@oh-my-pi/pi-catalog 18.2.0 → 18.2.1

package/src/provider-models/bundled-references.ts

@@ -1,3 +1,4 @@
+import { isBareIdReferenceProvider } from "../compat/behavior";
 import { isZeroCostXaiOAuthReference } from "../identity/reference";
 import { getBundledModels, getBundledProviders } from "../models";
 import type { Api, Model, ModelSpec } from "../types";
@@ -45,9 +46,9 @@ function getGlobalReferences(): Map<string, Model<Api>> {
 	for (const provider of getBundledProviders()) {
 		for (const model of getBundledModels(provider as Parameters<typeof getBundledModels>[0])) {
 			const candidate = model as Model<Api>;
-			// ClinePass limits, pricing, and reasoning controls are gateway-specific;
-			// matching them by bare id would contaminate unrelated proxy models.
-			if (candidate.provider === "cline-pass" || isZeroCostXaiOAuthReference(candidate)) {
+			// Gateway-specific metadata must remain provider-local when proxy
+			// discovery resolves references by bare model id.
+			if (!isBareIdReferenceProvider(candidate.provider) || isZeroCostXaiOAuthReference(candidate)) {
 				continue;
 			}
 			const existing = references.get(candidate.id);

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

@@ -89,15 +89,21 @@ export function resolveModelCacheProviderId(providerId: string, options: ModelCa
 		}
 		case "litellm": {
 			const baseUrl = options.baseUrl ?? getDefaultModelDiscoveryBaseUrl(providerId)!;
-			// rich-v9 unions compat across the management endpoints and keys the
-			// deployment's `supports_vision` declaration into it, so a warm
-			// rich-v8 row would keep retracting axes an earlier endpoint reported
-			// (issue #11982). rich-v8 invalidated rows whose `compatConfig`
-			// retained a colliding bundled model's provider-specific transport
-			// (e.g. Fireworks `wireModelIdMode`) before that leak was fixed
-			// (issue #9938).
-			return `litellm:rich-v9:${Bun.hash(baseUrl).toString(36)}`;
+			// rich-v11 invalidates rows that inherited ClinePass gateway metadata
+			// through generic models.dev bare-id enrichment (issue #10932). rich-v10
+			// filtered known non-conversational LiteLLM modes, unioned compat across
+			// the management endpoints, and keyed the deployment's `supports_vision`
+			// declaration into it; earlier versions invalidated rows whose
+			// `compatConfig` retained a colliding bundled model's provider-specific
+			// transport (e.g. Fireworks `wireModelIdMode`) (issue #9938).
+			return `litellm:rich-v11:${Bun.hash(baseUrl).toString(36)}`;
 		}
+		case "gmi-cloud":
+		case "siliconflow":
+		case "siliconflow-cn":
+			// models-v1 moves rows enriched before cross-provider reference
+			// isolation out of the legacy bare-provider namespaces (#10932).
+			return `${providerId}:models-v1`;
 		case "opencode-go":
 		case "opencode-zen": {
 			// v3: gateway-first rows cached before stencil enrichment carry null

package/src/provider-models/ollama.ts

@@ -1,9 +1,6 @@
 import { fetchWithRetry } from "@oh-my-pi/pi-utils";
-import { compareRevision, parseRevision } from "../compat/revision";
-import { classifyModel } from "../compat/taxonomy";
-import { Effort } from "../effort";
 import type { ModelManagerOptions } from "../model-manager";
-import type { FetchImpl, ModelSpec, ThinkingConfig } from "../types";
+import type { FetchImpl, ModelSpec } from "../types";
 import { discoveryFetch } from "../utils";
 import { createBundledReferenceMap, createReferenceResolver } from "./bundled-references";
 
@@ -54,11 +51,6 @@ export function isOllamaCloudOutputCapped(id: string): boolean {
 	return OLLAMA_CLOUD_OUTPUT_CAPPED_BASE_IDS[baseId] === true;
 }
 
