diff --git a/src/utils/textNormalization.ts b/src/utils/textNormalization.ts
index 8064fae0b..4c5600b8c 100644
--- a/src/utils/textNormalization.ts
+++ b/src/utils/textNormalization.ts
@@ -1,3 +1,5 @@
+import { parse as parseToml } from "smol-toml";
+
 /**
  * 将常见的中文/全角/弯引号统一为 ASCII 引号，以避免 TOML 解析失败。
  * - 双引号：” “ „ ‟ ＂ → "
@@ -15,8 +17,139 @@ export const normalizeQuotes = (text: string): string => {
   );
 };
 
+const SMART_DOUBLE_QUOTES = new Set(["“", "”", "„", "‟", "＂"]);
+const SMART_SINGLE_QUOTES = new Set(["‘", "’", "＇"]);
+
+/** 返回字符所属的引号族（ASCII 引号或弯/全角引号），非引号返回 null */
+const quoteFamily = (ch: string): '"' | "'" | null => {
+  if (ch === '"' || SMART_DOUBLE_QUOTES.has(ch)) return '"';
+  if (ch === "'" || SMART_SINGLE_QUOTES.has(ch)) return "'";
+  return null;
+};
+
 /**
- * 专用于 TOML 文本的归一化；目前等同于 normalizeQuotes，后续可扩展（如空白、行尾等）。
+ * 扫描 TOML 文本，把语法位置（字符串定界符）的弯/全角引号换成 ASCII 引号。
+ * - 字符串外的弯引号视为定界符，换成 ASCII 并开启字符串；
+ * - 由弯引号开启的字符串，同族的弯引号或 ASCII 引号都能关闭；
+ * - 由 ASCII 引号开启的字符串，严格模式下只有 ASCII 引号能关闭（里面的弯引号是内容），
+ *   宽松模式（lenient）下同族弯引号也能关闭；
+ * - 严格模式下，ASCII 开启的字符串到行尾（多行字符串到文末）都没关上时，回头按宽松
+ *   规则重扫：编辑器每次输入都会归一化，中文输入法先打出的左引号已变成 ASCII，
+ *   右引号要靠这一步识别；
+ * - 基本字符串里的 `\“` 这类转义不是合法 TOML，按 `\"` 处理；注释原样保留。
  */
-export const normalizeTomlText = (text: string): string =>
-  normalizeQuotes(text);
+const scanTomlQuotes = (text: string, lenient: boolean): string => {
+  let out = "";
+  let i = 0;
+  const n = text.length;
+
+  while (i < n) {
+    const ch = text[i];
+
+    if (ch === "#") {
+      const end = text.indexOf("\n", i);
+      const stop = end === -1 ? n : end;
+      out += text.slice(i, stop);
+      i = stop;
+      continue;
+    }
+
+    const family = quoteFamily(ch);
+    if (!family) {
+      out += ch;
+      i += 1;
+      continue;
+    }
+
+    // 开启字符串：判断是否为三引号（多行字符串）
+    const multiline =
+      i + 2 < n &&
+      quoteFamily(text[i + 1]) === family &&
+      quoteFamily(text[i + 2]) === family;
+    const openLen = multiline ? 3 : 1;
+    const openedBySmart = text.slice(i, i + openLen) !== family.repeat(openLen);
+    const bodyStart = i + openLen;
+    out += family.repeat(openLen);
+
+    const scan = (
+      acceptSmart: boolean,
+    ): { body: string; next: number } | null => {
+      let body = "";
+      let j = bodyStart;
+      const isCloser = (c: string) =>
+        c === family || (acceptSmart && quoteFamily(c) === family);
+
+      while (j < n) {
+        const c = text[j];
+        if (
+          family === '"' &&
+          c === "\\" &&
+          j + 1 < n &&
+          (multiline || text[j + 1] !== "\n")
+        ) {
+          const escaped = text[j + 1];
+          body += "\\" + (quoteFamily(escaped) === '"' ? '"' : escaped);
+          j += 2;
+          continue;
+        }
+        if (!multiline && c === "\n") return null;
+        if (!isCloser(c)) {
+          body += c;
+          j += 1;
+          continue;
+        }
+        if (!multiline) {
+          return { body: body + family, next: j + 1 };
+        }
+        // 多行字符串：连续 3 个及以上的定界符才关闭（多出的最多 2 个属于内容）
+        let run = 0;
+        while (j + run < n && isCloser(text[j + run])) run += 1;
+        if (run >= 3) {
+          return { body: body + family.repeat(run), next: j + run };
+        }
+        body += text.slice(j, j + run);
+        j += run;
+      }
+      return null;
+    };
+
+    const acceptSmart = lenient || openedBySmart;
+    const result = scan(acceptSmart) ?? (acceptSmart ? null : scan(true));
+    if (result) {
+      out += result.body;
+      i = result.next;
+      continue;
+    }
+
+    // 未闭合：原样输出到行尾（多行字符串到文末），交给解析器报错
+    const end = multiline ? n : text.indexOf("\n", bodyStart);
+    const stop = end === -1 ? n : end;
+    out += text.slice(bodyStart, stop);
+    i = stop;
+  }
+
+  return out;
+};
+
+const parsesAsToml = (text: string): boolean => {
+  try {
+    parseToml(text);
+    return true;
+  } catch {
+    return false;
+  }
+};
+
+/**
+ * 专用于 TOML 文本的归一化：只把语法位置的弯/全角引号换成 ASCII 引号，合法字符串
+ * 内部的内容原样保留（#8559）。严格结果解析不通过、宽松结果（字符串里的同族弯引号
+ * 也当作闭合符）能解析时才用宽松结果，所以合法文档里的字符串内容绝不会被改写。
