jinzd-ai-cli 0.4.58 → 0.4.60

package/dist/{chunk-7MQXQDVV.js → chunk-2DWWB4KD.js}

@@ -8,7 +8,7 @@ import { platform } from "os";
 import chalk from "chalk";
 
 // src/core/constants.ts
-var VERSION = "0.4.58";
+var VERSION = "0.4.60";
 var APP_NAME = "ai-cli";
 var CONFIG_DIR_NAME = ".aicli";
 var CONFIG_FILE_NAME = "config.json";

package/dist/{chunk-ZHURWJEW.js → chunk-3YVHYAXK.js}

@@ -6,8 +6,9 @@ import {
   ProviderError,
   ProviderNotFoundError,
   RateLimitError,
-  schemaToJsonSchema
-} from "./chunk-HDNVCYD6.js";
+  schemaToJsonSchema,
+  truncateForPersist
+} from "./chunk-C32FFHMY.js";
 import {
   APP_NAME,
   CONFIG_DIR_NAME,
@@ -20,7 +21,7 @@ import {
   MCP_TOOL_PREFIX,
   PLUGINS_DIR_NAME,
   VERSION
-} from "./chunk-7MQXQDVV.js";
+} from "./chunk-2DWWB4KD.js";
 
 // src/config/config-manager.ts
 import { readFileSync, writeFileSync, existsSync, mkdirSync } from "fs";
@@ -309,9 +310,12 @@ var BaseProvider = class {
   /**
    * 将 Message[] 转换为 OpenAI API 格式的消息数组。
    * content 为 string 时直接传递；为 MessageContentPart[] 时保留数组格式（vision 请求）。
+   *
+   * 自动跳过 role='tool' 和带 toolCalls 的 assistant 消息——
+   * 这些是 v0.4.60+ 持久化的工具历史，由 _extraMessages 机制单独注入。
    */
   normalizeMessages(messages) {
-    return messages.map((m) => ({ role: m.role, content: m.content }));
+    return messages.filter((m) => m.role !== "tool" && !m.toolCalls).map((m) => ({ role: m.role, content: m.content }));
   }
 };
 
