@kenkaiiii/gg-core 5.89.0 → 5.91.0

package/dist/index.d.cts

@@ -27,23 +27,6 @@ declare const PLAN_MODE_THINKING_CAP: ThinkingLevel;
  * `undefined` (thinking off) and already-low levels. */
 declare function clampThinkingForPlanMode(level: ThinkingLevel | undefined): ThinkingLevel | undefined;
 
-/**
- * Local model discovery — Ollama, LM Studio, llama.cpp (`llama-server`), vLLM,
- * and any other OpenAI-compatible server the user points us at.
- *
- * Everything rides the OpenAI-compatible `/v1` transport (see the `local`
- * provider in gg-ai's stream.ts); the only per-server difference is where the
- * *capabilities* come from, because `GET /v1/models` reports nothing useful:
- *
- *   - Ollama    → `/api/show` + `/api/ps` → capabilities + running context
- *   - LM Studio → `/api/v1/models` (v0 fallback) → capabilities + loaded context
- *   - llama.cpp → `GET /props`          → `default_generation_settings.n_ctx`
- *   - vLLM/other → nothing; `max_model_len` sometimes rides the model object.
- *
- * Probing never throws: an unreachable server is a normal state (the user just
- * doesn't have it running), not an error to surface.
- */
-
 /** Which capability API a local endpoint speaks, beyond plain `/v1/models`. */
 type LocalEndpointKind = "ollama" | "lmstudio" | "llamacpp" | "vllm" | "custom";
 interface LocalEndpoint {
@@ -91,7 +74,7 @@ declare const DEFAULT_LOCAL_ENDPOINTS: readonly LocalEndpoint[];
  * auto-compaction already sailed past, while under-guessing only compacts early.
  */
 declare const FALLBACK_CONTEXT_WINDOW = 4096;
-/** Placeholder token for endpoints with no key — these servers ignore it. */
+/** Token stored for endpoints with no key; never sent on the wire. */
 declare const LOCAL_API_KEY_PLACEHOLDER = "local";
 /**
  * `local/<endpointId>/<rawModelId>`. The raw id can itself contain slashes
@@ -102,6 +85,11 @@ declare function parseLocalModelId(id: string): {
     endpointId: string;
     rawId: string;
 } | undefined;
+/**
+ * The model's own name, for a UI that shows the endpoint separately: the last
+ * segment of the raw id, so `minimax/minimax-m2:free` reads `minimax-m2:free`.
+ */
+declare function localModelShortName(rawId: string): string;
 declare function isLocalModelId(id: string): boolean;
 /** Auth-storage key holding the credential (and baseUrl) for one local endpoint. */
 declare function localAuthStorageKey(endpointId: string): string;
@@ -117,6 +105,17 @@ interface ProbeOptions {
  * error toast.
  */
 declare function probeEndpoint(endpoint: LocalEndpoint, { timeoutMs, signal }?: ProbeOptions): Promise<LocalEndpointProbe>;
+/**
+ * The context Ollama serves `rawId` with, read from `/api/ps`. With `load`, first
+ * loads the model the way a chat request would (default options), since an
+ * unloaded model's size is only decided at load time: the OpenAI endpoint
+ * ignores `num_ctx`, so this is the window the agent will really get.
+ */
+declare function ollamaServedContext(endpoint: LocalEndpoint, rawId: string, { load, timeoutMs, signal, }?: {
+    load?: boolean;
+    timeoutMs?: number;
+    signal?: AbortSignal;
+}): Promise<number | undefined>;
 /** Convert a probed local model into the registry shape the whole app speaks. */
 declare function toModelInfo(model: LocalModel, endpoint: LocalEndpoint): ModelInfo;
 interface DiscoveryResult {
@@ -305,7 +304,12 @@ declare class AuthStorage {
     private snapshotIno;
     /** Per-provider lock to serialize concurrent refresh calls. */
     private refreshLocks;
-    constructor(filePath?: string);
+    private readonly detectMiniMaxRegion;
+    private readonly detectXiaomiCluster;
+    constructor(filePath?: string, options?: {
+        detectMiniMaxBaseUrl?: (apiKey: string) => Promise<string | undefined>;
+        detectXiaomiTokenPlanBaseUrl?: (apiKey: string) => Promise<string | undefined>;
+    });
     /** Path to the on-disk auth file. Useful for status output. */
     get path(): string;
     /** List provider keys with stored credentials. */
@@ -380,7 +384,22 @@ declare class AuthStorage {
     private mutateLatest;
     private reloadLatest;
     getCredentials(provider: string): Promise<OAuthCredentials | undefined>;
+    /**
+     * A MiniMax key saved without an endpoint (every key, until its first use)
+     * gets the endpoint of the region that issued it, saved so the check runs
+     * once. Without this a Mainland China key went to the Global API, got a 401,
+     * and was deleted as invalid.
