@prestyj/ai 5.7.0 → 5.9.0

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" | "sakana" | "palsu";
+type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "sakana" | "xai" | "palsu";
 type ThinkingLevel = "low" | "medium" | "high" | "xhigh" | "max" | "ultra";
 type CacheRetention = "none" | "short" | "long";
 interface TextContent {
@@ -41,6 +41,21 @@ interface ToolResult {
     toolCallId: string;
     content: ToolResultContent;
     isError?: boolean;
+    /**
+     * Set when the agent loop trimmed `content` to fit a per-result or per-turn
+     * budget. The provider (model input) and the persistent transcript both see
+     * the trimmed `content`, but the live `tool_call_end` event carried the FULL
+     * preview — so this marker makes that divergence explicit and reconcilable.
+     * Internal metadata only: it is never serialized onto the provider wire.
+     */
+    capped?: {
+        /** Length of the original, untrimmed string content. */
+        originalChars: number;
+        /** Length of the trimmed content actually sent to the model. */
+        keptChars: number;
+        /** Which budget triggered the trim. */
+        scope: "per-result" | "per-turn";
+    };
 }
 interface ServerToolCall {
     type: "server_tool_call";

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" | "sakana" | "palsu";
+type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "sakana" | "xai" | "palsu";
 type ThinkingLevel = "low" | "medium" | "high" | "xhigh" | "max" | "ultra";
 type CacheRetention = "none" | "short" | "long";
 interface TextContent {
@@ -41,6 +41,21 @@ interface ToolResult {
     toolCallId: string;
     content: ToolResultContent;
     isError?: boolean;
+    /**
+     * Set when the agent loop trimmed `content` to fit a per-result or per-turn
+     * budget. The provider (model input) and the persistent transcript both see
+     * the trimmed `content`, but the live `tool_call_end` event carried the FULL
+     * preview — so this marker makes that divergence explicit and reconcilable.
+     * Internal metadata only: it is never serialized onto the provider wire.
+     */
+    capped?: {
+        /** Length of the original, untrimmed string content. */
+        originalChars: number;
+        /** Length of the trimmed content actually sent to the model. */
+        keptChars: number;
+        /** Which budget triggered the trim. */
+        scope: "per-result" | "per-turn";
+    };
 }
 interface ServerToolCall {
     type: "server_tool_call";

package/dist/index.js

@@ -58,12 +58,14 @@ 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;
@@ -247,6 +249,9 @@ 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.`;
   }
@@ -259,6 +264,9 @@ function providerGuidance(provider, message, statusCode) {
   if (lower.includes("context_length_exceeded") || lower.includes("prompt is too long")) {
     return `Context window for this ${name} model is full. Compact the conversation to shrink history, or start a new session.`;
   }
+  if (lower.includes("many-image request") || lower.includes("image dimensions") && lower.includes("max allowed size")) {
+    return `An image in conversation history exceeds ${name}'s many-image limit. Restart EZ Coder so restored images are resized, then retry; if it persists, start a new session.`;
+  }
   if (statusCode === 413 || lower.includes("request_too_large") || lower.includes("request exceeds the maximum size")) {
     return `The request to ${name} is too large. Compact the conversation to shrink history, or start a new session.`;
   }
@@ -755,9 +763,14 @@ function toAnthropicMessages(messages, cacheControl) {
       continue;
     }
     if (msg.role === "user") {
+      if (typeof msg.content === "string") {
+        if (msg.content === "") continue;
+      } else if (!msg.content.some((p) => !(p.type === "text" && p.text === ""))) {
+        continue;
+      }
       out.push({
         role: "user",
-        content: typeof msg.content === "string" ? msg.content : msg.content.map((part) => {
+        content: typeof msg.content === "string" ? msg.content : msg.content.filter((part) => !(part.type === "text" && part.text === "")).map((part) => {
           if (part.type === "text") return { type: "text", text: part.text };
           if (part.type === "video") {
             return {
@@ -782,6 +795,7 @@ function toAnthropicMessages(messages, cacheControl) {
       continue;
     }
     if (msg.role === "assistant") {
+      if (typeof msg.content === "string" && msg.content === "") continue;
       const content = typeof msg.content === "string" ? msg.content : toAnthropicAssistantContent(msg.content, msgIdx > trajectoryStartIdx, idMap);
       if (Array.isArray(content) && content.length === 0) continue;
       out.push({ role: "assistant", content });
@@ -902,7 +916,7 @@ function remapToolCallId(id, idMap) {
   if (!id.startsWith("toolu_")) return id;
   const existing = idMap.get(id);
   if (existing) return existing;
-  const mapped = `call_${id.slice(5)}`;
+  const mapped = `call_${id.slice(6)}`;
   idMap.set(id, mapped);
   return mapped;
 }
@@ -1513,6 +1527,12 @@ async function* runStream(options) {
       statusCode: 504
     });
   }
+  if (stopReason === null) {
+    throw new ProviderError("anthropic", "Stream ended before completion (no stop_reason).", {
+      statusCode: 504,
+      cause: { partialContent: contentParts, outputTokens }
+    });
+  }
   const normalizedStop = normalizeAnthropicStopReason(stopReason);
   const response = {
     message: {
@@ -1779,6 +1799,17 @@ function getEnvironment() {
 }
 
