jinzd-ai-cli 0.4.110 → 0.4.112

package/dist/{batch-V255CIMB.js → batch-QQGRQ45S.js}

@@ -1,9 +1,9 @@
 #!/usr/bin/env node
 import {
   ConfigManager
-} from "./chunk-BLJIPN2F.js";
+} from "./chunk-FM5VYCXA.js";
 import "./chunk-2ZD3YTVM.js";
-import "./chunk-SQPICAHN.js";
+import "./chunk-AWZ63EVH.js";
 import "./chunk-PDX44BCA.js";
 
 // src/cli/batch.ts

package/dist/{chunk-M77QS5QW.js → chunk-5XHNTLRS.js}

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 import {
   TEST_TIMEOUT
-} from "./chunk-SQPICAHN.js";
+} from "./chunk-AWZ63EVH.js";
 
 // src/tools/builtin/run-tests.ts
 import { execSync, spawnSync } from "child_process";

package/dist/{chunk-LKWFPJIP.js → chunk-7XXZWPTN.js}

@@ -5,7 +5,7 @@ import {
 } from "./chunk-3BICTI5M.js";
 import {
   runTestsTool
-} from "./chunk-M77QS5QW.js";
+} from "./chunk-5XHNTLRS.js";
 import {
   EnvLoader,
   NetworkError,
@@ -18,7 +18,7 @@ import {
   SUBAGENT_ALLOWED_TOOLS,
   SUBAGENT_DEFAULT_MAX_ROUNDS,
   SUBAGENT_MAX_ROUNDS_LIMIT
-} from "./chunk-SQPICAHN.js";
+} from "./chunk-AWZ63EVH.js";
 import {
   fileCheckpoints
 } from "./chunk-4BKXL7SM.js";

package/dist/{chunk-SQPICAHN.js → chunk-AWZ63EVH.js}

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 
 // src/core/constants.ts
-var VERSION = "0.4.110";
+var VERSION = "0.4.112";
 var APP_NAME = "ai-cli";
 var CONFIG_DIR_NAME = ".aicli";
 var CONFIG_FILE_NAME = "config.json";

