jinzd-ai-cli 0.4.319 → 0.4.321

package/dist/{chunk-SZBSTSG4.js → chunk-ZDO3XCXL.js}

@@ -2,7 +2,7 @@
 import {
   evaluateCiGate,
   normalizeCiThresholds
-} from "./chunk-OQZIP4LU.js";
+} from "./chunk-LEA4AMXE.js";
 import {
   buildReviewPrompt,
   buildSecurityReviewPrompt,
@@ -13,13 +13,13 @@ import {
 } from "./chunk-6YZEZFX5.js";
 import {
   ProviderRegistry
-} from "./chunk-N2MHUA4B.js";
+} from "./chunk-XVB4RARX.js";
 import {
   ConfigManager
-} from "./chunk-KEOMNFBH.js";
+} from "./chunk-COHKPZ2E.js";
 import {
   VERSION
-} from "./chunk-A7YZLLWN.js";
+} from "./chunk-2R6Y5SJH.js";
 
 // src/cli/ci.ts
 import { execFileSync } from "child_process";

package/dist/{ci-HAPWBC37.js → ci-EZZD7CMS.js}

@@ -3,16 +3,16 @@ import {
   CI_COMMENT_MARKER,
   countSeverity,
   runCi
-} from "./chunk-SZBSTSG4.js";
-import "./chunk-OQZIP4LU.js";
+} from "./chunk-ZDO3XCXL.js";
+import "./chunk-LEA4AMXE.js";
 import "./chunk-HLWUDRBO.js";
 import "./chunk-6YZEZFX5.js";
-import "./chunk-N2MHUA4B.js";
-import "./chunk-KEOMNFBH.js";
-import "./chunk-GVAAGBHX.js";
+import "./chunk-XVB4RARX.js";
+import "./chunk-COHKPZ2E.js";
+import "./chunk-E2W2MI6H.js";
 import "./chunk-TU3L3PW7.js";
 import "./chunk-BMLQPSHP.js";
-import "./chunk-A7YZLLWN.js";
+import "./chunk-2R6Y5SJH.js";
 export {
   CI_COMMENT_MARKER,
   countSeverity,

package/dist/{ci-format-NGWP4GTQ.js → ci-format-ZWECV6XI.js}

@@ -6,8 +6,8 @@ import {
   formatCiResult,
   formatCiSarif,
   normalizeCiThresholds
-} from "./chunk-OQZIP4LU.js";
-import "./chunk-A7YZLLWN.js";
+} from "./chunk-LEA4AMXE.js";
+import "./chunk-2R6Y5SJH.js";
 export {
   DEFAULT_CI_THRESHOLDS,
   detectCiGateSignals,

package/dist/{constants-GSUK6ILC.js → constants-F2ULZ6W2.js}

@@ -39,7 +39,7 @@ import {
   TEST_TIMEOUT,
   VERSION,
   buildUserIdentityPrompt
-} from "./chunk-A7YZLLWN.js";
+} from "./chunk-2R6Y5SJH.js";
 export {
   AGENTIC_BEHAVIOR_GUIDELINE,
   APP_NAME,

package/dist/{constants-CPHZLOCQ.js → constants-HF5AYLBR.js}

@@ -38,7 +38,7 @@ import {
   TEST_TIMEOUT,
   VERSION,
   buildUserIdentityPrompt
-} from "./chunk-DCQDFQDW.js";
+} from "./chunk-VEO2MILP.js";
 export {
   AGENTIC_BEHAVIOR_GUIDELINE,
   APP_NAME,

package/dist/{doctor-cli-L6AK2GJ7.js → doctor-cli-QAKYBWLG.js}

@@ -2,16 +2,16 @@
 import {
   formatDoctorReport,
   runDoctorCli
-} from "./chunk-D6BLSFYD.js";
-import "./chunk-662ZJPJE.js";
+} from "./chunk-2GOBINY4.js";
+import "./chunk-VS43QMR2.js";
 import "./chunk-RS7SKYTS.js";
-import "./chunk-N2MHUA4B.js";
-import "./chunk-KEOMNFBH.js";
-import "./chunk-GVAAGBHX.js";
+import "./chunk-XVB4RARX.js";
+import "./chunk-COHKPZ2E.js";
+import "./chunk-E2W2MI6H.js";
 import "./chunk-TU3L3PW7.js";
 import "./chunk-HOSJZMQS.js";
 import "./chunk-BMLQPSHP.js";
-import "./chunk-A7YZLLWN.js";
+import "./chunk-2R6Y5SJH.js";
 export {
   formatDoctorReport,
   runDoctorCli

