From 1c037db0c23c6f9120d0be248054e9c94ce1348b Mon Sep 17 00:00:00 2001
From: pulsareonbot <pulsareon@qq.com>
Date: Sat, 4 Jul 2026 17:18:58 +0800
Subject: [PATCH] =?UTF-8?q?fix:=20=E8=A7=A3=E6=9E=90=20llama-cli=20?=
 =?UTF-8?q?=E8=BE=93=E5=87=BA,=E5=8F=AA=E6=8F=90=E5=8F=96=E7=94=9F?=
 =?UTF-8?q?=E6=88=90=E5=9B=9E=E5=A4=8D=20(=E5=8E=BB=E9=99=A4=20banner/?=
 =?UTF-8?q?=E8=BF=9B=E5=BA=A6=E5=8A=A8=E7=94=BB/=E7=BB=9F=E8=AE=A1)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 src/plugins/ai/backend.rs | 98 +++++++++++++++++++++++++++++++++++++--
 1 file changed, 95 insertions(+), 3 deletions(-)

diff --git a/src/plugins/ai/backend.rs b/src/plugins/ai/backend.rs
index e7feffd..e8085fd 100644
--- a/src/plugins/ai/backend.rs
+++ b/src/plugins/ai/backend.rs
@@ -194,9 +194,18 @@ impl AiBackend for LocalCliBackend {
             );
         }
 
-        // llama-cli 输出到 stdout，取生成部分
-        let reply = String::from_utf8_lossy(&output.stdout);
-        Ok(reply.trim().to_string())
+        // llama-cli --single-turn 输出格式：
+        //   [banner/logo]
+        //   > <user_input>      ← 用户输入回显
+        //   [进度动画 |/-\]
+        //   <生成的回复>        ← 我们要提取的部分
+        //   [空行]
+        //   [ Prompt: ... | Generation: ... ]  ← 性能统计
+        //
+        // 解析策略：找到 "> " 开头的行，跳过进度动画行，取实际回复行
+        let raw = String::from_utf8_lossy(&output.stdout);
+        let reply = parse_llama_output(&raw);
+        Ok(reply)
     }
 
     fn tts(&self, text: &str, model_path: &Path, output_path: &Path) -> Result<()> {
@@ -275,3 +284,86 @@ impl AiBackend for CloudBackend {
         "cloud"
     }
 }
+
+/// 解析 llama-cli --single-turn 的 stdout 输出，提取生成的回复文本
+///
+/// 输出格式：
+/// ```text
+/// [banner/logo ASCII art]
+/// build      : ...
+/// model      : ...
+/// ...
+/// > <user_input>
+///
+/// [进度动画 |/-\]
+/// <生成的回复>
+///
+/// [ Prompt: ... | Generation: ... ]
+///
+/// Exiting...
+/// ```
+fn parse_llama_output(raw: &str) -> String {
+    let lines: Vec<&str> = raw.lines().collect();
+    let mut reply_lines = Vec::new();
+    let mut after_user_input = false;
+
+    for line in &lines {
+        let trimmed = line.trim();
+
+        // 跳过空行（但在回复区内保留）
+        if trimmed.is_empty() && !after_user_input {
+            continue;
+        }
+
+        // 找到用户输入回显行 "> xxx"
+        if trimmed.starts_with('>') {
+            after_user_input = true;
+            continue;
+        }
+
+        if !after_user_input {
+            continue;
+        }
+
+        // 跳过进度动画行（只含 |/-\ 和空格）
+        let cleaned: String = trimmed.chars().filter(|c| !"|\\-/ ".contains(*c)).collect();
+        if cleaned.is_empty() {
+            continue;
+        }
+
+        // 遇到性能统计行 [ Prompt: ... ] 停止
+        if trimmed.starts_with('[') && trimmed.contains("Generation") {
+            break;
+        }
+        // 遇到 "Exiting..." 停止
+        if trimmed == "Exiting..." {
+            break;
+        }
+
+        reply_lines.push(trimmed);
+    }
+
+    reply_lines.join("\n").trim().to_string()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_parse_llama_output() {
+        let raw = "  █  █  build      : b1-fdb1db8\n\
+model      : model_store/llm/qwen2.5-0.5b-q4_k_m.gguf\n\
+\n\
+> hello\n\
+\n\
+|/-\\|/-\\ 汪！\n\
+\n\
+[ Prompt: 31.1 t/s | Generation: 26.3 t/s ]\n\
+\n\
+Exiting...\n";
+
+        let reply = parse_llama_output(raw);
+        assert_eq!(reply, "汪！");
+    }
+}