Merge pull request #623 from H-Chris233/fix/crash-guards-llm-endpoint-url-asr-model-validation

fix: crash guards for LLM endpoint URL + ASR model loading

openless-all/app/src-tauri/src/asr/local/test_run.rs

@@ -47,6 +47,43 @@ pub async fn run_test(model_id: ModelId) -> Result<TestResult> {
         anyhow::bail!("模型目录不存在：{}（请先下载）", dir.display());
     }
 
+    // ── 模型文件完整性检查 ────────────────────────────────────────────
+    // 在调 C FFI 之前先检查关键文件是否齐全、尺寸是否合理，避免因下载不完整
+    // 或文件损坏导致 C 端 qwen_load / qwen_transcribe_audio segfault 杀死进程。
+    // 需要的文件清单见 QwenAsrEngine::load() 注释。
+    let required_files = ["config.json", "vocab.json", "merges.txt"];
+    for fname in &required_files {
+        let path = dir.join(fname);
+        if !path.exists() {
+            anyhow::bail!("模型文件缺失：{fname}，请重新下载（预期路径：{}）", path.display());
+        }
+        let meta = std::fs::metadata(&path)
+            .map_err(|e| anyhow::anyhow!("读取 {fname} 元数据失败：{e}"))?;
+        if meta.len() == 0 {
+            anyhow::bail!("模型文件为空：{fname}，请重新下载");
+        }
+    }
+    // safetensors 可能是单文件 model.safetensors 或分片 model-00001-of-NNNN.safetensors
+    let has_safetensors: Vec<_> = std::fs::read_dir(&dir)
+        .map_err(|e| anyhow::anyhow!("读取模型目录失败：{e}"))?
+        .filter_map(|entry| entry.ok())
+        .filter(|e| e.path().extension().is_some_and(|ext| ext == "safetensors"))
+        .collect();
+    if has_safetensors.is_empty() {
+        anyhow::bail!("模型目录中没有 .safetensors 权重文件，请重新下载");
+    }
+    for entry in &has_safetensors {
+        let meta = std::fs::metadata(entry.path())
+            .map_err(|e| anyhow::anyhow!("读取 {} 元数据失败：{e}", entry.path().display()))?;
+        if meta.len() < 1024 {
+            anyhow::bail!(
+                "权重文件太小（{} bytes）：{}，请重新下载",
+                meta.len(),
+                entry.path().display()
+            );
+        }
+    }
+
     let samples = decode_wav_16k_mono(TEST_WAV)?;
     let audio_ms = (samples.len() as u64) * 1000 / 16_000;
 

openless-all/app/src-tauri/src/coordinator/llm_pipeline.rs

@@ -135,12 +135,19 @@ where
         );
         return StreamingPolishOutcome::UnsupportedFallback;
     }
-    let provider = match build_active_llm_provider(llm_thinking_enabled) {
-        Ok(p) => p,
-        Err(e) => {
+    let provider = match std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
+        build_active_llm_provider(llm_thinking_enabled)
+    })) {
+        Ok(Ok(p)) => p,
+        Ok(Err(e)) => {
             log::error!("[coord] streaming polish: build provider failed: {e}");
             return StreamingPolishOutcome::Failed(e.to_string());
         }
+        Err(panic) => {
+            let msg = format!("build_active_llm_provider panicked: {:?}", panic);
+            log::error!("[coord] {msg}");
+            return StreamingPolishOutcome::Failed(msg);
+        }
     };
     if !provider.supports_streaming_polish() {
         log::info!(
@@ -259,7 +266,17 @@ pub(crate) async fn polish_text(
             .await?);
     }
 
-    let provider = build_active_llm_provider(llm_thinking_enabled)?;
+    let provider = match std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
+        build_active_llm_provider(llm_thinking_enabled)
+    })) {
+        Ok(Ok(p)) => p,
+        Ok(Err(e)) => return Err(e),
+        Err(panic) => {
+            let msg = format!("build_active_llm_provider panicked: {:?}", panic);
+            log::error!("[coord] {msg}");
+            anyhow::bail!(msg);
+        }
+    };
     Ok(provider
         .polish(
             raw,
@@ -305,7 +322,17 @@ pub(crate) async fn translate_text(
             .await?);
     }
 
