记录下3月发现的两个东西（至于为什么5月才发出来你别管.jpg）
new-api nginx配置
起因是在RooCode中使用CLIProxyAPI反代的Codex GPT-5.2 xhigh（没错当时还没有GPT-5.4/5.5），前级还有一个new-api+nginx，发现总是思考到一半就报错重试，导致生成质量下降
看CLIProxyAPI的状态有很多“context canceled”（“stream error”应该就是Codex测返回的问题了），这玩意一半意味着客户端主动取消，检查日志最后发现是new-api测发现客户端关闭了连接
CLIProxyAPI和new-api管理界面截图：
log如下：
new-api | [INFO] 2026/03/11 - 16:57:11 | 202603110855175814086946gw4RK6c | client disconnected
new-api | [ERR] 2026/03/11 - 16:57:16 | 202603110855175814086946gw4RK6c | timeout waiting for goroutines to exit
new-api | [ERR] 2026/03/11 - 16:57:16 | 202603110855175814086946gw4RK6c | scanner error: read tcp 172.18.0.3:53944->172.18.0.10:8317: use of closed network connection
new-api | [INFO] 2026/03/11 - 16:57:16 | 202603110855175814086946gw4RK6c | 预扣费后补扣费：＄0.025010（实际消耗：＄0.025010，预扣费：＄0.000000）
new-api | [INFO] 2026/03/11 - 16:57:16 | 202603110855175814086946gw4RK6c | record consume log: userId=1, params={"channel_id":54,"prompt_tokens":137898,"completion_tokens":627,"model_name":"cpac/gpt-5.2(xhigh)","token_name":"xxx","quota":12505,"content":"","token_id":3,"use_time_seconds":119,"is_stream":true,"group":"svip","other":{"admin_info":{"local_count_tokens":true,"use_channel":["54"]},"billing_source":"wallet","cache_ratio":1,"cache_tokens":0,"completion_ratio":8,"frt":26282,"group_ratio":0.5,"model_price":-1,"model_ratio":0.175,"request_conversion":["OpenAI Compatible"],"request_path":"/v1/chat/completions","user_group_ratio":-1}}
new-api | [GIN] 2026/03/11 - 16:57:16 | relay | 202603110855175814086946gw4RK6c | 200 | 1m59.025445461s | xxx | POST /v1/chat/completions
new-api | [SYS] 2026/03/11 - 16:57:19 | batch update started
new-api | [SYS] 2026/03/11 - 16:57:19 | batch update finished cli-proxy-api | [2026-03-11 16:54:16] [3e4629b7] [debug] [conductor.go:2748] Use OAuth provider=codex auth_file=codex-xxx-free.json for model cpac/gpt-5.2(xhigh)
cli-proxy-api | [2026-03-11 16:54:16] [--------] [debug] [apply.go:146] thinking: config from model suffix | provider=codex model=gpt-5.2(xhigh) mode=level budget=0 level=xhigh
cli-proxy-api | [2026-03-11 16:54:16] [--------] [debug] [apply.go:197] thinking: processed config to apply | provider=codex model=gpt-5.2 mode=level budget=0 level=xhigh
cli-proxy-api | [2026-03-11 16:54:53] [--------] [info ] [gin_logger.go:93] 200 | 78ms | xxx | GET "/v0/management/usage"
cli-proxy-api | [2026-03-11 16:55:17] [9319b13e] [debug] [conductor.go:2748] Use OAuth provider=codex auth_file=codex-xxx-free.json for model cpac/gpt-5.2(xhigh)
cli-proxy-api | [2026-03-11 16:55:18] [--------] [debug] [apply.go:146] thinking: config from model suffix | provider=codex model=gpt-5.2(xhigh) mode=level budget=0 level=xhigh
cli-proxy-api | [2026-03-11 16:55:18] [--------] [debug] [apply.go:197] thinking: processed config to apply | provider=codex model=gpt-5.2 mode=level budget=0 level=xhigh
cli-proxy-api | [2026-03-11 16:55:35] [3e4629b7] [info ] [gin_logger.go:93] 200 | 1m19s | 172.18.0.3 | POST "/v1/chat/completions"
cli-proxy-api | [2026-03-11 16:57:16] [9319b13e] [info ] [gin_logger.go:93] 200 | 1m58s | 172.18.0.3 | POST "/v1/chat/completions"
cli-proxy-api | [2026-03-11 16:57:23] [bc46cb1c] [debug] [conductor.go:2748] Use OAuth provider=codex auth_file=codex-xxx-free.json for model cpac/gpt-5.2(xhigh)
cli-proxy-api | [2026-03-11 16:57:24] [--------] [debug] [apply.go:146] thinking: config from model suffix | provider=codex model=gpt-5.2(xhigh) mode=level budget=0 level=xhigh
cli-proxy-api | [2026-03-11 16:57:24] [--------] [debug] [apply.go:197] thinking: processed config to apply | provider=codex model=gpt-5.2 mode=level budget=0 level=xhigh
cli-proxy-api | [2026-03-11 16:57:26] [bc46cb1c] [info ] [gin_logger.go:93] 200 | 3.054s | 172.18.0.3 | POST "/v1/chat/completions"
cli-proxy-api | [2026-03-11 16:57:28] [6d0aabf8] [debug] [conductor.go:2748] Use OAuth provider=codex auth_file=codex-xxx-free.json for model cpac/gpt-5.2(xhigh)
cli-proxy-api | [2026-03-11 16:57:28] [--------] [debug] [apply.go:146] thinking: config from model suffix | provider=codex model=gpt-5.2(xhigh) mode=level budget=0 level=xhigh
cli-proxy-api | [2026-03-11 16:57:28] [--------] [debug] [apply.go:197] thinking: processed config to apply | provider=codex model=gpt-5.2 mode=level budget=0 level=xhigh后面切换了CLIProxyAPI直连就没有“context cancelled”的问题了，思来想去可能是因为GPT-5.2 xhigh在思考时是以一段一段输出的，如果段之间超过一定秒数（默认60秒）没有返回新的内容，nginx就有可能把这个连接掐掉，导致new-api也随之关闭连接
最后在nginx的location中加上了以下内容修复：
location / {
proxy_pass http://new-api;
client_max_body_size 200m;
proxy_set_header Referer $http_referer;
proxy_set_header X-Real-IP $remote_addr;
proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
proxy_set_header X-Forwarded-Proto $scheme;
proxy_set_header Host $http_host;
proxy_buffering off;
proxy_cache off;
proxy_read_timeout 300s;
proxy_send_timeout 300s;
chunked_transfer_encoding on;
}openwrt 硬件NAT
这个是因为在VR串流时，发现一旦PC有下载，串流就会卡到用不了（网络总流量只有20MiB/s左右，连接数在1000左右，我寻思这也不多啊…）
路由器是R3P（openwrt设备页面），我寻思虽然是N年前的老登了，但也不至于内网1G都跑不满吧…结果在使用时top看了一下，softirq直接占用25%（1个满物理核心）
最后搜索发现新版OpenWRT（好像也不是很新了，但是位置和OpenWRT 17相比也变了位置）的硬件NAT在Firewall下的一个选项里藏着：
修改了之后立即就好了，内网也轻松达到1Gbps，甚至Steam下载都能超过500M了（原来被告知这个B出租屋装的是300M，一直没在意过，结果合着是路由器一直软NAT导致的，真是无敌了）
Views: 0