@prestyj/ai 5.4.0 → 5.5.0

package/dist/index.d.cts

@@ -3,7 +3,7 @@ import Anthropic from '@anthropic-ai/sdk';
 import OpenAI from 'openai';
 
 type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "sakana" | "palsu";
-type ThinkingLevel = "low" | "medium" | "high" | "xhigh" | "max";
+type ThinkingLevel = "low" | "medium" | "high" | "xhigh" | "max" | "ultra";
 type CacheRetention = "none" | "short" | "long";
 interface TextContent {
     type: "text";
@@ -175,6 +175,10 @@ interface StreamOptions {
     serviceTier?: "auto" | "default" | "flex" | "priority";
     /** OpenAI ChatGPT account ID (from OAuth JWT) for codex endpoint */
     accountId?: string;
+    /** Stable conversation identity for Codex transport headers. This is distinct from
+     *  promptCacheKey: sessions with matching prefixes may share a cache key, but must
+     *  retain independent session/thread identities. */
+    transportSessionId?: string;
     /** Google Cloud/Code Assist project ID used by Gemini OAuth transport. */
     projectId?: string;
     /** Enable provider-native web search. Each provider uses its own format:

package/dist/index.d.ts

@@ -3,7 +3,7 @@ import Anthropic from '@anthropic-ai/sdk';
 import OpenAI from 'openai';
 
 type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "sakana" | "palsu";
-type ThinkingLevel = "low" | "medium" | "high" | "xhigh" | "max";
+type ThinkingLevel = "low" | "medium" | "high" | "xhigh" | "max" | "ultra";
 type CacheRetention = "none" | "short" | "long";
 interface TextContent {
     type: "text";
@@ -175,6 +175,10 @@ interface StreamOptions {
     serviceTier?: "auto" | "default" | "flex" | "priority";
     /** OpenAI ChatGPT account ID (from OAuth JWT) for codex endpoint */
     accountId?: string;
+    /** Stable conversation identity for Codex transport headers. This is distinct from
+     *  promptCacheKey: sessions with matching prefixes may share a cache key, but must
+     *  retain independent session/thread identities. */
+    transportSessionId?: string;
     /** Google Cloud/Code Assist project ID used by Gemini OAuth transport. */
     projectId?: string;
     /** Enable provider-native web search. Each provider uses its own format:

