Description
SessionEnd hooks that perform any non-trivial async work (e.g., an API call) are killed before they can complete. Claude Code appears to exit the process without waiting for the hook to finish, even when a generous timeout is configured (tested up to 90s).
Reproduction
- Configure a
SessionEnd hook in settings.json:
{
"hooks": {
"SessionEnd": [
{
"matcher": "",
"hooks": [
{
"type": "command",
"command": "/path/to/session-log.sh",
"timeout": 90
}
]
}
]
}
}
- The hook script calls
claude -p --model haiku to summarize the session transcript (takes ~5s):
#!/bin/bash
set -euo pipefail
INPUT=$(cat)
TRANSCRIPT_PATH=$(echo "$INPUT" | jq -r '.transcript_path // empty')
# ... extract condensed transcript with jq ...
SUMMARY=$(echo "$CONDENSED" | claude -p --bare --model haiku "Summarize this session" 2>/dev/null)
# ... write to log file ...
- Exit a Claude Code session normally
Expected: The hook runs to completion within the configured 90s timeout, then Claude Code exits.
Actual: Claude Code exits almost immediately. The claude -p subprocess is killed before it can return. Debug logging confirms the script reaches the claude -p call but never reaches the line after it.
Workaround
Detach the heavy work into a background process so it survives parent exit:
nohup bash -c '
# ... summarization work here ...
' _ "$TRANSCRIPT_PATH" "$SESSION_ID" "$CWD" >/dev/null 2>&1 &
disown
exit 0
This works reliably but means the hook has no way to report errors back to Claude Code, and users need to manage their own timeout (e.g., wrapping the API call in timeout 60).
Suggestion
Either:
- Wait for SessionEnd hooks to complete (up to the configured timeout) before exiting the process, consistent with how other hook events appear to work
- Document the limitation that SessionEnd hooks must not rely on completing before the parent exits, and recommend the
nohup/disown pattern for async work
Environment
- Claude Code: latest (CLI)
- macOS 15 (Darwin 24.6.0)
- Shell: zsh
Description
SessionEndhooks that perform any non-trivial async work (e.g., an API call) are killed before they can complete. Claude Code appears to exit the process without waiting for the hook to finish, even when a generous timeout is configured (tested up to 90s).Reproduction
SessionEndhook insettings.json:{ "hooks": { "SessionEnd": [ { "matcher": "", "hooks": [ { "type": "command", "command": "/path/to/session-log.sh", "timeout": 90 } ] } ] } }claude -p --model haikuto summarize the session transcript (takes ~5s):Expected: The hook runs to completion within the configured 90s timeout, then Claude Code exits.
Actual: Claude Code exits almost immediately. The
claude -psubprocess is killed before it can return. Debug logging confirms the script reaches theclaude -pcall but never reaches the line after it.Workaround
Detach the heavy work into a background process so it survives parent exit:
This works reliably but means the hook has no way to report errors back to Claude Code, and users need to manage their own timeout (e.g., wrapping the API call in
timeout 60).Suggestion
Either:
nohup/disownpattern for async workEnvironment