guardrails

A Claude Code guard mod that tells the model a denied Bash call ran nothing and must be re-run in full, and refuses git commit -F from shared /tmp.

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.

Lessons that fire at the moment of action: reminders, questions, or blocks attached to matching tool calls.

A Claude Code mod that enforces writing and git rules: rewrites em dashes, and asks before git commit --amend, unformatted pushes, or personal info in notes and commits.

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.

A Claude Code function-hook mod that flags hardcoded user home paths (/home/<user>, /Users/<user>) written into files and nudges $HOME / ~ instead. Machine-local (*.local.*) files are exempt, and the note is non-blocking.

The author shares a set of AI defense guidelines and interception guardrails for Claude Code, targeting real pitfalls such as force push and missed exceptions, with a GitHub project link.

A Claude Code mod that holds subagent fan-out, heavy-context prompts and retry loops before they spend your usage window. It reads your plan's usage, counts agents across every session, and asks in Claude Code's own dialog.

A Claude Code mod for the sprc Slurm cluster: live cluster/job dashboard pane, status line, job lifecycle tools, finish notifications and submit guardrails.

Project Sentinel is an open-source Claude Code mod for long autonomous runs that separately tracks work progress, context usage with automatic handoffs, and prompt-cache health/expiry. It adds autopilot fresh-context handoffs, cache protection, permissions/guardrails, resource limits, activity, profiles, and an optional companion, and can be installed via the plugin marketplace.

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.

A Claude Code plugin hook that warns when git add/stash/checkout/restore/reset may sweep dirty files this session did not write — a toast for the person, a note in the model's context; never blocks. Aimed at shared checkouts with several agents.