package/dist/electron-server.js

@@ -55,7 +55,7 @@ import {
   touchMemoryReferences,
   updateMemoryApproval,
   updateMemoryEntry
-} from "./chunk-GP24X3KJ.js";
+} from "./chunk-7QW3DFYV.js";
 import {
   DEFAULT_PATTERNS,
   redactJson,
@@ -81,7 +81,7 @@ import {
   recoverGbkMojibake,
   runTestsTool,
   shellInvocation
-} from "./chunk-AC44E4CV.js";
+} from "./chunk-JHSULP5D.js";
 import {
   getCurrentSessionKey,
   getExecutionConfigDir,
@@ -123,7 +123,7 @@ import {
   SUBAGENT_MAX_ROUNDS_LIMIT,
   VERSION,
   buildUserIdentityPrompt
-} from "./chunk-DCQDFQDW.js";
+} from "./chunk-VEO2MILP.js";
 
 // src/web/access.ts
 import { randomBytes, timingSafeEqual } from "crypto";
@@ -981,17 +981,21 @@ var PRICING_TABLE = {
   "glm-4.6": { input: 0.6, output: 2.2 },
   "glm-4.6v": { input: 0.6, output: 2.2 },
   "glm-5": { input: 0.85, output: 2.85 },
-  // GLM-5.3（2026-08-14 发布）官方公告口径「与 GLM-5.2 同价」，直连 API $1.40 / $4.40。
+  // GLM-5.3：官方定价页（open.bigmodel.cn/pricing，2026-10-08 浏览器读取）——输入 8 元、
+  // 输出 28 元、缓存命中 2 元（每百万 token），按 ¥6.9/$1 换算。此前记的 $1.40/$4.40 是
+  // Z.ai 国际站的美元价，而本 provider 打的是国内站、按人民币计费，高报约 20%；更大的
+  // 偏差在**缓存**：没有 cacheRead 时缓存命中按输入全价算，2026-10-07 审计会话 1474 万
+  // 缓存 token 被记成 $21.34，按官方缓存价约 $4.90（2026-10-07 审计 X-03）。
   // 注意它**始终思考**，思考 token 按输出价计费 —— 实际成本高度取决于 reasoning_effort。
-  "glm-5.3": { input: 1.4, output: 4.4 },
+  "glm-5.3": { input: 8 / CNY_PER_USD, output: 28 / CNY_PER_USD, cacheRead: 2 / CNY_PER_USD },
   // GLM-5.3-Flash：官方定价页（open.bigmodel.cn/pricing）一手数字——输入 0.8 元、
   // 输出 2.8 元、缓存命中 0.23 元（每百万 token），按记账汇率 ¥6.9/$1 换算。
   // **刻意记恢复价而非当前的 5 折限时价**（0.4/1.4/0.115 元）：官方只说「限时两周」、
   // 没公布截止日期，静态表表达不了这个时间维度；高估比低估安全——预算告警早触发
   // 好过晚触发。折扣期内 `/cost` 会比实际偏高约一倍，这是刻意的取舍。
   "glm-5.3-flash": { input: 0.116, output: 0.406, cacheRead: 0.033 },
-  // GLM-5.2（2026-06）官方直连 API 定价 $1.40 / $4.40。
-  "glm-5.2": { input: 1.4, output: 4.4 },
+  // GLM-5.2：与 GLM-5.3 同价（同一定价页：8 / 28 / 缓存命中 2 元），口径修正同上。
+  "glm-5.2": { input: 8 / CNY_PER_USD, output: 28 / CNY_PER_USD, cacheRead: 2 / CNY_PER_USD },
   // GLM-5V-Turbo（2026-04 视觉编程旗舰）官方直连 API 定价 $1.20 / $4.00。
   "glm-5v-turbo": { input: 1.2, output: 4 },
   "glm-5.1": { input: 0.95, output: 3.15 },
@@ -2433,6 +2437,14 @@ var HALLUCINATION_PATTERNS = [
   // 已创建！
   /File\s+(?:written|saved|created)\s+(?:to|as|at)/i,
   // File written to / saved as（要求介词）
+  // 2026-10-07 审计 B-02：中英文判据不对称。中文「已保存到」要求引号、英文只认 `File saved to`，
+  // 于是「已保存到 report.md。」「I saved the changes to src/index.ts.」双双漏检。
+  // 改为「后面跟一个带扩展名的文件名」——同样挡得住「已保存到数据库」「saved my time」这类误报。
+  /已(?:保存|写入)到?\s*[^\s，,。；;：:]{1,120}\.\w{1,6}(?![\w/])/,
+  // 动词后直接跟文件名（中间只许冠词 / both / the file 之类），或「动词 … to/into 文件名」。
+  // 刻意不放宽成「同句任意文件名」：「I saved you some time by checking README.md」不是写入声明。
+  /\b(?:I|We)(?:'ve| have)?\s+(?:now\s+|just\s+|also\s+|successfully\s+)?(?:created|written|wrote|saved|generated)\s+(?:(?:the|a|an|both|all|two|three|new|following)\s+)*(?:(?:files?|documents?|reports?|scripts?)\s+)?[`'"]?[\w./-]+\.(?:md|txt|json|ya?ml|csv|html?|css|[jt]sx?|py|sh|go|rs|java|sql|xml|toml|ini|cfg|pdf|docx?|xlsx?|pptx?)\b/i,
+  /\b(?:I|We)(?:'ve| have)?\s+(?:now\s+|just\s+|also\s+|successfully\s+)?(?:created|written|wrote|saved|generated)\b[^.\n]{0,40}?\b(?:to|into)\s+[`'"]?[\w./-]+\.(?:md|txt|json|ya?ml|csv|html?|css|[jt]sx?|py|sh|go|rs|java|sql|xml|toml|ini|cfg|pdf|docx?|xlsx?|pptx?)\b/i,