+     */
+    private withMiniMaxRegion;
     setCredentials(provider: string, creds: OAuthCredentials): Promise<void>;
+    /**
+     * After a 401, move a regional key to the cluster that issued it instead of
+     * treating it as invalid. Only the Xiaomi Token Plan is regional: every key
+     * used to go to Singapore, so Europe and China keys were rejected and deleted.
+     * Returns the new base URL, or undefined when no other cluster accepts the key.
+     */
+    relocateRegionalApiKey(storageKeys: string[]): Promise<string | undefined>;
+    clearRejectedApiKey(storageKey: string): Promise<boolean>;
     clearCredentials(provider: string): Promise<void>;
     /**
      * Mark the credential stored under `storageKey` as usage-exhausted until
@@ -430,6 +449,41 @@ declare class NotLoggedInError extends Error {
     constructor(provider: string);
 }
 
+/**
+ * MiniMax runs two separate platforms, and a key only works on the one that
+ * issued it: Global (platform.minimax.io) or Mainland China
+ * (platform.minimaxi.com). The user's location doesn't matter. Sending a key to
+ * the other region's API is rejected as an invalid key.
+ */
+declare const MINIMAX_GLOBAL_BASE_URL = "https://api.minimax.io/anthropic";
+declare const MINIMAX_CN_BASE_URL = "https://api.minimax.cn/anthropic";
+/**
+ * The Anthropic-compatible base URL for the region that issued `apiKey`, found
+ * by asking both regions about the key. Undefined when neither region gives a
+ * clear answer (offline, timeouts) — callers then keep the Global default and
+ * may ask again later.
+ */
+declare function detectMiniMaxBaseUrl(apiKey: string, options?: {
+    fetch?: typeof fetch;
+    timeoutMs?: number;
+}): Promise<string | undefined>;
+
+/**
+ * Xiaomi MiMo Token Plan runs separate clusters, and a `tp-` key only works on
+ * the cluster that issued it (shown on the user's Token Plan page). Any other
+ * cluster answers 401 "Invalid API Key". Pay-as-you-go (API Credits) keys use
+ * the single global host and are unaffected.
+ */
+declare const XIAOMI_TOKEN_PLAN_BASE_URLS: readonly ["https://token-plan-sgp.xiaomimimo.com/v1", "https://token-plan-ams.xiaomimimo.com/v1", "https://token-plan-cn.xiaomimimo.com/v1"];
+/**
+ * The Token Plan base URL whose cluster accepts `apiKey`, found by listing
+ * models (free, auth-only) on every cluster. Undefined when none accepts it.
+ */
+declare function detectXiaomiTokenPlanBaseUrl(apiKey: string, options?: {
+    fetch?: typeof fetch;
+    timeoutMs?: number;
+}): Promise<string | undefined>;
+
 type AuthMethod = "oauth" | "apikey";
 /**
  * One API-key option for a provider that splits auth across multiple distinct
@@ -836,4 +890,4 @@ interface AutoUpdater {
 }
 declare function createAutoUpdater(config: AutoUpdateConfig): AutoUpdater;
 
-export { AUTH_PROVIDERS, type ApiKeyVariant, type AuthMethod, type AuthMethodMeta, type AuthProviderMeta, AuthStorage, type AutoUpdateConfig, type AutoUpdater, DEFAULT_LOCAL_ENDPOINTS, type DiscoverOptions, type DiscoveryResult, type DualAuthProvider, FALLBACK_CONTEXT_WINDOW, GEMINI_API_KEY, GEMINI_API_KEY_BASE_URL, type InlineButton, LOCAL_API_KEY_PLACEHOLDER, LOCAL_AUTH_KEY_PREFIX, type LocalEndpoint, type LocalEndpointKind, type LocalEndpointProbe, type LocalModel, type LogLevel, MOONSHOT_OAUTH_KEY, ModelInfo, NotLoggedInError, type OAuthCredentials, type OAuthLoginCallbacks, PLAN_MODE_THINKING_CAP, type ProbeOptions, type ProgressCallback, SubscriptionUsageError, type SubscriptionUsageProvider, type SubscriptionUsageSnapshot, type SubscriptionUsageWindow, TelegramBot, type TelegramConfig, type TelegramMessage, type TelegramUpdate, type