copilot-api-plus 1.2.14 → 1.2.16

package/README.en.md

@@ -46,10 +46,10 @@ English | [简体中文](README.md)
 | 👥 **Multi-Account** | Multiple GitHub accounts with automatic failover on quota exhaustion/rate limiting/bans |
 | 🔀 **Model Routing** | Flexible model name mapping and per-model concurrency control |
 | 📱 **Visual Management** | Web dashboard for account management, model config, and runtime stats |
-| 🛡️ **Network Resilience** | 60s timeout + smart retry + instant stream recovery + proxy tunnel keepalive (45s heartbeat) |
+| 🛡️ **Network Resilience** | 120s timeout + smart retry + instant stream recovery + proxy tunnel keepalive (45s heartbeat) |
 | ✂️ **Context Passthrough** | Full context passthrough to upstream API; clients (e.g. Claude Code) manage compression |
 | 🔍 **Smart Model Matching** | Handles model name format differences (date suffixes, dash/dot versions, etc.) |
-| 🧠 **Thinking Chain** | Automatically enables deep thinking (thinking/reasoning) for supported models, improving code quality |
+| 🧠 **Thinking Chain** | Automatically enables deep thinking for supported models, improving code quality |
 
 ---
 
@@ -582,13 +582,13 @@ Each API request outputs a log line with model name, status code, and duration:
 
 Built-in connection timeout and smart retry for upstream API requests, minimizing Copilot request credit consumption:
 
-- **Connection timeout**: 60 seconds for the first attempt, 30 seconds for retries (headers typically arrive in 3–5s)
-- **Retry strategy**: Up to 2 retries (3 total attempts), 2-3 second delays
+- **Connection timeout**: 120 seconds for the first attempt, 30 seconds for retries (headers typically arrive in 3–5s)
+- **Retry strategy**: Up to 1 retry (2 total attempts), 2-second delay. **Timeout errors are never retried** — a timeout means the request likely reached Copilot and a credit was already consumed
 - **Instant stream recovery**: On SSE stream interruption, immediately destroys the connection pool so the next request uses fresh sockets — recovery drops from ~135s to seconds
 - **Connection pool reset**: Automatically destroys all pooled connections on the first network error and creates fresh instances, preventing retries from hitting stale sockets
 - **Proxy tunnel keepalive**: Sends lightweight heartbeat requests every 45s while SSE streams are active, preventing proxy nodes from killing CONNECT tunnels due to inactivity
 - **HTTP/2 support**: Enables HTTP/2 protocol for better multiplexing performance
-- Only retries network-layer errors (timeout, TLS disconnect, connection reset, etc.); HTTP error codes (e.g. 400/500) are not retried
+- Only retries network-layer connection errors (TLS disconnect, connection reset, etc.); timeout and HTTP error codes (e.g. 400/500) are not retried
 - SSE stream interruptions gracefully send error events to the client
 
 ---

package/README.md

@@ -47,10 +47,10 @@
 | 👥 **多账号管理** | 支持添加多个 GitHub 账号，额度耗尽/限流/封禁时自动切换下一个 |
 | 🔀 **模型路由** | 灵活的模型名映射和每模型并发控制 |
 | 📱 **可视化管理** | Web 仪表盘支持账号管理、模型管理、运行统计 |
-| 🛡️ **网络弹性** | 60s 连接超时 + 智能重试 + 流中断即时恢复 + 代理隧道保活（45s 心跳） |
+| 🛡️ **网络弹性** | 120s 连接超时 + 智能重试 + 流中断即时恢复 + 代理隧道保活（45s 心跳） |
 | ✂️ **上下文透传** | 全量透传上下文至上游 API，由客户端（如 Claude Code）自行管理压缩 |
 | 🔍 **智能模型匹配** | 自动处理模型名格式差异（日期后缀、dash/dot 版本号等） |
-| 🧠 **Thinking 思维链** | 自动为支持的模型启用深度思考（thinking/reasoning），提升代码质量 |
+| 🧠 **Thinking 思维链** | 自动为支持的模型启用深度思考，提升代码质量 |
 
