@kenkaiiii/gg-core 5.81.0 → 5.83.0

package/dist/model-registry.cjs

@@ -45,8 +45,10 @@ __export(model_registry_exports, {
   getSummaryModel: () => getSummaryModel,
   getToolResultCharLimit: () => getToolResultCharLimit,
   getVideoByteLimit: () => getVideoByteLimit,
+  isPaidModel: () => isPaidModel,
   registerRuntimeModels: () => registerRuntimeModels,
-  usesOpenAICodexTransport: () => usesOpenAICodexTransport
+  usesOpenAICodexTransport: () => usesOpenAICodexTransport,
+  withoutPaidTier: () => withoutPaidTier
 });
 module.exports = __toCommonJS(model_registry_exports);
 
@@ -122,10 +124,29 @@ var import_promises3 = require("timers/promises");
 // src/auth-storage.ts
 var MOONSHOT_OAUTH_KEY = "moonshot-oauth";
 var XIAOMI_CREDITS_KEY = "xiaomi-credits";
+var GEMINI_API_KEY = "gemini-api";
 var LOCAL_CREDENTIAL_LIFETIME_MS = 100 * 365 * 24 * 60 * 60 * 1e3;
 var USAGE_EXHAUSTED_DEFAULT_MS = 15 * 60 * 1e3;
 
 // src/model-registry.ts
+var CLAUDE_FAST_BILLING = {
+  label: "bills usage credits",
+  detail: "Fast mode bills your Claude usage credits ($8 in / $40 out per million tokens) from the first token, never your plan's included usage.",
+  paid: true,
+  standardModel: "claude-opus-5-5"
+};
+var XIAOMI_ULTRASPEED_BILLING = {
+  label: "bills API Credits",
+  detail: "UltraSpeed runs only on your Xiaomi API Credits, billed per token at a higher rate than MiMo-V2.6-Pro, never on the Token Plan.",
+  paid: true,
+  standardModel: "mimo-v2.6-pro"
+};
+var CODEX_FAST_BILLING = {
+  label: "2.5x plan usage",
+  detail: "Codex Fast mode uses your ChatGPT plan limits at 2.5x the standard rate (purchased credits at 2x).",
+  paid: false
+};
+var GEMINI_AUTH_KEYS = [GEMINI_API_KEY, "gemini"];
 var MODELS = [
   // ── Anthropic ──────────────────────────────────────────
   // NOTE: Claude Mythos 5.1 (`claude-mythos-5-1`, 2026-09-01) is kept commented
@@ -170,9 +191,8 @@ var MODELS = [
     // context, 128K output, image input. Same always-on adaptive thinking with
     // the full effort ladder (low→max, xhigh included; the API-side default
     // effort is `medium`) and the same Fable 5.1 constraints gg-ai already
-    // handles (no forced tool use, thinking can't be disabled). Also eligible
-    // for the gated `speed: "fast"` research preview (waitlist-only, premium
-    // $8/$40 pricing — not wired up). Opus 5 is retired here — a session that
+    // handles (no forced tool use, thinking can't be disabled). Fast mode is
+    // the separate `claude-opus-5-5-fast` entry below. Opus 5 is retired here — a session that
     // still has it saved falls back to the provider default on next start.
     id: "claude-opus-5-5",
     name: "Claude Opus 5.5",
@@ -189,6 +209,26 @@ var MODELS = [
     defaultThinkingLevel: "medium",
     maxThinkingLevel: "max"
   },
+  {
+    // Opus 5.5 in fast mode (research preview): the same model at up to 2.5x
+    // output tokens/sec, $8/$40 MTok. gg-ai's `stream()` sends it as
+    // `claude-opus-5-5` + `speed: "fast"` + the fast-mode beta. On a Claude
+    // subscription it bills usage credits only — never included plan usage —
+    // and Anthropic 429s "Usage credits are required for fast mode" until they
+    // are turned on. Its prompt cache is separate from standard-speed Opus.