TelegramVoiceMessage, XAI_OAUTH_KEY, XIAOMI_CREDITS_KEY, authPriorityNote, clampThinkingForPlanMode, clearLocalDiscoveryCache, closeLogger, createAutoUpdater, decodeOggOpus, describeAuthMethods, discoverLocalModels, downmixToMono, dualAuthProvider, dualAuthProviderByOAuthKey, endpointRoot, fetchSubscriptionUsage, findProbedModel, formatLocalModelId, generatePKCE, getAuthProvider, getLowestThinkingLevel, getNextThinkingLevel, getSessionId, getSupportedThinkingLevels, grokCliBaseUrl, grokCliHeaders, isGrokCliEndpoint, isKimiCodingEndpoint, isLocalModelId, isLoggerOpen, isModelLoaded, isThinkingLevelSupported, kimiCodeBaseUrl, kimiCodingHeaders, localAuthStorageKey, log, loginAnthropic, loginGemini, loginKimi, loginOpenAI, loginXai, oauthStorageKey, openLog, parseLocalModelId, probeEndpoint, providerStorageKeys, readStoredBaseUrlSync, refreshAnthropicToken, refreshGeminiToken, refreshKimiToken, refreshOpenAIToken, refreshXaiToken, registerLogCleanup, resample, setProgressCallback, toModelInfo, transcribeVoice, withFileLock };
+export { AUTH_PROVIDERS, type ApiKeyVariant, type AuthMethod, type AuthMethodMeta, type AuthProviderMeta, AuthStorage, type AutoUpdateConfig, type AutoUpdater, DEFAULT_LOCAL_ENDPOINTS, type DiscoverOptions, type DiscoveryResult, type DualAuthProvider, FALLBACK_CONTEXT_WINDOW, GEMINI_API_KEY, GEMINI_API_KEY_BASE_URL, type InlineButton, LOCAL_API_KEY_PLACEHOLDER, LOCAL_AUTH_KEY_PREFIX, type LocalEndpoint, type LocalEndpointKind, type LocalEndpointProbe, type LocalModel, type LogLevel, MINIMAX_CN_BASE_URL, MINIMAX_GLOBAL_BASE_URL, MOONSHOT_OAUTH_KEY, ModelInfo, NotLoggedInError, type OAuthCredentials, type OAuthLoginCallbacks, PLAN_MODE_THINKING_CAP, type ProbeOptions, type ProgressCallback, SubscriptionUsageError, type SubscriptionUsageProvider, type SubscriptionUsageSnapshot, type SubscriptionUsageWindow, TelegramBot, type TelegramConfig, type TelegramMessage, type TelegramUpdate, type TelegramVoiceMessage, XAI_OAUTH_KEY, XIAOMI_CREDITS_KEY, XIAOMI_TOKEN_PLAN_BASE_URLS, authPriorityNote, clampThinkingForPlanMode, clearLocalDiscoveryCache, closeLogger, createAutoUpdater, decodeOggOpus, describeAuthMethods, detectMiniMaxBaseUrl, detectXiaomiTokenPlanBaseUrl, discoverLocalModels, downmixToMono, dualAuthProvider, dualAuthProviderByOAuthKey, endpointRoot, fetchSubscriptionUsage, findProbedModel, formatLocalModelId, generatePKCE, getAuthProvider, getLowestThinkingLevel, getNextThinkingLevel, getSessionId, getSupportedThinkingLevels, grokCliBaseUrl, grokCliHeaders, isGrokCliEndpoint, isKimiCodingEndpoint, isLocalModelId, isLoggerOpen, isModelLoaded, isThinkingLevelSupported, kimiCodeBaseUrl, kimiCodingHeaders, localAuthStorageKey, localModelShortName, log, loginAnthropic, loginGemini, loginKimi, loginOpenAI, loginXai, oauthStorageKey, ollamaServedContext, openLog, parseLocalModelId, probeEndpoint, providerStorageKeys, readStoredBaseUrlSync, refreshAnthropicToken, refreshGeminiToken, refreshKimiToken, refreshOpenAIToken, refreshXaiToken, registerLogCleanup, resample, setProgressCallback, toModelInfo, transcribeVoice, withFileLock };

package/dist/index.d.ts

@@ -27,23 +27,6 @@ declare const PLAN_MODE_THINKING_CAP: ThinkingLevel;
  * `undefined` (thinking off) and already-low levels. */
 declare function clampThinkingForPlanMode(level: ThinkingLevel | undefined): ThinkingLevel | undefined;
 
-/**
- * Local model discovery — Ollama, LM Studio, llama.cpp (`llama-server`), vLLM,
- * and any other OpenAI-compatible server the user points us at.
- *
- * Everything rides the OpenAI-compatible `/v1` transport (see the `local`
- * provider in gg-ai's stream.ts); the only per-server difference is where the
- * *capabilities* come from, because `GET /v1/models` reports nothing useful:
- *
- *   - Ollama    → `/api/show` + `/api/ps` → capabilities + running context
- *   - LM Studio → `/api/v1/models` (v0 fallback) → capabilities + loaded context
- *   - llama.cpp → `GET /props`          → `default_generation_settings.n_ctx`
- *   - vLLM/other → nothing; `max_model_len` sometimes rides the model object.
- *
- * Probing never throws: an unreachable server is a normal state (the user just
- * doesn't have it running), not an error to surface.
