bnema.dev

Session Exit Classification Runbook

This runbook classifies browser sessions using existing P0 marker files in the log directory:

  • session_<id>.startup.marker
  • session_<id>.shutdown.marker
  • session_<id>.abrupt.marker

Marker directory (default):

LOG_DIR="${XDG_STATE_HOME:-$HOME/.local/state}/dumber/logs"

Crash report directory (auto-generated on startup when abrupt exits are detected):

CRASH_DIR="${XDG_STATE_HOME:-$HOME/.local/state}/dumber/logs/crashes"

Classification Model

  • clean_exit (marker-confirmed)
    • shutdown.marker is present.
  • main_process_crash_or_abrupt (marker-confirmed)
    • abrupt.marker is present and no shutdown.marker exists.
  • external_kill_or_oom_inferred (best-effort)
    • startup.marker exists, but shutdown.marker and abrupt.marker are both missing.
    • This is explicitly inferred, not confirmed.

Implementation path in code:

  • internal/bootstrap/session_exit_classification.go
    • ClassifySessionExitFromMarkers(lockDir, sessionID)
    • BuildSessionExitReport(lockDir)

Quick Marker Report (Shell)

LOG_DIR="${XDG_STATE_HOME:-$HOME/.local/state}/dumber/logs"

for startup in "$LOG_DIR"/session_*.startup.marker; do
  [ -e "$startup" ] || continue
  base="$(basename "$startup")"
  sid="${base#session_}"
  sid="${sid%.startup.marker}"
  shutdown="$LOG_DIR/session_${sid}.shutdown.marker"
  abrupt="$LOG_DIR/session_${sid}.abrupt.marker"

  if [ -f "$shutdown" ]; then
    class="clean_exit"
    inf="marker-confirmed"
  elif [ -f "$abrupt" ]; then
    class="main_process_crash_or_abrupt"
    inf="marker-confirmed"
  else
    class="external_kill_or_oom_inferred"
    inf="best-effort"
  fi

  printf "%s\t%s\t%s\n" "$sid" "$class" "$inf"
done | sort

Optional System-Log Correlation

Use this only when external_kill_or_oom_inferred needs more evidence.

  1. Get the startup timestamp for a session:
SID="<session-id>"
LOG_DIR="${XDG_STATE_HOME:-$HOME/.local/state}/dumber/logs"
cat "$LOG_DIR/session_${SID}.startup.marker"
  1. Inspect user journal around that window:
journalctl --user --since "2026-02-07 10:00:00" --until "2026-02-07 10:10:00" | rg -i "dumber|killed|oom|sigkill|signal"
  1. Inspect kernel/system OOM evidence:
journalctl -k --since "2026-02-07 10:00:00" --until "2026-02-07 10:10:00" | rg -i "oom|out of memory|killed process"
  1. Optional coredump check:
coredumpctl list | rg -i "dumber"
  1. Print GitHub-ready issue payload from generated report:
dumber crashes issue latest

Interpretation guidance:

  • OOM/killed lines near the startup window strengthen external_kill_or_oom_inferred.
  • No OOM evidence does not prove a clean exit; keep inference label explicit.