dsh-timesheet

Timesheet for DeepSeek Harness (dsh): turn-based time tracking from session logs.

Token dashboards tell you what you spent; this one tells you how much wall-clock time you spent with your agents — totals, per-day, per-project, per-provider, per-source, plus tool-call counts, failure rates, and time-to-first-token. Zero runtime dependencies; reads *.jsonl session logs only.

Install

dsh plugin add dsh-timesheet

Or run standalone:

npx dsh-timesheet ~/.codex/sessions

CLI

dsh-timesheet <dir> [--json]

• dir must contain *.jsonl session logs (e.g. ~/.codex/sessions, a dsh session directory, or any exported session folder).
• Prints a Markdown report; --json prints the machine-readable dsh-timesheet/v1 envelope.
• Exit codes: 0 report generated, 1 no turns found / warnings, 2 usage/IO error.

npx dsh-timesheet ~/.codex/sessions
npx dsh-timesheet ~/.codex/sessions --json

In-harness usage (agent-callable)

Ask your dsh agent:

给我出一份时间报表：timesheet，目录指向会话日志目录。 Give me a timesheet report: timesheet with dir set to your session-log directory.

The tool returns a dsh-timesheet/v1 report:

{
  "schema": "dsh-timesheet/v1",
  "target": "~/.codex/sessions",
  "ok": true,
  "scannedFiles": 42,
  "sessions": 42,
  "turns": 310,
  "totalDurationMs": 64800000,
  "failedTurns": 7,
  "toolCalls": 1284,
  "byDay": [ { "key": "2026-08-18", "turns": 22, "durationMs": 5400000, "ttftMs": 110000, "failed": 1, "avgMs": 245454 } ],
  "byProject": [ ... ],
  "byProvider": [ ... ],
  "bySource": [ ... ],
  "warnings": []
}

What it reads

Session logs use the standard dsh event envelope:

• session_meta → session id, workspace (cwd), model provider, source (desktop / CLI / VS Code / ...)
• event_msg / task_started → turn begin
• event_msg / task_complete → turn end, duration_ms, time_to_first_token_ms, error
• response_item / function_call → tool-call counts

Files ending in .jsonl are parsed; malformed lines are skipped with a warning. Compressed .jsonl.zstd logs are not read directly — export or decompress them first (the format is plain JSONL after decompression).

Why it exists

• The ecosystem tracks tokens and cost, but no plugin tracked wall-clock time — the WakaTime-shaped hole in the dsh plugin registry.
• Answering “where did my 6 hours go?” is the first step to spending agent time deliberately.
• Zero runtime dependencies, read-only by construction: it never writes to your sessions.

Development

pnpm install
pnpm typecheck
pnpm build
pnpm test
pnpm test:integration

CI runs the dsh-plugin-doctor preflight, unit tests, packed-artifact integration (real timesheet invocation), and a fresh-profile dsh web boot smoke on Windows.

License

MIT © 2026 zoahdev

dsh-timesheet（中文）

DeepSeek Harness（dsh）的时间报表插件：从会话日志做基于 turn 的时间跟踪。

Token 仪表盘告诉你花了多少钱；这个插件告诉你真实花了多少墙钟时间——总计、按天、按项目、按模型供应商、按来源，外加工具调用次数、失败率和首 token 延迟。零运行时依赖，只读 *.jsonl 会话日志。

安装

dsh plugin add dsh-timesheet

独立使用：

npx dsh-timesheet ~/.codex/sessions

CLI

dsh-timesheet <dir> [--json]

• dir 需要包含 *.jsonl 会话日志（如 ~/.codex/sessions、dsh 会话目录或导出的会话目录）。
• 输出 Markdown 报告；--json 输出机器可读的 dsh-timesheet/v1 报告。
• 退出码：0 生成成功，1 没有 turn / 有警告，2 用法/IO 错误。

npx dsh-timesheet ~/.codex/sessions
npx dsh-timesheet ~/.codex/sessions --json

在 harness 内使用（agent 可调用）

对 agent 说：

给我出一份时间报表：timesheet，目录指向会话日志目录。

工具返回 dsh-timesheet/v1 报告（结构见英文版 JSON 示例）。

读取什么

会话日志使用标准 dsh 事件结构：

• session_meta → 会话 id、工作区（cwd）、模型供应商、来源（desktop / CLI / VS Code / ...）
• event_msg / task_started → turn 开始
• event_msg / task_complete → turn 结束、duration_ms、time_to_first_token_ms、error
• response_item / function_call → 工具调用计数

只解析 .jsonl，畸形行跳过并告警。压缩的 .jsonl.zstd 不直接读取——先解压（解压后就是纯 JSONL）。

为什么需要它

• 生态里到处是 token 和成本统计，但没有人统计墙钟时间——这是 dsh 插件注册表里 WakaTime 形状的空洞。
• 回答“我 6 小时去哪了”是开始有意支配 agent 时间的第一步。
• 零运行时依赖、天然只读：永远不会写你的会话。

开发

pnpm install
pnpm typecheck
pnpm build
pnpm test
pnpm test:integration

CI 跑 dsh-plugin-doctor 预检、单元测试、打包集成（真实 timesheet 调用）、Windows 全新 profile 的 dsh web 启动冒烟。

许可证

MIT © 2026 zoahdev