@prestyj/ai 5.6.0 → 5.8.0

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 {
@@ -444,6 +444,27 @@ declare function formatErrorForDisplay(err: unknown): string;
  */
 declare function classifyProviderError(message: string): string;
 
+declare const REDACTED = "[REDACTED]";
+interface RedactionOptions {
+    /** Exact secret values to remove in addition to high-confidence formats. */
+    secrets?: Iterable<string>;
+    /** Maximum recursive object depth before a stable truncation marker is emitted. */
+    maxDepth?: number;
+    /** Maximum total array/object entries cloned before truncation markers are emitted. */
+    maxEntries?: number;
+    /** Maximum retained string length after sanitization. */
+    maxStringLength?: number;
+}
+/** Collect sufficiently distinctive secrets from security-sensitive environment variables. */
+declare function environmentSecrets(env: Record<string, string | undefined>): string[];
+/** Redact credentials from arbitrary text without mutating its source. */
+declare function redactText(text: string, options?: RedactionOptions): string;
+/**
+ * Recursively clone and sanitize transport/persistence payloads.
+ * Cycles, excessive depth, and excessive collection sizes become stable markers.
+ */
+declare function redactValue<T>(value: T, options?: RedactionOptions): T;
+
 /**
  * Provider-level diagnostic hook. Mirrors the pattern used by gg-agent's
  * setStreamDiagnostic — the host app wires a callback (typically writing to
@@ -454,6 +475,12 @@ type ProviderDiagnosticFn = (phase: string, data?: Record<string, unknown>) => v
 /** Register a diagnostic callback for provider-level tracing. */
 declare function setProviderDiagnostic(fn: ProviderDiagnosticFn | null): void;
 
+/**
+ * Cap historical images before provider dispatch, removing the oldest first.
+ * The persisted/live conversation is never mutated; only modified messages and
+ * tool results are cloned for the outgoing request.
+ */
+declare function clampProviderContextImages(messages: Message[], provider: Provider, supportsImages: boolean | undefined): Message[];
 declare function toAnthropicMessages(messages: Message[], cacheControl?: {
     type: "ephemeral";
     ttl?: "1h";
@@ -558,4 +585,4 @@ interface PalsuProviderConfig {
  */
 declare function registerPalsuProvider(config?: PalsuProviderConfig): PalsuProviderHandle;
 
-export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, EZCoderAIError, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, type ImageContent, type Message, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, type RawContent, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, classifyProviderError, formatError, formatErrorForDisplay, isHardBillingMessage, isUsageLimitError, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, registerPalsuProvider, setProviderDiagnostic, stream, toAnthropicMessages, toOpenAIMessages };
+export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, EZCoderAIError, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, type ImageContent, type Message, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, clampProviderContextImages, classifyProviderError, environmentSecrets, formatError, formatErrorForDisplay, isHardBillingMessage, isUsageLimitError, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, setProviderDiagnostic, stream, toAnthropicMessages, toOpenAIMessages };

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 {
@@ -444,6 +444,27 @@ declare function formatErrorForDisplay(err: unknown): string;
  */
 declare function classifyProviderError(message: string): string;
 
+declare const REDACTED = "[REDACTED]";
+interface RedactionOptions {
+    /** Exact secret values to remove in addition to high-confidence formats. */
+    secrets?: Iterable<string>;
+    /** Maximum recursive object depth before a stable truncation marker is emitted. */
+    maxDepth?: number;
+    /** Maximum total array/object entries cloned before truncation markers are emitted. */
+    maxEntries?: number;
+    /** Maximum retained string length after sanitization. */
+    maxStringLength?: number;
+}
+/** Collect sufficiently distinctive secrets from security-sensitive environment variables. */
+declare function environmentSecrets(env: Record<string, string | undefined>): string[];
+/** Redact credentials from arbitrary text without mutating its source. */
+declare function redactText(text: string, options?: RedactionOptions): string;
+/**
+ * Recursively clone and sanitize transport/persistence payloads.
+ * Cycles, excessive depth, and excessive collection sizes become stable markers.
