cleanup-terminals-browsers

Run the agent process tracking, audit, and cleanup script, including the portable Agent Broom package. Use before starting another localhost/dev server, while keeping track of what is running, before ending a turn that spawned long-runners, when CPU/memory is high, disk is low, or when the user says "Agent Broom", "agent-broom", "what is running", "track processes", "audit localhost", "high CPU", "clean up terminals/processes", "kill stray processes", "orphaned MCP", "devclean", "deep cleanup", "optimize crash reporters", "too many browsers", "localhosts out of control", "clear turbo/build cache", or "free disk space".

Category
Productivity
Skill
220 of 261
Source
Kevin wiki

This skill is intentionally a thin router. The durable behavior lives in scripts. The skill is the memory hook: it exists so agents remember when to run the script. Do not turn deterministic cleanup logic into prompt text. Run the script, inspect the dry run, then act only when ownership is clear.

This wiki keeps cleanup-terminals-browsers as the MECE owner for process hygiene. The complete portable Agent Broom package is vendored under references/agent-broom/ at commit 1bab484c38a3061826cbb93cec591c5d786a48db for version comparison and cross-harness installs.

Script entrypoint

Prefer the repo-root wrapper when it exists:

scripts/agent-hygiene.sh <command>

Fallback to the skill-local script:

cleanup_skill_dir="${CLAUDE_SKILL_DIR:-${CODEX_SKILL_DIR:-skills/productivity/cleanup-terminals-browsers}}"
[ -d "$cleanup_skill_dir" ] || cleanup_skill_dir=".agents/skills/cleanup-terminals-browsers"
bash "$cleanup_skill_dir/scripts/agent-hygiene.sh" <command>

Commands:

scripts/agent-hygiene.sh list
scripts/agent-hygiene.sh audit
scripts/agent-hygiene.sh add --pid <PID> --kind dev --port <PORT> --purpose "<why>" -- <command>
scripts/agent-hygiene.sh stop
scripts/agent-hygiene.sh stop --kill
scripts/agent-hygiene.sh prune
scripts/agent-hygiene.sh artifacts
scripts/agent-hygiene.sh devclean
scripts/agent-hygiene.sh devclean --deep
scripts/agent-hygiene.sh devclean --optimize
scripts/agent-hygiene.sh devclean --disk
scripts/agent-hygiene.sh doctor
scripts/devclean.sh --deep --optimize --disk

list, audit, stop, artifacts, devclean, and doctor are dry-run/reporting first. Cleanup requires an explicit cleanup command such as stop --kill or artifacts --clean. Dev-machine cleanup requires devclean --apply.

Lifecycle

  1. Before starting a localhost/dev server, run scripts/agent-hygiene.sh audit.
  2. When starting a long-running command, record it immediately with add.
  3. Before ending the turn, run audit again and report what remains running.
  4. Stop only recorded or clearly agent-owned process groups that are no longer needed.
  5. For broader machine leaks, run devclean dry-run first: · no flags: safe orphaned dev/MCP/frontend/mobile processes only. · --deep: heavy daemons such as Gradle, Kotlin LSP, Flutter, FVM, simulators, and Logi agents. · --optimize: IDE crash reporter settings, Crashpad dumps, and known background agents. · --disk: global dev caches plus project build artifacts.
  6. Keep the one dev server the user is actively using.

Never kill Cursor, VS Code, Codex, the user's shell, or shared MCP servers such as playwright-mcp and chrome-devtools-mcp.

Extension rule

If this needs more cleanup domains, add them to the scripts, not to this prompt. The script should stay dry-run-first, report sizes/processes before deleting or killing, and expose a narrow command such as artifacts, docker, simulators, or dev-caches.

The macOS cleanup gist from kibotu is the right shape for broader disk cleanup: a deterministic script with --dry-run, category flags, and a report. Use it as a reference for new script subcommands, not as more prose inside this skill: https://gist.github.com/kibotu/5a20336239bb4745e346ac113969d2cb

ImL1s/devclean is the reference shape for dev-machine leak cleanup: safe orphan detection, explicit deep mode, optimize mode, disk mode, and dry-run preview. This skill's local devclean.sh implements those capabilities with agent-safe defaults: dry-run unless --apply is passed, and no generic Crashpad-handler killing because that catches normal Codex/Cursor/Chrome helpers on macOS. https://github.com/ImL1s/devclean

  • references/agent-broom/ - full Agent Broom repo import and source receipt
  • scripts/agent-hygiene.sh - single entrypoint
  • scripts/devclean.sh - devclean-style machine cleanup
  • scripts/process-ledger.sh - process inventory
  • scripts/audit-processes.sh - localhost/listener/process audit
  • scripts/clean-artifacts.sh - build/cache audit and cleanup
  • config/cursor/rules/process-hygiene.mdc - always-on enforcement rule
  • wiki/concepts/agent-process-hygiene.md - rationale