package/dist/{chunk-YIHYQVV7.js → chunk-BWZJGCO6.js}

@@ -2,7 +2,7 @@
 import {
   schemaToJsonSchema,
   truncateForPersist
-} from "./chunk-LKWFPJIP.js";
+} from "./chunk-7XXZWPTN.js";
 import {
   AuthError,
   ProviderError,
@@ -18,7 +18,7 @@ import {
   MCP_PROTOCOL_VERSION,
   MCP_TOOL_PREFIX,
   VERSION
-} from "./chunk-SQPICAHN.js";
+} from "./chunk-AWZ63EVH.js";
 import {
   redactJson
 } from "./chunk-7ZJN4KLV.js";
@@ -1785,7 +1785,16 @@ var PSEUDO_TOOL_CALL_PATTERNS = [
   // ```tool_call\n...\n``` markdown fences (Kimi/Zhipu fallback)
   /```\s*tool_call\b/i,
   // Bare JSON tool-call block: lines starting with `{"name":"...","arguments":`
-  /^\s*\{\s*"name"\s*:\s*"[\w._-]+"\s*,\s*"arguments"\s*:/m
+  /^\s*\{\s*"name"\s*:\s*"[\w._-]+"\s*,\s*"arguments"\s*:/m,
+  // v0.4.112: <think> ... </think> reasoning blocks. The REPL renderer
+  // suppresses these from terminal output, but tee mode writes the raw
+  // delta to disk → reasoning leaks into the saved file. We saw a 600-line
+  // 审计报告.md whose first 57 lines were the model's planning monologue.
+  /<think\b[^>]*>/i,
+  // v0.4.112: leading ```markdown / ```md fence wrapping the entire document.
+  // DeepSeek V4 Pro Thinking sometimes "politely" wraps its document output
+  // in a markdown fence. The fence ends up literally in the saved file.
+  /^\s*```\s*(?:markdown|md|gfm)\b/im
 ];
 function detectPseudoToolCalls(content) {
   if (!content || content.length === 0) return null;
@@ -1794,6 +1803,71 @@ function detectPseudoToolCalls(content) {
   }
   return null;
 }
+function stripPseudoToolCalls(content) {
+  if (!content) return content;
+  let out = content;
+  out = out.replace(/<tool_call\b[^>]*>[\s\S]*?<\/tool_call>/gi, "");
+  out = out.replace(/<tool_call\b[^>]*\/>/gi, "");
+  out = out.replace(/<function_calls\b[^>]*>[\s\S]*?<\/function_calls>/gi, "");
+  out = out.replace(/<invoke\b[^>]*>[\s\S]*?<\/invoke>/gi, "");
+  out = out.replace(/<invoke\b[^>]*\/>/gi, "");
+  out = out.replace(/<tool_use(?:_id)?\b[^>]*>[\s\S]*?<\/tool_use(?:_id)?>/gi, "");
+  out = out.replace(/```\s*tool_call\b[\s\S]*?```/gi, "");
+  out = out.replace(/<think\b[^>]*>[\s\S]*?<\/think>/gi, "");
+  out = out.replace(/<think\b[^>]*>[\s\S]*?(?=^#{1,3}\s+\S|\n\s*\n)/im, "");
+  out = out.replace(/^\s*\{\s*"name"\s*:\s*"[\w._-]+"\s*,\s*"arguments"\s*:[\s\S]*?\}\s*$/gm, "");
+  out = unwrapDocumentFence(out);
+  out = peelMetaNarration(out);
+  out = out.replace(/\n{3,}/g, "\n\n").trim();
+  return out;
+}
+function unwrapDocumentFence(content) {
+  const trimmed = content.trim();
+  const open = trimmed.match(/^```\s*(markdown|md|gfm)?\s*\n/i);
+  if (!open) return content;
+  const afterOpen = trimmed.slice(open[0].length);
+  const closeMatch = afterOpen.match(/\n```\s*$/);
+  if (!closeMatch) return content;
+  const inner = afterOpen.slice(0, afterOpen.length - closeMatch[0].length);
+  if (inner.length < 200) return content;
+  return inner;
+}
+function peelMetaNarration(content) {
+  let out = content;
+  const firstHeadingMatch = out.match(/^#{1,3}\s+\S.*$/m);
+  if (firstHeadingMatch && firstHeadingMatch.index !== void 0) {
+    const before = out.slice(0, firstHeadingMatch.index);
+    const hasIntroMarker = /(?:以下(?:即为|是|就是)|这是|Here\s+is|Below\s+is|完整的?(?:审计报告|内容|文档)|审计报告(?:如下|的完整内容))/i.test(before);
+    if (before.length > 0 && before.length < 800 && hasIntroMarker) {
+      out = out.slice(firstHeadingMatch.index);
+    }
+    if (out.startsWith("---\n")) {
+      const headingAfterRule = out.slice(4).match(/^#{1,3}\s+\S/m);
+      if (headingAfterRule && headingAfterRule.index !== void 0 && headingAfterRule.index < 100) {
+        out = out.slice(4 + headingAfterRule.index);
+      }
+    }
+  }
+  const codaMatch = out.match(/\n[^\n]*?(?:以上(?:即为|就是|内容|为完整的?)|Above\s+is\s+the|本报告已经|该报告(?:已经|包含)|报告(?:已|至此)结束)[^\n]*$/i);
+  if (codaMatch && codaMatch.index !== void 0 && codaMatch.index > out.length / 2) {
+    out = out.slice(0, codaMatch.index);
+  }
+  return out.trim();
+}
+function looksLikeDocumentBody(content) {
+  if (!content || content.length < 200) return false;
+  if (/^#{1,6}\s+\S/m.test(content)) return true;
+  const paragraphs = content.split(/\n\s*\n/).filter((p) => p.trim().length > 30);
+  if (paragraphs.length >= 3) return true;
+  return false;
+}
+function stripToolCallReminder(systemPrompt) {
+  if (!systemPrompt) return systemPrompt;
+  const idx = systemPrompt.indexOf("[\u26A0\uFE0F Mandatory Tool Call Policy]");
+  if (idx === -1) return systemPrompt;
+  return systemPrompt.slice(0, idx).trimEnd();
+}
+var TEE_FINAL_USER_NUDGE = `\u26A0\uFE0F STOP using tools NOW. The save_last_response tee stream is open and capturing every token of THIS response. Output ONLY the requested document body, in markdown. The very first character of your response must be the document's top-level heading (e.g. "# \u5BA1\u8BA1\u62A5\u544A" / "# Audit Report"). Do NOT print <tool_call>, </tool_call>, <function_calls>, <invoke>, <tool_use>, <think>, or any other tool-call markup. Do NOT narrate that you will produce the document \u2014 just produce it. Do NOT pretend to call tools \u2014 there are none in this stream.`;
 var CONTENT_ONLY_STREAM_REMINDER = `
 
 [\u26A0\uFE0F CONTENT GENERATION MODE]
@@ -4081,6 +4155,10 @@ export {
   findPhantomClaims,
   buildPhantomCorrectionMessage,
   detectPseudoToolCalls,
+  stripPseudoToolCalls,
+  looksLikeDocumentBody,
+  stripToolCallReminder,
+  TEE_FINAL_USER_NUDGE,
   CONTENT_ONLY_STREAM_REMINDER,
   ProviderRegistry,
   getContentText,

package/dist/{chunk-GFDEAYPM.js → chunk-DEXCXFLP.js}

@@ -6,7 +6,7 @@ import { platform } from "os";
 import chalk from "chalk";
 
 // src/core/constants.ts
-var VERSION = "0.4.110";
+var VERSION = "0.4.112";
 var APP_NAME = "ai-cli";
 var CONFIG_DIR_NAME = ".aicli";
 var CONFIG_FILE_NAME = "config.json";

package/dist/{chunk-BLJIPN2F.js → chunk-FM5VYCXA.js}

@@ -8,7 +8,7 @@ import {
   CONFIG_FILE_NAME,
   HISTORY_DIR_NAME,
   PLUGINS_DIR_NAME
-} from "./chunk-SQPICAHN.js";
+} from "./chunk-AWZ63EVH.js";
 
