@prestyj/core 5.16.0 → 5.16.2

package/dist/index.cjs

@@ -2113,6 +2113,29 @@ var MODELS = [
     maxThinkingLevel: "high"
   },
   // ── OpenAI (Codex) ─────────────────────────────────────
+  {
+    // GPT-6 Astra — "Our most capable model for complex, demanding work."
+    // (Codex catalog priority 1, listed for every ChatGPT plan, requires a
+    // Codex client >= 0.153.0 — see CODEX_CLIENT_VERSION). Same split as 5.6:
+    // 1.05M on the public Responses API, 272K on the ChatGPT OAuth route
+    // (openai/codex models.json, `gpt-6-astra`). Reasoning ladder low → medium
+    // → high → xhigh → max → ultra; `ultra` is the Codex orchestration preset
+    // (multi_agent v2) and is Codex-only — the public API tops out at `max`.
+    // Note: through a plain API key OpenAI requires the Responses API for tool
+    // calling on Astra, so the Chat Completions path is text-only; the OAuth
+    // Codex route is the supported way to use it as an agent.
+    id: "gpt-6-astra",
+    name: "GPT-6 Astra",
+    provider: "openai",
+    contextWindow: 105e4,
+    codexContextWindow: 272e3,
+    maxOutputTokens: 128e3,
+    supportsThinking: true,
+    supportsImages: true,
+    supportsVideo: false,
+    costTier: "high",
+    maxThinkingLevel: "ultra"
+  },
   // 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
@@ -2166,24 +2189,11 @@ var MODELS = [
     costTier: "low",
     maxThinkingLevel: "max"
   },
-  {
-    id: "gpt-5.5",
-    name: "GPT-5.5",
-    provider: "openai",
-    contextWindow: 105e4,
-    codexContextWindow: 272e3,
-    maxOutputTokens: 128e3,
-    supportsThinking: true,
-    supportsImages: true,
-    supportsVideo: false,
-    costTier: "high",
-    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 and only accept "high"/"xhigh" reasoning effort, so the
-  // top tier is `xhigh`. `fugu` routes across all providers; `fugu-ultra` is
+  // 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",
@@ -2207,7 +2217,8 @@ var MODELS = [
     supportsImages: true,
     supportsVideo: false,
     costTier: "high",
-    maxThinkingLevel: "xhigh"
+    // The rolling alias now serves v1.1, which adds a distinct max effort.
+    maxThinkingLevel: "max"
   },
   // ── xAI (Grok) ─────────────────────────────────────────
   // Grok 4.6 (released 2026-08-12) is xAI's flagship for coding, agentic tasks,
@@ -2261,6 +2272,34 @@ var MODELS = [
     costTier: "low",
     maxThinkingLevel: "high"
   },
+  // 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.
+  {
+    id: "gemini-3.8-flash",
+    name: "Gemini 3.8 Flash",
+    provider: "gemini",
+    contextWindow: 1048576,
+    maxOutputTokens: 65536,
+    supportsThinking: true,
+    supportsImages: true,
+    supportsVideo: true,
+    maxVideoBytes: 20 * 1024 * 1024,
+    costTier: "low",
+    maxThinkingLevel: "high"
+  },
+  {
+    id: "gemini-3.5-flash-lite",
+    name: "Gemini 3.5 Flash Lite",
+    provider: "gemini",
+    contextWindow: 1048576,
+    maxOutputTokens: 65536,
+    supportsThinking: true,
+    supportsImages: true,
+    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
@@ -2268,7 +2307,7 @@ var MODELS = [
     // 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:
+    // Standard/Enterprise accounts get it. Kept after the working 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",
@@ -2460,21 +2499,19 @@ var MODELS = [
   {
     // `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.
+    // unchanged, same 1.6T/49B MoE). 1M context, text-only, low/high/max effort.
+    // 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",
     name: "DeepSeek V4 Pro",
     provider: "deepseek",
     contextWindow: 1048576,
-    maxOutputTokens: 393216,
+    maxOutputTokens: 384e3,
     supportsThinking: true,
     supportsImages: false,
     supportsVideo: false,
     costTier: "medium",
-    // DeepSeek V4 maps `xhigh` → its internal `max` tier.
