Repository navigation
Go 付费套餐 DeepSeek V4.1 Flash 严重响应延迟：99.99% 缓存命中仍等待 15–29 秒，原始 HTTP 可复现 #58003
Description
Description
我是 OpenCode Go 付费用户，DeepSeek V4.1 Flash 经常长时间没有思考或输出，已经严重拖慢正常编码和连续工具调用。以前的使用体验没有这么差。付费后仍反复空等，这种体验难以接受，请安排服务侧排查并给出处理进度。
2026-10-22 在同一电脑、相同合成输入下，绕过 Harness 直接请求两个服务，测得首个模型输出时间：
| 轮次 | DeepSeek 官方 deepseek-flash | Go deepseek-v4.1-flash |
|---|---|---|
| 1（均无缓存） | 5.997 s | 24.749 s |
| 2（均 >99.9% 缓存） | 1.896 s | 15.127 s |
| 3（均 >99.9% 缓存） | 1.831 s | 28.775 s |
每次均为 270354 input tokens、High、max_tokens=1024、stream=true，只要求回答 OK；均成功且仅输出 2 tokens，reasoning_tokens=0。两个热缓存样本中，Go 的平均起始等待约为官方的 11.8 倍。两端是各自服务公布的模型别名，未独立验证内部权重。
补测换两个新的稳定诊断会话：开启 Node 环境代理为 15.340 / 22.477 s；不启用该代理为 27.552 / 28.217 s；四次缓存均为 270336/270354，均正常 stop 和 DONE。两种模式均到 LAX，不能声称排除了所有网络因素，但这次换连接没有解决问题。
Steps to reproduce
- POST https://opencode.ai/inference/go/openai/v1/chat/completions，model=deepseek-v4.1-flash。
- 发送客户端自己的 User-Agent，以及每个诊断会话内保持不变的 x-opencode-session；本次也发送了 x-deepseek-harness-session-id。
- system 要求只回答 OK，user 为合成参考文本，将字符串
The project identifier is LATENCY_TEST. The only requested answer is OK.\n重复 18000 次；附一个无需调用的 read_file 函数工具。 - reasoning_effort=high、max_tokens=1024、stream=true、stream_options.include_usage=true；重复请求热缓存，测量从发请求到首个非空 content/reasoning delta，而非仅 HTTP 头。
- 用相同 messages/tools 对照 https://api.deepseek.com/chat/completions，model=deepseek-flash。
可直接定位的 Go 日志（均为合成数据探针）：
- 2026-10-22T02:39:46.962Z：x-opencode-log-id=
01a1617a-c606-76b8-8643-f5ced5dcef6b，15.340 s。 - 2026-10-22T02:40:02.304Z：x-opencode-log-id=
01a1617b-1aff-724c-8403-84d30d22f436，22.477 s。
OpenCode version / Operating System
使用 DeepSeek Harness 0.2.1-rc.2、pi-ai 1.1.0，Windows；以上实测使用独立 Node HTTP 脚本，不依赖 OpenCode 客户端、界面或插件。Harness 的压缩档位兼容问题已在本地修复，不作为本报告所称的 Go 延迟原因。
Expected behavior / 请求处理
请根据 log ID 区分网关排队、上游连接和首 token 等待，核对 Go 路由/容量，并给出恢复方案和时间预期。仅回复“换新会话或压缩”不足以解释本次高缓存、原始 HTTP 的对照差异。这里提供的是有限样本，不声称已经证明故意限速或具体上游根因。
相关：#57055、#51010。知道此前报告已关闭，但本账号今天仍能复现，请关联这些新日志并跟进实际恢复，而不是将关单视为解决。账户与订阅处理会另走支持邮件，不在公开 issue 披露私人信息。
Activity
opencode-agent commentedon Oct 22, 2026
Thanks for the careful measurements, @Xmage-x, and sorry that DeepSeek V4.1 Flash on OpenCode Go keeps leaving you waiting so long for the first output. Since you reproduced it with raw HTTP and no OpenCode client involved, this is happening on the hosted Go service or with an upstream provider, not in the OpenCode app.
Go service and account problems aren't handled in GitHub issues. Please email [email protected] with the details, including the two
x-opencode-log-idvalues and timestamps from this issue (01a1617a-c606-76b8-8643-f5ced5dcef6b,01a1617b-1aff-724c-8403-84d30d22f436), so the team can look up those requests. If you've already emailed, the team will get to it, especially now that we've seen it here. Please keep account, email and payment details out of this public issue and put them only in the email.You can follow live service status and past incidents at https://status.opencode.ai.
Related reports, most closely related first:
deepseek-v4.1-flashwith warm cache, reproduced with raw HTTPdeepseek-v4.1-flash, attributed to a temporary service or upstream problemdeepseek-v4.1-flashon Go than on DeepSeek's own API[opencode-go] deepseek-v4.1-flash: ~30% of API requests stall ~110s then fail with "unexpected EOF" before first token (tiny prompts included)#57270:deepseek-v4.1-flashrequests on Go stalling and then failing before the first tokenThe bot will close this issue in 1 day.
opencode-agent commentedon Oct 23, 2026
Closing this now, since OpenCode Go service and account problems are handled by email rather than in GitHub issues. Thanks again for the detailed measurements, @Xmage-x.
If you're still seeing the slow first output on
deepseek-v4.1-flash, please email [email protected]. Include the two log IDs and timestamps from this issue so the team can look up those requests. Please put any account details only in the email, not here. Live service status is at https://status.opencode.ai.