github / github/copilot-cli

session.shutdown event should persist the reason from SessionEndHookInput

未關閉
#2,852 0 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視
area:sessions
主要語言
Shell
星號
11.2k
分支
1.9k
平均合併
14 小時 16 分鐘
30 天內合併 PR
6

描述

### Problem

The runtime internally tracks session end reason via `SessionEndHookInput.reason`, but the `session.shutdown` event persisted to `events.jsonl` only records `shutdownType: "routine" | "error"`. That collapses several distinct non-error causes into the single label `"routine"` and makes post-hoc diagnosis of session lifecycle incidents much harder than it needs to be.

From the installed 1.0.32 type definitions:

```typescript
// sdk/index.d.ts
export declare interface SessionEndHookInput extends BaseHookInput {
reason: "complete" | "error" | "abort" | "timeout" | "user_exit";
finalMessage?: string;
error?: Error;
}

// sdk/index.d.ts
shutdown(shutdownType?: "routine" | "error", errorReason?: string): void;

// copilot-sdk/generated/session-events.d.ts
data: {
shutdownType: "routine" | "error";
errorReason?: string;
// ... usage metrics
}
```

So the runtime appears to have more granular reason information available internally, but the persisted shutdown event does not keep it.

### Concrete impact

In one 1.0.32/Linux session, I observed two `session.shutdown` events with `shutdownType: "routine"` that had materially different provenance:

| Shutdown | Provenance visible from nearby events |
| --- | --- |
| 1st | preceded by an `abort` event with `reason: "user initiated"`, resumed 18 seconds later |
| 2nd | no preceding abort, interrupted active tool execution, bash subprocess received SIGTERM / exit 143 163 ms later, resume happened 2h 44m later |

At the `session.shutdown` event level, those two cases are indistinguishable even though the surrounding event stream shows they are not the same kind of lifecycle transition.

### Context for the second shutdown

Single incident; not claiming a reproducible bug. Parameters at the time of the disruptive shutdown:

| Metric | Value |
| --- | --- |
| Copilot CLI | 1.0.32 |
| OS | Linux |
| Session age | 6h 12m 49.578s |
| Time since last user message | 5h 24m 08.244s |
| events.jsonl size through shutdown line | 45,016,499 bytes / 11,691 lines |
| `task` tool launches before shutdown | 40 |
| Cumulative API duration | 38,889,407 ms (~10.8h) |

I am **not** claiming the shutdown itself was necessarily a bug. It may have been policy, host lifecycle, resource pressure, or something else. The issue here is that the persisted shutdown event does not retain enough reason information to let users or maintainers tell which.

### Request

Please persist the internal session-end reason (or equivalent) into `session.shutdown` events. Even a small additive field would help a lot, for example:

```typescript
data: {
shutdownType: "routine" | "error";
reason?: "complete" | "abort" | "timeout" | "user_exit";
errorReason?: string;
}
```

If the runtime can distinguish other routine-shutdown causes such as external signal receipt or platform/session eviction, a richer optional field would be even better. But even persisting the existing `SessionEndHookInput.reason` values would be a meaningful improvement.

### Why this matters

This would make it much easier to triage session-lifecycle issues without reproducing them live, including cases that currently look similar in raw logs but may have very different causes.

### Related issues

- #2209 — large-session resume failures (different failure mode, same session-lifecycle diagnostic wall)
- #2609 — large-session write-path failures (different failure mode, same session-lifecycle diagnostic wall)
- #2132 — long-session crash under parallel agents (different mechanism: crash/OOM rather than a persisted `session.shutdown`)

### Environment

- Copilot CLI 1.0.32
- Node.js v24.11.1
- Linux x86_64

貢獻指南

開啟貢獻指南

研究方向

Start with the SessionEndHookInput and shutdown declarations in sdk/index.d.ts, then inspect the session event schema in copilot-sdk/generated/session-events.d.ts and trace where session.shutdown is written to events.jsonl. Determine how the existing end reason reaches persistence and update the event shape consistently. Done means shutdown events retain the applicable session-end reason without losing existing fields or behavior.

由索引模型根據 Issue 內容生成。

評估

技術堆疊
typescript
領域
cli, observability
Issue 類型
功能
難度
4/5
預估耗時
3-5 天
活躍度
冷清
描述清晰度
基本清楚
新手友好度
48/100

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。