 // src/config/config-manager.ts
 import { readFileSync, writeFileSync, existsSync, mkdirSync } from "fs";

package/dist/{constants-EVSUKVTR.js → constants-OCYK6U3N.js}

@@ -36,7 +36,7 @@ import {
   TEST_TIMEOUT,
   VERSION,
   buildUserIdentityPrompt
-} from "./chunk-SQPICAHN.js";
+} from "./chunk-AWZ63EVH.js";
 import "./chunk-PDX44BCA.js";
 export {
   AGENTIC_BEHAVIOR_GUIDELINE,

package/dist/electron-server.js

@@ -36,7 +36,7 @@ import {
   VERSION,
   buildUserIdentityPrompt,
   runTestsTool
-} from "./chunk-GFDEAYPM.js";
+} from "./chunk-DEXCXFLP.js";
 import {
   hasSemanticIndex,
   semanticSearch
@@ -2199,7 +2199,16 @@ var PSEUDO_TOOL_CALL_PATTERNS = [
   // ```tool_call\n...\n``` markdown fences (Kimi/Zhipu fallback)
   /```\s*tool_call\b/i,
   // Bare JSON tool-call block: lines starting with `{"name":"...","arguments":`
-  /^\s*\{\s*"name"\s*:\s*"[\w._-]+"\s*,\s*"arguments"\s*:/m
+  /^\s*\{\s*"name"\s*:\s*"[\w._-]+"\s*,\s*"arguments"\s*:/m,
+  // v0.4.112: <think> ... </think> reasoning blocks. The REPL renderer
+  // suppresses these from terminal output, but tee mode writes the raw
+  // delta to disk → reasoning leaks into the saved file. We saw a 600-line
+  // 审计报告.md whose first 57 lines were the model's planning monologue.
+  /<think\b[^>]*>/i,
+  // v0.4.112: leading ```markdown / ```md fence wrapping the entire document.
+  // DeepSeek V4 Pro Thinking sometimes "politely" wraps its document output
+  // in a markdown fence. The fence ends up literally in the saved file.
+  /^\s*```\s*(?:markdown|md|gfm)\b/im
 ];
 function detectPseudoToolCalls(content) {
   if (!content || content.length === 0) return null;
@@ -2208,6 +2217,71 @@ function detectPseudoToolCalls(content) {
   }
   return null;
 }
+function stripPseudoToolCalls(content) {
+  if (!content) return content;
+  let out = content;
+  out = out.replace(/<tool_call\b[^>]*>[\s\S]*?<\/tool_call>/gi, "");
+  out = out.replace(/<tool_call\b[^>]*\/>/gi, "");
+  out = out.replace(/<function_calls\b[^>]*>[\s\S]*?<\/function_calls>/gi, "");
+  out = out.replace(/<invoke\b[^>]*>[\s\S]*?<\/invoke>/gi, "");
+  out = out.replace(/<invoke\b[^>]*\/>/gi, "");
+  out = out.replace(/<tool_use(?:_id)?\b[^>]*>[\s\S]*?<\/tool_use(?:_id)?>/gi, "");
+  out = out.replace(/```\s*tool_call\b[\s\S]*?```/gi, "");
+  out = out.replace(/<think\b[^>]*>[\s\S]*?<\/think>/gi, "");
+  out = out.replace(/<think\b[^>]*>[\s\S]*?(?=^#{1,3}\s+\S|\n\s*\n)/im, "");
+  out = out.replace(/^\s*\{\s*"name"\s*:\s*"[\w._-]+"\s*,\s*"arguments"\s*:[\s\S]*?\}\s*$/gm, "");
+  out = unwrapDocumentFence(out);
+  out = peelMetaNarration(out);
+  out = out.replace(/\n{3,}/g, "\n\n").trim();
+  return out;
+}
+function unwrapDocumentFence(content) {
+  const trimmed = content.trim();
+  const open = trimmed.match(/^```\s*(markdown|md|gfm)?\s*\n/i);
+  if (!open) return content;
+  const afterOpen = trimmed.slice(open[0].length);
+  const closeMatch = afterOpen.match(/\n```\s*$/);
+  if (!closeMatch) return content;
+  const inner = afterOpen.slice(0, afterOpen.length - closeMatch[0].length);
+  if (inner.length < 200) return content;
+  return inner;
+}
+function peelMetaNarration(content) {
+  let out = content;
+  const firstHeadingMatch = out.match(/^#{1,3}\s+\S.*$/m);
+  if (firstHeadingMatch && firstHeadingMatch.index !== void 0) {
+    const before = out.slice(0, firstHeadingMatch.index);
+    const hasIntroMarker = /(?:以下(?:即为|是|就是)|这是|Here\s+is|Below\s+is|完整的?(?:审计报告|内容|文档)|审计报告(?:如下|的完整内容))/i.test(before);
+    if (before.length > 0 && before.length < 800 && hasIntroMarker) {
+      out = out.slice(firstHeadingMatch.index);
+    }
+    if (out.startsWith("---\n")) {
+      const headingAfterRule = out.slice(4).match(/^#{1,3}\s+\S/m);
+      if (headingAfterRule && headingAfterRule.index !== void 0 && headingAfterRule.index < 100) {
+        out = out.slice(4 + headingAfterRule.index);
+      }
+    }
+  }
+  const codaMatch = out.match(/\n[^\n]*?(?:以上(?:即为|就是|内容|为完整的?)|Above\s+is\s+the|本报告已经|该报告(?:已经|包含)|报告(?:已|至此)结束)[^\n]*$/i);
+  if (codaMatch && codaMatch.index !== void 0 && codaMatch.index > out.length / 2) {
+    out = out.slice(0, codaMatch.index);
+  }
+  return out.trim();
+}
+function looksLikeDocumentBody(content) {
+  if (!content || content.length < 200) return false;
+  if (/^#{1,6}\s+\S/m.test(content)) return true;
+  const paragraphs = content.split(/\n\s*\n/).filter((p) => p.trim().length > 30);
+  if (paragraphs.length >= 3) return true;
+  return false;
+}
+function stripToolCallReminder(systemPrompt) {
+  if (!systemPrompt) return systemPrompt;
+  const idx = systemPrompt.indexOf("[\u26A0\uFE0F Mandatory Tool Call Policy]");
+  if (idx === -1) return systemPrompt;
+  return systemPrompt.slice(0, idx).trimEnd();
+}
+var TEE_FINAL_USER_NUDGE = `\u26A0\uFE0F STOP using tools NOW. The save_last_response tee stream is open and capturing every token of THIS response. Output ONLY the requested document body, in markdown. The very first character of your response must be the document's top-level heading (e.g. "# \u5BA1\u8BA1\u62A5\u544A" / "# Audit Report"). Do NOT print <tool_call>, </tool_call>, <function_calls>, <invoke>, <tool_use>, <think>, or any other tool-call markup. Do NOT narrate that you will produce the document \u2014 just produce it. Do NOT pretend to call tools \u2014 there are none in this stream.`;
 var CONTENT_ONLY_STREAM_REMINDER = `
 