-const OLLAMA_CLOUD_GLM_52_THINKING: ThinkingConfig = {
-	mode: "effort",
-	efforts: [Effort.High, Effort.Max],
-};
-
 function trimTrailingSlash(value: string): string {
 	return value.endsWith("/") ? value.slice(0, -1) : value;
 }
@@ -93,27 +85,6 @@ function getContextWindow(modelInfo: Record<string, unknown> | undefined): numbe
 	}
 }
 
-function getThinkingConfig(modelId: string, capabilities: string[] | undefined): ThinkingConfig | undefined {
-	if (!capabilities?.includes("thinking")) {
-		return undefined;
-	}
-	const identity = classifyModel("ollama-cloud", modelId, { lenient: true });
-	const revision = identity.revision === undefined ? undefined : parseRevision(identity.revision);
-	const floor = parseRevision(identity.family === "flash" ? "5.3" : "5.2");
-	const isGlmEffortModel =
-		identity.class === "glm" &&
-		(identity.family === undefined ||
-			identity.family === "air" ||
-			identity.family === "turbo" ||
-			identity.family === "flash") &&
-		revision !== undefined &&
-		floor !== undefined &&
-		compareRevision(revision, floor) >= 0;
-	if (isGlmEffortModel) {
-		return OLLAMA_CLOUD_GLM_52_THINKING;
-	}
-	return { mode: "effort", efforts: [Effort.Minimal, Effort.Low, Effort.Medium, Effort.High] };
-}
 async function fetchShowMetadata(
 	baseUrl: string,
 	apiKey: string,
@@ -182,7 +153,16 @@ export function ollamaCloudModelManagerOptions(
 					// reference limit, falling back to the historical safe cap otherwise.
 					const contextWindow = discoveredContextWindow ?? 128000;
 					const reasoning = capabilities ? capabilities.includes("thinking") : (reference?.reasoning ?? false);
-					const thinking = capabilities ? getThinkingConfig(id, capabilities) : reference?.thinking;
+					// `/api/show` reports only a boolean `thinking` capability, never a
+					// tier vocabulary, so the effort ladder is left to the compat rules
+					// (which own every other host's ladder too). Synthesizing one here
+					// shadowed the rules: discovery rows carry explicit `thinking`, and
+					// `resolveThinkingPolicy` treats explicit metadata as authoritative
+					// over the KDL, so a fabricated `minimal..high` ladder overrode the
+					// DeepSeek V4 `low/high/max` contract and silently clamped `max`
+					// down to `high` (#8334 regression). Models whose ladder the rules
+					// do not know still fall back to the generic four-tier default.
+					const thinking = capabilities ? undefined : reference?.thinking;
 					const input = capabilities
 						? capabilities.includes("vision")
 							? (["text", "image"] as Array<"text" | "image">)

