@zqagent/core 0.3.10 → 0.3.12

package/dist/{TurnEngine-BPKhWuSI.d.cts → TurnEngine-BSciNywD.d.cts}

@@ -1,5 +1,6 @@
 import { T as TokenUsage, M as Message, a as ToolSchema, R as ReasoningEffort, S as StreamChunk, b as MessageAttachment, c as StoredAttachment } from './types-BYyXWWPW.cjs';
 import { a as ToolMetadata, T as ToolRegistry } from './ToolRegistry-DouDvn9E.cjs';
+import { ApiMode } from './internal/providers/diagnostics.cjs';
 import { ChatCompletionMessageParam } from 'openai/resources/chat/completions.js';
 import { RequestOptions } from 'openai/core.mjs';
 import { ModelCapability } from '@zqagent/shared/model-capabilities';
@@ -344,22 +345,40 @@ declare class OpenAIProvider implements ProviderClient {
     private _logDir?;
     /** 服务端点 baseURL（用于识别 DeepSeek 等特定供应商） */
     private _baseURL;
-    /** 模型 → API 模式缓存：某模型不支持 Responses 后，后续直走 Chat */
-    private _modeByModel;
+    /**
+     * 本实例的 API 模式。
+     *
+     * 显式配置 > 全局 `OPENAI_FORCE_API_MODE` > auto（先 Responses 再降级）。
+     * 早先只有全局环境变量一种表达方式，一个节点挂两家供应商时无法各自指定。
+     */
+    private _apiMode;
     constructor(opts: {
         apiKey?: string;
         baseURL?: string;
         /** LLM 请求日志根目录，按日期分目录，不传则不记录 */
         logDir?: string;
+        /**
+         * 单次请求超时（毫秒）。缺省/非法时回落到 `OPENAI_TIMEOUT_MS`，再回落 10 分钟。
+         *
+         * 必须能按供应商设置：推理模型的合理预算与普通模型差一个数量级，共用一个全局值
+         * 只能二选一（慢模型被掐断，或快模型失败时干等十分钟）。
+         */
+        timeoutMs?: number | null;
+        /**
+         * API 模式（auto / responses / chat）。缺省时读 `OPENAI_FORCE_API_MODE`，再缺省 auto。
+         */
+        apiMode?: ApiMode | null;
     });
     /**
      * 解析指定模型的 API 模式。
      *
-     * 优先级：环境变量 OPENAI_FORCE_API_MODE（responses|chat）强制指定
-     * > 运行时降级缓存（该模型曾因"不支持"失败而缓存为 chat）
+     * 优先级：本实例的显式配置（responses|chat）
+     * > 运行时降级缓存（该端点+模型曾因"不支持"失败而缓存为 chat）
      * > 默认 responses（Responses 优先）。
      */
     private _resolveMode;
+    /** 记录「该端点+模型不支持 Responses」，后续同进程内的请求直走 Chat */
+    private _markChatMode;
     /**
      * Responses 失败的分类（决定「能不能改走 Chat Completions」以及要不要缓存降级）。
      *

package/dist/{TurnEngine-DtGEuVLT.d.ts → TurnEngine-BU4R7q_0.d.ts}

@@ -1,5 +1,6 @@
 import { T as TokenUsage, M as Message, a as ToolSchema, R as ReasoningEffort, S as StreamChunk, b as MessageAttachment, c as StoredAttachment } from './types-BYyXWWPW.js';
 import { a as ToolMetadata, T as ToolRegistry } from './ToolRegistry-Cy0fQxrb.js';
+import { ApiMode } from './internal/providers/diagnostics.js';
 import { ChatCompletionMessageParam } from 'openai/resources/chat/completions.js';
 import { RequestOptions } from 'openai/core.mjs';
 import { ModelCapability } from '@zqagent/shared/model-capabilities';
@@ -344,22 +345,40 @@ declare class OpenAIProvider implements ProviderClient {
     private _logDir?;
     /** 服务端点 baseURL（用于识别 DeepSeek 等特定供应商） */
     private _baseURL;
-    /** 模型 → API 模式缓存：某模型不支持 Responses 后，后续直走 Chat */
-    private _modeByModel;
+    /**
+     * 本实例的 API 模式。
+     *
+     * 显式配置 > 全局 `OPENAI_FORCE_API_MODE` > auto（先 Responses 再降级）。
+     * 早先只有全局环境变量一种表达方式，一个节点挂两家供应商时无法各自指定。
+     */
+    private _apiMode;
     constructor(opts: {
         apiKey?: string;
         baseURL?: string;
         /** LLM 请求日志根目录，按日期分目录，不传则不记录 */
         logDir?: string;
+        /**
+         * 单次请求超时（毫秒）。缺省/非法时回落到 `OPENAI_TIMEOUT_MS`，再回落 10 分钟。
+         *
