jinzd-ai-cli 0.4.109 → 0.4.111

package/dist/{batch-BFOFTLFS.js → batch-XIDQVPRL.js}

@@ -1,9 +1,9 @@
 #!/usr/bin/env node
 import {
   ConfigManager
-} from "./chunk-JLHY6AOZ.js";
+} from "./chunk-VYDXL2PC.js";
 import "./chunk-2ZD3YTVM.js";
-import "./chunk-LZR3DYW3.js";
+import "./chunk-CPIQXP7Q.js";
 import "./chunk-PDX44BCA.js";
 
 // src/cli/batch.ts

package/dist/{chunk-N62X6OKJ.js → chunk-5EVKTBUU.js}

@@ -5,7 +5,7 @@ import {
 } from "./chunk-3BICTI5M.js";
 import {
   runTestsTool
-} from "./chunk-CUQ4HXHA.js";
+} from "./chunk-5LKW2GOF.js";
 import {
   EnvLoader,
   NetworkError,
@@ -18,7 +18,7 @@ import {
   SUBAGENT_ALLOWED_TOOLS,
   SUBAGENT_DEFAULT_MAX_ROUNDS,
   SUBAGENT_MAX_ROUNDS_LIMIT
-} from "./chunk-LZR3DYW3.js";
+} from "./chunk-CPIQXP7Q.js";
 import {
   fileCheckpoints
 } from "./chunk-4BKXL7SM.js";

package/dist/{chunk-CUQ4HXHA.js → chunk-5LKW2GOF.js}

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 import {
   TEST_TIMEOUT
-} from "./chunk-LZR3DYW3.js";
+} from "./chunk-CPIQXP7Q.js";
 
 // src/tools/builtin/run-tests.ts
 import { execSync, spawnSync } from "child_process";

package/dist/{chunk-OKRTYLB6.js → chunk-AVMGKU4E.js}

@@ -6,7 +6,7 @@ import { platform } from "os";
 import chalk from "chalk";
 
 // src/core/constants.ts
-var VERSION = "0.4.109";
+var VERSION = "0.4.111";
 var APP_NAME = "ai-cli";
 var CONFIG_DIR_NAME = ".aicli";
 var CONFIG_FILE_NAME = "config.json";

package/dist/{chunk-LZR3DYW3.js → chunk-CPIQXP7Q.js}

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 
 // src/core/constants.ts
-var VERSION = "0.4.109";
+var VERSION = "0.4.111";
 var APP_NAME = "ai-cli";
 var CONFIG_DIR_NAME = ".aicli";
 var CONFIG_FILE_NAME = "config.json";

package/dist/{chunk-3UJIJBKB.js → chunk-GRIEVPII.js}

@@ -2,7 +2,7 @@
 import {
   schemaToJsonSchema,
   truncateForPersist
-} from "./chunk-N62X6OKJ.js";
+} from "./chunk-5EVKTBUU.js";
 import {
   AuthError,
   ProviderError,
@@ -18,7 +18,7 @@ import {
   MCP_PROTOCOL_VERSION,
   MCP_TOOL_PREFIX,
   VERSION
-} from "./chunk-LZR3DYW3.js";
+} from "./chunk-CPIQXP7Q.js";
 import {
   redactJson
 } from "./chunk-7ZJN4KLV.js";
@@ -1772,6 +1772,68 @@ function buildPhantomCorrectionMessage(phantoms) {
   const list = phantoms.map((p) => `  - ${p}`).join("\n");
   return "You claimed to have written the following file(s), but no matching write_file tool call was actually made in this turn:\n" + list + '\n\nEach of these files does NOT exist on disk. You MUST now invoke write_file (via the function calling API) for every missing file listed above. Do NOT output another "completion summary" until the tool calls have actually been made.';
 }
+var PSEUDO_TOOL_CALL_PATTERNS = [
+  // <tool_call name="..."> ... </tool_call>  (DeepSeek V4 thinking, GLM)
+  /<tool_call\s+name\s*=\s*["'][\w._-]+["']/,
+  // <function_calls> ... </function_calls>  (Anthropic-style as text)
+  /<\/?function_calls\s*>/,
+  // <invoke name="..." />  (Anthropic XML tool-call, which is real for
+  // Claude API but is text/garbage for any other provider's plain stream)
+  /<invoke\s+name\s*=\s*["'][\w._-]+["']/,
+  // <tool_use> ... <tool_use_id>  (Claude flavor leaked into text)
+  /<tool_use(?:_id)?\b/,
+  // ```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
+];
+function detectPseudoToolCalls(content) {
+  if (!content || content.length === 0) return null;
+  for (const re of PSEUDO_TOOL_CALL_PATTERNS) {
+    if (re.test(content)) return re.source;
+  }
+  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(/^\s*\{\s*"name"\s*:\s*"[\w._-]+"\s*,\s*"arguments"\s*:[\s\S]*?\}\s*$/gm, "");
+  out = out.replace(/\n{3,}/g, "\n\n").trim();
+  return out;
+}
+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]
+You are now in a CONTENT-ONLY streaming pass. The file at the configured path will receive every token of THIS response.
+- Do NOT emit <tool_call>, </tool_call>, <function_calls>, <invoke>, <tool_use>, or any tool-call XML/JSON markup.
+- Do NOT print "I will now call ...", "let me read ...", "<think>" reasoning blocks (the surrounding REPL handles those separately \u2014 they should not enter the saved file).
+- Do NOT pretend to call tools. There are NO tools available in this stream \u2014 only your text output is captured.
+- Produce ONLY the requested document body. Markdown is fine. Code blocks are fine. Tool-call markup is NOT.
+- If you accidentally start a <tool_call>, STOP and produce the document body instead.
+
+The file is closed and named when this stream ends. If your output contains pseudo-tool-call markup, the save will be REJECTED and you will be asked to retry.`;
 