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

@@ -4,6 +4,8 @@ import { toClinePassPublicModelId } from "../cline-pass-model-id";
 import {
 	apiRouteExactModelIds,
 	apiRouteFor,
+	isBareIdReferenceProvider,
+	isExcludedDiscoveryMode,
 	isExcludedModel,
 	isLikelyOpenAIResponsesId,
 	modelLimitsFor,
@@ -579,6 +581,7 @@ type OpenAICompatibleModelManagerBuilderOptions<TApi extends Api> = {
 	dynamicModelsAuthoritative?: true;
 	requireApiKey?: true;
 	dropCachedModelIdsOnStaticMismatch?: readonly string[];
+	cacheProviderId?: string;
 	filterModel?: (
 		entry: OpenAICompatibleModelRecord,
 		model: ModelSpec<TApi>,
@@ -600,6 +603,7 @@ function createOpenAICompatibleModelManagerOptions<TApi extends Api>(
 	const filterModel = options.filterModel;
 	return {
 		providerId: options.providerId,
+		...(options.cacheProviderId && { cacheProviderId: options.cacheProviderId }),
 		...(options.dynamicModelsAuthoritative && { dynamicModelsAuthoritative: true }),
 		...(options.dropCachedModelIdsOnStaticMismatch && {
 			dropCachedModelIdsOnStaticMismatch: options.dropCachedModelIdsOnStaticMismatch,
@@ -982,6 +986,7 @@ export function gmiCloudModelManagerOptions(
 		api: "openai-completions",
 		providerId: "gmi-cloud",
 		defaultBaseUrl: GMI_CLOUD_BASE_URL,
+		cacheProviderId: resolveModelCacheProviderId("gmi-cloud"),
 		config,
 		requireApiKey: true,
 		mapModel: mapGmiCloudModel,
@@ -1201,6 +1206,13 @@ function mapDeepinfraModel(
 		return null;
 	}
 	const pricing = isRecord(metadata.pricing) ? metadata.pricing : {};
+	// `metadata.discount` is a promotional fraction in [0, 1): DeepInfra bills
+	// `pricing * (1 - discount)` (verified against the site — GLM-5.2 lists
+	// input 0.75 with discount 0.35 and charges 0.4875), while `pricing.*`
+	// stays at list price. Fold it into the rate card so cost reporting matches
+	// what the user is actually billed. Values outside (0, 1) are ignored.
+	const discount = toNumber(metadata.discount);
+	const discountMultiplier = discount !== undefined && discount > 0 && discount < 1 ? 1 - discount : 1;
 	// `reasoning_effort` marks models whose effort dial is advertised. The
 	// parameter itself is validated and accepted platform-wide on DeepInfra
 	// (verified: 200 on effort-tagged, reasoning-only, and plain-chat models;
@@ -1243,9 +1255,9 @@ function mapDeepinfraModel(
 		...(thinking ? { thinking } : {}),
 		input: tags.includes("vision") || tags.includes("vlm") ? ["text", "image"] : ["text"],
 		cost: {
-			input: toPositiveNumber(pricing.input_tokens, 0),
-			output: toPositiveNumber(pricing.output_tokens, 0),
-			cacheRead: toPositiveNumber(pricing.cache_read_tokens, 0),
+			input: toPositiveNumber(pricing.input_tokens, 0) * discountMultiplier,
+			output: toPositiveNumber(pricing.output_tokens, 0) * discountMultiplier,
+			cacheRead: toPositiveNumber(pricing.cache_read_tokens, 0) * discountMultiplier,
 			cacheWrite: 0,
 		},
 		contextWindow,
@@ -1711,6 +1723,7 @@ function createSiliconFlowModelManagerOptions(
 	const baseUrl = config?.baseUrl ?? defaultBaseUrl;
 	return {
 		providerId,
+		cacheProviderId: resolveModelCacheProviderId(providerId),
 		dynamicModelsAuthoritative: true,
 		...(apiKey && {
 			fetchDynamicModels: async () => {
@@ -2134,6 +2147,9 @@ function createModelsDevReferenceMap<TApi extends Api>(
 	const references = new Map<string, ModelSpec<TApi>>();
 	for (const model of models) {
 		const candidate = model as ModelSpec<TApi>;
+		if (!isBareIdReferenceProvider(candidate.provider)) {
+			continue;
+		}
 		const existing = references.get(candidate.id);
 		if (!existing) {
 			references.set(candidate.id, candidate);
@@ -3866,10 +3882,17 @@ export interface BasetenModelManagerConfig {
 	fetch?: FetchImpl;
 }
 