+ */
+export const normalizeTomlText = (text: string): string => {
+  if (!text || !/[“”„‟＂‘’＇]/.test(text)) return text;
+
+  const strict = scanTomlQuotes(text, false);
+  if (parsesAsToml(strict)) return strict;
+  const lenient = scanTomlQuotes(text, true);
+  return lenient !== strict && parsesAsToml(lenient) ? lenient : strict;
+};
diff --git a/tests/utils/textNormalization.test.ts b/tests/utils/textNormalization.test.ts
new file mode 100644
index 000000000..8d45e975b
--- /dev/null
+++ b/tests/utils/textNormalization.test.ts
@@ -0,0 +1,96 @@
+import { parse as parseToml } from "smol-toml";
+import { describe, expect, it } from "vitest";
+import { normalizeTomlText } from "@/utils/textNormalization";
+import { tomlToMcpServer, validateToml } from "@/utils/tomlUtils";
+
+describe("normalizeTomlText", () => {
+  it("normalizes smart quotes used as string delimiters", () => {
+    expect(normalizeTomlText("command = “npx”\nargs = [‘-y’]")).toBe(
+      "command = \"npx\"\nargs = ['-y']",
+    );
+    expect(normalizeTomlText("[mcp_servers.＂demo＂]")).toBe(
+      '[mcp_servers."demo"]',
+    );
+  });
+
+  it("keeps smart quotes inside strings opened by ASCII quotes", () => {
+    const input = [
+      'a = "don’t"',
+      'b = "say“hello”"',
+      "c = 'don’t'",
+      'd = """multi “line” ’text’"""',
+      "e = '''it’s'''",
+      'f = "escaped \\" then “x”"',
+    ].join("\n");
+    expect(normalizeTomlText(input)).toBe(input);
+  });
+
+  it("keeps quotes of the other family inside smart-quoted strings", () => {
+    expect(normalizeTomlText("a = “don’t”")).toBe('a = "don’t"');
+    expect(parseToml(normalizeTomlText("a = “don’t”")).a).toBe("don’t");
+  });
+
+  it("closes an ASCII-opened string with a smart quote when nothing else closes it", () => {
+    // 编辑器逐字归一化后，输入法打出的左引号已是 ASCII，右引号仍是弯引号
+    expect(normalizeTomlText('model = "gpt”\nx = 1')).toBe(
+      'model = "gpt"\nx = 1',
+    );
+    expect(normalizeTomlText("name = 'abc’")).toBe("name = 'abc'");
+  });
+
+  it("repairs smart closers that are followed by more ASCII quotes on the line", () => {
+    // 合法文档不会走到这里：只有严格结果解析失败时才把字符串里的弯引号当闭合符
+    expect(normalizeTomlText('args = ["npx”, "-y"]')).toBe(
+      'args = ["npx", "-y"]',
+    );
+    expect(normalizeTomlText('model = "gpt” # "example"')).toBe(
+      'model = "gpt" # "example"',
+    );
+  });
+
+  it("turns escaped smart double quotes into ASCII escapes", () => {
+    const normalized = normalizeTomlText("args = [“say \\“hello\\””]");
+    expect(normalized).toBe('args = ["say \\"hello\\""]');
+    expect(parseToml(normalized).args).toEqual(['say "hello"']);
+  });
+
+  it("leaves comments untouched", () => {
+    const input = "# “comment” it’s\nkey = “v” # don’t";
+    expect(normalizeTomlText(input)).toBe(
+      '# “comment” it’s\nkey = "v" # don’t',
+    );
+  });
+
+  it("does not touch unterminated strings beyond their opening quote", () => {
+    expect(normalizeTomlText("x = “abc")).toBe('x = "abc');
+  });
+});
+
+describe("tomlUtils with smart quotes in valid strings (#8559)", () => {
+  it("preserves a right single quotation mark in args", () => {
+    const input = 'command = "echo"\nargs = ["don’t"]';
+    expect(validateToml(input)).toBe("");
+    expect(tomlToMcpServer(input).args).toEqual(["don’t"]);
+  });
+
+  it("accepts curly double quotes inside a basic string", () => {
+    const input = 'command = "echo"\nargs = ["say“hello”"]';
+    expect(validateToml(input)).toBe("");
+    expect(tomlToMcpServer(input).args).toEqual(["say“hello”"]);
+  });
+
+  it("accepts curly single quotes inside a literal string", () => {
+    const input = "command = 'echo'\nargs = ['don’t']";
+    expect(validateToml(input)).toBe("");
+    expect(tomlToMcpServer(input).args).toEqual(["don’t"]);
+  });
+
+  it("still accepts configs written entirely with smart quotes", () => {
+    const input = "command = “npx”\nargs = [“-y”, “pkg”]";
+    expect(validateToml(input)).toBe("");
+    expect(tomlToMcpServer(input)).toMatchObject({
+      command: "npx",
+      args: ["-y", "pkg"],
+    });
+  });
+});