-    let provider = build_active_llm_provider(llm_thinking_enabled)?;
+    let provider = match std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
+        build_active_llm_provider(llm_thinking_enabled)
+    })) {
+        Ok(Ok(p)) => p,
+        Ok(Err(e)) => return Err(e),
+        Err(panic) => {
+            let msg = format!("build_active_llm_provider panicked: {:?}", panic);
+            log::error!("[coord] {msg}");
+            anyhow::bail!(msg);
+        }
+    };
     Ok(provider
         .translate_to(
             raw,
@@ -597,7 +624,17 @@ where
             .await?);
     }
 
-    let provider = build_active_llm_provider(llm_thinking_enabled)?;
+    let provider = match std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
+        build_active_llm_provider(llm_thinking_enabled)
+    })) {
+        Ok(Ok(p)) => p,
+        Ok(Err(e)) => return Err(e),
+        Err(panic) => {
+            let msg = format!("build_active_llm_provider panicked: {:?}", panic);
+            log::error!("[coord] {msg}");
+            anyhow::bail!(msg);
+        }
+    };
     Ok(provider
         .answer_chat_streaming(
             messages,
@@ -687,8 +724,17 @@ pub(crate) fn resolve_ark_endpoint_with_policy(
     if api_key.trim().is_empty() && endpoint.is_none() {
         anyhow::bail!("API Key 为空");
     }
-    let resolved = endpoint
+    let mut resolved = endpoint
         .unwrap_or_else(|| "https://ark.cn-beijing.volces.com/api/v3/chat/completions".to_string());
+    // 兜底：用户可能在自定义 endpoint 时只写了 "192.168.1.100:8080/v1" 漏了 scheme，
+    // 被 reqwest::IntoUrl 自动补 https:// 后对纯 HTTP 的 llama.cpp 发请求会 TLS 握手失败。
+    // 检测到无 scheme 时自动补 http://（局域网自托管服务最常见的形态）。
+    // 用 to_ascii_lowercase 防用户输大写如 HTTP://... 时被误判为缺 scheme。
+    let lower = resolved.to_ascii_lowercase();
+    if !lower.starts_with("http://") && !lower.starts_with("https://") {
+        resolved = format!("http://{resolved}");
+        log::info!("[llm] endpoint missing scheme, auto-prepended http://");
+    }
     // issue #609 F-01（SSRF）：用户自定义 endpoint 是 attacker-controlled，直接拿来发
     // 带 API Key 的请求等于把凭据指哪打哪。这里对 host/IP 段 + scheme 做配置时校验，
     // 默认官方 endpoint 也过一遍（它本就合法）。注意这只防住"配置即字面内网地址"，

openless-all/app/src-tauri/src/polish.rs

@@ -1252,9 +1252,16 @@ async fn send_with_transient_retry(
     request: reqwest::RequestBuilder,
 ) -> Result<reqwest::Response, LLMError> {
     const RETRY_DELAY_MS: u64 = 500;
-    let initial = request
-        .try_clone()
-        .expect("memory-backed body (json/form) must be clonable for retry");
+    let Some(initial) = request.try_clone() else {
+        // try_clone 失败（如 stream body 不可 clone）→ 不走重试，直接 send 一次。
+        // 用 expect 会 panic 杀死整个进程，这里兜底为单次发送。
+        log::warn!("[llm] request body not clonable, skipping retry");
+        return match request.send().await {
+            Ok(r) => Ok(r),
+            Err(e) if e.is_timeout() => Err(LLMError::Timeout),
+            Err(e) => Err(LLMError::Network(e.to_string())),
+        };
+    };
     match initial.send().await {
         Ok(r) => Ok(r),
         Err(e) if e.is_connect() || e.is_request() => {