+         * 必须能按供应商设置：推理模型的合理预算与普通模型差一个数量级，共用一个全局值
+         * 只能二选一（慢模型被掐断，或快模型失败时干等十分钟）。
+         */
+        timeoutMs?: number | null;
+        /**
+         * API 模式（auto / responses / chat）。缺省时读 `OPENAI_FORCE_API_MODE`，再缺省 auto。
+         */
+        apiMode?: ApiMode | null;
     });
     /**
      * 解析指定模型的 API 模式。
      *
-     * 优先级：环境变量 OPENAI_FORCE_API_MODE（responses|chat）强制指定
-     * > 运行时降级缓存（该模型曾因"不支持"失败而缓存为 chat）
+     * 优先级：本实例的显式配置（responses|chat）
+     * > 运行时降级缓存（该端点+模型曾因"不支持"失败而缓存为 chat）
      * > 默认 responses（Responses 优先）。
      */
     private _resolveMode;
+    /** 记录「该端点+模型不支持 Responses」，后续同进程内的请求直走 Chat */
+    private _markChatMode;
     /**
      * Responses 失败的分类（决定「能不能改走 Chat Completions」以及要不要缓存降级）。
      *

package/dist/index.cjs

@@ -125,6 +125,8 @@ __export(src_exports, {
   registerTerminalTools: () => registerTerminalTools,
   registerTodoTool: () => registerTodoTool,
   registerWebTools: () => registerWebTools,
+  repairToolCallHistory: () => repairToolCallHistory,
+  repairToolCallHistoryDetailed: () => repairToolCallHistoryDetailed,
   resetFfmpegCache: () => resetFfmpegCache,
   resetSandboxCache: () => resetSandboxCache,
   resolveMedia: () => resolveMedia,
@@ -345,10 +347,22 @@ function authErrorHints(message, code = "") {
     "\u51ED\u636E\u65E0\u6548\uFF1A\u82E5\u7528\u7684\u662F\u300C\u5FD7\u524D\u5927\u6A21\u578B\uFF08\u5185\u7F6E\uFF09\u300D\uFF0C\u5728\u8282\u70B9\u8BBE\u7F6E\u7684\u6A21\u578B\u9875\u70B9\u300C\u5237\u65B0\u5185\u7F6E\u6A21\u578B\u300D\u91CD\u65B0\u9886\u53D6 Key\uFF08\u5728\u63A7\u5236\u53F0\u91CD\u7F6E\u8FC7 Key \u540E\u9700\u8981\u8FD9\u4E00\u6B65\uFF09\uFF1B\u82E5\u662F\u81EA\u586B\u7684\u4F9B\u5E94\u5546\uFF0C\u8BF7\u68C0\u67E5 Key \u662F\u5426\u6B63\u786E\u6216\u5DF2\u8FC7\u671F"
   ];
 }
+function toolCallHistoryHints(message, code = "") {
+  const text = `${message} ${code}`;
+  if (!/tool_calls|tool[_ ]?call[_ ]?id/i.test(text) || !/must be followed by tool messages|insufficient tool messages|no tool output found|tool messages responding/i.test(
+    text
+  )) {
+    return [];
+  }
+  return [
+    "\u8FD9\u4E0D\u662F\u6A21\u578B\u6216\u4F59\u989D\u95EE\u9898\uFF1A\u4F1A\u8BDD\u5386\u53F2\u91CC\u6709\u4E00\u6761\u300C\u5DE5\u5177\u8C03\u7528\u6CA1\u6709\u62FF\u5230\u7ED3\u679C\u300D\u7684\u8BB0\u5F55\uFF08\u901A\u5E38\u662F\u4E0A\u6B21\u5728\u5DE5\u5177\u6267\u884C\u4E2D\u6309\u4E86 Ctrl+C\uFF0C\u6216\u8FDB\u7A0B\u88AB\u4E2D\u65AD\uFF09",
+    "\u5F15\u64CE\u5DF2\u81EA\u52A8\u6E05\u7406\u5E76\u91CD\u8BD5\uFF1B\u5982\u679C\u4ECD\u62A5\u540C\u6837\u7684\u9519\u8BEF\uFF0C\u7528 /new \u5F00\u4E00\u4E2A\u65B0\u4F1A\u8BDD\u5373\u53EF\uFF08\u65E7\u4F1A\u8BDD\u7684\u5386\u53F2\u91CC\u8FD8\u7559\u7740\u8FD9\u6761\u8BB0\u5F55\uFF09"
+  ];
+}
 function paramErrorHints(message, code = "") {
   const text = `${message} ${code}`;
   const hints = [];
-  if (/not[_\s]*(activated|enabled|purchased)|arrearage|quota|subscription|insufficient|未开通|欠费|额度|余额不足/i.test(
+  if (/not[_\s]*(activated|enabled|purchased)|arrearage|quota|subscription|billing|未开通|欠费|额度|余额不足|insufficient[_\s]*(quota|balance|credit|funds)/i.test(
     text
   )) {
     hints.push(
@@ -372,9 +386,23 @@ function paramErrorHints(message, code = "") {
   }
   return hints;
 }
-function llmTimeoutMs() {
+var DEFAULT_LLM_TIMEOUT_MS = 6e5;
+function normLlmTimeoutMs(v) {
+  if (v === null || v === void 0 || v === "") return null;
