: string): string {
 // rules (`runtime/behavior.kdl`; #887, #1617, #8957).
 // Runtime-discovered rows cached before model-identity corrections retain
 // stale capability metadata until the authoritative catalog TTL expires.
-const OPENCODE_CACHE_MIGRATION_MODEL_IDS = ["glm-5.3-flash", "longcat-2.5-preview-free", "space-bunny-free"] as const;
+const OPENCODE_CACHE_MIGRATION_MODEL_IDS = [
+	"deepseek-flash",
+	"glm-5.3-flash",
+	"longcat-2.5-preview-free",
+	"space-bunny-free",
+] as const;
 const OPENCODE_ZEN_CACHE_MIGRATION_MODEL_IDS = ["gemini-3.7-flash", "gemini-3.8-flash"] as const;
 
 // Billing-variant suffixes the OpenCode gateways append to a base model id
@@ -3269,7 +3331,10 @@ function openCodeModelManagerOptions(
 	const apiOverrides = (id: string): Api | undefined => {
 		// Route pins live in the `api-routes` behavior rules.
 		const route = apiRouteFor(providerId, id)?.api;
-		return route === "openai-completions" || route === "openai-responses" || route === "anthropic-messages"
+		return route === "openai-completions" ||
+			route === "openai-responses" ||
+			route === "anthropic-messages" ||
+			route === "google-generative-ai"
 			? route
 			: undefined;
 	};
@@ -3893,6 +3958,7 @@ export function kiloModelManagerOptions(config?: KiloModelManagerConfig): ModelM
 				provider: "kilo",
 				baseUrl,
 				apiKey,
+				mapModel: (entry, defaults) => ({ ...defaults, cost: perTokenPricingCost(entry.pricing) }),
 				fetch: config?.fetch,
 			}),
 	};
@@ -4570,11 +4636,7 @@ export function veniceModelManagerOptions(
 		config,
 		mapModel: (entry, defaults, reference) => {
 			const model = mapWithBundledReference(entry, defaults, reference);
-			return {
-				...model,
-				maxTokens: clampKimiK27CodeMaxTokens(defaults.id, model.maxTokens),
-				compat: { ...model.compat, supportsUsageInStreaming: false },
-			};
+			return { ...model, compat: { ...model.compat, supportsUsageInStreaming: false } };
 		},
 	});
 }
@@ -4727,7 +4789,9 @@ function museSparkLineageSpec(id: string): ModelSpec<"openai-responses"> | undef
 	let templateRevision: readonly [number, number, number] | undefined;
 	for (const model of seedModels<"openai-responses">("meta")) {
 		if ((billingVariantPlain(model.id) !== undefined) !== contributor) continue;
-		const candidate = parseRevision(classifyModel("meta", model.id, { lenient: true }).revision ?? "");
+		const seedIdentity = classifyModel("meta", model.id, { lenient: true });
+		if (seedIdentity.family !== "muse-spark") continue;
+		const candidate = parseRevision(seedIdentity.revision ?? "");
 		if (candidate === undefined) continue;
 		if (templateRevision === undefined || compareRevision(candidate, templateRevision) > 0) {
 			template = model;
@@ -6282,10 +6346,29 @@ export function nanoGptModelManagerOptions(
 					provider: "nanogpt",
 					baseUrl,
 					apiKey,
+					// Only the detailed roster carries per-model pricing, names, and limits.
+					query: { detailed: "true" },
 					mapModel: (entry, defaults) => {
 						const reference = resolveReference(defaults.id);
-						const mapped = mapWithBundledReference(entry, defaults, reference);
-						return { ...mapped, api: "openai-completions", provider: "nanogpt" };
+						// The detailed roster names its output cap `max_output_tokens`.