- */
-
 /** Which capability API a local endpoint speaks, beyond plain `/v1/models`. */
 type LocalEndpointKind = "ollama" | "lmstudio" | "llamacpp" | "vllm" | "custom";
 interface LocalEndpoint {
@@ -91,7 +74,7 @@ declare const DEFAULT_LOCAL_ENDPOINTS: readonly LocalEndpoint[];
  * auto-compaction already sailed past, while under-guessing only compacts early.
  */
 declare const FALLBACK_CONTEXT_WINDOW = 4096;
-/** Placeholder token for endpoints with no key — these servers ignore it. */
+/** Token stored for endpoints with no key; never sent on the wire. */
 declare const LOCAL_API_KEY_PLACEHOLDER = "local";
 /**
  * `local/<endpointId>/<rawModelId>`. The raw id can itself contain slashes
@@ -102,6 +85,11 @@ declare function parseLocalModelId(id: string): {
     endpointId: string;
     rawId: string;
 } | undefined;
+/**
+ * The model's own name, for a UI that shows the endpoint separately: the last
+ * segment of the raw id, so `minimax/minimax-m2:free` reads `minimax-m2:free`.
+ */
+declare function localModelShortName(rawId: string): string;
 declare function isLocalModelId(id: string): boolean;
 /** Auth-storage key holding the credential (and baseUrl) for one local endpoint. */
 declare function localAuthStorageKey(endpointId: string): string;
@@ -117,6 +105,17 @@ interface ProbeOptions {
  * error toast.
  */
 declare function probeEndpoint(endpoint: LocalEndpoint, { timeoutMs, signal }?: ProbeOptions): Promise<LocalEndpointProbe>;
+/**
+ * The context Ollama serves `rawId` with, read from `/api/ps`. With `load`, first
+ * loads the model the way a chat request would (default options), since an
+ * unloaded model's size is only decided at load time: the OpenAI endpoint
+ * ignores `num_ctx`, so this is the window the agent will really get.
+ */
+declare function ollamaServedContext(endpoint: LocalEndpoint, rawId: string, { load, timeoutMs, signal, }?: {
+    load?: boolean;
+    timeoutMs?: number;
+    signal?: AbortSignal;
+}): Promise<number | undefined>;
 /** Convert a probed local model into the registry shape the whole app speaks. */
 declare function toModelInfo(model: LocalModel, endpoint: LocalEndpoint): ModelInfo;
 interface DiscoveryResult {
@@ -305,7 +304,12 @@ declare class AuthStorage {
     private snapshotIno;
     /** Per-provider lock to serialize concurrent refresh calls. */
     private refreshLocks;
-    constructor(filePath?: string);
+    private readonly detectMiniMaxRegion;
+    private readonly detectXiaomiCluster;
+    constructor(filePath?: string, options?: {
+        detectMiniMaxBaseUrl?: (apiKey: string) => Promise<string | undefined>;
+        detectXiaomiTokenPlanBaseUrl?: (apiKey: string) => Promise<string | undefined>;
+    });
     /** Path to the on-disk auth file. Useful for status output. */
     get path(): string;
     /** List provider keys with stored credentials. */
@@ -380,7 +384,22 @@ declare class AuthStorage {
     private mutateLatest;
     private reloadLatest;
     getCredentials(provider: string): Promise<OAuthCredentials | undefined>;
+    /**
+     * A MiniMax key saved without an endpoint (every key, until its first use)
+     * gets the endpoint of the region that issued it, saved so the check runs
+     * once. Without this a Mainland China key went to the Global API, got a 401,
+     * and was deleted as invalid.
+     */
+    private withMiniMaxRegion;
     setCredentials(provider: string, creds: OAuthCredentials): Promise<void>;
+    /**
+     * After a 401, move a regional key to the cluster that issued it instead of
+     * treating it as invalid. Only the Xiaomi Token Plan is regional: every key
+     * used to go to Singapore, so Europe and China keys were rejected and deleted.
+     * Returns the new base URL, or undefined when no other cluster accepts the key.
+     */
+    relocateRegionalApiKey(storageKeys: string[]): Promise<string | undefined>;
+    clearRejectedApiKey(storageKey: string): Promise<boolean>;
     clearCredentials(provider: string): Promise<void>;
     /**
      * Mark the credential stored under `storageKey` as usage-exhausted until
@@ -430,6 +449,41 @@ declare class NotLoggedInError extends Error {
     constructor(provider: string);
 }
 
+/**
+ * MiniMax runs two separate platforms, and a key only works on the one that
+ * issued it: Global (platform.minimax.io) or Mainland China
+ * (platform.minimaxi.com). The user's location doesn't matter. Sending a key to
+ * the other region's API is rejected as an invalid key.
+ */
+declare const MINIMAX_GLOBAL_BASE_URL = "https://api.minimax.io/anthropic";
+declare const MINIMAX_CN_BASE_URL = "https://api.minimax.cn/anthropic";
+/**
+ * The Anthropic-compatible base URL for the region that issued `apiKey`, found
+ * by asking both regions about the key. Undefined when neither region gives a
+ * clear answer (offline, timeouts) — callers then keep the Global default and
+ * may ask again later.