+  const n = Number(v);
+  if (!Number.isFinite(n) || n <= 0) return null;
+  return Math.min(Math.round(n), 30 * 6e4);
+}
+function llmTimeoutMs(explicit) {
+  const fromArg = normLlmTimeoutMs(explicit);
+  if (fromArg) return fromArg;
   const n = Number(process.env.OPENAI_TIMEOUT_MS);
-  return Number.isFinite(n) && n > 0 ? n : 6e5;
+  return Number.isFinite(n) && n > 0 ? n : DEFAULT_LLM_TIMEOUT_MS;
+}
+function normApiMode(v) {
+  const s = typeof v === "string" ? v.trim().toLowerCase() : "";
+  if (s === "auto" || s === "responses" || s === "chat") return s;
+  return null;
 }
 function llmStreamIdleMs() {
   const n = Number(process.env.OPENAI_STREAM_IDLE_MS);
@@ -2220,6 +2248,76 @@ function applyToolResultBudget(tool, result, opts = {}) {
   return { content, truncated: true, savedPath };
 }
 
+// src/engine/history.ts
+function repairToolCallHistoryDetailed(messages) {
+  const out = [];
+  let droppedCalls = 0;
+  let droppedResults = 0;
+  let droppedMessages = 0;
+  let i = 0;
+  while (i < messages.length) {
+    const m = messages[i];
+    if (m.role === "tool") {
+      droppedResults += 1;
+      i += 1;
+      continue;
+    }
+    const calls = m.role === "assistant" ? m.tool_calls : void 0;
+    if (!calls?.length) {
+      if (m.role === "assistant" && m.tool_calls) {
+        const { tool_calls, ...rest } = m;
+        void tool_calls;
+        out.push(rest);
+      } else {
+        out.push(m);
+      }
+      i += 1;
+      continue;
+    }
+    const responders = /* @__PURE__ */ new Map();
+    const notices = [];
+    let j = i + 1;
+    while (j < messages.length) {
+      const t = messages[j];
+      if (t.role === "notice") {
+        notices.push(t);
+        j += 1;
+        continue;
+      }
+      if (t.role !== "tool") break;
+      if (t.tool_call_id && !responders.has(t.tool_call_id)) {
+        responders.set(t.tool_call_id, t);
+      }
+      j += 1;
+    }
+    const kept = calls.filter((tc) => tc.id && responders.has(tc.id));
+    droppedCalls += calls.length - kept.length;
+    droppedResults += responders.size - kept.length;
+    if (kept.length === 0) {
+      const { tool_calls, ...rest } = m;
+      void tool_calls;
+      if ((rest.content ?? "").trim() !== "") out.push(rest);
+      else droppedMessages += 1;
+      out.push(...notices);
+    } else {
+      out.push(kept.length === calls.length ? m : { ...m, tool_calls: kept });
+      for (const tc of kept) out.push(responders.get(tc.id));
+      out.push(...notices);
+    }
+    i = j;
+  }
+  const unchanged = out.length === messages.length && out.every((m, idx) => m === messages[idx]);
+  return {
+    messages: unchanged ? messages : out,
+    droppedCalls,
+    droppedResults,
+    droppedMessages
+  };
+}
+function repairToolCallHistory(messages) {
+  return repairToolCallHistoryDetailed(messages).messages;
+}
+
 // src/engine/TurnEngine.ts
 var crypto = __toESM(require("crypto"), 1);
 var fs3 = __toESM(require("fs"), 1);
@@ -2758,6 +2856,7 @@ ${notice}` : text,
     attachments = routed.attachments;
     this.taskId = taskId;
     this._turnStartTime = Date.now();
+    this.messages = repairToolCallHistory(this.messages);
     this.messages.push({
       role: "user",
       content: userInput,
@@ -3759,7 +3858,7 @@ ${notice}` : text,
   /** 过滤 notice + 剥离 display-only 字段，得到最终发送视图 */
