jinzd-ai-cli 0.4.180 → 0.4.181

package/dist/{run-tests-QFTKAWW6.js → run-tests-673ABQQE.js}

@@ -2,8 +2,8 @@
 import {
   executeTests,
   runTestsTool
-} from "./chunk-H7TXZO6D.js";
-import "./chunk-KRU4DFRH.js";
+} from "./chunk-D4ZTYYBU.js";
+import "./chunk-CPZ7KG3Y.js";
 import "./chunk-PDX44BCA.js";
 export {
   executeTests,

package/dist/{run-tests-BUII3HBU.js → run-tests-HZVKHQ33.js}

@@ -1,7 +1,7 @@
 import {
   executeTests,
   runTestsTool
-} from "./chunk-YDI22R3P.js";
+} from "./chunk-D774AWKW.js";
 import "./chunk-3RG5ZIWI.js";
 export {
   executeTests,

package/dist/{server-I6Y5TP2Z.js → server-RCUIX4R5.js}

@@ -1,13 +1,13 @@
 #!/usr/bin/env node
 import {
   ToolRegistry
-} from "./chunk-7WFMYFHC.js";
+} from "./chunk-ORXPLVM7.js";
 import "./chunk-HDSKW7Q3.js";
 import "./chunk-ZWVIDFGY.js";
-import "./chunk-H7TXZO6D.js";
+import "./chunk-D4ZTYYBU.js";
 import {
   runTool
-} from "./chunk-WY2DDJTH.js";
+} from "./chunk-4VB6UP4W.js";
 import {
   getDangerLevel,
   schemaToJsonSchema
@@ -15,7 +15,7 @@ import {
 import "./chunk-TZQHYZKT.js";
 import {
   VERSION
-} from "./chunk-KRU4DFRH.js";
+} from "./chunk-CPZ7KG3Y.js";
 import "./chunk-4BKXL7SM.js";
 import "./chunk-MM3F43H6.js";
 import "./chunk-KHYD3WXE.js";

