@prestyj/core 5.12.0 → 5.16.0

package/dist/index.cjs

@@ -1605,6 +1605,7 @@ var STATIC_API_KEY_PROVIDERS = /* @__PURE__ */ new Set([
   "minimax",
   "deepseek",
   "openrouter",
+  "huggingface",
   "sakana",
   "xai",
   // Local endpoints: a fixed (usually placeholder) key, never refreshable.
@@ -2040,8 +2041,13 @@ var MODELS = [
   // Project Glasswing (limited, invitation-only) model unavailable to most
   // users. Re-enable once it's generally available.
   {
-    id: "claude-fable-5",
-    name: "Claude Fable 5",
+    // Released 2026-09-01 — replaces Fable 5 at the same $10/$50 MTok (cache
+    // reads drop to $0.25). Always-on adaptive thinking steered by effort;
+    // forced tool use is rejected with a 400, which @prestyj/ai never sends on the
+    // Anthropic path. Fable 5 is retired here — a session that still has it
+    // saved falls back to the provider default on next start.
+    id: "claude-fable-5-1",
+    name: "Claude Fable 5.1",
     provider: "anthropic",
     contextWindow: 1e6,
     maxOutputTokens: 128e3,
@@ -2053,7 +2059,7 @@ var MODELS = [
   },
   // {
   //   // Mythos-class model offered through Project Glasswing (limited
-  //   // availability, invitation-only). Same underlying model as Fable 5 with
+  //   // availability, invitation-only). Same underlying model as Fable 5.1 with
   //   // some safeguards lifted; kept here so approved accounts can select it.
   //   id: "claude-mythos-5",
   //   name: "Claude Mythos 5",
@@ -2204,11 +2210,29 @@ var MODELS = [
     maxThinkingLevel: "xhigh"
   },
   // ── xAI (Grok) ─────────────────────────────────────────
-  // Grok 4.5 (released 2026-07-08) is xAI's flagship for coding, agentic
-  // tasks, and knowledge work — 500K context, text+image input, configurable
-  // `reasoning_effort` (low/medium/high, server default high; reasoning can't
-  // be fully disabled). Served over the OpenAI-compatible API at
-  // https://api.x.ai/v1 (API key from console.x.ai). xAI hasn't published an
+  // Grok 4.6 (released 2026-08-12) is xAI's flagship for coding, agentic tasks,
+  // and knowledge work, with a focus on long-running agents — 500K context,
+  // text+image input, and a `reasoning_effort` ladder that adds a new `xhigh`
+  // top rung (low/medium/high default/xhigh; reasoning still can't be fully
+  // disabled). $2/$6 per MTok under 200K prompt tokens ($4/$12 at or above),
+  // and it's the default model of the Grok Build coding agent. xAI advertises "no text output limit"; we keep the same
+  // 131K practical cap as 4.5 for budget predictability and input headroom.
+  {
+    id: "grok-4.6",
+    name: "Grok 4.6",
+    provider: "xai",
+    contextWindow: 5e5,
+    maxOutputTokens: 131072,
+    supportsThinking: true,
+    supportsImages: true,
+    supportsVideo: false,
+    costTier: "medium",
+    maxThinkingLevel: "xhigh"
+  },
+  // Grok 4.5 (released 2026-07-08) — superseded by 4.6 but retained as an explicit option. 500K context, text+image input,
+  // configurable `reasoning_effort` (low/medium/high, server default high;
+  // reasoning can't be fully disabled). Served over the OpenAI-compatible API
+  // at https://api.x.ai/v1 (API key from console.x.ai). xAI hasn't published an
   // official max-output cap for 4.5; 131K matches the Grok Responses ceiling
   // third-party integrations use.
   {
@@ -2223,7 +2247,7 @@ var MODELS = [
     costTier: "medium",
     maxThinkingLevel: "high"
   },
-  // ── Gemini ─────────────────────────────────────────────
+  // ── Gemini ─────────────────────────────────────────
   {
     id: "gemini-3.1-flash-lite",
     name: "Gemini 3.1 Flash Lite",
@@ -2237,6 +2261,28 @@ var MODELS = [
     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. Listed SECOND, after flash-lite:
+    // getFastModel picks the first low-tier entry, and flash-lite is the one
+    // that works on every account.