+ */
+declare function detectMiniMaxBaseUrl(apiKey: string, options?: {
+    fetch?: typeof fetch;
+    timeoutMs?: number;
+}): Promise<string | undefined>;
+
+/**
+ * Xiaomi MiMo Token Plan runs separate clusters, and a `tp-` key only works on
+ * the cluster that issued it (shown on the user's Token Plan page). Any other
+ * cluster answers 401 "Invalid API Key". Pay-as-you-go (API Credits) keys use
+ * the single global host and are unaffected.
+ */
+declare const XIAOMI_TOKEN_PLAN_BASE_URLS: readonly ["https://token-plan-sgp.xiaomimimo.com/v1", "https://token-plan-ams.xiaomimimo.com/v1", "https://token-plan-cn.xiaomimimo.com/v1"];
+/**
+ * The Token Plan base URL whose cluster accepts `apiKey`, found by listing
+ * models (free, auth-only) on every cluster. Undefined when none accepts it.
+ */
+declare function detectXiaomiTokenPlanBaseUrl(apiKey: string, options?: {
+    fetch?: typeof fetch;
+    timeoutMs?: number;
+}): Promise<string | undefined>;
+
 type AuthMethod = "oauth" | "apikey";
 /**
  * One API-key option for a provider that splits auth across multiple distinct
@@ -836,4 +890,4 @@ interface AutoUpdater {
 }
 declare function createAutoUpdater(config: AutoUpdateConfig): AutoUpdater;
 
-export { AUTH_PROVIDERS, type ApiKeyVariant, type AuthMethod, type AuthMethodMeta, type AuthProviderMeta, AuthStorage, type AutoUpdateConfig, type AutoUpdater, DEFAULT_LOCAL_ENDPOINTS, type DiscoverOptions, type DiscoveryResult, type DualAuthProvider, FALLBACK_CONTEXT_WINDOW, GEMINI_API_KEY, GEMINI_API_KEY_BASE_URL, type InlineButton, LOCAL_API_KEY_PLACEHOLDER, LOCAL_AUTH_KEY_PREFIX, type LocalEndpoint, type LocalEndpointKind, type LocalEndpointProbe, type LocalModel, type LogLevel, MOONSHOT_OAUTH_KEY, ModelInfo, NotLoggedInError, type OAuthCredentials, type OAuthLoginCallbacks, PLAN_MODE_THINKING_CAP, type ProbeOptions, type ProgressCallback, SubscriptionUsageError, type SubscriptionUsageProvider, type SubscriptionUsageSnapshot, type SubscriptionUsageWindow, TelegramBot, type TelegramConfig, type TelegramMessage, type TelegramUpdate, type TelegramVoiceMessage, XAI_OAUTH_KEY, XIAOMI_CREDITS_KEY, authPriorityNote, clampThinkingForPlanMode, clearLocalDiscoveryCache, closeLogger, createAutoUpdater, decodeOggOpus, describeAuthMethods, discoverLocalModels, downmixToMono, dualAuthProvider, dualAuthProviderByOAuthKey, endpointRoot, fetchSubscriptionUsage, findProbedModel, formatLocalModelId, generatePKCE, getAuthProvider, getLowestThinkingLevel, getNextThinkingLevel, getSessionId, getSupportedThinkingLevels, grokCliBaseUrl, grokCliHeaders, isGrokCliEndpoint, isKimiCodingEndpoint, isLocalModelId, isLoggerOpen, isModelLoaded, isThinkingLevelSupported, kimiCodeBaseUrl, kimiCodingHeaders, localAuthStorageKey, log, loginAnthropic, loginGemini, loginKimi, loginOpenAI, loginXai, oauthStorageKey, openLog, parseLocalModelId, probeEndpoint, providerStorageKeys, readStoredBaseUrlSync, refreshAnthropicToken, refreshGeminiToken, refreshKimiToken, refreshOpenAIToken, refreshXaiToken, registerLogCleanup, resample, setProgressCallback, toModelInfo, transcribeVoice, withFileLock };
+export { AUTH_PROVIDERS, type ApiKeyVariant, type AuthMethod, type AuthMethodMeta, type AuthProviderMeta, AuthStorage, type AutoUpdateConfig, type AutoUpdater, DEFAULT_LOCAL_ENDPOINTS, type DiscoverOptions, type DiscoveryResult, type DualAuthProvider, FALLBACK_CONTEXT_WINDOW, GEMINI_API_KEY, GEMINI_API_KEY_BASE_URL, type InlineButton, LOCAL_API_KEY_PLACEHOLDER, LOCAL_AUTH_KEY_PREFIX, type LocalEndpoint, type LocalEndpointKind, type LocalEndpointProbe, type LocalModel, type LogLevel, MINIMAX_CN_BASE_URL, MINIMAX_GLOBAL_BASE_URL, MOONSHOT_OAUTH_KEY, ModelInfo, NotLoggedInError, type OAuthCredentials, type OAuthLoginCallbacks, PLAN_MODE_THINKING_CAP, type ProbeOptions, type ProgressCallback, SubscriptionUsageError, type SubscriptionUsageProvider, type SubscriptionUsageSnapshot, type SubscriptionUsageWindow, TelegramBot, type TelegramConfig, type TelegramMessage, type TelegramUpdate, type