jinzd-ai-cli 0.4.186 → 0.4.187

package/dist/electron-server.js

@@ -36,7 +36,7 @@ import {
   VERSION,
   buildUserIdentityPrompt,
   runTestsTool
-} from "./chunk-7HMX2MTY.js";
+} from "./chunk-L5L5B7HD.js";
 import {
   hasSemanticIndex,
   semanticSearch
@@ -1552,625 +1552,1173 @@ Node.js does not automatically use system proxies. Try one of the following:
 // src/providers/openai-compatible.ts
 import OpenAI from "openai";
 
-// src/core/agent-loop.ts
-function partialTagTail(s, tag) {
-  const max = Math.min(s.length, tag.length - 1);
-  for (let len = max; len > 0; len--) {
-    if (s.endsWith(tag.slice(0, len))) return len;
-  }
-  return 0;
+// src/tools/hallucination.ts
+var HALLUCINATION_PATTERNS = [
+  /文件路径[：:]\s*`?[^\s`]+\.\w{1,5}/,
+  // 文件路径: `path/to/file.ext`（要求文件扩展名）
+  /已生成[：:!！]/,
+  // 已生成完成！
+  /已保存到?\s*[`'"]/,
+  // 已保存到 `path`（要求后跟路径引号）
+  /已写入[：:!！]/,
+  // 已写入！
+  /已创建[：:!！]/,
+  // 已创建！
+  /File\s+(?:written|saved|created)\s+(?:to|as|at)/i,
+  // File written to / saved as（要求介词）
+  /生成完成[！!]/,
+  // 生成完成！
+  /✅\s*(?:文件|已[生保写创]|第)\S*\.\w{1,5}/,
+  // ✅ 文件已保存 path.ext（要求文件扩展名）
+  /文件已[成功]?创建/,
+  // 文件已成功创建 / 文件已创建
+  /教案已[成功]?[生创保写]/,
+  // 教案已成功生成 / 教案已保存
+  /已成功[保写创生]入?[：:!！\s`'"]/,
+  // 已成功保存 / 已成功写入 / 已成功创建
+  /保存[到至]了?\s*[`'"]/,
+  // 保存到了 `path` / 保存至 'path'
+  /内容如下[：:]/,
+  // 内容如下：（后跟大段文件内容）
+  /以下是.*(?:教案|文件|内容)[：:]/
+  // 以下是xx教案内容：（Kimi 常见模式）
+];
+function detectsHallucinatedFileOp(content) {
+  return HALLUCINATION_PATTERNS.some((pattern) => pattern.test(content));
 }
-var ThinkTagFilter = class {
-  inThink = false;
-  buf = "";
-  push(raw) {
-    this.buf += raw;
-    let out = "";
-    while (this.buf.length > 0) {
-      if (!this.inThink) {
-        const open = this.buf.indexOf("<think>");
-        if (open === -1) {
-          const keep = partialTagTail(this.buf, "<think>");
-          out += this.buf.slice(0, this.buf.length - keep);
-          this.buf = this.buf.slice(this.buf.length - keep);
-          break;
-        }
-        out += this.buf.slice(0, open);
-        this.buf = this.buf.slice(open + "<think>".length);
-        this.inThink = true;
-      } else {
-        const close = this.buf.indexOf("</think>");
-        if (close === -1) {
-          const keep = partialTagTail(this.buf, "</think>");
-          this.buf = this.buf.slice(this.buf.length - keep);
-          break;
-        }
-        this.buf = this.buf.slice(close + "</think>".length);
-        this.inThink = false;
-      }
+function hadPreviousWriteToolCalls(extraMessages) {
+  const msgs = extraMessages;
+  return msgs.some((msg) => {
+    if (msg.role === "assistant" && Array.isArray(msg.tool_calls)) {
+      return msg.tool_calls.some((tc) => {
+        const fn = tc.function;
+        const name = fn?.name ?? "";
+        return name === "write_file" || name === "edit_file";
+      });
     }
-    return out;
-  }
-  /** 流结束：若仍持留可能的半截 '<think>' 前缀且并未进入 think 块，它是真实文本。 */
-  flush() {
-    if (!this.inThink && this.buf) {
-      const tail = this.buf;
-      this.buf = "";
-      return tail;
+    if (msg.role === "assistant" && Array.isArray(msg.content)) {
+      return msg.content.some((block) => {
+        if (block.type !== "tool_use") return false;
+        const name = block.name ?? "";
+        return name === "write_file" || name === "edit_file";
+      });
     }
-    this.buf = "";
-    return "";
-  }
-};
-function repairToolCallArguments(raw, onWarn) {
-  const argStr = raw || "{}";
-  try {
-    return JSON.parse(argStr);
-  } catch {
-    const truncated = argStr.trimEnd();
-    const lastComma = truncated.lastIndexOf(",");
-    const fixed = lastComma > 0 ? truncated.slice(0, lastComma) + "}" : truncated.slice(0, truncated.indexOf("{") + 1) + "}";
-    try {
-      const repaired = JSON.parse(fixed);
-      onWarn?.("Tool call JSON was truncated and auto-repaired. Some parameters may be missing.");
-      return repaired;
-    } catch {
-      onWarn?.("Tool call JSON could not be parsed, using empty arguments.");
-      return {};
+    if (msg.role === "model" && Array.isArray(msg.parts)) {
+      return msg.parts.some((part) => {
+        const fc = part.functionCall;
+        const name = fc?.name ?? "";
+        return name === "write_file" || name === "edit_file";
+      });
     }
-  }
+    return false;
+  });
 }
-async function consumeToolCallStream(stream, hooks = {}) {
-  const textParts = [];
-  const accumulators = /* @__PURE__ */ new Map();
-  let usage;
-  let rawContent;
-  let reasoningContent;
-  let finishReason;
-  let aborted = false;
-  const thinkFilter = new ThinkTagFilter();
-  const emitText = (raw) => {
-    const visible = thinkFilter.push(raw);
-    if (visible) {
-      textParts.push(visible);
-      hooks.onText?.(visible);
+var TOOL_CALL_REMINDER = `
+
+[\u26A0\uFE0F Mandatory Tool Call Policy]
+When you need to create, write, or modify files, you MUST use the function calling API to invoke write_file or edit_file.
+NEVER claim "file saved", "file created", "written to", etc. in your response text without actually calling the tool.
+Describing file content in text without calling the tool = the file does not exist = task failure.
+If multiple files need to be generated, you MUST call write_file separately for each file \u2014 do not skip any.
+Do NOT output fake "completion summaries" unless you have actually completed all file writes via tool_calls.
+
+CRITICAL \u2014 Batch file generation rules:
+1. You MUST call write_file once per file. There are NO shortcuts.
+2. After writing file N, immediately proceed to call write_file for file N+1. Do NOT stop to summarize.
+3. If you find yourself typing file content into your response text instead of into a write_file call, STOP and use the tool.
+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.";
+function extractClaimedFilePaths(content) {
+  const paths = /* @__PURE__ */ new Set();
+  const add = (p) => {
+    const trimmed = p.trim().replace(/[，,。、；;:：]+$/, "");
+    if (trimmed && /\.\w{1,6}$/.test(trimmed)) paths.add(trimmed);
+  };
+  let m;
+  const actionLineRe = /(?:已[生保写创]|saved|written|created|完成.*(?:写入|保存|创建|生成)|输出|file\s+(?:saved|written|created))/i;
+  const backtickRe = /`([^`\n]+?\.\w{1,6})`/g;
+  while ((m = backtickRe.exec(content)) !== null) {
+    let pos = m.index;
+    let linesBack = 0;
+    while (linesBack < 9 && pos > 0) {
+      pos--;
+      if (content[pos] === "\n") linesBack++;
+    }
+    const windowStart = pos === 0 ? 0 : pos + 1;
+    const lineEndIdx = content.indexOf("\n", m.index + m[0].length);
+    const window = content.slice(windowStart, lineEndIdx === -1 ? void 0 : lineEndIdx);
+    if (actionLineRe.test(window)) add(m[1]);
+  }
+  const zhRe = /(?:已保存(?:到)?|已写入(?:到)?|已创建|已生成|文件路径[：:]|保存为|写入到)\s*[`'”””]?([^\s`'”””，,。\n]+?\.\w{1,6})/g;
+  while ((m = zhRe.exec(content)) !== null) add(m[1]);
+  const enRe = /(?:saved|written|created)\s+(?:to|as|at)\s+[`'”]?([^\s`'”\n,]+?\.\w{1,6})/gi;
+  while ((m = enRe.exec(content)) !== null) add(m[1]);
+  const checkRe = /✅[^\n`]*?[`'”]?([^\s`'”\n,，。]+?\.\w{1,6})/g;
+  while ((m = checkRe.exec(content)) !== null) {
+    let pos = m.index;
+    let linesBack = 0;
+    while (linesBack < 9 && pos > 0) {
+      pos--;
+      if (content[pos] === "\n") linesBack++;
+    }
+    const windowStart = pos === 0 ? 0 : pos + 1;
+    const lineEndIdx = content.indexOf("\n", m.index + m[0].length);
+    const window = content.slice(windowStart, lineEndIdx === -1 ? void 0 : lineEndIdx);
+    if (actionLineRe.test(window)) add(m[1]);
+  }
+  return Array.from(paths);
+}
+function extractWrittenFilePaths(extraMessages) {
+  const paths = /* @__PURE__ */ new Set();
+  const msgs = extraMessages;
+  const addFromArgs = (raw) => {
+    if (typeof raw === "string") {
+      try {
+        const parsed = JSON.parse(raw);
+        if (typeof parsed.path === "string") paths.add(parsed.path);
+      } catch {
+      }
+    } else if (raw && typeof raw === "object") {
+      const p = raw.path;
+      if (typeof p === "string") paths.add(p);
     }
   };
-  try {
-    for await (const event of stream) {
-      if (hooks.signal?.aborted) {
-        aborted = true;
-        break;
+  for (const msg of msgs) {
+    if (msg.role === "assistant" && Array.isArray(msg.tool_calls)) {
+      for (const tc of msg.tool_calls) {
+        const fn = tc.function;
+        const name = fn?.name ?? "";
+        if (name === "write_file" || name === "edit_file") {
+          addFromArgs(fn?.arguments);
+        }
       }
-      switch (event.type) {
-        case "text_delta":
-          emitText(event.delta);
-          break;
-        case "thinking_start":
-          hooks.onThinkingStart?.();
-          break;
-        case "thinking_delta":
-          hooks.onThinkingDelta?.(event.delta);
-          break;
-        case "thinking_end":
-          hooks.onThinkingEnd?.();
-          break;
-        case "tool_call_start":
-          accumulators.set(event.index, { id: event.id, name: event.name, arguments: "" });
-          hooks.onToolCallStart?.(event.index, event.id, event.name);
-          break;
-        case "tool_call_delta": {
-          const acc = accumulators.get(event.index);
-          if (acc) acc.arguments += event.argumentsDelta;
-          break;
+    }
+    if (msg.role === "assistant" && Array.isArray(msg.content)) {
+      for (const block of msg.content) {
+        if (block.type !== "tool_use") continue;
+        const name = block.name ?? "";
+        if (name === "write_file" || name === "edit_file") {
+          addFromArgs(block.input);
         }
-        case "tool_call_end":
-          break;
-        case "done":
-          if (event.usage) usage = event.usage;
-          if (event.rawContent) rawContent = event.rawContent;
-          if (event.reasoningContent) reasoningContent = event.reasoningContent;
-          if (event.finishReason) finishReason = event.finishReason;
-          break;
       }
     }
-  } catch (err) {
-    if (err instanceof Error && (err.name === "AbortError" || err.message.includes("aborted"))) {
-      aborted = true;
-    } else {
-      throw err;
+    if (msg.role === "model" && Array.isArray(msg.parts)) {