 // src/providers/kimi.ts
 var KIMI_XML_REMINDER = `
@@ -4047,6 +4109,12 @@ export {
   extractWrittenFilePaths,
   findPhantomClaims,
   buildPhantomCorrectionMessage,
+  detectPseudoToolCalls,
+  stripPseudoToolCalls,
+  looksLikeDocumentBody,
+  stripToolCallReminder,
+  TEE_FINAL_USER_NUDGE,
+  CONTENT_ONLY_STREAM_REMINDER,
   ProviderRegistry,
   getContentText,
   SessionManager,

package/dist/{chunk-JLHY6AOZ.js → chunk-VYDXL2PC.js}

@@ -8,7 +8,7 @@ import {
   CONFIG_FILE_NAME,
   HISTORY_DIR_NAME,
   PLUGINS_DIR_NAME
-} from "./chunk-LZR3DYW3.js";
+} from "./chunk-CPIQXP7Q.js";
 
 // src/config/config-manager.ts
 import { readFileSync, writeFileSync, existsSync, mkdirSync } from "fs";

package/dist/{constants-K52FVIX5.js → constants-NPGSV4QY.js}

@@ -36,7 +36,7 @@ import {
   TEST_TIMEOUT,
   VERSION,
   buildUserIdentityPrompt
-} from "./chunk-LZR3DYW3.js";
+} from "./chunk-CPIQXP7Q.js";
 import "./chunk-PDX44BCA.js";
 export {
   AGENTIC_BEHAVIOR_GUIDELINE,

