Upstream Chat Completions Streaming Response Error, Claude Code

Quick fix

Check whether the issue is caused by local proxy conversion or the upstream endpoint; adjust the endpoint or disable the local proxy for the verified scenario.

Symptom

Error outputCopy

codex快速重试连接5次后返回报错“stream disconnected before completion: Upstream Chat Completions stream ended before sending finish_reason”

English translation

After 5 rapid connection retries, codex returns the error “stream disconnected before completion: Upstream Chat Completions stream ended before sending finish_reason”

Cause

These issues involve several types of proxy problems and cannot be treated as one failure: Chat Completions streaming content split into multiple Claude Code messages, an upstream stream ending before sending finish_reason, GLM-5.3 rejecting requests that disable thinking, and garbled tool-call arguments. The reporter for #3960 said switching back to the V4 baseurl restored service; #4785 was marked by a maintainer as a duplicate of another issue, with no fix provided here. Verify each symptom separately; this error alone does not indicate a uniform GLM proxy failure.

Fix

1. 1

   If you encounter “stream disconnected before completion” while using a Zhipu Anthropic endpoint, the issue reporter said switching back to the V4 baseurl restored service; the corpus does not provide the specific baseurl value.

2. 2

   If Claude Code splits a streaming response into multiple messages after passing through a proxy, #4785 only says the issue was closed as a duplicate and points to another issue for tracking; this corpus contains no confirmed fix.

3. 3

   If you use GLM-5.3 and see error 1210 when disabling thinking, the request form verified in the discussion sets thinking.type to enabled and reasoning_effort to low.

4. 4

   If Claude Code Auto Mode shows classifier unavailable, the temporary workaround reported in #1678 is to disable the CC Switch Anthropic Base URL proxy and restart the Claude Code session; later comments dispute whether it reproduces under normal conditions.

Affected Versions

ToolClaude Code、Codex

Versioncc-switch 3.12.3、3.13.0、3.16.2、3.16.3、3.16.4、3.19.1、3.20.0（各 issue 所报版本；并非统一受影响范围）

PlatformsmacOSWindows

Source Issues

This page is distilled from 13 real issues

• #1371SOS技术求助｜CC Switch 配置智谱 API 时，地址和 Key 被自动篡改
• #1544中转代理bug
• #1678Claude Code Auto Mode Classifier Compatibility Issue
• #2172vscode中Claude code插件的回答显示会被自动打断，该如何解决？
• #2460开启代理后，模型身份认知错乱： 被强制注入 Claude 系统提示词导致自我认知冲突
• #3960stream disconnected before completion: Upstream Chat Completions stream ended before sending finish_reason
• #4600在claude code聊天框中发送带粘贴图片的消息时，配置了视觉理解MCP的模型会直接报错，导致无法触发图片理解mcp
• #4785Bug: OpenAI Chat Completions streaming is split into multiple Claude assistant messages
• #4998feat(codex): decouple openai_responses wire API from native-safe catalog; add hub-managed model capability templates
• #6508[Feature] Codex 经本地路由接 Kimi 时，将内置 web_search 映射为 Kimi 服务端 $web_search（附已验证方案）
• #6737fix(proxy): GLM-5.3 rejects forced-thinking-off requests from Claude Code
• #6798质谱GLM 5.3乱码
• #6835使用zhipu模型，代理映射一直默认使用glm-5.3

FAQ

(No FAQ yet)

Related problems

The auto mode classifier is temporarily unavailableDisable the Anthropic Base URL proxy in CC Switch and restart Claude Code; one user in the issue reported that updating CLIProxyAPI resolved the problem.1 issues mergedusage null - Claude Code Qwen DashScopeUpgrade to cc-switch v3.16.3+, which fixes null streaming usage on the OpenAI-compatible interface automatically.2 issues mergedTool calls always hang: Codex function_call loses call_idA commenter confirmed non-empty id/name checks restore tool calls. The post also suggests bypassing CC Switch Responses translation to connect directly to…1 issues merged