+      for (const part of msg.parts) {
+        const fc = part.functionCall;
+        if (!fc) continue;
+        const name = fc.name ?? "";
+        if (name === "write_file" || name === "edit_file") {
+          addFromArgs(fc.args);
+        }
+      }
     }
   }
-  const tail = thinkFilter.flush();
-  if (tail && !aborted) {
-    textParts.push(tail);
-    hooks.onText?.(tail);
-  }
-  const textContent = textParts.join("");
-  if (aborted) {
-    return { textContent, toolCalls: [], usage, rawContent, reasoningContent, finishReason, aborted };
-  }
-  const toolCalls = [];
-  for (const [, acc] of accumulators) {
-    toolCalls.push({
-      id: acc.id,
-      name: acc.name,
-      arguments: repairToolCallArguments(acc.arguments, hooks.onWarn)
-    });
-  }
-  if (toolCalls.length > 0) {
-    if (rawContent) {
-      toolCalls._rawContent = rawContent;
+  return Array.from(paths);
+}
+function extractBashCommands(extraMessages) {
+  const cmds = [];
+  const msgs = extraMessages;
+  const addCmd = (raw) => {
+    if (typeof raw === "string") {
+      try {
+        const parsed = JSON.parse(raw);
+        if (typeof parsed.command === "string") cmds.push(parsed.command);
+      } catch {
+      }
+    } else if (raw && typeof raw === "object") {
+      const c = raw.command;
+      if (typeof c === "string") cmds.push(c);
     }
-    if (textContent) {
-      toolCalls._streamedText = textContent;
+  };
+  for (const msg of msgs) {
+    if (msg.role === "assistant" && Array.isArray(msg.tool_calls)) {
+      for (const tc of msg.tool_calls) {
+        const fn = tc.function;
+        if (fn?.name === "bash") addCmd(fn?.arguments);
+      }
     }
-  }
-  return { textContent, toolCalls, usage, rawContent, reasoningContent, finishReason, aborted };
-}
-var FREE_ROUND_TOOLS = /* @__PURE__ */ new Set(["write_todos"]);
-var MAX_CONSECUTIVE_FREE_ROUNDS = 3;
-var FreeRoundTracker = class {
-  consecutive = 0;
-  /** 返回 true 表示本轮不消耗有效轮次（调用方执行 round--）。 */
-  apply(toolNames) {
-    const allFree = toolNames.length > 0 && toolNames.every((n) => FREE_ROUND_TOOLS.has(n));
-    if (!allFree) {
-      this.consecutive = 0;
-      return false;
-    }
-    this.consecutive++;
-    return this.consecutive <= MAX_CONSECUTIVE_FREE_ROUNDS;
-  }
-};
-var BudgetWarner = class {
-  constructor(maxToolRounds) {
-    this.maxToolRounds = maxToolRounds;
-    this.noteAt = Math.max(10, Math.floor(maxToolRounds * 0.2));
-    const lowRaw = Math.max(5, Math.floor(maxToolRounds * 0.1));
-    const criticalRaw = Math.max(3, Math.floor(maxToolRounds * 0.05));
-    this.lowAt = Math.min(lowRaw, this.noteAt - 1);
-    this.criticalAt = Math.min(criticalRaw, this.lowAt - 1);
-  }
-  noteAt;
-  lowAt;
-  criticalAt;
-  warnedNote = false;
-  warnedLow = false;
-  warnedCritical = false;
-  check(roundsLeft) {
-    if (!this.warnedCritical && roundsLeft <= this.criticalAt) {
-      this.warnedCritical = true;
-      return {
-        level: "critical",
-        injectMessage: `\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.`,
-        displayMessage: `\u{1F6A8} Critical: ${roundsLeft} rounds remaining`
-      };
-    }
-    if (!this.warnedLow && roundsLeft <= this.lowAt) {
-      this.warnedLow = true;
-      return {
-        level: "low",
-        injectMessage: `\u26A0\uFE0F Budget warning: Only ${roundsLeft} tool rounds remaining. Prioritize completing the most critical task. Use efficient approaches (batch edits, fewer reads). If you cannot finish everything, summarize what's done and what remains.`,
-        displayMessage: `\u26A0\uFE0F Low budget: ${roundsLeft} rounds remaining`
-      };
+    if (msg.role === "assistant" && Array.isArray(msg.content)) {
+      for (const block of msg.content) {
+        if (block.type === "tool_use" && block.name === "bash") addCmd(block.input);
+      }
     }
-    if (!this.warnedNote && roundsLeft <= this.noteAt) {
-      this.warnedNote = true;
-      return {
-        level: "note",
-        injectMessage: `\u{1F4CA} Budget note: ${roundsLeft} tool rounds remaining out of ${this.maxToolRounds}. Plan your remaining work efficiently \u2014 use batch operations (e.g., replaceAll) when possible.`
-      };
+    if (msg.role === "model" && Array.isArray(msg.parts)) {
+      for (const part of msg.parts) {
+        const fc = part.functionCall;
+        if (fc && fc.name === "bash") addCmd(fc.args);
+      }
     }
-    return null;
   }
-};
-var EMPTY_RESPONSE_NUDGE = "Your previous response was empty \u2014 no text and no tool calls. This usually means the context window is nearly full. Please either: (1) continue the task by calling the next tool you need, or (2) give a concise final text summary of what has been accomplished so far and what remains. Do NOT repeat earlier long outputs.";
-function describeFinishReason(fr) {
-  if (fr === "length") return "output limit reached (finish_reason=length)";
-  if (fr === "content_filter") return "content blocked (finish_reason=content_filter)";
-  if (fr) return `empty response (finish_reason=${fr})`;
-  return "empty response";
+  return cmds;
 }
-function emptyResponseHint(fr) {
-  if (fr === "length") return "Output token limit hit \u2014 try /compact to reduce context, raise maxTokens, or /model to switch.";
-  if (fr === "content_filter") return "Content was blocked by the provider filter.";
-  return "Context window may be exhausted or max_tokens too low.";
+function findPhantomClaims(content, extraMessages) {
+  const claimed = extractClaimedFilePaths(content);
+  if (claimed.length === 0) return [];
+  const normalize = (p) => p.replace(/\\/g, "/").toLowerCase().replace(/^\.\//, "");
+  const basename6 = (p) => {
+    const parts = normalize(p).split("/");
+    return parts[parts.length - 1] ?? "";
+  };
+  const written = extractWrittenFilePaths(extraMessages).map(normalize);
+  const writtenBases = new Set(written.map(basename6));
+  const writtenFull = new Set(written);
+  const bashText = extractBashCommands(extraMessages).map((c) => c.replace(/\\/g, "/").toLowerCase()).join("\n");
+  return claimed.filter((raw) => {
+    const norm2 = normalize(raw);
+    if (writtenFull.has(norm2)) return false;
+    for (const w of writtenFull) {
+      if (w.endsWith("/" + norm2) || norm2.endsWith("/" + w)) return false;
+    }
+    if (writtenBases.has(basename6(norm2))) return false;
+    const base = basename6(norm2);
+    if (base && bashText.includes(base)) return false;
+    return true;
+  });
 }
-var EmptyResponseGuard = class {
-  retries = 0;
-  onEmpty(canRetry, finishReason) {
-    if (this.retries === 0 && canRetry) {
-      this.retries++;
-      return {
-        action: "nudge",
-        injectMessage: EMPTY_RESPONSE_NUDGE,
-        displayMessage: `\u26A0 ${describeFinishReason(finishReason)} \u2014 nudging AI to continue...`
-      };
-    }
-    return {
-      action: "stop",
-      displayMessage: "\u26A0 AI returned empty responses twice in a row. Stopping agentic loop.",
-      hint: emptyResponseHint(finishReason)
-    };
-  }
-  /** 非空响应到达 → 重置计数（下次空响应仍可 nudge 一次）。 */
-  onNonEmpty() {
-    this.retries = 0;
-  }
-};
-var ContextPressureMonitor = class {
-  warned80 = false;
-  check(requestTokens, contextWindow) {
-    if (contextWindow <= 0) return { action: "ok", ratio: 0 };
-    const ratio = requestTokens / contextWindow;
-    if (ratio >= 0.95) return { action: "abort", ratio };
-    if (ratio >= 0.8 && !this.warned80) {
-      this.warned80 = true;
-      return {
-        action: "warn",
-        ratio,
-        injectMessage: `\u26A0\uFE0F Context pressure: ~${Math.round(ratio * 100)}% of the ${contextWindow.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.`
-      };
-    }
-    return { action: "ok", ratio };
-  }
-};
-function accumulateUsage(total, delta) {
-  if (!delta) return;
-  total.inputTokens += delta.inputTokens;
-  total.outputTokens += delta.outputTokens;
-  total.cacheCreationTokens += delta.cacheCreationTokens ?? 0;
-  total.cacheReadTokens += delta.cacheReadTokens ?? 0;
+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.';
 }
-function buildRoundBudgetHint(opts) {
-  const pauseHint = opts.autoPauseInterval > 0 ? `
-- Every ${opts.autoPauseInterval} rounds the user will be asked whether to continue \u2014 use this as a natural checkpoint to report progress.` : "";
-  if (opts.planMode) {
-    return `
-
-[Tool Round Budget \u2014 Plan Mode]
-You have a maximum of ${opts.maxToolRounds} tool call rounds. You are in READ-ONLY Plan Mode:
-- Only use: read_file, list_dir, grep_files, glob_files, ask_user, write_todos
-- Do NOT attempt to call bash, write_file, edit_file \u2014 they are disabled
-- Do NOT write shell commands or code blocks as a substitute for tool calls
-- Do NOT read the same file more than once
-- Call write_todos ONCE to present your plan, then give a text summary
-- If the user asks you to execute anything, respond: "Please type /plan execute to switch to execute mode."${pauseHint}`;
+var DSML_PIPE_CLASS = "[|\\uFF5C\\u2502\\u2503\\u01C0]";
+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,
+  // 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,
+  // v0.4.173: DeepSeek V4 DSML pseudo-tool-call markup leaked as text. DeepSeek
+  // emits a fake tool call using its native special-token markup
+  //   <｜｜DSML｜｜tool_calls> <｜｜DSML｜｜invoke name="write"> <｜｜DSML｜｜parameter …>
+  // where the "pipe" is U+FF5C FULLWIDTH VERTICAL LINE (the same token family as
+  // <｜User｜>/<｜Assistant｜>). We saw an exam paper saved via save_last_response
+  // whose tee stream was preamble + this DSML wrapper + the real document body.
+  // The earlier <invoke …> pattern uses ASCII < > and does NOT match these.
+  new RegExp(`<\\/?\\s*${DSML_PIPE_CLASS}+\\s*DSML\\s*${DSML_PIPE_CLASS}+`, "i")
+];
+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 `
-
-[Tool Round Budget]
-You have a maximum of ${opts.maxToolRounds} tool call rounds for this task. Plan efficiently:
-- Prefer batch operations (e.g. global find-and-replace) over repetitive single edits.
-- Do NOT read the same file more than once \u2014 use the content from previous reads.
-- Prioritize the most critical tasks first in case rounds run out.
