VSCode Codex Speed

在状态栏查看 Codex 输出速度，包括 Remote SSH 服务器上的会话。

See Codex output speed in your VS Code status bar, including sessions on a Remote SSH server.

简体中文 · English · GitHub · 反馈问题 / Report an issue

Codex 85.0 tok/s          输出速度 / Measured rate
Codex ~85.0 tok/s         生成中估算 / Generation estimate
Codex 0.0 tok/s          无当前输出 / No current output

简体中文

能帮你做什么

• 直接在状态栏查看 Codex 的输出 token/s，悬停查看最近调用的模型、输出数量和耗时。
• 在 Remote SSH 窗口读取服务器上的 Codex 日志，不用手动同步日志或再次配置 SSH 连接。
• 汇总多个会话和子 agent，也可以只查看当前工作区中的会话。
• 点击状态栏打开仪表盘：数字读数、1／2／5 分钟曲线、会话列表及可切换范围的用量。

这是独立的社区扩展，不属于 OpenAI 官方产品。界面默认跟随 VS Code 显示语言，支持英文、简体中文和繁体中文；其他语言使用英文。

设置 codexSpeed.language 可选 auto（默认）、en、zh-cn 或 zh-tw。手动切换立即更新状态栏、悬浮提示、仪表盘及诊断提示，无需重载。Remote SSH 跟随客户端语言，手动选择请写入远程设置。命令面板标题和设置说明始终跟随 VS Code 显示语言。

仪表盘

悬停显示简洁摘要，点击状态栏打开完整仪表盘。核心测速采用与 Agent Speed 相同的 Codex 口径：三秒窗口合计实测与在途估算；完成计数到达后按调用区间重建历史，替换此前估算。曲线为最近三百个一秒桶，蓝线是实测部分，蓝绿色虚线为含估算的合计；悬停或方向键可查看每秒组成与占比。主读数是三秒平均，不必与某个一秒值完全相等。

用量支持近 30 分钟、今天、近 7 天和近 30 天，默认“今天”，按客户端时区划分。近 7／30 天是滚动小时窗口。输入可能含缓存 token，平均速度取耗时已知的完成调用；五分钟监测峰值可能含估算，并会随曲线重建而校正。

首次使用会分批索引历史，显示进度及不完整提示。缓存保存计数与读取断点，重启后继续增量读取；只支持 sessions 目录下近 30 天内更新的未压缩 JSONL。缓存文件保存在扩展存储目录，Codex 日志保持只读。

监控台采用蓝色实测线和浅蓝绿色估算线：宽屏为并排趋势区、横向用量数据和会话表格，窄窗口转为紧凑卡片。顶部可选择语言、重新扫描或打开设置，连接信息在下方按需展开。关闭面板后仍按已有监控周期采样，未增加额外扫描计时器；没有费用或网络流量统计。

安装后怎么用

1. 在扩展市场搜索 VSCode Codex Speed，或使用 @id:NO00.vscode-codex-speed 安装。
2. 使用 Remote SSH 时，确认它安装在 SSH: 服务器名 一侧。
3. 在同一台主机上，以同一用户运行 Codex。状态栏会自动出现速度读数。

默认读取 CODEX_HOME，未设置时使用 ~/.codex。新会话通常在约五秒内被发现，之后默认每半秒刷新。

如果使用独立的 Codex 数据目录，请在 VS Code 的用户或远程设置中填写包含 sessions 子目录的路径：

{
  "codexSpeed.codexHome": "/home/your-user/.codex"
}

数字代表什么

~ 表示包含估算，并非逐 token 遥测。 完成记录按响应耗时均匀重建历史，因此校正后的实线也可能呈水平。工具中仍有刚完成响应的三秒窗口贡献可以正常出现；窗口排空后归零。速度含首 token 等待，输出含 reasoning。

默认统计此主机的全部近期会话，所以当前聊天停下来时，其他会话仍可能贡献速度。如果缺少结束事件，生成估算可能持续到最近一条日志后的五分钟。

常用设置

看不到速度时

先悬停状态栏查看提示，或在命令面板执行 Codex Speed: 查看诊断日志。确认扩展和 Codex 位于同一主机，运行用户有权读取日志；codexHome 应指向 .codex 数据目录，而非 sessions 本身。此版本仅读取未压缩的会话 JSONL。

