根据提供的code differences信息，由于没有具体的代码变更内容，我将生成一个通用的commit message模板：

```
docs(readme): 更新文档说明

- 添加了项目使用指南
- 完善了API接口说明
- 修正了一些文字错误
```

注：由于未提供具体的代码差异信息，以上为示例格式。请提供具体的代码变更内容以便生成准确的commit message。

康康/Services/HealthExportService.swift

@@ -155,7 +155,7 @@ struct HealthExportService {
                         // 用「全文累计 + 每 chunk 重清 + diff yield」:
                         // - thinking 阶段,UI 看到的 generated 始终为空
                         // - 看到 </think> 后,真实内容流式出现
-                        var rawAccum = ""
+                        var stripper = ThinkStripper()
                         let stream = await AIRuntime.shared.generate(
                             prompt: genPrompt,
                             maxTokens: 1024
@@ -163,21 +163,15 @@ struct HealthExportService {
                         for try await chunk in stream {
                             try Task.checkCancellation()
                             if chunk.decodeRate > 0 { lastRate = chunk.decodeRate }
-                            rawAccum += chunk.text
-                            let clean = Self.stripThinkBlocks(rawAccum)
-                            if clean.count > generated.count, clean.hasPrefix(generated) {
-                                let delta = String(clean.dropFirst(generated.count))
-                                generated = clean
+                            let delta = stripper.feed(chunk.text)
+                            if !delta.isEmpty {
                                 continuation.yield(.token(TokenChunk(
                                     text: delta,
                                     decodeRate: chunk.decodeRate
                                 )))
-                            } else if clean != generated {
-                                // 极少:清理后比上次还短(模型补了开标签)。让 UI 不要回退,
-                                // 直接对齐 generated = clean 但不 yield(避免显示倒退)。
-                                generated = clean
                             }
                         }
+                        generated = stripper.output
                     }
 
                     guard !generated.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty else {
@@ -250,21 +244,16 @@ struct HealthExportService {
                         dataJSON: dataJSON
                     )
 
-                    var displayed = ""
-                    var rawAccum = ""
+                    var stripper = ThinkStripper()
                     let stream = await AIRuntime.shared.generate(prompt: prompt, maxTokens: 480)
                     for try await chunk in stream {
                         try Task.checkCancellation()
-                        rawAccum += chunk.text
-                        let clean = Self.stripThinkBlocks(rawAccum)
-                        if clean.count > displayed.count, clean.hasPrefix(displayed) {
-                            let delta = String(clean.dropFirst(displayed.count))
-                            displayed = clean
+                        let delta = stripper.feed(chunk.text)
+                        if !delta.isEmpty {
                             continuation.yield(.token(TokenChunk(text: delta, decodeRate: chunk.decodeRate)))
-                        } else if clean != displayed {
-                            displayed = clean
                         }
                     }
+                    let displayed = stripper.output
 
                     guard !displayed.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty else {
                         throw ServiceError.generationFailed("模型未输出任何内容")
@@ -307,23 +296,18 @@ struct HealthExportService {
                         dataJSON: dataJSON
                     )
 
-                    var generated = ""
-                    var rawAccum = ""
                     var lastRate: Double = 0
+                    var stripper = ThinkStripper()
                     let stream = await AIRuntime.shared.generate(prompt: genPrompt, maxTokens: 1200)
                     for try await chunk in stream {
                         try Task.checkCancellation()
                         if chunk.decodeRate > 0 { lastRate = chunk.decodeRate }
-                        rawAccum += chunk.text
-                        let clean = Self.stripThinkBlocks(rawAccum)
-                        if clean.count > generated.count, clean.hasPrefix(generated) {
-                            let delta = String(clean.dropFirst(generated.count))
-                            generated = clean
+                        let delta = stripper.feed(chunk.text)
+                        if !delta.isEmpty {
                             continuation.yield(.token(TokenChunk(text: delta, decodeRate: chunk.decodeRate)))
-                        } else if clean != generated {
-                            generated = clean
                         }
                     }
+                    var generated = stripper.output
 
                     guard !generated.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty else {
                         throw ServiceError.generationFailed("模型未输出任何内容")
@@ -786,3 +770,38 @@ struct HealthExportService {
         return s
     }
 }
+
+/// 流式 `<think>` 去除器:逐 chunk 喂入,返回这次应增量 yield 的 delta。
+///
+/// 旧实现每个 token 都对「整段累计文本」重跑 `stripThinkBlocks` + `count`/`hasPrefix`/`dropFirst`,
+/// 全是 O(n) grapheme 操作,1024/1200 token 长报告随长度二次方增长(且都在 MainActor 上)。
+/// 这里一旦思考段闭合(出现 `</think>`)或确定不存在(首个非空字符不是 `<`,Qwen 思考必在最前),
+/// 就切到纯增量拼接,把生成主体阶段的每 token 成本降到 O(1)。最坏情况退化为旧行为,无正确性风险。
+private struct ThinkStripper {
+    private var rawAccum = ""
+    private(set) var output = ""
+    private var resolved = false
+
+    mutating func feed(_ piece: String) -> String {
+        rawAccum += piece
+        if resolved {
+            output += piece        // 快路径:思考段已解析,直接增量
+            return piece
+        }
+        let clean = HealthExportService.stripThinkBlocks(rawAccum)
+        var delta = ""
+        if clean.count > output.count, clean.hasPrefix(output) {
+            delta = String(clean.dropFirst(output.count))
+            output = clean
+        } else if clean != output {
+            output = clean         // 清理后比上次短(模型补了开标签):对齐但不回退显示
+        }
+        // 判定能否对「后续」token 切快路径(本 token 已由上面的完整清理处理)。
+        if rawAccum.contains("</think>") {
+            resolved = true                                                   // 思考段已闭合,其后纯增量
+        } else if let c = rawAccum.first(where: { !$0.isWhitespace }), c != "<" {
+            resolved = true                                                   // 开头非 '<' → 不存在 <think>
+        }
+        return delta
+    }
+}