@sayknow-cli/ai 0.3.7 → 0.3.9

package/dist/types/auth-broker/remote-store.d.ts

@@ -70,6 +70,7 @@ export declare class RemoteAuthCredentialStore implements AuthCredentialStore {
     getCache(key: string): string | null;
     setCache(key: string, value: string, expiresAtSec: number): void;
     cleanExpiredCache(): void;
+    deleteCachePrefix(prefix: string): void;
     /**
      * Store-level hook consumed by `AuthStorage` — routes refresh through the
      * broker so the actual refresh token never leaves the broker host. Returns

package/dist/types/auth-broker/wire-schemas.d.ts

@@ -420,6 +420,7 @@ export declare const credentialIfAbsentUploadResponseSchema: z.ZodObject<{
         "skipped-existing-fallback": "skipped-existing-fallback";
         "skipped-existing-runtime": "skipped-existing-runtime";
         "skipped-invalid": "skipped-invalid";
+        "updated-existing": "updated-existing";
     }>;
     entries: z.ZodArray<z.ZodObject<{
         id: z.ZodNumber;

package/dist/types/auth-storage.d.ts

@@ -102,7 +102,7 @@ export interface AuthCredentialSnapshotEntry {
     credential: SnapshotCredential;
     identityKey: string | null;
 }
-export type AuthCredentialIfAbsentReason = "inserted" | "skipped-existing" | "skipped-existing-runtime" | "skipped-existing-config" | "skipped-existing-env" | "skipped-existing-fallback" | "skipped-invalid";
+export type AuthCredentialIfAbsentReason = "inserted" | "updated-existing" | "skipped-existing" | "skipped-existing-runtime" | "skipped-existing-config" | "skipped-existing-env" | "skipped-existing-fallback" | "skipped-invalid";
 export interface AuthCredentialIfAbsentResult {
     inserted: boolean;
     reason: AuthCredentialIfAbsentReason;
@@ -147,6 +147,7 @@ export interface AuthCredentialStore {
         includeExpired?: boolean;
     }): string | null;
     setCache(key: string, value: string, expiresAtSec: number): void;
+    deleteCachePrefix?(prefix: string): void;
     cleanExpiredCache(): void;
     /**
      * Optional store-supplied OAuth refresh. When present, `AuthStorage` uses
@@ -646,6 +647,7 @@ export declare class SqliteAuthCredentialStore implements AuthCredentialStore {
         includeExpired?: boolean;
     }): string | null;
     setCache(key: string, value: string, expiresAtSec: number): void;
+    deleteCachePrefix(prefix: string): void;
     cleanExpiredCache(): void;
     /**
      * Save OAuth credentials for a provider.

package/dist/types/provider-models/openai-compat.d.ts

@@ -58,6 +58,11 @@ export interface DeepSeekModelManagerConfig {
     baseUrl?: string;
 }
 export declare function deepseekModelManagerOptions(config?: DeepSeekModelManagerConfig): ModelManagerOptions<"openai-completions">;
+export interface DeepInfraModelManagerConfig {
+    apiKey?: string;
+    baseUrl?: string;
+}
+export declare function deepinfraModelManagerOptions(config?: DeepInfraModelManagerConfig): ModelManagerOptions<"openai-completions">;
 export interface FireworksModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;

package/dist/types/providers/cursor/client-version.d.ts

@@ -0,0 +1,10 @@
+/**
+ * The Cursor client version reported to api2.cursor.sh.
+ *
+ * Every call against the Cursor backend must send the same
+ * x-cursor-client-version: the backend gates features and minimum versions on
+ * it, so a drift between the agent Run path and model discovery makes one of
+ * them fail while the other keeps working. Keep this as the single source of
+ * truth for the header value.
+ */
+export declare const CURSOR_CLIENT_VERSION = "cli-2026.02.13-41ac335";

package/dist/types/providers/cursor.d.ts

@@ -1,7 +1,8 @@
 import { type JsonValue } from "@bufbuild/protobuf";
 import type { CursorExecHandlerResult, CursorExecHandlers, CursorToolResultHandler, Message, StreamFunction, StreamOptions, ToolResultMessage } from "../types";
+import { CURSOR_CLIENT_VERSION } from "./cursor/client-version";
 export declare const CURSOR_API_URL = "https://api2.cursor.sh";
-export declare const CURSOR_CLIENT_VERSION = "cli-2026.01.09-231024f";
+export { CURSOR_CLIENT_VERSION };
 /** Drop all cached state + blob bytes for a conversation (F15 bound + session-teardown hook). */
 export declare function disposeCursorConversation(conversationId: string): void;
 export interface CursorOptions extends StreamOptions {

package/dist/types/providers/openai-bounded-rate-limits.d.ts

@@ -0,0 +1,3 @@
+import type { FetchImpl } from "../types";
+export declare function isOpenAIUsageExhaustionResponse(bodyText: string, retryAfterMs: number | undefined, retryDelayCapMs: number): boolean;
+export declare function wrapOpenAIFetchForBoundedRateLimits(baseFetch: FetchImpl, maxRetryDelayMs: number | undefined): FetchImpl;