+    id: "claude-opus-5-5-fast",
+    name: "Claude Opus 5.5 Fast",
+    provider: "anthropic",
+    contextWindow: 1e6,
+    maxOutputTokens: 128e3,
+    supportsThinking: true,
+    supportsImages: true,
+    supportsVideo: false,
+    costTier: "high",
+    defaultThinkingLevel: "medium",
+    maxThinkingLevel: "max",
+    billing: CLAUDE_FAST_BILLING
+  },
   {
     // Released 2026-09-28 — replaces Sonnet 5 at $2/$10 MTok, with the same
     // 1M context / 128K output and adaptive thinking, now including xhigh.
@@ -201,9 +241,10 @@ var MODELS = [
     supportsImages: true,
     supportsVideo: false,
     costTier: "medium",
-    // Anthropic's API and Claude Code both start the 5.5 family at medium; at a
-    // given level 5.5 thinks more than 5, so the ladder ceiling (max) made fresh
-    // sessions slow and expensive. Users can still opt up.
+    // We start Sonnet 5.5 at medium like the rest of the 5.5 family (the API's
+    // own default for Sonnet 5.5 is high). At a given level 5.5 thinks more
+    // than 5, so the ladder ceiling (max) made fresh sessions slow and
+    // expensive. Users can still opt up.
     defaultThinkingLevel: "medium",
     maxThinkingLevel: "max"
   },
@@ -256,6 +297,25 @@ var MODELS = [
     costTier: "high",
     maxThinkingLevel: "ultra"
   },
+  {
+    // Codex Fast mode for Astra (catalog `service_tiers: priority`, "2x speed,
+    // increased usage"). gg-ai's `stream()` sends it as `gpt-6-astra` +
+    // `service_tier: "priority"`, exactly what the Codex CLI sends. On a
+    // ChatGPT plan it draws included limits at 2.5x the standard rate.
+    id: "gpt-6-astra-fast",
+    name: "GPT-6 Astra Fast",
+    provider: "openai",
+    contextWindow: 105e4,
+    codexContextWindow: 272e3,
+    maxOutputTokens: 128e3,
+    supportsThinking: true,
+    defaultThinkingLevel: "low",
+    supportsImages: true,
+    supportsVideo: false,
+    costTier: "high",
+    maxThinkingLevel: "ultra",
+    billing: CODEX_FAST_BILLING
+  },
   // GPT-6 Sol + Luna — released 2026-09-22 below Astra, replacing the whole
   // GPT-5.6 family (Sol/Terra/Luna; there is no GPT-6 Terra — OpenAI's Codex
   // catalog upgrades 5.6 Terra to 6 Sol). GPT-6.1 Sol (2026-09-29) then replaced
@@ -284,6 +344,23 @@ var MODELS = [
     costTier: "medium",
     maxThinkingLevel: "ultra"
   },
+  {
+    // Codex Fast mode for 6.1 Sol ("2x speed, increased usage") — see
+    // `gpt-6-astra-fast` for the wire mapping and billing.
+    id: "gpt-6.1-sol-fast",
+    name: "GPT-6.1 Sol Fast",
+    provider: "openai",
+    contextWindow: 105e4,
+    codexContextWindow: 272e3,
+    maxOutputTokens: 128e3,
+    supportsThinking: true,
+    defaultThinkingLevel: "low",
+    supportsImages: true,
+    supportsVideo: false,
+    costTier: "medium",
+    maxThinkingLevel: "ultra",
+    billing: CODEX_FAST_BILLING
+  },
   {
     // Luna — "Fast and affordable model for easier tasks." (Codex priority 4,
     // default medium, needs a Codex client >= 0.155.0). $0.10/$0.50 MTok.
@@ -301,6 +378,24 @@ var MODELS = [
     costTier: "low",
     maxThinkingLevel: "max"
   },
+  {
+    // Codex Fast mode for Luna ("1.5x speed") — see `gpt-6-astra-fast` for the
+    // wire mapping and billing. Listed after Luna so compaction summaries
+    // (first `low` OpenAI model) stay on standard-speed Luna.