   /生成完成[！!]/,
   // 生成完成！
   /✅\s*(?:文件|已[生保写创]|第)\S*\.\w{1,5}/,
@@ -2483,6 +2495,8 @@ var QUICK_WRITE_KEYWORDS = [
   "saved",
   "written",
   "created",
+  "wrote",
+  "generated",
   "File",
   "file",
   // emoji 勾选标记
@@ -2599,8 +2613,13 @@ function extractClaimedFilePaths(content) {
   }
   const zhRe = /(?:已保存(?:到)?|已写入(?:到)?|已创建|已生成|文件路径[：:]|保存为|写入到)\s*[`'”””]?([^\s`'”””，,。\n]+\.\w{1,6})/g;
   while ((m = zhRe.exec(content)) !== null) add(m[1]);
-  const enRe = /(?:saved|written|created)\s+(?:to|as|at)\s+[`'”]?([^\s`'”\n,]+\.\w{1,6})/gi;
+  const enRe = /\b(?:saved|written|wrote|created)\b[^.\n]{0,40}?\b(?:to|as|at|into)\s+[`'”]?([^\s`'”\n,]+\.\w{1,6})/gi;
   while ((m = enRe.exec(content)) !== null) add(m[1]);
+  const enClaimRe = /\b(?:I|We)(?:'ve| have)?\s+(?:now\s+|just\s+|also\s+|successfully\s+)?(?:created|written|wrote|saved|generated)\b([^\n;]*)/gi;
+  while ((m = enClaimRe.exec(content)) !== null) {
+    const clause = m[1].split(/[.!?](?:\s|$)|\b(?:based on|from|using|per|according to|referencing|following|after|by|while|when|checking|reading|reviewing)\b/i)[0] ?? "";
+    for (const f of clause.matchAll(/[`'"]?([\w./-]+\.\w{1,6})\b/g)) add(f[1]);
+  }
   const checkRe = /✅[^\n`]*?[`'”]?([^\s`'”\n,，。]+?\.\w{1,6})/g;
   while ((m = checkRe.exec(content)) !== null) {
     const pathStart = m.index + m[0].lastIndexOf(m[1]);
@@ -2875,8 +2894,27 @@ var META_NARRATION_HEURISTICS = [
   /(?:我来|让我|容我)\s*(?:整理|梳理|归纳|想想|看看|重新|检查|核实)/u,
   // 过渡语（需要 2+ hits）："以下是：" / "如下：" 作为独立行。文档中可作小节标题，
   // 但配合另一个对话模式（如"好的"）时是典型的 AI 内容包装。
-  /^.{0,60}(?:以下是|下面是|如下)\s*[：:，,]\s*$/mu
+  /^.{0,60}(?:以下是|下面是|如下)\s*[：:，,]\s*$/mu,
+  // 叙述自己要调用工具（2026-10-07 审计 B-05）：文档正文不会这样写自己。
+  /(?:让我|我来|我将|我会|接下来)\s*(?:调用|使用|用)\s*`?\w+`?\s*工具/u,
+  /\bI(?:'ll| will| am going to)\s+(?:now\s+)?(?:use|call|invoke)\s+the\s+`?\w+`?\s+tool\b/i
+];
+var ANNOUNCEMENT_PATTERNS = [
+  /^\s*(?:好的|好|收到|明白|了解|嗯|OK|Okay|Sure|Certainly)[，,。.!！\s]/iu,
+  /(?:我(?:这就|马上|现在|立即|立刻|来|将|会)|让我|下面我|接下来我)\s*(?:为[您你])?\s*(?:开始|着手|撰写|写|生成|整理|输出|编写|准备|调用|使用)/u,
+  /(?:请稍[候等]|稍等|马上就好|这就开始)/u,
+  /\b(?:let me|I(?:'ll| will| am going to)|I'm going to)\s+(?:now\s+)?(?:write|draft|generate|create|start|prepare|produce|compose|use|call)\b/i,
+  /\b(?:please wait|one moment|stand by)\b/i
 ];
+function looksLikeAnnouncementOnly(content) {
+  const text = content.trim();
+  if (!text || text.length > 300 || text.split("\n").length > 4) return null;
+  if (/^#{1,6}\s+\S/m.test(text)) return null;
+  for (const re of ANNOUNCEMENT_PATTERNS) {
+    if (re.test(text)) return re.source;
+  }
+  return null;
+}
 var CHINESE_TAIL_CLOSING_MARKERS = [
   /(?:需要你|请你|麻烦你|你看)\s*(?:确认|决定|选择|帮忙|看[一看下]|过目|觉得|认为)/u,
   /(?:随时|有问题|有需要|需要调整|需要修改)\s*(?:联系|找我|告诉|沟通|调整|修改)/u,
@@ -3013,6 +3051,15 @@ This fresh stream has NO tools \u2014 output is captured verbatim. STOP emitting
       summary: `[save_last_response REJECTED] The content-only stream produced NO text at all (output budget likely spent on reasoning). ${saveToFile} was NOT saved.`
     };
   }
+  const announced = looksLikeAnnouncementOnly(rawContent);
+  if (announced) {
+    return fallback(`announcement only: ${announced}`) ?? {
+      kind: "reject",
+      reason: "not-a-document",
+      matched: announced,
