984 lines
45 KiB
Rust
984 lines
45 KiB
Rust
//! 流式接收 LLM 响应
|
||
|
||
use std::collections::HashMap;
|
||
use std::sync::Arc;
|
||
use std::sync::atomic::AtomicBool;
|
||
use std::time::Duration;
|
||
|
||
use tauri::{AppHandle, Emitter, Manager};
|
||
use tracing::warn;
|
||
|
||
use df_ai::provider::{CompletionRequest, LlmProvider};
|
||
|
||
use super::{AiChatEvent, ToolCallDraft};
|
||
|
||
/// 从 anyhow 错误中尽力提取 HTTP 状态码/错误分类，供诊断拼接。
|
||
///
|
||
/// `provider.stream()` 失败有两类来源：
|
||
/// - 业务层：provider 在 non-2xx 时 `bail!("LLM 流式 API 错误 {status}: {body}")`，原始文本已含状态码，
|
||
/// 此处用正则从文本抠 `HTTP <code>` 或裸 `<code>`（provider 串已含 status code）。
|
||
/// - 传输层（reqwest，不直接命名以避免给 src-tauri 加依赖）：从 Display 文本识别
|
||
/// timeout/connect 关键词分类。
|
||
///
|
||
/// 返回 `(status_or_class, raw)`：status 取文本中首个三位数；否则按关键词给 timeout/connect/unknown。
|
||
/// 从字符串(非 anyhow)提取诊断信息,用于 mpsc channel 中来自专用线程的错误字符串。
|
||
fn extract_error_diag_from_str(raw: &str) -> (String, String) {
|
||
// 1) 抠 HTTP 状态码
|
||
const HTTP_CODES: &[&str] = &[
|
||
"400", "401", "403", "404", "408", "409", "413",
|
||
"422", "429", "500", "502", "503", "504",
|
||
];
|
||
let chars: Vec<char> = raw.chars().collect();
|
||
let n = chars.len();
|
||
let mut i = 0;
|
||
while i + 3 <= n {
|
||
if chars[i].is_ascii_digit() && chars[i + 1].is_ascii_digit() && chars[i + 2].is_ascii_digit() {
|
||
let code: String = chars[i..i + 3].iter().collect();
|
||
let prev_ok = i == 0 || !chars[i - 1].is_ascii_digit();
|
||
let next_ok = i + 3 == n || !chars[i + 3].is_ascii_digit();
|
||
if prev_ok && next_ok && HTTP_CODES.contains(&code.as_str()) {
|
||
return (format!("HTTP {}", code), raw.to_string());
|
||
}
|
||
i += 3;
|
||
continue;
|
||
}
|
||
i += 1;
|
||
}
|
||
|
||
let lower = raw.to_lowercase();
|
||
let class = if lower.contains("timeout") || lower.contains("超时") {
|
||
"timeout"
|
||
} else if lower.contains("connect") || lower.contains("dns") || lower.contains("resolve") {
|
||
"connect"
|
||
} else if lower.contains("timed out") {
|
||
"timeout"
|
||
} else {
|
||
"unknown"
|
||
};
|
||
(class.to_string(), raw.to_string())
|
||
}
|
||
|
||
/// 从 anyhow::Error 提取诊断信息
|
||
#[allow(dead_code)]
|
||
fn extract_error_diag(e: &anyhow::Error) -> (String, String) {
|
||
extract_error_diag_from_str(&e.to_string())
|
||
}
|
||
|
||
/// AiError 诊断消息的上下文，区分「建连/首字节失败」与「流中途断」两类。
|
||
#[derive(Copy, Clone, Eq, PartialEq, Debug)]
|
||
pub(crate) enum DiagKind {
|
||
/// `provider.stream()` 直接返回 Err：连接/鉴权/HTTP non-2xx 等
|
||
#[allow(dead_code)]
|
||
Init,
|
||
/// 流已建立，next() 返回 Err：SSE 传输断/解析错等
|
||
MidStream,
|
||
}
|
||
|
||
/// 拼接 AiError 的可读诊断文本（纯函数，便于单测）。
|
||
///
|
||
/// 格式：`[<provider_name>] <上下文>(<status_or_class>): <raw>`
|
||
/// - provider_name：`provider.name()`（anthropic 协议为 "anthropic-compat"，
|
||
/// openai 兼容为模型名）。当前 `LlmProvider` trait 未暴露 base_url/provider_type，
|
||
/// 不改签名的前提下这是唯一可得的 provider 标识。
|
||
/// - status_or_class：`HTTP 4xx/5xx` 或 `timeout`/`connect`/`unknown` 分类。
|
||
/// - raw：anyhow 原始错误文本。
|
||
pub(crate) fn fmt_diag(provider_name: &str, kind: DiagKind, status_or_class: &str, raw: &str) -> String {
|
||
let ctx = match kind {
|
||
DiagKind::Init => "AI 调用失败",
|
||
DiagKind::MidStream => "流式接收错误",
|
||
};
|
||
format!("[{}] {}({}): {}", provider_name, ctx, status_or_class, raw)
|
||
}
|
||
|
||
/// 流式接收结果。显式区分三类出口，避免调用方对 None 的歧义推断
|
||
/// (UX-2025-04 / CR-30-1 / CR-30-2 / 决策 F-260616-07 a1)。
|
||
pub(crate) enum StreamResult {
|
||
/// 正常完成（流尽 + finished 信号，或用户主动停止）。incomplete=false。
|
||
/// 调用方按正常流程入库 + emit AiCompleted。
|
||
Complete {
|
||
text: String,
|
||
tool_calls: HashMap<u32, ToolCallDraft>,
|
||
usage: df_ai::provider::TokenUsage,
|
||
/// DeepSeek thinking 模式推理内容（需回传到下一轮请求）
|
||
reasoning_content: Option<String>,
|
||
},
|
||
/// 流中途断（MidStream chunk Err / idle timeout / provider stream-error / 有 partial_text
|
||
/// 但流尽未收到 finished）。已 emit 过 AiTextDelta（前端 currentText 已累积），
|
||
/// **不 emit AiError**——调用方保文入库 + emit AiCompleted(incomplete=true) + 系统提示。
|
||
/// 调用方**不重试**（决策 a1）。
|
||
Partial {
|
||
text: String,
|
||
tool_calls: HashMap<u32, ToolCallDraft>,
|
||
usage: df_ai::provider::TokenUsage,
|
||
/// DeepSeek thinking 模式推理内容（部分累积）
|
||
reasoning_content: Option<String>,
|
||
},
|
||
/// Init 失败（provider.stream() Err：建连/鉴权/HTTP non-2xx,或流建立后空文本各类中断）。
|
||
/// **不再 emit AiError**——重试可恢复路径的 emit 权交调用方(agentic/mod.rs),
|
||