+						const maxOutputTokens = entry.max_output_tokens ?? entry.max_completion_tokens;
+						const mapped = mapWithBundledReference(
+							{ ...entry, max_completion_tokens: maxOutputTokens },
+							reference
+								? defaults
+								: {
+										...defaults,
+										contextWindow: toPositiveNumberOrNull(entry.context_length),
+										maxTokens: toPositiveNumberOrNull(maxOutputTokens),
+									},
+							reference,
+						);
+						return {
+							...mapped,
+							api: "openai-completions",
+							provider: "nanogpt",
+							cost: { ...mapped.cost, ...nanoGptPricingCost(entry.pricing, mapped.cost) },
+						};
 					},
 					filterModel: (_entry, model) => {
 						const match = NANO_GPT_THINKING_SUFFIX_RE.exec(model.id);
@@ -6969,13 +7052,10 @@ function createOpenCodeApiResolution(
 	rules: ApiResolutionRule[];
 } {
 	const completionsBaseUrl = `${basePath}/v1`;
-	// Per-API base URLs on the OpenCode-style endpoint:
-	// - openai-completions / openai-responses / google-generative-ai → /v1
-	// - anthropic-messages → bare basePath (the Anthropic client appends /v1/messages)
-	const baseUrlForApi = (api: Api): string => (api === "anthropic-messages" ? basePath : completionsBaseUrl);
+	const baseUrlForApi = (api: Api): string => openCodeBaseUrlForApi(api, basePath);
 	// Per-id route pins live in the `api-routes` behavior rules.
 	const overrideRules: ApiResolutionRule[] = (
-		["openai-completions", "openai-responses", "anthropic-messages"] as const
+		["openai-completions", "openai-responses", "anthropic-messages", "google-generative-ai"] as const
 	).map(api => ({
 		matches: modelId => apiRouteFor(providerId, modelId)?.api === api,
 		resolved: { api, baseUrl: baseUrlForApi(api) },
@@ -7004,8 +7084,17 @@ function createOpenCodeApiResolution(
 
 // Resolver rules for the models.dev descriptor path; the per-id pins live in
 // the `api-routes` behavior rules, which also drive the live-discovery mapper.
-const OPENCODE_ZEN_API_RESOLUTION = createOpenCodeApiResolution("https://opencode.ai/zen", "opencode-zen");
-const OPENCODE_GO_API_RESOLUTION = createOpenCodeApiResolution("https://opencode.ai/zen/go", "opencode-go");
+const OPENCODE_ZEN_API_RESOLUTION = createOpenCodeApiResolution(OPENCODE_ZEN_BASE_PATH, "opencode-zen");
+const OPENCODE_GO_BASE_PATH = "https://opencode.ai/inference/go";
+const OPENCODE_GO_API_RESOLUTION = createOpenCodeApiResolution(OPENCODE_GO_BASE_PATH, "opencode-go");
+
+/**
+ * OpenCode Go endpoint for one request API. The model generator re-bases
+ * previous-snapshot rows with it so retained rows follow endpoint moves.
+ */
+export function openCodeGoBaseUrlForApi(api: Api): string {
+	return openCodeBaseUrlForApi(api, OPENCODE_GO_BASE_PATH);
+}
 
 const COPILOT_BASE_URL = "https://api.githubcopilot.com";
 
@@ -7383,7 +7472,7 @@ const OPENCODE_MODELS_DEV_DESCRIPTORS: readonly ModelsDevProviderDescriptor[] =
 				OPENCODE_ZEN_API_RESOLUTION.defaultResolution,
 			),
 	}),
-	openAiCompletionsDescriptor("opencode-go", "opencode-go", "https://opencode.ai/zen/go/v1", {
+	openAiCompletionsDescriptor("opencode-go", "opencode-go", `${OPENCODE_GO_BASE_PATH}/v1`, {
 		filterModel: filterActiveToolCallModels,
 		resolveApi: (modelId, raw) =>
 			resolveApiByRules(
@@ -7464,12 +7553,7 @@ const MODELS_DEV_PROVIDER_DESCRIPTORS_SPECIALIZED: readonly ModelsDevProviderDes
 	// --- Synthetic ---
 	openAiCompletionsDescriptor("synthetic", "synthetic", "https://api.synthetic.new/openai/v1"),
 	// --- Venice AI ---
-	openAiCompletionsDescriptor("venice", "venice", "https://api.venice.ai/api/v1", {
-		transformModel: model => {
-			const maxTokens = clampKimiK27CodeMaxTokens(model.id, model.maxTokens);
-			return maxTokens === model.maxTokens ? model : { ...model, maxTokens };
-		},
-	}),
+	openAiCompletionsDescriptor("venice", "venice", "https://api.venice.ai/api/v1"),
 	// --- Ollama Cloud ---
 	simpleModelsDevDescriptor("ollama-cloud", "ollama-cloud", "ollama-chat", "https://ollama.com"),
 	// --- Xiaomi Token Plan ---

package/src/provider-models/special.ts

@@ -64,6 +64,7 @@ export function openaiCodexModelManagerOptions(
 		providerId: "openai-codex",
 		cacheProviderId: resolveModelCacheProviderId("openai-codex", { baseUrl }),
 		dynamicModelsAuthoritative: providerEntry("openai-codex")?.dynamicModelsAuthoritative === true,
+		keepStaticNames: providerEntry("openai-codex")?.keepBundledNames === true,
 		...(resolveAccounts
 			? {
 					fetchDynamicModels: async () => {

package/src/types.d.ts

@@ -93,6 +93,14 @@ export interface ThinkingConfig {
      * provides an explicit wire off-path.