   _finalizeOutbound(sourceMessages) {
     const _STRIP = /* @__PURE__ */ new Set(["source", "_display", "ts", "reasoning", "usage", "attachmentRefs"]);
-    return sourceMessages.filter((m) => m.role !== "notice").map((m) => {
+    const stripped = sourceMessages.filter((m) => m.role !== "notice").map((m) => {
       if (m.role === "tool") {
         return {
           role: "tool",
@@ -3768,15 +3867,16 @@ ${notice}` : text,
         };
       }
       const staleAttachments = Boolean(m.attachments?.length) && (m.ts ?? 0) < this._turnStartTime;
-      const stripped = {};
+      const plain = {};
       for (const [k, v] of Object.entries(m)) {
         if (staleAttachments && k === "attachments") continue;
         if (!_STRIP.has(k)) {
-          stripped[k] = v;
+          plain[k] = v;
         }
       }
-      return stripped;
+      return plain;
     });
+    return repairToolCallHistory(stripped);
   }
   /** 软阈值（token 数）：达到它只淘汰旧工具结果 */
   _softTriggerTokens() {
@@ -4529,6 +4629,11 @@ var ResponsesApiError = class extends Error {
   type;
   requestId;
 };
+var NO_AUTH_KEY = "not-needed";
+var degradedModes = /* @__PURE__ */ new Map();
+function modeKey(baseURL, model) {
+  return `${baseURL}::${model}`;
+}
 var OpenAIProvider = class {
   /** OpenAI SDK 客户端 */
   client;
@@ -4536,23 +4641,31 @@ var OpenAIProvider = class {
   _logDir;
   /** 服务端点 baseURL（用于识别 DeepSeek 等特定供应商） */
   _baseURL;
-  /** 模型 → API 模式缓存：某模型不支持 Responses 后，后续直走 Chat */
-  _modeByModel = /* @__PURE__ */ new Map();
+  /**
+   * 本实例的 API 模式。
+   *
+   * 显式配置 > 全局 `OPENAI_FORCE_API_MODE` > auto（先 Responses 再降级）。
+   * 早先只有全局环境变量一种表达方式，一个节点挂两家供应商时无法各自指定。
+   */
+  _apiMode;
   constructor(opts) {
+    const explicitKey = opts.apiKey !== void 0;
     const apiKey = cleanLlmValue(opts.apiKey ?? process.env.OPENAI_API_KEY);
     const baseURL = normalizeBaseURL(opts.baseURL ?? process.env.OPENAI_BASE_URL);
-    if (!apiKey) {
+    if (!apiKey && !explicitKey) {
       throw new Error(sanitizeEmoji("\u274C \u672A\u8BBE\u7F6E OPENAI_API_KEY \u73AF\u5883\u53D8\u91CF"));
     }
+    const effectiveKey = apiKey ?? NO_AUTH_KEY;
     this._baseURL = baseURL ?? "";
+    this._apiMode = normApiMode(opts.apiMode) ?? normApiMode(process.env.OPENAI_FORCE_API_MODE) ?? "auto";
     const agent = socksAgent();
     const proxyHint = unsupportedProxyHint();
     if (proxyHint) console.warn(`[llm] ${proxyHint}`);
     const baseFetch = agent ? makeSocksFetch(agent) : ((input, init) => globalThis.fetch(input, init));
     this.client = new import_openai.default({
-      apiKey,
+      apiKey: effectiveKey,
       ...baseURL ? { baseURL } : {},
-      timeout: llmTimeoutMs(),
+      timeout: llmTimeoutMs(opts.timeoutMs),
       maxRetries: llmMaxRetries(),
       // 类型：SDK 期望 (RequestInfo, RequestInit) => Promise<Response>，
       // node-fetch 的 Response 与 DOM Response 结构不同，此处显式放宽
@@ -4569,14 +4682,17 @@ var OpenAIProvider = class {
   /**
    * 解析指定模型的 API 模式。
    *
-   * 优先级：环境变量 OPENAI_FORCE_API_MODE（responses|chat）强制指定
-   * > 运行时降级缓存（该模型曾因"不支持"失败而缓存为 chat）
+   * 优先级：本实例的显式配置（responses|chat）
+   * > 运行时降级缓存（该端点+模型曾因"不支持"失败而缓存为 chat）
    * > 默认 responses（Responses 优先）。
    */
   _resolveMode(model) {
-    const forced = process.env.OPENAI_FORCE_API_MODE?.toLowerCase();
-    if (forced === "responses" || forced === "chat") return forced;