@@ -475,7 +479,7 @@ var ClaudeProvider = class extends BaseProvider {
   }
   async chat(request) {
     try {
-      const messages = request.messages.filter((m) => m.role !== "system").map((m) => ({
+      const messages = request.messages.filter((m) => m.role !== "system" && m.role !== "tool" && !m.toolCalls).map((m) => ({
         role: m.role,
         content: this.contentToClaudeParts(m.content)
       }));
@@ -500,7 +504,7 @@ var ClaudeProvider = class extends BaseProvider {
   }
   async *chatStream(request) {
     try {
-      const messages = request.messages.filter((m) => m.role !== "system").map((m) => ({
+      const messages = request.messages.filter((m) => m.role !== "system" && m.role !== "tool" && !m.toolCalls).map((m) => ({
         role: m.role,
         content: this.contentToClaudeParts(m.content)
       }));
@@ -557,7 +561,7 @@ var ClaudeProvider = class extends BaseProvider {
           }
         }))
       );
-      const baseMessages = request.messages.filter((m) => m.role !== "system").map((m) => ({ role: m.role, content: this.contentToClaudeParts(m.content) }));
+      const baseMessages = request.messages.filter((m) => m.role !== "system" && m.role !== "tool" && !m.toolCalls).map((m) => ({ role: m.role, content: this.contentToClaudeParts(m.content) }));
       const extraMessages = request._extraMessages ?? [];
       const allMessages = [...baseMessages, ...extraMessages];
       const { thinking, temperature } = this.buildThinkingParams(request);
@@ -869,7 +873,7 @@ var GeminiProvider = class extends BaseProvider {
     return parts.length > 0 ? parts : [{ text: "" }];
   }
   toGeminiHistory(messages) {
-    return messages.filter((m) => m.role !== "system").map((m) => ({
+    return messages.filter((m) => m.role !== "system" && m.role !== "tool" && !m.toolCalls).map((m) => ({
       role: m.role === "assistant" ? "model" : "user",
       parts: this.contentToGeminiParts(m.content)
     }));
@@ -2521,10 +2525,19 @@ var Session = class _Session {
         messageIndex: c.messageIndex,
         timestamp: c.timestamp.toISOString()
       })),
-      messages: this.messages.map((m) => ({
-        ...m,
-        timestamp: m.timestamp.toISOString()
-      }))
+      messages: this.messages.map((m) => {
+        const out = {
+          role: m.role,
+          content: m.content,
+          timestamp: m.timestamp.toISOString()
+        };
+        if (m.toolCalls) out.toolCalls = m.toolCalls;
+        if (m.reasoningContent !== void 0) out.reasoningContent = m.reasoningContent;
+        if (m.toolCallId) out.toolCallId = m.toolCallId;
+        if (m.toolName) out.toolName = m.toolName;
+        if (m.isError !== void 0) out.isError = m.isError;
+        return out;
+      })
     };
   }
   /**
@@ -2594,11 +2607,17 @@ var Session = class _Session {
     }
     session.messages = d.messages.map((m) => {
       const ts = new Date(m.timestamp);
-      return {
+      const msg = {
         role: m.role ?? "user",
-        content: m.content,
+        content: Array.isArray(m.content) ? m.content : String(m.content ?? ""),
         timestamp: isNaN(ts.getTime()) ? /* @__PURE__ */ new Date() : ts
       };
+      if (Array.isArray(m.toolCalls)) msg.toolCalls = m.toolCalls;
+      if (typeof m.reasoningContent === "string") msg.reasoningContent = m.reasoningContent;
+      if (typeof m.toolCallId === "string") msg.toolCallId = m.toolCallId;
+      if (typeof m.toolName === "string") msg.toolName = m.toolName;
+      if (typeof m.isError === "boolean") msg.isError = m.isError;
+      return msg;
     });
     return session;
   }
@@ -3667,6 +3686,78 @@ function formatCost(amount) {
   return `$${amount.toFixed(2)}`;
 }
 
+// src/session/tool-history.ts
+function persistToolRound(session, toolCalls, toolResults, opts) {
+  session.addMessage({
+    role: "assistant",
+    content: opts?.assistantContent ?? "",
+    toolCalls,
+    reasoningContent: opts?.reasoningContent,
+    timestamp: /* @__PURE__ */ new Date()
+  });
+  for (let i = 0; i < toolCalls.length; i++) {
+    const tc = toolCalls[i];
+    const tr = toolResults[i];
+    if (!tr) continue;
+    session.addMessage({
+      role: "tool",
+      content: truncateForPersist(tr.content),
+      toolCallId: tr.callId,
+      toolName: tc.name,
+      isError: tr.isError,
+      timestamp: /* @__PURE__ */ new Date()
+    });
+  }
+}
+function isToolMessage(m) {
+  return m.role === "tool" || !!(m.toolCalls && m.toolCalls.length > 0);
+}
+function extractToolHistory(messages) {
+  const baseMessages = [];
+  const toolHistory = [];
+  for (const m of messages) {
+    if (isToolMessage(m)) {
+      toolHistory.push(m);
+    } else {
+      baseMessages.push(m);
+    }
+  }
+  return { baseMessages, toolHistory };
+}
+function rebuildExtraMessages(provider, toolHistory) {
+  if (toolHistory.length === 0) return [];
+  const result = [];
+  let i = 0;
+  while (i < toolHistory.length) {
+    const msg = toolHistory[i];
+    if (msg.role === "assistant" && msg.toolCalls && msg.toolCalls.length > 0) {
+      const toolCalls = msg.toolCalls;
+      const toolResults = [];
+      let j = i + 1;
+      while (j < toolHistory.length && toolHistory[j].role === "tool") {
+        const tm = toolHistory[j];
+        toolResults.push({
+          callId: tm.toolCallId ?? "",
+          content: typeof tm.content === "string" ? tm.content : getContentText(tm.content),
+          isError: tm.isError ?? false
+        });
+        j++;
+      }
+      result.push(
+        ...provider.buildToolResultMessages(toolCalls, toolResults, msg.reasoningContent)
+      );
+      i = j;
+    } else {
+      result.push({
+        role: msg.role,
+        content: typeof msg.content === "string" ? msg.content : getContentText(msg.content)
+      });
+      i++;
+    }
+  }
+  return result;
+}
+
 // src/repl/dev-state.ts
 import { existsSync as existsSync5, readFileSync as readFileSync4, writeFileSync as writeFileSync3, unlinkSync as unlinkSync2, mkdirSync as mkdirSync4 } from "fs";
 import { join as join5 } from "path";
@@ -3777,6 +3868,9 @@ export {
   computeCost,
   formatCost,
   parseSimpleYaml,
+  persistToolRound,
+  extractToolHistory,
+  rebuildExtraMessages,
   SNAPSHOT_PROMPT,
   sessionHasMeaningfulContent,
   saveDevState,