      */
     requiresEffort?: boolean;
+    /**
+     * The catalog has no evidence this model reasons (unclassified id, no
+     * capability metadata). The ladder is offered for the user to opt into:
+     * sessions leave thinking unset until the user picks an effort, and a
+     * request without an effort goes out as non-reasoning (see
+     * `withoutUnverifiedThinking`).
+     */
+    unverified?: boolean;
 }
 export type Provider = string;
 /** Token budgets for each thinking level (token-based providers only) */
@@ -443,6 +451,8 @@ export interface OpenAICompat {
     reasoningDeltasMayBeCumulative?: boolean;
     /** Strip leaked DeepSeek chat-template special tokens from visible content deltas. Default: auto-detected. */
     stripDeepseekSpecialTokens?: boolean;
+    /** Drop assistant thinking blocks from replayed chat-completions history. Default: false. */
+    stripThinkingHistory?: boolean;
     /** Heal leaked chat-template/tool-call/thinking markup from visible content deltas. Default: auto-detected. */
     streamMarkupHealingPattern?: OpenAIStreamMarkupHealingPattern;
     /** Whether this wire may revise already-streamed text (`stream-revision` axis). Unassigned: append-only. */
@@ -492,6 +502,14 @@ export interface AnthropicCompat {
      * contract instead. Default: false.
      */
     firstPartyProvider?: boolean;
+    /**
+     * Whether the provider honors Claude Code's `CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS`
+     * kill switch, which withholds pre-release fields (`context_management`, beta
+     * tool fields, task budgets, structured-output format, `thinking.block_binding`,
+     * server compaction) with their betas. Rule-owned on the
+     * provider whose `ANTHROPIC_*` env routing Claude Code shares. Default: false.
+     */
+    honorsClaudeCodeBetaKillSwitch?: boolean;
     /**
      * Whether requests may carry `output_config.effort` (and its effort beta
      * header). Vertex AI rejects the field/header. Default: true.
@@ -858,12 +876,15 @@ export interface ResolvedOpenAISharedCompat {
  * `buildModel`; request handlers read fields and never detect, resolve, or
  * allocate.
