[Bug] Infinite hanging and quota surge with gemini-claude-opus-4-5-thinking in Claude Code

[KnightCat9]

Is it a request payload issue?
[ ] Yes, this is a request payload issue. I am using a client/cURL to send a request payload, but I received an unexpected error.
[x] No, it's another issue.

Describe the bug
When using the specified model in Claude Code, the terminal occasionally hangs in an infinite waiting state. However, in the CLI Proxy API Management Center (CPAMC), the request is marked as "Success," and the quota/usage statistics continue to rise indefinitely. This issue is sporadic but becomes persistent in long-context conversations.

CLI Type
claude code

Model Name
gemini-claude-opus-4-5-thinking (via Antigravity)

LLM Client
claude code

Request Information
Using CLI Proxy API Version: v6.6.80. (No specific cURL provided, issue observed during streaming in Claude Code).

Expected behavior
The response should complete normally, and the usage/quota statistics in CPAMC should stop increasing once the model finishes generating.

Screenshots
N/A

OS Type

OS: Linux Ubuntu
Version: 24.04
Additional context
It is suspected that the 2API format provided by Antigravity is incompatible with the response format required by Claude Code, causing the client to wait indefinitely for a stream end signal that never arrives, while the proxy keeps billing.

[Tuanm]

Yes, I got the same issue. Sometimes, restarting the CLIProxyAPI server can temporarily fix the issue.

[PancakeZik]

Mine started doing that yesterday. I rolled back a few commits and it's working normally again

[leaph]

@PancakeZik to which version did you roll back?

[PancakeZik]

@PancakeZik to which version did you roll back?

I build from source, but I just picked a commit from Friday and it's fine. Sorry don't know the version

[KkkzzkkK]

I got the same issue！！！

[vanphappi]

I got the same issue

[rezalast4m]

+1

[KnightCat9]

我在同样的使用环境下使用su-kaka/gcli2api最新的版本的 流式抗截断/claude-opus-4-5-thinking 时则不会再出现这个问题
Under the same environment, I do not encounter this issue when using the latest version of su-kaka/gcli2api with the 流式抗截断/claude-opus-4-5-thinking model.

[luispater]

What is the ANTHROPIC_DEFAULT_HAIKU_MODEL in your settings.json?

[luispater]

#848

[KnightCat9]

What is the ANTHROPIC_DEFAULT_HAIKU_MODEL in your settings.json?

I have tried setting the ANTHROPIC_DEFAULT_HAIKU_MODEL to Gemini 3 Flash (from Gemini CLI), and it responds normally without any issues.

However, the problem specifically persists when using any Claude model under Antigravity in Claude Code. As described in the original issue:

The client hangs indefinitely, waiting for a response.
During this period, the logs show multiple successful API responses have already been processed.
This is different from issue #848. While #848 mentions a quota surge, my issue is a functional failure where the model is completely unusable (infinite waiting), not just an issue with billing/usage statistics.

[KnightCat9]

I have found the root cause of the issue after checking the logs in the Claude Code CLI (rather than the VS Code plugin).

I encountered a 554 error on the client side. It turns out that my proxy was behind EdgeOne, which has a default origin timeout of 15 seconds. If the origin server takes longer than 15s to respond, EdgeOne terminates the connection.

After following a community guide (https://linux.do/t/topic/923049) and increasing the origin timeout to 600s, the issue no longer occurs.

However, this raises a question:
Why do Gemini models work perfectly fine within the 15s limit, while Claude models (via Antigravity) frequently trigger this timeout? Does this imply that the Claude models provided via this route are not using "true" streaming (i.e., the response is being buffered instead of being flushed to the client immediately)? If the proxy waits for a large chunk of data before sending it out, it would easily exceed the 15s threshold.

我在使用命令行版的 Claude Code（而非 VS Code 插件）时发现了客户端报 554 错误。

经排查，这可能是因为我在代理前套了一层 EdgeOne，而 EdgeOne 的默认回源超时时间是 15秒。如果回源响应超过 15 秒没有数据返回，连接就会断开。我参考了L站的教程 (https://linux.do/t/topic/923049) 将回源超时改为 600秒 后，问题就不再复现了。
但是为什么我调用 Gemini 系列模型时不会触发这个超时，而使用 Antigravity 下的 Claude 模型时却频繁触发？这是否意味着这些 Claude 模型实现的并不是“真流式”传输？导致回源时间超过 15 秒从而触发报错。

[GitCourser]

@KnightCat9
你直接写中文不就完了,怎么还写双语,作者是自己人,而且也在L站...