-- When remaining rounds are low, focus on completing the current task and summarizing.${pauseHint}`;
-}
-function buildRoundsExhaustedPrompt(maxToolRounds) {
-  return `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 (e.g. send another request to continue)`;
+  return null;
 }
-function buildUserStopMessage(effectiveRound, maxToolRounds) {
-  return `The user has stopped the task at round ${effectiveRound}/${maxToolRounds}. Do not call any more tools. Summarize what has been completed and what remains.`;
+function stripPseudoToolCalls(content) {
+  if (!content) return content;
+  let out = content;
+  const dsmlBody = extractDsmlContent(out);
+  if (dsmlBody !== null) {
+    out = dsmlBody;
+  } else {
+    out = stripDsmlTags(out);
+  }
+  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*$/gm, "");
+  out = unwrapDocumentFence(out);
+  out = peelMetaNarration(out);
+  out = out.replace(/\n{3,}/g, "\n\n").trim();
+  return out;
 }
-function summarizeRecentTools(history, interval) {
-  const recent = history.slice(-interval);
-  const counts = /* @__PURE__ */ new Map();
-  for (const rh of recent) {
-    for (const t of rh.tools) counts.set(t, (counts.get(t) || 0) + 1);
+function extractDsmlContent(content) {
+  if (!content) return null;
+  const P = DSML_PIPE_CLASS;
+  const re = new RegExp(
+    `<\\s*${P}+\\s*DSML\\s*${P}+\\s*parameter\\b[^>]*\\bname\\s*=\\s*["']content["'][^>]*>([\\s\\S]*?)<\\s*/\\s*${P}+\\s*DSML\\s*${P}+\\s*parameter\\s*>`,
+    "i"
+  );
+  const m = content.match(re);
+  if (m && typeof m[1] === "string") {
+    const body = m[1].trim();
+    return body.length > 0 ? body : null;
   }
-  return [...counts.entries()].sort((a, b) => b[1] - a[1]).map(([name, count]) => count > 1 ? `${name}\xD7${count}` : name).join(", ");
+  return null;
 }
-
-// src/providers/openai-compatible.ts
-function toUsage(u) {
-  if (!u) return void 0;
-  const cached = u.prompt_tokens_details?.cached_tokens ?? 0;
-  const usage = {
-    inputTokens: Math.max(0, u.prompt_tokens - cached),
-    outputTokens: u.completion_tokens
-  };
-  if (cached > 0) usage.cacheReadTokens = cached;
-  return usage;
+function stripDsmlTags(content) {
+  const P = DSML_PIPE_CLASS;
+  let out = content;
+  out = out.replace(
+    new RegExp(
+      `<\\s*${P}+\\s*DSML\\s*${P}+\\s*tool_calls\\b[\\s\\S]*?<\\s*/\\s*${P}+\\s*DSML\\s*${P}+\\s*tool_calls\\s*>`,
+      "gi"
+    ),
+    ""
+  );
+  out = out.replace(new RegExp(`<\\s*/?\\s*${P}+\\s*DSML\\s*${P}+[^>]*>`, "gi"), "");
+  return out;
 }
-var OpenAICompatibleProvider = class extends BaseProvider {
-  client;
-  defaultTimeout = 6e4;
-  // ms
-  /** 子类设为 false 可禁用流式工具调用（虚假声明检测需要完整响应） */
-  enableStreamingToolCalls = true;
-  async initialize(apiKey, options) {
-    if (options?.timeout !== void 0) {
-      this.defaultTimeout = options.timeout;
+function unwrapDocumentFence(content) {
+  const trimmed = content.trim();
+  const open = trimmed.match(/^```\s*(markdown|md|gfm)?\s*\n/i);
+  if (!open) return content;
+  const afterOpen = trimmed.slice(open[0].length);
+  const closeMatch = afterOpen.match(/\n```\s*$/);
+  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);
     }
-    const clientOptions = {
-      apiKey,
-      baseURL: options?.baseUrl ?? this.defaultBaseUrl,
-      timeout: this.defaultTimeout
-    };
-    const proxyUrl = options?.proxy;
-    try {
-      const { Agent: Agent2, ProxyAgent, fetch: undiciFetch } = await import("undici");
-      const STREAM_BODY_TIMEOUT = 30 * 60 * 1e3;
-      const STREAM_HEADERS_TIMEOUT = 5 * 60 * 1e3;
-      const dispatcher = proxyUrl ? new ProxyAgent({
-        uri: proxyUrl,
-        bodyTimeout: STREAM_BODY_TIMEOUT,
-        headersTimeout: STREAM_HEADERS_TIMEOUT
-      }) : new Agent2({
-        bodyTimeout: STREAM_BODY_TIMEOUT,
-        headersTimeout: STREAM_HEADERS_TIMEOUT
-      });
-      clientOptions.fetch = ((url, init) => undiciFetch(url, { ...init, dispatcher }));
-    } catch {
+    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);
+      }
     }
-    this.client = new OpenAI(clientOptions);
   }
-  /**
-   * 将 systemPrompt + messages 合并为 OpenAI messages 数组（system 消息放首位）。
-   *
-   * v0.4.100+：按原始顺序保留工具消息（assistant.toolCalls 和 role='tool'），
-   * 不再剥离到 _extraMessages 末尾——之前的剥离会让历史工具往返被插到当前用户消息之后，
-   * 导致模型把"过去的工具调用结果"当作"对当前问题的回应"，DeepSeek V4 Flash 上尤其明显
-   * （会复读上一轮的"完成汇总"）。
-   *
-   * DeepSeek V4 thinking 模式：所有 assistant 消息（含带 toolCalls 的）必须有
-   * reasoning_content 字段，缺失则 API 400。
-   */
-  buildMessages(request) {
-    const msgs = [];
-    for (const m of request.messages) {
-      if (m.role === "tool") {
-        if (!m.toolCallId) continue;
-        msgs.push({
-          role: "tool",
-          tool_call_id: m.toolCallId,
-          content: typeof m.content === "string" ? m.content : ""
-        });
-        continue;
-      }
-      if (m.role === "assistant" && m.toolCalls && m.toolCalls.length > 0) {
-        const assistantMsg = {
-          role: "assistant",
-          content: typeof m.content === "string" && m.content ? m.content : null,
-          tool_calls: m.toolCalls.map((tc) => ({
-            id: tc.id,
-            type: "function",
-            function: { name: tc.name, arguments: JSON.stringify(tc.arguments) }
-          })),
-          reasoning_content: m.reasoningContent ?? ""
-        };
-        msgs.push(assistantMsg);
-        continue;
-      }
-      const base = { role: m.role, content: m.content };
-      if (m.role === "assistant") {
-        base.reasoning_content = m.reasoningContent ?? "";
+  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();
+}
+var META_NARRATION_HARD_MARKERS = [
+  /\[⚠️\s*CONTENT GENERATION MODE\]/,
+  /CONTENT_ONLY_STREAM_REMINDER\b/,
+  /<system-reminder>/i
+];
+var META_NARRATION_HEURISTICS = [
+  /\bthe user (?:is asking me|wants me|is requesting|expects me)\b/i,
+  /\blet me (?:re-?read|re-?consider|reconsider|think about|carefully (?:re-?read|consider))\b/i,
+  /\bI'?m (?:in (?:a )?content-only|in CONTENT-ONLY|currently in)\b/i,
+  /\bI think (?:there might be|I should|I cannot|the (?:user|best)|maybe)\b/i,
+  /\bWait,?\s+let me\b/i,
+  /\bActually,?\s+I\b/i,
+  /\bI need to be honest with the user\b/i,
+  /\bI(?:'m| am) in a special mode\b/i,
+  /\bGiven that I cannot\b/i
+];
+function detectMetaNarration(content) {
+  if (!content) return null;
+  const head = content.slice(0, 2e3);
+  for (const re of META_NARRATION_HARD_MARKERS) {
+    if (re.test(head)) return re.source;
+  }
+  if (/^#{1,3}\s+\S/m.test(head)) return null;
+  let hits = 0;
+  let firstMatch = "";
+  for (const re of META_NARRATION_HEURISTICS) {
+    if (re.test(head)) {
+      hits++;
+      if (!firstMatch) firstMatch = re.source;
+      if (hits >= 2) return `meta-narration:${firstMatch}`;
+    }
+  }
+  return null;
+}
+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/core/agent-loop.ts
+function partialTagTail(s, tag) {
+  const max = Math.min(s.length, tag.length - 1);
+  for (let len = max; len > 0; len--) {
+    if (s.endsWith(tag.slice(0, len))) return len;
+  }
+  return 0;
+}
+var ThinkTagFilter = class {
+  inThink = false;
+  buf = "";
+  push(raw) {
+    this.buf += raw;
+    let out = "";
+    while (this.buf.length > 0) {
+      if (!this.inThink) {
+        const open = this.buf.indexOf("<think>");
+        if (open === -1) {
+          const keep = partialTagTail(this.buf, "<think>");
+          out += this.buf.slice(0, this.buf.length - keep);
+          this.buf = this.buf.slice(this.buf.length - keep);
+          break;
+        }
+        out += this.buf.slice(0, open);
+        this.buf = this.buf.slice(open + "<think>".length);
+        this.inThink = true;
+      } else {
+        const close = this.buf.indexOf("</think>");
+        if (close === -1) {
+          const keep = partialTagTail(this.buf, "</think>");
+          this.buf = this.buf.slice(this.buf.length - keep);
+          break;
+        }
+        this.buf = this.buf.slice(close + "</think>".length);
+        this.inThink = false;
       }
-      msgs.push(base);
     }
-    const systemContent = [request.systemPrompt, request.systemPromptVolatile].filter(Boolean).join("\n\n---\n\n");
-    if (systemContent) {
-      return [{ role: "system", content: systemContent }, ...msgs];
+    return out;
+  }
+  /** 流结束：若仍持留可能的半截 '<think>' 前缀且并未进入 think 块，它是真实文本。 */
+  flush() {
+    if (!this.inThink && this.buf) {
+      const tail = this.buf;
+      this.buf = "";
+      return tail;
     }
-    return msgs;
+    this.buf = "";
+    return "";
   }
-  async chat(request) {
+};
+function repairToolCallArguments(raw, onWarn) {
+  const argStr = raw || "{}";
+  try {
+    return JSON.parse(argStr);
+  } catch {
+    const truncated = argStr.trimEnd();
+    const lastComma = truncated.lastIndexOf(",");
+    const fixed = lastComma > 0 ? truncated.slice(0, lastComma) + "}" : truncated.slice(0, truncated.indexOf("{") + 1) + "}";
     try {
-      const response = await this.client.chat.completions.create({
-        model: request.model,
-        messages: this.buildMessages(request),
-        temperature: request.temperature,
-        max_tokens: request.maxTokens,
-        stream: false,
-        ...request.thinking ? { thinking: { type: "enabled" } } : {}
-      }, {
-        timeout: request.timeout ?? this.defaultTimeout
-      });
-      const firstChoice = response.choices?.[0];
-      if (!firstChoice) {
-        return { content: "", model: response.model, usage: void 0 };
-      }
-      return {
-        content: firstChoice.message.content ?? "",
-        model: response.model,
-        usage: toUsage(response.usage)
-      };
-    } catch (err) {
-      throw this.wrapError(err);
+      const repaired = JSON.parse(fixed);
+      onWarn?.("Tool call JSON was truncated and auto-repaired. Some parameters may be missing.");
+      return repaired;
+    } catch {
+      onWarn?.("Tool call JSON could not be parsed, using empty arguments.");
+      return {};
     }
   }
-  async *chatStream(request) {
-    try {
-      const stream = await this.client.chat.completions.create({
-        model: request.model,
-        messages: this.buildMessages(request),
-        temperature: request.temperature,
-        max_tokens: request.maxTokens,
-        stream: true,
-        // 请求末尾 usage chunk，供 token 统计使用
-        stream_options: { include_usage: true },
-        ...request.thinking ? { thinking: { type: "enabled" } } : {}
-      }, {
-        timeout: request.timeout ?? this.defaultTimeout,
-        signal: request.signal
-      });
-      let thinkingStarted = false;
-      let reasoningAccumulator = "";
-      for await (const chunk of stream) {
-        const choice = chunk.choices[0];
-        const done = choice?.finish_reason != null;
-        if (!choice && chunk.usage) {
-          yield {
-            delta: "",
-            done: true,
-            usage: toUsage(chunk.usage),
-            ...reasoningAccumulator ? { reasoningContent: reasoningAccumulator } : {}
-          };
-          continue;
-        }
-        const reasoningDelta = choice?.delta?.reasoning_content;
-        if (reasoningDelta) {
-          if (!thinkingStarted) {
-            yield { delta: "<think>", done: false };
-            thinkingStarted = true;
-          }
-          reasoningAccumulator += reasoningDelta;
-          yield { delta: reasoningDelta, done: false };
-          continue;
-        }
-        const delta = choice?.delta?.content ?? "";
-        if (thinkingStarted && delta) {
-          thinkingStarted = false;
-          yield { delta: "</think>", done: false };
-        }
-        if (done) {
-          yield { delta, done, ...reasoningAccumulator ? { reasoningContent: reasoningAccumulator } : {} };
-        } else {
-          yield { delta, done };
+}
+async function consumeToolCallStream(stream, hooks = {}) {
+  const textParts = [];
+  const accumulators = /* @__PURE__ */ new Map();
+  let usage;
+  let rawContent;
+  let reasoningContent;
+  let finishReason;
+  let aborted = false;
+  const thinkFilter = new ThinkTagFilter();
+  const emitText = (raw) => {
+    const visible = thinkFilter.push(raw);
+    if (visible) {
+      textParts.push(visible);
+      hooks.onText?.(visible);
+    }
+  };
+  try {
+    for await (const event of stream) {
+      if (hooks.signal?.aborted) {
+        aborted = true;
+        break;