+ */
+declare function redactValue<T>(value: T, options?: RedactionOptions): T;
+
 /**
  * Provider-level diagnostic hook. Mirrors the pattern used by gg-agent's
  * setStreamDiagnostic — the host app wires a callback (typically writing to
@@ -454,6 +475,12 @@ type ProviderDiagnosticFn = (phase: string, data?: Record<string, unknown>) => v
 /** Register a diagnostic callback for provider-level tracing. */
 declare function setProviderDiagnostic(fn: ProviderDiagnosticFn | null): void;
 
+/**
+ * Cap historical images before provider dispatch, removing the oldest first.
+ * The persisted/live conversation is never mutated; only modified messages and
+ * tool results are cloned for the outgoing request.
+ */
+declare function clampProviderContextImages(messages: Message[], provider: Provider, supportsImages: boolean | undefined): Message[];
 declare function toAnthropicMessages(messages: Message[], cacheControl?: {
     type: "ephemeral";
     ttl?: "1h";
@@ -558,4 +585,4 @@ interface PalsuProviderConfig {
  */
 declare function registerPalsuProvider(config?: PalsuProviderConfig): PalsuProviderHandle;
 
-export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, EZCoderAIError, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, type ImageContent, type Message, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, type RawContent, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, classifyProviderError, formatError, formatErrorForDisplay, isHardBillingMessage, isUsageLimitError, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, registerPalsuProvider, setProviderDiagnostic, stream, toAnthropicMessages, toOpenAIMessages };
+export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, EZCoderAIError, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, type ImageContent, type Message, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, clampProviderContextImages, classifyProviderError, environmentSecrets, formatError, formatErrorForDisplay, isHardBillingMessage, isUsageLimitError, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, setProviderDiagnostic, stream, toAnthropicMessages, toOpenAIMessages };

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;
@@ -102,13 +104,20 @@ function isRawJsonErrorEcho(message) {
     return false;
   }
 }
+function isRawHtmlErrorEcho(message) {
+  const withoutStatus = message.trimStart().replace(/^\d{3}\s+/, "").trimStart();
+  return /^<!doctype\s+html(?:\s|>)/i.test(withoutStatus) || /^<html(?:\s|>)/i.test(withoutStatus);
+}
+function providerHtmlErrorMessage(statusCode) {
+  return statusCode ? `The provider returned an HTML error page (HTTP ${statusCode}) instead of an API response.` : "The provider returned an HTML error page instead of an API response.";
+}
 function emptyProviderErrorMessage(statusCode) {
   return statusCode ? `The provider returned an empty error response (HTTP ${statusCode}), with no further detail.` : "The provider returned an empty error response, with no further detail.";
 }
 function formatError(err) {
   if (err instanceof ProviderError) {
     const name = providerDisplayName(err.provider);
-    const cleanMessage = cleanProviderMessage(err.message);
+    const cleanMessage = cleanProviderMessage(err.message, err.statusCode);
     if (isMythosAccessError(cleanMessage)) {
       return {
         headline: "Claude Mythos 5 is invitation-only.",
@@ -203,8 +212,9 @@ function formatErrorForDisplay(err) {
   lines.push(`  \u2192 ${f.guidance}`);
   return lines.join("\n");
 }
-function cleanProviderMessage(message) {
-  return message.replace(/^\[[^\]]+\]\s*/, "").trim();
+function cleanProviderMessage(message, statusCode) {
+  const clean = message.replace(/^\[[^\]]+\]\s*/, "").trim();
+  return isRawHtmlErrorEcho(clean) ? providerHtmlErrorMessage(statusCode) : clean;