+    id: "gpt-6-luna-fast",
+    name: "GPT-6 Luna Fast",
+    provider: "openai",
+    contextWindow: 105e4,
+    codexContextWindow: 272e3,
+    maxOutputTokens: 128e3,
+    supportsThinking: true,
+    defaultThinkingLevel: "medium",
+    supportsImages: true,
+    supportsVideo: false,
+    costTier: "low",
+    maxThinkingLevel: "max",
+    billing: CODEX_FAST_BILLING
+  },
   // ── Sakana (Fugu) ──────────────────────────────────────
   // Sakana Fugu is a multi-agent system surfaced as a standard LLM via the
   // OpenAI-compatible Sakana API (https://api.sakana.ai/v1). All three take
@@ -383,10 +478,14 @@ var MODELS = [
     supportsVideo: true,
     maxVideoBytes: 20 * 1024 * 1024,
     costTier: "low",
-    maxThinkingLevel: "high"
+    maxThinkingLevel: "high",
+    authStorageKeys: GEMINI_AUTH_KEYS
   },
   // Keep 3.1 Flash Lite first for the working OAuth default and fast-model routing.
-  // New GA models are opt-in; Code Assist access varies by account.
+  // New GA models are opt-in; Code Assist access varies by account. Gemini 3.7
+  // Flash is retired: Google deprecated it on 2026-10-08 and routes every
+  // `gemini-3.7-flash` request to 3.8 Flash (ai.google.dev/gemini-api/docs/
+  // changelog), so a saved session on it falls back to the provider default.
   {
     id: "gemini-3.8-flash",
     name: "Gemini 3.8 Flash",
@@ -398,7 +497,8 @@ var MODELS = [
     supportsVideo: true,
     maxVideoBytes: 20 * 1024 * 1024,
     costTier: "low",
-    maxThinkingLevel: "high"
+    maxThinkingLevel: "high",
+    authStorageKeys: GEMINI_AUTH_KEYS
   },
   {
     id: "gemini-3.5-flash-lite",
@@ -411,27 +511,8 @@ var MODELS = [
     supportsVideo: true,
     maxVideoBytes: 20 * 1024 * 1024,
     costTier: "low",
-    maxThinkingLevel: "high"
-  },
-  {
-    // Gemini 3.7 Flash (released 2026-08-13) — Google's most capable Flash for
-    // coding, agents, and multi-step execution; GA-stable on the Gemini API as
-    // `gemini-3.7-flash`. 1M context, 64K output, thinking low/medium/high.
-    // Sent over our Code Assist (OAuth) transport ahead of gemini-cli — upstream
-    // hasn't listed 3.7 yet (google-gemini/gemini-cli#28802, still open) — so
-    // free/personal accounts 404 (entitlement-gated) while Code Assist
-    // Standard/Enterprise accounts get it.
-    id: "gemini-3.7-flash",
-    name: "Gemini 3.7 Flash",
-    provider: "gemini",
-    contextWindow: 1048576,
-    maxOutputTokens: 65536,
-    supportsThinking: true,
-    supportsImages: true,
-    supportsVideo: true,
-    maxVideoBytes: 20 * 1024 * 1024,
-    costTier: "low",
-    maxThinkingLevel: "high"
+    maxThinkingLevel: "high",
+    authStorageKeys: GEMINI_AUTH_KEYS
   },
   {
     // Wire name `gemini-3-flash` — the Code Assist (OAuth) backend rejects the
@@ -447,7 +528,8 @@ var MODELS = [
     supportsVideo: true,
     maxVideoBytes: 20 * 1024 * 1024,
     costTier: "low",
-    maxThinkingLevel: "high"
+    maxThinkingLevel: "high",
+    authStorageKeys: GEMINI_AUTH_KEYS
   },
   {
     // Gemini 3.1 Pro is public preview — gated behind Code Assist preview
@@ -463,7 +545,8 @@ var MODELS = [
     supportsVideo: true,
     maxVideoBytes: 20 * 1024 * 1024,
     costTier: "high",
-    maxThinkingLevel: "high"
+    maxThinkingLevel: "high",
+    authStorageKeys: GEMINI_AUTH_KEYS
   },
   // ── Moonshot (Kimi) ────────────────────────────────────
   // K3 is Kimi's 2.8T-parameter flagship for long-horizon coding, knowledge
@@ -652,9 +735,13 @@ var MODELS = [
     maxVideoBytes: 36 * 1024 * 1024,
     costTier: "high",
     maxThinkingLevel: "high",
-    authStorageKeys: [XIAOMI_CREDITS_KEY]
+    authStorageKeys: [XIAOMI_CREDITS_KEY],
+    billing: XIAOMI_ULTRASPEED_BILLING
   },
   // ── DeepSeek ───────────────────────────────────────────
+  // Window 1,000,000 tokens (decimal "1M"), as DeepSeek's own agent sizes it
+  // (deepseek-ai/deepseek-harness DEFAULT_CONTEXT_WINDOW = 1_000_000; models.dev
+  // via earendil-works/pi agrees) — not 1,048,576.