TelegramVoiceMessage, XAI_OAUTH_KEY, XIAOMI_CREDITS_KEY, XIAOMI_TOKEN_PLAN_BASE_URLS, authPriorityNote, clampThinkingForPlanMode, clearLocalDiscoveryCache, closeLogger, createAutoUpdater, decodeOggOpus, describeAuthMethods, detectMiniMaxBaseUrl, detectXiaomiTokenPlanBaseUrl, discoverLocalModels, downmixToMono, dualAuthProvider, dualAuthProviderByOAuthKey, endpointRoot, fetchSubscriptionUsage, findProbedModel, formatLocalModelId, generatePKCE, getAuthProvider, getLowestThinkingLevel, getNextThinkingLevel, getSessionId, getSupportedThinkingLevels, grokCliBaseUrl, grokCliHeaders, isGrokCliEndpoint, isKimiCodingEndpoint, isLocalModelId, isLoggerOpen, isModelLoaded, isThinkingLevelSupported, kimiCodeBaseUrl, kimiCodingHeaders, localAuthStorageKey, localModelShortName, log, loginAnthropic, loginGemini, loginKimi, loginOpenAI, loginXai, oauthStorageKey, ollamaServedContext, openLog, parseLocalModelId, probeEndpoint, providerStorageKeys, readStoredBaseUrlSync, refreshAnthropicToken, refreshGeminiToken, refreshKimiToken, refreshOpenAIToken, refreshXaiToken, registerLogCleanup, resample, setProgressCallback, toModelInfo, transcribeVoice, withFileLock };

package/dist/index.js

@@ -5,13 +5,18 @@ import {
   GEMINI_API_KEY_BASE_URL,
   HIDDEN_PROVIDERS,
   LOCAL_AUTH_KEY_PREFIX,
+  MINIMAX_CN_BASE_URL,
+  MINIMAX_GLOBAL_BASE_URL,
   MODELS,
   MOONSHOT_OAUTH_KEY,
   NotLoggedInError,
   XAI_OAUTH_KEY,
   XIAOMI_CREDITS_KEY,
+  XIAOMI_TOKEN_PLAN_BASE_URLS,
   clearRuntimeModels,
   closeLogger,
+  detectMiniMaxBaseUrl,
+  detectXiaomiTokenPlanBaseUrl,
   dualAuthProvider,
   dualAuthProviderByOAuthKey,
   generatePKCE,
@@ -61,7 +66,7 @@ import {
   usesOpenAICodexTransport,
   withFileLock,
   withoutPaidTier
-} from "./chunk-5ZYOOESD.js";
+} from "./chunk-M53JVW2V.js";
 import {
   getAppPaths
 } from "./chunk-2X5L6U4Z.js";
@@ -94,6 +99,13 @@ var ANTHROPIC_ADAPTIVE_THINKING_LEVELS = [
 ];
 var MOONSHOT_K3_THINKING_LEVELS = ["low", "high", "max"];
 var GLM_THINKING_LEVELS = ["low", "medium", "high", "xhigh", "max"];
+var MINIMAX_EFFORT_THINKING_LEVELS = [
+  "low",
+  "medium",
+  "high",
+  "xhigh",
+  "max"
+];
 var LOCAL_THINKING_LEVELS = ["low", "medium", "high", "max"];
 function isOpenAIGptModel(provider, model) {
   return provider === "openai" && model.startsWith("gpt-");
@@ -110,6 +122,9 @@ function isMoonshotK3Model(provider, model) {
 function isGlmModel(provider) {
   return provider === "glm";
 }
+function isMiniMaxEffortModel(provider, model) {
+  return provider === "minimax" && model.startsWith("MiniMax-M3.1");
+}
 function isAnthropicXhighModel(provider, model) {
   return provider === "anthropic" && /opus-5|opus-4-8|opus-4-7|sonnet-5[-.]5|haiku-5[-.]5/.test(model);
 }
@@ -142,6 +157,7 @@ function getSupportedThinkingLevels(provider, model) {
   }
   if (isMoonshotK3Model(provider, model)) return MOONSHOT_K3_THINKING_LEVELS;
   if (provider === "deepseek") return DEEPSEEK_THINKING_LEVELS;
+  if (isMiniMaxEffortModel(provider, model)) return MINIMAX_EFFORT_THINKING_LEVELS;
   if (isGlmModel(provider)) {
     const maxIndex2 = GLM_THINKING_LEVELS.indexOf(maxLevel);
     if (maxIndex2 === -1) return GLM_THINKING_LEVELS;
@@ -162,7 +178,7 @@ function isThinkingLevelSupported(provider, model, level) {
 }
 function getNextThinkingLevel(provider, model, current) {
   const supportedLevels = getSupportedThinkingLevels(provider, model);
-  const shouldCycleLevels = isOpenAIGptModel(provider, model) || isAnthropicAdaptiveModel(provider, model) || isSakanaModel(provider) || isXaiModel(provider) || isMoonshotK3Model(provider, model) || isGlmModel(provider) || provider === "deepseek" || // Local servers take a real effort level, not just on/off: Ollama accepts
+  const shouldCycleLevels = isOpenAIGptModel(provider, model) || isAnthropicAdaptiveModel(provider, model) || isSakanaModel(provider) || isXaiModel(provider) || isMoonshotK3Model(provider, model) || isGlmModel(provider) || isMiniMaxEffortModel(provider, model) || provider === "deepseek" || // Local servers take a real effort level, not just on/off: Ollama accepts
   // low/medium/high on `reasoning_effort` (verified against 0.32) and the
   // other OpenAI-compatible servers use the same three. A model that can't
   // reason at all already has no supported levels, so it never gets here.