package/dist/{server-KT6GRCM7.js → server-SLDE3AML.js}

@@ -21,7 +21,7 @@ import {
   loadDevState,
   persistToolRound,
   setupProxy
-} from "./chunk-FQHEXIYP.js";
+} from "./chunk-6SY45V62.js";
 import {
   ToolExecutor,
   ToolRegistry,
@@ -39,20 +39,29 @@ import {
   spawnAgentContext,
   truncateOutput,
   undoStack
-} from "./chunk-7WFMYFHC.js";
+} from "./chunk-ORXPLVM7.js";
 import "./chunk-HDSKW7Q3.js";
 import "./chunk-ZWVIDFGY.js";
-import "./chunk-H7TXZO6D.js";
+import "./chunk-D4ZTYYBU.js";
 import {
   SessionManager,
   getContentText
 } from "./chunk-RIVZNS3K.js";
 import {
+  BudgetWarner,
   CONTENT_ONLY_STREAM_REMINDER,
+  ContextPressureMonitor,
+  EmptyResponseGuard,
+  FreeRoundTracker,
   HALLUCINATION_CORRECTION_MESSAGE,
   ProviderRegistry,
   TEE_FINAL_USER_NUDGE,
   TOOL_CALL_REMINDER,
+  accumulateUsage,
+  buildRoundBudgetHint,
+  buildRoundsExhaustedPrompt,
+  buildUserStopMessage,
+  consumeToolCallStream,
   detectMetaNarration,
   detectPseudoToolCalls,
   detectsHallucinatedFileOp,
@@ -60,17 +69,18 @@ import {
   hadPreviousWriteToolCalls,
   looksLikeDocumentBody,
   stripPseudoToolCalls,
-  stripToolCallReminder
-} from "./chunk-2TWARH5X.js";
+  stripToolCallReminder,
+  summarizeRecentTools
+} from "./chunk-5LK7H45B.js";
 import {
   runTool
-} from "./chunk-WY2DDJTH.js";
+} from "./chunk-4VB6UP4W.js";
 import {
   getDangerLevel
 } from "./chunk-HIU2SH4V.js";
 import {
   ConfigManager
-} from "./chunk-2UGK2RXK.js";
+} from "./chunk-P2VFMUR5.js";
 import "./chunk-TZQHYZKT.js";
 import {
   AGENTIC_BEHAVIOR_GUIDELINE,
@@ -90,7 +100,7 @@ import {
   SKILLS_DIR_NAME,
   VERSION,
   buildUserIdentityPrompt
-} from "./chunk-KRU4DFRH.js";
+} from "./chunk-CPZ7KG3Y.js";
 import {
   formatGitContextForPrompt,
   getGitContext,
@@ -516,8 +526,6 @@ function loadMemoryContent(configDir) {
 import { existsSync as existsSync3, readFileSync as readFileSync3, appendFileSync, writeFileSync, mkdirSync, readdirSync, statSync, createWriteStream, unlinkSync } from "fs";
 import { join as join2, resolve, dirname } from "path";
 import { execSync } from "child_process";
-var FREE_ROUND_TOOLS = /* @__PURE__ */ new Set(["write_todos"]);
-var MAX_CONSECUTIVE_FREE_ROUNDS = 5;
 var SessionHandler = class _SessionHandler {
   ws;
   config;
@@ -1002,15 +1010,7 @@ var SessionHandler = class _SessionHandler {
     const autoPauseIntervalRaw = this.config.get("autoPauseInterval");
     const autoPauseInterval = typeof autoPauseIntervalRaw === "number" ? autoPauseIntervalRaw : 50;
     const { stable: toolStable, volatile: toolVolatile } = this.buildSystemPrompt();
-    const pauseHint = autoPauseInterval > 0 ? `
-- Every ${autoPauseInterval} rounds the user will be asked whether to continue \u2014 use this as a natural checkpoint to report progress.` : "";
-    const roundBudgetHint = `
-
-[Tool Round Budget]
-You have a maximum of ${maxToolRounds} tool call rounds for this task. Plan efficiently:
-- Prefer batch operations (e.g. global find-and-replace) over repetitive single edits.
-- Prioritize the most critical tasks first in case rounds run out.
-- When remaining rounds are low, focus on completing the current task and summarizing.${pauseHint}`;
+    const roundBudgetHint = buildRoundBudgetHint({ maxToolRounds, autoPauseInterval });
     const systemPrompt = toolStable + TOOL_CALL_REMINDER + roundBudgetHint + (mcpBudgetNote ? `
 