+      }
+      switch (event.type) {
+        case "text_delta":
+          emitText(event.delta);
+          break;
+        case "thinking_start":
+          hooks.onThinkingStart?.();
+          break;
+        case "thinking_delta":
+          hooks.onThinkingDelta?.(event.delta);
+          break;
+        case "thinking_end":
+          hooks.onThinkingEnd?.();
+          break;
+        case "tool_call_start":
+          accumulators.set(event.index, { id: event.id, name: event.name, arguments: "" });
+          hooks.onToolCallStart?.(event.index, event.id, event.name);
+          break;
+        case "tool_call_delta": {
+          const acc = accumulators.get(event.index);
+          if (acc) acc.arguments += event.argumentsDelta;
+          break;
         }
+        case "tool_call_end":
+          break;
+        case "done":
+          if (event.usage) usage = event.usage;
+          if (event.rawContent) rawContent = event.rawContent;
+          if (event.reasoningContent) reasoningContent = event.reasoningContent;
+          if (event.finishReason) finishReason = event.finishReason;
+          break;
       }
-    } catch (err) {
-      throw this.wrapError(err);
+    }
+  } catch (err) {
+    if (err instanceof Error && (err.name === "AbortError" || err.message.includes("aborted"))) {
+      aborted = true;
+    } else {
+      throw err;
     }
   }
-  /**
-   * 请求 AI 并获取工具调用列表（不执行，只解析）。
-   * 返回 { toolCalls, usage? } 时说明 AI 想要调用工具，
-   * 返回 { content, usage? } 时说明 AI 给出了最终回答。
-   */
-  async chatWithTools(request, tools) {
-    try {
-      const openaiTools = tools.map((t) => ({
-        type: "function",
-        function: {
-          name: t.name,
-          description: t.description,
-          parameters: {
-            type: "object",
-            properties: Object.fromEntries(
-              Object.entries(t.parameters).map(([key, schema]) => [
-                key,
-                schemaToJsonSchema(schema)
-              ])
-            ),
-            required: Object.entries(t.parameters).filter(([, s]) => s.required).map(([k]) => k)
-          }
-        }
-      }));
-      const baseMessages = this.buildMessages(request);
-      const extraMessages = request._extraMessages ?? [];
-      const allMessages = [...baseMessages, ...extraMessages];
-      const response = await this.client.chat.completions.create({
-        model: request.model,
-        messages: allMessages,
-        tools: openaiTools,
-        tool_choice: "auto",
-        temperature: request.temperature,
-        max_tokens: request.maxTokens,
-        stream: false,
-        ...request.thinking ? { thinking: { type: "enabled" } } : {}
-      }, {
-        timeout: request.timeout ?? this.defaultTimeout
-      });
-      const firstChoice = response.choices?.[0];
-      if (!firstChoice) {
-        return { content: "", usage: void 0 };
-      }
-      const message = firstChoice.message;
-      const finishReason = firstChoice.finish_reason;
-      const usage = toUsage(response.usage);
-      const contentStr = typeof message.content === "string" ? message.content : "";
-      const hasToolCalls = !!(message.tool_calls && message.tool_calls.length > 0);
-      const reasoningContent = message.reasoning_content;
-      if (message.tool_calls && message.tool_calls.length > 0) {
-        const toolCalls = message.tool_calls.map((tc) => {
-          const parsedArgs = repairToolCallArguments(
-            tc.function.arguments || "{}",
-            (m) => process.stderr.write(`[warn] ${m}
-`)
-          );
-          return {
-            id: tc.id,
-            name: tc.function.name,
-            arguments: parsedArgs
-          };
-        });
-        return { toolCalls, usage, reasoningContent };
-      }
+  const tail = thinkFilter.flush();
+  if (tail && !aborted) {
+    textParts.push(tail);
+    hooks.onText?.(tail);
+  }
+  const textContent = textParts.join("");
+  if (aborted) {
+    return { textContent, toolCalls: [], usage, rawContent, reasoningContent, finishReason, aborted };
+  }
+  const toolCalls = [];
+  for (const [, acc] of accumulators) {
+    toolCalls.push({
+      id: acc.id,
+      name: acc.name,
+      arguments: repairToolCallArguments(acc.arguments, hooks.onWarn)
+    });
+  }
+  if (toolCalls.length > 0) {
+    if (rawContent) {
+      toolCalls._rawContent = rawContent;
+    }
+    if (textContent) {
+      toolCalls._streamedText = textContent;
+    }
+  }
+  return { textContent, toolCalls, usage, rawContent, reasoningContent, finishReason, aborted };
+}
+var FREE_ROUND_TOOLS = /* @__PURE__ */ new Set(["write_todos"]);
+var MAX_CONSECUTIVE_FREE_ROUNDS = 3;
+var FreeRoundTracker = class {
+  consecutive = 0;
+  /** 返回 true 表示本轮不消耗有效轮次（调用方执行 round--）。 */
+  apply(toolNames) {
+    const allFree = toolNames.length > 0 && toolNames.every((n) => FREE_ROUND_TOOLS.has(n));
+    if (!allFree) {
+      this.consecutive = 0;
+      return false;
+    }
+    this.consecutive++;
+    return this.consecutive <= MAX_CONSECUTIVE_FREE_ROUNDS;
+  }
+};
+var BudgetWarner = class {
+  constructor(maxToolRounds) {
+    this.maxToolRounds = maxToolRounds;
+    this.noteAt = Math.max(10, Math.floor(maxToolRounds * 0.2));
+    const lowRaw = Math.max(5, Math.floor(maxToolRounds * 0.1));
+    const criticalRaw = Math.max(3, Math.floor(maxToolRounds * 0.05));
+    this.lowAt = Math.min(lowRaw, this.noteAt - 1);
+    this.criticalAt = Math.min(criticalRaw, this.lowAt - 1);
+  }
+  noteAt;
+  lowAt;
+  criticalAt;
+  warnedNote = false;
+  warnedLow = false;
+  warnedCritical = false;
+  check(roundsLeft) {
+    if (!this.warnedCritical && roundsLeft <= this.criticalAt) {
+      this.warnedCritical = true;
       return {
-        content: message.content ?? "",
-        usage,
-        ...reasoningContent ? { reasoningContent } : {},
-        ...!hasToolCalls && (finishReason ?? "") ? { finishReason } : {}
+        level: "critical",
+        injectMessage: `\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.`,
+        displayMessage: `\u{1F6A8} Critical: ${roundsLeft} rounds remaining`
+      };
+    }
+    if (!this.warnedLow && roundsLeft <= this.lowAt) {
+      this.warnedLow = true;
+      return {
+        level: "low",
+        injectMessage: `\u26A0\uFE0F Budget warning: Only ${roundsLeft} tool rounds remaining. Prioritize completing the most critical task. Use efficient approaches (batch edits, fewer reads). If you cannot finish everything, summarize what's done and what remains.`,
+        displayMessage: `\u26A0\uFE0F Low budget: ${roundsLeft} rounds remaining`
+      };
+    }
+    if (!this.warnedNote && roundsLeft <= this.noteAt) {
+      this.warnedNote = true;
+      return {
+        level: "note",
+        injectMessage: `\u{1F4CA} Budget note: ${roundsLeft} tool rounds remaining out of ${this.maxToolRounds}. Plan your remaining work efficiently \u2014 use batch operations (e.g., replaceAll) when possible.`
       };
-    } catch (err) {
-      throw this.wrapError(err);
     }
+    return null;
   }
-  /**
-   * 流式工具调用：文本内容实时输出、工具名称/参数逐块发射。
-   * 子类（DeepSeek / Kimi）因虚假声明检测需要完整响应，故不继承此方法。
-   */
-  async *chatWithToolsStream(request, tools) {
-    if (!this.enableStreamingToolCalls) {
-      const result = await this.chatWithTools(request, tools);
-      if ("toolCalls" in result) {
-        for (let i = 0; i < result.toolCalls.length; i++) {
-          const tc = result.toolCalls[i];
-          yield { type: "tool_call_start", index: i, id: tc.id, name: tc.name };
-          yield { type: "tool_call_delta", index: i, argumentsDelta: JSON.stringify(tc.arguments) };
-          yield { type: "tool_call_end", index: i };
-        }
-      } else {
-        yield { type: "text_delta", delta: result.content };
-      }
-      const rc = "reasoningContent" in result ? result.reasoningContent : void 0;
-      const fr = "finishReason" in result ? result.finishReason : void 0;
-      yield {
-        type: "done",
-        usage: result.usage,
-        ...rc ? { reasoningContent: rc } : {},
-        ...fr ? { finishReason: fr } : {}
+};
+var EMPTY_RESPONSE_NUDGE = "Your previous response was empty \u2014 no text and no tool calls. This usually means the context window is nearly full. Please either: (1) continue the task by calling the next tool you need, or (2) give a concise final text summary of what has been accomplished so far and what remains. Do NOT repeat earlier long outputs.";
+function describeFinishReason(fr) {
+  if (fr === "length") return "output limit reached (finish_reason=length)";
+  if (fr === "content_filter") return "content blocked (finish_reason=content_filter)";
+  if (fr) return `empty response (finish_reason=${fr})`;
+  return "empty response";
+}
+function emptyResponseHint(fr) {
+  if (fr === "length") return "Output token limit hit \u2014 try /compact to reduce context, raise maxTokens, or /model to switch.";
+  if (fr === "content_filter") return "Content was blocked by the provider filter.";
+  return "Context window may be exhausted or max_tokens too low.";
+}
+var EmptyResponseGuard = class {
+  retries = 0;
+  onEmpty(canRetry, finishReason) {
+    if (this.retries === 0 && canRetry) {
+      this.retries++;
+      return {
+        action: "nudge",
+        injectMessage: EMPTY_RESPONSE_NUDGE,
+        displayMessage: `\u26A0 ${describeFinishReason(finishReason)} \u2014 nudging AI to continue...`
       };
-      return;
     }
-    const openaiTools = tools.map((t) => ({
-      type: "function",
-      function: {
-        name: t.name,
-        description: t.description,
-        parameters: {
-          type: "object",
-          properties: Object.fromEntries(
-            Object.entries(t.parameters).map(([key, schema]) => [
-              key,
-              schemaToJsonSchema(schema)
-            ])
-          ),
-          required: Object.entries(t.parameters).filter(([, s]) => s.required).map(([k]) => k)
+    return {
+      action: "stop",
+      displayMessage: "\u26A0 AI returned empty responses twice in a row. Stopping agentic loop.",
+      hint: emptyResponseHint(finishReason)
+    };
+  }
+  /** 非空响应到达 → 重置计数（下次空响应仍可 nudge 一次）。 */
+  onNonEmpty() {
+    this.retries = 0;
+  }
+};
+var ContextPressureMonitor = class {
+  warned80 = false;
+  check(requestTokens, contextWindow) {
+    if (contextWindow <= 0) return { action: "ok", ratio: 0 };
+    const ratio = requestTokens / contextWindow;
+    if (ratio >= 0.95) return { action: "abort", ratio };
+    if (ratio >= 0.8 && !this.warned80) {
+      this.warned80 = true;
+      return {
+        action: "warn",
+        ratio,
+        injectMessage: `\u26A0\uFE0F Context pressure: ~${Math.round(ratio * 100)}% of the ${contextWindow.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.`
+      };
+    }
+    return { action: "ok", ratio };
+  }
+};
+function accumulateUsage(total, delta) {
+  if (!delta) return;
+  total.inputTokens += delta.inputTokens;
+  total.outputTokens += delta.outputTokens;
+  total.cacheCreationTokens += delta.cacheCreationTokens ?? 0;
+  total.cacheReadTokens += delta.cacheReadTokens ?? 0;
+}
+function buildRoundBudgetHint(opts) {
+  const pauseHint = opts.autoPauseInterval > 0 ? `
+- Every ${opts.autoPauseInterval} rounds the user will be asked whether to continue \u2014 use this as a natural checkpoint to report progress.` : "";
+  if (opts.planMode) {
+    return `
+
+[Tool Round Budget \u2014 Plan Mode]
+You have a maximum of ${opts.maxToolRounds} tool call rounds. You are in READ-ONLY Plan Mode:
+- Only use: read_file, list_dir, grep_files, glob_files, ask_user, write_todos
+- Do NOT attempt to call bash, write_file, edit_file \u2014 they are disabled
+- Do NOT write shell commands or code blocks as a substitute for tool calls
+- Do NOT read the same file more than once
+- Call write_todos ONCE to present your plan, then give a text summary
+- If the user asks you to execute anything, respond: "Please type /plan execute to switch to execute mode."${pauseHint}`;
+  }
+  return `
+
+[Tool Round Budget]
+You have a maximum of ${opts.maxToolRounds} tool call rounds for this task. Plan efficiently:
+- Prefer batch operations (e.g. global find-and-replace) over repetitive single edits.
+- Do NOT read the same file more than once \u2014 use the content from previous reads.
+- Prioritize the most critical tasks first in case rounds run out.
