agent-reap

Stars1

Forks1

UpdatedJuly 9, 2026 at 14:55

Clean up runaway/orphaned agent processes: stuck zip/tail/python/node, high CPU, dead Claude/Codex/OMX sessions.

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Source

voidmatcha

voidmatcha/dotfiles

View GitHub Repository View Creator Repositories

Download

Related occupationsSOC

Based on SOC occupation classification

Network and Computer Systems AdministratorsComputer and Mathematical Occupations·SOC 15-1244

SKILL.md

readonly

name	agent-reap
description	Clean up runaway/orphaned agent processes: stuck zip/tail/python/node, high CPU, dead Claude/Codex/OMX sessions.

Agent Reap

Dead agent sessions leave behind heredoc interpreters and shell-snapshot jobs reparented to launchd, spinning at 100% CPU with nobody left to read their output. This skill finds them with explainable rules and kills only what the user explicitly approves.

Workflow

Resolve the checkout explicitly; never assume the caller's cwd:

DOTFILES_DIR="${DOTFILES_DIR:-$HOME/work/dotfiles}"
test -f "$DOTFILES_DIR/scripts/agent_reap.py"

Scan (read-only): python3 "$DOTFILES_DIR/scripts/agent_reap.py".
- Lower the CPU floor when hunting quieter leaks: --threshold 30
- --json for machine-readable output.
Show the candidate table to the user. Each row carries a rule tag explaining why it was flagged:
- ORPHAN-INLINE — old, high-CPU ppid 1 interpreter running inline code (python -c, sh -c, …); both the CPU and --min-age gates must pass.
- AGENT-CHILD — long-running high-CPU descendant of a live agent process or a Claude shell-snapshot job.
- ORPHAN-CPU — ppid 1 CLI tool (non-app, non-system) pegging the CPU.
Never kill without explicit user confirmation. Present the exact PID list (AskUserQuestion works well) — AGENT-CHILD rows belong to a session that is still alive, so name which session loses the work before asking.
Kill only the approved pids: python3 "$DOTFILES_DIR/scripts/agent_reap.py" --kill --pids <approved,comma,separated>
- Add --force only if SIGTERM survivors remain.
Re-run the scan to verify, and report reclaimed CPU.

Safety properties (already enforced by the script)

--kill refuses any pid not in the current candidate set.
Agent main processes (claude/codex/omx/omc) are never candidates.
The script's own session subtree is never a candidate, but sibling agent sessions under the same multiplexer (cmux) are still scanned.

More from this repository

same repository

korean-technical-terminology

voidmatcha/dotfiles

Review and edit Korean technical docs, READMEs, UI copy, and prompts when terminology sounds machine-translated or a canonical English term was unnecessarily translated. Classify each disputed term as canonical English to preserve, established Korean to keep, or an awkward calque to rewrite naturally. Also strips surface AI-tell symbols from Korean prose — em/en dashes and spaced --, arrows (→), curly/smart quotes, ellipsis (…), and middle-dot (·) bullets or noun enumerations. Use for 기술 용어, 번역투, 직역, 영어로 써야 할 표현, 원어 유지, 한글화가 어색함, 용어 통일, 화살표 제거, em대시 없애, 가운뎃점 없애, AI 티 기호, AI slop 기호, Korean technical writing, or terminology review. Do not use as a general spelling or tone checker.

2026-07-121

asset-improver

voidmatcha/dotfiles

Auto-research and improve a repo's skills, hooks, and scripts: external real-world web research + internal audit + Codex cross-model verification, behind a propose then approve then apply gate with freeze-on-success and auto-revert. Use when asked to improve, audit, harden, or research-improve skills/hooks/scripts in the current repo.

2026-07-091

agent-reach

voidmatcha/dotfiles

Use public web, GitHub, social/community, video, RSS, and career search surfaces from the CLI/MCP when local repo evidence is not enough. Routes to public-only research tools while avoiding internal/company URLs and secret-bearing pages.

2026-07-091

agent-usage-audit

voidmatcha/dotfiles

Audit installed Claude/Codex agents, skills, commands, metadata-level reference counts, and source-level token pressure to decide what to prune, keep, or route. This inventory skill does not mine historical correction content or decide the current session's handoff state.

2026-07-091

code-intel-doctor

voidmatcha/dotfiles

Diagnose codegraph/serena MCP setup, Codex config, and per-repo indexes when code intelligence is broken or stale.

2026-07-091

context-check

voidmatcha/dotfiles

Check Claude/Codex/OMX context, cache, headroom, and handover pressure; advise continue, compact, clear, or hand off.

2026-07-091

name	agent-reap
description	Clean up runaway/orphaned agent processes: stuck zip/tail/python/node, high CPU, dead Claude/Codex/OMX sessions.

Agent Reap

Workflow

Resolve the checkout explicitly; never assume the caller's cwd:

DOTFILES_DIR="${DOTFILES_DIR:-$HOME/work/dotfiles}"
test -f "$DOTFILES_DIR/scripts/agent_reap.py"

Scan (read-only): python3 "$DOTFILES_DIR/scripts/agent_reap.py".
- Lower the CPU floor when hunting quieter leaks: --threshold 30
- --json for machine-readable output.
Show the candidate table to the user. Each row carries a rule tag explaining why it was flagged:
- ORPHAN-INLINE — old, high-CPU ppid 1 interpreter running inline code (python -c, sh -c, …); both the CPU and --min-age gates must pass.
- AGENT-CHILD — long-running high-CPU descendant of a live agent process or a Claude shell-snapshot job.
- ORPHAN-CPU — ppid 1 CLI tool (non-app, non-system) pegging the CPU.
Never kill without explicit user confirmation. Present the exact PID list (AskUserQuestion works well) — AGENT-CHILD rows belong to a session that is still alive, so name which session loses the work before asking.
Kill only the approved pids: python3 "$DOTFILES_DIR/scripts/agent_reap.py" --kill --pids <approved,comma,separated>
- Add --force only if SIGTERM survivors remain.
Re-run the scan to verify, and report reclaimed CPU.

Safety properties (already enforced by the script)

--kill refuses any pid not in the current candidate set.
Agent main processes (claude/codex/omx/omc) are never candidates.
The script's own session subtree is never a candidate, but sibling agent sessions under the same multiplexer (cmux) are still scanned.