Guards

Before a risky Bash command runs, says what it will touch. A plain deletion is judged by git: files outside git or tracked and clean get only a short label, a Ukrainian dialog (with the rm line, the number of files and a typed 'Other' decision) is for uncommitted changes, the home folder, more than 20 files and other hard risks. Static analysis plus read-only git, nothing is executed to preview.

Saves every image (screenshot) returned by a tool to a folder the user picks: the first screenshot asks for the path, the answer is remembered. /evidence-dir shows, sets or resets it.

A quick pop-up (toast, plus a macOS system notification) when a long Bash command or a long turn finishes, so you can stop watching it.

Live plan progress bars above the Claude Code prompt: stages, steps, a pixel fill and soft sounds for decision, error and done (qa-mods fork: a wider title, the full title on hover, Ukrainian interface)

Asks before git push and wrangler --remote, showing the folder, the branch, the remote and the commits that will go out. Understands git -C dir and cd dir &&. Fail closed.

/replay opens a pane that steps through the session's file edits (Edit, Write, MultiEdit) one at a time, showing each as a diff with Prev / Next.

Detects Claude repeating the same failing tool call (3+ identical failures, or 4+ with the same error) and tells it to change strategy and tell the user. Never blocks.

Denies Edit/Write/Bash writes to Claude config paths the sandbox cannot write (settings, skills, plugins, hooks, projects, ...) and tells Claude to hand the user a ready Terminal command instead of trying workarounds.

Masks API keys, tokens, private keys and password assignments in tool results before the model sees them.

Checks Claude's claims (tests pass, verified, fixed, no regression, API returns N) against the tool calls that actually ran. Unsupported claims get a visible warning; /evidence lists claims and the evidence behind each. Warns only, never blocks.