  */
-export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "trustExplicitThinkingOnly" | "reasoningEffortMap" | "supportsReasoningParams" | "supportsReasoningSummary" | "statefulResponses" | "supportsSamplingParams" | "supportsPenaltyAndStopParams" | "thinkingFormat" | "kimiApiFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "disableReasoningWithTools" | "supportsToolChoice" | "supportsForcedToolChoice" | "supportsNamedToolChoice" | "reasoningContentField" | "mistralReasoningContentParts" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "syntheticReasoningContentFallback" | "replayReasoningContent" | "qwenPreserveThinking" | "qwenTemplateReasoningEffort" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "sessionHeader" | "supportsPromptCacheBreakpoints" | "promptCacheBreakpointTtl" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "clampOutputToModelMax" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "streamRevision" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "supportsConfigurationUpdate" | "supportsSteering" | "stripImageInput" | "thinkingLoopGuard" | "whenThinking">> & {
+export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "trustExplicitThinkingOnly" | "reasoningEffortMap" | "supportsReasoningParams" | "supportsReasoningSummary" | "statefulResponses" | "supportsSamplingParams" | "supportsPenaltyAndStopParams" | "thinkingFormat" | "kimiApiFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "disableReasoningWithTools" | "supportsToolChoice" | "supportsForcedToolChoice" | "supportsNamedToolChoice" | "reasoningContentField" | "mistralReasoningContentParts" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "syntheticReasoningContentFallback" | "replayReasoningContent" | "qwenPreserveThinking" | "qwenTemplateReasoningEffort" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "stripThinkingHistory" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "sessionHeader" | "supportsPromptCacheBreakpoints" | "promptCacheBreakpointTtl" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "clampOutputToModelMax" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "streamRevision" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "supportsConfigurationUpdate" | "supportsSteering" | "stripImageInput" | "thinkingLoopGuard" | "whenThinking">> & {
     vercelGatewayRouting?: OpenAICompat["vercelGatewayRouting"];
     extraBody?: OpenAICompat["extraBody"];
     cacheControlFormat?: OpenAICompat["cacheControlFormat"];
     thinkingKeep?: OpenAICompat["thinkingKeep"];
     streamIdleTimeoutMs?: number;
+    stripThinkingHistory?: boolean;
+    /** The configured `baseUrl` is the official OpenAI API (provider `openai`); unset otherwise. */
+    officialEndpoint?: boolean;
     toolStrictMode: ResolvedToolStrictMode;
     /** The model sits behind Vercel AI Gateway. */
     isVercelGatewayHost: boolean;
@@ -928,9 +949,11 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
     isVercelGatewayHost: boolean;
     /**
      * The configured endpoint is first-party OpenAI (`provider === "openai"` on
-     * an `api.openai.com` or unset baseUrl). Gates official-only Responses
-     * behavior: default-on stateful `previous_response_id` chaining and the
-     * `text.verbosity` field.
+     * an `api.openai.com` or unset baseUrl), or the Codex backend
+     * (`openai-codex-responses` on the default Codex baseUrl). Gates platform
+     * OpenAI's official-only Responses behavior (default-on stateful
+     * `previous_response_id` chaining and the `text.verbosity` field; the Codex
+     * adapter is not governed by either gate) and host request-size budgets.
      */
     officialEndpoint: boolean;
     /** Run Harmony-protocol leak detection/mitigation on streamed output. */
@@ -1055,23 +1078,43 @@ export interface GoogleCompat {
     antigravityUsageLabel?: string;
     /** Strip image inputs before encoding (text-only serving of a multimodal id). */
     stripImageInput?: boolean;
+    /**
+     * Largest width or height a single image may carry for a Claude reader on
+     * this host (see {@link AnthropicCompat.maxImageDimension}). Setting this or
+     * {@link maxImagePayloadBytes} turns on Claude's image clamp, many-image
+     * cap included; unset leaves images as sent.
+     */
+    maxImageDimension?: number;
+    /**
+     * Largest base64 payload a single image may carry, in bytes (see
+     * {@link AnthropicCompat.maxImagePayloadBytes}).