package/dist/types/types.d.ts

@@ -48,7 +48,7 @@ export interface ThinkingConfig {
     /** Provider-specific transport used to encode the selected effort. */
     mode: ThinkingControlMode;
 }
-export type KnownProvider = "alibaba-coding-plan" | "amazon-bedrock" | "azure-openai" | "anthropic" | "google" | "google-gemini-cli" | "google-antigravity" | "google-vertex" | "openai" | "openai-codex" | "kimi-code" | "minimax-code" | "minimax-code-cn" | "github-copilot" | "fireworks" | "firepass" | "fugu" | "gitlab-duo" | "cursor" | "deepseek" | "xai" | "groq" | "cerebras" | "openrouter" | "kilo" | "vercel-ai-gateway" | "zai" | "glm-zcode" | "mistral" | "minimax" | "opencode-go" | "opencode-zen" | "synthetic" | "cloudflare-ai-gateway" | "huggingface" | "litellm" | "moonshot" | "nvidia" | "nanogpt" | "ollama" | "ollama-cloud" | "qianfan" | "qwen-portal" | "together" | "venice" | "vllm" | "xiaomi" | "xiaomi-token-plan-sgp" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "zenmux" | "lm-studio";
+export type KnownProvider = "alibaba-coding-plan" | "amazon-bedrock" | "azure-openai" | "anthropic" | "google" | "google-gemini-cli" | "google-antigravity" | "google-vertex" | "openai" | "openai-codex" | "kimi-code" | "minimax-code" | "minimax-code-cn" | "github-copilot" | "fireworks" | "firepass" | "fugu" | "gitlab-duo" | "cursor" | "deepseek" | "deepinfra" | "xai" | "groq" | "cerebras" | "openrouter" | "kilo" | "vercel-ai-gateway" | "zai" | "glm-zcode" | "mistral" | "minimax" | "opencode-go" | "opencode-zen" | "synthetic" | "cloudflare-ai-gateway" | "huggingface" | "litellm" | "moonshot" | "nvidia" | "nanogpt" | "ollama" | "ollama-cloud" | "qianfan" | "qwen-portal" | "together" | "venice" | "vllm" | "xiaomi" | "xiaomi-token-plan-sgp" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "zenmux" | "lm-studio";
 export type Provider = KnownProvider | string;
 import type { Effort } from "./model-thinking";
 /** Token budgets for each thinking level (token-based providers only) */
@@ -106,9 +106,9 @@ export type ResolvedServiceTier = Exclude<ServiceTier, "openai-only" | "claude-o
 export declare function resolveServiceTier(serviceTier: ServiceTier | null | undefined, provider: Provider | undefined): ResolvedServiceTier | undefined;
 /**
  * True when the (possibly scoped) tier should be sent as OpenAI's
- * `service_tier` request field for the given provider. Non-OpenAI
- * providers, unsupported tiers (`"auto"`, `"default"`), and scope
- * mismatches all return false.
+ * `service_tier` request field for the given provider. OpenAI accepts
+ * `flex`, `scale`, and `priority`; DeepInfra accepts `priority`.
+ * Unsupported tiers (`"auto"`, `"default"`) and scope mismatches return false.
  */
 export declare function shouldSendServiceTier(serviceTier: ServiceTier | null | undefined, provider: Provider | undefined): boolean;
 /**

package/dist/types/utils/oauth/types.d.ts

@@ -7,7 +7,7 @@ export type OAuthCredentials = {
     email?: string;
     accountId?: string;
 };
-export type OAuthProvider = "alibaba-coding-plan" | "anthropic" | "cerebras" | "cloudflare-ai-gateway" | "cursor" | "deepseek" | "fireworks" | "firepass" | "fugu" | "github-copilot" | "google-gemini-cli" | "google-antigravity" | "gitlab-duo" | "huggingface" | "kimi-code" | "kilo" | "kagi" | "litellm" | "lm-studio" | "minimax-code" | "minimax-code-cn" | "moonshot" | "nvidia" | "nanogpt" | "ollama" | "ollama-cloud" | "openai-codex" | "openai-codex-device" | "opencode-go" | "opencode-zen" | "parallel" | "perplexity" | "qianfan" | "qwen-portal" | "synthetic" | "tavily" | "together" | "venice" | "vercel-ai-gateway" | "vllm" | "xai" | "glm-zcode" | "xiaomi" | "xiaomi-token-plan-sgp" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "zenmux" | "zai";
+export type OAuthProvider = "alibaba-coding-plan" | "anthropic" | "cerebras" | "cloudflare-ai-gateway" | "cursor" | "deepseek" | "deepinfra" | "fireworks" | "firepass" | "fugu" | "github-copilot" | "google-gemini-cli" | "google-antigravity" | "gitlab-duo" | "huggingface" | "kimi-code" | "kilo" | "kagi" | "litellm" | "lm-studio" | "minimax-code" | "minimax-code-cn" | "moonshot" | "nvidia" | "nanogpt" | "ollama" | "ollama-cloud" | "openai-codex" | "openai-codex-device" | "opencode-go" | "opencode-zen" | "parallel" | "perplexity" | "qianfan" | "qwen-portal" | "synthetic" | "tavily" | "together" | "venice" | "vercel-ai-gateway" | "vllm" | "xai" | "glm-zcode" | "xiaomi" | "xiaomi-token-plan-sgp" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "zenmux" | "zai";
 export type OAuthProviderId = OAuthProvider | (string & {});
 export type OAuthPrompt = {
     message: string;

package/package.json

@@ -1,9 +1,9 @@
 {
 	"type": "module",
 	"name": "@sayknow-cli/ai",
-	"version": "0.3.7",
+	"version": "0.3.9",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
-	"homepage": "https://github.com/jaybeyond/Sayknow_CLI",
+	"homepage": "https://sayknow-cli.com",
 	"author": "jaybeyond",
 	"contributors": [
 		"Mario Zechner"
@@ -43,7 +43,7 @@
 	"dependencies": {
 		"@anthropic-ai/sdk": "^0.94.0",
 		"@bufbuild/protobuf": "^2.12.0",
-		"@sayknow-cli/utils": "0.3.7",
+		"@sayknow-cli/utils": "0.3.9",
 		"openai": "^6.36.0",
 		"partial-json": "^0.1.7",
 		"zod": "4.4.3"

package/src/auth-broker/remote-store.ts

@@ -86,6 +86,7 @@ export class RemoteAuthCredentialStore implements AuthCredentialStore {
 	#cache: Map<string, CacheEntry> = new Map();
 	#usageCache?: UsageCacheEntry;
 	#usageInflight?: Promise<UsageReport[] | null>;
+	#usageCacheEpoch = 0;
 	#closed = false;
 	/**
 	 * `true` once the SSE consumer received its first frame and hasn't dropped
@@ -465,6 +466,19 @@ export class RemoteAuthCredentialStore implements AuthCredentialStore {
 		}
 	}
 