 // src/providers/openai.ts
+function toKimiK3Effort(level) {
+  switch (level) {
+    case "low":
+      return "low";
+    case "medium":
+    case "high":
+      return "high";
+    default:
+      return "max";
+  }
+}
 function extractOpenAIUsage(usage) {
   let cacheRead = 0;
   let cacheWrite = 0;
@@ -1833,7 +1864,12 @@ 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 k3Effort = options.thinking ? toKimiK3Effort(options.thinking) : void 0;
+  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") {
@@ -1845,7 +1881,11 @@ async function* runStream2(options) {
   }
   const messages = toOpenAIMessages(downgradedMessages, {
     provider: options.provider,
-    thinking: !!options.thinking,
+    // K2.7 preserves reasoning even when the user hides thinking in the UI;
+    // keep assistant tool-call history wire-valid in that display mode. A
+    // disabled K3 must NOT carry placeholder reasoning_content (mirrors the
+    // official CLI: reasoning is preserved only while thinking is enabled).
+    thinking: isKimiK27 || !!options.thinking,
     supportsImages: options.supportsImages
   });
   const defaultTemp = options.provider === "glm" ? 0.6 : void 0;
@@ -1855,10 +1895,10 @@ 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 } } : {}
@@ -1868,13 +1908,23 @@ 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 = k3Effort ? { type: "enabled", effort: k3Effort, keep: "all" } : { type: "disabled" };
+    } else if (k3Effort) {
+      paramsAny.reasoning_effort = k3Effort;
+    } else {
+      paramsAny.thinking = { type: "disabled" };
+    }
+  }
   if (usesThinkingParam) {
     if (options.thinking) {
       params.thinking = { type: "enabled" };
@@ -1978,6 +2028,12 @@ async function* runStream2(options) {
       statusCode: 504
     });
   }
+  if (finishReason === null) {
+    throw new ProviderError(providerName, "Stream ended before completion (no finish_reason).", {
+      statusCode: 504,
+      cause: { partialText: textAccum, outputTokens }
+    });
+  }
   if (thinkingAccum) {
     contentParts.push({ type: "thinking", text: thinkingAccum });
   }
@@ -2170,6 +2226,7 @@ function toError2(err, provider = "openai") {
 
 // src/providers/openai-codex.ts
 import os from "os";
+import * as zstd from "@bokuweb/zstd-wasm";
 
 // src/utils/sse.ts
 function parseSseBuffer(buffer) {
@@ -2225,6 +2282,50 @@ 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-");
 }
@@ -2305,10 +2406,17 @@ 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) {
@@ -3347,6 +3455,18 @@ 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,