+     */
+    maxImagePayloadBytes?: number;
     /** Thinking-loop watchdog guard family applied to streamed reasoning. */
     thinkingLoopGuard?: "gemini" | "deepseek" | "xai";
     /** See {@link BedrockCompat.supportsSamplingParams}. */
     supportsSamplingParams?: boolean;
 }
 /** Fully-resolved google-API compat view, materialized once by `buildModel`. */
-export type ResolvedGoogleCompat = Required<Omit<GoogleCompat, "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "thinkingLoopGuard" | "antigravityUsageLabel" | "supportsSamplingParams">> & {
+export type ResolvedGoogleCompat = Required<Omit<GoogleCompat, "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "thinkingLoopGuard" | "antigravityUsageLabel" | "supportsSamplingParams" | "maxImageDimension" | "maxImagePayloadBytes">> & {
     streamFirstEventTimeoutMs?: number;
     streamIdleTimeoutMs?: number;
     thinkingLoopGuard?: GoogleCompat["thinkingLoopGuard"];
     antigravityUsageLabel?: string;
     supportsSamplingParams?: boolean;
+    maxImageDimension?: number;
+    maxImagePayloadBytes?: number;
 };
+/** Ollama chat history compatibility resolved by the catalog. */
+export interface OllamaCompat {
+    /** Omit assistant thinking when the deployment rejects it in history. */
+    stripThinkingHistory?: boolean;
+}
+export type ResolvedOllamaCompat = Required<OllamaCompat>;
 /** Sparse, user-authored compat overrides for a given API (models.json / config vocabulary). */
-export type CompatConfigOf<TApi extends Api> = TApi extends "openai-completions" | "openrouter" | "openai-responses" | "azure-openai-responses" | "openai-codex-responses" ? OpenAICompat : TApi extends "anthropic-messages" ? AnthropicCompat : TApi extends "bedrock-converse-stream" ? BedrockCompat : TApi extends "devin-agent" ? DevinCompat : TApi extends "google-generative-ai" | "google-vertex" | "google-gemini-cli" ? GoogleCompat : undefined;
+export type CompatConfigOf<TApi extends Api> = TApi extends "openai-completions" | "openrouter" | "openai-responses" | "azure-openai-responses" | "openai-codex-responses" ? OpenAICompat : TApi extends "anthropic-messages" ? AnthropicCompat : TApi extends "bedrock-converse-stream" ? BedrockCompat : TApi extends "devin-agent" ? DevinCompat : TApi extends "google-generative-ai" | "google-vertex" | "google-gemini-cli" ? GoogleCompat : TApi extends "ollama-chat" ? OllamaCompat : undefined;
 /** Resolved compat for a given API: complete record, materialized once by `buildModel`. */
-export type CompatOf<TApi extends Api> = TApi extends "openrouter" ? ResolvedOpenRouterCompat : TApi extends "openai-completions" ? ResolvedOpenAICompat : TApi extends "openai-responses" | "azure-openai-responses" | "openai-codex-responses" ? ResolvedOpenAIResponsesCompat : TApi extends "anthropic-messages" ? ResolvedAnthropicCompat : TApi extends "bedrock-converse-stream" ? ResolvedBedrockCompat : TApi extends "devin-agent" ? ResolvedDevinCompat : TApi extends "google-generative-ai" | "google-vertex" | "google-gemini-cli" ? ResolvedGoogleCompat : undefined;
+export type CompatOf<TApi extends Api> = TApi extends "openrouter" ? ResolvedOpenRouterCompat : TApi extends "openai-completions" ? ResolvedOpenAICompat : TApi extends "openai-responses" | "azure-openai-responses" | "openai-codex-responses" ? ResolvedOpenAIResponsesCompat : TApi extends "anthropic-messages" ? ResolvedAnthropicCompat : TApi extends "bedrock-converse-stream" ? ResolvedBedrockCompat : TApi extends "devin-agent" ? ResolvedDevinCompat : TApi extends "google-generative-ai" | "google-vertex" | "google-gemini-cli" ? ResolvedGoogleCompat : TApi extends "ollama-chat" ? ResolvedOllamaCompat : undefined;
 /** Provider-native compaction endpoint configuration for one model. */
 export interface RemoteCompactionConfig<TApi extends Api = Api> {
     /** Enables provider-native compaction for providers not enabled by built-in policy. */
@@ -1145,7 +1188,7 @@ export interface ModelCost extends TokenCost {
  * generations rather than providers: DeepSeek V3 through V4 share
  * `"deepseek-v3"`; Kimi K2 through K3 share `"kimi-k2"`.