/// 由 agentic 在重试耗尽/Fatal 时统一 emit 最终 AiError(单气泡),避免 N 次重试 push N+1 气泡。
|
||
/// `retryable`: 据状态码分类(retry::is_status_retryable 镜像)：true=5xx/429/timeout/connect
|
||
/// 可重试；false=4xx(非429)/鉴权/参数错 Fatal 立即放弃。调用方据此决定是否重试。
|
||
/// `error`: 已格式化的错误文本(fmt_diag 拼接 provider 标识 + 上下文 + 状态 + 原文),
|
||
/// 供 agentic 重试耗尽/Fatal 时 emit AiError 用,保留原诊断信息不丢失。
|
||
InitFailed { retryable: bool, error: String },
|
||
}
|
||
|
||
/// 将 chunk 携带的 tool_calls delta 累加到累积表。
|
||
///
|
||
/// 每个 `ToolCallDelta` 按 `index` 归位到同一个 `ToolCallDraft`：
|
||
/// - `id` 覆盖(后到的 id 视为权威,匹配 OpenAI/Anthropic 协议行为)
|
||
/// - `function_name` / `function_arguments` 增量拼接(流式分片到达)
|
||
///
|
||
/// 抽取自 stream_llm 的 chunk match arm(原最深 7 层嵌套点),纯状态累加无 emit/return 副作用。
|
||
fn accumulate_tool_calls(
|
||
tc_deltas: &[df_ai::provider::ToolCallDelta],
|
||
tool_calls_acc: &mut HashMap<u32, ToolCallDraft>,
|
||
) {
|
||
for tc_delta in tc_deltas {
|
||
let draft = tool_calls_acc.entry(tc_delta.index).or_default();
|
||
if let Some(id) = &tc_delta.id {
|
||
draft.id = id.clone();
|
||
}
|
||
if let Some(name) = &tc_delta.function_name {
|
||
draft.name.push_str(name);
|
||
}
|
||
if let Some(args) = &tc_delta.function_arguments {
|
||
draft.args.push_str(args);
|
||
}
|
||
}
|
||
}
|
||
|
||
/// 双写 emit: Tauri 事件 + ai_event_bus 总线(跨端透传)。消费 ev。
|
||
fn emit_ai_chat_event(app_handle: &AppHandle, ev: AiChatEvent) {
|
||
let _ = app_handle.emit("ai-chat-event", ev.clone());
|
||
let _ = app_handle.state::<crate::state::AppState>().ai_event_bus.publish_event(ev);
|
||
}
|
||
|
||
/// flush 合批的 AiTextDelta(双写 emit + event_bus)。空 buffer 不 emit。
|
||
fn flush_delta(app_handle: &AppHandle, conv_id: &str, pending_delta: &mut String) {
|
||
if pending_delta.is_empty() {
|
||
return;
|
||
}
|
||
let delta = std::mem::take(pending_delta);
|
||
emit_ai_chat_event(
|
||
app_handle,
|
||
AiChatEvent::AiTextDelta {
|
||
delta,
|
||
conversation_id: Some(conv_id.to_string()),
|
||
},
|
||
);
|
||
}
|
||
|
||
/// 流式接收 LLM 响应。
|
||
///
|
||
/// 三类异常处理（返回 StreamResult 显式区分出口）：
|
||
/// - Init 失败（provider.stream() Err）：UX-260618-15 **不 emit AiError**，返回
|
||
/// InitFailed{retryable, error}。retryable 据状态码分类(retry::is_status_retryable 镜像)：
|
||
/// 5xx/429/timeout/connect=true，4xx(非429)/鉴权/参数错=false Fatal。调用方走流前重试(仅
|
||
/// retryable=true,≤max_retries 次, 复用 retry::backoff_delay 退避 + 30s 总预算),重试耗尽或
|
||
/// Fatal 时由调用方 emit 最终 AiError(error 字段携带诊断文本,单气泡聚合)。
|
||
/// - MidStream 失败（流已建立后断/timeout/err 事件/有 partial_text 但流尽未 finished）：
|
||
/// **不 emit AiError**，返回 Partial{...} 保文。调用方不重试，入库 + AiCompleted(incomplete)
|
||
/// + 系统提示网络中断。空文本仍返回 InitFailed{retryable=true}(无文可保,交重试,emit 权归调用方)。
|
||
/// - 用户停止（stop_flag）：返回 Complete（incomplete 语义非异常中断）。
|
||
///
|
||
/// 心跳与停止响应（B-260615-02 / B-260615-04）：单 `tokio::select!` 三分支
|
||
/// - `stream.next()`：正常 chunk 处理
|
||
/// - `heartbeat.tick()`（30s）：静默期（如工具执行后等下一轮首 chunk）emit `AiHeartbeat`，
|
||
/// 前端 watchdog 据此 reset，区分「LLM 在跑」与「真断」（避免空气泡误报中断）
|
||
/// - `stop_notify.notified()`：用户点停止即时打断，不再等 chunk 到或 120s idle timeout
|
||
pub(crate) async fn stream_llm(
|
||
provider: Arc<dyn LlmProvider>,
|
||
request: CompletionRequest,
|
||
app_handle: &AppHandle,
|
||
stop_flag: &AtomicBool,
|
||
notify: &tokio::sync::Notify,
|
||
conv_id: &str,
|
||
) -> StreamResult {
|
||
/// 流式读取空闲超时:超过此时长无任何 chunk 即判定连接已断
|
||
/// 原 120s 太长,后端挂死时用户等太久。改为 45s,正常流式每 chunk 间隔
|
||
/// 远不到 45s(即使模型思考间隙),超时即判为连接异常走保文/报错路径
|
||
const STREAM_IDLE_TIMEOUT: Duration = Duration::from_secs(45);
|
||
/// 首 chunk 超时(等首字节)。
|
||
/// 正常 provider 首 token <3s(即使 48 工具定义 + 长上下文),10s 足够宽容;
|
||
/// 已有数据后中途静默仍用 STREAM_IDLE_TIMEOUT(120s,允许模型思考间隙)。
|
||
const FIRST_CHUNK_TIMEOUT: Duration = Duration::from_secs(10);
|
||
/// 心跳间隔：静默期向前端报「LLM 仍在跑」，reset watchdog
|
||
const HEARTBEAT_INTERVAL: Duration = Duration::from_secs(30);
|
||
/// 后端 delta 合批窗口。50ms 内的多次 AiTextDelta 合并为一条再 emit,
|
||
/// 降低 IPC 事件风暴(前端 rAF 只节流渲染不节流 IPC);合并后前端 currentText
|
||
/// 逐条累加,最终一致。
|
||
const DELTA_FLUSH_INTERVAL: Duration = Duration::from_millis(50);
|
||
|
||
// ================================================================
|
||
// BUG-2026-07-17 根治: 整个 LLM 流式请求运行在独立 OS 线程的
|
||
// 专用 tokio runtime 上,通过 mpsc channel 将 chunk 流式送回主 runtime。
|
||
// ================================================================
|
||
//
|
||
// 根因: Tauri 在 Windows 上使用单线程 tokio runtime。provider.stream()
|
||
// 内部(eventsource-stream/reqwest TLS/网络建连)存在同步阻塞操作,
|
||