 }
 function inferSource(err) {
   const msg = err.message.toLowerCase();
@@ -239,6 +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.`;
   }
@@ -565,6 +578,66 @@ function toAnthropicAssistantContent(content, preserveThinking, idMap) {
     return true;
   }).map((part) => toAnthropicAssistantPart(part, idMap)).filter((b) => b !== null);
 }
+var PROVIDER_IMAGE_LIMIT_PLACEHOLDER = "[image omitted: provider image limit]";
+var PROVIDER_IMAGE_BUDGETS = {
+  anthropic: 90,
+  minimax: 90,
+  openai: 200,
+  gemini: 200,
+  openrouter: 90
+};
+function countContextImages(messages) {
+  let count = 0;
+  for (const message of messages) {
+    if (message.role === "user" && Array.isArray(message.content)) {
+      count += message.content.filter((part) => part.type === "image").length;
+    } else if (message.role === "tool") {
+      for (const result of message.content) {
+        if (Array.isArray(result.content)) {
+          count += result.content.filter((part) => part.type === "image").length;
+        }
+      }
+    }
+  }
+  return count;
+}
+function clampProviderContextImages(messages, provider, supportsImages) {
+  if (supportsImages === false) return messages;
+  const budget = PROVIDER_IMAGE_BUDGETS[provider] ?? 5;
+  let remainingToRemove = countContextImages(messages) - budget;
+  if (remainingToRemove <= 0) return messages;
+  return messages.map((message) => {
+    if (message.role === "user" && Array.isArray(message.content)) {
+      const content = message.content.filter((part) => {
+        if (part.type !== "image" || remainingToRemove <= 0) return true;
+        remainingToRemove--;
+        return false;
+      });
+      return {
+        ...message,
+        content: content.length > 0 ? content : [{ type: "text", text: PROVIDER_IMAGE_LIMIT_PLACEHOLDER }]
+      };
+    }
+    if (message.role === "tool") {
+      return {
+        ...message,
+        content: message.content.map((result) => {
+          if (!Array.isArray(result.content)) return result;
+          const content = result.content.filter((part) => {
+            if (part.type !== "image" || remainingToRemove <= 0) return true;
+            remainingToRemove--;
+            return false;
+          });
+          return {
+            ...result,
+            content: content.length > 0 ? content : [{ type: "text", text: PROVIDER_IMAGE_LIMIT_PLACEHOLDER }]
+          };
+        })
+      };
+    }
+    return message;
+  });
+}
 var NON_VISION_USER_IMAGE_PLACEHOLDER = "(image omitted: model does not support images)";
 var NON_VISION_TOOL_IMAGE_PLACEHOLDER = "(tool image omitted: model does not support images)";
 var NON_VIDEO_USER_PLACEHOLDER = "(video omitted: model does not support video)";
@@ -1582,7 +1655,8 @@ function toError(err) {
     const bodyMessage = typeof nestedError?.message === "string" && nestedError.message.trim() ? nestedError.message.trim() : typeof errorBody?.message === "string" && errorBody.message.trim() ? errorBody.message.trim() : void 0;
     const bodyType = typeof nestedError?.type === "string" ? nestedError.type : typeof errorBody?.type === "string" ? errorBody.type : typeof err.type === "string" ? err.type : void 0;
     const fallbackMessage = isRawJsonErrorEcho(err.message) ? emptyProviderErrorMessage(err.status) : err.message;
-    const message = bodyType && bodyMessage ? `${bodyType}: ${bodyMessage}` : bodyMessage ?? fallbackMessage;
+    const messageCandidate = bodyMessage ?? err.message;
+    const message = isRawHtmlErrorEcho(messageCandidate) ? providerHtmlErrorMessage(err.status) : bodyType && bodyMessage ? `${bodyType}: ${bodyMessage}` : bodyMessage ?? fallbackMessage;
     if (err.status === 429) {
       const limit = readUnifiedRateLimit(err.headers);
