@prestyj/core 5.24.0 → 5.25.1

package/dist/model-registry.cjs

@@ -76,7 +76,7 @@ function isKimiCodingEndpoint(baseUrl) {
 
 // src/auth-storage.ts
 var import_promises4 = __toESM(require("fs/promises"), 1);
-var import_node_fs3 = require("fs");
+var import_node_fs4 = require("fs");
 var import_node_crypto6 = __toESM(require("crypto"), 1);
 
 // src/oauth/anthropic.ts
@@ -84,6 +84,7 @@ var import_node_crypto3 = __toESM(require("crypto"), 1);
 
 // src/claude-code-version.ts
 var import_promises = __toESM(require("fs/promises"), 1);
+var import_node_fs3 = require("fs");
 var import_node_path4 = __toESM(require("path"), 1);
 
 // src/logger.ts
@@ -95,7 +96,8 @@ var MAX_BYTES = 10 * 1024 * 1024;
 var MIN_HEADROOM_BYTES = 64 * 1024;
 
 // src/claude-code-version.ts
-var CACHE_TTL_MS = 24 * 60 * 60 * 1e3;
+var CACHE_FRESH_MS = 60 * 60 * 1e3;
+var CACHE_HARD_MS = 24 * 60 * 60 * 1e3;
 
 // src/oauth/anthropic.ts
 var CLIENT_ID = atob("OWQxYzI1MGEtZTYxYi00NGQ5LTg4ZWQtNTk0NGQxOTYyZjVl");
@@ -163,11 +165,37 @@ var MODELS = [
   //   costTier: "high",
   //   maxThinkingLevel: "max",
   // },
+  {
+    // Released 2026-09-22 — "For long-running agentic coding and knowledge
+    // work". Fable-class capability at $4/$20 MTok (cheaper than the Opus 5 it
+    // replaces, $5/$25). Adaptive thinking with the full effort ladder
+    // (low→max, xhigh included), but thinking can no longer be disabled: a
+    // `thinking: {type: "disabled"}` or budget_tokens request 400s. @prestyj/ai
+    // omits the field entirely when thinking is off, so that path is safe.
+    // Forced tool use (`tool_choice` any/tool) also 400s — see
+    // `toAnthropicToolChoice`, which downgrades it to `auto` for this model.
+    // Anthropic declares the server-side default effort as `medium` (Opus 5
+    // ran `high`), and 5.5 thinks more per turn at a given level, so a fresh
+    // session starts at `medium` rather than the ladder ceiling.
+    id: "claude-opus-5-5",
+    name: "Claude Opus 5.5",
+    provider: "anthropic",
+    contextWindow: 1e6,
+    maxOutputTokens: 128e3,
+    supportsThinking: true,
+    defaultThinkingLevel: "medium",
+    supportsImages: true,
+    supportsVideo: false,
+    costTier: "high",
+    maxThinkingLevel: "max"
+  },
   {
     // Released 2026-07-24 — "For complex agentic coding and enterprise work".
     // Near-Fable capability at half the price ($5/$25 vs $10/$50). Adaptive
     // thinking with the full effort ladder (low→max, xhigh included); dateless
-    // ID is the canonical pinned snapshot (post-4.6 naming scheme).
+    // ID is the canonical pinned snapshot (post-4.6 naming scheme). Kept as a
+    // legacy option now that Opus 5.5 leads the line: it's the last Opus that
+    // accepts disabled thinking and forced tool use.
     id: "claude-opus-5",
     name: "Claude Opus 5",
     provider: "anthropic",
@@ -316,16 +344,21 @@ var MODELS = [
     maxThinkingLevel: "max"
   },
   // ── xAI (Grok) ─────────────────────────────────────────
-  // 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
+  // Grok 4.7 (released 2026-09-21) — xAI's flagship for coding, agentic tasks,
+  // and knowledge work: a new, larger base model with a longer RL run weighted
+  // toward hours-long tasks, plus stronger self-verification and long-context
+  // management. 500K context, text+image input, and a `reasoning_effort`
+  // ladder of 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",
+  // and it's the default model of the Grok Build coding agent. xAI advertises
+  // "no fixed text output limit"; we keep the 131K practical cap for budget
+  // predictability and input headroom. (A faster "Grok 4.7 Fast" variant
+  // exists but is Cursor/Grok Build-only — not on the public API — so it isn't
+  // registered.) Only the newest Grok ships — 4.6/4.5 are superseded and
+  // retired; saved sessions on them fall back to this default.
+  {
+    id: "grok-4.7",
+    name: "Grok 4.7",
     provider: "xai",
     contextWindow: 5e5,
     maxOutputTokens: 131072,
@@ -335,24 +368,6 @@ var MODELS = [
     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.
-  {
-    id: "grok-4.5",
-    name: "Grok 4.5",