 [\u26A0\uFE0F CONTENT GENERATION MODE]
@@ -10696,7 +10770,8 @@ ${summaryResult.content}`,
     try {
       mkdirSync9(dirname4(saveToFile), { recursive: true });
       fileStream = createWriteStream(saveToFile, { encoding: "utf-8" });
-      const teeSystemPrompt = (systemPrompt ?? "") + CONTENT_ONLY_STREAM_REMINDER;
+      const teeSystemPrompt = stripToolCallReminder(systemPrompt ?? "") + CONTENT_ONLY_STREAM_REMINDER;
+      const teeExtraMessages = extraMessages.length > 0 ? [...extraMessages, { role: "user", content: TEE_FINAL_USER_NUDGE }] : [{ role: "user", content: TEE_FINAL_USER_NUDGE }];
       const chatRequest = {
         messages: apiMessages,
         model: this.currentModel,
@@ -10709,7 +10784,7 @@ ${summaryResult.content}`,
         thinking: modelParams.thinking,
         thinkingBudget: modelParams.thinkingBudget,
         signal: ac.signal,
-        ...extraMessages.length > 0 ? { _extraMessages: extraMessages } : {}
+        _extraMessages: teeExtraMessages
       };
       const stream = provider.chatStream(chatRequest);
       for await (const chunk of stream) {
@@ -10727,12 +10802,22 @@ ${summaryResult.content}`,
       });
       const pseudoMatch = detectPseudoToolCalls(fullContent);
       if (pseudoMatch) {
-        try {
-          unlinkSync4(saveToFile);
-        } catch {
+        const cleaned = stripPseudoToolCalls(fullContent);
+        if (looksLikeDocumentBody(cleaned)) {
+          writeFileSync8(saveToFile, cleaned, "utf-8");
+          fullContent = cleaned;
+          const lines = cleaned.split("\n").length;
+          const bytes = Buffer.byteLength(cleaned, "utf-8");
+          summary = `File saved (with cleanup): ${saveToFile} (${lines} lines, ${bytes} bytes; pseudo-tool-call markup matching ${pseudoMatch} was stripped before save)`;
+          undoStack.push(saveToFile, `save_last_response: ${saveToFile}`);
+        } else {
+          try {
+            unlinkSync4(saveToFile);
+          } catch {
+          }
+          isError = true;
+          summary = `[save_last_response REJECTED] Your output was tool-call markup with no usable document body (matched: ${pseudoMatch}). ${saveToFile} was NOT saved. This fresh stream has NO tools \u2014 output is captured verbatim. STOP emitting <tool_call>, <function_calls>, <invoke>, <think>, or JSON tool blocks. Produce the document body NOW: start with a markdown heading and write the full content.`;
         }
-        isError = true;
-        summary = `[save_last_response REJECTED] Your fresh streaming output contained pseudo-tool-call markup (matched: ${pseudoMatch}). The file at "${saveToFile}" was NOT saved (deleted). This fresh stream has NO tools attached \u2014 output is captured verbatim. Do NOT print <tool_call>, <function_calls>, <invoke>, or {"name":"...","arguments":...} JSON. Try again: produce ONLY the document body.`;
       } else {
         const lines = fullContent.split("\n").length;
         const bytes = Buffer.byteLength(fullContent, "utf-8");
@@ -11833,7 +11918,7 @@ ${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-XVK66TGP.js");
+          const { executeTests } = await import("./run-tests-V73IZCNW.js");
           const argStr = args.join(" ").trim();
           let testArgs = {};
           if (argStr) {

