Skip to content

gproxy v2.6.6

Choose a tag to compare

View all tags
@LeenHawk LeenHawk released this 15 Aug 12:23
v2.6.6
a516160

v2.6.6

Adds first-class video and media routing, Kimi API and subscription OAuth channels, and safer Claude hosted-tool conversion for OpenAI Responses.

English

Added

  • Video protocol and routing. OpenAI-compatible video create, retrieve, list, delete, download, remix, character, edit, and extend operations are modeled and routed where supported. OpenRouter video generation, Gemini Veo through AI Studio and Vertex AI, AWS Bedrock Nova Reel, xAI, and Grok Build receive provider-specific endpoint and payload shaping.
  • Cross-provider image and audio routing. Provider media adapters now route supported image generation/editing, speech, and transcription operations with the required JSON, multipart, authentication, and response handling.
  • Kimi channels. Added the API-key based Kimi API channel and the Kimi Code subscription channel with OAuth login, token refresh, account usage discovery, and credential health handling.
  • Service-tier pricing. Pricing rules can apply model-specific multipliers or per-token rates to provider service tiers, and final settlement uses the tier actually reported by the upstream when available.

Fixed

  • Claude hosted-tool results. Claude WebSearch and WebFetch results now complete their matching OpenAI Responses web_search_call; search result URLs are retained as action sources, fetch URLs remain open_page actions, failures remain failed, and streams emit the dedicated in-progress, searching, and completed lifecycle events.
  • Explicit semantic-loss diagnostics. Unsupported WebSearch, WebFetch, TextEditor, Memory, citation, and hosted-result fields no longer disappear silently. Detailed request, response, and stream conversion APIs report structured diagnostics while still returning the best representable target value.

Changed

  • Web tool approximation visibility. Collapsing Claude WebSearch/WebFetch definitions into one Responses tool, or expanding one Responses WebSearch tool into Claude WebSearch and WebFetch, is now reported as a lossy conversion.

简体中文

新增

  • 视频协议与路由。 新增 OpenAI 兼容的视频创建、查询、列表、删除、下载、混剪、角色、编辑及续写操作模型，并按渠道能力路由。OpenRouter 视频生成、AI Studio 与 Vertex AI 的 Gemini Veo、AWS Bedrock Nova Reel、xAI 及 Grok Build 均会执行各自所需的端点和请求体转换。
  • 跨供应商生图与音频路由。 支持的渠道现在可以路由图片生成/编辑、语音合成及音频转写，并处理对应的 JSON、multipart、认证和响应格式。
  • Kimi 渠道。 新增使用 API Key 的 Kimi API 渠道，以及支持 OAuth 登录、token 刷新、账户用量查询和凭据健康管理的 Kimi Code 订阅渠道。
  • 服务层级计价。 价格规则可按模型为供应商 service tier 配置倍率或各 token 类别价格；上游报告实际层级时，最终结算会优先采用该层级。

修复

  • Claude 托管工具结果。 Claude WebSearch 与 WebFetch 结果现在会完成对应的 OpenAI Responses web_search_call；搜索结果 URL 会保留到 action sources，抓取 URL 保持为 open_page，失败状态不会被抹平，流式响应也会输出专用的 in-progress、searching 与 completed 生命周期事件。
  • 显式语义损失诊断。 无法表达的 WebSearch、WebFetch、TextEditor、Memory、citation 及托管工具结果字段不再静默消失。详细请求、响应和流式转换接口会在返回最佳可表达结果的同时给出结构化 diagnostics。

变更

  • Web 工具近似转换可见性。 Claude WebSearch/WebFetch 合并为一个 Responses 工具，或一个 Responses WebSearch 展开为 Claude WebSearch 与 WebFetch 时，现在会明确报告为有损转换。
Assets 59
Loading