 ${mcpBudgetNote}` : "");
@@ -1018,18 +1018,11 @@ ${mcpBudgetNote}` : "");
     const modelParams = this.getModelParams();
     const roundUsage = { inputTokens: 0, outputTokens: 0, cacheCreationTokens: 0, cacheReadTokens: 0 };
     const supportsStreamingTools = typeof provider.chatWithToolsStream === "function";
-    let consecutiveFreeRounds = 0;
     const roundToolHistory = [];
-    const warnNoteAt = Math.max(10, Math.floor(maxToolRounds * 0.2));
-    const warnLowAt = Math.max(5, Math.floor(maxToolRounds * 0.1));
-    const warnCriticalAt = Math.max(3, Math.floor(maxToolRounds * 0.05));
-    const warnLowEff = Math.min(warnLowAt, warnNoteAt - 1);
-    const warnCriticalEff = Math.min(warnCriticalAt, warnLowEff - 1);
-    let warnedNote = false;
-    let warnedLow = false;
-    let warnedCritical = false;
-    let emptyResponseRetries = 0;
-    let warnedCtx80 = false;
+    const budgetWarner = new BudgetWarner(maxToolRounds);
+    const emptyGuard = new EmptyResponseGuard();
+    const ctxMonitor = new ContextPressureMonitor();
+    const freeRounds = new FreeRoundTracker();
     const ac = new AbortController();
     this.abortController = ac;
     try {
@@ -1037,27 +1030,12 @@ ${mcpBudgetNote}` : "");
         if (ac.signal.aborted) break;
         this.toolExecutor.setRoundInfo(round + 1, maxToolRounds);
         this.send({ type: "round_progress", current: round + 1, total: maxToolRounds });
-        const roundsLeft = maxToolRounds - round;
-        if (!warnedCritical && roundsLeft <= warnCriticalEff) {
-          warnedCritical = true;
-          extraMessages.push({
-            role: "user",
-            content: `\u{1F6A8} Critical budget: Only ${roundsLeft} rounds left! Wrap up NOW \u2014 complete the current operation and give a final summary. Do NOT start new tasks.`
-          });
-          this.send({ type: "info", message: `\u{1F6A8} Critical: ${roundsLeft} rounds remaining` });
-        } else if (!warnedLow && roundsLeft <= warnLowEff) {
-          warnedLow = true;
-          extraMessages.push({
-            role: "user",
-            content: `\u26A0\uFE0F Budget warning: Only ${roundsLeft} tool rounds remaining. Prioritize completing the most critical task. If you cannot finish everything, summarize what's done and what remains.`
-          });
-          this.send({ type: "info", message: `\u26A0\uFE0F Low budget: ${roundsLeft} rounds remaining` });
-        } else if (!warnedNote && roundsLeft <= warnNoteAt) {
-          warnedNote = true;
-          extraMessages.push({
-            role: "user",
-            content: `\u{1F4CA} Budget note: ${roundsLeft} tool rounds remaining out of ${maxToolRounds}. Plan your remaining work efficiently \u2014 use batch operations (e.g., replaceAll) when possible.`
-          });
+        const budgetWarning = budgetWarner.check(maxToolRounds - round);
+        if (budgetWarning) {
+          extraMessages.push({ role: "user", content: budgetWarning.injectMessage });
+          if (budgetWarning.displayMessage) {
+            this.send({ type: "info", message: budgetWarning.displayMessage });
+          }
         }
         if (this.userInterjection) {
           const msg = this.userInterjection;
@@ -1068,11 +1046,11 @@ ${mcpBudgetNote}` : "");
         const ctxWindow = this.getContextWindowSize();
         if (ctxWindow > 0) {
           const reqTokens = this.estimateRequestTokens(systemPrompt, extraMessages);
-          const reqRatio = reqTokens / ctxWindow;
-          if (reqRatio >= 0.95) {
+          const pressure = ctxMonitor.check(reqTokens, ctxWindow);
+          if (pressure.action === "abort") {
             this.send({
               type: "response_done",
-              content: `\u26A0 Context at ${Math.round(reqRatio * 100)}% of ${ctxWindow.toLocaleString()} tokens \u2014 aborting before API rejection.
+              content: `\u26A0 Context at ${Math.round(pressure.ratio * 100)}% of ${ctxWindow.toLocaleString()} tokens \u2014 aborting before API rejection.
 
 Too much tool output accumulated this turn. Your work so far is preserved.
 