package/dist/electron-server.js

@@ -36,7 +36,7 @@ import {
   VERSION,
   buildUserIdentityPrompt,
   runTestsTool
-} from "./chunk-OKRTYLB6.js";
+} from "./chunk-AVMGKU4E.js";
 import {
   hasSemanticIndex,
   semanticSearch
@@ -2186,6 +2186,68 @@ CRITICAL \u2014 Batch file generation rules:
 4. Only produce a text summary AFTER all write_file calls have been made and returned success.
 5. The system compares every "file saved" claim against actual tool calls. Phantom claims trigger an automatic retry \u2014 do not waste rounds.`;
 var HALLUCINATION_CORRECTION_MESSAGE = "You did NOT actually call the write_file tool \u2014 the file was NOT created! Please immediately use the write_file tool via the function calling API to perform the actual file write. Do NOT describe file content in text \u2014 you MUST invoke write_file through the tool_calls mechanism.";
+var PSEUDO_TOOL_CALL_PATTERNS = [
+  // <tool_call name="..."> ... </tool_call>  (DeepSeek V4 thinking, GLM)
+  /<tool_call\s+name\s*=\s*["'][\w._-]+["']/,
+  // <function_calls> ... </function_calls>  (Anthropic-style as text)
+  /<\/?function_calls\s*>/,
+  // <invoke name="..." />  (Anthropic XML tool-call, which is real for
+  // Claude API but is text/garbage for any other provider's plain stream)
+  /<invoke\s+name\s*=\s*["'][\w._-]+["']/,
+  // <tool_use> ... <tool_use_id>  (Claude flavor leaked into text)
+  /<tool_use(?:_id)?\b/,
+  // ```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
+];
+function detectPseudoToolCalls(content) {
+  if (!content || content.length === 0) return null;
+  for (const re of PSEUDO_TOOL_CALL_PATTERNS) {
+    if (re.test(content)) return re.source;
+  }
+  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(/^\s*\{\s*"name"\s*:\s*"[\w._-]+"\s*,\s*"arguments"\s*:[\s\S]*?\}\s*$/gm, "");
+  out = out.replace(/\n{3,}/g, "\n\n").trim();
+  return out;
+}
+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]
+You are now in a CONTENT-ONLY streaming pass. The file at the configured path will receive every token of THIS response.
+- Do NOT emit <tool_call>, </tool_call>, <function_calls>, <invoke>, <tool_use>, or any tool-call XML/JSON markup.
+- Do NOT print "I will now call ...", "let me read ...", "<think>" reasoning blocks (the surrounding REPL handles those separately \u2014 they should not enter the saved file).
+- Do NOT pretend to call tools. There are NO tools available in this stream \u2014 only your text output is captured.
+- Produce ONLY the requested document body. Markdown is fine. Code blocks are fine. Tool-call markup is NOT.
+- If you accidentally start a <tool_call>, STOP and produce the document body instead.
+
+The file is closed and named when this stream ends. If your output contains pseudo-tool-call markup, the save will be REJECTED and you will be asked to retry.`;
 
 // src/providers/kimi.ts
 var KIMI_XML_REMINDER = `
@@ -9787,7 +9849,7 @@ function autoTrimSessionIfNeeded(session, sizeLimit = SESSION_SIZE_LIMIT) {
 }
 
 // src/web/session-handler.ts
-import { existsSync as existsSync20, readFileSync as readFileSync13, appendFileSync as appendFileSync3, writeFileSync as writeFileSync8, mkdirSync as mkdirSync9, readdirSync as readdirSync9, statSync as statSync8, createWriteStream } from "fs";
+import { existsSync as existsSync20, readFileSync as readFileSync13, appendFileSync as appendFileSync3, writeFileSync as writeFileSync8, mkdirSync as mkdirSync9, readdirSync as readdirSync9, statSync as statSync8, createWriteStream, unlinkSync as unlinkSync4 } from "fs";
 import { join as join13, resolve as resolve5, dirname as dirname4 } from "path";
 import { execSync as execSync3 } from "child_process";
 
@@ -10663,10 +10725,12 @@ ${summaryResult.content}`,
     try {
       mkdirSync9(dirname4(saveToFile), { recursive: true });
       fileStream = createWriteStream(saveToFile, { encoding: "utf-8" });
+      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,
-        systemPrompt,
+        systemPrompt: teeSystemPrompt,
         systemPromptVolatile,
         stream: true,
         temperature: modelParams.temperature,
@@ -10675,7 +10739,7 @@ ${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) {
@@ -10691,10 +10755,30 @@ ${summaryResult.content}`,
       await new Promise((resolve7, reject) => {
         fileStream.end((err) => err ? reject(err) : resolve7());
       });
-      const lines = fullContent.split("\n").length;
-      const bytes = Buffer.byteLength(fullContent, "utf-8");
-      summary = `File saved: ${saveToFile} (${lines} lines, ${bytes} bytes)`;
-      undoStack.push(saveToFile, `save_last_response: ${saveToFile}`);
+      const pseudoMatch = detectPseudoToolCalls(fullContent);
+      if (pseudoMatch) {
+        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.`;
+        }
+      } else {
+        const lines = fullContent.split("\n").length;
+        const bytes = Buffer.byteLength(fullContent, "utf-8");
+        summary = `File saved: ${saveToFile} (${lines} lines, ${bytes} bytes)`;
+        undoStack.push(saveToFile, `save_last_response: ${saveToFile}`);
+      }
       if (teeUsage) {
         roundUsage.inputTokens += teeUsage.inputTokens;
         roundUsage.outputTokens += teeUsage.outputTokens;
@@ -11789,7 +11873,7 @@ ${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-TTP4QJVN.js");
+          const { executeTests } = await import("./run-tests-TTU6DNLI.js");
           const argStr = args.join(" ").trim();
           let testArgs = {};
           if (argStr) {