package/dist/{chunk-HDNVCYD6.js → chunk-C32FFHMY.js}

@@ -10,7 +10,7 @@ import {
   SUBAGENT_DEFAULT_MAX_ROUNDS,
   SUBAGENT_MAX_ROUNDS_LIMIT,
   runTestsTool
-} from "./chunk-7MQXQDVV.js";
+} from "./chunk-2DWWB4KD.js";
 
 // src/tools/builtin/bash.ts
 import { execSync } from "child_process";
@@ -1098,6 +1098,20 @@ function snapToLineBoundary(content, target, direction) {
     return target;
   }
 }
+var PERSIST_MAX_CHARS = 8192;
+function truncateForPersist(content, maxChars = PERSIST_MAX_CHARS) {
+  if (content.length <= maxChars) return content;
+  const headSize = Math.floor(maxChars * 0.7);
+  const tailSize = Math.floor(maxChars * 0.2);
+  const head = content.slice(0, headSize);
+  const tail = content.slice(-tailSize);
+  const omitted = content.length - headSize - tailSize;
+  return `${head}
+
+[... ${omitted} chars omitted for storage ...]
+
+${tail}`;
+}
 function truncateOutput(content, toolName, maxChars) {
   const limit = maxChars ?? activeMaxChars;
   if (content.length <= limit) return content;
@@ -4208,6 +4222,7 @@ export {
   checkPermission,
   setMaxOutputCap,
   setContextWindow,
+  truncateForPersist,
   truncateOutput,
   ToolExecutor,
   lastResponseStore,

package/dist/{chunk-5CQPX74I.js → chunk-X4GL6D5L.js}

@@ -6,7 +6,7 @@ import { platform } from "os";
 import chalk from "chalk";
 
 // src/core/constants.ts
-var VERSION = "0.4.58";
+var VERSION = "0.4.60";
 var APP_NAME = "ai-cli";
 var CONFIG_DIR_NAME = ".aicli";
 var CONFIG_FILE_NAME = "config.json";

package/dist/{hub-MRK53S5O.js → hub-JTMNY7JR.js}

@@ -385,7 +385,7 @@ ${content}`);
   }
 }
 async function runTaskMode(config, providers, configManager, topic) {
-  const { TaskOrchestrator } = await import("./task-orchestrator-BWFYT4Q5.js");
+  const { TaskOrchestrator } = await import("./task-orchestrator-Z4IK3UEA.js");
   const orchestrator = new TaskOrchestrator(config, providers, configManager);
   let interrupted = false;
   const onSigint = () => {