   // The live /models list (2026-09-28) serves exactly `deepseek-flash` and
   // `deepseek-v4-pro`. V4 Flash and V4 Flash Vision Exp are retired; their old
   // ids only temporarily route to V4.1 Flash, so they are retired here too.
@@ -668,7 +755,7 @@ var MODELS = [
     id: "deepseek-v4-pro",
     name: "DeepSeek V4 Pro",
     provider: "deepseek",
-    contextWindow: 1048576,
+    contextWindow: 1e6,
     maxOutputTokens: 384e3,
     supportsThinking: true,
     supportsImages: false,
@@ -682,7 +769,7 @@ var MODELS = [
     id: "deepseek-flash",
     name: "DeepSeek V4.1 Flash",
     provider: "deepseek",
-    contextWindow: 1048576,
+    contextWindow: 1e6,
     maxOutputTokens: 384e3,
     supportsThinking: true,
     supportsImages: true,
@@ -746,7 +833,7 @@ var MODELS = [
     id: "deepseek-ai/DeepSeek-V4.1-Flash",
     name: "DeepSeek V4.1 Flash",
     provider: "huggingface",
-    contextWindow: 1048576,
+    contextWindow: 1e6,
     maxOutputTokens: 131072,
     supportsThinking: false,
     supportsImages: true,
@@ -793,6 +880,19 @@ function getModel(id) {
 function getModelsForProvider(provider) {
   return getAllModels().filter((m) => m.provider === provider);
 }
+function isPaidModel(modelId) {
+  return getModel(modelId)?.billing?.paid === true;
+}
+function withoutPaidTier(modelId) {
+  const model = getModel(modelId);
+  if (model?.billing?.paid !== true) return modelId;
+  const usable = (id) => id !== void 0 && getModel(id) !== void 0 && !isPaidModel(id);
+  const standard = model.billing.standardModel;
+  if (usable(standard)) return standard;
+  const providerDefault = DEFAULT_MODEL_IDS[model.provider];
+  if (usable(providerDefault)) return providerDefault;
+  return FALLBACK_DEFAULT_MODEL_ID;
+}
 function getAuthStorageKeys(provider, modelId) {
   const model = getAllModels().find((m) => m.id === modelId && m.provider === provider);
   return model?.authStorageKeys ?? [provider];
@@ -874,7 +974,7 @@ var SUMMARY_COST_TIER = {
 };
 function getSummaryModel(provider, currentModelId) {
   const tier = SUMMARY_COST_TIER[provider];
-  const tagged = tier ? getModelsForProvider(provider).find((m) => m.costTier === tier) : void 0;
+  const tagged = tier ? getModelsForProvider(provider).find((m) => m.costTier === tier && m.billing === void 0) : void 0;
   return tagged ?? getModel(currentModelId) ?? getDefaultModel(provider);
 }
 // Annotate the CommonJS export names for ESM import in node:
@@ -894,7 +994,9 @@ function getSummaryModel(provider, currentModelId) {
   getSummaryModel,
   getToolResultCharLimit,
   getVideoByteLimit,
+  isPaidModel,
   registerRuntimeModels,
-  usesOpenAICodexTransport
+  usesOpenAICodexTransport,
+  withoutPaidTier
 });
 //# sourceMappingURL=model-registry.cjs.map