+- When remaining rounds are low, focus on completing the current task and summarizing.${pauseHint}`;
+}
+function buildRoundsExhaustedPrompt(maxToolRounds) {
+  return `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 (e.g. send another request to continue)`;
+}
+function buildUserStopMessage(effectiveRound, maxToolRounds) {
+  return `The user has stopped the task at round ${effectiveRound}/${maxToolRounds}. Do not call any more tools. Summarize what has been completed and what remains.`;
+}
+async function runAgentLoop(host) {
+  const { maxToolRounds, autoPauseInterval, usage } = host;
+  const extraMessages = [];
+  const budgetWarner = new BudgetWarner(maxToolRounds);
+  const emptyGuard = new EmptyResponseGuard();
+  const ctxMonitor = new ContextPressureMonitor();
+  const freeRounds = new FreeRoundTracker();
+  const roundToolHistory = [];
+  const hasWriteTools = host.toolDefs.some((t) => t.name === "write_file" || t.name === "edit_file");
+  for (let round = 0; round < maxToolRounds; round++) {
+    if (host.signal?.aborted) return { reason: "aborted", usage };
+    host.onRoundStart?.(round, maxToolRounds);
+    if (await host.beforeRound?.(round, extraMessages) === "stop") {
+      return { reason: "stopped", usage };
+    }
+    const warning = budgetWarner.check(maxToolRounds - round);
+    if (warning) {
+      extraMessages.push({ role: "user", content: warning.injectMessage });
+      host.onBudgetWarning?.(warning);
+    }
+    const interjection = host.pollInterjection?.();
+    if (interjection) extraMessages.push({ role: "user", content: interjection });
+    const ctxWindow = host.getContextWindow();
+    if (ctxWindow > 0) {
+      const pressure = ctxMonitor.check(host.estimateRequestTokens(extraMessages), ctxWindow);
+      if (pressure.action === "abort") {
+        host.onContextPressure?.(pressure, ctxWindow, round);
+        return { reason: "context-overflow", usage };
+      } else if (pressure.action === "warn") {
+        host.onContextPressure?.(pressure, ctxWindow, round);
+        extraMessages.push({ role: "user", content: pressure.injectMessage });
+      }
+    }
+    const outcome = await host.callModel(round, extraMessages);
+    if (host.signal?.aborted) return { reason: "aborted", usage };
+    accumulateUsage(usage, outcome.usage);
+    if (outcome.stopLoop) return { reason: "host-stop", usage };
+    const toolCalls = outcome.toolCalls ?? [];
+    if (toolCalls.length === 0) {
+      const content = outcome.content ?? "";
+      const alreadyRendered = !!outcome.alreadyRendered;
+      const alreadyWrote = hadPreviousWriteToolCalls(extraMessages);
+      const coarseHallucination = !host.planMode && hasWriteTools && !alreadyWrote && !!content && detectsHallucinatedFileOp(content);
+      const phantomPaths = (coarseHallucination || alreadyWrote) && !host.planMode && hasWriteTools && content ? findPhantomClaims(content, extraMessages) : [];
+      const bashRanThisTurn = extractBashCommands(extraMessages).length > 0;
+      const coarseShouldFire = coarseHallucination && !bashRanThisTurn;
+      if ((phantomPaths.length > 0 || coarseShouldFire) && round < maxToolRounds - 1) {
+        host.onHallucinationRetry?.({ phantomPaths, round, alreadyRendered });
+        const correctionMsg = phantomPaths.length > 0 ? buildPhantomCorrectionMessage(phantomPaths) : HALLUCINATION_CORRECTION_MESSAGE;
+        const reasoningField = outcome.reasoningContent ? { reasoning_content: outcome.reasoningContent } : host.providerId === "deepseek" ? { reasoning_content: "" } : {};
+        extraMessages.push(
+          { role: "assistant", content, ...reasoningField },
+          { role: "user", content: correctionMsg }
+        );
+        continue;
+      }
+      if (!content || content.trim() === "") {
+        const decision = emptyGuard.onEmpty(round < maxToolRounds - 1, outcome.finishReason);
+        host.onEmptyResponse?.(decision, { alreadyRendered, round });
+        if (decision.action === "nudge") {
+          extraMessages.push({ role: "user", content: decision.injectMessage });
+          continue;
         }
+        return { reason: "empty-response", usage };
+      }
+      emptyGuard.onNonEmpty();
+      await host.onFinalContent(content, { reasoningContent: outcome.reasoningContent, alreadyRendered });
+      return { reason: "final", usage };
+    }
+    emptyGuard.onNonEmpty();
+    const saveCall = toolCalls.find((tc) => tc.name === "save_last_response");
+    const savePath = saveCall ? String(saveCall.arguments["path"] ?? "") : "";
+    if (saveCall && savePath && host.runSaveLastResponseTee) {
+      const directive = await host.runSaveLastResponseTee({
+        toolCalls,
+        call: saveCall,
+        saveToFile: savePath,
+        extraMessages,
+        reasoningContent: outcome.reasoningContent
+      });
+      if (directive === "stop") return { reason: "tee-stop", usage };
+      freeRounds.apply(toolCalls.map((tc) => tc.name));
+      continue;
+    }
+    const toolResults = await host.executeTools(toolCalls, extraMessages);
+    if (host.isInterrupted?.() || host.signal?.aborted) {
+      host.onInterrupted?.();
+      return { reason: "aborted", usage };
+    }
+    roundToolHistory.push({ round: round + 1, tools: toolCalls.map((tc) => tc.name) });
+    host.onToolsExecuted?.(toolCalls, toolResults, extraMessages);
+    extraMessages.push(...host.buildToolResultMessages(toolCalls, toolResults, outcome.reasoningContent));
+    for (const tc of toolCalls) {
+      if (tc.name.startsWith("mcp__")) host.onMcpToolUsed?.(tc.name);
+    }
+    host.persistRound?.(toolCalls, toolResults, {
+      assistantContent: outcome.content,
+      reasoningContent: outcome.reasoningContent
+    });
+    host.afterToolRoundPersist?.(toolCalls, toolResults, extraMessages);
+    if (freeRounds.apply(toolCalls.map((tc) => tc.name))) {
+      round--;
+    }
+    if (host.checkLoopHealth?.(toolCalls, extraMessages) === "skip-checkpoint") continue;
+    const postInterjection = host.pollInterjection?.();
+    if (postInterjection) extraMessages.push({ role: "user", content: postInterjection });
+    const effectiveRound = round + 1;
+    const remaining = maxToolRounds - effectiveRound;
+    if (autoPauseInterval > 0 && effectiveRound > 0 && effectiveRound % autoPauseInterval === 0 && remaining > 0 && !host.signal?.aborted && host.requestAutoPause) {
+      const toolSummary = summarizeRecentTools(roundToolHistory, autoPauseInterval);
+      const resp = await host.requestAutoPause({ effectiveRound, maxToolRounds, remaining, toolSummary });
+      if (host.signal?.aborted) return { reason: "aborted", usage };
+      if (resp.action === "stop") {
+        extraMessages.push({ role: "user", content: buildUserStopMessage(effectiveRound, maxToolRounds) });
+        break;
+      } else if (resp.action === "redirect" && resp.message) {
+        extraMessages.push({ role: "user", content: resp.message });
       }
-    }));
-    const baseMessages = this.buildMessages(request);
-    const extraMessages = request._extraMessages ?? [];
-    const allMessages = [...baseMessages, ...extraMessages];
+    }
+    host.onRoundEnd?.(round);
+  }
+  let summaryContent = null;
+  try {
+    const summaryExtra = [
+      ...extraMessages,
+      { role: "user", content: buildRoundsExhaustedPrompt(maxToolRounds) }
+    ];
+    const summary = await host.callSummary(summaryExtra);
+    accumulateUsage(usage, summary.usage);
+    summaryContent = summary.content && summary.content.trim() ? summary.content : null;
+  } catch {
+    summaryContent = null;
+  }
+  await host.onRoundsExhausted(summaryContent);
+  return { reason: "rounds-exhausted", usage };
+}
+function summarizeRecentTools(history, interval) {
+  const recent = history.slice(-interval);
+  const counts = /* @__PURE__ */ new Map();
+  for (const rh of recent) {
+    for (const t of rh.tools) counts.set(t, (counts.get(t) || 0) + 1);
+  }
+  return [...counts.entries()].sort((a, b) => b[1] - a[1]).map(([name, count]) => count > 1 ? `${name}\xD7${count}` : name).join(", ");
+}
+
+// src/providers/openai-compatible.ts
+function toUsage(u) {
+  if (!u) return void 0;
+  const cached = u.prompt_tokens_details?.cached_tokens ?? 0;
+  const usage = {
+    inputTokens: Math.max(0, u.prompt_tokens - cached),
+    outputTokens: u.completion_tokens
+  };
+  if (cached > 0) usage.cacheReadTokens = cached;
+  return usage;
+}
+var OpenAICompatibleProvider = class extends BaseProvider {
+  client;
+  defaultTimeout = 6e4;
+  // ms
+  /** 子类设为 false 可禁用流式工具调用（虚假声明检测需要完整响应） */
+  enableStreamingToolCalls = true;
+  async initialize(apiKey, options) {
+    if (options?.timeout !== void 0) {
+      this.defaultTimeout = options.timeout;
+    }
+    const clientOptions = {
+      apiKey,
+      baseURL: options?.baseUrl ?? this.defaultBaseUrl,
+      timeout: this.defaultTimeout
+    };
+    const proxyUrl = options?.proxy;
     try {
-      const stream = await this.client.chat.completions.create({
+      const { Agent: Agent2, ProxyAgent, fetch: undiciFetch } = await import("undici");
+      const STREAM_BODY_TIMEOUT = 30 * 60 * 1e3;
+      const STREAM_HEADERS_TIMEOUT = 5 * 60 * 1e3;
+      const dispatcher = proxyUrl ? new ProxyAgent({
+        uri: proxyUrl,
+        bodyTimeout: STREAM_BODY_TIMEOUT,
+        headersTimeout: STREAM_HEADERS_TIMEOUT
+      }) : new Agent2({
+        bodyTimeout: STREAM_BODY_TIMEOUT,
+        headersTimeout: STREAM_HEADERS_TIMEOUT
+      });
+      clientOptions.fetch = ((url, init) => undiciFetch(url, { ...init, dispatcher }));
+    } catch {
+    }
+    this.client = new OpenAI(clientOptions);
+  }
+  /**
+   * 将 systemPrompt + messages 合并为 OpenAI messages 数组（system 消息放首位）。
+   *
+   * v0.4.100+：按原始顺序保留工具消息（assistant.toolCalls 和 role='tool'），
+   * 不再剥离到 _extraMessages 末尾——之前的剥离会让历史工具往返被插到当前用户消息之后，
+   * 导致模型把"过去的工具调用结果"当作"对当前问题的回应"，DeepSeek V4 Flash 上尤其明显
+   * （会复读上一轮的"完成汇总"）。
+   *
+   * DeepSeek V4 thinking 模式：所有 assistant 消息（含带 toolCalls 的）必须有
+   * reasoning_content 字段，缺失则 API 400。
+   */
+  buildMessages(request) {
+    const msgs = [];
+    for (const m of request.messages) {
+      if (m.role === "tool") {
+        if (!m.toolCallId) continue;
+        msgs.push({
+          role: "tool",
+          tool_call_id: m.toolCallId,
+          content: typeof m.content === "string" ? m.content : ""
+        });
+        continue;
+      }
+      if (m.role === "assistant" && m.toolCalls && m.toolCalls.length > 0) {
+        const assistantMsg = {
+          role: "assistant",
+          content: typeof m.content === "string" && m.content ? m.content : null,
+          tool_calls: m.toolCalls.map((tc) => ({
+            id: tc.id,
+            type: "function",
+            function: { name: tc.name, arguments: JSON.stringify(tc.arguments) }
+          })),
+          reasoning_content: m.reasoningContent ?? ""
+        };
+        msgs.push(assistantMsg);
+        continue;
+      }
+      const base = { role: m.role, content: m.content };
+      if (m.role === "assistant") {
+        base.reasoning_content = m.reasoningContent ?? "";
+      }
+      msgs.push(base);
+    }
+    const systemContent = [request.systemPrompt, request.systemPromptVolatile].filter(Boolean).join("\n\n---\n\n");
+    if (systemContent) {
+      return [{ role: "system", content: systemContent }, ...msgs];
+    }
+    return msgs;
+  }
+  async chat(request) {
+    try {
+      const response = await this.client.chat.completions.create({
         model: request.model,
-        messages: allMessages,
-        tools: openaiTools,
-        tool_choice: "auto",
+        messages: this.buildMessages(request),
         temperature: request.temperature,
         max_tokens: request.maxTokens,
-        stream: true,
-        stream_options: { include_usage: true },
+        stream: false,
         ...request.thinking ? { thinking: { type: "enabled" } } : {}
       }, {
-        timeout: request.timeout ?? this.defaultTimeout,
+        timeout: request.timeout ?? this.defaultTimeout
+      });
+      const firstChoice = response.choices?.[0];
+      if (!firstChoice) {
+        return { content: "", model: response.model, usage: void 0 };
+      }
+      return {
+        content: firstChoice.message.content ?? "",
+        model: response.model,
+        usage: toUsage(response.usage)
+      };
+    } catch (err) {
+      throw this.wrapError(err);
+    }
+  }
+  async *chatStream(request) {
+    try {
+      const stream = await this.client.chat.completions.create({