-// A previous version of OMP shipped this model without reasoning levels. We've
-// since fixed that. This const lets us bust the cache so that users on that
-// version of OMP pick up the reasoning levels immediately.
-const BASETEN_CACHE_MIGRATION_MODEL_IDS = ["zai-org/GLM-5.3", "zai-org/GLM-5.3-Flash"] as const;
+// A previous version of OMP shipped these models without reasoning levels.
+// We've since fixed that (V4-generation whitelist). This const lets us bust
+// the cache so that users on that version of OMP pick up the reasoning levels
+// immediately.
+const BASETEN_CACHE_MIGRATION_MODEL_IDS = [
+	"zai-org/GLM-5.3",
+	"zai-org/GLM-5.3-Flash",
+	"deepseek-ai/DeepSeek-V4-Flash-0731",
+	"deepseek-ai/DeepSeek-V4.1-Flash",
+	"deepseek-ai/DeepSeek-V4-Pro-0813",
+] as const;
 
 export function basetenModelManagerOptions(
 	config?: BasetenModelManagerConfig,
@@ -3900,7 +3923,7 @@ export function basetenModelManagerOptions(
 				(identity.class === "kimi" && identity.family === "k3") ||
 				isGlmReasoningIdentity("baseten", defaults.id, "5.2") ||
 				defaults.id === "openai/gpt-oss-120b" ||
-				defaults.id === "deepseek-ai/DeepSeek-V4-Pro";
+				isDeepseekV4Generation("baseten", defaults.id);
 			const reasoning =
 				isSupportedBasetenReasoningModel &&
 				(features.includes("reasoning") || features.includes("reasoning_effort"));
@@ -4678,7 +4701,11 @@ type LiteLLMRichEndpointFailure = {
 	error?: unknown;
 };
 type LiteLLMRichEndpointResult<TApi extends Api> =
-	| { models: LiteLLMRichEndpointModel<TApi>[]; incompleteVisionMetadata: boolean }
+	| {
+			models: LiteLLMRichEndpointModel<TApi>[];
+			excludedModelIds: ReadonlySet<string>;
+			incompleteVisionMetadata: boolean;
+	  }
 	| { failure: LiteLLMRichEndpointFailure };
 
 const LITELLM_RICH_ENDPOINTS = ["/model_group/info", "/v2/model/info", "/model/info", "/v1/model/info"] as const;
@@ -4707,6 +4734,11 @@ function warnLiteLLMMetadataFallback(managementBaseUrl: string, failure: LiteLLM
 	});
 }
 