+      summary: `[save_last_response REJECTED] The content-only stream produced only an announcement (e.g. "OK, I'll write the report now\u2026"), not the document itself. ${saveToFile} was NOT saved.`
+    };
+  }
   const { lines, bytes } = teeFileStats(rawContent);
   return {
     kind: "ok",
@@ -3034,6 +3081,7 @@ function teeStreamErrorSummary(saveToFile, errMsg) {
 // src/core/stalled-intent.ts
 var WAITING_FOR_USER_RE = /[？?]\s*$|是否|要不要|需不需要|需要我|要我|可以吗|好吗|行吗|您看|你看看|请确认|确认后|同意后|如果(?:你|您)|若(?:你|您)|如(?:你|您)需要|等你|等待(?:你|您)|待(?:你|您)|我可以|我也可以|还可以|(?:你|您)可以|告诉我|或者(?:让我|我)|可以为(?:你|您)|可以帮|随时|\bwould you\b|\bshall i\b|\bshould i\b|\bdo you want\b|\blet me know\b|\bif you(?:'d| would)? like\b|\bonce you confirm\b|\byour (?:approval|confirmation)\b|\bif you want\b|\bi can also\b/i;
 var COMPLETED_RE = /已|完成|好了|完毕|写好|搞定|\b(?:done|finished|completed|have (?:written|created|generated))\b/i;
+var NEGATION_RE = /没有?|未|无法|不(?:会|能|再|打算|需要|用|必|要)|n't\b|\b(?:not|never|no longer|cannot)\b/i;
 var CN_VERBS = "\u5F00\u5199|\u52A8\u624B|\u7F16\u5199|\u5199\u5165|\u5199|\u751F\u6210|\u521B\u5EFA|\u5236\u4F5C|\u6267\u884C|\u8FD0\u884C|\u5B89\u88C5|\u4FEE\u6539|\u66F4\u65B0|\u4FEE\u590D|\u5B9E\u73B0|\u6784\u5EFA|\u8F93\u51FA|\u4FDD\u5B58|\u8BFB\u53D6|\u67E5\u770B|\u68C0\u67E5|\u8C03\u7528|\u5904\u7406|\u6574\u7406|\u8865\u5145|\u5B8C\u5584|\u91CD\u5199|\u6539\u5199|\u843D\u76D8";
 var CN_FIRST_PERSON_INTENT_RE = new RegExp(
   `(?:\u6211|\u54B1\u4EEC|\u6211\u4EEC|\u8BA9\u6211)(?:\u73B0\u5728|\u8FD9\u5C31|\u9A6C\u4E0A|\u7ACB\u5373|\u7ACB\u523B|\u76F4\u63A5|\u63A5\u4E0B\u6765|\u4E0B\u9762|\u5148|\u518D|\u968F\u540E|\u7136\u540E|\u5C31)*(?:\u5C06|\u4F1A|\u8981|\u6765|\u53BB|\u5F00\u59CB|\u7740\u624B|\u51C6\u5907)?(?:[^\uFF0C,\u3002\uFF01!\uFF1B;\uFF1A:\\n]{0,4})?(?:${CN_VERBS})`
@@ -3055,11 +3103,11 @@ function detectStalledIntent(content) {
   if (WAITING_FOR_USER_RE.test(tail)) return null;
   const clauses = tail.split(/[，,；;。！!：:\n]/).map((c) => c.trim()).filter(Boolean);
   for (const clause of clauses) {
-    if (COMPLETED_RE.test(clause)) continue;
+    if (COMPLETED_RE.test(clause) || NEGATION_RE.test(clause)) continue;
     if (CN_FIRST_PERSON_INTENT_RE.test(clause) || CN_BARE_INTENT_RE.test(clause)) return clause;
   }
   for (const sentence of tail.split(/(?<=[.!])\s+|\n/)) {
-    if (COMPLETED_RE.test(sentence)) continue;
+    if (COMPLETED_RE.test(sentence) || NEGATION_RE.test(sentence)) continue;
     const m = EN_INTENT_RE.exec(sentence);
     if (m) return sentence.trim();
   }
@@ -3507,7 +3555,12 @@ async function runAgentLoop(host) {
       if (tc.name.startsWith("mcp__")) host.onMcpToolUsed?.(tc.name);
     }
     host.persistRound?.(toolCalls, toolResults, {
-      assistantContent: outcome.content,
+      // 工具调用前模型说的话（"A 方向汇报：…"、"先看一下配置"）挂在 _streamedText 上：
+      // 本轮上下文里有（buildToolResultMessages 会带上），但此前存档只取 outcome.content，
+      // 而工具调用轮次的 outcome 没有 content——2026-10-07 审计会话 157 条带工具调用的
+      // assistant 消息正文**全部为空**：下一条用户消息时模型看不到自己之前的阶段性汇报，
+      // /session show、导出、回放也全缺这部分（审计 X-01）。
+      assistantContent: outcome.content ?? toolCalls._streamedText,
       reasoningContent: outcome.reasoningContent
     });
     host.afterToolRoundPersist?.(toolCalls, toolResults, extraMessages);
@@ -3562,6 +3615,7 @@ async function runLeanAgentLoop(opts) {
   let roundsUsed = 0;
   let toolCallCount = 0;