-    return this._modeByModel.get(model) ?? "responses";
+    if (this._apiMode !== "auto") return this._apiMode;
+    return degradedModes.get(modeKey(this._baseURL, model)) ?? "responses";
+  }
+  /** 记录「该端点+模型不支持 Responses」，后续同进程内的请求直走 Chat */
+  _markChatMode(model) {
+    degradedModes.set(modeKey(this._baseURL, model), "chat");
   }
   /**
    * Responses 失败的分类（决定「能不能改走 Chat Completions」以及要不要缓存降级）。
@@ -4652,7 +4768,7 @@ var OpenAIProvider = class {
         } catch (err) {
           const kind = this._responsesFailureKind(err, true);
           if (kind) {
-            if (kind === "endpoint") this._modeByModel.set(model, "chat");
+            if (kind === "endpoint") this._markChatMode(model);
             console.warn(
               `[llm] ${model} Responses API \u4E0D\u53EF\u7528(${kind})\uFF0C\u6539\u7528 Chat Completions (${String(
                 err?.message ?? err
@@ -4726,7 +4842,7 @@ var OpenAIProvider = class {
       } catch (err) {
         const kind = this._responsesFailureKind(err, true);
         if (kind) {
-          if (kind === "endpoint") this._modeByModel.set(model, "chat");
+          if (kind === "endpoint") this._markChatMode(model);
           console.warn(
             `[llm] ${model} Responses API \u4E0D\u53EF\u7528(${kind})\uFF0C\u6539\u7528 Chat Completions (${String(
               err?.message ?? err
@@ -4772,7 +4888,7 @@ var OpenAIProvider = class {
     } catch (err) {
       const kind = emitted ? null : this._responsesFailureKind(err, false);
       if (!kind) throw enrichLlmError(err, this._baseURL);
-      if (kind === "endpoint") this._modeByModel.set(ctx.model, "chat");
+      if (kind === "endpoint") this._markChatMode(ctx.model);
       console.warn(
         `[llm] ${ctx.model} Responses \u6D41\u5185\u5931\u8D25(${kind})\uFF0C\u6539\u7528 Chat Completions (${String(
           err?.message ?? err
@@ -7296,6 +7412,9 @@ var ConsoleHook = class {
     for (const hint of authErrorHints(error, errorType)) {
       cliLog(`   \u{1F4A1} ${hint}`);
     }
+    for (const hint of toolCallHistoryHints(error, errorType)) {
+      cliLog(`   \u{1F4A1} ${hint}`);
+    }
     for (const hint of paramErrorHints(error)) {
       cliLog(`   \u{1F4A1} ${hint}`);
     }
@@ -13298,6 +13417,8 @@ function createDefaultEngine(opts) {
     mode = "interactive" /* INTERACTIVE */,
     maxIterations = DEFAULT_MAX_ITERATIONS,
     reasoningEffort,
+    timeoutMs,
+    apiMode,
     allowedCommands = [],
     shellPolicy = "approve",
     sandbox,
@@ -13318,7 +13439,7 @@ function createDefaultEngine(opts) {
   } = opts;
   const envEffort = process.env.OPENAI_REASONING_EFFORT?.toLowerCase();
   const effectiveEffort = reasoningEffort ?? (envEffort === "low" || envEffort === "medium" || envEffort === "high" ? envEffort : void 0);
-  const effectiveProvider = provider ?? new OpenAIProvider({ logDir });
+  const effectiveProvider = provider ?? new OpenAIProvider({ logDir, timeoutMs, apiMode });
   const sandboxPlan = resolveSandbox({ ...sandbox, workspace });
   const registry = new ToolRegistry();
   const conflictGuard = new ConflictGuard();
@@ -13489,6 +13610,8 @@ function createDefaultEngine(opts) {
   registerTerminalTools,
   registerTodoTool,
   registerWebTools,
+  repairToolCallHistory,
+  repairToolCallHistoryDetailed,
   resetFfmpegCache,
   resetSandboxCache,
   resolveMedia,