从旧的 local-codex-tools 版本迁移时，请先卸载旧扩展，避免重复状态栏。已有 codexSpeed.* 设置可继续使用。

隐私与资源占用

扩展只读会话日志，不修改 Codex 设置，不读取登录凭证，不上传对话内容，也不发起模型请求或额外消耗 token。监控会产生文件读取和计算开销。更多设置、资源限制和开发步骤请看完整中文说明。

English

What you get

• Codex output tokens per second in the status bar. Hover for the latest measured call's model, output count, and duration.
• Direct log monitoring on your Remote SSH host, with no manual log synchronization or extra SSH connection setup.
• Combined readings for multiple sessions and subagents, or a view limited to your workspace folders.
• A dashboard with a digital speed readout, 1/2/5-minute charts, session summaries and selectable usage periods.

This is an independent community extension, not an official OpenAI product. The UI follows the VS Code display language by default. English, Simplified Chinese and Traditional Chinese are supported; other languages fall back to English.

Set codexSpeed.language to auto (default), en, zh-cn or zh-tw in User or Remote settings. Manual changes immediately update the status bar, hover details, dashboard and diagnostic messages without reloading. Remote SSH follows the client's display language, not the server's system locale. Command Palette titles and Settings descriptions always follow VS Code's display language.

Dashboard

Hover for a compact summary, or click to open the dashboard. Core Codex speed follows Agent Speed's rules: a three-second window combines measured records and in-flight estimates. Completed usage reconstructs its interval and replaces the earlier estimate. The chart has 300 one-second buckets: blue for measured components, teal dashed totals when estimates contribute. Hover or use arrow keys to inspect each second. A one-second chart value need not equal the three-second main reading.

Usage supports the last 30 minutes, today, the last 7 days and the last 30 days. Today follows the client time zone; 7/30-day periods use rolling hours. Input may include cached tokens; averages use timed completed calls. The five-minute monitored peak can include estimates and is corrected with reconstructed history.

The console uses blue measured lines and teal estimate lines. Wide windows show the reading beside its trend, a horizontal usage strip and a session table; narrow windows use compact cards. Language, rescan and Settings are in the toolbar, with connection metadata in expandable details. Closing the dashboard keeps the existing status-bar sampling active, without another scanning timer. Cost and network traffic tracking are not included.

Get started

1. Install VSCode Codex Speed from the Marketplace, or search for @id:NO00.vscode-codex-speed.
2. In a Remote SSH window, confirm installation on SSH: your-server.
3. Run Codex on that host as the same user. The status item appears automatically.

The extension uses CODEX_HOME, falling back to ~/.codex. New sessions are normally discovered within about five seconds, followed by half-second polling by default.

For a custom data directory, set codexSpeed.codexHome in VS Code User or Remote settings to the directory containing sessions, for example /home/your-user/.codex.

Read the numbers

~ means estimates contribute; it is not per-token telemetry. Completed records distribute output uniformly over their response intervals, so corrected measured lines can be flat. A tools phase can still show a just-completed response's contribution to the three-second window; it returns to zero when the window empties. Duration includes time to first token; output includes reasoning.

The default host-wide scope includes other sessions, even when the chat you are viewing has stopped. Missing completion events can leave generation estimates active until five minutes after the latest log event.

Historical usage is indexed in batches from uncompressed JSONL files under sessions modified within the last 30 days. Initial indexing and incomplete results are clearly marked. Counters and read cursors are cached in extension storage so indexing can resume after a restart; Codex logs remain read-only.

Useful settings

Troubleshooting

Hover over the status item for a diagnostic message, or run Codex Speed: Show Diagnostic Logs from the Command Palette. Check that the extension and Codex run on the same host and that the extension user can read the logs. Point codexHome to the data directory, not the sessions folder. This version reads uncompressed session JSONL only.

If upgrading from an old local-codex-tools build, uninstall that extension first to avoid duplicate status items. Existing codexSpeed.* settings remain compatible.

Privacy and resource use

The extension reads logs without modifying Codex settings, reading login credentials, or uploading conversation text. It makes no model requests and consumes no extra model tokens. File reads and calculations have some resource overhead. See the full English documentation for additional settings, limits, and development instructions.