@prestyj/ai 4.13.3 → 5.0.0

package/dist/index.cjs

@@ -112,6 +112,7 @@ var PROVIDER_DISPLAY = {
   moonshot: "Moonshot",
   deepseek: "DeepSeek",
   openrouter: "OpenRouter",
+  sakana: "Sakana",
   xiaomi: "Xiaomi (MiMo)",
   minimax: "MiniMax"
 };
@@ -1020,8 +1021,12 @@ function toOpenAIToolChoice(choice) {
   if (choice === "required") return "required";
   return { type: "function", function: { name: choice.name } };
 }
-function toOpenAIReasoningEffort(level, _model) {
-  return level === "max" ? "xhigh" : level;
+function toOpenAIReasoningEffort(level, model) {
+  const effort = level === "max" ? "xhigh" : level;
+  if (model.startsWith("fugu") && (effort === "low" || effort === "medium")) {
+    return "high";
+  }
+  return effort;
 }
 function normalizeAnthropicStopReason(reason) {
   switch (reason) {
@@ -1144,7 +1149,13 @@ async function prewarmAnthropicCache(options) {
       },
       {
         signal: options.signal ?? void 0,
-        ...isOAuth ? { headers: { "anthropic-beta": "claude-code-20250219,oauth-2025-04-20" } } : {}
+        ...(() => {
+          const betas = [
+            ...isOAuth ? ["claude-code-20250219", "oauth-2025-04-20"] : [],
+            ...cacheControl?.ttl === "1h" ? ["extended-cache-ttl-2025-04-11"] : []
+          ];
+          return betas.length ? { headers: { "anthropic-beta": betas.join(",") } } : {};
+        })()
       }
     );
   } catch {
@@ -1228,7 +1239,12 @@ async function* runStream(options) {
     ...options.compaction ? ["compact-2026-01-12"] : [],
     ...options.clearToolUses ? ["context-management-2025-06-27"] : [],
     "fine-grained-tool-streaming-2025-05-14",
-    ...!hasAdaptiveThinking ? ["interleaved-thinking-2025-05-14"] : []
+    ...!hasAdaptiveThinking ? ["interleaved-thinking-2025-05-14"] : [],
+    // The 1-h cache TTL (cacheRetention "long") is gated behind this beta. Without
+    // it Anthropic silently ignores ttl:"1h" and falls back to the 5-min default,
+    // so a pre-warmed cache expires before the user's first turn. cacheControl.ttl
+    // is only "1h" on the first-party endpoint (see toAnthropicCacheControl).
+    ...cacheControl?.ttl === "1h" ? ["extended-cache-ttl-2025-04-11"] : []
   ];
   const requestOptions = {
     signal: options.signal ?? void 0,
@@ -3184,6 +3200,16 @@ providerRegistry.register("openrouter", {
     baseUrl: options.baseUrl ?? "https://openrouter.ai/api/v1"
   })
 });
+providerRegistry.register("sakana", {
+  // Sakana Fugu is a multi-agent system exposed as a standard LLM through the
+  // OpenAI-compatible Sakana API. We ride the Chat Completions transport (the
+  // Responses API is also offered). Fugu models only accept "high"/"xhigh"
+  // reasoning effort — clamped centrally in toOpenAIReasoningEffort.
+  stream: (options) => streamOpenAI({
+    ...options,
+    baseUrl: options.baseUrl ?? "https://api.sakana.ai/v1"
+  })
+});
 providerRegistry.register("minimax", {
   stream: (options) => streamAnthropic({
     ...options,