@prestyj/core 5.26.0 → 5.28.0

package/dist/model-registry.cjs

@@ -123,6 +123,7 @@ var import_promises2 = __toESM(require("fs/promises"), 1);
 var import_promises3 = require("timers/promises");
 
 // src/auth-storage.ts
+var MOONSHOT_OAUTH_KEY = "moonshot-oauth";
 var XIAOMI_CREDITS_KEY = "xiaomi-credits";
 var LOCAL_CREDENTIAL_LIFETIME_MS = 100 * 365 * 24 * 60 * 60 * 1e3;
 var USAGE_EXHAUSTED_DEFAULT_MS = 15 * 60 * 1e3;
@@ -208,8 +209,10 @@ var MODELS = [
     maxThinkingLevel: "max"
   },
   {
-    id: "claude-sonnet-5",
-    name: "Claude Sonnet 5",
+    // Released 2026-09-28 — replaces Sonnet 5 at $2/$10 MTok, with the same
+    // 1M context / 128K output and adaptive thinking, now including xhigh.
+    id: "claude-sonnet-5-5",
+    name: "Claude Sonnet 5.5",
     provider: "anthropic",
     contextWindow: 1e6,
     maxOutputTokens: 128e3,
@@ -256,11 +259,18 @@ var MODELS = [
     costTier: "high",
     maxThinkingLevel: "ultra"
   },
-  {
-    // GPT-6 Sol — "Workhorse model for coding and everyday work." (Codex
-    // catalog priority 2, default medium, requires Codex client >= 0.155.0).
-    // Launched Sep 22 2026 at $2/$10 per 1M tokens — half of GPT-5.6 Sol.
-    // Same 1.05M public / 272K Codex split and low → ultra ladder as Astra.
+  // 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). Both need a Codex client >= 0.155.0
+  // on the ChatGPT OAuth route. Same window split as Astra: 1.05M on the public
+  // Responses API, 272K on the Codex route; 128K output, text+image input,
+  // freeform apply_patch, responses-lite transport. The 5.6 ids are retired —
+  // a saved session on one falls back to the provider default on next start.
+  {
+    // Sol — "Workhorse model for coding and everyday work." (Codex priority 2,
+    // default medium). $2/$10 MTok. Ladder low → medium → high → xhigh → max →
+    // ultra; ultra is the Codex orchestration preset (max effort on the wire +
+    // proactive local subagent delegation).
     id: "gpt-6-sol",
     name: "GPT-6 Sol",
     provider: "openai",
@@ -275,10 +285,8 @@ var MODELS = [
     maxThinkingLevel: "ultra"
   },
   {
-    // GPT-6 Luna — "Fast and affordable model for easier tasks." (Codex
-    // catalog priority 3, default medium, client >= 0.155.0). $0.10/$0.50 per
-    // 1M tokens. Reasoning tops out at `max` (no ultra preset). Listed ahead of
-    // GPT-5.6 Luna so getFastModel picks it as the OpenAI fast tier.
+    // Luna — "Fast and affordable model for easier tasks." (Codex priority 3,
+    // default medium). $0.10/$0.50 MTok. Reasoning tops out at `max`.
     id: "gpt-6-luna",
     name: "GPT-6 Luna",
     provider: "openai",
@@ -292,71 +300,29 @@ var MODELS = [
     costTier: "low",
     maxThinkingLevel: "max"
   },
-  // GPT-5.6 family — three agentic coding tiers launched July 2026. The public
-  // Responses API advertises a 1.05M context window; OpenAI's Codex product
-  // catalog advertises 272K on the ChatGPT OAuth route (corrected from the
-  // initially advertised 372K — openai/codex PR #33972, Jul 18 2026 hotfix). All three take
-  // text+image input, freeform apply_patch, text+image web search, and parallel
-  // tool calls.
-  {
-    // Sol — "Latest frontier agentic coding model." (priority 1, default low).
-    // Reasoning ladder: low → medium → high → xhigh → max → ultra. Ultra is a
-    // Codex orchestration preset: the request uses max effort while the local
-    // runtime proactively delegates suitable independent work to subagents.