@@ -1085,16 +1063,12 @@ Too much tool output accumulated this turn. Your work so far is preserved.
             this.addWebSessionUsage(roundUsage);
             session.addTokenUsage(roundUsage);
             return;
-          } else if (reqRatio >= 0.8 && !warnedCtx80) {
-            warnedCtx80 = true;
+          } else if (pressure.action === "warn") {
             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.`
+              message: `\u26A0 Context at ${Math.round(pressure.ratio * 100)}% \u2014 asking AI to wrap up`
             });
+            extraMessages.push({ role: "user", content: pressure.injectMessage });
           }
         }
         const chatRequest = {
@@ -1142,37 +1116,29 @@ Details: ${errMsg.split("\n")[0]}
           throw providerErr;
         }
         if (ac.signal.aborted) break;
-        if (result.usage) {
-          roundUsage.inputTokens += result.usage.inputTokens;
-          roundUsage.outputTokens += result.usage.outputTokens;
-          roundUsage.cacheCreationTokens += result.usage.cacheCreationTokens ?? 0;
-          roundUsage.cacheReadTokens += result.usage.cacheReadTokens ?? 0;
-        }
+        accumulateUsage(roundUsage, result.usage);
         const hasToolCalls = !!(result.toolCalls && result.toolCalls.length > 0);
         const contentBlank = !result.content || result.content.trim() === "";
         if (!hasToolCalls && contentBlank) {
-          if (emptyResponseRetries === 0 && round < maxToolRounds - 1) {
-            emptyResponseRetries++;
-            this.send({
-              type: "info",
-              message: "\u26A0 AI returned an empty response. Nudging to continue..."
-            });
-            extraMessages.push({
-              role: "user",
-              content: "Your previous response was empty \u2014 no text and no tool calls. This usually means the context window is nearly full. Please either: (1) continue the task by calling the next tool you need, or (2) give a concise final text summary of what has been accomplished so far and what remains. Do NOT repeat earlier long outputs."
-            });
+          const decision = emptyGuard.onEmpty(round < maxToolRounds - 1, result.finishReason);
+          if (decision.action === "nudge") {
+            this.send({ type: "info", message: decision.displayMessage });
+            extraMessages.push({ role: "user", content: decision.injectMessage });
             continue;
           }
           this.send({
             type: "response_done",
-            content: "\u26A0 AI returned empty responses twice in a row. Stopping agentic loop.\n\nLikely causes: context window exhausted, max_tokens too low, or content filter.\nTry: /compact to reduce context, /clear to reset, or switch to a larger-context model.",
+            content: `${decision.displayMessage}
+
+${decision.hint}
+Try: /compact to reduce context, /clear to reset, or switch to a larger-context model.`,
             usage: roundUsage
           });
           this.addWebSessionUsage(roundUsage);
           session.addTokenUsage(roundUsage);
           return;
         }
-        emptyResponseRetries = 0;
+        emptyGuard.onNonEmpty();
         if (result.content && !result.toolCalls) {
           const hasWriteTools = toolDefs.some((t) => t.name === "write_file" || t.name === "edit_file");
           const alreadyWrote = hadPreviousWriteToolCalls(extraMessages);
@@ -1252,7 +1218,7 @@ ${systemPromptVolatile}` : systemPrompt;
               assistantContent: teeResult.content,
               reasoningContent: reasoningContent2
             });
-            consecutiveFreeRounds = 0;
+            freeRounds.apply(result.toolCalls.map((tc) => tc.name));
             continue;
           }
           const toolResults = await this.toolExecutor.executeAll(result.toolCalls);
@@ -1263,14 +1229,8 @@ ${systemPromptVolatile}` : systemPrompt;
             assistantContent: result.content,
             reasoningContent
           });
-          const allFree = result.toolCalls.every((tc) => FREE_ROUND_TOOLS.has(tc.name));
-          if (allFree) {
-            consecutiveFreeRounds++;
-            if (consecutiveFreeRounds <= MAX_CONSECUTIVE_FREE_ROUNDS) {
-              round--;
-            }
-          } else {
-            consecutiveFreeRounds = 0;
+          if (freeRounds.apply(result.toolCalls.map((tc) => tc.name))) {
+            round--;
           }
           if (this.userInterjection) {
             const msg = this.userInterjection;
@@ -1282,12 +1242,7 @@ ${systemPromptVolatile}` : systemPrompt;
         const effectiveRound = round + 1;
         const remaining = maxToolRounds - effectiveRound;
         if (autoPauseInterval > 0 && effectiveRound > 0 && effectiveRound % autoPauseInterval === 0 && remaining > 0 && !ac.signal.aborted) {
-          const recentHistory = roundToolHistory.slice(-autoPauseInterval);
-          const toolCounts = /* @__PURE__ */ new Map();
-          for (const rh of recentHistory) {
-            for (const t of rh.tools) toolCounts.set(t, (toolCounts.get(t) || 0) + 1);
-          }
-          const toolSummary = [...toolCounts.entries()].sort((a, b) => b[1] - a[1]).map(([name, count]) => count > 1 ? `${name}\xD7${count}` : name).join(", ");
+          const toolSummary = summarizeRecentTools(roundToolHistory, autoPauseInterval);
           const requestId = `pause_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`;
           const pauseResp = await new Promise((resolve3) => {
