`/new` discards in-memory usage statistics without writing `session.shutdown` to `events.jsonl`
- 主要语言
- Shell
- 星标
- 11.2k
- 派生
- 1.9k
- 平均合并
- 14 小时 16 分钟
- 30 天内合并 PR
- 6
描述
### Describe the bug
When starting a new session with `/new`, the CLI abandons the current session's in-memory `modelMetrics` without triggering a `session.shutdown` event. All token usage accumulated in the outgoing session is permanently lost — it is never written to `events.jsonl` and never appears in subsequent `/usage` output or `/chronicle` queries.
The only reliable way to persist usage statistics is to exit the CLI entirely with Ctrl+D, which is disruptive when the intent is simply to start a fresh session.
### Affected version
GitHub Copilot CLI 1.0.67.
### Steps to reproduce the behavior
1. Start the Copilot CLI and do some work (several turns, tool calls, etc.)
2. Run `/usage` — note the token totals shown
3. Run `/new` to start a fresh session
4. In the new session, run `/resume` to return to the previous session
5. Run `/usage`
`/usage` shows zero / only the new segment's usage.
The previous session's token counts are gone.
Inspecting `~/.copilot/session-state//events.jsonl` confirms no `session.shutdown` event was written for the outgoing session.
### Expected behavior
`/usage` after step 5 shows the token totals from before `/new` was called (same as step 2).
### Additional context
Running `/usage` before `/new` shows the numbers correctly — they exist in memory at that point. The data is only lost because `/new` does not flush the outgoing session before switching.
The fix appears straightforward: trigger `session.shutdown` (and flush `modelMetrics` to `events.jsonl`) for the outgoing session when `/new` is invoked, before initialising the new one — the same cleanup path that Ctrl+D takes.
**Environment**
• OS: Linux (dev container)
• CLI version: 1.0.67
贡献指南
调研方向
从 `/new` 的 CLI 入口点开始,将其会话切换路径与 Ctrl+D 的关闭清理进行比较。跟踪 `session.shutdown` 和 `modelMetrics` 如何写入 `events.jsonl`;完成标准是:使用 `/new` 重启、恢复之前的会话,并在 `/usage` 和 `/chronicle` 中看到之前的总数。
由索引模型根据 Issue 内容生成。
评估
- 技术栈
- shell
- 领域
- cli
- Issue 类型
- 缺陷
- 难度
- 3/5
- 预计耗时
- 1-2 天
- 活跃度
- 冷清
- 描述清晰度
- 基本清楚
- 新手友好度
- 68/100