+	deleteCachePrefix(prefix: string): void {
+		for (const key of this.#cache.keys()) {
+			if (key.startsWith(prefix)) this.#cache.delete(key);
+		}
+		if (prefix.startsWith("usage_cache:")) this.#invalidateUsageCache();
+	}
+
+	#invalidateUsageCache(): void {
+		this.#usageCache = undefined;
+		this.#usageInflight = undefined;
+		this.#usageCacheEpoch += 1;
+	}
+
 	/**
 	 * Store-level hook consumed by `AuthStorage` — routes refresh through the
 	 * broker so the actual refresh token never leaves the broker host. Returns
@@ -561,10 +575,11 @@ export class RemoteAuthCredentialStore implements AuthCredentialStore {
 			return Promise.resolve(cached.reports);
 		}
 		if (this.#usageInflight) return this.#usageInflight;
+		const epoch = this.#usageCacheEpoch;
 		const inflight = this.#client
 			.fetchUsage()
 			.then(body => {
-				this.#usageCache = { reports: body.reports, fetchedAt: Date.now() };
+				if (this.#usageCacheEpoch === epoch) this.#usageCache = { reports: body.reports, fetchedAt: Date.now() };
 				return body.reports;
 			})
 			.catch(error => {
@@ -572,7 +587,7 @@ export class RemoteAuthCredentialStore implements AuthCredentialStore {
 				return null;
 			})
 			.finally(() => {
-				this.#usageInflight = undefined;
+				if (this.#usageCacheEpoch === epoch) this.#usageInflight = undefined;
 			});
 		this.#usageInflight = inflight;
 		return inflight;

package/src/auth-broker/wire-schemas.ts

@@ -187,6 +187,7 @@ export const credentialDisableResponseSchema = z
 // ─── Upload ────────────────────────────────────────────────────────────────
 const credentialIfAbsentReasonSchema = z.enum([
 	"inserted",
+	"updated-existing",
 	"skipped-existing",
 	"skipped-existing-runtime",
 	"skipped-existing-config",

package/src/auth-storage.ts

@@ -32,6 +32,7 @@ import { kimiUsageProvider } from "./usage/kimi";
 import { codexRankingStrategy, openaiCodexUsageProvider } from "./usage/openai-codex";
 import { zaiUsageProvider } from "./usage/zai";
 import { getOAuthApiKey, getOAuthProvider, refreshOAuthToken, resolveOAuthStorageProvider } from "./utils/oauth";
+import { loginDeepInfra } from "./utils/oauth/deepinfra";
 import { loginDeepSeek } from "./utils/oauth/deepseek";
 import { loginOpenAICodexDevice } from "./utils/oauth/openai-codex";
 import type { OAuthController, OAuthCredentials, OAuthProvider, OAuthProviderId } from "./utils/oauth/types";
@@ -153,6 +154,7 @@ export interface AuthCredentialSnapshotEntry {
 
 export type AuthCredentialIfAbsentReason =
 	| "inserted"
+	| "updated-existing"
 	| "skipped-existing"
 	| "skipped-existing-runtime"
 	| "skipped-existing-config"
@@ -209,6 +211,7 @@ export interface AuthCredentialStore {
 	deleteAuthCredentialsForProvider(provider: string, disabledCause: string): void;
 	getCache(key: string, options?: { includeExpired?: boolean }): string | null;
 	setCache(key: string, value: string, expiresAtSec: number): void;
+	deleteCachePrefix?(prefix: string): void;
 	cleanExpiredCache(): void;
 	/**
 	 * Optional store-supplied OAuth refresh. When present, `AuthStorage` uses
@@ -471,6 +474,7 @@ interface UsageCache {
 	get<T>(key: string): UsageCacheEntry<T> | undefined;
 	getStale<T>(key: string): UsageCacheEntry<T> | undefined;
 	set<T>(key: string, entry: UsageCacheEntry<T>): void;
+	deletePrefix?(prefix: string): void;
 	cleanup?(): void;
 }
 
@@ -661,6 +665,10 @@ class AuthStorageUsageCache implements UsageCache {
 		this.store.setCache(`${USAGE_CACHE_PREFIX}${key}`, payload, Math.floor(durableExpiresAt / 1000));
 	}
 
+	deletePrefix(prefix: string): void {
+		this.store.deleteCachePrefix?.(`${USAGE_CACHE_PREFIX}${prefix}`);
+	}
+
 	cleanup(): void {
 		this.store.cleanExpiredCache();
 	}
@@ -1296,8 +1304,6 @@ export class AuthStorage {
 			return this.#snapshotSkipResult(storageProvider, "skipped-existing-runtime");
 		if (this.#configOverrides.has(storageProvider))
 			return this.#snapshotSkipResult(storageProvider, "skipped-existing-config");
-		if (this.#getCredentialsForProvider(storageProvider).length > 0)
-			return this.#snapshotSkipResult(storageProvider, "skipped-existing");
 		if (getEnvApiKey(storageProvider)) return this.#snapshotSkipResult(storageProvider, "skipped-existing-env");
 		if (this.#fallbackResolver?.(storageProvider))
 			return this.#snapshotSkipResult(storageProvider, "skipped-existing-fallback");
@@ -1310,6 +1316,7 @@ export class AuthStorage {
 			result.entries.map(entry => ({ id: entry.id, credential: entry.credential })),
 		);
 		this.#resetProviderAssignments(storageProvider);
+		if (result.inserted) this.#invalidateUsageCacheForProvider(storageProvider);
 		return {
 			inserted: result.inserted,
 			reason: result.reason,
@@ -1327,6 +1334,13 @@ export class AuthStorage {
 			stored.map(record => ({ id: record.id, credential: record.credential })),
 		);
 		this.#resetProviderAssignments(provider);
+		this.#invalidateUsageCacheForProvider(provider);
+	}
+
+	#invalidateUsageCacheForProvider(provider: string): void {
+		this.#usageRequestInFlight.clear();
+		this.#usageReportsInFlight.clear();
+		this.#usageCache.deletePrefix?.(`report:${provider}:`);
 	}
 