package/dist/index.js

@@ -818,7 +818,7 @@ function toAnthropicThinking(level, maxTokens, model) {
     };
   }
   const VISIBLE_FLOOR = 1024;
-  const effectiveLevel = level === "xhigh" || level === "max" ? "high" : level;
+  const effectiveLevel = level === "xhigh" || level === "max" || level === "ultra" ? "high" : level;
   const budgetMap = {
     low: Math.max(1024, Math.floor(maxTokens * 0.2)),
     medium: Math.max(2048, Math.floor(maxTokens * 0.45)),
@@ -986,7 +986,7 @@ function toOpenAIToolChoice(choice) {
   return { type: "function", function: { name: choice.name } };
 }
 function toOpenAIReasoningEffort(level, model) {
-  const effort = level === "max" ? "xhigh" : level;
+  const effort = level === "max" || level === "ultra" ? "xhigh" : level;
   if (model.startsWith("fugu") && (effort === "low" || effort === "medium")) {
     return "high";
   }
@@ -1712,11 +1712,16 @@ function getEnvironment() {
 // src/providers/openai.ts
 function extractOpenAIUsage(usage) {
   let cacheRead = 0;
+  let cacheWrite = 0;
   const details = usage.prompt_tokens_details;
   if (details?.cached_tokens) {
     cacheRead = details.cached_tokens;
   }
   const usageAny = usage;
+  const detailsAny = details;
+  if (typeof detailsAny?.cache_write_tokens === "number") {
+    cacheWrite = detailsAny.cache_write_tokens;
+  }
   if (!cacheRead && typeof usageAny.cached_tokens === "number" && usageAny.cached_tokens > 0) {
     cacheRead = usageAny.cached_tokens;
   }
@@ -1724,9 +1729,10 @@ function extractOpenAIUsage(usage) {
     cacheRead = usageAny.prompt_cache_hit_tokens;
   }
   return {
-    inputTokens: usage.prompt_tokens - cacheRead,
+    inputTokens: usage.prompt_tokens - cacheRead - cacheWrite,
     outputTokens: usage.completion_tokens,
-    cacheRead
+    cacheRead,
+    cacheWrite
   };
 }
 var openaiClientCache = /* @__PURE__ */ new Map();
@@ -1791,8 +1797,9 @@ async function* runStream2(options) {
   if (options.provider === "openai" || options.provider === "moonshot") {
     const paramsAny = params;
     paramsAny.prompt_cache_key = normalizePromptCacheKey(options.promptCacheKey ?? "ezcoder");
-    const retention = options.cacheRetention ?? "short";
-    if (retention === "long") {
+    if (options.provider === "openai" && options.model.startsWith("gpt-5.6")) {
+      paramsAny.prompt_cache_options = { mode: "implicit", ttl: "30m" };
+    } else if ((options.cacheRetention ?? "short") === "long") {
       paramsAny.prompt_cache_retention = "24h";
     }
   }
@@ -1843,6 +1850,7 @@ async function* runStream2(options) {
   let inputTokens = 0;
   let outputTokens = 0;
   let cacheRead = 0;
+  let cacheWrite = 0;
   let finishReason = null;
   let receivedAnyChunk = false;
   try {
@@ -1850,7 +1858,7 @@ async function* runStream2(options) {
       receivedAnyChunk = true;
       const choice = chunk.choices?.[0];
       if (chunk.usage) {
-        ({ inputTokens, outputTokens, cacheRead } = extractOpenAIUsage(chunk.usage));
+        ({ inputTokens, outputTokens, cacheRead, cacheWrite } = extractOpenAIUsage(chunk.usage));
       }
       if (!choice) continue;
       if (choice.finish_reason) {
@@ -1930,7 +1938,12 @@ async function* runStream2(options) {
       content: contentParts.length > 0 ? contentParts : textAccum || ""
     },
     stopReason,
-    usage: { inputTokens, outputTokens, ...cacheRead > 0 && { cacheRead } }
+    usage: {
+      inputTokens,
+      outputTokens,
+      ...cacheRead > 0 && { cacheRead },
+      ...cacheWrite > 0 && { cacheWrite }
+    }
   };
   yield { type: "done", stopReason };
   return response;
@@ -2003,8 +2016,9 @@ function completionToResponse(completion) {
   let inputTokens = 0;
   let outputTokens = 0;
   let cacheRead = 0;
+  let cacheWrite = 0;
   if (completion.usage) {
-    ({ inputTokens, outputTokens, cacheRead } = extractOpenAIUsage(completion.usage));
+    ({ inputTokens, outputTokens, cacheRead, cacheWrite } = extractOpenAIUsage(completion.usage));
   }
   const stopReason = normalizeOpenAIStopReason(choice?.finish_reason ?? null);
   return {
@@ -2013,7 +2027,12 @@ function completionToResponse(completion) {
       content: contentParts.length > 0 ? contentParts : textAccum
     },
     stopReason,
-    usage: { inputTokens, outputTokens, ...cacheRead > 0 && { cacheRead } }
+    usage: {
+      inputTokens,
+      outputTokens,
+      ...cacheRead > 0 && { cacheRead },
+      ...cacheWrite > 0 && { cacheWrite }
+    }
   };
 }
 function classifyOpenAICompatLimit(args) {
@@ -2135,6 +2154,10 @@ function extractRequestIdFromMessage(message) {
 
