codex-sweep

Codex Sweep

Scans the process table for stale OpenAI Codex plugin processes and kills them, then removes their dead cxc-* temp dirs.

Why they accumulate: the codex plugin spawns one detached broker (app-server-broker.mjs + a codex app-server child) per workspace, and cleanup relies on Claude's SessionEnd hook — which never fires on killed terminals, crashes, or worktree-teammate sessions. Orphans pile up (~3 processes each); on WSL2 they exhaust fs.inotify.max_user_instances=128 and Vite starts failing with EMFILE.

Safe by design: the plugin respawns a broker on demand (ensureBrokerSession self-heal), so a false-positive kill costs at most a ~2s respawn. Brokers belonging to live Claude sessions are detected and kept.

Run

Run directly — no confirmation step needed (unlike /dev-clean-wsl's deletions, kills here cost nothing to redo):

node $HOME/.claude/scripts/codex-sweep.js

Pass through user args:

• --dry-run — when the user asks "what would it kill" (lists each broker: pid, age, tier, workspace)
• --age <hours> — tier-2 idle threshold (default 2h; a SessionStart hook also runs this automatically with --auto = 6h threshold, throttled to once per 6h)
• --json — machine-readable

What it kills:

• Tier 1 (always): brokers whose --cwd workspace was deleted; codex app-server orphaned to PPID 1; codex exec orphaned to PPID 1 for >60 min
• Tier 2: brokers older than the age threshold with no live claude session at or above their workspace

Report

Relay the script's summary line to the user. On --dry-run, show the per-broker table.

If the user hit EMFILE / inotify exhaustion on WSL2

After sweeping, verify the pressure dropped:

find /proc/*/fd -lname 'anon_inode:inotify' 2>/dev/null | wc -l

If it recurs, raise the ceiling (see the inotify note in /dev-clean-wsl):

echo 'fs.inotify.max_user_instances=1024' | sudo tee /etc/sysctl.d/60-inotify.conf
sudo sysctl --system

Troubleshooting

• "kept N brokers" is normal — those belong to live Claude sessions or are under the age threshold.
• "tier 2 skipped: liveness lookup failed" — lsof (Mac) or /proc (Linux) lookup failed; tier 1 still ran. Rerun manually.
• Some cxc-* dirs survive one cycle after a sweep (10-min mtime floor guards freshly-spawned sessions); the next sweep removes them.
• A workspace on an unmounted volume looks deleted → its broker gets killed. Harmless: respawns on next use.