 	/**
@@ -1593,6 +1607,11 @@ export class AuthStorage {
 				await saveApiKeyCredential(apiKey);
 				return;
 			}
+			case "deepinfra": {
+				const apiKey = await loginDeepInfra(ctrl);
+				await saveApiKeyCredential(apiKey);
+				return;
+			}
 			case "xai": {
 				const { loginXai } = await import("./utils/oauth/xai");
 				credentials = await loginXai({
@@ -3699,6 +3718,7 @@ export class SqliteAuthCredentialStore implements AuthCredentialStore {
 	#getCacheStmt: Statement;
 	#getCacheIncludingExpiredStmt: Statement;
 	#upsertCacheStmt: Statement;
+	#deleteCachePrefixStmt: Statement;
 	#deleteExpiredCacheStmt: Statement;
 	#closed = false;
 
@@ -3738,6 +3758,7 @@ export class SqliteAuthCredentialStore implements AuthCredentialStore {
 		this.#upsertCacheStmt = this.#db.prepare(
 			"INSERT INTO cache (key, value, expires_at) VALUES (?, ?, ?) ON CONFLICT(key) DO UPDATE SET value = excluded.value, expires_at = excluded.expires_at",
 		);
+		this.#deleteCachePrefixStmt = this.#db.prepare("DELETE FROM cache WHERE substr(key, 1, ?) = ?");
 		this.#deleteExpiredCacheStmt = this.#db.prepare(`DELETE FROM cache WHERE expires_at <= ${SQLITE_NOW_EPOCH}`);
 	}
 
@@ -4096,16 +4117,58 @@ export class SqliteAuthCredentialStore implements AuthCredentialStore {
 		}
 
 		const writeIfAbsent = this.#db.transaction(
-			(providerName: string, record: SerializedCredentialRecord): AuthCredentialIfAbsentResult => {
+			(
+				providerName: string,
+				item: AuthCredential,
+				record: SerializedCredentialRecord,
+			): AuthCredentialIfAbsentResult => {
 				const existingRows = this.#listActiveByProviderStmt.all(providerName) as AuthRow[];
-				const existing: StoredAuthCredential[] = [];
+				const existing: Array<{
+					id: number;
+					credential: AuthCredential;
+					identityKey: string | null;
+				}> = [];
 				for (const row of existingRows) {
 					const activeCredential = deserializeCredential(row);
 					if (!activeCredential) continue;
-					existing.push(toStoredAuthCredential(row, activeCredential));
+					existing.push({
+						id: row.id,
+						credential: activeCredential,
+						identityKey: resolveRowCredentialIdentityKey(providerName, row),
+					});
 				}
 				if (existing.length > 0) {
-					return { inserted: false, reason: "skipped-existing", provider: providerName, entries: existing };
+					let targetId: number | null = null;
+					for (const row of existing) {
+						if (!matchesReplacementCredential(providerName, row.credential, row.identityKey, item)) continue;
+						if (targetId === null) {
+							targetId = row.id;
+							this.#updateStmt.run(record.credentialType, record.data, record.identityKey, row.id);
+						} else {
+							this.#deleteStmt.run("replaced by newer credential", row.id);
+						}
+					}
+
+					if (targetId !== null) {
+						return {
+							inserted: true,
+							reason: "updated-existing",
+							provider: providerName,
+							entries: this.listAuthCredentials(providerName),
+						};
+					}
+
+					return {
+						inserted: false,
+						reason: "skipped-existing",
+						provider: providerName,
+						entries: existing.map(row => ({
+							id: row.id,
+							provider: providerName,
+							credential: row.credential,
+							disabledCause: null,
+						})),
+					};
 				}
 
 				this.#insertStmt.get(providerName, record.credentialType, record.data, record.identityKey);
@@ -4118,7 +4181,9 @@ export class SqliteAuthCredentialStore implements AuthCredentialStore {
 			},
 		);
 
-		return writeIfAbsent.immediate(provider, serialized);
+		const result = writeIfAbsent.immediate(provider, credential, serialized);
+		if (result.inserted) this.#purgeSupersededDisabledRows(provider, result.entries);
+		return result;
 	}
 
 	/**
@@ -4215,6 +4280,13 @@ export class SqliteAuthCredentialStore implements AuthCredentialStore {
 		}
 	}
 
+	deleteCachePrefix(prefix: string): void {
+		if (prefix.length === 0) return;
+		try {
+			this.#deleteCachePrefixStmt.run(prefix.length, prefix);
+		} catch {}
+	}
+
 	cleanExpiredCache(): void {
 		try {
 			this.#deleteExpiredCacheStmt.run();
@@ -4305,6 +4377,7 @@ export class SqliteAuthCredentialStore implements AuthCredentialStore {
 		this.#getCacheStmt.finalize();
 		this.#getCacheIncludingExpiredStmt.finalize();
 		this.#upsertCacheStmt.finalize();
+		this.#deleteCachePrefixStmt.finalize();
 		this.#deleteExpiredCacheStmt.finalize();
 		this.#db.close();
 	}

package/src/provider-models/descriptors.ts

@@ -13,6 +13,7 @@ import {
 	anthropicModelManagerOptions,
 	cerebrasModelManagerOptions,
 	cloudflareAiGatewayModelManagerOptions,
+	deepinfraModelManagerOptions,
 	deepseekModelManagerOptions,
 	firepassModelManagerOptions,
 	fireworksModelManagerOptions,
@@ -168,6 +169,12 @@ export const PROVIDER_DESCRIPTORS: readonly ProviderDescriptor[] = [
 		config => deepseekModelManagerOptions(config),
 		catalog("DeepSeek", ["DEEPSEEK_API_KEY"]),
 	),
+	catalogDescriptor(
+		"deepinfra",
+		"deepseek-ai/DeepSeek-V3.2",
+		config => deepinfraModelManagerOptions(config),
+		catalog("DeepInfra", ["DEEPINFRA_API_KEY"]),
+	),
 	descriptor("mistral", "devstral-medium-latest", config => mistralModelManagerOptions(config)),
 	catalogDescriptor(
 		"nvidia",