 // src/providers/openai-codex.ts
 var DEFAULT_BASE_URL = "https://chatgpt.com/backend-api";
+var CODEX_CLIENT_VERSION = "0.144.1";
+function usesResponsesLite(model) {
+  return model.startsWith("gpt-5.6-");
+}
 function outputTextKey(itemId, contentIndex) {
   return `${itemId ?? ""}:${contentIndex ?? 0}`;
 }
@@ -2150,6 +2173,7 @@ async function* runStream3(options) {
   const downgradedImages = downgradeUnsupportedImages(options.messages, options.supportsImages);
   const downgraded = downgradeUnsupportedVideos(downgradedImages, options.supportsVideo);
   const { system, input } = toCodexInput(downgraded, { supportsImages: options.supportsImages });
+  const responsesLite = usesResponsesLite(options.model);
   const body = {
     model: options.model,
     store: false,
@@ -2157,7 +2181,7 @@ async function* runStream3(options) {
     instructions: system,
     input,
     tool_choice: "auto",
-    parallel_tool_calls: true,
+    parallel_tool_calls: !responsesLite,
     include: ["reasoning.encrypted_content"]
   };
   if (options.tools?.length) {
@@ -2168,24 +2192,29 @@ async function* runStream3(options) {
     body.temperature = options.temperature;
   }
   body.reasoning = {
-    effort: options.thinking ?? "none",
-    summary: "auto"
+    // `ultra` is a client orchestration preset, not a Codex API effort.
+    effort: options.thinking === "ultra" ? "max" : options.thinking ?? "none",
+    summary: "auto",
+    ...responsesLite ? { context: "all_turns" } : {}
   };
   const headers = {
     "Content-Type": "application/json",
     Accept: "text/event-stream",
     Authorization: `Bearer ${options.apiKey}`,
     "OpenAI-Beta": "responses=experimental",
-    originator: "ezcoder",
-    "User-Agent": `ezcoder (${os.platform()} ${os.release()}; ${os.arch()})`
+    originator: responsesLite ? "codex_cli_rs" : "ezcoder",
+    "User-Agent": responsesLite ? `codex_cli_rs/${CODEX_CLIENT_VERSION}` : `ezcoder (${os.platform()} ${os.release()}; ${os.arch()})`,
+    ...responsesLite ? {
+      version: CODEX_CLIENT_VERSION,
+      "X-OpenAI-Internal-Codex-Responses-Lite": "true"
+    } : {}
   };
   if (options.accountId) {
     headers["chatgpt-account-id"] = options.accountId;
   }
-  const cacheScopeId = body.prompt_cache_key;
-  if (cacheScopeId) {
-    headers["session_id"] = cacheScopeId;
-    headers["x-client-request-id"] = cacheScopeId;
+  if (options.transportSessionId) {
+    headers["session_id"] = options.transportSessionId;
+    headers["x-client-request-id"] = options.transportSessionId;
   }
   const response = await fetch(url, {
     method: "POST",
@@ -2229,6 +2258,7 @@ async function* runStream3(options) {
   let inputTokens = 0;
   let outputTokens = 0;
   let cacheRead = 0;
+  let cacheWrite = 0;
   const diagStart = Date.now();
   const diagSeen = /* @__PURE__ */ new Set();
   for await (const event of parseSSE(response.body)) {
@@ -2405,7 +2435,8 @@ async function* runStream3(options) {
       const usage = resp?.usage;
       if (usage) {
         cacheRead = usage.input_tokens_details?.cached_tokens ?? 0;
-        inputTokens = (usage.input_tokens ?? 0) - cacheRead;
+        cacheWrite = usage.input_tokens_details?.cache_write_tokens ?? 0;
+        inputTokens = (usage.input_tokens ?? 0) - cacheRead - cacheWrite;
         outputTokens = usage.output_tokens ?? 0;
       }
     }
@@ -2453,7 +2484,12 @@ async function* runStream3(options) {
       content: contentParts.length > 0 ? contentParts : textAccum || ""
     },
     stopReason,
-    usage: { inputTokens, outputTokens, ...cacheRead > 0 && { cacheRead } }
+    usage: {
+      inputTokens,
+      outputTokens,
+      ...cacheRead > 0 && { cacheRead },
+      ...cacheWrite > 0 && { cacheWrite }
+    }
   };
   yield { type: "done", stopReason };
   return streamResponse;
@@ -2816,6 +2852,7 @@ function toGemini3ThinkingLevel(level) {
     case "high":
     case "xhigh":
     case "max":
+    case "ultra":
       return "HIGH";
   }
 }
@@ -2828,6 +2865,7 @@ function toThinkingBudget(level) {
     case "high":
     case "xhigh":
     case "max":
+    case "ultra":
       return 8192;
   }
 }