// 阻塞唯一工作线程导致 tokio 计时器系统无法推进,timeout 不触发。
|
||
//
|
||
// 方案:
|
||
// 1. 独立 OS 线程 + 多线程 tokio runtime 执行完整 LLM 请求
|
||
// (provider.stream + stream.next 循环),隔离同步阻塞。
|
||
// 2. 通过 mpsc channel 将 StreamChunk 逐块传回主 runtime。
|
||
// 3. reqwest Response/Stream 的生命周期不跨 runtime,避免
|
||
// "error decoding response body"(runtime 状态不兼容)。
|
||
//
|
||
// 资源: 专用线程运行至 LLM 流结束或 mpsc receiver 被 drop(超时/停止)。
|
||
let (chunk_tx, mut chunk_rx) = tokio::sync::mpsc::channel::<
|
||
std::result::Result<df_ai::provider::StreamChunk, String>
|
||
>(64);
|
||
let provider_for_thread = provider.clone();
|
||
let conv_id_owned = conv_id.to_string();
|
||
|
||
std::thread::spawn(move || {
|
||
let rt = match tokio::runtime::Runtime::new() {
|
||
Ok(rt) => rt,
|
||
Err(e) => {
|
||
tracing::error!(
|
||
conv_id = %conv_id_owned,
|
||
"[ai] 创建专用 tokio runtime 失败: {}",
|
||
e,
|
||
);
|
||
return;
|
||
}
|
||
};
|
||
// BUG-2026-07-17 双重保险:在专用 runtime 内部也加 120s 超时。
|
||
// 即使 provider.stream() 在专用线程上意外阻塞(理论不应发生,因专用 runtime
|
||
// 多线程可推进独立 timer wheel),此 timeout 保证线程最终退出不永久挂起。
|
||
rt.block_on(async move {
|
||
use futures::StreamExt;
|
||
let stream_result = tokio::time::timeout(
|
||
std::time::Duration::from_secs(120),
|
||
provider_for_thread.stream(request),
|
||
).await;
|
||
match stream_result {
|
||
Ok(Ok(mut stream)) => {
|
||
while let Some(chunk_result) = stream.next().await {
|
||
let sent = match chunk_result {
|
||
Ok(chunk) => chunk_tx.send(Ok(chunk)).await,
|
||
Err(e) => chunk_tx.send(Err(e.to_string())).await,
|
||
};
|
||
if sent.is_err() {
|
||
// receiver dropped (timeout/stop/error) — 停止发送
|
||
break;
|
||
}
|
||
}
|
||
}
|
||
Ok(Err(e)) => {
|
||
let _ = chunk_tx.send(Err(e.to_string())).await;
|
||
}
|
||
Err(_elapsed) => {
|
||
tracing::error!(
|
||
conv_id = %conv_id_owned,
|
||
"[ai] 专用 runtime 内部超时(120s),provider.stream 未返回"
|
||
);
|
||
let _ = chunk_tx.send(Err("LLM 调用超时(120s)".to_string())).await;
|
||
}
|
||
}
|
||
});
|
||
});
|
||
|
||
// 在主 runtime 上通过 mpsc 接收 chunk,应用 timeout/heartbeat/stop 逻辑
|
||
let mut full_text = String::new();
|
||
let mut tool_calls_acc: HashMap<u32, ToolCallDraft> = HashMap::new();
|
||
let mut finished_received = false;
|
||
let mut stopped = false;
|
||
let mut final_usage: Option<df_ai::provider::TokenUsage> = None;
|
||
// DeepSeek thinking 模式推理内容累积（多轮需回传）
|
||
let mut reasoning_content_acc: Option<String> = None;
|
||
// per-conv delta 合批累加器。pending_delta 累积未 flush 的文本,
|
||
// next_flush_at 记录下次 flush 的绝对时刻(首个入 buffer 时置 now+50ms)。
|
||
let mut pending_delta = String::new();
|
||
let mut next_flush_at: Option<tokio::time::Instant> = None;
|
||
|
||
// heartbeat interval 提至 loop 外复用,避免每轮重建计时器
|
||
// (每轮重建会丢已积累的节拍,且 interval 首次 tick 立即返回的特性会被误用)。
|
||
// tokio interval 首 tick 立即返回——此处先丢弃首 tick,让心跳等满首个 30s 静默期才发
|
||
// (心跳语义是"静默期仍在跑",循环入口立即报无意义且会与 rx.recv() 抢分支错过首 chunk)。
|
||
let mut heartbeat = tokio::time::interval(HEARTBEAT_INTERVAL);
|
||
heartbeat.tick().await; // 丢弃首 tick
|
||
|
||
// BUG-2026-07-14 根治: wall-clock deadline 替代 select! 内 timeout 包裹。
|
||
// idle_deadline 是绝对时刻:"下一次 chunk 必须在此前到达,否则判定超时"。
|
||
// - 首 chunk 阶段: idle_deadline = now + 10s (FIRST_CHUNK_TIMEOUT)
|
||
// - 收到首 chunk 后: idle_deadline = now + 120s (STREAM_IDLE_TIMEOUT),每次收到 chunk 更新
|
||
// 心跳打断 select! 后回到循环顶部,deadline 检查在 select! 之前(不受竞争影响)。
|
||
let mut idle_deadline = tokio::time::Instant::now() + FIRST_CHUNK_TIMEOUT;
|
||
let mut first_chunk_done = false;
|
||
|
||
loop {
|
||
// stop 即时打断。
|
||
if stop_flag.load(std::sync::atomic::Ordering::SeqCst) {
|
||
stopped = true;
|
||
break;
|
||
}
|
||
|
||
// BUG-2026-07-14 根治: wall-clock deadline 检查(不依赖 select! timeout 语义)。
|
||
if tokio::time::Instant::now() >= idle_deadline {
|
||
if !first_chunk_done {
|
||
flush_delta(app_handle, conv_id, &mut pending_delta);
|
||
return StreamResult::InitFailed {
|
||
retryable: true,
|
||
error: format!(
|
||
"流式响应超时（{}秒未收到首字节，服务无响应或请求体过大）",
|
||
FIRST_CHUNK_TIMEOUT.as_secs()
|
||
),
|
||
};
|
||
}
|
||
// 已收到 finish_reason 仅剩等 usage 帧:慢 provider 拖过 idle 窗口不降 Partial,
|
||
// 按正常完成收尾(usage 缺失用默认 0,文本/tool_calls 已完整)。
|
||
if finished_received {
|
||
break;
|
||
}
|
||
warn!(
|
||
provider = %provider.name(),
|
||
conv_id = %conv_id,
|
||
text_len = full_text.len(),
|
||
"[ai] 流中途 idle timeout(deadline),保文不重试(incomplete)",
|
||
);
|
||
flush_delta(app_handle, conv_id, &mut pending_delta);
|
||
return StreamResult::Partial {
|
||
text: full_text,
|
||