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

@@ -665,6 +665,21 @@ export function deepseekModelManagerOptions(
 ): ModelManagerOptions<"openai-completions"> {
 	return createSimpleOpenAICompletionsOptions("deepseek", "https://api.deepseek.com", config);
 }
+export interface DeepInfraModelManagerConfig {
+	apiKey?: string;
+	baseUrl?: string;
+}
+
+export function deepinfraModelManagerOptions(
+	config?: DeepInfraModelManagerConfig,
+): ModelManagerOptions<"openai-completions"> {
+	return createSimpleOpenAICompletionsOptions(
+		"deepinfra" as Parameters<typeof getBundledModels>[0],
+		"https://api.deepinfra.com/v1/openai",
+		config,
+	);
+}
+
 // ---------------------------------------------------------------------------
 // 7.5 Fireworks
 // ---------------------------------------------------------------------------
@@ -2254,6 +2269,8 @@ const MODELS_DEV_PROVIDER_DESCRIPTORS_CORE: readonly ModelsDevProviderDescriptor
 	}),
 	// --- xAI ---
 	openAiCompletionsDescriptor("xai", "xai", "https://api.x.ai/v1"),
+	// --- DeepInfra ---
+	openAiCompletionsDescriptor("deepinfra", "deepinfra", "https://api.deepinfra.com/v1/openai"),
 	// --- DeepSeek ---
 	openAiCompletionsDescriptor("deepseek", "deepseek", "https://api.deepseek.com", {
 		// Only ship the v4 family as built-ins; older deepseek-chat / deepseek-reasoner

package/src/providers/anthropic.ts

@@ -2281,8 +2281,6 @@ function buildParams(
 					params.output_config = { effort } as typeof params.output_config;
 				}
 			}
-		} else if (options?.thinkingEnabled === false) {
-			params.thinking = { type: "disabled" };
 		}
 	}
 

package/src/providers/azure-openai-responses.ts

@@ -35,6 +35,7 @@ import {
 	markToolChoiceIncapability,
 	resolveToolChoice,
 } from "../utils/tool-choice-capability";
+import { wrapOpenAIFetchForBoundedRateLimits } from "./openai-bounded-rate-limits";
 import { normalizeOpenAIResponsesPromptCacheKey, supportsDeveloperRole } from "./openai-responses";
 import {
 	appendResponsesToolResultMessages,
@@ -272,7 +273,7 @@ function createClient(model: Model<"azure-openai-responses">, apiKey: string, op
 
 	const { baseUrl, apiVersion } = resolveAzureConfig(model, options);
 
-	const baseFetch = options?.fetch ?? fetch;
+	const baseFetch = wrapOpenAIFetchForBoundedRateLimits(options?.fetch ?? fetch, options?.maxRetryDelayMs);
 	const onSseEvent = options?.onSseEvent;
 	return new AzureOpenAI({
 		apiKey,

package/src/providers/cursor/client-version.ts

@@ -0,0 +1,10 @@
+/**
+ * The Cursor client version reported to api2.cursor.sh.
+ *
+ * Every call against the Cursor backend must send the same
+ * x-cursor-client-version: the backend gates features and minimum versions on
+ * it, so a drift between the agent Run path and model discovery makes one of
+ * them fail while the other keeps working. Keep this as the single source of
+ * truth for the header value.
+ */
+export const CURSOR_CLIENT_VERSION = "cli-2026.02.13-41ac335";

package/src/providers/cursor.ts

@@ -31,6 +31,7 @@ import { parseStreamingJson } from "../utils/json-parse";
 import { formatErrorMessageWithRetryAfter } from "../utils/retry-after";
 import { flattenToolRootCombinators, toolWireSchema } from "../utils/schema";
 import { COMPOSER_EDIT_DISCIPLINE_PROMPT, isComposerHarnessModel } from "./composer-discipline";
+import { CURSOR_CLIENT_VERSION } from "./cursor/client-version";
 import type { McpToolDefinition } from "./cursor/gen/agent_pb";
 import {
 	AgentClientMessageSchema,
@@ -131,7 +132,7 @@ import {
 } from "./cursor/gen/agent_pb";
 
 export const CURSOR_API_URL = "https://api2.cursor.sh";
-export const CURSOR_CLIENT_VERSION = "cli-2026.01.09-231024f";
+export { CURSOR_CLIENT_VERSION };
 
 const conversationStateCache = new Map<string, ConversationStateStructure>();
 const conversationBlobStores = new Map<string, Map<string, Uint8Array>>();

package/src/providers/google-gemini-headers.ts

@@ -4,7 +4,7 @@
  * GeminiCLI/VERSION/MODEL (PLATFORM; ARCH; SURFACE)
  */
 export function getGeminiCliUserAgent(modelId = "gemini-3.1-pro-preview"): string {
-	const version = process.env.PI_AI_GEMINI_CLI_VERSION || "0.46.0";
+	const version = process.env.SKC_AI_GEMINI_CLI_VERSION || process.env.PI_AI_GEMINI_CLI_VERSION || "0.49.0";
 	const platform = process.platform === "win32" ? "win32" : process.platform;
 	const arch = process.arch === "x64" ? "x64" : process.arch;
 	return `GeminiCLI/${version}/${modelId} (${platform}; ${arch}; terminal)`;