       const farOff = limit.resetsAt != null && limit.resetsAt * 1e3 - Date.now() > 6e4;
@@ -1764,7 +1838,11 @@ async function* runStream2(options) {
   const providerName = options.provider ?? "openai";
   const useStreaming = options.streaming !== false;
   const client = createClient2(options);
-  const usesThinkingParam = options.provider === "glm" || options.provider === "moonshot" || options.provider === "xiaomi";
+  const isKimiK3 = options.provider === "moonshot" && options.model === "kimi-k3";
+  const isManagedKimiK3 = isKimiK3 && options.baseUrl?.replace(/\/+$/, "").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") {
@@ -1776,7 +1854,9 @@ 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;
@@ -1786,10 +1866,10 @@ 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 } } : {}
@@ -1799,13 +1879,21 @@ 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" };
@@ -2050,7 +2138,8 @@ function toError2(err, provider = "openai") {
     const body = err.error;
     const bodyMessage = typeof body?.message === "string" && body.message.trim() ? body.message.trim() : void 0;
     const modelName = typeof body?.model === "string" ? body.model : "";
-    const cleanMessage = bodyMessage ?? (isRawJsonErrorEcho(err.message) ? emptyProviderErrorMessage(err.status) : err.message);
+    const messageCandidate = bodyMessage ?? err.message;
+    const cleanMessage = isRawHtmlErrorEcho(messageCandidate) ? providerHtmlErrorMessage(err.status) : bodyMessage ? bodyMessage : isRawJsonErrorEcho(err.message) ? emptyProviderErrorMessage(err.status) : err.message;
     let hint;
     if (modelName === "codex-mini-latest" || cleanMessage.includes("codex-mini-latest")) {
       hint = "codex-mini-latest requires an OpenAI Pro or Max subscription. Your account currently has access to GPT-5.4 and GPT-5.4 Mini.";
@@ -2100,6 +2189,7 @@ 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) {
@@ -2155,6 +2245,50 @@ 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-");
 }
@@ -2164,6 +2298,24 @@ function outputTextKey(itemId, contentIndex) {
 function isVisibleOutputItem(itemType) {
   return itemType === "message";
 }
+function toCodexToolChoice(choice, tools) {
+  const resolved = choice ?? "auto";
+  if (typeof resolved === "object") {
+    throw new EZCoderAIError(
+      `OpenAI Codex does not support selecting the named tool \`${resolved.name}\`; use auto, none, or required.`,
+      { source: "capability" }
+    );
+  }
+  if (resolved === "required" && !tools?.length) {
+    throw new EZCoderAIError(
+      "OpenAI Codex cannot require a tool call when no tools are configured.",
+      {
+        source: "capability"
+      }
+    );
+  }
+  return resolved;
+}
 function streamOpenAICodex(options) {
   return new StreamResult(runStream3(options), options.signal);
 }
@@ -2180,7 +2332,7 @@ async function* runStream3(options) {
     stream: true,
     instructions: system,
     input,
-    tool_choice: "auto",
+    tool_choice: toCodexToolChoice(options.toolChoice, options.tools),
     parallel_tool_calls: !responsesLite,
     include: ["reasoning.encrypted_content"]
   };
@@ -2213,18 +2365,26 @@ async function* runStream3(options) {
     headers["chatgpt-account-id"] = options.accountId;
   }
   if (options.transportSessionId) {
-    headers["session_id"] = options.transportSessionId;
-    headers["x-client-request-id"] = options.transportSessionId;
-  }
+    const transportSessionId = normalizePromptCacheKey(options.transportSessionId);
+    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) {
     const text = await response.text().catch(() => "");
-    const parsed = parseCodexErrorBody(text);
+    const parsed = parseCodexErrorBody(text, response.status);
     const message = parsed.message ?? `Codex API returned HTTP ${response.status}.`;