   let stalledIntentNudges = 0;
+  let hallucinationRetries = 0;
   try {
     for (let round = 0; round < opts.maxRounds; round++) {
       roundsUsed = round + 1;
@@ -3577,6 +3631,7 @@ async function runLeanAgentLoop(opts) {
           timeout: opts.chatParams?.timeout,
           thinking: opts.chatParams?.thinking,
           thinkingEffort: opts.chatParams?.thinkingEffort,
+          thinkingBudget: opts.chatParams?.thinkingBudget,
           ...extraMessages.length > 0 ? { _extraMessages: extraMessages } : {}
         },
         opts.toolDefsForRound(round)
@@ -3596,6 +3651,24 @@ async function runLeanAgentLoop(opts) {
           );
           continue;
         }
+        const nextTools = round < opts.maxRounds - 1 ? opts.toolDefsForRound(round + 1) : [];
+        const canWrite = nextTools.some((t) => t.name === "write_file" || t.name === "edit_file");
+        if (canWrite && hallucinationRetries < 1 && result.content) {
+          const alreadyWrote = hadPreviousWriteToolCalls(extraMessages);
+          const coarse = !alreadyWrote && detectsHallucinatedFileOp(result.content) && extractBashCommands(extraMessages).length === 0;
+          const phantom = coarse || alreadyWrote ? findPhantomClaims(result.content, extraMessages) : [];
+          if (phantom.length > 0 || coarse) {
+            hallucinationRetries++;
+            extraMessages.push(
+              { role: "assistant", content: result.content },
+              {
+                role: "user",
+                content: phantom.length > 0 ? buildPhantomCorrectionMessage(phantom) : HALLUCINATION_CORRECTION_MESSAGE
+              }
+            );
+            continue;
+          }
+        }
         content = result.content;
         break;
       }
@@ -4704,6 +4777,7 @@ function ollamaNativeBaseUrl(openAiBaseUrl) {
   return openAiBaseUrl.replace(/\/+$/, "").replace(/\/v1$/, "");
 }
 var NATIVE_API_TIMEOUT_MS = 3e3;
+var SMALL_CONTEXT_WARN = 16384;
 var OllamaProvider = class extends OpenAICompatibleProvider {
   // 用 127.0.0.1 而非 localhost：Node.js 18+ 在 Windows 上优先解析 localhost 为 IPv6 (::1)，
   // 但 Ollama 默认只监听 IPv4，导致连接失败。
@@ -4782,12 +4856,13 @@ var OllamaProvider = class extends OpenAICompatibleProvider {
       if (show?.capabilities?.includes("thinking")) thinkingModels.add(id);
       if (running.size === 0 && !show) return;
       const digest = digests.get(id);
-      const { contextWindow } = resolveOllamaContextWindow({
+      const { contextWindow, source: source2 } = resolveOllamaContextWindow({
         runningContext: digest && running.get(digest) || running.get(id),
         show,
         fallback: this.estimateContextWindow(id)
       });
       result.set(id, contextWindow);
+      this.contextSources.set(id, source2);
     }));
     return { contextWindows: result, thinkingModels };
   }
@@ -4883,6 +4958,77 @@ var OllamaProvider = class extends OpenAICompatibleProvider {
     const target = norm2(model);
     return this.dynamicModels.some((m) => m.supportsThinking === true && norm2(m.id) === target);
   }
+  // ── 加载后补读真实窗口（2026-10-07 审计 C-01）────────────────────────────
+  //
+  // 上下文窗口只在启动时读一次。模型当时若还没加载，`/api/ps` 里没有它，只能取训练上限
+  // （Qwen3.8 是 262K）；而服务端实际开的窗口可能只有它的几十分之一。第一次请求把模型
+  // 加载进来之后若不再读，上下文压力监控、自动压缩与工具预算就整场按那个大数在算——
+  // aicli 以为还有余量，Ollama 早已在悄悄截掉最早的消息（v0.4.311 那次说的正是这个后果）。
