Terminator

Generated for Codex from the canonical source plugin. Run just update at the marketplace root to refresh it.

Run this workflow from Codex. Stop hooks that terminate a Claude Code session when the agent's final message contains a kill phrase. Two scripts: single-kill (ends Claude) and double-kill (ends Claude plus the terminal shell that launched it), each bound to its own phrase. Installs at local (project) or global (user) scope; scripts require LOCAL or GLOBAL arg and exit 1 if missing. Contains-match (robust to model preambles), case-sensitivity configurable, and a kill path that walks the process tree to the owning claude (and, for double-kill, the top-most shell ancestor) and never signals pid <= 2. Install, remove, update, info, and whendone skills via a router; composes safely with existing Stop hooks.

Compatibility scope: This plugin runs as a Codex skill but intentionally operates on Claude Code files, sessions, processes, or service behavior. Those source-domain references are preserved rather than relabeled as unsupported Codex runtime behavior.

Install with Codex

Register the repository marketplace with codex plugin marketplace add ., then run codex plugin add terminator@florianbuetow-codex-plugins or install it from Florian Buetow Codex Plugins in the Plugins Directory. Invoke its workflows with $<skill-name>.

Claude Code Stop hooks that terminate a session — and optionally its terminal — when the agent's final message contains a kill phrase.

Two scripts, you choose by installing

Behaviour is chosen by which script you install, each bound to its own phrase — not by a runtime flag. Install one or both.

Matching

The hook fires when the agent's final message contains the configured phrase (case-sensitivity configurable). Contains-match is used deliberately: models often prepend text like "I'll provide the requested output," which would defeat a whole-message exact match.

Choose a memorable, rare phrase — a movie quote or a sentence unlikely to appear in normal output. The install and update skills warn you if you pick something too common; single common words like done, finished, or stop will cause sessions to terminate at seemingly random moments.

How termination works

• Walks up the process tree to the owning claude process and signals it (SIGTERM, then SIGKILL).
• double-kill.sh additionally walks to the top-most shell ancestor (the terminal's controlling shell — robust to any wrapper/launcher between it and Claude) and signals that too, ~1s after Claude.
• Never signals pid ≤ 2 (init/launchd/kernel) — the one hard safety rule.

macOS note: killing the shell ends the shell session. Whether the Terminal window then closes depends on the terminal app's profile ("When the shell exits" → "Close the window"). A process cannot close a GUI window with a signal.

Usage (skills)

Files it writes (project-local)

.claude/hooks/single-kill.sh   and/or   .claude/hooks/double-kill.sh
.claude/terminator.json        # { single_killphrase, double_killphrase, case_sensitive }
.claude/settings.local.json    # Stop hook entries (merged in, never clobbering others)

Requirements

bash, jq, and standard POSIX tools (ps, kill). macOS and Linux (the macOS path is covered by tests; the Linux path uses the same POSIX ps interface).