-    maxThinkingLevel: "xhigh"
+    maxThinkingLevel: "max"
   },
   {
     id: "deepseek-v4-flash",
@@ -2486,7 +2523,20 @@ var MODELS = [
     supportsImages: false,
     supportsVideo: false,
     costTier: "low",
-    maxThinkingLevel: "xhigh"
+    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)",
+    provider: "deepseek",
+    contextWindow: 1048576,
+    maxOutputTokens: 384e3,
+    supportsThinking: true,
+    supportsImages: true,
+    supportsVideo: false,
+    costTier: "low",
+    maxThinkingLevel: "max"
   },
   // ── OpenRouter ─────────────────────────────────────────
   {
@@ -2496,8 +2546,10 @@ var MODELS = [
     contextWindow: 1e6,
     maxOutputTokens: 65536,
     supportsThinking: true,
-    supportsImages: false,
-    supportsVideo: false,
+    supportsImages: true,
+    supportsVideo: true,
+    // Practical inline-payload cap, not an asserted provider maximum.
+    maxVideoBytes: 20 * 1024 * 1024,
     costTier: "medium",
     maxThinkingLevel: "high"
   },
@@ -2653,7 +2705,8 @@ var OPENAI_GPT_56_THINKING_LEVELS = [
   "max",
   "ultra"
 ];
-var SAKANA_THINKING_LEVELS = ["high", "xhigh"];
+var SAKANA_THINKING_LEVELS = ["high", "xhigh", "max"];
+var DEEPSEEK_THINKING_LEVELS = ["low", "high", "max"];
 var XAI_THINKING_LEVELS = ["low", "medium", "high", "xhigh"];
 var ANTHROPIC_XHIGH_THINKING_LEVELS = [
   "low",
@@ -2717,13 +2770,14 @@ function getSupportedThinkingLevels(provider, model) {
     return XAI_THINKING_LEVELS.slice(0, maxIndex2 + 1);
   }
   if (isMoonshotK3Model(provider, model)) return MOONSHOT_K3_THINKING_LEVELS;
+  if (provider === "deepseek") return DEEPSEEK_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 levels = model.startsWith("gpt-5.6-") || model.startsWith("gpt-6-") ? OPENAI_GPT_56_THINKING_LEVELS : OPENAI_GPT_THINKING_LEVELS;
   const maxIndex = levels.indexOf(maxLevel);
   if (maxIndex === -1) return ["medium"];
   return levels.slice(0, maxIndex + 1);
@@ -2733,7 +2787,7 @@ 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) || // 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) || 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.
@@ -2751,7 +2805,7 @@ function getNextThinkingLevel(provider, model, current) {
 var DEFAULT_LOCAL_ENDPOINTS = [
   { id: "ollama", label: "Ollama", baseUrl: "http://127.0.0.1:11434/v1", kind: "ollama" }
 ];
-var FALLBACK_CONTEXT_WINDOW = 8192;
+var FALLBACK_CONTEXT_WINDOW = 4096;
 var LOCAL_API_KEY_PLACEHOLDER = "local";
 var DEFAULT_PROBE_TIMEOUT_MS = 1200;
 var ENRICH_CONCURRENCY = 6;
@@ -2844,7 +2898,7 @@ async function enrich(endpoint, entries, options) {
   return entries.map((entry) => genericModel(endpoint, entry));
 }
 function genericModel(endpoint, entry) {
-  const declared = typeof entry.max_model_len === "number" ? entry.max_model_len : void 0;
+  const declared = contextLength(entry.max_model_len);
   return {
     rawId: entry.id,
     endpointId: endpoint.id,
@@ -2857,16 +2911,23 @@ function genericModel(endpoint, entry) {
 }
 async function enrichOllama(endpoint, entries, options) {
   const showUrl = `${endpointRoot(endpoint.baseUrl)}/api/show`;
+  const running = await fetchJson(
+    `${endpointRoot(endpoint.baseUrl)}/api/ps`,
+    endpoint,
+    options
+  );
+  const runningModels = Array.isArray(running?.models) ? running.models : [];
   const enriched = await mapLimited(entries, ENRICH_CONCURRENCY, async (entry) => {
     const show = await fetchJson(showUrl, endpoint, {
       ...options,
       method: "POST",
       body: { model: entry.id }
     });
-    if (!show) return genericModel(endpoint, entry);
-    const caps = show.capabilities ?? [];
+    const caps = Array.isArray(show?.capabilities) ? show.capabilities : [];
     if (caps.includes("embedding") && !caps.includes("completion")) return void 0;
-    const ctx = ollamaContextLength(show.model_info);
+    const ctx = contextLength(
+      runningModels.find((model) => model?.name === entry.id || model?.model === entry.id)?.context_length
+    );
     return {
       rawId: entry.id,
       endpointId: endpoint.id,
@@ -2874,38 +2935,61 @@ async function enrichOllama(endpoint, entries, options) {
       contextWindowKnown: ctx !== void 0,
       // Ollama reports capabilities honestly, so trust it here rather than
       // using the optimistic generic default.