+        model: request.model,
+        messages: this.buildMessages(request),
+        temperature: request.temperature,
+        max_tokens: request.maxTokens,
+        stream: true,
+        // 请求末尾 usage chunk，供 token 统计使用
+        stream_options: { include_usage: true },
+        ...request.thinking ? { thinking: { type: "enabled" } } : {}
+      }, {
+        timeout: request.timeout ?? this.defaultTimeout,
+        signal: request.signal
+      });
+      let thinkingStarted = false;
+      let reasoningAccumulator = "";
+      for await (const chunk of stream) {
+        const choice = chunk.choices[0];
+        const done = choice?.finish_reason != null;
+        if (!choice && chunk.usage) {
+          yield {
+            delta: "",
+            done: true,
+            usage: toUsage(chunk.usage),
+            ...reasoningAccumulator ? { reasoningContent: reasoningAccumulator } : {}
+          };
+          continue;
+        }
+        const reasoningDelta = choice?.delta?.reasoning_content;
+        if (reasoningDelta) {
+          if (!thinkingStarted) {
+            yield { delta: "<think>", done: false };
+            thinkingStarted = true;
+          }
+          reasoningAccumulator += reasoningDelta;
+          yield { delta: reasoningDelta, done: false };
+          continue;
+        }
+        const delta = choice?.delta?.content ?? "";
+        if (thinkingStarted && delta) {
+          thinkingStarted = false;
+          yield { delta: "</think>", done: false };
+        }
+        if (done) {
+          yield { delta, done, ...reasoningAccumulator ? { reasoningContent: reasoningAccumulator } : {} };
+        } else {
+          yield { delta, done };
+        }
+      }
+    } catch (err) {
+      throw this.wrapError(err);
+    }
+  }
+  /**
+   * 请求 AI 并获取工具调用列表（不执行，只解析）。
+   * 返回 { toolCalls, usage? } 时说明 AI 想要调用工具，
+   * 返回 { content, usage? } 时说明 AI 给出了最终回答。
+   */
+  async chatWithTools(request, tools) {
+    try {
+      const openaiTools = tools.map((t) => ({
+        type: "function",
+        function: {
+          name: t.name,
+          description: t.description,
+          parameters: {
+            type: "object",
+            properties: Object.fromEntries(
+              Object.entries(t.parameters).map(([key, schema]) => [
+                key,
+                schemaToJsonSchema(schema)
+              ])
+            ),
+            required: Object.entries(t.parameters).filter(([, s]) => s.required).map(([k]) => k)
+          }
+        }
+      }));
+      const baseMessages = this.buildMessages(request);
+      const extraMessages = request._extraMessages ?? [];
+      const allMessages = [...baseMessages, ...extraMessages];
+      const response = await this.client.chat.completions.create({
+        model: request.model,
+        messages: allMessages,
+        tools: openaiTools,
+        tool_choice: "auto",
+        temperature: request.temperature,
+        max_tokens: request.maxTokens,
+        stream: false,
+        ...request.thinking ? { thinking: { type: "enabled" } } : {}
+      }, {
+        timeout: request.timeout ?? this.defaultTimeout
+      });
+      const firstChoice = response.choices?.[0];
+      if (!firstChoice) {
+        return { content: "", usage: void 0 };
+      }
+      const message = firstChoice.message;
+      const finishReason = firstChoice.finish_reason;
+      const usage = toUsage(response.usage);
+      const contentStr = typeof message.content === "string" ? message.content : "";
+      const hasToolCalls = !!(message.tool_calls && message.tool_calls.length > 0);
+      const reasoningContent = message.reasoning_content;
+      if (message.tool_calls && message.tool_calls.length > 0) {
+        const toolCalls = message.tool_calls.map((tc) => {
+          const parsedArgs = repairToolCallArguments(
+            tc.function.arguments || "{}",
+            (m) => process.stderr.write(`[warn] ${m}
+`)
+          );
+          return {
+            id: tc.id,
+            name: tc.function.name,
+            arguments: parsedArgs
+          };
+        });
+        return { toolCalls, usage, reasoningContent };
+      }
+      return {
+        content: message.content ?? "",
+        usage,
+        ...reasoningContent ? { reasoningContent } : {},
+        ...!hasToolCalls && (finishReason ?? "") ? { finishReason } : {}
+      };
+    } catch (err) {
+      throw this.wrapError(err);
+    }
+  }
+  /**
+   * 流式工具调用：文本内容实时输出、工具名称/参数逐块发射。
+   * 子类（DeepSeek / Kimi）因虚假声明检测需要完整响应，故不继承此方法。
+   */
+  async *chatWithToolsStream(request, tools) {
+    if (!this.enableStreamingToolCalls) {
+      const result = await this.chatWithTools(request, tools);
+      if ("toolCalls" in result) {
+        for (let i = 0; i < result.toolCalls.length; i++) {
+          const tc = result.toolCalls[i];
+          yield { type: "tool_call_start", index: i, id: tc.id, name: tc.name };
+          yield { type: "tool_call_delta", index: i, argumentsDelta: JSON.stringify(tc.arguments) };
+          yield { type: "tool_call_end", index: i };
+        }
+      } else {
+        yield { type: "text_delta", delta: result.content };
+      }
+      const rc = "reasoningContent" in result ? result.reasoningContent : void 0;
+      const fr = "finishReason" in result ? result.finishReason : void 0;
+      yield {
+        type: "done",
+        usage: result.usage,
+        ...rc ? { reasoningContent: rc } : {},
+        ...fr ? { finishReason: fr } : {}
+      };
+      return;
+    }
+    const openaiTools = tools.map((t) => ({
+      type: "function",
+      function: {
+        name: t.name,
+        description: t.description,
+        parameters: {
+          type: "object",
+          properties: Object.fromEntries(
+            Object.entries(t.parameters).map(([key, schema]) => [
+              key,
+              schemaToJsonSchema(schema)
+            ])
+          ),
+          required: Object.entries(t.parameters).filter(([, s]) => s.required).map(([k]) => k)
+        }
+      }
+    }));
+    const baseMessages = this.buildMessages(request);
+    const extraMessages = request._extraMessages ?? [];
+    const allMessages = [...baseMessages, ...extraMessages];
+    try {
+      const stream = await this.client.chat.completions.create({
+        model: request.model,
+        messages: allMessages,
+        tools: openaiTools,
+        tool_choice: "auto",
+        temperature: request.temperature,
+        max_tokens: request.maxTokens,
+        stream: true,
+        stream_options: { include_usage: true },
+        ...request.thinking ? { thinking: { type: "enabled" } } : {}
+      }, {
+        timeout: request.timeout ?? this.defaultTimeout,
         signal: request.signal
       });
       const toolCallAccumulators = /* @__PURE__ */ new Map();
@@ -2488,336 +3036,39 @@ var ZhipuProvider = class extends OpenAICompatibleProvider {
       {
         id: "glm-z1-air",
         displayName: "GLM-Z1 Air (Lightweight Reasoning)",
-        contextWindow: 131072,
-        supportsStreaming: true,
-        supportsThinking: true
-      },
-      {
-        id: "glm-z1-flash",
-        displayName: "GLM-Z1 Flash (Free Reasoning)",
-        contextWindow: 128e3,
-        supportsStreaming: true,
-        supportsThinking: true
-      },
-      // ── GLM-4 系列（稳定，价格低） ──
-      {
-        id: "glm-4-plus",
-        displayName: "GLM-4 Plus",
-        contextWindow: 128e3,
-        supportsStreaming: true
-      },
-      {
-        id: "glm-4-air",
-        displayName: "GLM-4 Air",
-        contextWindow: 128e3,
-        supportsStreaming: true
-      },
-      {
-        id: "glm-4-flash",
-        displayName: "GLM-4 Flash (Free)",
-        contextWindow: 128e3,
-        supportsStreaming: true
-      }
-    ]
-  };
-};
-
-// src/tools/hallucination.ts
-var HALLUCINATION_PATTERNS = [
-  /文件路径[：:]\s*`?[^\s`]+\.\w{1,5}/,
-  // 文件路径: `path/to/file.ext`（要求文件扩展名）
-  /已生成[：:!！]/,
-  // 已生成完成！
-  /已保存到?\s*[`'"]/,
-  // 已保存到 `path`（要求后跟路径引号）
-  /已写入[：:!！]/,
-  // 已写入！
-  /已创建[：:!！]/,
-  // 已创建！
-  /File\s+(?:written|saved|created)\s+(?:to|as|at)/i,
-  // File written to / saved as（要求介词）
-  /生成完成[！!]/,
-  // 生成完成！
-  /✅\s*(?:文件|已[生保写创]|第)\S*\.\w{1,5}/,
-  // ✅ 文件已保存 path.ext（要求文件扩展名）
-  /文件已[成功]?创建/,
-  // 文件已成功创建 / 文件已创建
-  /教案已[成功]?[生创保写]/,
-  // 教案已成功生成 / 教案已保存
-  /已成功[保写创生]入?[：:!！\s`'"]/,
-  // 已成功保存 / 已成功写入 / 已成功创建
-  /保存[到至]了?\s*[`'"]/,
-  // 保存到了 `path` / 保存至 'path'
-  /内容如下[：:]/,
-  // 内容如下：（后跟大段文件内容）
-  /以下是.*(?:教案|文件|内容)[：:]/
-  // 以下是xx教案内容：（Kimi 常见模式）
-];
-function detectsHallucinatedFileOp(content) {
-  return HALLUCINATION_PATTERNS.some((pattern) => pattern.test(content));
-}
-function hadPreviousWriteToolCalls(extraMessages) {
-  const msgs = extraMessages;
-  return msgs.some((msg) => {
-    if (msg.role === "assistant" && Array.isArray(msg.tool_calls)) {
-      return msg.tool_calls.some((tc) => {
-        const fn = tc.function;
-        const name = fn?.name ?? "";
-        return name === "write_file" || name === "edit_file";
-      });
-    }
-    if (msg.role === "assistant" && Array.isArray(msg.content)) {
-      return msg.content.some((block) => {
-        if (block.type !== "tool_use") return false;
-        const name = block.name ?? "";
-        return name === "write_file" || name === "edit_file";
-      });
-    }
-    if (msg.role === "model" && Array.isArray(msg.parts)) {
-      return msg.parts.some((part) => {
-        const fc = part.functionCall;
-        const name = fc?.name ?? "";
-        return name === "write_file" || name === "edit_file";
-      });
-    }
-    return false;
-  });
-}
-var TOOL_CALL_REMINDER = `
-
-[\u26A0\uFE0F Mandatory Tool Call Policy]
-When you need to create, write, or modify files, you MUST use the function calling API to invoke write_file or edit_file.
-NEVER claim "file saved", "file created", "written to", etc. in your response text without actually calling the tool.
-Describing file content in text without calling the tool = the file does not exist = task failure.
-If multiple files need to be generated, you MUST call write_file separately for each file \u2014 do not skip any.
-Do NOT output fake "completion summaries" unless you have actually completed all file writes via tool_calls.
-
-CRITICAL \u2014 Batch file generation rules:
-1. You MUST call write_file once per file. There are NO shortcuts.
-2. After writing file N, immediately proceed to call write_file for file N+1. Do NOT stop to summarize.
-3. If you find yourself typing file content into your response text instead of into a write_file call, STOP and use the tool.
-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.";
-function extractBashCommands(extraMessages) {
-  const cmds = [];
-  const msgs = extraMessages;
-  const addCmd = (raw) => {
-    if (typeof raw === "string") {
-      try {
-        const parsed = JSON.parse(raw);
-        if (typeof parsed.command === "string") cmds.push(parsed.command);
-      } catch {
-      }
-    } else if (raw && typeof raw === "object") {
-      const c = raw.command;
-      if (typeof c === "string") cmds.push(c);
-    }
-  };
-  for (const msg of msgs) {
-    if (msg.role === "assistant" && Array.isArray(msg.tool_calls)) {
-      for (const tc of msg.tool_calls) {
-        const fn = tc.function;
-        if (fn?.name === "bash") addCmd(fn?.arguments);
-      }
-    }
-    if (msg.role === "assistant" && Array.isArray(msg.content)) {
-      for (const block of msg.content) {
-        if (block.type === "tool_use" && block.name === "bash") addCmd(block.input);
-      }
-    }
-    if (msg.role === "model" && Array.isArray(msg.parts)) {
-      for (const part of msg.parts) {
-        const fc = part.functionCall;
-        if (fc && fc.name === "bash") addCmd(fc.args);
-      }
-    }
-  }
-  return cmds;
-}
-var DSML_PIPE_CLASS = "[|\\uFF5C\\u2502\\u2503\\u01C0]";
-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,
-  // 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,
-  // v0.4.173: DeepSeek V4 DSML pseudo-tool-call markup leaked as text. DeepSeek
-  // emits a fake tool call using its native special-token markup
-  //   <｜｜DSML｜｜tool_calls> <｜｜DSML｜｜invoke name="write"> <｜｜DSML｜｜parameter …>
-  // where the "pipe" is U+FF5C FULLWIDTH VERTICAL LINE (the same token family as
-  // <｜User｜>/<｜Assistant｜>). We saw an exam paper saved via save_last_response
-  // whose tee stream was preamble + this DSML wrapper + the real document body.
-  // The earlier <invoke …> pattern uses ASCII < > and does NOT match these.