package/dist/{hub-4XJ5ZTM4.js → hub-T34FCE5J.js}

@@ -386,7 +386,7 @@ ${content}`);
   }
 }
 async function runTaskMode(config, providers, configManager, topic) {
-  const { TaskOrchestrator } = await import("./task-orchestrator-XSNTJYNR.js");
+  const { TaskOrchestrator } = await import("./task-orchestrator-IJFF7XLQ.js");
   const orchestrator = new TaskOrchestrator(config, providers, configManager);
   let interrupted = false;
   const onSigint = () => {

package/dist/index.js

@@ -1,17 +1,20 @@
 #!/usr/bin/env node
 import {
+  CONTENT_ONLY_STREAM_REMINDER,
   HALLUCINATION_CORRECTION_MESSAGE,
   McpManager,
   ProviderRegistry,
   SNAPSHOT_PROMPT,
   SessionManager,
   SkillManager,
+  TEE_FINAL_USER_NUDGE,
   TOOL_CALL_REMINDER,
   autoTrimSessionIfNeeded,
   buildPhantomCorrectionMessage,
   buildWriteRoundReminder,
   clearDevState,
   computeCost,
+  detectPseudoToolCalls,
   detectsHallucinatedFileOp,
   extractWrittenFilePaths,
   findPhantomClaims,
@@ -20,15 +23,18 @@ import {
   getPricing,
   hadPreviousWriteToolCalls,
   loadDevState,
+  looksLikeDocumentBody,
   parseSimpleYaml,
   persistToolRound,
   saveDevState,
   sessionHasMeaningfulContent,
-  setupProxy
-} from "./chunk-3UJIJBKB.js";
+  setupProxy,
+  stripPseudoToolCalls,
+  stripToolCallReminder
+} from "./chunk-GRIEVPII.js";
 import {
   ConfigManager
-} from "./chunk-JLHY6AOZ.js";
+} from "./chunk-VYDXL2PC.js";
 import {
   ToolExecutor,
   ToolRegistry,
@@ -47,10 +53,10 @@ import {
   spawnAgentContext,
   theme,
   undoStack
-} from "./chunk-N62X6OKJ.js";
+} from "./chunk-5EVKTBUU.js";
 import "./chunk-3BICTI5M.js";
 import "./chunk-2DXY7UGF.js";
-import "./chunk-CUQ4HXHA.js";
+import "./chunk-5LKW2GOF.js";
 import "./chunk-2ZD3YTVM.js";
 import {
   AGENTIC_BEHAVIOR_GUIDELINE,
@@ -73,7 +79,7 @@ import {
   SKILLS_DIR_NAME,
   VERSION,
   buildUserIdentityPrompt
-} from "./chunk-LZR3DYW3.js";
+} from "./chunk-CPIQXP7Q.js";
 import {
   formatGitContextForPrompt,
   getGitContext,
@@ -100,7 +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 } 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";
 