package/dist/index.js

@@ -13,6 +13,7 @@ import {
   clearDevState,
   computeCost,
   detectsHallucinatedFileOp,
+  extractToolHistory,
   extractWrittenFilePaths,
   findPhantomClaims,
   formatCost,
@@ -24,10 +25,12 @@ import {
   hadPreviousWriteToolCalls,
   loadDevState,
   parseSimpleYaml,
+  persistToolRound,
+  rebuildExtraMessages,
   saveDevState,
   sessionHasMeaningfulContent,
   setupProxy
-} from "./chunk-ZHURWJEW.js";
+} from "./chunk-3YVHYAXK.js";
 import {
   ToolExecutor,
   ToolRegistry,
@@ -41,7 +44,7 @@ import {
   spawnAgentContext,
   theme,
   undoStack
-} from "./chunk-HDNVCYD6.js";
+} from "./chunk-C32FFHMY.js";
 import {
   fileCheckpoints
 } from "./chunk-4BKXL7SM.js";
@@ -66,7 +69,7 @@ import {
   SKILLS_DIR_NAME,
   VERSION,
   buildUserIdentityPrompt
-} from "./chunk-7MQXQDVV.js";
+} from "./chunk-2DWWB4KD.js";
 
 // src/index.ts
 import { program } from "commander";
@@ -2161,7 +2164,7 @@ ${hint}` : "")
       usage: "/test [command|filter]",
       async execute(args, ctx) {
         try {
-          const { executeTests } = await import("./run-tests-ZC6WLEE4.js");
+          const { executeTests } = await import("./run-tests-QGJHXL5Z.js");
           const argStr = args.join(" ").trim();
           let testArgs = {};
           if (argStr) {
@@ -4322,7 +4325,29 @@ Session '${this.resumeSessionId}' not found.
         sendNotification("ai-cli", `Task completed in ${Math.round(elapsed / 1e3)}s`);
       }
     } catch (err) {
-      this.renderer.renderError(err);
+      const errMsg = err instanceof Error ? err.message : String(err);
+      const isCtxLengthError = /maximum context length|context_length_exceeded|context window|too many tokens|reduce the length of the messages/i.test(errMsg);
+      if (isCtxLengthError) {
+        process.stderr.write(
+          theme.error(`
+\u26A0 Context length exceeded \u2014 the conversation is too long for this model.
+`)
+        );
+        process.stderr.write(theme.dim(`  Details: ${errMsg.split("\n")[0]}
+`));
+        process.stderr.write(
+          theme.dim(
+            `  Recovery options:
+    1. Run /compact to summarize old messages and free context
+    2. Run /clear to start a fresh session (keeps this terminal)
+    3. Run /model to switch to a model with a larger context window
+
+`
+          )
+        );
+      } else {
+        this.renderer.renderError(err);
+      }
     }
   }
   /**
@@ -4397,6 +4422,23 @@ Session '${this.resumeSessionId}' not found.
     }
     return total;
   }
+  /**
+   * 估算 agentic 循环当前请求的 token 数（包含 session messages + extraMessages + system prompt）。
+   * extraMessages 结构复杂（含 tool_calls、tool 结果等），这里用 JSON.stringify 后除以字符/token 比。
+   * 用于 handleChatWithTools 循环内每轮发 API 前做上下文压力检查。
+   */
+  estimateRequestTokens(systemPrompt, extraMessages) {
+    let total = this.estimateConversationTokens();
+    if (extraMessages.length > 0) {
+      try {
+        const serialized = JSON.stringify(extraMessages);
+        total += this.estimateTokens(serialized);
+      } catch {
+      }
+    }
+    void systemPrompt;
+    return total;
+  }
   /**
    * 获取当前模型的 context window 大小。
    */
@@ -4830,8 +4872,9 @@ Session '${this.resumeSessionId}' not found.
     if (this.blockedTools) {
       toolDefs = toolDefs.filter((t) => !this.blockedTools.has(t.name));
     }
-    const apiMessages = [...messages];
-    const extraMessages = [];
+    const { baseMessages: cleanMessages, toolHistory } = extractToolHistory(messages);
+    const apiMessages = [...cleanMessages];
+    const extraMessages = toolHistory.length > 0 ? rebuildExtraMessages(provider, toolHistory) : [];
     const maxToolRounds = this.maxToolRoundsOverride ?? this.config.get("maxToolRounds") ?? DEFAULT_MAX_TOOL_ROUNDS;
     const autoPauseIntervalRaw = this.config.get("autoPauseInterval");
     const autoPauseInterval = typeof autoPauseIntervalRaw === "number" ? autoPauseIntervalRaw : DEFAULT_AUTO_PAUSE_INTERVAL;
@@ -4865,6 +4908,7 @@ You have a maximum of ${maxToolRounds} tool call rounds for this task. Plan effi
     let lastToolCallSignature = "";
