Skip to content

gproxy v2.2.1

Choose a tag to compare

View all tags
@LeenHawk LeenHawk released this 31 Jul 08:53
v2.2.1
21b9f1a

v2.2.1

Adds native DeepSeek Responses API routing for buffered, SSE, and WebSocket-bridged clients; refreshes the OpenAI Responses schema and Claude block parsing; and fixes legacy-Claude trailing system turns plus GPT-5.6 Luna/Terra pricing.

English

Added

  • DeepSeek Responses API. The deepseek channel now sends OpenAI Responses requests directly to DeepSeek's native /responses endpoint with Bearer authentication, including non-streaming and HTTP/SSE streaming calls. Responses WebSocket clients are bridged to the same SSE transport, and settings_json.endpoints.openai_responses is available as an exact endpoint override. Existing providers retain their stored routing rules and must reset routing defaults to opt in to the new routes.

Fixed

  • Legacy Claude trailing system turns. When OpenAI Chat requests targeting pre-Opus-4.8 Claude models end with system or developer messages, those trailing messages now become user turns instead of an accidental assistant prefill that unsupported models reject. Mid-conversation system messages keep their existing assistant-turn downgrade.
  • GPT-5.6 pricing. Corrects the bundled OpenRouter price rules for GPT-5.6 Luna and Terra, including input, output, cache-read, and 30-minute cache-creation rates.

Changed

  • OpenAI Responses schema. Refreshes protocol and transform coverage for the fast service tier, the data_residency_mismatch response error, and function-call-output name / namespace fields across Responses and Compact conversions.
  • Claude protocol parsing. Claude content-block unions now dispatch once on their type tag instead of repeatedly trying every untagged variant. Unknown tags and known tags with unmodeled shapes still fall back to raw forwarding, preserving compatibility while reducing parsing work.
  • Console dependencies. Updates React and React DOM to 19.2.8.

简体中文

新增

  • DeepSeek Responses API。 deepseek channel 现会使用 Bearer 鉴权将 OpenAI Responses 请求直接发送到 DeepSeek 原生 /responses endpoint，覆盖非流式与 HTTP/SSE 流式调用。Responses WebSocket 客户端会桥接到同一 SSE transport，并可通过 settings_json.endpoints.openai_responses 配置精确 endpoint override。已有 provider 会保留数据库中的路由规则，需要重置默认路由后才能启用这些新 route。

修复

  • 旧版 Claude 的尾部 System Turn。 OpenAI Chat 请求以 system 或 developer message 结尾且目标为 Opus 4.8 之前的 Claude 模型时，这些尾部消息现在会降级为 user turn，不再形成模型不支持并会拒绝的意外 assistant prefill；mid-conversation system message 仍沿用原有的 assistant-turn 降级行为。
  • GPT-5.6 价格。 修正内置 OpenRouter 价格规则中的 GPT-5.6 Luna 与 Terra 输入、输出、Cache Read 和 30 分钟 Cache Creation 费率。

调整

  • OpenAI Responses Schema。 更新 Responses 与 Compact 的协议及转换覆盖，加入 fast service tier、data_residency_mismatch response error，以及 function-call-output 的 name / namespace 字段。
  • Claude 协议解析。 Claude content-block union 改为按 type tag 一次分派，不再依次尝试每个 untagged variant。未知 tag 或已知 tag 的未建模结构仍会回退为 raw forwarding，在减少解析开销的同时保持兼容性。
  • Console 依赖。 React 与 React DOM 更新至 19.2.8。
Assets 59
Loading