+  // 所以每个模型在首次请求之后补读一次 `/api/ps`；读到的窗口若小得连工具定义都放不下，
+  // 提示一次怎么调大。都是后台做，不拖慢这次请求。
+  /** 每个模型的窗口是从哪一级读到的（只有 'running' 是服务端实际值） */
+  contextSources = /* @__PURE__ */ new Map();
+  refreshedAfterLoad = /* @__PURE__ */ new Set();
+  warnedSmallContext = /* @__PURE__ */ new Set();
+  async chat(request) {
+    try {
+      return await super.chat(request);
+    } finally {
+      this.scheduleContextRefresh(request.model);
+    }
+  }
+  async *chatStream(request) {
+    try {
+      yield* super.chatStream(request);
+    } finally {
+      this.scheduleContextRefresh(request.model);
+    }
+  }
+  async chatWithTools(request, tools) {
+    try {
+      return await super.chatWithTools(request, tools);
+    } finally {
+      this.scheduleContextRefresh(request.model);
+    }
+  }
+  async *chatWithToolsStream(request, tools) {
+    try {
+      yield* super.chatWithToolsStream(request, tools);
+    } finally {
+      this.scheduleContextRefresh(request.model);
+    }
+  }
+  scheduleContextRefresh(model) {
+    const key = model.replace(/:latest$/, "");
+    if (this.refreshedAfterLoad.has(key)) return;
+    this.refreshedAfterLoad.add(key);
+    void this.refreshLoadedContext(model).catch(() => {
+    });
+  }
+  /** 补读已加载模型的实际窗口，更新同一份权重（digest 相同）的所有别名。测试可直接 await。 */
+  async refreshLoadedContext(model) {
+    const norm2 = (id2) => id2.replace(/:latest$/, "");
+    const [running, digests] = await Promise.all([this.fetchRunningContexts(), this.fetchDigests()]);
+    const self = this.dynamicModels.find((m) => norm2(m.id) === norm2(model));
+    const id = self?.id ?? model;
+    const digest = digests.get(id);
+    const actual = digest && running.get(digest) || running.get(id);
+    if (!actual) return;
+    for (const m of this.dynamicModels) {
+      if (m.id === id || digest && digests.get(m.id) === digest) {
+        m.contextWindow = actual;
+        this.contextSources.set(m.id, "running");
+      }
+    }
+    if (actual < SMALL_CONTEXT_WARN && !this.warnedSmallContext.has(norm2(id))) {
+      this.warnedSmallContext.add(norm2(id));
+      process.stderr.write(
+        `
+[ollama] ${id} is loaded with a ${actual.toLocaleString()}-token context window. aicli's tool definitions alone take about 8K tokens, so earlier messages will be silently cut off. Raise it: set OLLAMA_CONTEXT_LENGTH=32768 (or more) for the Ollama server, or add "PARAMETER num_ctx 32768" to the model's Modelfile.
+`
+      );
+    }
+  }
   /** 实时刷新模型列表 */
   async listModels() {
     await this.tryConnect();
@@ -13418,21 +13564,37 @@ function estimateTokens(text) {
   return Math.ceil(tokens) + 4;
 }
 function estimateToolDefinitionTokens(def) {
-  let charCount = def.name.length + def.description.length;
-  for (const [key, param] of Object.entries(def.parameters)) {
-    charCount += key.length + param.type.length + param.description.length;
-    if (param.enum) {
-      charCount += param.enum.join(" ").length;
-    }
-  }
-  const structuralOverhead = 40 + Object.keys(def.parameters).length * 30;
-  return Math.ceil((charCount + structuralOverhead) * 0.25) + 10;
+  const json = JSON.stringify({ name: def.name, description: def.description, parameters: def.parameters });
+  const cjk = json.match(CJK_REGEX)?.length ?? 0;
+  return Math.ceil((json.length - cjk + 40) / 3.7 + cjk * 1.5);
 }
 