-    provider: "xai",
-    contextWindow: 5e5,
-    maxOutputTokens: 131072,
-    supportsThinking: true,
-    supportsImages: true,
-    supportsVideo: false,
-    costTier: "medium",
-    maxThinkingLevel: "high"
-  },
   // ── Gemini ─────────────────────────────────────────
   {
     id: "gemini-3.1-flash-lite",
@@ -541,11 +556,15 @@ var MODELS = [
     maxThinkingLevel: "high"
   },
   // ── Xiaomi (MiMo) ──────────────────────────────────────
-  // V2.6 series (2026-09) supersedes V2.5 one-for-one: pro → pro, the omni
-  // `mimo-v2.5` → flash, ultraspeed → ultraspeed. The whole series is now
-  // full-modality, so unlike V2.5-Pro the flagship no longer needs a separate
-  // omni sibling for attachments. V2.5 entries are retired here — a session
-  // that still has one saved falls back to the provider default on next start.
+  // V2.6 series (released 2026-09-22, open-weight: Pro 1.02T/42B-A, Flash
+  // 309B/15B-A, plus a 9B Qwen distill not served over the API) supersedes V2.5
+  // one-for-one: pro → pro, the omni `mimo-v2.5` → flash, ultraspeed →
+  // ultraspeed. Every V2.6 text model is natively full-modality, so unlike
+  // V2.5-Pro the flagship no longer needs a separate omni sibling for
+  // attachments — image/video ride the same OpenAI-compatible base64 transport
+  // the old omni model used. API prices are unchanged from V2.5. The V2.5 ids
+  // deprecate on the platform 2026-10-21 and are retired here — a session that
+  // still has one saved falls back to the provider default on next start.
   //
   // Capabilities below are measured against the Token Plan host, not taken
   // from marketing copy: image and video both come back with `image_tokens` /
@@ -555,6 +574,8 @@ var MODELS = [
   // binary 1M below (2^20) and not the decimal 1e6 V2.5 was listed with — the
   // few-token gap is the chat envelope the server adds on top of the content.
   {
+    // Coding/agentic flagship — highest open-weight score on Artificial
+    // Analysis at launch (46, tied with Grok 4.7).
     id: "mimo-v2.6-pro",
     name: "MiMo-V2.6-Pro",
     provider: "xiaomi",
@@ -568,9 +589,10 @@ var MODELS = [
     maxThinkingLevel: "high",
     authStorageKeys: ["xiaomi", XIAOMI_CREDITS_KEY]
   },
-  // Flash: the cheap, high-frequency sibling at the same modality surface and
-  // window as Pro. It is the provider's `low` tier, so scout sub-agents and
-  // compaction summaries route here instead of paying Pro rates.
+  // Flash: the cheap, high-frequency sibling (~10% of Pro's price class) at the
+  // same modality surface and window as Pro. It is the provider's `low` tier,
+  // so scout sub-agents and compaction summaries route here instead of paying
+  // Pro rates.
   {
     id: "mimo-v2.6-flash",
     name: "MiMo-V2.6-Flash",
@@ -591,9 +613,9 @@ var MODELS = [
   // authStorageKeys doc). The Token Plan host rejects it with "Not supported
   // model" — the known-model/wrong-host reply — where an invented id gets
   // "Unsupported model", which is how this id was confirmed without a
-  // Credits key. Attachment support is inferred from the series announcement
-  // ("full modality across the series") rather than measured: images are
-  // enabled, video stays off until it can be verified on the platform host.
+  // Credits key. Attachments can't be probed directly for the same reason, so
+  // this entry tracks the rest of the V2.6 series ("full modality across the
+  // series") and mirrors Pro's verified image+video surface.
   {
     id: "mimo-v2.6-pro-ultraspeed",
     name: "MiMo-V2.6-Pro-UltraSpeed",
@@ -602,7 +624,8 @@ var MODELS = [
     maxOutputTokens: 131072,
     supportsThinking: true,
     supportsImages: true,
-    supportsVideo: false,
+    supportsVideo: true,
+    maxVideoBytes: 36 * 1024 * 1024,
     costTier: "high",
     maxThinkingLevel: "high",
     authStorageKeys: [XIAOMI_CREDITS_KEY]
@@ -752,7 +775,7 @@ function getDefaultModel(provider) {
     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-4.6");
+  if (provider === "xai") return MODELS.find((m) => m.id === "grok-4.7");
   if (provider === "local") {
     return getModelsForProvider("local")[0] ?? PLACEHOLDER_LOCAL_MODEL;
   }