deepseek check failed: Not found (404), the endpoint does not exist - Codex
Quick fix
Codex needs the Responses API while DeepSeek only supports the Chat API — use a third-party proxy to convert between the protocols.
Symptom
Section titled “Symptom”deepseek 检查失败: Not found (404)接口地址不存在,请检查 Base URL 和 API 路径。deepseek check failed: Not found (404), the endpoint does not exist. Check the Base URL and API path.
Codex natively uses the OpenAI Responses API format while DeepSeek models only support the standard Chat Completions API. Configuring DeepSeek for Codex directly in cc-switch means the requests Codex sends are rejected by DeepSeek with a 404 or 400. cc-switch does not yet include this protocol conversion officially (tracked by #2461). The community also reports that even with a third-party conversion, unstable JSON serialisation can leave the DeepSeek prefix cache hit rate extremely low, spiking costs and misbehaving sub-agents.
Deploy a third-party protocol conversion proxy (codex-deepseek, for example) that converts the OpenAI Responses API to the DeepSeek Chat Completions API in real time.
In cc-switch, change the DeepSeek API endpoint to the local proxy address (http://127.0.0.1:15721/v1, for example).
Affected Versions
Section titled “Affected Versions”Source Issues
Section titled “Source Issues”This page is distilled from 4 real issues
- Why is the DeepSeek prefix cache hit rate still low and the cost still spiking after using a conversion proxy?
- Because the Responses-to-Chat format conversion can make byte-level JSON serialisation unstable, breaking the exact prefix cache DeepSeek relies on. That is a known defect in the conversion layer today.
- When will cc-switch support DeepSeek protocol conversion natively?
- That request is tracked in official issue #2461. Until then, use codex-deepseek, ccx, or a similar third-party tool as an interim solution.
Related problems
Section titled “Related problems”This is an unofficial community wiki with no affiliation to the cc-switch authors or the project itself. Its content is compiled from the project's public GitHub issues. This site distributes no software.