tool_calls: tool_calls_acc,
|
||
usage: final_usage.unwrap_or_default(),
|
||
reasoning_content: reasoning_content_acc,
|
||
};
|
||
}
|
||
|
||
// 到点 flush 合批的 delta(单条 AiTextDelta 双写 emit + event_bus)。
|
||
if let Some(deadline) = next_flush_at {
|
||
if tokio::time::Instant::now() >= deadline {
|
||
next_flush_at = None;
|
||
flush_delta(app_handle, conv_id, &mut pending_delta);
|
||
}
|
||
}
|
||
|
||
// 多分支 select!:
|
||
// 1) chunk_rx.recv():来自专用线程的 LLM chunk (15s 保底 timeout,防 select! 死等)
|
||
// 2) heartbeat.tick():静默期发 AiHeartbeat reset 前端 watchdog
|
||
// 3) notify.notified():用户停止即时打断
|
||
// 4) flush 定时器(sleep_until next_flush_at):尾部 delta 不因流静默而延迟。
|
||
let flush_deadline = next_flush_at
|
||
.unwrap_or_else(|| tokio::time::Instant::now() + Duration::from_secs(3600));
|
||
let flush_fut = tokio::time::sleep_until(flush_deadline);
|
||
tokio::pin!(flush_fut);
|
||
|
||
tokio::select! {
|
||
chunk_result = tokio::time::timeout(
|
||
Duration::from_secs(15),
|
||
chunk_rx.recv()
|
||
) => {
|
||
match chunk_result {
|
||
Err(_elapsed) => {
|
||
// 15s 保底 timeout 触发。正常路径不会到这里:心跳 30s 会先触发 select! 返回。
|
||
if !first_chunk_done {
|
||
flush_delta(app_handle, conv_id, &mut pending_delta);
|
||
return StreamResult::InitFailed {
|
||
retryable: true,
|
||
error: format!(
|
||
"流式响应超时（{}秒未收到首字节，服务无响应或请求体过大）",
|
||
FIRST_CHUNK_TIMEOUT.as_secs()
|
||
),
|
||
};
|
||
}
|
||
// 已收到 finish_reason 仅剩等 usage 帧:慢/坏 provider 拖过 15s 不降 Partial,
|
||
// 按正常完成收尾(与 idle_deadline 分支同口径,usage 缺失用默认 0)。
|
||
if finished_received {
|
||
break;
|
||
}
|
||
warn!(
|
||
provider = %provider.name(),
|
||
conv_id = %conv_id,
|
||
text_len = full_text.len(),
|
||
"[ai] 流中途 idle timeout,保文不重试(incomplete)",
|
||
);
|
||
flush_delta(app_handle, conv_id, &mut pending_delta);
|
||
return StreamResult::Partial {
|
||
text: full_text,
|
||
tool_calls: tool_calls_acc,
|
||
usage: final_usage.unwrap_or_default(),
|
||
reasoning_content: reasoning_content_acc,
|
||
};
|
||
}
|
||
Ok(None) => break, // mpsc channel 关闭 = 专用线程 stream 结束
|
||
Ok(Some(Ok(chunk))) => {
|
||
// BUG-2026-07-14 根治: 收到 chunk 更新 idle_deadline。
|
||
if !first_chunk_done {
|
||
first_chunk_done = true;
|
||
}
|
||
idle_deadline = tokio::time::Instant::now() + STREAM_IDLE_TIMEOUT;
|
||
if !chunk.delta.is_empty() {
|
||
full_text.push_str(&chunk.delta);
|
||
// 合批入 buffer,首个入 buffer 时置 flush 时刻(到点统一 emit 单条)。
|
||
pending_delta.push_str(&chunk.delta);
|
||
if next_flush_at.is_none() {
|
||
next_flush_at = Some(tokio::time::Instant::now() + DELTA_FLUSH_INTERVAL);
|
||
}
|
||
}
|
||
if let Some(tc_deltas) = &chunk.tool_calls {
|
||
accumulate_tool_calls(tc_deltas, &mut tool_calls_acc);
|
||
}
|
||
if let Some(u) = &chunk.usage {
|
||
final_usage = Some(u.clone());
|
||
}
|
||
if let Some(ref rc) = chunk.reasoning_content {
|
||
reasoning_content_acc.get_or_insert_with(String::new).push_str(rc);
|
||
}
|
||
// provider 流式错误事件
|
||
if let Some(err_msg) = &chunk.error {
|
||
// A2-B12 / G4.2:error 帧按 message 文本分类 retryable。
|
||
// 原硬编码 retryable:true,确定性请求错(鉴权/context_length/invalid_request)
|
||
// 被当瞬态空耗重试。现抠 HTTP 状态码/传输类走既有 classify_status_or_class
|
||
// 单一分类源;无码则按错误类型关键词保守降级(仅明确非重试降 Fatal)。
|
||
let (status_or_class, _raw) = extract_error_diag_from_str(err_msg);
|
||
let retryable = classify_error_frame_retryable(err_msg, &status_or_class);
|
||
let diag_class = if status_or_class == "unknown" {
|
||
"stream-error"
|
||
} else {
|
||
status_or_class.as_str()
|
||
};
|
||
warn!(
|
||
provider = %provider.name(),
|
||
conv_id = %conv_id,
|
||
error = %err_msg,
|
||
status_or_class = %status_or_class,
|
||
retryable = %retryable,
|
||
text_len = full_text.len(),
|
||
"[ai] provider 流式错误事件",
|
||
);
|
||
if full_text.is_empty() && tool_calls_acc.is_empty() {
|
||
flush_delta(app_handle, conv_id, &mut pending_delta);
|
||
return StreamResult::InitFailed {
|
||
retryable,
|
||
error: fmt_diag(
|
||
provider.name(),
|
||
DiagKind::MidStream,
|
||
diag_class,
|
||
err_msg,
|
||
),
|
||
};
|
||
}
|
||
flush_delta(app_handle, conv_id, &mut pending_delta);
|
||
return StreamResult::Partial {
|
||
text: full_text,
|
||
tool_calls: tool_calls_acc,
|
||
usage: final_usage.unwrap_or_default(),
|
||
reasoning_content: reasoning_content_acc,
|
||
};
|
||
}
|
||
if chunk.finished {
|
||
finished_received = true;
|
||
// usage 携带点:OpenAI 兼容流中 usage 挂在 [DONE] 帧(或 usage-only 帧),
|
||
// 而 finish_reason 帧(finished=true, usage=None)在其之前到达。