+    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"
+  },
   {
     // Wire name `gemini-3-flash` — the Code Assist (OAuth) backend rejects the
     // display string `gemini-3.5-flash` with a 404, so gemini-cli keeps this
@@ -2304,11 +2350,15 @@ var MODELS = [
     maxThinkingLevel: "high"
   },
   // ── Z.AI (GLM) ─────────────────────────────────────────
-  // GLM-5.2: coding-first flagship with a usable 1M-token context window
-  // (5x jump over GLM-5.1's ~200K) and 131K max output. Released 2026-06-13.
+  // Two GLM entries, both live on the coding endpoint (verified against its
+  // /models list). The pre-5.3 ids stay retired: they routed to strictly worse
+  // coding for the same plan quota, and the endpoint already answers `glm-5.2`
+  // requests as glm-5.3.
+  // `max` is both the ceiling and Z.AI's own default — the rungs below it live
+  // in thinking-level.ts.
   {
-    id: "glm-5.2",
-    name: "GLM-5.2",
+    id: "glm-5.3",
+    name: "GLM-5.3",
     provider: "glm",
     contextWindow: 1e6,
     maxOutputTokens: 131072,
@@ -2316,43 +2366,31 @@ var MODELS = [
     supportsImages: false,
     supportsVideo: false,
     costTier: "medium",
-    maxThinkingLevel: "high"
-  },
-  {
-    id: "glm-5.1",
-    name: "GLM-5.1",
-    provider: "glm",
-    contextWindow: 204800,
-    maxOutputTokens: 131072,
-    supportsThinking: true,
-    supportsImages: false,
-    supportsVideo: false,
-    costTier: "medium",
-    maxThinkingLevel: "high"
-  },
-  {
-    id: "glm-4.7",
-    name: "GLM-4.7",
-    provider: "glm",
-    contextWindow: 2e5,
-    maxOutputTokens: 131072,
-    supportsThinking: true,
-    supportsImages: false,
-    supportsVideo: false,
-    costTier: "low",
-    maxThinkingLevel: "high"
+    maxThinkingLevel: "max"
   },
+  // GLM-5.3-Flash (released 2026-08-26): 320B-A18B natively multimodal sibling
+  // at ~1/20th of 5.3's API price with 3× the coding-plan quota, so it is the
+  // provider's `low` tier — scout sub-agents and compaction summaries route
+  // here instead of paying 5.3 rates.
+  // Images are native on the coding endpoint (verified: base64 data URL in an
+  // `image_url` block answers correctly), which also means GLM image
+  // attachments go inline for this model rather than through the zai_vision MCP
+  // detour that `supportsImages: false` triggers.
+  // Video/file input is documented but unverified on this transport, so it
+  // stays off until measured. Thinking cannot be disabled server-side (Z.AI
+  // maps a `disabled` toggle to the `low` rung and answers 200), and unlike
+  // 5.3 it accepts any reasoning_effort string without a 400.
   {
-    id: "glm-4.7-flash",
-    name: "GLM-4.7 Flash",
+    id: "glm-5.3-flash",
+    name: "GLM-5.3-Flash",
     provider: "glm",
-    contextWindow: 2e5,
+    contextWindow: 1e6,
     maxOutputTokens: 131072,
     supportsThinking: true,
-    supportsImages: false,
+    supportsImages: true,
     supportsVideo: false,
     costTier: "low",
-    maxThinkingLevel: "high"
+    maxThinkingLevel: "max"
   },
   // ── MiniMax ────────────────────────────────────────────
   {
@@ -2420,15 +2458,21 @@ var MODELS = [
   },
   // ── DeepSeek ───────────────────────────────────────────
   {
+    // `deepseek-v4-pro` now serves DeepSeek-V4-Pro-0813 (released 2026-08-13,
+    // first STABLE V4 Pro — supersedes the April preview; calling name
+    // unchanged, same 1.6T/49B MoE). 1M context, 384K (393,216) max output,
+    // text-only, reasoning ladder low/high plus Think Max — mapped from our
+    // `xhigh`. ~$0.43/$0.87 per MTok on DeepSeek's own API, so a mid-tier
+    // price band rather than the preview's top band.