package/src/providers/openai-bounded-rate-limits.ts

@@ -0,0 +1,57 @@
+import type { FetchImpl } from "../types";
+import { getRetryAfterMsFromHeaders } from "../utils/retry-after";
+
+const OPENAI_RETRY_DELAY_CAP_MS = 60_000;
+
+// Mirror of `wrapAnthropicFetchForBoundedRateLimits`: OpenAI-compatible providers
+// (e.g. opencode-go) return HTTP 429 for *permanent* usage/quota exhaustion — a
+// monthly-cap reset that can be days away. The OpenAI SDK treats 429 as transient
+// and retries up to `maxRetries`, honoring an out-of-range `Retry-After`; the
+// `create()` call then hangs before the error can surface to the agent loop, so
+// no assistant error is produced and the session-level retry/fallback never runs.
+// Detect exhaustion and set `x-should-retry: false` so the SDK gives up at once
+// and the session retry layer applies its own fail-fast (retry-after > maxDelayMs).
+//
+// Shared by every adapter that drives a raw OpenAI SDK client — openai-completions,
+// openai-responses, and azure-openai-responses. Adapters that route through
+// `fetchWithRetry` (codex, bedrock, ollama, gemini-cli) already bound 429 retries
+// themselves and do not need this wrapper.
+export function isOpenAIUsageExhaustionResponse(
+	bodyText: string,
+	retryAfterMs: number | undefined,
+	retryDelayCapMs: number,
+): boolean {
+	if (retryAfterMs !== undefined && retryAfterMs > retryDelayCapMs) return true;
+	return /monthly usage limit|usage limit reached|usage_limit_reached|out_of_credits|insufficient_quota|quota[ _]?exceeded/i.test(
+		bodyText,
+	);
+}
+
+export function wrapOpenAIFetchForBoundedRateLimits(
+	baseFetch: FetchImpl,
+	maxRetryDelayMs: number | undefined,
+): FetchImpl {
+	const retryDelayCapMs = maxRetryDelayMs ?? OPENAI_RETRY_DELAY_CAP_MS;
+	return Object.assign(
+		async (input: string | URL | Request, init?: RequestInit): Promise<Response> => {
+			const response = await baseFetch(input, init);
+			if (response.status !== 429 || retryDelayCapMs === 0) return response;
+
+			const headers = new Headers(response.headers);
+			const retryAfterMs = getRetryAfterMsFromHeaders(headers);
+			const bodyText = await response
+				.clone()
+				.text()
+				.catch(() => "");
+			if (!isOpenAIUsageExhaustionResponse(bodyText, retryAfterMs, retryDelayCapMs)) return response;
+
+			headers.set("x-should-retry", "false");
+			return new Response(bodyText, {
+				status: response.status,
+				statusText: response.statusText,
+				headers,
+			});
+		},
+		baseFetch.preconnect ? { preconnect: baseFetch.preconnect } : {},
+	);
+}

package/src/providers/openai-completions-compat.ts

@@ -104,6 +104,10 @@ export function detectOpenAICompat(model: Model<"openai-completions">, resolvedB
 		baseUrl.includes("opencode.ai");
 	const isOpenCodeProvider = provider === "opencode-go" || provider === "opencode-zen";
 	const isOpenCodeGoReasoning = provider === "opencode-go" && Boolean(model.reasoning);
+	const isOpenCodeGoKimiReasoning = provider === "opencode-go" && isKimiModel && Boolean(model.reasoning);
+	const isOpenCodeGoKimi25Reasoning = isOpenCodeGoKimiReasoning && model.id === "kimi-k2.5";
+	const isOpenCodeGoKimi27CodeReasoning = isOpenCodeGoKimiReasoning && model.id === "kimi-k2.7-code";
+	const needsOpenCodeGoKimiEffortMap = isOpenCodeGoKimi25Reasoning || isOpenCodeGoKimi27CodeReasoning;
 
 	const useMaxTokens =
 		provider === "mistral" ||
@@ -170,22 +174,31 @@ export function detectOpenAICompat(model: Model<"openai-completions">, resolvedB
 					xhigh: "default",
 					max: "default",
 				} satisfies Partial<Record<OpenAIReasoningEffort, string>>)
-			: isDeepseekFamily && model.reasoning
+			: needsOpenCodeGoKimiEffortMap
 				? ({
-						minimal: "high",
-						low: "high",
-						medium: "high",
-						high: "high",
-						xhigh: "max",
-						max: "max",
+						// Live Go probes (2026-07-06) showed model-specific effort gaps:
+						// kimi-k2.5 rejects "minimal", while kimi-k2.7-code rejects
+						// OpenAI-style "xhigh" and "max"; all other Kimi efforts tested
+						// successfully and should pass through unchanged.
+						...(isOpenCodeGoKimi25Reasoning ? { minimal: "low" } : {}),
+						...(isOpenCodeGoKimi27CodeReasoning ? { xhigh: "high", max: "high" } : {}),
 					} satisfies Partial<Record<OpenAIReasoningEffort, string>>)
-				: isFireworks
+				: isDeepseekFamily && model.reasoning
 					? ({
-							// Fireworks' OpenAI-compatible endpoint rejects OpenAI's
-							// `minimal` literal but accepts `none` for the lowest setting.
-							minimal: "none",
+							minimal: "high",
+							low: "high",
+							medium: "high",
+							high: "high",
+							xhigh: "max",
+							max: "max",
 						} satisfies Partial<Record<OpenAIReasoningEffort, string>>)
-					: {};
+					: isFireworks
+						? ({
+								// Fireworks' OpenAI-compatible endpoint rejects OpenAI's
+								// `minimal` literal but accepts `none` for the lowest setting.
+								minimal: "none",
+							} satisfies Partial<Record<OpenAIReasoningEffort, string>>)
+						: {};
 
 	return {
 		supportsStore: !isNonStandard,
@@ -198,7 +211,7 @@ export function detectOpenAICompat(model: Model<"openai-completions">, resolvedB
 		disableReasoningOnForcedToolChoice: isKimiModel || isAnthropicModel || isOpenCodeGoReasoning,
 		disableReasoningOnToolChoice: isDeepseekFamily && Boolean(model.reasoning) && !isOpenRouter,
 		supportsToolChoice: !isDirectDeepseekReasoning,
-		supportsForcedToolChoice: true,
+		supportsForcedToolChoice: !isOpenCodeGoKimiReasoning,
 		maxTokensField: useMaxTokens ? "max_tokens" : "max_completion_tokens",
 		requiresToolResultName: isMistral,
 		requiresAssistantAfterToolResult: false,

