Cockpit Tools v1.3.72

更新日志（中文）

[1.3.72] - 2026-10-21

新增

• API 服务官方优先、本地兜底：优先使用已绑定的 ChatGPT 官方授权，遇到可恢复且可安全重试的故障后，尝试同账号本地 OAuth 通道或允许的其他账号。保留明确选择的“仅本地”、对话与工具历史、账号权限和额度限制；官方通道使用默认速度，本地回退保留请求的速度。请求详情展示实际通道及恢复结果。
• 账号测智：“鹈鹕测智”更名，新增糖果推理、不同考点的公开精选题，以及数学、代码追踪、长文账本和依赖规划固定答案题。题名明确考点，参考解法默认展开，逐字段展示结果并区分格式符合与最终答案正确性。退役题目保留原题干与判分规则供历史记录查看；支持纯文本与 HTML 并排比较，未完成请求不计分，本地结果不等同于官方榜单成绩。
• API 服务用量概览：紧凑、清晰的弹框按所选模型类别展示 GPT 账号额度、DeepSeek 余额或 Grok 配额，同时显示当前会话指标和服务统计。请求记录可切换当前会话／全部服务并继续加载，保留滚动位置；账号名称、已知套餐标签、请求与实际上游模型、授权通道、Token 和性能与 API 服务口径保持一致。GPT API Key 账号数独立于请求统计读取，按当前服务和 Key 的账号范围计数，切换会话或断连保留已确认数量；弹框收窄至 660px，请求记录紧凑排版，必要时在列表内横向滚动。余额按供应商、计入余额的账号数和金额对齐展示，两页共用按概览内容自适应的高度，上限 700px 且不超过窗口高度的 85%，指标单位缩小，概览合并重复模型，点击弹框外部可关闭。无法关联或缺少有效计时的数据保持未知，历史派发依据不完整时标记为已核实统计。
• Codex 启动预览常驻“重新授权”入口，账号未出现授权失败时也可主动授权，覆盖原账号凭据。

变更

• API 服务额度、速度和缓存改为统一浅色底状态条，左侧以单个紫色小圆点点缀；速度显示蓝色，额度按显示值沿用账号总览配色：≥80% 绿、≥40% 黄、≥10% 橙、低于 10% 红，无数据显示“—”，点击可查看详情并刷新。 客户端输出速度统一使用 API 服务的首响后耗时口径，按当前会话近期有效请求加权，无法关联会话或缺少有效计时时显示“—”。
• API 服务客户端隐藏 ChatGPT 订阅额度提示；普通账号仍保留原生提示及真实限制。

移除

• 取消 API 服务启动、页面进入和普通账号添加时的官方授权自动提示；成员管理仍可手动授权，批量加入时继续处理正常账号并报告逐项失败，保留已保存的接入方式。
• 移除两种 API 服务引擎的 OAuth 额度保留、阈值和安全垫；旧保留设置不再参与调度，真实额度耗尽、Key 范围和账号并发限制继续生效。

修复

• 修正 Windows 更新包的安装格式标识，避免自动更新误选安装包类型。
• API 服务绑定的 Codex 客户端恢复原生听写入口与转写流程，遵循客户端原生能力和设置，不再接管听写；保留通用音频 API 与实时语音。
• Rust 与 Go API 请求在流握手后收到额度错误时也可自动恢复，覆盖未绑定官方授权的账号；先冷却耗尽的通道，再尝试安全回退，并遵守上游恢复时间，保留对话和工具安全校验。已产生输出或执行状态不明的请求不重放；未发送的候选不记失败、不占上游重试次数。
• 修复混合模式下绑定的 ChatGPT 额度耗尽后，独立 API 会话的发送按钮仍被禁用的问题；保留未保存草稿、富文本、活动对话及真实额度限制。
• Cockpit 管理的 Codex 实时语音连接在重新接管后同步当前 API 服务端口，保留自定义语音地址。

Changelog (English)

[1.3.72] - 2026-10-21

Added

• Official-first API Service routing: Prefer linked official ChatGPT access, then recover through the same account's local OAuth route or another permitted account after a safe, recoverable failure. Preserve explicit local-only choices, conversation and tool history, account permissions and quota limits. Official access uses default speed; local fallback retains the requested speed. Request details show the actual route and recovery results.
• Account tests: Rename Pelican test and add candy reasoning, distinct public benchmark cases, and fixed-answer mathematics, code tracing, long-context ledger and dependency-planning tasks. Show descriptive task names, expanded references and per-field results, distinguishing format compliance from final-answer correctness. Retired prompts and their original graders remain available in history. Plain-text replies can be compared alongside HTML; incomplete requests are not scored, and local results are separate from official leaderboards.
• API Service usage overview: A compact, readable dialog shows the selected model family’s GPT account quotas, DeepSeek balance or Grok quota alongside current-chat metrics and service statistics. Switch between current-chat and service-wide request records and load more without losing the scroll position. Account names, known subscription labels, requested and upstream models, authorization routes, token usage and performance follow API Service conventions. GPT API Key account counts load independently of request statistics and follow the current service/key account scope, retaining confirmed counts across chat changes and disconnections. The 660px dialog keeps request fields compact with internal horizontal scrolling when needed. Balance rows align provider, contributing account count and amount; dialog height adapts to overview content across both tabs, capped at 700px and 85% of the viewport, metric units are smaller, repeated overview models are collapsed, and clicking outside closes the dialog. Unlinked or unmeasured data stays unknown; incomplete historical statistics are marked as verified subsets.
• Keep Reauthorize available in the Codex launch preview, allowing credentials to be replaced for the existing account before an authorization failure.

Changed

• Show API Service quota, speed and cache in one status strip with a light surface and a single purple accent dot; use blue for speed and account-overview colors for displayed quota percentages: green at ≥80%, yellow at ≥40%, orange at ≥10%, red below 10%, and “—” for missing data. Click to view details and refresh. Client output speed uses the same post-first-response timing as API Service, weighted over the current chat’s recent valid requests; unmatched chats or missing timing show “—”.
• Hide the ChatGPT subscription usage banner in API Service clients while preserving native notices and actual limits for ordinary accounts.

Removed

• Remove automatic official-authorization prompts on API Service startup, page entry and ordinary account additions. Manual authorization remains in member management; batch additions continue for healthy accounts and report individual failures without changing saved access choices.
• Remove OAuth quota reserves, thresholds and safety margins from both API Service engines. Existing reserve settings no longer affect routing; actual quota exhaustion, key scopes and account concurrency still apply.

Fixed

• Correct Windows update-package installer identification so automatic updates can select the appropriate installer format.
• Restore native dictation controls and transcription in Codex clients bound to API Service, following the client’s native capabilities and settings instead of overriding dictation. General audio APIs and realtime voice remain available.
• Recover Rust and Go API requests when quota errors arrive after stream handshakes, including accounts without official authorization. Cool down exhausted access before trying a safe fallback, respect upstream reset times and retain conversation and tool safety checks. Requests with delivered output or uncertain execution are not replayed; unsent candidates create no failure records and consume no upstream retry attempts.
• Fix the send button remaining disabled for independent API conversations when the linked ChatGPT quota is exhausted in mixed mode. Preserve unsaved drafts, rich content, active conversations and actual quota limits.
• Keep Cockpit-managed Codex realtime voice connections on the current API Service port after reconnection, while preserving custom voice endpoints.