|
||
// 若此刻 break 会错过 [DONE] 帧的 usage → 真实 completion_tokens 丢失,
|
||
// 前端 token 显示 0。拿到真实 usage 才停;usage 仍 None 则继续读到
|
||
// usage-only/[DONE] 帧或通道关闭(Ok(None) 兜底退出,不断连误判)。
|
||
if chunk.usage.is_some() {
|
||
break;
|
||
}
|
||
}
|
||
}
|
||
Ok(Some(Err(err_str))) => {
|
||
// 专用线程上报的流错误(SSE 解析失败/HTTP 错误等)
|
||
let (status_or_class, raw) = extract_error_diag_from_str(&err_str);
|
||
warn!(
|
||
provider = %provider.name(),
|
||
status = %status_or_class,
|
||
conv_id = %conv_id,
|
||
error = %raw,
|
||
text_len = full_text.len(),
|
||
"[ai] 流式接收中途错误(from 专用线程)",
|
||
);
|
||
if full_text.is_empty() && tool_calls_acc.is_empty() {
|
||
flush_delta(app_handle, conv_id, &mut pending_delta);
|
||
return StreamResult::InitFailed {
|
||
retryable: classify_status_with_body(&status_or_class, &raw),
|
||
error: fmt_diag(
|
||
provider.name(),
|
||
DiagKind::MidStream,
|
||
&status_or_class,
|
||
&raw,
|
||
),
|
||
};
|
||
}
|
||
flush_delta(app_handle, conv_id, &mut pending_delta);
|
||
return StreamResult::Partial {
|
||
text: full_text,
|
||
tool_calls: tool_calls_acc,
|
||
usage: final_usage.unwrap_or_default(),
|
||
reasoning_content: reasoning_content_acc,
|
||
};
|
||
}
|
||
}
|
||
}
|
||
// 2) 心跳:静默期 30s 发 AiHeartbeat,前端 watchdog reset
|
||
_ = heartbeat.tick() => {
|
||
emit_ai_chat_event(app_handle, AiChatEvent::AiHeartbeat {
|
||
conversation_id: Some(conv_id.to_string()),
|
||
});
|
||
}
|
||
// 3) 即时停止唤醒
|
||
_ = notify.notified() => {
|
||
if stop_flag.load(std::sync::atomic::Ordering::SeqCst) {
|
||
stopped = true;
|
||
break;
|
||
}
|
||
}
|
||
// 4) delta 合批到点 flush(流静默/工具执行等待期也能按时发出尾部文本)
|
||
_ = &mut flush_fut => {
|
||
if next_flush_at.is_some() {
|
||
next_flush_at = None;
|
||
flush_delta(app_handle, conv_id, &mut pending_delta);
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
// 用户停止
|
||
if stopped {
|
||
flush_delta(app_handle, conv_id, &mut pending_delta);
|
||
return StreamResult::Complete {
|
||
text: full_text,
|
||
tool_calls: tool_calls_acc,
|
||
usage: final_usage.unwrap_or_default(),
|
||
reasoning_content: reasoning_content_acc,
|
||
};
|
||
}
|
||
|
||
// 断连检测
|
||
if !finished_received {
|
||
if full_text.is_empty() && tool_calls_acc.is_empty() {
|
||
return StreamResult::InitFailed {
|
||
retryable: true,
|
||
error: "流式响应意外中断（未收到完成信号，已丢弃残缺响应）".to_string(),
|
||
};
|
||
}
|
||
warn!(
|
||
provider = %provider.name(),
|
||
conv_id = %conv_id,
|
||
text_len = full_text.len(),
|
||
"[ai] 流尽未收到 finished 但有 partial_text,保文不重试(incomplete)",
|
||
);
|
||
flush_delta(app_handle, conv_id, &mut pending_delta);
|
||
return StreamResult::Partial {
|
||
text: full_text,
|
||
tool_calls: tool_calls_acc,
|
||
usage: final_usage.unwrap_or_default(),
|
||
reasoning_content: reasoning_content_acc,
|
||
};
|
||
}
|
||
|
||
// 正常完成
|
||
flush_delta(app_handle, conv_id, &mut pending_delta);
|
||
StreamResult::Complete {
|
||
text: full_text,
|
||
tool_calls: tool_calls_acc,
|
||
usage: final_usage.unwrap_or_default(),
|
||
reasoning_content: reasoning_content_acc,
|
||
}
|
||
}
|
||
|
||
/// 据 extract_error_diag 产出的 status_or_class 字符串镜像 retry::is_status_retryable 分类。
|
||
///
|
||
/// status_or_class 取值: "HTTP 5xx"/"HTTP 4xx"(白名单码) / "timeout" / "connect" / "unknown"。
|
||
/// - HTTP 429 / 5xx → retryable=true(复用 retry::is_status_retryable 逻辑)
|
||
/// - HTTP 4xx(非429) → retryable=false Fatal(401/403/404/422 等)
|
||
/// - timeout / connect → retryable=true(瞬态,可重试)
|
||
/// - unknown / 其他 → retryable=true(保守可重试,避免误判 Fatal 错杀)
|
||
///
|
||
/// 抽公共分类函数(决策 F-260616-07 a1 "复用 retry.rs 错误分类逻辑或镜像"),
|
||
/// 避免 stream_recv 与 agentic 各自重写状态码解析。
|
||
fn classify_status_or_class(status_or_class: &str) -> bool {
|
||
use df_ai::retry::is_status_retryable;
|
||
// 抠 status_or_class 内三位数字状态码("HTTP 401" → 401)
|
||
let chars: Vec<char> = status_or_class.chars().collect();
|
||
let mut i = 0;
|
||
let n = chars.len();
|
||
while i + 3 <= n {
|
||
if chars[i].is_ascii_digit()
|
||
&& chars[i + 1].is_ascii_digit()
|
||
&& chars[i + 2].is_ascii_digit()
|
||
{
|
||
let code: String = chars[i..i + 3].iter().collect();
|
||
if let Ok(status) = code.parse::<u16>() {
|
||
return is_status_retryable(status);
|
||
}
|
||
i += 3;
|
||
continue;
|
||
}
|
||
i += 1;
|
||
}
|
||
// 无三位数字码:按文本分类(timeout/connect/transport/decode 可重试,其余 unknown 保守不重试)
|
||
// Anthropic 中文参数错误(如 `[1214][messages 参数非法]`)无 HTTP 数字前缀,
|
||
// 走此文本分支,原 || true 致 retryable=true 空耗 4 次重试浪费 token。