package/src/providers/openai-completions.ts

@@ -35,7 +35,7 @@ import {
 	type ToolChoice,
 	type ToolResultMessage,
 } from "../types";
-import { normalizeSystemPrompts } from "../utils";
+import { normalizeSystemPrompts, sanitizeJsonStrings } from "../utils";
 import { createAbortSourceTracker } from "../utils/abort";
 import { AssistantMessageEventStream } from "../utils/event-stream";
 import { toFirepassWireModelId, toFireworksWireModelId } from "../utils/fireworks-model-id";
@@ -166,7 +166,7 @@ function normalizeStreamingContentText(content: unknown): string {
 function serializeToolArguments(value: unknown): string {
 	if (value && typeof value === "object" && !Array.isArray(value)) {
 		try {
-			return JSON.stringify(value);
+			return JSON.stringify(sanitizeJsonStrings(value));
 		} catch {
 			return "{}";
 		}
@@ -178,7 +178,7 @@ function serializeToolArguments(value: unknown): string {
 		try {
 			const parsed = JSON.parse(trimmed);
 			if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
-				return JSON.stringify(parsed);
+				return JSON.stringify(sanitizeJsonStrings(parsed));
 			}
 		} catch {}
 		return "{}";

package/src/providers/openai-responses.ts

@@ -71,6 +71,7 @@ import {
 	resolveGitHubCopilotBaseUrl,
 } from "./github-copilot-headers";
 import { compactGrammarDefinition } from "./grammar";
+import { wrapOpenAIFetchForBoundedRateLimits } from "./openai-bounded-rate-limits";
 import {
 	applyOpenAIRequestTransformBody,
 	applyOpenAIRequestTransformHeaders,
@@ -274,6 +275,7 @@ export const streamOpenAIResponses: StreamFunction<"openai-responses"> = (
 				options?.fetch,
 				options?.authCredentialType,
 				options?.requestMaxRetries,
+				options?.maxRetryDelayMs,
 			);
 			const premiumRequestsTotal = copilotPremiumRequests;
 			const providerSessionState = getOpenAIResponsesProviderSessionState(model, options?.providerSessionState);
@@ -400,6 +402,7 @@ function createClient(
 	fetchOverride?: FetchImpl,
 	authCredentialType?: OpenAIResponsesOptions["authCredentialType"],
 	requestMaxRetries?: number,
+	maxRetryDelayMs?: number,
 ): {
 	client: OpenAI;
 	copilotPremiumRequests: number | undefined;
@@ -446,8 +449,9 @@ function createClient(
 		headers["x-client-request-id"] ??= sessionId;
 	}
 	const baseFetch = fetchOverride ?? fetch;
+	const boundedFetch = wrapOpenAIFetchForBoundedRateLimits(baseFetch, maxRetryDelayMs);
 	const transformedFetch = wrapFetchForOpenAIRequestTransform(
-		baseFetch,
+		boundedFetch,
 		model.requestTransform,
 		`Sayknow-CLI/${packageJson.version}`,
 	);

package/src/rate-limit-utils.ts

@@ -13,7 +13,7 @@ export type RateLimitReason =
 const QUOTA_EXHAUSTED_BACKOFF_MS = 30 * 60 * 1000; // 30 min
 const RATE_LIMIT_EXCEEDED_BACKOFF_MS = 30 * 1000; // 30s
 const MODEL_CAPACITY_BASE_MS = 45 * 1000; // 45s base
-const MODEL_CAPACITY_JITTER_MS = 30 * 1000; // ±15s
+const MODEL_CAPACITY_JITTER_MS = 30 * 1000; // uniform +0–30s above base → 45–75s total
 const SERVER_ERROR_BACKOFF_MS = 20 * 1000; // 20s
 
 /**
@@ -93,8 +93,10 @@ export function calculateRateLimitBackoffMs(reason: RateLimitReason): number {
 }
 
 /** Detect usage/quota limit errors in error messages (persistent, requires credential switch). */
+// ZAI reports durable token exhaustion as "[1310][Weekly/Monthly Limit Exhausted...]".
+// Keep this explicit so generic "rate limit exhausted, retry..." throttles remain retryable.
 const USAGE_LIMIT_PATTERN =
-	/usage.?limit|usage_limit_reached|usage_not_included|limit_reached|model.?limit|model_limit_reached|message.?limit|message_limit_reached|limit for this model|quota.?exceeded|out_of_credits|request would exceed your account.?s rate limit|resource has been exhausted[^\n]*(?:quota|limit)/i;
+	/usage.?limit|usage_limit_reached|usage_not_included|limit_reached|model.?limit|model_limit_reached|message.?limit|message_limit_reached|limit for this model|weekly\/monthly\s+limit\s+exhausted|quota.?exceeded|out_of_credits|request would exceed your account.?s rate limit|resource has been exhausted[^\n]*(?:quota|limit)/i;
 export function isUsageLimitError(errorMessage: string): boolean {
 	return USAGE_LIMIT_PATTERN.test(errorMessage);
 }

package/src/stream.ts

@@ -98,6 +98,7 @@ const serviceProviderMap: Record<string, KeyResolver> = {
 	"opencode-zen": "OPENCODE_API_KEY",
 	cursor: "CURSOR_ACCESS_TOKEN",
 	deepseek: "DEEPSEEK_API_KEY",
+	deepinfra: "DEEPINFRA_API_KEY",
 	"openai-codex": "OPENAI_CODEX_OAUTH_TOKEN",
 	"azure-openai": "AZURE_OPENAI_API_KEY",
 	"azure-openai-responses": "AZURE_OPENAI_API_KEY",

package/src/types.ts

@@ -116,6 +116,7 @@ export type KnownProvider =
 	| "gitlab-duo"
 	| "cursor"
 	| "deepseek"
+	| "deepinfra"
 	| "xai"
 	| "groq"
 	| "cerebras"
@@ -224,17 +225,20 @@ export function resolveServiceTier(
 