-      supportsTools: caps.includes("tools"),
+      supportsTools: show ? caps.includes("tools") : true,
       supportsImages: caps.includes("vision"),
       supportsThinking: caps.includes("thinking")
     };
   });
   return enriched.filter((model) => model !== void 0);
 }
-function ollamaContextLength(info) {
-  if (!info) return void 0;
-  for (const [key, value] of Object.entries(info)) {
-    if (key.endsWith(".context_length") && typeof value === "number" && value > 0) return value;
-  }
-  return void 0;
+function contextLength(value) {
+  return typeof value === "number" && Number.isSafeInteger(value) && value > 0 ? value : void 0;
 }
 async function enrichLmStudio(endpoint, entries, options) {
-  const detail = await fetchJson(
-    `${endpointRoot(endpoint.baseUrl)}/api/v0/models`,
-    endpoint,
-    options
-  );
-  if (!detail?.data) return entries.map((entry) => genericModel(endpoint, entry));
-  const byId = new Map(detail.data.filter((m) => m.id).map((m) => [m.id, m]));
+  const root = endpointRoot(endpoint.baseUrl);
+  const detail = await fetchJson(`${root}/api/v1/models`, endpoint, options);
+  const byId = /* @__PURE__ */ new Map();
+  if (Array.isArray(detail?.models)) {
+    for (const model of detail.models) {
+      if (typeof model?.key !== "string") continue;
+      const instances = Array.isArray(model.loaded_instances) ? model.loaded_instances : [];
+      const contexts = instances.map((instance) => contextLength(instance?.config?.context_length));
+      const ctx = contexts.length && contexts.every((value) => value !== void 0) ? contexts.reduce((min, value) => Math.min(min, value)) : void 0;
+      const info = {
+        id: model.key,
+        type: model.type === "llm" && model.capabilities?.vision === true ? "vlm" : model.type,
+        state: instances.length ? "loaded" : "not-loaded",
+        loaded_context_length: ctx
+      };
+      byId.set(model.key, info);
+      for (const instance of instances) {
+        if (typeof instance?.id === "string" && instance.id !== model.key) {
+          byId.set(instance.id, {
+            ...info,
+            id: instance.id,
+            loaded_context_length: contextLength(instance.config?.context_length)
+          });
+        }
+      }
+    }
+  } else {
+    const legacy = await fetchJson(`${root}/api/v0/models`, endpoint, options);
+    if (Array.isArray(legacy?.data)) {
+      for (const model of legacy.data) {
+        if (typeof model?.id === "string") byId.set(model.id, model);
+      }
+    }
+  }
   const models = [];
   for (const entry of entries) {
     const info = byId.get(entry.id);
     if (info && info.type !== "llm" && info.type !== "vlm") continue;
-    const ctx = info?.max_context_length;
+    const ctx = info?.state === "loaded" ? contextLength(info.loaded_context_length) : void 0;
     models.push({
       rawId: entry.id,
       endpointId: endpoint.id,
-      contextWindow: typeof ctx === "number" && ctx > 0 ? ctx : FALLBACK_CONTEXT_WINDOW,
-      contextWindowKnown: typeof ctx === "number" && ctx > 0,
+      contextWindow: ctx ?? FALLBACK_CONTEXT_WINDOW,
+      contextWindowKnown: ctx !== void 0,
       // LM Studio doesn't report tool support; it gates per-model at request time.
       supportsTools: true,
       supportsImages: info?.type === "vlm",
@@ -2921,8 +3005,8 @@ async function enrichLlamaCpp(endpoint, entries, options) {
     endpoint,
     options
   );
-  const nCtx = props?.default_generation_settings?.n_ctx;
-  const known = typeof nCtx === "number" && nCtx > 0;
+  const nCtx = contextLength(props?.default_generation_settings?.n_ctx);
+  const known = nCtx !== void 0;
   return entries.map((entry) => ({
     ...genericModel(endpoint, entry),
     contextWindow: known ? nCtx : FALLBACK_CONTEXT_WINDOW,