更新日志（中文）
[1.3.76] - 2026-10-30
新增
- Anthropic 接入：Codex 模型供应商和 API 服务支持 Anthropic Messages，可配置鉴权、获取或手动填写模型，并使用 Claude 模型对话。支持流式回复、工具调用、网页搜索及上下文压缩；APIKEY.FUN 的 Anthropic 账号可查看余额和用量，实际能力取决于上游服务。
- OpenAI 官方 API 用量：使用现有 API Key 查询按 UTC 统计的消费金额、请求数和 Token；缺少用量读取权限时提供管理员密钥配置指引。官方消费不显示为剩余余额。
- Codex 项目与会话修复：快速修复可恢复当前本地项目和会话归属；完整修复还支持历史快照和会话索引恢复，保留云项目、凭据与聊天正文，并提供本地备份和中断恢复。感谢 @harryzhou0106-svg 提供恢复参考（#2955）。
- 我的皮肤：保存、切换、重命名、更换和删除本地背景图片，每张皮肤独立保存显示参数；恢复默认不会删除已保存的皮肤。
- 请求诊断：无需记录请求正文即可查看请求、历史、工具和指令的体积，以及账号满载恢复情况，便于排查大请求失败。
- macOS 正式包签名：Apple Silicon、Intel 和 Universal 正式包支持 Developer ID 签名与 Apple 公证，签名版本可正常安装，无需手动移除隔离属性。
变更
- API 服务：统一使用 Rust 引擎，原有引擎选择自动迁移，保留账号、密钥、端口和会话历史；实例网关和供应商测试也使用 Rust，安装包不再依赖 Go sidecar。
- API 协议兼容：增强 Responses、HTTP/SSE 和 WebSocket 对话兼容性，支持连接恢复、续接与 steering，改善工具调用、压缩响应、旧式请求和用量统计的处理；已产生输出或工具执行迹象的请求不会被自动重放。
- 账号与供应商管理：账号总览和模型供应商共用完整编辑器，可分别配置供应商默认模型和单个 Key 的模型，保留账号绑定和模型同步设置。已有供应商在后台同步时仍可使用，保存失败可在弹框内重试。
- 账号用量展示：官方 OpenAI API 账号展示今日消耗、请求和 Token，并按官方地址提供独立标签与筛选；其他供应商保留余额和额度展示。详情继续展示上游提供的用量、套餐和模型统计；额度查询失败且无缓存时显示失败提示。
- 账号卡片：APIKEY.FUN 卡片采用金色品牌名称和钥匙背景，Codex 账号订阅标识靠右对齐。
- 更新记录：打开更新记录或更新弹框时读取 GitHub Release 最新说明，远端修改或删除在下次获取后生效；加载失败可在弹框内重试。
- 项目主页：APIKEY.FUN 赞助商链接更新为 apikey.fan，保留赞助商名称和邀请参数。
修复
- Codex 速度控件：适配新版 ChatGPT/Codex 客户端，恢复模型菜单中的原生 Fast 控件；额度提示暂不兼容时，速度控件仍可独立启用。
- API 服务续聊与换号：响应未完成、断线重连、服务重启或重试失败后，可恢复已验证的完整历史；原账号额度耗尽时，在安全条件满足且有可用账号时自动换号。续接引用失效或 GPT 拒绝可定位的推理密文时，可在输出前安全恢复并重试；保留指令、可读摘要和配对工具结果，避免重复执行。
- API 服务额度与满载恢复：原账号额度重置时间已知时，可在等待预算内重新检查；满载时优先尝试可用账号，无替代账号时可进行有限的同账号重试。超过预算或无法安全重试时返回相应错误，不误报账号绑定失效。参考 #2962。
- API 服务连接恢复：识别 HTTP/2 TLS 断流并清理故障连接；同一路径短时间连续失败时暂时使用 HTTP/1.1，再恢复自动协议协商。保留代理设置和已输出的工具调用，不自动重放已执行操作。
- API 服务续接兼容：识别旧式
session_id、合法的桌面和终端宿主工具结果，以及已验证历史中省略的空工具参数，减少续聊和换号时的历史误判，同时保留调用方与账号隔离。 - Codex 模型：官方 DeepSeek 账号使用原生 Responses 并保留思考档位，兼容第三方供应商后续 DeepSeek 型号；为 GPT-6.1 Terra 补齐 Ultra 推理档位。启用供应商模型同步时可恢复被受管速度目录接管的模型列表，保留用户自定义目录及模型上下文设置。
- Codex 统计与启动：补齐续写和分叉会话漏算的 Token；仪表盘按实际存在的额度窗口推荐账号。启动预览打开后若实例配置变化，启动和重启使用当前配置，不覆盖过期草稿。
- 后台刷新与编辑：供应商卡片刷新时保持可见和滚动位置；API 服务调度、价格、超时、模型规则、Key、端口、代理、分组选择和 OAuth 绑定的未保存编辑不再被后台刷新重置，保存失败在当前弹框内提示。
- Antigravity 额度与唤醒：悬浮窗完整显示已有额度窗口，修正用量趋势图的日期命中和提示位置；自动唤醒的模型列表与请求参数对齐官方客户端，覆盖 Gemini 3.8 Flash、Gemini 4 Argon 和 Claude 5.5 各档位。
Changelog (English)
[1.3.76] - 2026-10-30
Added
- Anthropic access: Codex model providers and API Service support Anthropic Messages, with configurable authentication, model discovery or manual catalogs, and Claude conversations. Streaming, tools, web search and context compaction are supported; APIKEY.FUN Anthropic accounts can display balance and usage. Actual capabilities depend on the upstream service.