             this.pendingAutoPause.set(requestId, resolve3);
@@ -1302,10 +1257,7 @@ ${systemPromptVolatile}` : systemPrompt;
           if (ac.signal.aborted) break;
           if (pauseResp.action === "stop") {
             this.send({ type: "info", message: `\u23F8  Stopped by user at ${effectiveRound}/${maxToolRounds}` });
-            extraMessages.push({
-              role: "user",
-              content: `The user has stopped the task at round ${effectiveRound}/${maxToolRounds}. Do not call any more tools. Summarize what has been completed and what remains.`
-            });
+            extraMessages.push({ role: "user", content: buildUserStopMessage(effectiveRound, maxToolRounds) });
             break;
           } else if (pauseResp.action === "redirect" && pauseResp.message) {
             this.send({ type: "info", message: `\u26A1 Redirect: "${pauseResp.message}"` });
@@ -1316,13 +1268,7 @@ ${systemPromptVolatile}` : systemPrompt;
       try {
         const summaryExtra = [
           ...extraMessages,
-          {
-            role: "user",
-            content: `You have used all ${maxToolRounds} tool call rounds. Do not call any more tools. Summarize in text:
-1. What work has been completed so far
-2. What tasks remain unfinished
-3. What the user can do next`
-          }
+          { role: "user", content: buildRoundsExhaustedPrompt(maxToolRounds) }
         ];
         const summaryResult = await provider.chatWithTools(
           {
@@ -1509,69 +1455,33 @@ This fresh stream has NO tools. Produce ONLY the document body: start with a mar
     });
     return { content: fullContent, summary, isError };
   }
-  /** Consume streaming tool call events and forward to client */
+  /**
+   * Consume streaming tool call events and forward to client.
+   *
+   * v0.4.181: 委托给 core/agent-loop 的统一消费器（与 REPL 同一实现）。
+   * Web 端由此获得三个此前只在 REPL 修过的行为：内联 <think> 折叠
+   * （MiniMax 推理泄漏不再渲染进浏览器）、截断 JSON 自动修复、
+   * 工具调用按 event.index 键累积（稀疏 index 不错位）。
+   */
   async consumeToolStream(streamGen, ac) {
-    let textContent = "";
-    const toolCalls = [];
-    const toolArgBuffers = /* @__PURE__ */ new Map();
-    let usage;
-    let rawContent;
-    let reasoningContent;
-    for await (const event of streamGen) {
-      if (ac.signal.aborted) break;
-      switch (event.type) {
-        case "text_delta":
-          textContent += event.delta;
-          this.send({ type: "text_delta", delta: event.delta });
-          break;
-        case "thinking_start":
-          this.send({ type: "thinking_start" });
-          break;
-        case "thinking_delta":
-          this.send({ type: "thinking_delta", delta: event.delta });
-          break;
-        case "thinking_end":
-          this.send({ type: "thinking_end" });
-          break;
-        case "tool_call_start":
-          toolArgBuffers.set(event.index, "");
-          this.send({ type: "info", message: `\u2699 Streaming: ${event.name}...` });
-          toolCalls.push({ id: event.id, name: event.name, arguments: {} });
-          break;
-        case "tool_call_delta": {
-          const buf = (toolArgBuffers.get(event.index) ?? "") + event.argumentsDelta;
-          toolArgBuffers.set(event.index, buf);
-          break;
-        }
-        case "tool_call_end": {
-          const tc = toolCalls.find((_, i) => i === event.index);
-          const argStr = toolArgBuffers.get(event.index) ?? "{}";
-          if (tc) {
-            try {
-              tc.arguments = JSON.parse(argStr);
-            } catch {
-              tc.arguments = { _raw: argStr };
-            }
-          }
-          break;
-        }
-        case "done":
-          if (event.usage) usage = event.usage;
-          if (event.rawContent) rawContent = event.rawContent;
-          if (event.reasoningContent) reasoningContent = event.reasoningContent;
-          break;
-      }
-    }
-    if (toolCalls.length > 0) {
-      if (rawContent) {
-        toolCalls._rawContent = rawContent;
-      }
-      if (textContent) {
-        toolCalls._streamedText = textContent;
-      }
-      return { toolCalls, usage, reasoningContent };
+    const result = await consumeToolCallStream(streamGen, {
+      signal: ac.signal,
+      onText: (visible) => this.send({ type: "text_delta", delta: visible }),
+      onThinkingStart: () => this.send({ type: "thinking_start" }),
+      onThinkingDelta: (delta) => this.send({ type: "thinking_delta", delta }),
+      onThinkingEnd: () => this.send({ type: "thinking_end" }),
+      onToolCallStart: (_index, _id, name) => this.send({ type: "info", message: `\u2699 Streaming: ${name}...` }),
+      onWarn: (message) => this.send({ type: "info", message: `\u26A0 ${message}` })
+    });
+    if (result.toolCalls.length > 0) {
+      return { toolCalls: result.toolCalls, usage: result.usage, reasoningContent: result.reasoningContent };
     }
-    return { content: textContent, usage, reasoningContent };
+    return {
+      content: result.textContent,
+      usage: result.usage,
+      reasoningContent: result.reasoningContent,
+      finishReason: result.finishReason
+    };
   }
   // ── Commands ─────────────────────────────────────────────────────
   async handleCommand(name, args) {
@@ -2577,7 +2487,7 @@ ${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-QFTKAWW6.js");
+          const { executeTests } = await import("./run-tests-673ABQQE.js");
           const argStr = args.join(" ").trim();
           let testArgs = {};
           if (argStr) {