     const requestId = parsed.requestId ?? readHeader(response.headers, "x-request-id", "openai-request-id", "x-oai-request-id");
     const usageLimit = codexUsageLimitError(parsed.errorObj, response.status, requestId);
@@ -2618,13 +2778,14 @@ function toCodexTools(tools) {
     strict: null
   }));
 }
-function parseCodexErrorBody(text) {
+function parseCodexErrorBody(text, statusCode) {
   if (!text) return {};
   try {
     const parsed = JSON.parse(text);
     const error = parsed.error;
     const detail = parsed.detail;
-    const message = error?.message ?? parsed.message ?? (typeof detail === "string" ? detail : void 0);
+    const rawMessage = error?.message ?? parsed.message ?? (typeof detail === "string" ? detail : void 0);
+    const message = rawMessage && isRawHtmlErrorEcho(rawMessage) ? providerHtmlErrorMessage(statusCode) : rawMessage;
     const requestId = parsed.request_id ?? error?.request_id ?? (message ? extractRequestIdFromMessage(message) : void 0);
     const errorObj = error ?? parsed;
     return {
@@ -2633,8 +2794,12 @@ function parseCodexErrorBody(text) {
       ...errorObj ? { errorObj } : {}
     };
   } catch {
-    const trimmed = text.trim().slice(0, 240);
-    return trimmed ? { message: trimmed } : {};
+    const trimmed = text.trim();
+    if (isRawHtmlErrorEcho(trimmed)) {
+      return { message: providerHtmlErrorMessage(statusCode) };
+    }
+    const bounded = trimmed.slice(0, 240);
+    return bounded ? { message: bounded } : {};
   }
 }
 var CODEX_USAGE_LIMIT_CODE = /usage_limit_reached|usage_not_included/i;
@@ -3253,6 +3418,18 @@ 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,
@@ -3275,7 +3452,12 @@ function stream(options) {
   if (options.supportsVideo !== true && messagesContainVideo(options.messages)) {
     throw new VideoUnsupportedError();
   }
-  return entry.stream(options);
+  const messages = clampProviderContextImages(
+    options.messages,
+    options.provider,
+    options.supportsImages
+  );
+  return entry.stream(messages === options.messages ? options : { ...options, messages });
 }
 function messagesContainVideo(messages) {
   for (const msg of messages) {
@@ -3380,6 +3562,125 @@ Original: ${message}`;
   return message;
 }
 
+// src/redaction.ts
+var REDACTED = "[REDACTED]";
+var TRUNCATED = "[TRUNCATED]";
+var CIRCULAR = "[CIRCULAR]";
+var SENSITIVE_NAME = /(?:^|[_-])(?:api[_-]?key|access[_-]?token|refresh[_-]?token|token|key|auth(?:orization)?|bearer|cookie|credential|private[_-]?key|password|passwd|secret)(?:$|[_-])/i;
+var SENSITIVE_ASSIGNMENT = /\b((?:api[_-]?key|access[_-]?token|refresh[_-]?token|token|key|auth(?:orization)?|bearer|cookie|credential|private[_-]?key|password|passwd|secret))\b(\s*[=:]\s*)(["']?)([^\s,"';}]+)\3/gi;
+function escaped(value) {
+  return value.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+}
+function normalizedSecrets(secrets) {
+  if (!secrets) return [];
+  return [...new Set([...secrets].filter((value) => value.length >= 8 && value !== REDACTED))].sort(
+    (a, b) => b.length - a.length
+  );
+}
+function environmentSecrets(env) {
+  const values = /* @__PURE__ */ new Set();
+  for (const [name, value] of Object.entries(env)) {
+    if (!value || value.length < 8 || value === REDACTED || !SENSITIVE_NAME.test(name)) continue;
+    values.add(value);
+  }
+  return [...values].sort((a, b) => b.length - a.length);
+}
+function redactText(text, options = {}) {
+  let result = text;
+  result = result.replace(
+    /-----BEGIN (?:[A-Z0-9 ]+ )?PRIVATE KEY-----[\s\S]*?-----END (?:[A-Z0-9 ]+ )?PRIVATE KEY-----/g,
+    REDACTED
+  );
+  result = result.replace(/\b([a-z][a-z0-9+.-]*:\/\/)[^\s/@:]+:[^\s/@]+@/gi, `$1${REDACTED}@`);
+  result = result.replace(
+    /\b(authorization\s*[:=]\s*)(?:bearer|basic)\s+[^\s,;]+/gi,
+    `$1${REDACTED}`