|
||
// 删 || true 对齐 retry::is_status_retryable 未知码不可重试语义;timeout/connect 瞬态仍可重试。
|
||
//
|
||
// BUG-2026-07-17 根治:SSE 解码/传输错误(error decoding response body / Transport error)
|
||
// 是传输层错误,不是 Fatal。原实现未识别这些关键词,返 unknown→Fatal,致连续重试 4 次均
|
||
// 立即放弃。现补全传输层关键词,让这类错误走可重试路径。
|
||
let lower = status_or_class.to_lowercase();
|
||
lower.contains("timeout")
|
||
|| lower.contains("connect")
|
||
|| lower.contains("transport")
|
||
|| lower.contains("decoding")
|
||
|| lower.contains("decode")
|
||
|| lower.contains("body")
|
||
|| lower.contains("eof")
|
||
|| lower.contains("reset")
|
||
|| lower.contains("broken pipe")
|
||
}
|
||
|
||
/// 429 额度/配额耗尽特征检测:区分「额度耗尽型 429」(确定性,重试必失败)与
|
||
/// 「普通瞬时限流 429」(可重试)。实证:anthropic/GLM 周月额度耗尽返回
|
||
/// `[1310] 您已达到每周/每月使用上限,限额将在 X 重置`(code 1310)。
|
||
/// 特征命中 → Fatal 不重试;未命中 → 交 is_status_retryable 正常判定。
|
||
/// 入参 combined:状态码串 + 错误体原文(如 "HTTP 429 [1310] 已达到上限...")。
|
||
fn is_quota_exhaustion(combined: &str) -> bool {
|
||
let lower = combined.to_lowercase();
|
||
let quota_markers = [
|
||
"额度", "限额", "达到上限", "使用上限", "重置",
|
||
"quota", "limit exceeded", "usage limit", "quota exhausted",
|
||
"insufficient_quota",
|
||
// 注意:不含 rate_limit_error——它是标准 429 错误类型(普通瞬时限流也带),
|
||
// 误配会把可恢复的限流当额度耗尽降 Fatal。额度语义用上面明确词。
|
||
];
|
||
// 仅当文本同时含「429」与额度特征,才判定额度耗尽(避免误伤普通限流)。
|
||
lower.contains("429")
|
||
&& quota_markers.iter().any(|m| lower.contains(m))
|
||
}
|
||
|
||
/// 额度耗尽型 429 → Fatal(确定性,不重试);否则交 classify_status_or_class 正常判定。
|
||
/// 入参 body:错误体原文(provider bail 文本 / error 帧 message)。
|
||
fn classify_status_with_body(status_or_class: &str, body: &str) -> bool {
|
||
if is_quota_exhaustion(&format!("{status_or_class} {body}")) {
|
||
return false;
|
||
}
|
||
classify_status_or_class(status_or_class)
|
||
}
|
||
|
||
/// 据 provider 流式 error 帧的 message 文本分类是否可重试(A2-B12 / G4.2)。
|
||
///
|
||
/// openai/anthropic helper 仅将 error 帧的 `message` 字段透传进 chunk.error
|
||
/// (结构化 `type` 未上抛),故此处按 message 文本保守分类,对齐 Err 路径单一分类源:
|
||
/// - 内含显式 HTTP 状态码或传输类(timeout/connect) → 复用 classify_status_or_class
|
||
/// - 无状态码时按错误类型关键词:明确非重试签名(authentication / context_length /
|
||
/// invalid_request / 鉴权 / 参数错等) → Fatal
|
||
/// - 其余 → retryable=true(保守,防误判瞬态为 Fatal)
|
||
fn classify_error_frame_retryable(msg: &str, status_or_class: &str) -> bool {
|
||
// 已有状态码/传输类 → 走既有单一分类源(显式 4xx Fatal,5xx/429/timeout/connect 可重试)
|
||
// 额度耗尽型 429(如 [1310] 已达周/月上限)确定性失败,经 body 感知识别降 Fatal,
|
||
// 不再被普通 429 路径误判可重试空耗。
|
||
if status_or_class != "unknown" {
|
||
return classify_status_with_body(status_or_class, msg);
|
||
}
|
||
// 无状态码:按错误类型关键词明确非重试签名降级 Fatal
|
||
let lower = msg.to_lowercase();
|
||
const FATAL_KEYWORDS: &[&str] = &[
|
||
// OpenAI / Anthropic error.type
|
||
"authentication_error",
|
||
"invalid_request_error",
|
||
"context_length_exceeded",
|
||
"permission_error",
|
||
"not_found_error",
|
||
"request_too_large",
|
||
"insufficient_quota",
|
||
// OpenAI 消息文本签名
|
||
"authentication",
|
||
"invalid_request",
|
||
"invalid request",
|
||
"context length",
|
||
"api key",
|
||
"api-key",
|
||
"api_key",
|
||
"unauthorized",
|
||
// 中文签名Anthropic 中文参数错误)
|
||
"鉴权",
|
||
"认证失败",
|
||
"未授权",
|
||
"密钥",
|
||
"参数非法",
|
||
"参数错误",
|
||
"上下文长度",
|
||
"输入超长",
|
||
];
|
||
!FATAL_KEYWORDS.iter().any(|k| lower.contains(k))
|
||
}
|
||
|
||
// ============================================================
|
||
// 单测：诊断提取/格式化（纯函数，不发 HTTP、不依赖 app_handle）
|
||
// ============================================================
|
||
|
||
#[cfg(test)]
|
||
mod tests {
|
||
use super::*;
|
||
|
||
// ---- classify_status_or_class：CR-30-1 retryable 分类(镜像 retry::is_status_retryable) ----
|
||
|
||
/// 5xx → retryable=true
|
||
#[test]
|
||
fn classify_5xx_retryable() {
|
||
assert!(classify_status_or_class("HTTP 500"));
|
||
assert!(classify_status_or_class("HTTP 503"));
|
||
}
|
||
|
||
/// 429 → retryable=true(限流可重试)
|
||
#[test]
|
||
fn classify_429_retryable() {
|
||
assert!(classify_status_or_class("HTTP 429"));
|
||
}
|
||
|
||
/// 4xx(非429) → retryable=false Fatal(鉴权/参数错立即放弃)
|
||
#[test]
|
||
fn classify_4xx_fatal() {
|
||
assert!(!classify_status_or_class("HTTP 401"));
|
||
assert!(!classify_status_or_class("HTTP 403"));
|
||
assert!(!classify_status_or_class("HTTP 404"));
|
||
assert!(!classify_status_or_class("HTTP 422"));
|
||
}
|
||
|
||
/// timeout / connect → retryable=true(瞬态)
|
||
#[test]
|
||