@@ -1594,7 +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-K52FVIX5.js");
+          const { MCP_PROJECT_CONFIG_NAME: MCP_PROJECT_CONFIG_NAME2 } = await import("./constants-NPGSV4QY.js");
           const { approveProject, hashMcpFile } = await import("./project-trust-IFM7FXEV.js");
           const cwd = process.cwd();
           const projectRoot = getGitRoot2(cwd) ?? cwd;
@@ -2644,7 +2650,7 @@ ${hint}` : "")
       usage: "/test [command|filter]",
       async execute(args, ctx) {
         try {
-          const { executeTests } = await import("./run-tests-ZBDJYOB2.js");
+          const { executeTests } = await import("./run-tests-QBNLSYCO.js");
           const argStr = args.join(" ").trim();
           let testArgs = {};
           if (argStr) {
@@ -6237,10 +6243,12 @@ ${mcpBudgetNote}` : "");
           } else {
             const teeAc = this.setupStreamInterrupt();
             try {
+              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,
-                systemPrompt,
+                systemPrompt: teeSystemPrompt,
                 systemPromptVolatile,
                 stream: true,
                 temperature: modelParams.temperature,
@@ -6249,13 +6257,87 @@ ${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(
                 genStream,
                 { saveToFile, showTokens: teeShowTokens, sessionTotal: teeShowTokens ? { ...this.sessionTokenUsage } : void 0, signal: teeAc.signal }
               );
+              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 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 output was tool-call XML/JSON with no document body. ${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 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;
+                const newMsgs3 = provider.buildToolResultMessages(result.toolCalls, errorResults, reasoningContent3);
+                extraMessages.push(...newMsgs3);
+                if (genUsage) {
+                  roundUsage.inputTokens += genUsage.inputTokens;
+                  roundUsage.outputTokens += genUsage.outputTokens;
+                  roundUsage.cacheCreationTokens += genUsage.cacheCreationTokens ?? 0;
+                  roundUsage.cacheReadTokens += genUsage.cacheReadTokens ?? 0;
+                }
+                continue;
+              }
               lastResponseStore.content = genContent;
               if (genUsage) {
                 roundUsage.inputTokens += genUsage.inputTokens;
@@ -6800,7 +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-VDMQC43J.js");
+  const { startWebServer } = await import("./server-42ZKNS56.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) => {
@@ -6923,7 +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-BFOFTLFS.js");
+    const batch = await import("./batch-XIDQVPRL.js");
     switch (action) {
       case "submit":
         if (!arg) {
@@ -6966,7 +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-LEPDI3OL.js");
+  const { startMcpServer } = await import("./server-5TGJOWML.js");
   await startMcpServer({
     allowDestructive: !!options.allowDestructive,
     allowOutsideCwd: !!options.allowOutsideCwd,
@@ -7093,7 +7175,7 @@ program.command("hub [topic]").description("Start multi-agent hub (discuss / bra
     }),
     config.get("customProviders")
   );
-  const { startHub } = await import("./hub-4XJ5ZTM4.js");
+  const { startHub } = await import("./hub-T34FCE5J.js");
   await startHub(
     {
       topic: topic ?? "",

package/dist/{run-tests-ZBDJYOB2.js → run-tests-QBNLSYCO.js}

@@ -2,8 +2,8 @@
 import {
   executeTests,
   runTestsTool
-} from "./chunk-CUQ4HXHA.js";
-import "./chunk-LZR3DYW3.js";
+} from "./chunk-5LKW2GOF.js";
+import "./chunk-CPIQXP7Q.js";
 import "./chunk-PDX44BCA.js";
 export {
   executeTests,

package/dist/{run-tests-TTP4QJVN.js → run-tests-TTU6DNLI.js}

@@ -1,7 +1,7 @@
 import {
   executeTests,
   runTestsTool
-} from "./chunk-OKRTYLB6.js";
+} from "./chunk-AVMGKU4E.js";
 import "./chunk-3RG5ZIWI.js";
 export {
   executeTests,

package/dist/{server-VDMQC43J.js → server-42ZKNS56.js}

@@ -3,25 +3,31 @@ import {
   AuthManager
 } from "./chunk-BYNY5JPB.js";
 import {
+  CONTENT_ONLY_STREAM_REMINDER,
   HALLUCINATION_CORRECTION_MESSAGE,
   McpManager,
   ProviderRegistry,
   SessionManager,
   SkillManager,
+  TEE_FINAL_USER_NUDGE,
   TOOL_CALL_REMINDER,
   autoTrimSessionIfNeeded,
   computeCost,
+  detectPseudoToolCalls,
   detectsHallucinatedFileOp,
   formatCost,
   getContentText,
   hadPreviousWriteToolCalls,
   loadDevState,
+  looksLikeDocumentBody,
   persistToolRound,
-  setupProxy
-} from "./chunk-3UJIJBKB.js";
+  setupProxy,
+  stripPseudoToolCalls,
+  stripToolCallReminder
+} from "./chunk-GRIEVPII.js";
 import {
   ConfigManager
-} from "./chunk-JLHY6AOZ.js";
+} from "./chunk-VYDXL2PC.js";
 import {
   ToolExecutor,
   ToolRegistry,
@@ -39,10 +45,10 @@ import {
   spawnAgentContext,
   truncateOutput,
   undoStack
-} from "./chunk-N62X6OKJ.js";
+} from "./chunk-5EVKTBUU.js";
 import "./chunk-3BICTI5M.js";
 import "./chunk-2DXY7UGF.js";
-import "./chunk-CUQ4HXHA.js";
+import "./chunk-5LKW2GOF.js";
 import "./chunk-2ZD3YTVM.js";
 import {
   AGENTIC_BEHAVIOR_GUIDELINE,
@@ -62,7 +68,7 @@ import {
   SKILLS_DIR_NAME,
   VERSION,
   buildUserIdentityPrompt
-} from "./chunk-LZR3DYW3.js";
+} from "./chunk-CPIQXP7Q.js";
 import {
   formatGitContextForPrompt,
   getGitContext,
@@ -473,7 +479,7 @@ function loadMemoryContent(configDir) {
 }
 
