@prestyj/ai 5.7.0 → 5.8.0

package/README.md

@@ -43,7 +43,7 @@ Tool parameters are Zod schemas. Converted to JSON Schema at the provider bounda
 | `anthropic` | Claude Opus 4.8, Sonnet 5, Haiku 4.5 | Extended thinking, prompt caching, server-side compaction |
 | `openai` | GPT-4.1, o3, o4-mini | Supports OAuth (codex endpoint) and API keys |
 | `glm` | GLM-5.1, GLM-4.7 | Z.AI platform, OpenAI-compatible |
-| `moonshot` | Kimi K2.7 | Moonshot platform, OpenAI-compatible |
+| `moonshot` | Kimi K3, Kimi K2.7 Code | Moonshot platform, OpenAI-compatible |
 
 ---
 

package/dist/index.cjs

@@ -118,12 +118,14 @@ var PROVIDER_DISPLAY = {
   deepseek: "DeepSeek",
   openrouter: "OpenRouter",
   sakana: "Sakana",
+  xai: "xAI (Grok)",
   xiaomi: "Xiaomi (MiMo)",
   minimax: "MiniMax"
 };
 var PROVIDER_STATUS_URL = {
   openai: "status.openai.com",
-  anthropic: "status.anthropic.com"
+  anthropic: "status.anthropic.com",
+  xai: "status.x.ai"
 };
 function providerDisplayName(provider) {
   return PROVIDER_DISPLAY[provider] ?? provider;
@@ -307,6 +309,9 @@ function providerGuidance(provider, message, statusCode) {
   if (statusCode === 503 || lower.includes("service unavailable")) {
     return `${name} is temporarily unavailable. Retry shortly \u2014 not a EZ Coder issue.`;
   }
+  if (statusCode === 507 || lower.includes("exceeded request buffer limit while retrying upstream")) {
+    return `${name}'s proxy could not retry this large request. EZ Coder already retried automatically \u2014 compact the conversation, then retry.`;
+  }
   if (statusCode === 500 || lower.includes("server_error") || lower.includes("500") && lower.includes("internal server error")) {
     return status ? `This is an error from ${name}, not EZ Coder. Retry \u2014 if it keeps happening, check ${status}.` : `This is an error from ${name}, not EZ Coder. Retry \u2014 if it keeps happening, try a different model via the model selector.`;
   }
@@ -1893,7 +1898,11 @@ async function* runStream2(options) {
   const providerName = options.provider ?? "openai";
   const useStreaming = options.streaming !== false;
   const client = createClient2(options);
-  const usesThinkingParam = options.provider === "glm" || options.provider === "moonshot" || options.provider === "xiaomi";
+  const isKimiK3 = options.provider === "moonshot" && options.model === "kimi-k3";
+  const isManagedKimiK3 = isKimiK3 && options.baseUrl?.replace(/\/+$/, "").endsWith("/coding/v1") === true;
+  const isKimiK27 = options.provider === "moonshot" && options.model.startsWith("kimi-k2.7-code");
+  const hasFixedKimiSampling = isKimiK3 || isKimiK27;
+  const usesThinkingParam = options.provider === "glm" || options.provider === "moonshot" && !isKimiK3 && !isKimiK27 || options.provider === "xiaomi";
   const downgradedImages = downgradeUnsupportedImages(options.messages, options.supportsImages);
   const downgradedMessages = downgradeUnsupportedVideos(downgradedImages, options.supportsVideo);
   if (options.provider === "moonshot") {
@@ -1905,7 +1914,9 @@ async function* runStream2(options) {
   }
   const messages = toOpenAIMessages(downgradedMessages, {
     provider: options.provider,
-    thinking: !!options.thinking,
+    // K3 and K2.7 preserve reasoning even when the user hides thinking in the
+    // UI; keep assistant tool-call history wire-valid in that display mode.
+    thinking: isKimiK3 || isKimiK27 || !!options.thinking,
     supportsImages: options.supportsImages
   });
   const defaultTemp = options.provider === "glm" ? 0.6 : void 0;
@@ -1915,10 +1926,10 @@ async function* runStream2(options) {
     messages,
     stream: useStreaming,
     ...options.maxTokens ? { max_completion_tokens: options.maxTokens } : {},
-    ...effectiveTemp != null && !options.thinking ? { temperature: effectiveTemp } : {},
-    ...options.topP != null ? { top_p: options.topP } : {},
+    ...effectiveTemp != null && !options.thinking && !hasFixedKimiSampling ? { temperature: effectiveTemp } : {},
+    ...options.topP != null && !hasFixedKimiSampling ? { top_p: options.topP } : {},
     ...options.stop ? { stop: options.stop } : {},
-    ...options.thinking && !usesThinkingParam ? { reasoning_effort: toOpenAIReasoningEffort(options.thinking, options.model) } : {},
+    ...options.thinking && !usesThinkingParam && !isKimiK3 && !isKimiK27 ? { reasoning_effort: toOpenAIReasoningEffort(options.thinking, options.model) } : {},
     ...options.tools?.length ? { tools: toOpenAITools(options.tools) } : {},
     ...options.toolChoice && options.tools?.length ? { tool_choice: toOpenAIToolChoice(options.toolChoice) } : {},
     ...useStreaming ? { stream_options: { include_usage: true } } : {}
@@ -1928,13 +1939,21 @@ async function* runStream2(options) {
     paramsAny.prompt_cache_key = normalizePromptCacheKey(options.promptCacheKey ?? "ezcoder");
     if (options.provider === "openai" && options.model.startsWith("gpt-5.6")) {
       paramsAny.prompt_cache_options = { mode: "implicit", ttl: "30m" };
-    } else if ((options.cacheRetention ?? "short") === "long") {
+    } else if (!isKimiK3 && (options.cacheRetention ?? "short") === "long") {
       paramsAny.prompt_cache_retention = "24h";
     }
   }
   if (options.provider === "openai" && options.serviceTier) {
     params.service_tier = options.serviceTier;
   }
+  if (isKimiK3) {
+    const paramsAny = params;
+    if (isManagedKimiK3) {
+      paramsAny.thinking = { type: "enabled", effort: "max", keep: "all" };
+    } else {
+      paramsAny.reasoning_effort = "max";
+    }
+  }
   if (usesThinkingParam) {
     if (options.thinking) {
       params.thinking = { type: "enabled" };
@@ -2230,6 +2249,7 @@ function toError2(err, provider = "openai") {
 