     let repeatedToolCallCount = 0;
     let emptyResponseRetries = 0;
+    let warnedCtx80 = false;
     const roundToolHistory = [];
     this.setupInterjectionListener();
     try {
@@ -4923,6 +4967,52 @@ You have a maximum of ${maxToolRounds} tool call rounds for this task. Plan effi
 `));
           extraMessages.push({ role: "user", content: msg });
         }
+        const ctxWindow = this.getContextWindowSize();
+        if (ctxWindow > 0) {
+          const reqTokens = this.estimateRequestTokens(systemPrompt, extraMessages);
+          const reqRatio = reqTokens / ctxWindow;
+          if (reqRatio >= 0.95) {
+            spinner.stop();
+            process.stderr.write(
+              theme.error(
+                `
+\u26A0 Context at ${Math.round(reqRatio * 100)}% of ${fmtTokens(ctxWindow)} \u2014 aborting agentic loop before API rejection.
+`
+              )
+            );
+            process.stderr.write(
+              theme.dim(
+                `  Too much tool output accumulated this turn. Your work so far is preserved.
+  Recovery: run /compact to shrink history, then ask the AI to continue.
+
+`
+              )
+            );
+            if (roundUsage.inputTokens > 0 || roundUsage.outputTokens > 0) {
+              this.addSessionUsage(roundUsage);
+              session.addTokenUsage(roundUsage);
+              if (this.shouldShowTokens()) {
+                this.renderer.renderUsage(roundUsage, this.sessionTokenUsage);
+              }
+            }
+            return;
+          } else if (reqRatio >= 0.8 && !warnedCtx80) {
+            warnedCtx80 = true;
+            spinner.stop();
+            process.stdout.write(
+              theme.warning(
+                `
+\u26A0 Context at ${Math.round(reqRatio * 100)}% of ${fmtTokens(ctxWindow)} \u2014 asking AI to wrap up.
+`
+              )
+            );
+            extraMessages.push({
+              role: "user",
+              content: `\u26A0\uFE0F Context pressure: ~${Math.round(reqRatio * 100)}% of the ${fmtTokens(ctxWindow)} context window is used. Avoid reading more files or running broad scans. Finish the current critical step, then produce a final summary. Every unnecessary tool call now risks breaking the conversation.`
+            });
+            spinner.start(`Thinking... (round ${round + 1}/${maxToolRounds})`);
+          }
+        }
         let result;
         let alreadyRendered = false;
         const chatRequest = {
@@ -5174,6 +5264,11 @@ You have a maximum of ${maxToolRounds} tool call rounds for this task. Plan effi
         const reasoningContent = "reasoningContent" in result ? result.reasoningContent : void 0;
         const newMsgs = provider.buildToolResultMessages(result.toolCalls, toolResults, reasoningContent);
         extraMessages.push(...newMsgs);
+        const streamedContent = "content" in result ? result.content : void 0;
+        persistToolRound(session, result.toolCalls, toolResults, {
+          assistantContent: streamedContent,
+          reasoningContent
+        });
         const thisRoundHadWrite = result.toolCalls.some(
           (tc) => tc.name === "write_file" || tc.name === "edit_file"
         );
@@ -5609,7 +5704,7 @@ program.command("web").description("Start Web UI server with browser-based chat
     console.error("Error: Invalid port number. Must be between 1 and 65535.");
     process.exit(1);
   }
-  const { startWebServer } = await import("./server-SFDOVFUN.js");
+  const { startWebServer } = await import("./server-L2XJYXMB.js");
   await startWebServer({ port, host: options.host });
 });
 program.command("user [action] [username]").description("Manage Web UI users (list | create <name> | delete <name> | reset-password <name> | migrate <name>)").action(async (action, username) => {
@@ -5842,7 +5937,7 @@ program.command("hub [topic]").description("Start multi-agent hub (discuss / bra
     }),
     config.get("customProviders")
   );
-  const { startHub } = await import("./hub-MRK53S5O.js");
+  const { startHub } = await import("./hub-JTMNY7JR.js");
   await startHub(
     {
       topic: topic ?? "",

