[Config]: kimi-k2.7-code fails with Moonshot invalid thinking via Codex++ proxy

[xuanweijiang]

接入模式

纯 API

上游协议

Responses API

Base URL

https://your-proxy.example.com/v1

相关 config.toml / auth.json 片段

model_provider = "custom"
model = "kimi-k2.7-code"
model_reasoning_effort = "high"
review_model = "gpt-5.5"

[model_providers.custom]
name = "custom"
wire_api = "responses"
requires_openai_auth = true
base_url = "http://127.0.0.1:57321/v1" # Codex++ local proxy; upstream custom proxy domain is configured in Codex++ and redacted

问题描述

Summary

When using kimi-k2.7-code through Codex++ with my own upstream proxy, Codex can list the model but chat requests fail with a Moonshot provider error:

Error from provider (Moonshot AI): invalid thinking: only type=enabled is allowed for this model

Environment

• OS: Windows
• Codex++: v1.2.18
• Access mode: pure API / custom upstream proxy
• Upstream protocol: Responses API
• Codex local Base URL: http://127.0.0.1:57321/v1 (Codex++ local proxy)
• Actual upstream proxy domain: self-hosted/custom domain, redacted as https://your-proxy.example.com/v1
• Model: kimi-k2.7-code
• Reasoning effort: high

What happens

1. Start Codex++.
2. Configure Codex++ to route Codex requests through my own upstream proxy domain.
3. Codex uses the local Codex++ proxy at http://127.0.0.1:57321/v1.
4. Select kimi-k2.7-code.
5. Send a normal chat message.
6. The request fails with the Moonshot error above.

Expected behavior

Codex++ should translate or sanitize the reasoning/thinking fields for Moonshot/Kimi code models before forwarding the request to the upstream proxy/provider.

For this model, Moonshot appears to reject incompatible thinking payloads and expects only an enabled thinking form, or no unsupported reasoning/thinking fields depending on the exact upstream endpoint.

Notes

• Other Kimi models such as kimi-k2.6 do not appear to trigger the same error in this setup.
• A local test proxy placed before Codex++ that only patched kimi-k2.7-code requests by removing incompatible reasoning fields and forcing thinking: { "type": "enabled" } allowed minimal test requests to pass. This suggests the failure is likely in Codex++'s request conversion/forwarding for this model family.
• No API keys, account tokens, or real upstream domains are included in this report.

日志 / 状态文件

Observed error:

Error from provider (Moonshot AI): invalid thinking: only type=enabled is allowed for this model

Codex local Base URL:
http://127.0.0.1:57321/v1

Actual upstream Base URL:
self-hosted/custom proxy domain, redacted as https://your-proxy.example.com/v1

The model list endpoint can return kimi-k2.7-code, but conversation requests fail.

提交前确认

• 我已经脱敏了 API Key、Token、账号等敏感信息。