 // src/providers/openai-codex.ts
 var import_node_os = __toESM(require("os"), 1);
+var zstd = __toESM(require("@bokuweb/zstd-wasm"), 1);
 
 // src/utils/sse.ts
 function parseSseBuffer(buffer) {
@@ -2285,6 +2305,50 @@ function extractRequestIdFromMessage(message) {
 // src/providers/openai-codex.ts
 var DEFAULT_BASE_URL = "https://chatgpt.com/backend-api";
 var CODEX_CLIENT_VERSION = "0.144.1";
+var CODEX_REQUEST_COMPRESSION_MIN_BYTES = 16 * 1024;
+var zstdInitPromise;
+async function encodeCodexRequest(body) {
+  const json = JSON.stringify(body);
+  const raw = new TextEncoder().encode(json);
+  if (raw.byteLength < CODEX_REQUEST_COMPRESSION_MIN_BYTES) {
+    return {
+      body: json,
+      compressed: false,
+      rawBytes: raw.byteLength,
+      encodedBytes: raw.byteLength
+    };
+  }
+  try {
+    zstdInitPromise ??= zstd.init();
+    await zstdInitPromise;
+    const compressed = Uint8Array.from(zstd.compress(raw));
+    if (compressed.byteLength >= raw.byteLength) {
+      return {
+        body: json,
+        compressed: false,
+        rawBytes: raw.byteLength,
+        encodedBytes: raw.byteLength
+      };
+    }
+    return {
+      body: compressed,
+      compressed: true,
+      rawBytes: raw.byteLength,
+      encodedBytes: compressed.byteLength
+    };
+  } catch (error) {
+    providerDiag("codex_request_compression_failed", {
+      error: error instanceof Error ? error.message : String(error),
+      rawBytes: raw.byteLength
+    });
+    return {
+      body: json,
+      compressed: false,
+      rawBytes: raw.byteLength,
+      encodedBytes: raw.byteLength
+    };
+  }
+}
 function usesResponsesLite(model) {
   return model.startsWith("gpt-5.6-");
 }
@@ -2365,10 +2429,17 @@ async function* runStream3(options) {
     headers["session_id"] = transportSessionId;
     headers["x-client-request-id"] = transportSessionId;
   }
+  const encodedRequest = await encodeCodexRequest(body);
+  if (encodedRequest.compressed) headers["Content-Encoding"] = "zstd";
+  providerDiag("codex_request_body", {
+    rawBytes: encodedRequest.rawBytes,
+    encodedBytes: encodedRequest.encodedBytes,
+    compressed: encodedRequest.compressed
+  });
   const response = await fetch(url, {
     method: "POST",
     headers,
-    body: JSON.stringify(body),
+    body: encodedRequest.body,
     signal: options.signal
   });
   if (!response.ok) {
@@ -3407,6 +3478,18 @@ providerRegistry.register("sakana", {
     baseUrl: options.baseUrl ?? "https://api.sakana.ai/v1"
   })
 });
+providerRegistry.register("xai", {
+  // xAI's public API (console.x.ai key) is OpenAI-compatible — ride the Chat
+  // Completions transport like Moonshot/DeepSeek. Grok reasoning models take
+  // top-level `reasoning_effort` (low/medium/high), which the shared thinking
+  // path already sends. xAI's OAuth path exists but only via the Grok CLI's
+  // private Responses proxy (cli-chat-proxy.grok.com) with reverse-engineered
+  // attribution headers and account-tier gating — intentionally not wired.
+  stream: (options) => streamOpenAI({
+    ...options,
+    baseUrl: options.baseUrl ?? "https://api.x.ai/v1"
+  })
+});
 providerRegistry.register("minimax", {
   stream: (options) => streamAnthropic({
     ...options,