claude-code - 💡(How to fix) Fix SessionEnd hook event — fires once on true session exit, supports agent/prompt hook types [1 comments, 2 participants]

Problem

The current Stop event fires after every response turn, making it unsuitable for end-of-session work. There is no hook event that fires once when the user genuinely exits or closes a Claude Code session. The agent and prompt hook types — which can invoke Claude reasoning — are restricted to tool events (PreToolUse, PostToolUse, PermissionRequest), leaving session-lifecycle events limited to dumb shell commands.

Requested Changes

1. Add a SessionEnd event (or rename/repurpose a variant of Stop) that fires exactly once when the user closes the session, not after every turn. Distinguish it from the existing intra-session Stop event.
2. Allow agent and prompt hook types on SessionEnd, so an intelligent summary/checkpoint can be written automatically without user intervention.

Concrete Use Case

A user wants an automatic session checkpoint — Claude writes a "where we left off" summary to a notes file on exit. This requires reasoning, not just a shell script. Currently impossible without a manual slash command.

Suggested Companion Feature — GDG-Style Output Versioning

For any hook or slash command that writes a file, consider supporting a generation-versioned filename pattern (inspired by IBM OS/360 Generation Data Groups) as an option — e.g., NOTES.G0001V00.md, NOTES.G0002V00.md — so writes never overwrite prior state. Users concerned about destructive overwrites could opt into this naming scheme via a hook output field or slash command convention. This gives a recoverable audit trail without requiring git for every personal notes file.

Workaround Today

Manual /checkpoint slash command in ~/.claude/commands/checkpoint.md. Works, but requires user discipline at session close.

Suggested by KG5ZSH — filed via Claude Code