     id: "deepseek-v4-pro",
     name: "DeepSeek V4 Pro",
     provider: "deepseek",
     contextWindow: 1048576,
-    maxOutputTokens: 384e3,
+    maxOutputTokens: 393216,
     supportsThinking: true,
     supportsImages: false,
     supportsVideo: false,
-    costTier: "high",
+    costTier: "medium",
     // DeepSeek V4 maps `xhigh` → its internal `max` tier.
     maxThinkingLevel: "xhigh"
   },
@@ -2456,6 +2500,45 @@ var MODELS = [
     supportsVideo: false,
     costTier: "medium",
     maxThinkingLevel: "high"
+  },
+  // ── Hugging Face (Inference Providers router) ────────
+  // One HF token (hf.co/settings/tokens, "Make calls to Inference Providers"
+  // permission) routes to whichever hosted backend serves each open model;
+  // billing follows each backend's rates on the HF account (small free tier).
+  // Model ids are Hub repo paths, so they intentionally contain a slash — the
+  // same shape local/ vLLM ids already use (`local/vllm/Qwen/Qwen3-32B`).
+  {
+    // Qwen's open flagship for agentic coding — tool-calling native, non-thinking
+    // (the Coder line dropped the <think> block). 262K native context (1M needs
+    // YaRN, which the router doesn't apply), 131K max output. :auto suffix lets
+    // HF pick the backend with capacity; we keep the bare repo id so the picker
+    // matches what GET /v1/models reports.
+    id: "Qwen/Qwen3-Coder-480B-A35B-Instruct",
+    name: "Qwen3 Coder 480B",
+    provider: "huggingface",
+    contextWindow: 262144,
+    maxOutputTokens: 131072,
+    supportsThinking: false,
+    supportsImages: false,
+    supportsVideo: false,
+    costTier: "medium",
+    maxThinkingLevel: "low"
+  },
+  {
+    // OpenAI's open-weight 120B MoE (5.1B active) — general-purpose, tool-calling
+    // native, adjustable reasoning effort (low/medium/high, default medium) over
+    // the router's Chat Completions API. Cheap enough to be the low-tier sibling
+    // for summaries and fast sub-agents.
+    id: "openai/gpt-oss-120b",
+    name: "GPT-OSS 120B",
+    provider: "huggingface",
+    contextWindow: 131072,
+    maxOutputTokens: 65536,
+    supportsThinking: true,
+    supportsImages: false,
+    supportsVideo: false,
+    costTier: "low",
+    maxThinkingLevel: "high"
   }
 ];
 var runtimeModels = /* @__PURE__ */ new Map();
@@ -2497,13 +2580,15 @@ function getDefaultModel(provider) {
   if (provider === "xiaomi") return MODELS.find((m) => m.id === "mimo-v2.5-pro");
   if (provider === "openai") return MODELS.find((m) => m.id === "gpt-5.6-sol");
   if (provider === "gemini") return MODELS.find((m) => m.id === "gemini-3.1-flash-lite");
-  if (provider === "glm") return MODELS.find((m) => m.id === "glm-5.2");
+  if (provider === "glm") return MODELS.find((m) => m.id === "glm-5.3");
   if (provider === "moonshot") return MODELS.find((m) => m.id === "kimi-k3");
   if (provider === "minimax") return MODELS.find((m) => m.id === "MiniMax-M3");
   if (provider === "deepseek") return MODELS.find((m) => m.id === "deepseek-v4-pro");
+  if (provider === "huggingface")
+    return MODELS.find((m) => m.id === "Qwen/Qwen3-Coder-480B-A35B-Instruct");
   if (provider === "openrouter") return MODELS.find((m) => m.id === "qwen/qwen3.6-plus");
