safety

A Claude Code plugin that detects when Claude edits or deletes files outside your request. Clear drift pauses on a question dialog; milder drift is flagged above the prompt with pull-back, fine, and allow-folder keys.

A Claude Code plugin that intercepts Bash tool calls and forces an explicit user dialog before any command that changes the machine (sudo, brew install, global config writes), even in auto mode, with per-project ask/block rules.

A Reddit user shares five small Claude Code hook mods they built after repeated close calls: a destructive-command guard that pauses on git reset --hard or rm -rf and lists uncommitted files at risk, a context-usage weather-forecast indicator above the prompt, an API-key filter for pasted secrets, a push-to-main and Friday-afternoon block, and a task board Claude can read and append to. The author notes hooks fail open — a crashing hook is skipped, so they added a catch that converts any crash into a deny — and admits the guards are pattern-matching only and not bulletproof.

Clickable safety rules for Claude Code: block rm -rf, force-push, mass kubectl deletes, secret file access, sudo and network calls, with presets and jail-to-project modes.

An in-process `tool.call` mod that denies Bash git commands which silently discard uncommitted work. It is the mod form of `hooks/git-guard.sh` and replaces it once cut over, saving the shell forks that script cost on every Bash call (about 34 ms per call, against under a microsecond here).

A Claude Code plugin that blocks destructive shell commands and edits to secrets before they run, using function hooks on Bash, Edit, Write and NotebookEdit tool calls.

The author nearly wiped a Kubernetes namespace after approving an agent’s `oc delete --all` without reading it, then built a suite of 11 Claude Code plugins using the new function hooks API. The main mod, guardrails, blocks destructive commands like mass kubectl/oc deletes, rm -rf, force-push, and secret reads, and tells Claude the user blocked it. Other mods cover activity monitoring, desktop notifications, context tracking, usage limits, prompt coaching, command discovery, and a mod manager. Nine of eleven never call a model, there is no telemetry, and the project is free and MIT licensed.

A Claude Code plugin that shows a red warning band when you are on a protected branch or in a production environment, and blocks irreversible git operations like force-push and reset --hard there.

A Claude Code function-hook plugin that blocks a `git add` from staging your whole `$HOME` work-tree (`dot add -A`, `dot add .`, `dot add *`). Ships with install instructions via `CLAUDE_CODE_PLUGIN_DIRS` or `--plugin-dir`, plus a configurable `dot-guard.wrappers` setting for custom dotfiles aliases.

Dry-runs risky Bash commands (recursive deletes, destructive git, kubectl delete, SQL wipes) and asks Proceed or Cancel first.

Holds a risky shell command before Claude runs it, works out what it would change, and opens a pane with **Proceed** and **Cancel**. Cancel refuses the command and tells Claude why, so it knows nothing ran.

48 open-source Claude Code mods just shipped, terminal and desktop. Subagents as planets, usage limits, a heartbeat per tool call, a crab. The one I'd turn on first: the guard that blocks git push --force.

▶ ♫