+  );
+  result = result.replace(/\b(bearer|basic)\s+[A-Za-z0-9+/_.=-]{8,}/gi, `$1 ${REDACTED}`);
+  result = result.replace(/\b(cookie|set-cookie)(\s*[:=]\s*)[^\r\n]+/gi, `$1$2${REDACTED}`);
+  result = result.replace(
+    /\beyJ[A-Za-z0-9_-]{6,}\.[A-Za-z0-9_-]{6,}\.[A-Za-z0-9_-]{6,}\b/g,
+    REDACTED
+  );
+  result = result.replace(
+    /\b(?:sk-(?:ant-|proj-)?|xox[baprs]-|gh[pousr]_|github_pat_|AIza)[A-Za-z0-9_-]{12,}\b/g,
+    REDACTED
+  );
+  result = result.replace(
+    SENSITIVE_ASSIGNMENT,
+    (_match, name, separator) => `${name}${separator}${REDACTED}`
+  );
+  for (const secret of normalizedSecrets(options.secrets)) {
+    result = result.replace(new RegExp(escaped(secret), "g"), REDACTED);
+  }
+  const maxStringLength = options.maxStringLength ?? 1e6;
+  if (result.length > maxStringLength) {
+    result = `${result.slice(0, maxStringLength)}${TRUNCATED}`;
+  }
+  return result;
+}
+function isBinary(value) {
+  return value instanceof ArrayBuffer || ArrayBuffer.isView(value) || typeof Blob !== "undefined" && value instanceof Blob;
+}
+function isMediaObject(value) {
+  return (value.type === "image" || value.type === "video") && (typeof value.data === "string" || typeof value.url === "string");
+}
+function redactValue(value, options = {}) {
+  const maxDepth = options.maxDepth ?? 20;
+  const maxEntries = options.maxEntries ?? 1e4;
+  const seen = /* @__PURE__ */ new WeakSet();
+  let entries = 0;
+  const visit = (current, depth, sensitive = false) => {
+    if (typeof current === "string") {
+      if (sensitive && current.length > 0 && current !== REDACTED) return REDACTED;
+      return redactText(current, options);
+    }
+    if (current === null || current === void 0 || typeof current === "number" || typeof current === "boolean" || typeof current === "bigint") {
+      return current;
+    }
+    if (typeof current !== "object") return current;
+    if (isBinary(current)) return current;
+    if (current instanceof Date) return new Date(current.getTime());
+    if (depth >= maxDepth) return TRUNCATED;
+    if (seen.has(current)) return CIRCULAR;
+    seen.add(current);
+    if (current instanceof Error) {
+      const error = {
+        name: current.name,
+        message: visit(current.message, depth + 1),
+        stack: visit(current.stack, depth + 1)
+      };
+      for (const [key, child] of Object.entries(current)) {
+        error[key] = visit(child, depth + 1, SENSITIVE_NAME.test(key));
+      }
+      return error;
+    }
+    if (Array.isArray(current)) {
+      const clone2 = [];
+      for (const child of current) {
+        if (++entries > maxEntries) {
+          clone2.push(TRUNCATED);
+          break;
+        }
+        clone2.push(visit(child, depth + 1));
+      }
+      return clone2;
+    }
+    const record = current;
+    if (isMediaObject(record)) return { ...record };
+    const clone = {};
+    for (const [key, child] of Object.entries(record)) {
+      if (++entries > maxEntries) {
+        clone[TRUNCATED] = true;
+        break;
+      }
+      clone[key] = visit(child, depth + 1, SENSITIVE_NAME.test(key));
+    }
+    return clone;
+  };
+  return visit(value, 0);
+}
+
 // src/providers/palsu.ts
 function palsuText(text) {
   return { role: "assistant", content: text ? [{ type: "text", text }] : [] };
@@ -3535,8 +3836,11 @@ export {
   EZCoderAIError,
   EventStream,
   ProviderError,
+  REDACTED as REDACTION_MARKER,
   StreamResult,
+  clampProviderContextImages,
   classifyProviderError,
+  environmentSecrets,
   formatError,
   formatErrorForDisplay,
   isHardBillingMessage,
@@ -3547,6 +3851,8 @@ export {
   palsuToolCall,
   prewarmAnthropicCache,
   providerRegistry,
+  redactText,
+  redactValue,
   registerPalsuProvider,
   setProviderDiagnostic,
   stream,