package/dist/{hub-O3EQ2LDY.js → hub-7KYFVLHM.js}

@@ -386,7 +386,7 @@ ${content}`);
   }
 }
 async function runTaskMode(config, providers, configManager, topic) {
-  const { TaskOrchestrator } = await import("./task-orchestrator-2TGWMUUT.js");
+  const { TaskOrchestrator } = await import("./task-orchestrator-RPYZT2WL.js");
   const orchestrator = new TaskOrchestrator(config, providers, configManager);
   let interrupted = false;
   const onSigint = () => {

package/dist/index.js

@@ -7,6 +7,7 @@ import {
   SNAPSHOT_PROMPT,
   SessionManager,
   SkillManager,
+  TEE_FINAL_USER_NUDGE,
   TOOL_CALL_REMINDER,
   autoTrimSessionIfNeeded,
   buildPhantomCorrectionMessage,
@@ -22,15 +23,18 @@ import {
   getPricing,
   hadPreviousWriteToolCalls,
   loadDevState,
+  looksLikeDocumentBody,
   parseSimpleYaml,
   persistToolRound,
   saveDevState,
   sessionHasMeaningfulContent,
-  setupProxy
-} from "./chunk-YIHYQVV7.js";
+  setupProxy,
+  stripPseudoToolCalls,
+  stripToolCallReminder
+} from "./chunk-BWZJGCO6.js";
 import {
   ConfigManager
-} from "./chunk-BLJIPN2F.js";
+} from "./chunk-FM5VYCXA.js";
 import {
   ToolExecutor,
   ToolRegistry,
@@ -49,10 +53,10 @@ import {
   spawnAgentContext,
   theme,
   undoStack
-} from "./chunk-LKWFPJIP.js";
+} from "./chunk-7XXZWPTN.js";
 import "./chunk-3BICTI5M.js";
 import "./chunk-2DXY7UGF.js";
-import "./chunk-M77QS5QW.js";
+import "./chunk-5XHNTLRS.js";
 import "./chunk-2ZD3YTVM.js";
 import {
   AGENTIC_BEHAVIOR_GUIDELINE,
@@ -75,7 +79,7 @@ import {
   SKILLS_DIR_NAME,
   VERSION,
   buildUserIdentityPrompt
-} from "./chunk-SQPICAHN.js";
+} from "./chunk-AWZ63EVH.js";
 import {
   formatGitContextForPrompt,
   getGitContext,
@@ -102,7 +106,7 @@ import { program } from "commander";
 
 // src/repl/repl.ts
 import * as readline from "readline";
-import { existsSync as existsSync5, readFileSync as readFileSync4, readdirSync as readdirSync3, statSync as statSync3, unlinkSync as unlinkSync2 } from "fs";
+import { existsSync as existsSync5, readFileSync as readFileSync4, readdirSync as readdirSync3, statSync as statSync3, unlinkSync as unlinkSync2, writeFileSync as writeFileSync3 } from "fs";
 import { join as join5, resolve as resolve2, extname as extname2, dirname as dirname3, basename as basename2 } from "path";
 import chalk4 from "chalk";
 