 /**
  * True when the (possibly scoped) tier should be sent as OpenAI's
- * `service_tier` request field for the given provider. Non-OpenAI
- * providers, unsupported tiers (`"auto"`, `"default"`), and scope
- * mismatches all return false.
+ * `service_tier` request field for the given provider. OpenAI accepts
+ * `flex`, `scale`, and `priority`; DeepInfra accepts `priority`.
+ * Unsupported tiers (`"auto"`, `"default"`) and scope mismatches return false.
  */
 export function shouldSendServiceTier(
 	serviceTier: ServiceTier | null | undefined,
 	provider: Provider | undefined,
 ): boolean {
-	if (provider !== "openai" && provider !== "openai-codex") return false;
 	const resolved = resolveServiceTier(serviceTier, provider);
-	return resolved === "flex" || resolved === "scale" || resolved === "priority";
+	if (provider === "openai" || provider === "openai-codex") {
+		return resolved === "flex" || resolved === "scale" || resolved === "priority";
+	}
+	if (provider === "deepinfra") return resolved === "priority";
+	return false;
 }
 
 /**
@@ -252,7 +256,9 @@ export function getPriorityPremiumRequests(
 	if (resolveServiceTier(serviceTier, provider) !== "priority") return 0;
 	// Only providers that realize `priority` on the wire bill the user.
 	// Everywhere else, the field is silently dropped and nothing is charged.
-	return provider === "openai" || provider === "openai-codex" || provider === "anthropic" ? 1 : 0;
+	return provider === "openai" || provider === "openai-codex" || provider === "anthropic" || provider === "deepinfra"
+		? 1
+		: 0;
 }
 
 export interface ProviderSessionState {

package/src/utils/discovery/cursor.ts

@@ -2,11 +2,11 @@ import * as http2 from "node:http2";
 import { create, fromBinary, toBinary } from "@bufbuild/protobuf";
 import * as z from "zod/v4";
 import { getBundledModels } from "../../models";
+import { CURSOR_CLIENT_VERSION } from "../../providers/cursor/client-version";
 import { GetUsableModelsRequestSchema, GetUsableModelsResponseSchema } from "../../providers/cursor/gen/agent_pb";
 import type { Model } from "../../types";
 
 const CURSOR_DEFAULT_BASE_URL = "https://api2.cursor.sh";
-const CURSOR_DEFAULT_CLIENT_VERSION = "cli-2026.02.13-41ac335";
 const CURSOR_GET_USABLE_MODELS_PATH = "/agent.v1.AgentService/GetUsableModels";
 
 const DEFAULT_CONTEXT_WINDOW = 200_000;
@@ -91,7 +91,7 @@ function buildRequestHeaders(options: CursorModelDiscoveryOptions): Record<strin
 		te: "trailers",
 		authorization: `Bearer ${options.apiKey}`,
 		"x-ghost-mode": "true",
-		"x-cursor-client-version": options.clientVersion ?? CURSOR_DEFAULT_CLIENT_VERSION,
+		"x-cursor-client-version": options.clientVersion ?? CURSOR_CLIENT_VERSION,
 		"x-cursor-client-type": "cli",
 	};
 }

package/src/utils/oauth/fugu.ts

@@ -3,10 +3,10 @@ import { createApiKeyLogin } from "./api-key-login";
 
 export const loginFugu = createApiKeyLogin({
 	providerLabel: "Sakana Fugu",
-	authUrl: "https://fugu.sakana.ai/",
+	authUrl: "https://console.sakana.ai/api-keys",
 	instructions: "Create or copy your Sakana Fugu API key",
 	promptMessage: "Paste your Sakana Fugu API key",
-	placeholder: "fugu_...",
+	placeholder: "fish_...",
 	validation: {
 		kind: "models-endpoint",
 		provider: "Sakana Fugu",

package/src/utils/oauth/index.ts

@@ -60,6 +60,11 @@ const builtInOAuthProviders: OAuthProviderInfo[] = [
 		name: "DeepSeek",
 		available: true,
 	},
+	{
+		id: "deepinfra",
+		name: "DeepInfra",
+		available: true,
+	},
 	{
 		id: "xai",
 		name: "xAI",
@@ -359,6 +364,7 @@ export async function refreshOAuthToken(
 		case "fireworks":
 		case "firepass":
 		case "fugu":
+		case "deepinfra":
 		case "nvidia":
 		case "nanogpt":
 		case "synthetic":

package/src/utils/oauth/types.ts

@@ -15,6 +15,7 @@ export type OAuthProvider =
 	| "cloudflare-ai-gateway"
 	| "cursor"
 	| "deepseek"
+	| "deepinfra"
 	| "fireworks"
 	| "firepass"
 	| "fugu"

package/src/utils.ts

@@ -215,7 +215,11 @@ function sanitizeOpenAIResponsesHistoryItemForReplay(
 	// providerPayload stores raw output items; replay strips fields that are output-only.
 	const { id: _id, ...itemWithoutId } = item;
 	const sanitizedItem =
-		item.type === "computer_call" ? sanitizeComputerCallForResponsesInput(itemWithoutId) : itemWithoutId;
+		item.type === "computer_call"
+			? sanitizeComputerCallForResponsesInput(itemWithoutId)
+			: item.type === "image_generation_call"
+				? sanitizeImageGenerationCallForResponsesInput(itemWithoutId)
+				: itemWithoutId;
 	if (typeof item.call_id === "string") {
 		sanitizedItem.call_id = normalizeReplayedResponsesHistoryCallId(item.call_id, normalizedCallIds);
 	}
@@ -231,6 +235,22 @@ function sanitizeComputerCallForResponsesInput(item: Record<string, unknown>): R
 	return inputSafeItem;
 }
 
+function sanitizeImageGenerationCallForResponsesInput(item: Record<string, unknown>): Record<string, unknown> {
+	// Image generation output items include request-time knobs that are not part of
+	// the Responses input replay schema. Replaying them verbatim makes OpenAI-compatible
+	// endpoints reject the next turn, e.g. `Unknown parameter: input[n].action`.
+	const {
+		action: _action,
+		background: _background,
+		output_format: _outputFormat,
+		quality: _quality,
+		revised_prompt: _revisedPrompt,
+		size: _size,
+		...inputSafeItem
+	} = item;
+	return inputSafeItem;
+}
+
 function normalizeReplayedResponsesHistoryCallId(value: string, normalizedValues: Map<string, string>): string {
 	const normalized = normalizedValues.get(value);
 	if (normalized) return normalized;