+/** Exclude only known non-conversational modes; unknown and non-string modes remain selectable for aliases. */
+export function isSelectableLiteLLMModelMode(mode: unknown): boolean {
+	return typeof mode !== "string" || !isExcludedDiscoveryMode("litellm", mode);
+}
+
 export function normalizeLiteLLMManagementBaseUrl(baseUrl: string): string {
 	const trimmed = baseUrl.trim().replace(/\/+$/g, "");
 	if (!trimmed) {
@@ -4747,7 +4779,10 @@ function mapLiteLLMOpenAICompatibleModel(
 	entry: OpenAICompatibleModelRecord,
 	defaults: ModelSpec<Api>,
 	reference: ModelSpec<Api> | undefined,
-): ModelSpec<Api> {
+): ModelSpec<Api> | null {
+	if (!isSelectableLiteLLMModelMode(entry.mode)) {
+		return null;
+	}
 	const model = mapWithBundledReference(entry, defaults, reference);
 	return {
 		...model,
@@ -4943,7 +4978,10 @@ function mapLiteLLMRichEntry<TApi extends Api>(
 	options: FetchLiteLLMRichModelsOptions<TApi>,
 	runtimeBaseUrl: string,
 ): ModelSpec<TApi> | null {
-	if (isLiteLLMUnusableSentinelPlaceholder(entry)) {
+	if (
+		!isSelectableLiteLLMModelMode(getLiteLLMMetadataValue(entry, "mode")) ||
+		isLiteLLMUnusableSentinelPlaceholder(entry)
+	) {
 		return null;
 	}
 	const id = getLiteLLMRichModelId(entry);
@@ -5151,7 +5189,22 @@ async function fetchLiteLLMRichEndpoint<TApi extends Api>(
 		return null;
 	}
 	const deduped = new Map<string, LiteLLMRichEndpointModel<TApi>>();
+	const excludedModelIds = new Set<string>();
 	for (const entry of entries) {
+		if (isLiteLLMUnusableSentinelPlaceholder(entry)) {
+			continue;
+		}
+		const modelId = getLiteLLMRichModelId(entry);
+		if (!isSelectableLiteLLMModelMode(getLiteLLMMetadataValue(entry, "mode"))) {
+			if (modelId) {
+				excludedModelIds.add(modelId);
+				deduped.delete(modelId);
+			}
+			continue;
+		}
+		if (modelId && excludedModelIds.has(modelId)) {
+			continue;
+		}
 		const model = mapLiteLLMRichEntry(entry, options, runtimeBaseUrl);
 		if (model) {
 			const supportsVision = getLiteLLMMetadataValue(entry, "supports_vision");
@@ -5177,12 +5230,13 @@ async function fetchLiteLLMRichEndpoint<TApi extends Api>(
 			deduped.set(model.id, existing ? mergeLiteLLMRichEndpointModels(existing, next) : next);
 		}
 	}
-	if (deduped.size === 0) {
+	if (deduped.size === 0 && excludedModelIds.size === 0) {
 		return null;
 	}
 	const models = Array.from(deduped.values()).sort((left, right) => left.model.id.localeCompare(right.model.id));
 	return {
 		models,
+		excludedModelIds,
 		incompleteVisionMetadata: models.some(entry => entry.supportsVision !== true && entry.supportsVision !== false),
 	};
 }
@@ -5197,6 +5251,7 @@ async function fetchLiteLLMRichModelsInternal<TApi extends Api>(
 	}
 	const fetchModels = async (signal?: AbortSignal): Promise<ModelSpec<TApi>[] | null> => {
 		const deduped = new Map<string, LiteLLMRichEndpointModel<TApi>>();
+		const excludedModelIds = new Set<string>();
 		let metadataFailure: LiteLLMRichEndpointFailure | undefined;
 		for (const endpoint of LITELLM_RICH_ENDPOINTS) {
 			const result = await fetchLiteLLMRichEndpoint(endpoint, options, managementBaseUrl, runtimeBaseUrl, signal);
@@ -5218,8 +5273,15 @@ async function fetchLiteLLMRichModelsInternal<TApi extends Api>(
 				}
 				continue;
 			}
+			for (const modelId of result.excludedModelIds) {
+				excludedModelIds.add(modelId);
+				deduped.delete(modelId);
+			}
 			const hadPriorModels = deduped.size > 0;
 			for (const next of result.models) {
+				if (excludedModelIds.has(next.model.id)) {
