@prestyj/ai 4.12.1 → 4.14.0

package/dist/index.d.cts

@@ -2,7 +2,7 @@ import { z } from 'zod';
 import Anthropic from '@anthropic-ai/sdk';
 import OpenAI from 'openai';
 
-type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "palsu";
+type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "sakana" | "palsu";
 type ThinkingLevel = "low" | "medium" | "high" | "xhigh" | "max";
 type CacheRetention = "none" | "short" | "long";
 interface TextContent {
@@ -255,6 +255,18 @@ declare class StreamResult implements AsyncIterable<StreamEvent> {
     private resolveResponse;
     private rejectResponse;
     private resolveWait;
+    /**
+     * High-water mark: when the buffer exceeds this many unconsumed events,
+     * the pump pauses until the consumer drains below the low-water mark.
+     * Prevents unbounded memory growth when a consumer is slow.
+     * Only active when someone IS iterating — if nobody iterates (the `then()`
+     * path), backpressure is skipped so the pump can complete and resolve.
+     */
+    private static readonly HIGH_WATER;
+    private static readonly LOW_WATER;
+    private iterating;
+    private paused;
+    private resolveDrain;
     constructor(generator: AsyncGenerator<StreamEvent, StreamResponse>, signal?: AbortSignal);
     private pump;
     private _nextWithAbort;

package/dist/index.d.ts

@@ -2,7 +2,7 @@ import { z } from 'zod';
 import Anthropic from '@anthropic-ai/sdk';
 import OpenAI from 'openai';
 
-type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "palsu";
+type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "sakana" | "palsu";
 type ThinkingLevel = "low" | "medium" | "high" | "xhigh" | "max";
 type CacheRetention = "none" | "short" | "long";
 interface TextContent {
@@ -255,6 +255,18 @@ declare class StreamResult implements AsyncIterable<StreamEvent> {
     private resolveResponse;
     private rejectResponse;
     private resolveWait;
+    /**
+     * High-water mark: when the buffer exceeds this many unconsumed events,
+     * the pump pauses until the consumer drains below the low-water mark.
+     * Prevents unbounded memory growth when a consumer is slow.
+     * Only active when someone IS iterating — if nobody iterates (the `then()`
+     * path), backpressure is skipped so the pump can complete and resolve.
+     */
+    private static readonly HIGH_WATER;
+    private static readonly LOW_WATER;
+    private iterating;
+    private paused;
+    private resolveDrain;
     constructor(generator: AsyncGenerator<StreamEvent, StreamResponse>, signal?: AbortSignal);
     private pump;
     private _nextWithAbort;

package/dist/index.js

@@ -57,6 +57,7 @@ var PROVIDER_DISPLAY = {
   moonshot: "Moonshot",
   deepseek: "DeepSeek",
   openrouter: "OpenRouter",
+  sakana: "Sakana",
   xiaomi: "Xiaomi (MiMo)",
   minimax: "MiniMax"
 };
@@ -281,7 +282,7 @@ var EventStream = class {
     }
   }
 };
-var StreamResult = class {
+var StreamResult = class _StreamResult {
   response;
   buffer = [];
   done = false;
@@ -289,6 +290,18 @@ var StreamResult = class {
   resolveResponse;
   rejectResponse;
   resolveWait = null;
+  /**
+   * High-water mark: when the buffer exceeds this many unconsumed events,
+   * the pump pauses until the consumer drains below the low-water mark.
+   * Prevents unbounded memory growth when a consumer is slow.
+   * Only active when someone IS iterating — if nobody iterates (the `then()`
+   * path), backpressure is skipped so the pump can complete and resolve.
+   */
+  static HIGH_WATER = 5e3;
+  static LOW_WATER = 1e3;
+  iterating = false;
+  paused = false;
+  resolveDrain = null;
   constructor(generator, signal) {
     this.response = new Promise((resolve, reject) => {
       this.resolveResponse = resolve;
@@ -303,6 +316,13 @@ var StreamResult = class {
         this.buffer.push(next.value);
         this.resolveWait?.();
         this.resolveWait = null;
+        if (this.iterating && this.buffer.length > _StreamResult.HIGH_WATER) {
+          this.paused = true;
+          await new Promise((r) => {
+            this.resolveDrain = r;
+          });
+          this.paused = false;
+        }
         next = await this._nextWithAbort(generator, signal);
       }
       this.done = true;
@@ -341,11 +361,20 @@ var StreamResult = class {
     }
   }
   async *[Symbol.asyncIterator]() {
+    this.iterating = true;
     let index = 0;
     while (true) {
       while (index < this.buffer.length) {
         yield this.buffer[index++];
       }
+      if (this.paused && index > _StreamResult.LOW_WATER) {
+        this.resolveDrain?.();
+        this.resolveDrain = null;
+      }
+      if (index > 0 && !this.paused) {
+        this.buffer.splice(0, index);
+        index = 0;
+      }
       if (this.error) throw this.error;
       if (this.done) return;
       await new Promise((r) => {
@@ -358,6 +387,11 @@ var StreamResult = class {
     }
   }
   then(onfulfilled, onrejected) {
+    if (this.paused) {
+      this.paused = false;
+      this.resolveDrain?.();
+      this.resolveDrain = null;
+    }
     return this.response.then(onfulfilled, onrejected);
   }
 };
@@ -932,8 +966,12 @@ 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) {
@@ -3096,6 +3134,16 @@ 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,