 // src/tools/registry.ts
 import { pathToFileURL } from "url";
 import { existsSync as existsSync19, mkdirSync as mkdirSync9, readdirSync as readdirSync9 } from "fs";
 import { join as join13 } from "path";
+var TRIMMABLE_BUILTIN_ORDER = [
+  "google_search",
+  "recall_memory",
+  "search_code",
+  "notebook_edit",
+  "spawn_agent",
+  "run_interactive",
+  "save_memory",
+  "find_references",
+  "get_outline",
+  "find_symbol",
+  "task_stop",
+  "task_list",
+  "task_create",
+  "git_log",
+  "git_diff",
+  "git_commit",
+  "git_status",
+  "web_fetch",
+  "web_search",
+  "run_tests"
+];
 var ToolRegistry = class {
   tools = /* @__PURE__ */ new Map();
   pluginToolNames = /* @__PURE__ */ new Set();
@@ -13500,7 +13662,7 @@ var ToolRegistry = class {
       totalTokens += estimateToolDefinitionTokens(def);
     }
     if (totalTokens <= tokenBudget) {
-      return { definitions: allDefs, trimmedCount: 0, systemNote: null };
+      return { definitions: allDefs, trimmedCount: 0, trimmedBuiltins: [], systemNote: null };
     }
     const builtinDefs = [];
     const mcpDefs = [];
@@ -13512,10 +13674,19 @@ var ToolRegistry = class {
       }
     }
     let budget = tokenBudget;
-    for (const def of builtinDefs) {
-      budget -= estimateToolDefinitionTokens(def);
-    }
-    const kept = [...builtinDefs];
+    const trimmedBuiltins = [];
+    let builtinCost = builtinDefs.reduce((n, d) => n + estimateToolDefinitionTokens(d), 0);
+    const keptBuiltins = new Set(builtinDefs.map((d) => d.name));
+    for (const name of TRIMMABLE_BUILTIN_ORDER) {
+      if (builtinCost <= budget) break;
+      const def = builtinDefs.find((d) => d.name === name);
+      if (!def || this.pluginToolNames.has(name)) continue;
+      keptBuiltins.delete(name);
+      trimmedBuiltins.push(name);
+      builtinCost -= estimateToolDefinitionTokens(def);
+    }
+    budget -= builtinCost;
+    const kept = builtinDefs.filter((d) => keptBuiltins.has(d.name));
     const used = usedToolNames ?? /* @__PURE__ */ new Set();
     const remaining = [];
     for (const def of mcpDefs) {
@@ -13537,6 +13708,12 @@ var ToolRegistry = class {
         trimmed++;
       }
     }
+    const notes = [];
+    if (trimmedBuiltins.length > 0) {
+      notes.push(
+        `[Tool Budget] To fit this model's context window, these tools are not available in this session: ${trimmedBuiltins.join(", ")}. Use bash for the equivalent (git commands, grep for symbols, curl for web pages) instead of calling them.`
+      );
+    }
     let systemNote = null;
     if (trimmed > 0) {
       const trimmedServers = /* @__PURE__ */ new Set();
@@ -13549,7 +13726,13 @@ var ToolRegistry = class {
       }
       systemNote = `[MCP Tool Budget] ${trimmed} MCP tool(s) from server(s) [${[...trimmedServers].join(", ")}] were excluded to fit the context window. If you need a specific MCP tool that isn't listed, tell the user to ask for it explicitly.`;
     }
-    return { definitions: kept, trimmedCount: trimmed, systemNote };
+    if (systemNote) notes.push(systemNote);
+    return {
+      definitions: kept,
+      trimmedCount: trimmed,
+      trimmedBuiltins,
+      systemNote: notes.length > 0 ? notes.join("\n\n") : null
+    };
   }
   listAll() {
     return [...this.tools.values()];
@@ -17526,7 +17709,7 @@ Cache:  write=${cacheCreate} read=${cacheRead}` : "";
 Cost:   ${formatCost(cost)}` : "";
   let memoryLine = "";
   try {
-    const { countPendingMemories: countPendingMemories2 } = await import("./persistent-memory-IPNMYXU4.js");
+    const { countPendingMemories: countPendingMemories2 } = await import("./persistent-memory-EVZX7SLR.js");
     const pending = countPendingMemories2(ctx.config.getConfigDir());
     if (pending > 0) memoryLine = `
 Memory: \u26A0 ${pending} pending approval \u2014 see the Memory panel or /memory`;
@@ -18514,7 +18697,7 @@ async function handleTest(args, ctx) {
       const isCommand = argStr.includes(" ") || /^(mvn|gradle|npm|pytest|cargo|go)\b/.test(argStr);