fn classify_transient_retryable() {
|
||
assert!(classify_status_or_class("timeout"));
|
||
assert!(classify_status_or_class("connect"));
|
||
}
|
||
|
||
/// unknown → 保守 retryable=false防 Anthropic 中文参数错误等 Fatal 空耗重试)
|
||
#[test]
|
||
fn classify_unknown_not_retryable() {
|
||
assert!(!classify_status_or_class("unknown"));
|
||
}
|
||
|
||
// ---- classify_error_frame_retryable：A2-B12 / G4.2 error 帧分类 ----
|
||
|
||
/// error 帧 message 含显式 4xx → Fatal(接入 classify_status_or_class 单一分类源)
|
||
#[test]
|
||
fn error_frame_http_4xx_fatal() {
|
||
let (s, _) = extract_error_diag_from_str("LLM 流式 API 错误 401: Unauthorized");
|
||
assert!(!classify_error_frame_retryable("LLM 流式 API 错误 401: Unauthorized", &s));
|
||
}
|
||
|
||
/// error 帧 message 含显式 5xx / 429 → retryable
|
||
#[test]
|
||
fn error_frame_http_5xx_retryable() {
|
||
let (s, _) = extract_error_diag_from_str("LLM 流式 API 错误 500: Internal Server Error");
|
||
assert!(classify_error_frame_retryable("LLM 流式 API 错误 500: Internal Server Error", &s));
|
||
let (s2, _) = extract_error_diag_from_str("rate limit 429");
|
||
assert!(classify_error_frame_retryable("rate limit 429", &s2));
|
||
}
|
||
|
||
/// error 帧 message 含 timeout → retryable(瞬态)
|
||
#[test]
|
||
fn error_frame_timeout_retryable() {
|
||
let (s, _) = extract_error_diag_from_str("operation timed out");
|
||
assert!(classify_error_frame_retryable("operation timed out", &s));
|
||
}
|
||
|
||
/// authentication / api key 签名 → Fatal
|
||
#[test]
|
||
fn error_frame_authentication_fatal() {
|
||
let (s, _) = extract_error_diag_from_str("Incorrect API key provided: sk-xxx");
|
||
assert!(!classify_error_frame_retryable("Incorrect API key provided: sk-xxx", &s));
|
||
let (s2, _) = extract_error_diag_from_str("invalid x-api-key");
|
||
assert!(!classify_error_frame_retryable("invalid x-api-key", &s2));
|
||
}
|
||
|
||
/// context_length_exceeded / context length 签名 → Fatal
|
||
#[test]
|
||
fn error_frame_context_length_fatal() {
|
||
let (s, _) = extract_error_diag_from_str("This model's maximum context length is 128000 tokens");
|
||
assert!(!classify_error_frame_retryable("This model's maximum context length is 128000 tokens", &s));
|
||
}
|
||
|
||
/// invalid_request 签名 → Fatal
|
||
#[test]
|
||
fn error_frame_invalid_request_fatal() {
|
||
let (s, _) = extract_error_diag_from_str("invalid_request_error: the model is not supported");
|
||
assert!(!classify_error_frame_retryable("invalid_request_error: the model is not supported", &s));
|
||
}
|
||
|
||
/// 无明确签名(overloaded / stream error 等瞬态语义)→ 保守 retryable=true(防误判 Fatal 错杀)
|
||
#[test]
|
||
fn error_frame_unknown_conservative_retryable() {
|
||
let (s, _) = extract_error_diag_from_str("overloaded_error: Overloaded");
|
||
assert!(classify_error_frame_retryable("overloaded_error: Overloaded", &s));
|
||
let (s2, _) = extract_error_diag_from_str("stream error");
|
||
assert!(classify_error_frame_retryable("stream error", &s2));
|
||
}
|
||
|
||
// ---- is_quota_exhaustion / classify_status_with_body：额度耗尽型 429 不重试 ----
|
||
|
||
/// 额度耗尽型 429（anthropic/GLM 周月上限，含 [1310] 与「使用上限/重置」特征）→ 判定额度耗尽
|
||
#[test]
|
||
fn quota_exhaustion_429_detected() {
|
||
let combined = "HTTP 429 [1310] 您已达到每周/每月使用上限,限额将在 2026-08-14 重置";
|
||
assert!(is_quota_exhaustion(combined));
|
||
assert!(!classify_status_with_body("HTTP 429", combined));
|
||
}
|
||
|
||
/// 普通瞬时限流 429（无额度特征）→ 不判定额度耗尽,仍可重试
|
||
#[test]
|
||
fn transient_429_not_quota() {
|
||
let combined = "HTTP 429 请求过于频繁,请稍后重试";
|
||
assert!(!is_quota_exhaustion(combined));
|
||
assert!(classify_status_with_body("HTTP 429", combined));
|
||
}
|
||
|
||
/// 额度关键词但无 429 状态码 → 不判定额度耗尽(防误伤普通 5xx)
|
||
#[test]
|
||
fn quota_marker_without_429_not_detected() {
|
||
assert!(!is_quota_exhaustion("HTTP 500 达到上限"));
|
||
}
|
||
|
||
/// classify_status_with_body 非 429 路径行为不变(5xx 仍可重试)
|
||
#[test]
|
||
fn classify_status_with_body_5xx_retryable() {
|
||
assert!(classify_status_with_body("HTTP 500", "Internal Server Error"));
|
||
}
|
||
|
||
/// 普通瞬时限流 429 的标准 type=rate_limit_error 不应误判额度耗尽(仍可重试,防误伤)
|
||
#[test]
|
||
fn standard_rate_limit_error_not_quota() {
|
||
let combined = r#"HTTP 429 {"type":"rate_limit_error","message":"请求过于频繁,请稍后重试"}"#;
|
||
assert!(!is_quota_exhaustion(combined));
|
||
assert!(classify_status_with_body("HTTP 429", combined));
|
||
}
|
||
|
||
/// error 帧 message 含额度耗尽型 429 → 经 classify_error_frame_retryable 降 Fatal 不重试
|
||
#[test]
|