+					continue;
+				}
 				const existing = deduped.get(next.model.id);
 				if (!existing) {
 					if (!hadPriorModels) {
@@ -5229,6 +5291,9 @@ async function fetchLiteLLMRichModelsInternal<TApi extends Api>(
 				}
 				deduped.set(next.model.id, mergeLiteLLMRichEndpointModels(existing, next));
 			}
+			if (deduped.size === 0) {
+				continue;
+			}
 			let needsMoreMetadata = false;
 			for (const entry of deduped.values()) {
 				if (
@@ -5249,6 +5314,9 @@ async function fetchLiteLLMRichModelsInternal<TApi extends Api>(
 			}
 		}
 		if (deduped.size === 0) {
+			if (excludedModelIds.size > 0) {
+				return [];
+			}
 			if (metadataFailure) {
 				warnLiteLLMMetadataFallback(managementBaseUrl, metadataFailure);
 			}
@@ -5275,17 +5343,18 @@ export function litellmModelManagerOptions(config?: LiteLLMModelManagerConfig):
 	const baseUrl = config?.baseUrl ?? getDefaultModelDiscoveryBaseUrl("litellm")!;
 	return {
 		providerId: "litellm",
-		// rich-v9 keys the deployment's `supports_vision` declaration into the
-		// cached compat and unions compat across management endpoints instead of
-		// letting a later one retract what an earlier one reported (issue
-		// #11982). rich-v8 invalidated rows whose `compatConfig` retained a
-		// colliding bundled model's provider-specific transport (e.g. Fireworks
-		// `wireModelIdMode`) before that leak was fixed. Earlier versions added
-		// bundled reference fallback, moved OpenAI models to Responses, continued
-		// past incomplete vision/API metadata and endpoints omitting cache
-		// pricing, stripped reseller usage suffixes, filtered placeholder rows,
-		// and mapped rich pricing. Bump the version whenever these mappers change,
-		// or warm authoritative caches keep serving pre-change rows for the full TTL.
+		// rich-v11 invalidates rows that inherited ClinePass gateway metadata
+		// through generic models.dev bare-id enrichment (issue #10932). rich-v10
+		// filtered known non-conversational LiteLLM modes, keyed the deployment's
+		// `supports_vision` declaration into cached compat, and unioned compat
+		// across management endpoints instead of letting a later endpoint retract
+		// what an earlier one reported (issue #11982). Earlier versions fixed
+		// provider-specific transport leakage, added bundled reference fallback,
+		// moved OpenAI models to Responses, continued past incomplete vision/API
+		// metadata and endpoints omitting cache pricing, stripped reseller usage
+		// suffixes, filtered placeholder rows, and mapped rich pricing. Bump the
+		// version whenever these mappers change, or warm authoritative caches keep
+		// serving pre-change rows for the full TTL.
 		cacheProviderId: resolveModelCacheProviderId("litellm", { baseUrl }),
 		// litellm is a local-only proxy and is never bundled in models.json (that
 		// would leak the machine's localhost catalog). Prefer the proxy's richer
@@ -5304,7 +5373,7 @@ export function litellmModelManagerOptions(config?: LiteLLMModelManagerConfig):
 				resolveApi: resolveLiteLLMApi,
 				timeoutMs: 10_000,
 			});
-			if (richModels && richModels.length > 0) {
+			if (richModels !== null) {
 				return richModels;
 			}
 			return fetchOpenAICompatibleModels<Api>({
@@ -6308,6 +6377,7 @@ const MODELS_DEV_PROVIDER_DESCRIPTORS_CORE: readonly ModelsDevProviderDescriptor
 			return {
 				...model,
 				id,
+				name: id,
 				thinking: model.reasoning ? buildClinePassThinking(raw, model) : undefined,
 			};
 		},