  */
-export type ModelTokenizer = "claude-v3" | "claude-v47" | "claude-v5" | "claude-v5-sonnet" | "qwen3" | "deepseek-v3" | "kimi-k2" | "glm5";
+export type ModelTokenizer = "claude-v3" | "claude-v47" | "claude-v5" | "claude-v5-sonnet" | "qwen3" | "deepseek-v3" | "kimi-k2" | "glm5" | "o200k";
 /** One account's discovered entitlements on a model; see {@link Model.accountAccess}. */
 export interface ModelAccountAccess {
     /**
@@ -1198,6 +1241,8 @@ export interface Model<TApi extends Api = Api> {
     requiresGlyphTokenization?: boolean;
     /** Whether this model requires Cursor's tool-schema combiner projection. */
     requiresCursorToolSchemaProjection?: boolean;
+    /** Whether advertised tool schemas occupy every model request's context (false for deferred catalogs). */
+    toolSchemasInContext?: boolean;
     /** Whether this model requires tool-result images hoisted into sibling user content blocks. */
     requiresToolResultImageHoisting?: boolean;
     /**
@@ -1246,6 +1291,12 @@ export interface Model<TApi extends Api = Api> {
      * than ZETA's general image pipeline. `stb` local backends reject WebP.
      */
     imageInputDecoder?: "stb";
+    /**
+     * Most images one request may carry, from user configuration. Replaces the
+     * catalog's `image-count-budget` for hosts no rule names, such as a custom
+     * gateway in front of a vendor with a higher limit.
+     */
+    imageCountBudget?: number;
     /**
      * Native provider tool-call support. `false` is the only unsupported signal:
      * `true` and `undefined` both mean callers may use native tools. Catalog and
@@ -1328,6 +1379,8 @@ export interface Model<TApi extends Api = Api> {
     contextWindow: number | null;
     /** Preserve the host's supplied window instead of applying inferred expansion or reference-price caps. */
     contextWindowAuthoritative?: boolean;
+    /** The long-context price tier does not cap the runtime context window when extended context is off. */
+    longContextWindowUncapped?: boolean;
     /** Optional larger prompt window available when extended context is enabled. */
     maxContextWindow?: number;
     maxTokens: number | null;
@@ -1354,7 +1407,7 @@ export interface Model<TApi extends Api = Api> {
      * Streaming transport override. When `"pi-native"`, `streamSimple` routes
      * the request to the model's `baseUrl` via the auth-gateway's
      * `POST /v1/pi/stream` endpoint instead of dispatching the per-API
-     * provider client. The `baseUrl` must point at an `zeta-c auth-gateway`
+     * provider client. The `baseUrl` must point at an `zetacode auth-gateway`
      * (or compatible) host; `headers.Authorization` (or `apiKey` resolved by
      * the registry) carries the gateway bearer.
      *
@@ -1369,8 +1422,6 @@ export interface Model<TApi extends Api = Api> {
     preferWebsockets?: boolean;
     /** Codex Responses Lite transport: send the lite marker and carry instructions/tools as input items (mirrors codex-rs `use_responses_lite`). */
     useResponsesLite?: boolean;
-    /** Codex Code Mode restriction: model expects tools routed through a programmatic exec surface (mirrors codex-rs `tool_mode`). */
-    toolMode?: "code_mode_only";
     /**
      * Service-tier ids the provider advertises for this model (Codex discovery
      * `service_tiers[].id`, e.g. `priority`, `ultrafast`). Absent when the
@@ -1463,7 +1514,7 @@ export interface Model<TApi extends Api = Api> {
  * vocabulary of `buildModel`. Identical to `Model` except `compat` carries the
  * sparse override shape and nothing is resolved yet.
  */
-export interface ModelSpec<TApi extends Api = Api> extends Omit<Model<TApi>, "compat" | "identity" | "compatConfig" | "requiresGlyphTokenization" | "requiresCursorToolSchemaProjection" | "requiresToolResultImageHoisting" | "supportsAssistantPrefill" | "supportsComputerUseConfig"> {
+export interface ModelSpec<TApi extends Api = Api> extends Omit<Model<TApi>, "compat" | "identity" | "compatConfig" | "requiresGlyphTokenization" | "requiresCursorToolSchemaProjection" | "toolSchemasInContext" | "requiresToolResultImageHoisting" | "supportsAssistantPrefill" | "supportsComputerUseConfig"> {
     /** Sparse compatibility overrides; resolved into `Model.compat` by `buildModel`. */
     compat?: CompatConfigOf<TApi>;
 }