||
fn error_frame_quota_429_fatal() {
|
||
let msg = "HTTP 429 [1310] 您已达到每周/每月使用上限,限额将在 2026-08-14 重置";
|
||
let (s, _) = extract_error_diag_from_str(msg);
|
||
assert_eq!(s, "HTTP 429");
|
||
assert!(!classify_error_frame_retryable(msg, &s));
|
||
}
|
||
|
||
// ---- extract_error_diag：业务层 bail（provider 串已含状态码）----
|
||
|
||
/// provider 在 non-2xx bail 的典型串：抠出 401（鉴权失败/Key 错）
|
||
#[test]
|
||
fn diag_extracts_401_from_provider_bail() {
|
||
let e = anyhow::anyhow!("LLM 流式 API 错误 401: Unauthorized");
|
||
let (status, raw) = extract_error_diag(&e);
|
||
assert_eq!(status, "HTTP 401");
|
||
assert!(raw.contains("401"));
|
||
}
|
||
|
||
/// 404 = base_url/endpoint 不对
|
||
#[test]
|
||
fn diag_extracts_404_from_provider_bail() {
|
||
let e = anyhow::anyhow!("LLM 流式 API 错误 404: Not Found");
|
||
assert_eq!(extract_error_diag(&e).0, "HTTP 404");
|
||
}
|
||
|
||
/// 429 = 限流
|
||
#[test]
|
||
fn diag_extracts_429_from_provider_bail() {
|
||
let e = anyhow::anyhow!("LLM 流式 API 错误 429: rate limit");
|
||
assert_eq!(extract_error_diag(&e).0, "HTTP 429");
|
||
}
|
||
|
||
/// 5xx = 上游服务端错
|
||
#[test]
|
||
fn diag_extracts_500_from_provider_bail() {
|
||
let e = anyhow::anyhow!("LLM 流式 API 错误 500: Internal Server Error");
|
||
assert_eq!(extract_error_diag(&e).0, "HTTP 500");
|
||
}
|
||
|
||
// ---- extract_error_diag：传输层（reqwest Display 文本，不命名 reqwest）----
|
||
|
||
/// 超时：连接成功但响应慢/静默断
|
||
#[test]
|
||
fn diag_classifies_timeout() {
|
||
let e = anyhow::anyhow!("error sending request for url (https://api.x.com/v1/chat/completions): operation timed out");
|
||
let (status, _) = extract_error_diag(&e);
|
||
assert_eq!(status, "timeout");
|
||
}
|
||
|
||
/// 中文「超时」关键词
|
||
#[test]
|
||
fn diag_classifies_timeout_cn() {
|
||
let e = anyhow::anyhow!("请求超时");
|
||
assert_eq!(extract_error_diag(&e).0, "timeout");
|
||
}
|
||
|
||
/// 连接失败：DNS 解析失败 / 端点不通
|
||
#[test]
|
||
fn diag_classifies_connect() {
|
||
let e = anyhow::anyhow!("dns error: failed to lookup address information");
|
||
assert_eq!(extract_error_diag(&e).0, "connect");
|
||
}
|
||
|
||
/// resolve 关键词也归 connect
|
||
#[test]
|
||
fn diag_classifies_connect_resolve() {
|
||
let e = anyhow::anyhow!("error connecting: resolve failed");
|
||
assert_eq!(extract_error_diag(&e).0, "connect");
|
||
}
|
||
|
||
// ---- extract_error_diag：边界 ----
|
||
|
||
/// 三位数但不在已知 HTTP 状态白名单（如 "200"）→ 不当状态码，走分类
|
||
#[test]
|
||
fn diag_ignores_non_http_status_number() {
|
||
let e = anyhow::anyhow!("成功 200 条记录");
|
||
// 200 不在白名单，应落到 unknown
|
||
assert_eq!(extract_error_diag(&e).0, "unknown");
|
||
}
|
||
|
||
/// 完全无特征文本 → unknown，且 raw 原样返回
|
||
#[test]
|
||
fn diag_unknown_preserves_raw() {
|
||
let e = anyhow::anyhow!("奇怪的错误 xyz");
|
||
let (status, raw) = extract_error_diag(&e);
|
||
assert_eq!(status, "unknown");
|
||
assert_eq!(raw, "奇怪的错误 xyz");
|
||
}
|
||
|
||
/// 状态码出现在更长数字串里也不误匹配（边界判断：前后必须非数字）
|
||
#[test]
|
||
fn diag_does_not_match_status_inside_longer_digits() {
|
||
// 边界判断后,白名单码嵌在长数字串内不再误命中(原窗口切片会从 14012 抠出 401)
|
||
let e = anyhow::anyhow!("port 14012 used");
|
||
assert_eq!(extract_error_diag(&e).0, "unknown");
|
||
// 纯长数字串(无白名单码)同样不命中
|
||
let e2 = anyhow::anyhow!("port 99999 used");
|
||
assert_eq!(extract_error_diag(&e2).0, "unknown");
|
||
}
|
||
|
||
// ---- fmt_diag：两个上下文 ----
|
||
|
||
#[test]
|
||
fn fmt_diag_init_branch() {
|
||
let s = fmt_diag("anthropic-compat", DiagKind::Init, "HTTP 401", "Unauthorized");
|
||
assert_eq!(s, "[anthropic-compat] AI 调用失败(HTTP 401): Unauthorized");
|
||
}
|
||
|
||
/// 状态码紧贴中文字符(多字节 UTF-8)时仍能正确抠出
|
||
/// (原按字节窗口切片在中文边界脆弱,改为 char 迭代后安全)
|
||
#[test]
|
||
fn diag_extracts_code_adjacent_to_multibyte_chars() {
|
||
let e = anyhow::anyhow!("错误401Unauthorized");
|
||
assert_eq!(extract_error_diag(&e).0, "HTTP 401");
|
||
// 中文在前:状态码前字符是多字节中文,char 迭代边界判断正确
|
||
let e2 = anyhow::anyhow!("请求失败：401");
|
||
assert_eq!(extract_error_diag(&e2).0, "HTTP 401");
|
||
}
|
||
|
||
#[test]
|
||
fn fmt_diag_midstream_branch() {
|
||
let s = fmt_diag("gpt-4o", DiagKind::MidStream, "timeout", "operation timed out");
|
||
assert_eq!(s, "[gpt-4o] 流式接收错误(timeout): operation timed out");
|
||
}
|
||
|
||
/// provider 名/错误文本含特殊字符仍按模板拼接（无格式注入风险）
|
||
#[test]
|
||
fn fmt_diag_handles_special_chars() {
|
||
let s = fmt_diag("model/x", DiagKind::Init, "unknown", "err: {json} \"quoted\"");
|
||
assert_eq!(s, "[model/x] AI 调用失败(unknown): err: {json} \"quoted\"");
|
||
}
|
||
}
|