- OpenAI official API usage: Use the existing API Key to query spend, requests and tokens reported in UTC. An admin-key setup guide is available when usage-read permission is missing. Official spend is not displayed as remaining credit.
- Codex project and chat repair: Quick repair restores current local projects and chat assignments; full repair also supports historical snapshots and chat-index recovery. Cloud projects, credentials and chat contents are preserved, with local backups and interrupted-write recovery. Thanks @harryzhou0106-svg for the recovery reference (#2955).
- My skins: Save, switch, rename, replace and delete local background images, with independent settings per skin. Restoring defaults keeps saved skins.
- Request diagnostics: Inspect request, history, tool and instruction sizes and account capacity recovery without recording request bodies, making large-request failures easier to investigate.
- macOS release signing: Apple Silicon, Intel and Universal releases support Developer ID signing and Apple notarization. Signed releases can be installed normally without manually removing quarantine attributes.
Changed
- API Service: Use the Rust engine exclusively. Existing engine selections migrate while accounts, keys, ports and conversation history are preserved. Instance gateways and provider tests also use Rust, and installers no longer depend on the Go sidecar.
- API protocol compatibility: Improve Responses, HTTP/SSE and WebSocket conversations with connection recovery, continuations and steering, plus better handling of tools, compressed responses, legacy requests and usage accounting. Requests with output or evidence of tool execution are not automatically replayed.
- Account and provider management: Account overview and model providers share a complete editor for provider defaults and per-key models, while retaining account bindings and model synchronization settings. Existing providers remain available during background reconciliation, and failed saves can be retried in the dialog.
- Account usage displays: Official OpenAI API accounts show today's spend, requests and tokens, with a separate badge and filter based on the official endpoint. Other providers retain balance and quota displays. Details continue to show available usage, subscription and model statistics; a failed quota query without cached data shows an error.
- Account cards: APIKEY.FUN cards use a gold brand name and key background, and Codex subscription labels align to the right.
- Release notes: Read current GitHub Release notes when opening release history or the update dialog, so remote edits and removals appear on the next fetch. Failed loads can be retried in the dialog.
- Project homepage: Update the APIKEY.FUN sponsor link to apikey.fan while retaining the sponsor name and referral parameter.
Fixed
- Codex speed controls: Restore the native Fast control in the model menu for newer ChatGPT/Codex clients. Speed controls remain available independently when quota UI compatibility is unsupported.
- API Service continuations and account switching: Recover verified complete history after incomplete responses, reconnects, service restarts or failed retries. When the original account runs out of quota, switch to an available account when safe. Invalid continuation references and identifiable GPT reasoning ciphertext rejections can be safely recovered and retried before output; instructions, readable summaries and paired tool results are preserved to avoid duplicate execution.
- API Service quota and capacity recovery: When the original account's reset time is known, recheck it within the configured wait budget. On capacity errors, try an available account first and use bounded same-account retries when no alternate is available. Return an appropriate error when the budget is exceeded or a retry is unsafe, rather than misreporting an invalid account binding. Based on #2962.
- API Service connection recovery: Detect HTTP/2 TLS stream failures and retire the affected connection. Repeated failures on the same route temporarily use HTTP/1.1 before restoring automatic protocol negotiation. Proxy settings and emitted tool calls are preserved, without automatically replaying executed actions.
- API Service continuation compatibility: Recognize legacy
session_id, valid desktop and terminal host tool results, and omitted empty tool arguments in verified history, reducing false rejections during continuations and account switching while preserving caller and account isolation. - Codex models: Official DeepSeek accounts use native Responses and retain reasoning levels; later DeepSeek models from third-party providers are recognized. GPT-6.1 Terra gains the Ultra reasoning option. Provider model synchronization can recover model lists taken over by the managed speed catalog while preserving user catalogs and model context settings.
- Codex statistics and launch: Include missing Token usage from continuation and branch sessions, and base dashboard recommendations on quota windows actually present. If instance settings change after opening the launch preview, start and restart use the current settings without overwriting a stale draft.
- Background refresh and editing: Keep provider cards and scroll position stable during refresh. Unsaved API Service routing, pricing, timeout, model, key, port, proxy, group selection and OAuth binding edits survive background updates; save failures appear in the active dialog.
- Antigravity quotas and wakeup: Show all available quota windows in floating cards and correct date hit-testing and tooltip positions in usage trends. Align automatic wakeup model lists and request parameters with the official client, covering Gemini 3.8 Flash, Gemini 4 Argon and Claude 5.5 effort tiers.