-    id: "gpt-5.6-sol",
-    name: "GPT-5.6 Sol",
-    provider: "openai",
-    contextWindow: 105e4,
-    codexContextWindow: 272e3,
-    maxOutputTokens: 128e3,
-    supportsThinking: true,
-    defaultThinkingLevel: "low",
-    supportsImages: true,
-    supportsVideo: false,
-    costTier: "high",
-    maxThinkingLevel: "ultra"
-  },
+  // ── 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
+  // text + image input (verified against the live /models list, 2026-09-28).
+  // `fugu` balances latency and quality; `fugu-max` (v1.0, 2026-09-11) is the
+  // cost tier over the largest open-weight pool ($2/$6 per 1M); `fugu-ultra`
+  // is the heavier quality tier (may need larger client timeouts). Plain Fugu
+  // and Fugu Max stop at xhigh — Sakana documents max as the same effort there.
   {
-    // Terra — "Balanced agentic coding model for everyday work." (priority 2,
-    // default medium).
-    id: "gpt-5.6-terra",
-    name: "GPT-5.6 Terra",
-    provider: "openai",
-    contextWindow: 105e4,
-    codexContextWindow: 272e3,
+    id: "fugu",
+    name: "Fugu",
+    provider: "sakana",
+    contextWindow: 1e6,
     maxOutputTokens: 128e3,
     supportsThinking: true,
-    defaultThinkingLevel: "medium",
     supportsImages: true,
     supportsVideo: false,
     costTier: "medium",
-    maxThinkingLevel: "ultra"
-  },
-  {
-    // Luna — "Fast and affordable agentic coding model." (priority 3, default
-    // medium). Reasoning tops out at `max`.
-    id: "gpt-5.6-luna",
-    name: "GPT-5.6 Luna",
-    provider: "openai",
-    contextWindow: 105e4,
-    codexContextWindow: 272e3,
-    maxOutputTokens: 128e3,
-    supportsThinking: true,
-    defaultThinkingLevel: "medium",
-    supportsImages: true,
-    supportsVideo: false,
-    costTier: "low",
-    maxThinkingLevel: "max"
+    maxThinkingLevel: "xhigh"
   },
-  // ── 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). Both models take
-  // text + image input. Plain Fugu stops at xhigh; Ultra v1.1 also supports max.
-  // `fugu` routes across all providers; `fugu-ultra` is
-  // the heavier tier (may need larger client timeouts on complex tasks).
   {
-    id: "fugu",
-    name: "Fugu",
+    id: "fugu-max",
+    name: "Fugu Max",
     provider: "sakana",
     contextWindow: 1e6,
     maxOutputTokens: 128e3,
@@ -376,7 +342,7 @@ var MODELS = [
     supportsImages: true,
     supportsVideo: false,
     costTier: "high",
-    // The rolling alias now serves v1.1, which adds a distinct max effort.
+    // The rolling alias now serves v2.0 (2026-09-11), which keeps max effort.
     maxThinkingLevel: "max"
   },
   // ── xAI (Grok) ─────────────────────────────────────────
@@ -520,7 +486,27 @@ var MODELS = [
     costTier: "high",
     maxThinkingLevel: "max"
   },
-  // Retain the cheaper dedicated coding model as an explicit alternative.
+  // K2.8 Preview (2026-09-11) is served only on the Kimi For Coding OAuth
+  // endpoint, under its rolling `kimi-for-coding` id (live /models, 2026-09-28:
+  // display_name "K2.8 Preview", 1M context, image + video input, efforts
+  // low/high/max default max). The public API-key endpoint does not serve it,
+  // so it resolves from the Kimi sign-in credential only.
