Skip to content

gproxy v2.9.8

Choose a tag to compare

View all tags
@LeenHawk LeenHawk released this 02 Sep 03:17
v2.9.8
7d47e2b

v2.9.8

Adds the September 2026 Claude Messages protocol controls and preserves their
semantics across OpenAI and Gemini transforms.

同步 2026 年 9 月 Claude Messages 协议控制项，并在 OpenAI 与 Gemini 转换中
保留对应语义。

English

Added

  • Claude Fable 5.1 and Mythos 5.1 wire support. The Claude protocol recognizes both new model IDs and the four beta headers for thinking binding controls, thinking progress updates, per-message effort, and turn-scoped system messages.
  • Typed September Messages fields. Requests model thinking.display: "updates", thinking.block_binding.prefix_mismatch_behavior, message-level output_config.effort, and clear_at. Buffered and streaming responses model input_transformations, including every documented thinking-drop reason and the message_start/message_delta override lifecycle.

Fixed

  • Cross-protocol turn rendering. Claude-to-OpenAI Chat, OpenAI Responses, and Gemini transforms no longer render system reminders after clear_at: "next_user_message" has cleared them, or emit effort-only empty system messages as prompts.
  • Per-message effort conversion. The last effort change that applies to the current user turn now becomes the target protocol's reasoning effort for content generation and token counting.

Changed

  • Input transformation diagnostics remain observable. Buffered Claude responses preserve input_transformations as a target-protocol extension instead of losing the API's explanation for dropped thinking blocks.

简体中文

新增

  • Claude Fable 5.1 与 Mythos 5.1 Wire 支持。 Claude 协议现已识别两个新模型 ID，以及 thinking binding controls、thinking progress updates、消息级 effort 和 turn-scoped system message 对应的四个 Beta Header。
  • 9 月 Messages 字段类型化。 请求现已建模 thinking.display: "updates"、thinking.block_binding.prefix_mismatch_behavior、消息级 output_config.effort 与 clear_at；非流式及流式响应现已建模 input_transformations、全部已公开的 thinking drop 原因，以及 message_start/message_delta 覆盖生命周期。

修复

  • 跨协议 Turn 渲染。 Claude 转 OpenAI Chat、OpenAI Responses 与 Gemini 时，不再渲染已被 clear_at: "next_user_message" 清除的 system 提醒，也不会把仅用于调整 effort 的空 system 消息误发成 Prompt。
  • 消息级 Effort 转换。 当前用户 Turn 已生效的最后一次 effort 调整，会映射为目标协议的 reasoning effort，同时覆盖内容生成与 Token Counting。

变更

  • 保留输入转换诊断。 Claude 非流式响应的 input_transformations 会作为目标协议扩展继续保留，不再丢失 API 对 Thinking Block 被移除原因的说明。
Assets 59
Loading