       testArgs = isCommand ? { command: argStr } : { filter: argStr };
     }
-    const runTests = ctx.runTests ?? (await import("./run-tests-PDD4KUQL.js")).executeTests;
+    const runTests = ctx.runTests ?? (await import("./run-tests-3SYXE2PJ.js")).executeTests;
     const report = await runTests(testArgs);
     ctx.send({ type: "info", message: report });
   } catch (err) {
@@ -18905,7 +19088,7 @@ async function handleMemory(args, ctx) {
   }
   if (intent.kind === "status") {
     const { getChatIndexStatus: getChatIndexStatus2 } = await import("./chat-index-RLVLQRRC.js");
-    const { listMemoryEntries: listMemoryEntries2, isMemoryExpired: isMemoryExpired2, formatMemoryForPrompt: formatMemoryForPrompt2 } = await import("./persistent-memory-IPNMYXU4.js");
+    const { listMemoryEntries: listMemoryEntries2, isMemoryExpired: isMemoryExpired2, formatMemoryForPrompt: formatMemoryForPrompt2 } = await import("./persistent-memory-EVZX7SLR.js");
     const memories = listMemoryEntries2(configDir, { includeExpired: true, includeRejected: true });
     const injected = formatMemoryForPrompt2(configDir, process.cwd(), ctx.config.get("memory")?.maxChars);
     const view = buildMemoryStatusView({
@@ -18928,7 +19111,7 @@ async function handleMemory(args, ctx) {
     return;
   }
   if (intent.kind === "recall") {
-    const { searchPersistentMemories: searchPersistentMemories2, touchMemoryReferences: touchMemoryReferences2 } = await import("./persistent-memory-IPNMYXU4.js");
+    const { searchPersistentMemories: searchPersistentMemories2, touchMemoryReferences: touchMemoryReferences2 } = await import("./persistent-memory-EVZX7SLR.js");
     const { searchChatMemory: searchChatMemory2 } = await import("./chat-index-RLVLQRRC.js");
     const persistentHits = searchPersistentMemories2(configDir, intent.query, process.cwd(), 3);
     if (persistentHits.length > 0) {
@@ -19049,7 +19232,7 @@ async function handleMemory(args, ctx) {
     ctx.handleMemoryManage(intent.action, args[1], intent.action === "expire" ? args[2] : void 0);
   } else if (intent.kind === "export") {
     if (args[1] === "md") {
-      const { exportMemoryEntries: exportMemoryEntries2 } = await import("./persistent-memory-IPNMYXU4.js");
+      const { exportMemoryEntries: exportMemoryEntries2 } = await import("./persistent-memory-EVZX7SLR.js");
       ctx.send({
         type: "export_data",
         format: "md",
@@ -20054,7 +20237,7 @@ No tools match "${filter}".
           const { join: join26 } = await import("path");
           const { existsSync: existsSync32 } = await import("fs");
           const { getGitRoot: getGitRoot2 } = await import("./git-context-VNCA2GU7.js");
-          const { MCP_PROJECT_CONFIG_NAME: MCP_PROJECT_CONFIG_NAME2 } = await import("./constants-CPHZLOCQ.js");
+          const { MCP_PROJECT_CONFIG_NAME: MCP_PROJECT_CONFIG_NAME2 } = await import("./constants-HF5AYLBR.js");
           const { approveProject, hashMcpFile } = await import("./project-trust-6CJB5ZMX.js");
           const cwd = process.cwd();
           const projectRoot = getGitRoot2(cwd) ?? cwd;
@@ -20973,7 +21156,7 @@ function createReviewCommands() {
       usage: "/test [command|filter]",
       async execute(args, ctx) {
         try {
-          const { executeTests } = await import("./run-tests-PDD4KUQL.js");
+          const { executeTests } = await import("./run-tests-3SYXE2PJ.js");
           const argStr = args.join(" ").trim();
           let testArgs = {};
           if (argStr) {

package/dist/{hub-TCZFG2ZG.js → hub-OYH2UGER.js}

@@ -158,7 +158,7 @@ ${content}`);
   }
 }
 async function runTaskMode(config, providers, configManager, topic) {
-  const { TaskOrchestrator } = await import("./task-orchestrator-544JETWF.js");
+  const { TaskOrchestrator } = await import("./task-orchestrator-QHJEDRDS.js");
   const orchestrator = new TaskOrchestrator(config, providers, configManager);
   let interrupted = false;
   const onSigint = () => {