+  {
+    id: "kimi-for-coding",
+    name: "Kimi K2.8 Preview",
+    provider: "moonshot",
+    contextWindow: 1048576,
+    maxOutputTokens: 131072,
+    supportsThinking: true,
+    supportsImages: true,
+    supportsVideo: true,
+    maxVideoBytes: 100 * 1024 * 1024,
+    costTier: "medium",
+    maxThinkingLevel: "max",
+    authStorageKeys: [MOONSHOT_OAUTH_KEY]
+  },
+  // K2.7 Code is requested by its pinned id (not the `kimi-for-coding` alias
+  // that moved to K2.8), so it stays the real K2.7 on both endpoints.
   {
     id: "kimi-k2.7-code",
     name: "Kimi K2.7 Code",
@@ -667,10 +653,14 @@ var MODELS = [
     authStorageKeys: [XIAOMI_CREDITS_KEY]
   },
   // ── DeepSeek ───────────────────────────────────────────
+  // 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.
   {
     // `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, text-only, low/high/max effort.
+    // DeepSeek reversed its planned 2026-09-14 retirement, so it stays served.
     // Docs abbreviate output as 384K; use the same conservative 384,000-token
     // application cap across V4 models rather than mixing decimal/binary units.
     id: "deepseek-v4-pro",
@@ -685,21 +675,10 @@ var MODELS = [
     maxThinkingLevel: "max"
   },
   {
-    id: "deepseek-v4-flash",
-    name: "DeepSeek V4 Flash",
-    provider: "deepseek",
-    contextWindow: 1048576,
-    maxOutputTokens: 384e3,
-    supportsThinking: true,
-    supportsImages: false,
-    supportsVideo: false,
-    costTier: "low",
-    maxThinkingLevel: "max"
-  },
-  // Opt-in experimental vision sibling; never replaces the stable summary model.
-  {
-    id: "deepseek-v4-flash-vision-exp",
-    name: "DeepSeek V4 Flash Vision (Experimental)",
+    // `deepseek-flash` is the rolling alias for the latest Flash — currently
+    // V4.1 Flash (2026-09-10): native image input, 1M context, 384K output.
+    id: "deepseek-flash",
+    name: "DeepSeek V4.1 Flash",
     provider: "deepseek",
     contextWindow: 1048576,
     maxOutputTokens: 384e3,
@@ -711,11 +690,13 @@ var MODELS = [
   },
   // ── OpenRouter ─────────────────────────────────────────
   {
-    id: "qwen/qwen3.6-plus",
-    name: "Qwen3.6-Plus",
+    // Qwen3.8 Max — Alibaba's flagship (live /endpoints, 2026-09-28): 1M
+    // context, 131,072 output, text + image + video input, reasoning on.
+    id: "qwen/qwen3.8-max",
+    name: "Qwen3.8 Max",
     provider: "openrouter",
     contextWindow: 1e6,
-    maxOutputTokens: 65536,
+    maxOutputTokens: 131072,
     supportsThinking: true,
     supportsImages: true,
     supportsVideo: true,
@@ -809,13 +790,13 @@ function getDefaultModel(provider) {
   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 === "openrouter") return MODELS.find((m) => m.id === "qwen/qwen3.8-max");
   if (provider === "sakana") return MODELS.find((m) => m.id === "fugu");
   if (provider === "xai") return MODELS.find((m) => m.id === "grok-4.7");
   if (provider === "local") {
     return getModelsForProvider("local")[0] ?? PLACEHOLDER_LOCAL_MODEL;
   }
-  return MODELS.find((m) => m.id === "claude-sonnet-5");
+  return MODELS.find((m) => m.id === "claude-sonnet-5-5");
 }
 var PLACEHOLDER_LOCAL_MODEL = {
   id: "local/none/none",
@@ -853,7 +834,7 @@ function getDefaultThinkingLevel(modelId, options) {
 }
 function getSummaryModel(provider, currentModelId) {
   if (provider === "anthropic") {
-    return MODELS.find((m) => m.id === "claude-sonnet-5");
+    return MODELS.find((m) => m.id === "claude-sonnet-5-5");
   }
   if (provider === "openai" || provider === "glm" || provider === "deepseek" || provider === "huggingface") {
     const low = getModelsForProvider(provider).find((m) => m.costTier === "low");