package/dist/{task-orchestrator-6RLY5JAL.js → task-orchestrator-42D5LBAX.js}

@@ -3,20 +3,20 @@ import {
   ToolRegistry,
   googleSearchContext,
   truncateOutput
-} from "./chunk-7WFMYFHC.js";
+} from "./chunk-ORXPLVM7.js";
 import "./chunk-HDSKW7Q3.js";
 import "./chunk-ZWVIDFGY.js";
-import "./chunk-H7TXZO6D.js";
+import "./chunk-D4ZTYYBU.js";
 import {
   runTool
-} from "./chunk-WY2DDJTH.js";
+} from "./chunk-4VB6UP4W.js";
 import {
   getDangerLevel
 } from "./chunk-HIU2SH4V.js";
 import "./chunk-TZQHYZKT.js";
 import {
   SUBAGENT_ALLOWED_TOOLS
-} from "./chunk-KRU4DFRH.js";
+} from "./chunk-CPZ7KG3Y.js";
 import "./chunk-4BKXL7SM.js";
 import "./chunk-MM3F43H6.js";
 import "./chunk-KHYD3WXE.js";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "jinzd-ai-cli",
-  "version": "0.4.180",
+  "version": "0.4.181",
   "description": "Cross-platform REPL-style AI CLI with multi-provider support",
   "type": "module",
   "main": "./dist/index.js",