Stop Claude Code repeating the same mistakes with mechanical PreToolUse guards

Their default path: add the marketplace and install the plugin so a PreToolUse Bash hook runs python3 hooks/guard_bash.py. It reads the pending command from stdin and denies when a rule matches. Only self_kill is on in the plugin. The bug it blocks: pkill -f server.js matches the shell that runs it (pattern appears in that shell's own command line), so the tool call dies with exit 143/144 and nothing after it runs; a while pgrep -f … wait never ends. Fix the guard tells Claude: pkill -A -f server.js (-A = ignore ancestors, procps-ng 4.0.3+). Override one command with GUARD_OK=self_kill in it.

Full set: clone the repo, mkdir -p .claude/hooks tools, copy hooks/*.py and tools/*.py, then merge the hooks block from their settings.json into your project's .claude/settings.json. Do not overwrite an existing file. If you already have a Bash PreToolUse entry, put guard_bash.py inside that same entry — with three separate Bash entries their blocking inside one long session was unreliable; merged into one, it blocked every time (observed, cause unknown). Then edit values at the top of each file / via env: GUARD_SAFE_PULL, GUARD_BLOCKED_SITES, GUARD_UPLOAD_ROOT, GUARD_FIRST_COMMIT_LINE, short-save matcher, Stop-hook RULES, and the tables inside tools/*.py.

Their table (condensed): (1) guard_bash.py — self-kill pkill/pgrep -f, optional bare git pull, blocked marketplace headless pages, upload-root scatter, first-commit said: line; (2) guard_short_save.py — save under ~300 chars to an overwrite MCP tool (placeholder body wiped a record twice); (3) stop_check.py — replies that judge eligibility from web search or read bot traffic as demand without the required fact; (4–7) pre-ship scripts for tag limits / banned words / preview video, generation fact gates, duplicate titles, sprite baseline feet. Principle they state: one rule = one mistake that actually happened; fail open on checker bugs; overrides are explicit tokens.

After install: type one command that must be blocked in a real Claude Code session and watch it get denied. For the full copy install, run bash run_all.sh and expect == result: ALL PASS. Each test_*.sh checks both directions — what must block blocks, and ordinary replies like "I set up the users table" still pass. They verified live on Claude Code 2.1.310 / 2.1.316 (2026-10/11): bare git pull denied when configured; Stop hook forced a rewrite of a "traffic grew, so there's demand" reply; on Linux 2.1.316 a plain pkill -f still exited 144 without the plugin, and the plugin denied it before run.