@@ -190,11 +206,12 @@ function clampThinkingForPlanMode(level) {
 }
 
 // src/local-models.ts
+import { LOCAL_NO_API_KEY, localBearerKey } from "@kenkaiiii/gg-ai";
 var DEFAULT_LOCAL_ENDPOINTS = [
   { id: "ollama", label: "Ollama", baseUrl: "http://127.0.0.1:11434/v1", kind: "ollama" }
 ];
 var FALLBACK_CONTEXT_WINDOW = 4096;
-var LOCAL_API_KEY_PLACEHOLDER = "local";
+var LOCAL_API_KEY_PLACEHOLDER = LOCAL_NO_API_KEY;
 var DEFAULT_PROBE_TIMEOUT_MS = 1200;
 var ENRICH_CONCURRENCY = 6;
 var CACHE_TTL_MS = 3e4;
@@ -210,6 +227,9 @@ function parseLocalModelId(id) {
   if (slash <= 0 || slash === rest.length - 1) return void 0;
   return { endpointId: rest.slice(0, slash), rawId: rest.slice(slash + 1) };
 }
+function localModelShortName(rawId) {
+  return rawId.slice(rawId.lastIndexOf("/") + 1) || rawId;
+}
 function isLocalModelId(id) {
   return parseLocalModelId(id) !== void 0;
 }
@@ -220,8 +240,9 @@ function endpointRoot(baseUrl) {
   return baseUrl.replace(/\/+$/, "").replace(/\/v1$/, "");
 }
 function authHeaders(endpoint) {
+  const key = localBearerKey(endpoint.apiKey);
   return {
-    Authorization: `Bearer ${endpoint.apiKey ?? LOCAL_API_KEY_PLACEHOLDER}`,
+    ...key ? { Authorization: `Bearer ${key}` } : {},
     Accept: "application/json"
   };
 }
@@ -283,10 +304,22 @@ async function enrich(endpoint, entries, options) {
   if (endpoint.kind === "lmstudio") return enrichLmStudio(endpoint, entries, options);
   if (endpoint.kind === "ollama") return enrichOllama(endpoint, entries, options);
   if (endpoint.kind === "llamacpp") return enrichLlamaCpp(endpoint, entries, options);
+  if (endpoint.kind === "custom") return enrichCustom(endpoint, entries, options);
   return entries.map((entry) => genericModel(endpoint, entry));
 }
+async function enrichCustom(endpoint, entries, options) {
+  const generic = entries.map((entry) => genericModel(endpoint, entry));
+  if (generic.every((model) => model.contextWindowKnown)) return generic;
+  const root = endpointRoot(endpoint.baseUrl);
+  const props = await fetchJson(`${root}/props`, endpoint, options);
+  const nCtx = contextLength(props?.default_generation_settings?.n_ctx);
+  if (nCtx !== void 0) return llamaCppModels(endpoint, entries, nCtx);
+  const lmStudio = await fetchJson(`${root}/api/v1/models`, endpoint, options);
+  if (Array.isArray(lmStudio?.models)) return enrichLmStudio(endpoint, entries, options, lmStudio);
+  return generic;
+}
 function genericModel(endpoint, entry) {
-  const declared = contextLength(entry.max_model_len);
+  const declared = contextLength(entry.max_model_len) ?? contextLength(entry.context_length);
   return {
     rawId: entry.id,
     endpointId: endpoint.id,
@@ -313,9 +346,7 @@ async function enrichOllama(endpoint, entries, options) {
     });
     const caps = Array.isArray(show?.capabilities) ? show.capabilities : [];
     if (caps.includes("embedding") && !caps.includes("completion")) return void 0;
-    const ctx = contextLength(
-      runningModels.find((model) => model?.name === entry.id || model?.model === entry.id)?.context_length
-    );
+    const ctx = runningContext(runningModels, entry.id) ?? modelfileNumCtx(show?.parameters);
     return {
       rawId: entry.id,
       endpointId: endpoint.id,
@@ -330,12 +361,41 @@ async function enrichOllama(endpoint, entries, options) {
   });
   return enriched.filter((model) => model !== void 0);
 }