-  new RegExp(`<\\/?\\s*${DSML_PIPE_CLASS}+\\s*DSML\\s*${DSML_PIPE_CLASS}+`, "i")
-];
-function detectPseudoToolCalls(content) {
-  if (!content || content.length === 0) return null;
-  for (const re of PSEUDO_TOOL_CALL_PATTERNS) {
-    if (re.test(content)) return re.source;
-  }
-  return null;
-}
-function stripPseudoToolCalls(content) {
-  if (!content) return content;
-  let out = content;
-  const dsmlBody = extractDsmlContent(out);
-  if (dsmlBody !== null) {
-    out = dsmlBody;
-  } else {
-    out = stripDsmlTags(out);
-  }
-  out = out.replace(/<tool_call\b[^>]*>[\s\S]*?<\/tool_call>/gi, "");
-  out = out.replace(/<tool_call\b[^>]*\/>/gi, "");
-  out = out.replace(/<function_calls\b[^>]*>[\s\S]*?<\/function_calls>/gi, "");
-  out = out.replace(/<invoke\b[^>]*>[\s\S]*?<\/invoke>/gi, "");
-  out = out.replace(/<invoke\b[^>]*\/>/gi, "");
-  out = out.replace(/<tool_use(?:_id)?\b[^>]*>[\s\S]*?<\/tool_use(?:_id)?>/gi, "");
-  out = out.replace(/```\s*tool_call\b[\s\S]*?```/gi, "");
-  out = out.replace(/<think\b[^>]*>[\s\S]*?<\/think>/gi, "");
-  out = out.replace(/<think\b[^>]*>[\s\S]*?(?=^#{1,3}\s+\S|\n\s*\n)/im, "");
-  out = out.replace(/^\s*\{\s*"name"\s*:\s*"[\w._-]+"\s*,\s*"arguments"\s*:[\s\S]*?\}\s*$/gm, "");
-  out = unwrapDocumentFence(out);
-  out = peelMetaNarration(out);
-  out = out.replace(/\n{3,}/g, "\n\n").trim();
-  return out;
-}
-function extractDsmlContent(content) {
-  if (!content) return null;
-  const P = DSML_PIPE_CLASS;
-  const re = new RegExp(
-    `<\\s*${P}+\\s*DSML\\s*${P}+\\s*parameter\\b[^>]*\\bname\\s*=\\s*["']content["'][^>]*>([\\s\\S]*?)<\\s*/\\s*${P}+\\s*DSML\\s*${P}+\\s*parameter\\s*>`,
-    "i"
-  );
-  const m = content.match(re);
-  if (m && typeof m[1] === "string") {
-    const body = m[1].trim();
-    return body.length > 0 ? body : null;
-  }
-  return null;
-}
-function stripDsmlTags(content) {
-  const P = DSML_PIPE_CLASS;
-  let out = content;
-  out = out.replace(
-    new RegExp(
-      `<\\s*${P}+\\s*DSML\\s*${P}+\\s*tool_calls\\b[\\s\\S]*?<\\s*/\\s*${P}+\\s*DSML\\s*${P}+\\s*tool_calls\\s*>`,
-      "gi"
-    ),
-    ""
-  );
-  out = out.replace(new RegExp(`<\\s*/?\\s*${P}+\\s*DSML\\s*${P}+[^>]*>`, "gi"), "");
-  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);
+        contextWindow: 131072,
+        supportsStreaming: true,
+        supportsThinking: true
+      },
+      {
+        id: "glm-z1-flash",
+        displayName: "GLM-Z1 Flash (Free Reasoning)",
+        contextWindow: 128e3,
+        supportsStreaming: true,
+        supportsThinking: true
+      },
+      // ── GLM-4 系列（稳定，价格低） ──
+      {
+        id: "glm-4-plus",
+        displayName: "GLM-4 Plus",
+        contextWindow: 128e3,
+        supportsStreaming: true
+      },
+      {
+        id: "glm-4-air",
+        displayName: "GLM-4 Air",
+        contextWindow: 128e3,
+        supportsStreaming: true
+      },
+      {
+        id: "glm-4-flash",
+        displayName: "GLM-4 Flash (Free)",
+        contextWindow: 128e3,
+        supportsStreaming: true
       }
-    }
-  }
-  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();
-}
-var META_NARRATION_HARD_MARKERS = [
-  /\[⚠️\s*CONTENT GENERATION MODE\]/,
-  /CONTENT_ONLY_STREAM_REMINDER\b/,
-  /<system-reminder>/i
-];
-var META_NARRATION_HEURISTICS = [
-  /\bthe user (?:is asking me|wants me|is requesting|expects me)\b/i,
-  /\blet me (?:re-?read|re-?consider|reconsider|think about|carefully (?:re-?read|consider))\b/i,
-  /\bI'?m (?:in (?:a )?content-only|in CONTENT-ONLY|currently in)\b/i,
-  /\bI think (?:there might be|I should|I cannot|the (?:user|best)|maybe)\b/i,
-  /\bWait,?\s+let me\b/i,
-  /\bActually,?\s+I\b/i,
-  /\bI need to be honest with the user\b/i,
-  /\bI(?:'m| am) in a special mode\b/i,
-  /\bGiven that I cannot\b/i
-];
-function detectMetaNarration(content) {
-  if (!content) return null;
-  const head = content.slice(0, 2e3);
-  for (const re of META_NARRATION_HARD_MARKERS) {
-    if (re.test(head)) return re.source;
-  }
-  if (/^#{1,3}\s+\S/m.test(head)) return null;
-  let hits = 0;
-  let firstMatch = "";
-  for (const re of META_NARRATION_HEURISTICS) {
-    if (re.test(head)) {
-      hits++;
-      if (!firstMatch) firstMatch = re.source;
-      if (hits >= 2) return `meta-narration:${firstMatch}`;
-    }
-  }
-  return null;
-}
-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 = `
@@ -12384,7 +12635,6 @@ var SessionHandler = class _SessionHandler {
   async handleChatWithTools(provider, messages, toolDefs, mcpBudgetNote) {
     const session = this.sessions.current;
     const apiMessages = [...messages];
-    const extraMessages = [];
     const maxToolRounds = this.config.get("maxToolRounds") ?? DEFAULT_MAX_TOOL_ROUNDS;
     const autoPauseIntervalRaw = this.config.get("autoPauseInterval");
     const autoPauseInterval = typeof autoPauseIntervalRaw === "number" ? autoPauseIntervalRaw : 50;
@@ -12395,37 +12645,113 @@ var SessionHandler = class _SessionHandler {
 ${mcpBudgetNote}` : "");
     const systemPromptVolatile = toolVolatile;
     const modelParams = this.getModelParams();
-    const roundUsage = { inputTokens: 0, outputTokens: 0, cacheCreationTokens: 0, cacheReadTokens: 0 };
+    const usage = { inputTokens: 0, outputTokens: 0, cacheCreationTokens: 0, cacheReadTokens: 0 };
     const supportsStreamingTools = typeof provider.chatWithToolsStream === "function";
-    const roundToolHistory = [];
-    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 {
-      for (let round = 0; round < maxToolRounds; round++) {
-        if (ac.signal.aborted) break;
-        this.toolExecutor.setRoundInfo(round + 1, maxToolRounds);
-        this.send({ type: "round_progress", current: round + 1, total: maxToolRounds });
-        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 });
+      const loopResult = await runAgentLoop({
+        maxToolRounds,
+        autoPauseInterval,
+        planMode: this.planMode,
+        providerId: this.currentProvider,
+        toolDefs,
+        signal: ac.signal,
+        usage,
+        callModel: async (_round, extraMessages) => {
+          const chatRequest = {
+            messages: apiMessages,
+            model: this.currentModel,
+            systemPrompt,
+            systemPromptVolatile,
+            stream: false,
+            temperature: modelParams.temperature,
+            maxTokens: modelParams.maxTokens,
+            timeout: modelParams.timeout,
+            thinking: modelParams.thinking,
+            thinkingBudget: modelParams.thinkingBudget,
+            signal: ac.signal,
+            ...extraMessages.length > 0 ? { _extraMessages: extraMessages } : {}
+          };
+          try {
+            if (supportsStreamingTools) {
+              const streamGen = provider.chatWithToolsStream(chatRequest, toolDefs);
+              return await this.consumeToolStream(streamGen, ac);
+            }
+            const result = await provider.chatWithTools(chatRequest, toolDefs);
+            return result;
+          } catch (providerErr) {
+            const errMsg = providerErr instanceof Error ? providerErr.message : String(providerErr);
+            const isCtxLengthError = /maximum context length|context_length_exceeded|context window|too many tokens|reduce the length of the messages/i.test(errMsg);
+            if (isCtxLengthError) {
+              this.send({
+                type: "response_done",
+                content: `\u26A0 Context length exceeded \u2014 the conversation is too long for this model.
+
+Details: ${errMsg.split("\n")[0]}
+
+**Recovery options**:
+1. Run \`/compact\` to summarize old messages and free context
+2. Run \`/clear\` to start a fresh session
+3. Run \`/model\` to switch to a model with a larger context window`,
+                usage
+              });
+              return { stopLoop: true };
+            }
+            throw providerErr;
           }
-        }
-        if (this.userInterjection) {
+        },
+        callSummary: async (summaryExtra) => {
+          const summaryResult = await provider.chatWithTools(
+            {
+              messages: apiMessages,
+              model: this.currentModel,
+              systemPrompt,
+              systemPromptVolatile,
+              stream: false,
+              temperature: modelParams.temperature,
+              maxTokens: modelParams.maxTokens,
+              timeout: modelParams.timeout,
+              _extraMessages: summaryExtra
+            },
+            []
+          );
+          return "content" in summaryResult ? { content: summaryResult.content, usage: summaryResult.usage } : { usage: summaryResult.usage };
+        },
+        executeTools: async (toolCalls) => {
+          googleSearchContext.configManager = this.config;
+          spawnAgentContext.provider = provider;
+          spawnAgentContext.model = this.currentModel;
+          spawnAgentContext.systemPrompt = systemPromptVolatile ? `${systemPrompt}
+
+---
+
+${systemPromptVolatile}` : systemPrompt;
+          spawnAgentContext.modelParams = modelParams;
+          spawnAgentContext.configManager = this.config;
+          ToolExecutor.currentMessageIndex = this.sessions.current?.messages.length ?? 0;
+          return this.toolExecutor.executeAll(toolCalls);
+        },
+        buildToolResultMessages: (toolCalls, results, reasoningContent) => provider.buildToolResultMessages(toolCalls, results, reasoningContent),
+        getContextWindow: () => this.getContextWindowSize(),
+        estimateRequestTokens: (extraMessages) => this.estimateRequestTokens(systemPrompt, extraMessages),
+        pollInterjection: () => {
+          if (!this.userInterjection) return null;
           const msg = this.userInterjection;
           this.userInterjection = null;
           this.send({ type: "info", message: `\u26A1 Interjection: "${msg}"` });
-          extraMessages.push({ role: "user", content: msg });
-        }
-        const ctxWindow = this.getContextWindowSize();
-        if (ctxWindow > 0) {
-          const reqTokens = this.estimateRequestTokens(systemPrompt, extraMessages);
-          const pressure = ctxMonitor.check(reqTokens, ctxWindow);
+          return msg;
+        },
+        onRoundStart: (round, total) => {
+          this.toolExecutor.setRoundInfo(round + 1, total);
+          this.send({ type: "round_progress", current: round + 1, total });
+        },
+        onBudgetWarning: (warning) => {
+          if (warning.displayMessage) {
+            this.send({ type: "info", message: warning.displayMessage });
+          }
+        },
+        onContextPressure: (pressure, ctxWindow) => {
           if (pressure.action === "abort") {
             this.send({
               type: "response_done",
@@ -12437,191 +12763,48 @@ Too much tool output accumulated this turn. Your work so far is preserved.
 1. Run \`/compact\` to shrink history, then ask the AI to continue
 2. Run \`/clear\` to start fresh
 3. Switch to a larger-context model`,
-              usage: roundUsage
+              usage
             });
-            this.addWebSessionUsage(roundUsage);
-            session.addTokenUsage(roundUsage);
-            return;
-          } else if (pressure.action === "warn") {
+          } else {
             this.send({
               type: "info",
               message: `\u26A0 Context at ${Math.round(pressure.ratio * 100)}% \u2014 asking AI to wrap up`
             });
-            extraMessages.push({ role: "user", content: pressure.injectMessage });
           }
-        }
-        const chatRequest = {
-          messages: apiMessages,
-          model: this.currentModel,
-          systemPrompt,
-          systemPromptVolatile,
-          stream: false,
-          temperature: modelParams.temperature,
-          maxTokens: modelParams.maxTokens,
-          timeout: modelParams.timeout,
-          thinking: modelParams.thinking,
-          thinkingBudget: modelParams.thinkingBudget,
-          signal: ac.signal,
-          ...extraMessages.length > 0 ? { _extraMessages: extraMessages } : {}
-        };
-        let result;
-        try {
-          if (supportsStreamingTools) {
-            const streamGen = provider.chatWithToolsStream(chatRequest, toolDefs);
-            result = await this.consumeToolStream(streamGen, ac);
+        },
+        onHallucinationRetry: ({ phantomPaths }) => {
+          const detail = phantomPaths.length > 0 ? ` (phantom files: ${phantomPaths.join(", ")})` : "";
+          this.send({ type: "info", message: `\u26A0 Hallucinated completion detected, forcing retry...${detail}` });
+        },
+        onEmptyResponse: (decision) => {
+          if (decision.action === "nudge") {
+            this.send({ type: "info", message: decision.displayMessage });
           } else {
-            result = await provider.chatWithTools(chatRequest, toolDefs);
-          }
-        } catch (providerErr) {
-          const errMsg = providerErr instanceof Error ? providerErr.message : String(providerErr);
-          const isCtxLengthError = /maximum context length|context_length_exceeded|context window|too many tokens|reduce the length of the messages/i.test(errMsg);
-          if (isCtxLengthError) {
             this.send({
               type: "response_done",
-              content: `\u26A0 Context length exceeded \u2014 the conversation is too long for this model.