package/dist/{run-tests-ZC6WLEE4.js → run-tests-QGJHXL5Z.js}

@@ -2,7 +2,7 @@
 import {
   executeTests,
   runTestsTool
-} from "./chunk-7MQXQDVV.js";
+} from "./chunk-2DWWB4KD.js";
 export {
   executeTests,
   runTestsTool

package/dist/{run-tests-DL6HGIK3.js → run-tests-WGBDMO4H.js}

@@ -1,7 +1,7 @@
 import {
   executeTests,
   runTestsTool
-} from "./chunk-5CQPX74I.js";
+} from "./chunk-X4GL6D5L.js";
 export {
   executeTests,
   runTestsTool

package/dist/{server-SFDOVFUN.js → server-L2XJYXMB.js}

@@ -9,6 +9,7 @@ import {
   TOOL_CALL_REMINDER,
   computeCost,
   detectsHallucinatedFileOp,
+  extractToolHistory,
   formatCost,
   formatGitContextForPrompt,
   getContentText,
@@ -16,8 +17,10 @@ import {
   getGitRoot,
   hadPreviousWriteToolCalls,
   loadDevState,
+  persistToolRound,
+  rebuildExtraMessages,
   setupProxy
-} from "./chunk-ZHURWJEW.js";
+} from "./chunk-3YVHYAXK.js";
 import {
   AuthManager
 } from "./chunk-BYNY5JPB.js";
@@ -36,7 +39,7 @@ import {
   spawnAgentContext,
   truncateOutput,
   undoStack
-} from "./chunk-HDNVCYD6.js";
+} from "./chunk-C32FFHMY.js";
 import "./chunk-4BKXL7SM.js";
 import {
   AGENTIC_BEHAVIOR_GUIDELINE,
@@ -56,7 +59,7 @@ import {
   SKILLS_DIR_NAME,
   VERSION,
   buildUserIdentityPrompt
-} from "./chunk-7MQXQDVV.js";
+} from "./chunk-2DWWB4KD.js";
 