+function runningContext(models, rawId) {
+  return contextLength(
+    models.find((model) => model?.name === rawId || model?.model === rawId)?.context_length
+  );
+}
+function modelfileNumCtx(parameters) {
+  const match = /^num_ctx\s+(\d+)\s*$/m.exec(parameters ?? "");
+  return match ? contextLength(Number(match[1])) : void 0;
+}
+async function ollamaServedContext(endpoint, rawId, {
+  load = false,
+  timeoutMs = 12e4,
+  signal
+} = {}) {
+  const root = endpointRoot(endpoint.baseUrl);
+  if (load) {
+    await fetchJson(`${root}/api/generate`, endpoint, {
+      timeoutMs,
+      signal,
+      method: "POST",
+      body: { model: rawId }
+    });
+  }
+  const running = await fetchJson(`${root}/api/ps`, endpoint, {
+    timeoutMs: DEFAULT_PROBE_TIMEOUT_MS,
+    signal
+  });
+  return runningContext(Array.isArray(running?.models) ? running.models : [], rawId);
+}
 function contextLength(value) {
   return typeof value === "number" && Number.isSafeInteger(value) && value > 0 ? value : void 0;
 }
-async function enrichLmStudio(endpoint, entries, options) {
+async function enrichLmStudio(endpoint, entries, options, fetched) {
   const root = endpointRoot(endpoint.baseUrl);
-  const detail = await fetchJson(`${root}/api/v1/models`, endpoint, options);
+  const detail = fetched ?? await fetchJson(`${root}/api/v1/models`, endpoint, options);
   const byId = /* @__PURE__ */ new Map();
   if (Array.isArray(detail?.models)) {
     for (const model of detail.models) {
@@ -393,12 +453,17 @@ async function enrichLlamaCpp(endpoint, entries, options) {
     endpoint,
     options
   );
-  const nCtx = contextLength(props?.default_generation_settings?.n_ctx);
-  const known = nCtx !== void 0;
+  return llamaCppModels(
+    endpoint,
+    entries,
+    contextLength(props?.default_generation_settings?.n_ctx)
+  );
+}
+function llamaCppModels(endpoint, entries, nCtx) {
   return entries.map((entry) => ({
     ...genericModel(endpoint, entry),
-    contextWindow: known ? nCtx : FALLBACK_CONTEXT_WINDOW,
-    contextWindowKnown: known
+    contextWindow: nCtx ?? FALLBACK_CONTEXT_WINDOW,
+    contextWindowKnown: nCtx !== void 0
   }));
 }
 async function mapLimited(items, limit, fn) {
@@ -555,7 +620,7 @@ var ALL_AUTH_PROVIDERS = [
   {
     value: "minimax",
     label: "MiniMax",
-    description: "MiniMax M3",
+    description: "MiniMax M3.1 Flash, M3",
     methods: ["apikey"],
     apiKeyLabel: "MiniMax"
   },
@@ -1387,6 +1452,8 @@ export {
   HIDDEN_PROVIDERS,
   LOCAL_API_KEY_PLACEHOLDER,
   LOCAL_AUTH_KEY_PREFIX,
+  MINIMAX_CN_BASE_URL,
+  MINIMAX_GLOBAL_BASE_URL,
   MODELS,
   MOONSHOT_OAUTH_KEY,
   NotLoggedInError,
@@ -1395,6 +1462,7 @@ export {
   TelegramBot,
   XAI_OAUTH_KEY,
   XIAOMI_CREDITS_KEY,
+  XIAOMI_TOKEN_PLAN_BASE_URLS,
   authPriorityNote,
   clampThinkingForPlanMode,
   clearLocalDiscoveryCache,
@@ -1403,6 +1471,8 @@ export {
   createAutoUpdater,
   decodeOggOpus,
   describeAuthMethods,
+  detectMiniMaxBaseUrl,
+  detectXiaomiTokenPlanBaseUrl,
   discoverLocalModels,
   downmixToMono,
   dualAuthProvider,
@@ -1446,6 +1516,7 @@ export {
   kimiCodeBaseUrl,
   kimiCodingHeaders,
   localAuthStorageKey,
+  localModelShortName,
   log,
   loginAnthropic,
   loginGemini,
@@ -1453,6 +1524,7 @@ export {
   loginOpenAI,
   loginXai,
   oauthStorageKey,
+  ollamaServedContext,
   openLog,
   parseLocalModelId,
   pendingClaudeCodeVersionRefresh,