-
-Details: ${errMsg.split("\n")[0]}
-
-**Recovery options**:
-1. Run \`/compact\` to summarize old messages and free context
-2. Run \`/clear\` to start a fresh session
-3. Run \`/model\` to switch to a model with a larger context window`,
-              usage: roundUsage
-            });
-            this.addWebSessionUsage(roundUsage);
-            session.addTokenUsage(roundUsage);
-            return;
-          }
-          throw providerErr;
-        }
-        if (ac.signal.aborted) break;
-        accumulateUsage(roundUsage, result.usage);
-        const hasToolCalls = !!(result.toolCalls && result.toolCalls.length > 0);
-        const contentBlank = !result.content || result.content.trim() === "";
-        if (!hasToolCalls && contentBlank) {
-          const decision = emptyGuard.onEmpty(round < maxToolRounds - 1, result.finishReason);
-          if (decision.action === "nudge") {
-            this.send({ type: "info", message: decision.displayMessage });
-            extraMessages.push({ role: "user", content: decision.injectMessage });
-            continue;
-          }
-          this.send({
-            type: "response_done",
-            content: `${decision.displayMessage}
+              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;
-        }
-        emptyGuard.onNonEmpty();
-        if (result.content && !result.toolCalls) {
-          const hasWriteTools = toolDefs.some((t) => t.name === "write_file" || t.name === "edit_file");
-          const alreadyWrote = hadPreviousWriteToolCalls(extraMessages);
-          const bashRanThisTurn = extractBashCommands(extraMessages).length > 0;
-          if (hasWriteTools && !alreadyWrote && !bashRanThisTurn && detectsHallucinatedFileOp(result.content) && round < maxToolRounds - 1) {
-            this.send({ type: "info", message: "\u26A0 Hallucinated completion detected, forcing retry..." });
-            const reasoningField = result.reasoningContent ? { reasoning_content: result.reasoningContent } : this.currentProvider === "deepseek" ? { reasoning_content: "" } : {};
-            extraMessages.push(
-              { role: "assistant", content: result.content, ...reasoningField },
-              { role: "user", content: HALLUCINATION_CORRECTION_MESSAGE }
-            );
-            continue;
+              usage
+            });
           }
-          this.send({ type: "response_done", content: result.content, usage: roundUsage });
+        },
+        onFinalContent: (content, { reasoningContent }) => {
+          this.send({ type: "response_done", content, usage });
           session.addMessage({
             role: "assistant",
-            content: result.content,
+            content,
             timestamp: /* @__PURE__ */ new Date(),
-            ...result.reasoningContent ? { reasoningContent: result.reasoningContent } : {}
-          });
-          this.addWebSessionUsage(roundUsage);
-          session.addTokenUsage(roundUsage);
-          return;
-        }
-        if (result.toolCalls && result.toolCalls.length > 0) {
-          roundToolHistory.push({
-            round: round + 1,
-            tools: result.toolCalls.map((tc) => tc.name)
-          });
-          for (const tc of result.toolCalls) {
-            if (tc.name.startsWith("mcp__")) this.usedMcpToolNames.add(tc.name);
-          }
-          googleSearchContext.configManager = this.config;
-          spawnAgentContext.provider = provider;
-          spawnAgentContext.model = this.currentModel;
-          spawnAgentContext.systemPrompt = systemPromptVolatile ? `${systemPrompt}
-
----
-
-${systemPromptVolatile}` : systemPrompt;
-          spawnAgentContext.modelParams = modelParams;
-          spawnAgentContext.configManager = this.config;
-          ToolExecutor.currentMessageIndex = this.sessions.current?.messages.length ?? 0;
-          const saveLastResponseCall = result.toolCalls.find((tc) => tc.name === "save_last_response");
-          const saveLastResponsePath = saveLastResponseCall ? String(saveLastResponseCall.arguments["path"] ?? "") : "";
-          if (saveLastResponseCall && saveLastResponsePath) {
-            const teeResult = await this.runSaveLastResponseTee(
-              provider,
-              saveLastResponseCall,
-              saveLastResponsePath,
-              apiMessages,
-              extraMessages,
-              systemPrompt,
-              systemPromptVolatile,
-              modelParams,
-              ac,
-              roundUsage
-            );
-            const teeToolResults = result.toolCalls.map((tc) => {
-              if (tc.id === saveLastResponseCall.id) {
-                return {
-                  callId: tc.id,
-                  content: teeResult.summary,
-                  isError: teeResult.isError
-                };
-              }
-              return {
-                callId: tc.id,
-                content: "[skipped: file already saved by tee streaming]",
-                isError: false
-              };
-            });
-            const reasoningContent2 = result.reasoningContent;
-            const newMsgs2 = provider.buildToolResultMessages(result.toolCalls, teeToolResults, reasoningContent2);
-            extraMessages.push(...newMsgs2);
-            persistToolRound(session, result.toolCalls, teeToolResults, {
-              assistantContent: teeResult.content,
-              reasoningContent: reasoningContent2
-            });
-            freeRounds.apply(result.toolCalls.map((tc) => tc.name));
-            continue;
-          }
-          const toolResults = await this.toolExecutor.executeAll(result.toolCalls);
-          const reasoningContent = result.reasoningContent;
-          const newMsgs = provider.buildToolResultMessages(result.toolCalls, toolResults, reasoningContent);
-          extraMessages.push(...newMsgs);
-          persistToolRound(session, result.toolCalls, toolResults, {
-            assistantContent: result.content,
-            reasoningContent
+            ...reasoningContent ? { reasoningContent } : {}
           });
-          if (freeRounds.apply(result.toolCalls.map((tc) => tc.name))) {
-            round--;
-          }
-          if (this.userInterjection) {
-            const msg = this.userInterjection;
-            this.userInterjection = null;
-            this.send({ type: "info", message: `\u26A1 Interjection: "${msg}"` });
-            extraMessages.push({ role: "user", content: msg });
-          }
-        }
-        const effectiveRound = round + 1;
-        const remaining = maxToolRounds - effectiveRound;
-        if (autoPauseInterval > 0 && effectiveRound > 0 && effectiveRound % autoPauseInterval === 0 && remaining > 0 && !ac.signal.aborted) {
-          const toolSummary = summarizeRecentTools(roundToolHistory, autoPauseInterval);
+        },
+        persistRound: (toolCalls, results, info) => {
+          persistToolRound(session, toolCalls, results, info);
+        },
+        // Track MCP tool usage for next-turn budget prioritization (C1)
+        onMcpToolUsed: (name) => this.usedMcpToolNames.add(name),
+        requestAutoPause: async ({ effectiveRound, maxToolRounds: totalRounds, toolSummary }) => {
           const requestId = `pause_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`;
           const pauseResp = await new Promise((resolve7) => {
             this.pendingAutoPause.set(requestId, resolve7);
@@ -12629,58 +12812,85 @@ ${systemPromptVolatile}` : systemPrompt;
               type: "auto_pause_request",
               requestId,
               currentRound: effectiveRound,
-              totalRounds: maxToolRounds,
+              totalRounds,
               toolSummary
             });
           });
-          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: buildUserStopMessage(effectiveRound, maxToolRounds) });
-            break;
+            this.send({ type: "info", message: `\u23F8  Stopped by user at ${effectiveRound}/${totalRounds}` });
           } else if (pauseResp.action === "redirect" && pauseResp.message) {
             this.send({ type: "info", message: `\u26A1 Redirect: "${pauseResp.message}"` });
-            extraMessages.push({ role: "user", content: pauseResp.message });
           }
-        }
-      }
-      try {
-        const summaryExtra = [
-          ...extraMessages,
-          { role: "user", content: buildRoundsExhaustedPrompt(maxToolRounds) }
-        ];
-        const summaryResult = await provider.chatWithTools(
-          {
-            messages: apiMessages,
-            model: this.currentModel,
+          return pauseResp;
+        },
+        onRoundsExhausted: (summaryContent) => {
+          if (summaryContent !== null) {
+            this.send({
+              type: "response_done",
+              content: `\u26A0 Reached maximum tool call rounds (${maxToolRounds}).
+
+${summaryContent}`,
+              usage
+            });
+            session.addMessage({ role: "assistant", content: summaryContent, timestamp: /* @__PURE__ */ new Date() });
+          } else {
+            this.send({
+              type: "error",
+              message: `Reached maximum tool call rounds (${maxToolRounds}). You can continue by asking the AI to proceed.`
+            });
+          }
+        },
+        // ─── save_last_response tee-streaming（v0.4.102+）─────────────────
+        // AI 在 Web 模式下调用 save_last_response 时，复刻 REPL 的 tee 流式路径：
+        // 发起一次新的 chatStream → 实时通过 WS 推送文本 + 同步写盘 → 注入合成
+        // 工具结果。否则该工具会落到默认 executor，读到空的 lastResponseStore
+        // 直接报错；用户被迫退到 write_file，又因 tool_call arguments 截断
+        // (~2KB) 只能写出片段，再用 edit_file 反复 insert 才能补全（v0.4.101 报告）。
+        // 与 REPL 不同：Web 端 tee 成功后继续 agentic 循环（返回 'continue'），
+        // 让模型基于工具结果给出最终文本。
+        runSaveLastResponseTee: async ({ toolCalls, call, saveToFile, extraMessages, reasoningContent }) => {
+          const teeResult = await this.runSaveLastResponseTee(
+            provider,
+            call,
+            saveToFile,
+            apiMessages,
+            extraMessages,
             systemPrompt,
             systemPromptVolatile,
-            stream: false,
-            temperature: modelParams.temperature,
-            maxTokens: modelParams.maxTokens,
-            timeout: modelParams.timeout,
-            _extraMessages: summaryExtra
-          },
-          []
-        );
-        if ("content" in summaryResult && summaryResult.content) {
-          this.send({
-            type: "response_done",
-            content: `\u26A0 Reached maximum tool call rounds (${maxToolRounds}).
-
-${summaryResult.content}`,
-            usage: roundUsage
+            modelParams,
+            ac,
+            usage
+          );
+          const teeToolResults = toolCalls.map((tc) => {
+            if (tc.id === call.id) {
+              return {
+                callId: tc.id,
+                content: teeResult.summary,
+                isError: teeResult.isError
+              };
+            }
+            return {
+              callId: tc.id,
+              content: "[skipped: file already saved by tee streaming]",
+              isError: false
+            };
+          });
+          const newMsgs = provider.buildToolResultMessages(toolCalls, teeToolResults, reasoningContent);
+          extraMessages.push(...newMsgs);
+          persistToolRound(session, toolCalls, teeToolResults, {
+            assistantContent: teeResult.content,
+            reasoningContent
           });
-          session.addMessage({ role: "assistant", content: summaryResult.content, timestamp: /* @__PURE__ */ new Date() });
+          return "continue";
         }
-      } catch {
-        this.send({
-          type: "error",
-          message: `Reached maximum tool call rounds (${maxToolRounds}). You can continue by asking the AI to proceed.`
-        });
+      });
+      if (loopResult.reason !== "tee-stop") {
+        this.addWebSessionUsage(usage);
+        session.addTokenUsage(usage);
+      }
+      if (loopResult.reason === "aborted") {
+        this.send({ type: "info", message: "[interrupted]" });
       }
-      this.addWebSessionUsage(roundUsage);
-      session.addTokenUsage(roundUsage);
     } catch (err) {
       if (err.name === "AbortError") {
         this.send({ type: "info", message: "[interrupted]" });
@@ -13866,7 +14076,7 @@ ${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-625NA546.js");
+          const { executeTests } = await import("./run-tests-NWRKAVFS.js");
           const argStr = args.join(" ").trim();
           let testArgs = {};
           if (argStr) {