package/src/types.ts

@@ -327,6 +327,12 @@ export interface OpenAICompat {
 	 * Default: auto-detected (DeepSeek reasoning models).
 	 */
 	disableReasoningOnToolChoice?: boolean;
+	/**
+	 * Disable reasoning whenever the request advertises function tools.
+	 * Use for model surfaces that reject every tools-plus-reasoning combination.
+	 * Default: false.
+	 */
+	disableReasoningWithTools?: boolean;
 	/** OpenRouter-specific routing preferences. Only used when baseUrl points to OpenRouter. */
 	openRouterRouting?: OpenRouterRouting;
 	/** Vercel AI Gateway routing preferences. Only used when baseUrl points to Vercel AI Gateway. */
@@ -421,12 +427,24 @@ export interface OpenAICompat {
 	strictResponsesPairing?: boolean;
 	/** Whether the Responses API accepts the `detail: "original"` image hint. Default: auto-detected (false for GitHub Copilot, which rejects it with a 400). */
 	supportsImageDetailOriginal?: boolean;
+	/**
+	 * Whether the Responses endpoint accepts `configuration_update` input items
+	 * that change `reasoning.effort` mid-conversation while the request-level
+	 * effort stays pinned for prompt caching (GPT-6 Astra). Default:
+	 * rule-detected (`true` for `gpt-6-astra` on any host, `false` otherwise).
+	 * Set `false` for custom `openai-responses` / `openai-codex-responses`
+	 * endpoints that reject the item type with HTTP 400; effort changes are then
+	 * sent as the top-level `reasoning.effort`.
+	 */
+	supportsConfigurationUpdate?: boolean;
 	/** Whether streamed reasoning deltas for the same field may repeat the full cumulative text snapshot. Default: false. */
 	reasoningDeltasMayBeCumulative?: boolean;
 	/** Strip leaked DeepSeek chat-template special tokens from visible content deltas. Default: auto-detected. */
 	stripDeepseekSpecialTokens?: 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. */
+	streamRevision?: "none" | "possible";
 	/** Treat an empty length-finished stream as a context-window error. Default: auto-detected. */
 	emptyLengthFinishIsContextError?: boolean;
 	/** Normalize tool call ids to OpenAI's 40-character limit. Default: auto-detected. */
@@ -599,6 +617,8 @@ export interface AnthropicCompat {
 export interface BedrockCompat {
 	/** Whether this endpoint accepts no checkpoints, automatic caching, or explicit cachePoint blocks. */
 	promptCacheMode?: "none" | "automatic" | "explicit";
+	/** Whether this wire may revise already-streamed text (`stream-revision` axis). Unassigned: append-only. */
+	streamRevision?: "none" | "possible";
 	/** Whether explicit cachePoint blocks accept `ttl: "1h"`; omitted TTL means Bedrock's 5-minute default. */
 	supportsLongPromptCacheRetention?: boolean;
 	/**
@@ -623,6 +643,8 @@ export interface BedrockCompat {
 /** Fully-resolved Bedrock Converse prompt-cache capabilities, materialized once by `buildModel`. */
 export interface ResolvedBedrockCompat {
 	promptCacheMode: NonNullable<BedrockCompat["promptCacheMode"]>;
+	/** See {@link BedrockCompat.streamRevision}. */
+	streamRevision?: BedrockCompat["streamRevision"];
 	supportsLongPromptCacheRetention: boolean;
 	promptCacheMinimumTokens: number;
 	promptCacheMaximumCheckpoints: number;
@@ -688,6 +710,7 @@ export interface ResolvedOpenAISharedCompat {
 	filterReasoningHistory: boolean;
 	disableReasoningOnForcedToolChoice: boolean;
 	disableReasoningOnToolChoice: boolean;
+	disableReasoningWithTools?: boolean;
 	supportsToolChoice: boolean;
 	supportsForcedToolChoice: boolean;
 	supportsNamedToolChoice: boolean;
@@ -705,6 +728,8 @@ export interface ResolvedOpenAISharedCompat {
 	requiresAssistantContentForToolCalls: boolean;
 	stripDeepseekSpecialTokens: boolean;
 	streamMarkupHealingPattern?: OpenAIStreamMarkupHealingPattern;
+	/** See {@link OpenAICompat.streamRevision}. */
+	streamRevision?: OpenAICompat["streamRevision"];
 	/** See {@link OpenAICompat.streamFirstEventTimeoutMs}. */
 	streamFirstEventTimeoutMs?: number;
 	reasoningDeltasMayBeCumulative: boolean;
@@ -765,6 +790,7 @@ export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat &
 			| "filterReasoningHistory"
 			| "disableReasoningOnForcedToolChoice"
 			| "disableReasoningOnToolChoice"
+			| "disableReasoningWithTools"
 			| "supportsToolChoice"
 			| "supportsForcedToolChoice"
 			| "supportsNamedToolChoice"
@@ -801,10 +827,12 @@ export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat &
 			| "toolSchemaFlavor"
 			| "streamFirstEventTimeoutMs"
 			| "streamIdleTimeoutMs"
+			| "streamRevision"
 			| "cacheControlFormat"
 			| "thinkingKeep"
 			| "strictResponsesPairing"
 			| "supportsImageDetailOriginal"
+			| "supportsConfigurationUpdate"
 			| "stripImageInput"
 			| "thinkingLoopGuard"
 			| "whenThinking"