 // src/web/server.ts
 import express from "express";
@@ -642,6 +645,49 @@ var SessionHandler = class _SessionHandler {
     } catch {
     }
   }
+  /** 获取当前模型的 context window 大小（0 表示未知）*/
+  getContextWindowSize() {
+    try {
+      const provider = this.providers.get(this.currentProvider);
+      const modelInfo = provider?.info.models.find((m) => m.id === this.currentModel);
+      return modelInfo?.contextWindow ?? 0;
+    } catch {
+      return 0;
+    }
+  }
+  /** 粗略估算文本 token 数（2.5 chars/token）*/
+  estTokens(text) {
+    return Math.ceil(text.length / 2.5);
+  }
+  /**
+   * 估算当前 agentic 请求总 token 数（session messages + extraMessages + system prompt）。
+   * 用于 handleChat 循环内每轮发 API 前的压力检查。
+   */
+  estimateRequestTokens(systemPrompt, extraMessages) {
+    let total = 0;
+    if (systemPrompt) total += this.estTokens(systemPrompt);
+    const session = this.sessions.current;
+    if (session) {
+      for (const msg of session.messages) {
+        if (typeof msg.content === "string") {
+          total += this.estTokens(msg.content);
+        } else if (Array.isArray(msg.content)) {
+          for (const part of msg.content) {
+            if (part.type === "text" && part.text) {
+              total += this.estTokens(part.text);
+            }
+          }
+        }
+      }
+    }
+    if (extraMessages.length > 0) {
+      try {
+        total += this.estTokens(JSON.stringify(extraMessages));
+      } catch {
+      }
+    }
+    return total;
+  }
   /** Save session only if it exists and has messages (never persist empty "Untitled" sessions). */
   saveIfNeeded() {
     if (this.sessions.current && this.sessions.current.messages.length > 0) {
@@ -771,8 +817,9 @@ var SessionHandler = class _SessionHandler {
   }
   async handleChatWithTools(provider, messages, toolDefs) {
     const session = this.sessions.current;
-    const apiMessages = [...messages];
-    const extraMessages = [];
+    const { baseMessages: cleanMessages, toolHistory } = extractToolHistory(messages);
+    const apiMessages = [...cleanMessages];
+    const extraMessages = toolHistory.length > 0 ? rebuildExtraMessages(provider, toolHistory) : [];
     const maxToolRounds = this.config.get("maxToolRounds") ?? DEFAULT_MAX_TOOL_ROUNDS;
     const autoPauseIntervalRaw = this.config.get("autoPauseInterval");
     const autoPauseInterval = typeof autoPauseIntervalRaw === "number" ? autoPauseIntervalRaw : 50;
@@ -801,6 +848,7 @@ You have a maximum of ${maxToolRounds} tool call rounds for this task. Plan effi
     let warnedLow = false;
     let warnedCritical = false;
     let emptyResponseRetries = 0;
+    let warnedCtx80 = false;
     const ac = new AbortController();
     this.abortController = ac;
     try {
@@ -836,6 +884,38 @@ You have a maximum of ${maxToolRounds} tool call rounds for this task. Plan effi
           this.send({ type: "info", message: `\u26A1 Interjection: "${msg}"` });
           extraMessages.push({ role: "user", content: msg });
         }
+        const ctxWindow = this.getContextWindowSize();
+        if (ctxWindow > 0) {
+          const reqTokens = this.estimateRequestTokens(systemPrompt, extraMessages);
+          const reqRatio = reqTokens / ctxWindow;
+          if (reqRatio >= 0.95) {
+            this.send({
+              type: "response_done",
+              content: `\u26A0 Context at ${Math.round(reqRatio * 100)}% of ${ctxWindow.toLocaleString()} tokens \u2014 aborting before API rejection.
+
+Too much tool output accumulated this turn. Your work so far is preserved.
+
+**Recovery**:
+1. Run \`/compact\` to shrink history, then ask the AI to continue
+2. Run \`/clear\` to start fresh
+3. Switch to a larger-context model`,
+              usage: roundUsage
+            });
+            this.addWebSessionUsage(roundUsage);
+            session.addTokenUsage(roundUsage);
+            return;
+          } else if (reqRatio >= 0.8 && !warnedCtx80) {
+            warnedCtx80 = true;
+            this.send({
+              type: "info",
+              message: `\u26A0 Context at ${Math.round(reqRatio * 100)}% \u2014 asking AI to wrap up`
+            });