@@ -1596,7 +1600,7 @@ ${text}
           const { join: join6 } = await import("path");
           const { existsSync: existsSync6 } = await import("fs");
           const { getGitRoot: getGitRoot2 } = await import("./git-context-7KIP4X2V.js");
-          const { MCP_PROJECT_CONFIG_NAME: MCP_PROJECT_CONFIG_NAME2 } = await import("./constants-EVSUKVTR.js");
+          const { MCP_PROJECT_CONFIG_NAME: MCP_PROJECT_CONFIG_NAME2 } = await import("./constants-OCYK6U3N.js");
           const { approveProject, hashMcpFile } = await import("./project-trust-IFM7FXEV.js");
           const cwd = process.cwd();
           const projectRoot = getGitRoot2(cwd) ?? cwd;
@@ -2646,7 +2650,7 @@ ${hint}` : "")
       usage: "/test [command|filter]",
       async execute(args, ctx) {
         try {
-          const { executeTests } = await import("./run-tests-M73756VV.js");
+          const { executeTests } = await import("./run-tests-JMKGA7XU.js");
           const argStr = args.join(" ").trim();
           let testArgs = {};
           if (argStr) {
@@ -6239,7 +6243,8 @@ ${mcpBudgetNote}` : "");
           } else {
             const teeAc = this.setupStreamInterrupt();
             try {
-              const teeSystemPrompt = (systemPrompt ?? "") + CONTENT_ONLY_STREAM_REMINDER;
+              const teeSystemPrompt = stripToolCallReminder(systemPrompt ?? "") + CONTENT_ONLY_STREAM_REMINDER;
+              const teeExtraMessages = extraMessages.length > 0 ? [...extraMessages, { role: "user", content: TEE_FINAL_USER_NUDGE }] : [{ role: "user", content: TEE_FINAL_USER_NUDGE }];
               const genStream = provider.chatStream({
                 messages: apiMessages,
                 model: effectiveModel,
@@ -6252,7 +6257,7 @@ ${mcpBudgetNote}` : "");
                 thinking: modelParams.thinking,
                 thinkingBudget: modelParams.thinkingBudget,
                 signal: teeAc.signal,
-                ...extraMessages.length > 0 ? { _extraMessages: extraMessages } : {}
+                _extraMessages: teeExtraMessages
               });
               const teeShowTokens = this.shouldShowTokens();
               const { content: genContent, usage: genUsage, tokensShown: teeTokShown } = await this.renderer.renderStream(
@@ -6261,22 +6266,65 @@ ${mcpBudgetNote}` : "");
               );
               const pseudoMatch = detectPseudoToolCalls(genContent);
               if (pseudoMatch) {
+                const cleaned = stripPseudoToolCalls(genContent);
+                if (looksLikeDocumentBody(cleaned)) {
+                  try {
+                    writeFileSync3(saveToFile, cleaned, "utf-8");
+                    process.stdout.write(theme.warning(
+                      `
+  \u26A0 Salvaged save: stripped pseudo-tool-call markup (matched: ${pseudoMatch})
+  ${saveToFile} now contains the cleaned document (${cleaned.length} chars; was ${genContent.length}).
+
+`
+                    ));
+                    lastResponseStore.content = cleaned;
+                    if (genUsage) {
+                      roundUsage.inputTokens += genUsage.inputTokens;
+                      roundUsage.outputTokens += genUsage.outputTokens;
+                      roundUsage.cacheCreationTokens += genUsage.cacheCreationTokens ?? 0;
+                      roundUsage.cacheReadTokens += genUsage.cacheReadTokens ?? 0;
+                    }
+                    session.addMessage({ role: "assistant", content: cleaned, timestamp: /* @__PURE__ */ new Date() });
+                    this.events.emit("message.after", { content: cleaned });
+                    const lines2 = cleaned.split("\n").length;
+                    const bytes2 = Buffer.byteLength(cleaned, "utf-8");
+                    const okResults = result.toolCalls.map((tc) => ({
+                      callId: tc.id,
+                      content: tc.name === "save_last_response" ? `File saved (with cleanup): ${saveToFile} (${lines2} lines, ${bytes2} bytes; pseudo-tool-call markup was stripped before save)` : `[skipped: file already saved by tee streaming]`,
+                      isError: false
+                    }));
+                    const reasoningContent4 = "reasoningContent" in result ? result.reasoningContent : void 0;
+                    const newMsgs4 = provider.buildToolResultMessages(result.toolCalls, okResults, reasoningContent4);
+                    extraMessages.push(...newMsgs4);
+                    if (roundUsage.inputTokens > 0 || roundUsage.outputTokens > 0) {
+                      this.addSessionUsage(roundUsage, effectiveModel);
+                      session.addTokenUsage(roundUsage);
+                      if (teeShowTokens && !teeTokShown) {
+                        this.renderer.renderUsage(roundUsage, this.sessionTokenUsage);
+                      }
+                    }
+                    return;
+                  } catch (writeErr) {
+                    process.stderr.write(`[tee] salvage write failed: ${writeErr.message ?? writeErr}
+`);
+                  }
+                }
                 try {
                   unlinkSync2(saveToFile);
                 } catch {
                 }
                 process.stdout.write(theme.error(
                   `
-  \u2717 Rejected save: response contained pseudo-tool-call markup (matched: ${pseudoMatch})
-  ${saveToFile} was deleted; the model will be asked to regenerate without tool markup.
+  \u2717 Rejected save: response was pseudo-tool-call markup with no usable document body (matched: ${pseudoMatch})
+  ${saveToFile} was deleted; asking model to retry.
 
 `
                 ));
                 const errorResults = result.toolCalls.map((tc) => ({
                   callId: tc.id,
-                  content: tc.name === "save_last_response" ? `[save_last_response REJECTED] Your fresh streaming output contained pseudo-tool-call markup (e.g. <tool_call>, <function_calls>, or {"name":"...","arguments":...}). The file at "${saveToFile}" was NOT saved (it was deleted to prevent garbage on disk).
+                  content: tc.name === "save_last_response" ? `[save_last_response REJECTED] Your output was tool-call XML/JSON with no document body. ${saveToFile} was NOT saved.
 