   if (provider === "sakana") return MODELS.find((m) => m.id === "fugu");
-  if (provider === "xai") return MODELS.find((m) => m.id === "grok-4.5");
+  if (provider === "xai") return MODELS.find((m) => m.id === "grok-4.6");
   if (provider === "local") {
     return getModelsForProvider("local")[0] ?? PLACEHOLDER_LOCAL_MODEL;
   }
@@ -2547,7 +2632,7 @@ function getSummaryModel(provider, currentModelId) {
   if (provider === "anthropic") {
     return MODELS.find((m) => m.id === "claude-sonnet-5");
   }
-  if (provider === "openai" || provider === "glm" || provider === "deepseek") {
+  if (provider === "openai" || provider === "glm" || provider === "deepseek" || provider === "huggingface") {
     const low = getModelsForProvider(provider).find((m) => m.costTier === "low");
     if (low) return low;
   }
@@ -2569,7 +2654,7 @@ var OPENAI_GPT_56_THINKING_LEVELS = [
   "ultra"
 ];
 var SAKANA_THINKING_LEVELS = ["high", "xhigh"];
-var XAI_THINKING_LEVELS = ["low", "medium", "high"];
+var XAI_THINKING_LEVELS = ["low", "medium", "high", "xhigh"];
 var ANTHROPIC_XHIGH_THINKING_LEVELS = [
   "low",
   "medium",
@@ -2584,6 +2669,7 @@ var ANTHROPIC_ADAPTIVE_THINKING_LEVELS = [
   "max"
 ];
 var MOONSHOT_K3_THINKING_LEVELS = ["low", "high", "max"];
+var GLM_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-");
@@ -2597,6 +2683,9 @@ function isXaiModel(provider) {
 function isMoonshotK3Model(provider, model) {
   return provider === "moonshot" && model === "kimi-k3";
 }
+function isGlmModel(provider) {
+  return provider === "glm";
+}
 function isAnthropicXhighModel(provider, model) {
   return provider === "anthropic" && /opus-5|opus-4-8|opus-4-7/.test(model);
 }
@@ -2628,6 +2717,11 @@ function getSupportedThinkingLevels(provider, model) {
     return XAI_THINKING_LEVELS.slice(0, maxIndex2 + 1);
   }
   if (isMoonshotK3Model(provider, model)) return MOONSHOT_K3_THINKING_LEVELS;
+  if (isGlmModel(provider)) {
+    const maxIndex2 = GLM_THINKING_LEVELS.indexOf(maxLevel);
+    if (maxIndex2 === -1) return GLM_THINKING_LEVELS;
+    return GLM_THINKING_LEVELS.slice(0, maxIndex2 + 1);
+  }
   if (!isOpenAIGptModel(provider, model)) return [maxLevel];
   const levels = model.startsWith("gpt-5.6-") ? OPENAI_GPT_56_THINKING_LEVELS : OPENAI_GPT_THINKING_LEVELS;
   const maxIndex = levels.indexOf(maxLevel);
@@ -2639,7 +2733,7 @@ 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) || // 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) || // 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.
@@ -2655,10 +2749,7 @@ function getNextThinkingLevel(provider, model, current) {
 
 // src/local-models.ts
 var DEFAULT_LOCAL_ENDPOINTS = [
-  { id: "ollama", label: "Ollama", baseUrl: "http://127.0.0.1:11434/v1", kind: "ollama" },
-  { id: "lmstudio", label: "LM Studio", baseUrl: "http://127.0.0.1:1234/v1", kind: "lmstudio" },
-  { id: "llamacpp", label: "llama.cpp", baseUrl: "http://127.0.0.1:8080/v1", kind: "llamacpp" },
-  { id: "vllm", label: "vLLM", baseUrl: "http://127.0.0.1:8000/v1", kind: "vllm" }
+  { id: "ollama", label: "Ollama", baseUrl: "http://127.0.0.1:11434/v1", kind: "ollama" }
 ];
 var FALLBACK_CONTEXT_WINDOW = 8192;
 var LOCAL_API_KEY_PLACEHOLDER = "local";