+            extraMessages.push({
+              role: "user",
+              content: `\u26A0\uFE0F Context pressure: ~${Math.round(reqRatio * 100)}% of the ${ctxWindow.toLocaleString()}-token context window is used. Avoid reading more files or running broad scans. Finish the current critical step, then produce a final summary. Every unnecessary tool call now risks breaking the conversation.`
+            });
+          }
+        }
         const chatRequest = {
           messages: apiMessages,
           model: this.currentModel,
@@ -850,11 +930,34 @@ You have a maximum of ${maxToolRounds} tool call rounds for this task. Plan effi
           ...extraMessages.length > 0 ? { _extraMessages: extraMessages } : {}
         };
         let result;
-        if (supportsStreamingTools) {
-          const streamGen = provider.chatWithToolsStream(chatRequest, toolDefs);
-          result = await this.consumeToolStream(streamGen, ac);
-        } else {
-          result = await provider.chatWithTools(chatRequest, toolDefs);
+        try {
+          if (supportsStreamingTools) {
+            const streamGen = provider.chatWithToolsStream(chatRequest, toolDefs);
+            result = await this.consumeToolStream(streamGen, ac);
+          } else {
+            result = await provider.chatWithTools(chatRequest, toolDefs);
+          }
+        } catch (providerErr) {
+          const errMsg = providerErr instanceof Error ? providerErr.message : String(providerErr);
+          const isCtxLengthError = /maximum context length|context_length_exceeded|context window|too many tokens|reduce the length of the messages/i.test(errMsg);
+          if (isCtxLengthError) {
+            this.send({
+              type: "response_done",
+              content: `\u26A0 Context length exceeded \u2014 the conversation is too long for this model.
+
+Details: ${errMsg.split("\n")[0]}
+
+**Recovery options**:
+1. Run \`/compact\` to summarize old messages and free context
+2. Run \`/clear\` to start a fresh session
+3. Run \`/model\` to switch to a model with a larger context window`,
+              usage: roundUsage
+            });
+            this.addWebSessionUsage(roundUsage);
+            session.addTokenUsage(roundUsage);
+            return;
+          }
+          throw providerErr;
         }
         if (ac.signal.aborted) break;
         if (result.usage) {
@@ -921,6 +1024,10 @@ You have a maximum of ${maxToolRounds} tool call rounds for this task. Plan effi
           const reasoningContent = result.reasoningContent;
           const newMsgs = provider.buildToolResultMessages(result.toolCalls, toolResults, reasoningContent);
           extraMessages.push(...newMsgs);
+          persistToolRound(session, result.toolCalls, toolResults, {
+            assistantContent: result.content,
+            reasoningContent
+          });
           const allFree = result.toolCalls.every((tc) => FREE_ROUND_TOOLS.has(tc.name));
           if (allFree) {
             consecutiveFreeRounds++;
@@ -1816,7 +1923,7 @@ ${undoResults.map((r) => `  \u2022 ${r}`).join("\n")}` });
       case "test": {
         this.send({ type: "info", message: "\u{1F9EA} Running tests..." });
         try {
-          const { executeTests } = await import("./run-tests-ZC6WLEE4.js");
+          const { executeTests } = await import("./run-tests-QGJHXL5Z.js");
           const argStr = args.join(" ").trim();
           let testArgs = {};
           if (argStr) {

package/dist/{task-orchestrator-BWFYT4Q5.js → task-orchestrator-Z4IK3UEA.js}

@@ -4,11 +4,11 @@ import {
   getDangerLevel,
   googleSearchContext,
   truncateOutput
-} from "./chunk-HDNVCYD6.js";
+} from "./chunk-C32FFHMY.js";
 import "./chunk-4BKXL7SM.js";
 import {
   SUBAGENT_ALLOWED_TOOLS
-} from "./chunk-7MQXQDVV.js";
+} from "./chunk-2DWWB4KD.js";
 
 // src/hub/task-orchestrator.ts
 import { createInterface } from "readline";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "jinzd-ai-cli",
-  "version": "0.4.58",
+  "version": "0.4.60",
   "description": "Cross-platform REPL-style AI CLI with multi-provider support",
   "type": "module",
   "main": "./dist/index.js",