-This fresh stream has NO tools attached \u2014 your output is captured verbatim. Do NOT print tool-call XML/JSON. Try again: produce ONLY the document body (markdown is fine; tool markup is NOT).` : `[skipped: save_last_response was rejected and other parallel calls are abandoned]`,
+This fresh stream has NO tools \u2014 output is captured verbatim. STOP emitting <tool_call>, <function_calls>, <invoke>, <think>, or JSON tool blocks. Produce the document body NOW: start with a markdown heading like "# \u5BA1\u8BA1\u62A5\u544A" and write the full report.` : `[skipped: save_last_response was rejected and other parallel calls are abandoned]`,
                   isError: tc.name === "save_last_response"
                 }));
                 const reasoningContent3 = "reasoningContent" in result ? result.reasoningContent : void 0;
@@ -6834,7 +6882,7 @@ 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-YRX5VO5N.js");
+  const { startWebServer } = await import("./server-YY6CUP3K.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) => {
@@ -6957,7 +7005,7 @@ program.command("sessions").description("List recent conversation sessions").act
 });
 program.command("batch <action> [arg] [arg2]").description("Anthropic Message Batches: submit | list | status <id> | results <id> [out] | cancel <id>").option("--dry-run", "Parse and validate input without submitting (submit only)").action(async (action, arg, arg2, options) => {
   try {
-    const batch = await import("./batch-V255CIMB.js");
+    const batch = await import("./batch-QQGRQ45S.js");
     switch (action) {
       case "submit":
         if (!arg) {
@@ -7000,7 +7048,7 @@ program.command("batch <action> [arg] [arg2]").description("Anthropic Message Ba
   }
 });
 program.command("mcp-serve").description("Start an MCP server over STDIO, exposing aicli's built-in tools to Claude Desktop / Cursor / other MCP clients").option("--allow-destructive", "Allow bash / run_interactive / task_create (always destructive in MCP mode)").option("--allow-outside-cwd", "Allow tool path arguments to escape the sandbox root \u2014 disabled by default").option("--tools <list>", "Comma-separated whitelist of tools to expose (default: all eligible tools)").option("--cwd <path>", "Working directory AND sandbox root (default: current directory)").action(async (options) => {
-  const { startMcpServer } = await import("./server-54OXGXJ6.js");
+  const { startMcpServer } = await import("./server-WWFOPWWJ.js");
   await startMcpServer({
     allowDestructive: !!options.allowDestructive,
     allowOutsideCwd: !!options.allowOutsideCwd,
@@ -7127,7 +7175,7 @@ program.command("hub [topic]").description("Start multi-agent hub (discuss / bra
     }),
     config.get("customProviders")
   );
-  const { startHub } = await import("./hub-O3EQ2LDY.js");
+  const { startHub } = await import("./hub-7KYFVLHM.js");
   await startHub(
     {
       topic: topic ?? "",

package/dist/{run-tests-M73756VV.js → run-tests-JMKGA7XU.js}

@@ -2,8 +2,8 @@
 import {
   executeTests,
   runTestsTool
-} from "./chunk-M77QS5QW.js";
-import "./chunk-SQPICAHN.js";
+} from "./chunk-5XHNTLRS.js";
+import "./chunk-AWZ63EVH.js";
 import "./chunk-PDX44BCA.js";
 export {
   executeTests,

package/dist/{run-tests-XVK66TGP.js → run-tests-V73IZCNW.js}

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

package/dist/{server-54OXGXJ6.js → server-WWFOPWWJ.js}

@@ -3,14 +3,14 @@ import {
   ToolRegistry,
   getDangerLevel,
   schemaToJsonSchema
-} from "./chunk-LKWFPJIP.js";
+} from "./chunk-7XXZWPTN.js";
 import "./chunk-3BICTI5M.js";
 import "./chunk-2DXY7UGF.js";
-import "./chunk-M77QS5QW.js";
+import "./chunk-5XHNTLRS.js";
 import "./chunk-2ZD3YTVM.js";
 import {
   VERSION
-} from "./chunk-SQPICAHN.js";
+} from "./chunk-AWZ63EVH.js";
 import "./chunk-4BKXL7SM.js";
 import "./chunk-7ZJN4KLV.js";
 import "./chunk-KHYD3WXE.js";