 // src/web/session-handler.ts
-import { existsSync as existsSync3, readFileSync as readFileSync3, appendFileSync, writeFileSync, mkdirSync, readdirSync, statSync, createWriteStream } from "fs";
+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"]);
@@ -1257,10 +1263,12 @@ ${summaryResult.content}`,
     try {
       mkdirSync(dirname(saveToFile), { recursive: true });
       fileStream = createWriteStream(saveToFile, { encoding: "utf-8" });
+      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,
-        systemPrompt,
+        systemPrompt: teeSystemPrompt,
         systemPromptVolatile,
         stream: true,
         temperature: modelParams.temperature,
@@ -1269,7 +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) {
@@ -1285,10 +1293,30 @@ ${summaryResult.content}`,
       await new Promise((resolve3, reject) => {
         fileStream.end((err) => err ? reject(err) : resolve3());
       });
-      const lines = fullContent.split("\n").length;
-      const bytes = Buffer.byteLength(fullContent, "utf-8");
-      summary = `File saved: ${saveToFile} (${lines} lines, ${bytes} bytes)`;
-      undoStack.push(saveToFile, `save_last_response: ${saveToFile}`);
+      const pseudoMatch = detectPseudoToolCalls(fullContent);
+      if (pseudoMatch) {
+        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.`;
+        }
+      } else {
+        const lines = fullContent.split("\n").length;
+        const bytes = Buffer.byteLength(fullContent, "utf-8");
+        summary = `File saved: ${saveToFile} (${lines} lines, ${bytes} bytes)`;
+        undoStack.push(saveToFile, `save_last_response: ${saveToFile}`);
+      }
       if (teeUsage) {
         roundUsage.inputTokens += teeUsage.inputTokens;
         roundUsage.outputTokens += teeUsage.outputTokens;
@@ -2383,7 +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-ZBDJYOB2.js");
+          const { executeTests } = await import("./run-tests-QBNLSYCO.js");
           const argStr = args.join(" ").trim();
           let testArgs = {};
           if (argStr) {

package/dist/{server-LEPDI3OL.js → server-5TGJOWML.js}

@@ -3,14 +3,14 @@ import {
   ToolRegistry,
   getDangerLevel,
   schemaToJsonSchema
-} from "./chunk-N62X6OKJ.js";
+} from "./chunk-5EVKTBUU.js";
 import "./chunk-3BICTI5M.js";
 import "./chunk-2DXY7UGF.js";
-import "./chunk-CUQ4HXHA.js";
+import "./chunk-5LKW2GOF.js";
 import "./chunk-2ZD3YTVM.js";
 import {
   VERSION
-} from "./chunk-LZR3DYW3.js";
+} from "./chunk-CPIQXP7Q.js";
 import "./chunk-4BKXL7SM.js";
 import "./chunk-7ZJN4KLV.js";
 import "./chunk-KHYD3WXE.js";

package/dist/{task-orchestrator-XSNTJYNR.js → task-orchestrator-IJFF7XLQ.js}

@@ -4,14 +4,14 @@ import {
   getDangerLevel,
   googleSearchContext,
   truncateOutput
-} from "./chunk-N62X6OKJ.js";
+} from "./chunk-5EVKTBUU.js";
 import "./chunk-3BICTI5M.js";
 import "./chunk-2DXY7UGF.js";
-import "./chunk-CUQ4HXHA.js";
+import "./chunk-5LKW2GOF.js";
 import "./chunk-2ZD3YTVM.js";
 import {
   SUBAGENT_ALLOWED_TOOLS
-} from "./chunk-LZR3DYW3.js";
+} from "./chunk-CPIQXP7Q.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": "0.4.109",
+  "version": "0.4.111",
   "description": "Cross-platform REPL-style AI CLI with multi-provider support",
   "type": "module",
   "main": "./dist/index.js",