 ---
 
@@ -745,13 +745,13 @@ Anthropic 格式的模型名（如 `claude-opus-4-6`）和 Copilot 的模型列
 
 对上游 API 的请求内置了连接超时和智能重试，以最小化 Copilot 请求次数消耗：
 
-- **连接超时**：首次请求 60 秒，重试请求 30 秒（响应头通常 3~5 秒到达）
-- **重试策略**：最多重试 2 次（共 3 次尝试），间隔 2-3 秒
+- **连接超时**：首次请求 120 秒，重试请求 30 秒（响应头通常 3~5 秒到达）
+- **重试策略**：最多重试 1 次（共 2 次尝试），间隔 2 秒，**超时错误不重试**（超时意味着请求已到达 Copilot，额度已消耗）
 - **流中断即时恢复**：SSE 流中断时立刻销毁连接池，下一个请求使用全新连接，恢复时间从 ~135 秒降至几秒
 - **连接池重置**：首次网络错误后自动销毁所有连接并创建新实例，避免后续请求复用坏连接
 - **代理隧道保活**：SSE 流传输期间每 45 秒发送一次轻量心跳请求，防止代理节点因空闲而杀断 CONNECT 隧道
 - **HTTP/2 支持**：启用 HTTP/2 协议，提升多路复用性能
-- 仅重试网络层错误（超时、TLS 断开、连接重置等），HTTP 错误码（如 400/500）不重试
+- 仅重试网络层连接错误（TLS 断开、连接重置等），超时和 HTTP 错误码（如 400/500）不重试
 - SSE 流传输中断时，优雅地向客户端发送错误事件
 
 ---

package/dist/main.js

@@ -1768,12 +1768,14 @@ async function checkRateLimit(state) {
 const FETCH_TIMEOUT_MS = 12e4;
 /**
 * Retry delays in ms.  After the first failure the connection pool is reset
-* (see `resetConnections`), so retries use fresh sockets.  We allow up to
-* 2 retries because SSE streams through HTTP proxies are frequently
-* interrupted during long model thinking phases (~60 s idle timeout on
-* many proxy nodes).  Keeping the delay short avoids wasting wall-clock time.
+* (see `resetConnections`), so retries use fresh sockets.
+*
+* We only allow 1 retry (2 total attempts) to minimize credit waste.
+* Timeout errors are NOT retried at all — they indicate the request likely
+* reached Copilot (consuming a credit) and the upstream is slow or the
+* proxy killed the connection mid-flight.
 */
-const RETRY_DELAYS = [2e3, 3e3];
+const RETRY_DELAYS = [2e3];
 /**
 * Timeout for retry attempts (waiting for response headers only).
 * Response headers typically arrive within 3–5 s, even on slow models.
@@ -1820,6 +1822,11 @@ async function fetchWithRetry(url, buildInit) {
 		return await fetchWithTimeout(url, buildInit(), timeout);
 	} catch (error) {
 		lastError = error;
+		const msg = error instanceof Error ? error.message : String(error);
+		if (msg.includes("timed out")) {
+			consola.warn(`Request timed out on attempt ${attempt + 1}/${maxAttempts} — not retrying (credit likely consumed):`, msg);
+			break;
+		}
 		if (attempt === 0) resetConnections();
 		if (attempt < maxAttempts - 1) {
 			const delay = RETRY_DELAYS[attempt];
@@ -1905,7 +1912,7 @@ function injectThinking(payload, resolvedModel) {
 	return payload;
 }
 function logThinkingInjection(original, injected, resolvedModel) {
-	if (original.reasoning_effort || original.thinking_budget) consola.debug(`Thinking: client-specified (reasoning_effort=${original.reasoning_effort ?? "none"} / thinking_budget=${original.thinking_budget ?? "none"})`);
+	if (original.reasoning_effort || original.thinking_budget) consola.debug(`Thinking: translated (reasoning_effort=${original.reasoning_effort ?? "none"} / thinking_budget=${original.thinking_budget ?? "none"})`);
 	else if (injected.thinking_budget && injected.thinking_budget !== original.thinking_budget) consola.debug(`Thinking: injected thinking_budget=${injected.thinking_budget} for "${resolvedModel}"`);
 	else if (injected.reasoning_effort === "high") consola.debug(`Thinking: injected reasoning_effort=high for "${resolvedModel}"`);
 	else if (reasoningUnsupportedModels.has(resolvedModel)) consola.debug(`Thinking: skipped — "${resolvedModel}" does not support reasoning`);
@@ -2463,8 +2470,7 @@ function translateToOpenAI(payload) {
 		user: payload.metadata?.user_id,
 		tools: translateAnthropicToolsToOpenAI(payload.tools),
 		tool_choice: toolChoice,
-		...!isForced && payload.thinking && { thinking: payload.thinking },
-		...!isForced && payload.thinking?.type === "enabled" && { reasoning_effort: "high" },
+		...!isForced && payload.thinking && { reasoning_effort: "high" },
 		...!isForced && payload.thinking?.budget_tokens && { thinking_budget: payload.thinking.budget_tokens }
 	};
 }