package/dist/{server-YRX5VO5N.js → server-YY6CUP3K.js}

@@ -9,6 +9,7 @@ import {
   ProviderRegistry,
   SessionManager,
   SkillManager,
+  TEE_FINAL_USER_NUDGE,
   TOOL_CALL_REMINDER,
   autoTrimSessionIfNeeded,
   computeCost,
@@ -18,12 +19,15 @@ import {
   getContentText,
   hadPreviousWriteToolCalls,
   loadDevState,
+  looksLikeDocumentBody,
   persistToolRound,
-  setupProxy
-} from "./chunk-YIHYQVV7.js";
+  setupProxy,
+  stripPseudoToolCalls,
+  stripToolCallReminder
+} from "./chunk-BWZJGCO6.js";
 import {
   ConfigManager
-} from "./chunk-BLJIPN2F.js";
+} from "./chunk-FM5VYCXA.js";
 import {
   ToolExecutor,
   ToolRegistry,
@@ -41,10 +45,10 @@ import {
   spawnAgentContext,
   truncateOutput,
   undoStack
-} from "./chunk-LKWFPJIP.js";
+} from "./chunk-7XXZWPTN.js";
 import "./chunk-3BICTI5M.js";
 import "./chunk-2DXY7UGF.js";
-import "./chunk-M77QS5QW.js";
+import "./chunk-5XHNTLRS.js";
 import "./chunk-2ZD3YTVM.js";
 import {
   AGENTIC_BEHAVIOR_GUIDELINE,
@@ -64,7 +68,7 @@ import {
   SKILLS_DIR_NAME,
   VERSION,
   buildUserIdentityPrompt
-} from "./chunk-SQPICAHN.js";
+} from "./chunk-AWZ63EVH.js";
 import {
   formatGitContextForPrompt,
   getGitContext,
@@ -1259,7 +1263,8 @@ ${summaryResult.content}`,
     try {
       mkdirSync(dirname(saveToFile), { recursive: true });
       fileStream = createWriteStream(saveToFile, { encoding: "utf-8" });
-      const teeSystemPrompt = (systemPrompt ?? "") + CONTENT_ONLY_STREAM_REMINDER;
+      const teeSystemPrompt = stripToolCallReminder(systemPrompt ?? "") + CONTENT_ONLY_STREAM_REMINDER;
+      const teeExtraMessages = extraMessages.length > 0 ? [...extraMessages, { role: "user", content: TEE_FINAL_USER_NUDGE }] : [{ role: "user", content: TEE_FINAL_USER_NUDGE }];
       const chatRequest = {
         messages: apiMessages,
         model: this.currentModel,
@@ -1272,7 +1277,7 @@ ${summaryResult.content}`,
         thinking: modelParams.thinking,
         thinkingBudget: modelParams.thinkingBudget,
         signal: ac.signal,
-        ...extraMessages.length > 0 ? { _extraMessages: extraMessages } : {}
+        _extraMessages: teeExtraMessages
       };
       const stream = provider.chatStream(chatRequest);
       for await (const chunk of stream) {
@@ -1290,12 +1295,22 @@ ${summaryResult.content}`,
       });
       const pseudoMatch = detectPseudoToolCalls(fullContent);
       if (pseudoMatch) {
-        try {
-          unlinkSync(saveToFile);
-        } catch {
+        const cleaned = stripPseudoToolCalls(fullContent);
+        if (looksLikeDocumentBody(cleaned)) {
+          writeFileSync(saveToFile, cleaned, "utf-8");
+          fullContent = cleaned;
+          const lines = cleaned.split("\n").length;
+          const bytes = Buffer.byteLength(cleaned, "utf-8");
+          summary = `File saved (with cleanup): ${saveToFile} (${lines} lines, ${bytes} bytes; pseudo-tool-call markup matching ${pseudoMatch} was stripped before save)`;
+          undoStack.push(saveToFile, `save_last_response: ${saveToFile}`);
+        } else {
+          try {
+            unlinkSync(saveToFile);
+          } catch {
+          }
+          isError = true;
+          summary = `[save_last_response REJECTED] Your output was tool-call markup with no usable document body (matched: ${pseudoMatch}). ${saveToFile} was NOT saved. This fresh stream has NO tools \u2014 output is captured verbatim. STOP emitting <tool_call>, <function_calls>, <invoke>, <think>, or JSON tool blocks. Produce the document body NOW: start with a markdown heading and write the full content.`;
         }
-        isError = true;
-        summary = `[save_last_response REJECTED] Your fresh streaming output contained pseudo-tool-call markup (matched: ${pseudoMatch}). The file at "${saveToFile}" was NOT saved (deleted). This fresh stream has NO tools attached \u2014 output is captured verbatim. Do NOT print <tool_call>, <function_calls>, <invoke>, or {"name":"...","arguments":...} JSON. Try again: produce ONLY the document body.`;
       } else {
         const lines = fullContent.split("\n").length;
         const bytes = Buffer.byteLength(fullContent, "utf-8");
@@ -2396,7 +2411,7 @@ ${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-M73756VV.js");
+          const { executeTests } = await import("./run-tests-JMKGA7XU.js");
           const argStr = args.join(" ").trim();
           let testArgs = {};
           if (argStr) {

package/dist/{task-orchestrator-2TGWMUUT.js → task-orchestrator-RPYZT2WL.js}

@@ -4,14 +4,14 @@ import {
   getDangerLevel,
   googleSearchContext,
   truncateOutput
-} from "./chunk-LKWFPJIP.js";
+} from "./chunk-7XXZWPTN.js";
 import "./chunk-3BICTI5M.js";
 import "./chunk-2DXY7UGF.js";
-import "./chunk-M77QS5QW.js";
+import "./chunk-5XHNTLRS.js";
 import "./chunk-2ZD3YTVM.js";
 import {
   SUBAGENT_ALLOWED_TOOLS
-} from "./chunk-SQPICAHN.js";
+} from "./chunk-AWZ63EVH.js";
 import "./chunk-4BKXL7SM.js";
 import "./chunk-7ZJN4KLV.js";
 import "./chunk-KHYD3WXE.js";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "jinzd-ai-cli",
-  "version": "0.4.110",
+  "version": "0.4.112",
   "description": "Cross-platform REPL-style AI CLI with multi-provider support",
   "type": "module",
   "main": "./dist/index.js",