Skip to content

gproxy v2.3.2

Choose a tag to compare

@LeenHawk LeenHawk released this 07 Aug 07:35

v2.3.2

Adds route affinity, richer model capability metadata, cross-provider Fast mode, and Codex subscription surfaces, with fixes for Claude cache accounting and the Console dark theme.

English

Added

  • Route-member affinity. Routes can keep requests from the same user key, or from an explicitly selected request header, on a consistent route member while retaining normal health checks and failover behavior.
  • Richer provider-model capabilities. Provider models can now expose context-window, input/output token limits, and thinking/reasoning capabilities through the database, admin API, model catalog, and Console editor.
  • Cross-provider Fast mode. Fast/service-tier intent is translated across OpenAI, Claude, and Gemini protocol conversions instead of being limited to a single provider family.
  • Codex subscription/control surfaces. Codex channels expose the subscription-related upstream operations needed by supported clients, together with additional request shaping and profile metadata handling.
  • Richer Claude Code OAuth profile metadata. Claude Code credentials now retain more subscription/profile information for management and display.

Fixed

  • Claude streaming cache TTL accounting. One-hour cache creation tokens are no longer folded into the five-minute cache bucket when usage is collected from streamed responses.
  • Codex reasoning input sanitization. Unsupported status fields are stripped from reasoning items before forwarding them upstream, avoiding request rejection while preserving the reasoning content.
  • Console dark mode. The dark palette is softer and the dashboard usage chart now consumes OKLCH theme tokens correctly, restoring readable axes, grids, tooltips, and request-count curves.
  • Release target metadata. Cross-built binaries now receive the intended build version/target metadata used by self-update reporting.

Changed

  • Channel metadata no longer carries provider_family. Provider-family classification was removed from the public channel metadata path; protocol transforms continue to operate from the actual endpoint/protocol semantics instead.
  • Star-history automation is isolated from normal development branches. The scheduled chart refresh now publishes from its dedicated branch rather than generating routine README churn on main/staging.

简体中文

新增

  • 路由成员粘性。 路由现在可以按同一 user key,或按显式指定的请求 header,将请求稳定地落到同一路由成员,同时继续保留正常的健康检查与故障切换行为。
  • 更完整的 Provider Model 能力信息。 Provider Model 现在可在数据库、Admin API、模型目录和 Console 编辑器中记录并展示总上下文、输入/输出 token 上限,以及 thinking/reasoning 能力。
  • 跨 Provider Fast mode。 Fast/service-tier 语义现在会在 OpenAI、Claude、Gemini 协议转换之间正确映射,不再局限于单一 provider family。
  • Codex 订阅/控制接口。 Codex 渠道补齐受支持客户端所需的订阅相关上游操作,并完善请求整形与 profile metadata 处理。
  • 更丰富的 Claude Code OAuth profile metadata。 Claude Code 凭证会保留更多订阅与 profile 信息,供管理与展示使用。

修复

  • Claude 流式缓存 TTL 用量统计。 从流式响应采集用量时,1 小时缓存写入 token 不再错误计入 5 分钟缓存桶。
  • Codex reasoning 输入清理。 转发上游前会移除 reasoning item 中不受支持的 status 字段,避免请求被拒绝,同时保留 reasoning 内容本身。
  • Console 深色模式。 深色主题改为更柔和的配色;首页用量图表也改为正确直接使用 OKLCH 主题变量,坐标轴、网格、tooltip 与请求数曲线恢复清晰可读。
  • Release target 元数据。 Cross 构建现在会正确注入 build version/target 信息,供自更新与版本报告使用。

变更

  • Channel metadata 移除 provider_family 公共渠道元数据路径不再携带 provider family 分类;协议转换仍按实际 endpoint/protocol 语义正常工作。
  • Star History 自动化与开发分支隔离。 定时图表刷新改为在独立分支发布,不再给 main / staging 制造周期性的 README 更新提交。