Commands

Your 5h/7d subscription limits, a burn-pace forecast and a cold-cache warning, shown above the prompt only when there is something to act on.

Blocks shell commands that wipe your home folder, the project, disks, protected branches or databases, and snapshots uncommitted work to a git ref before commands that would destroy it.

Stop Claude from editing files through Bash (sed -i, heredocs, echo >, tee, python -c) and point it to Edit/Write, so every change shows as a diff and /rewind can undo it.

Retries a turn that died on a transient API error (500, 529 overloaded, timeouts, connection resets) with backoff, and resumes work when the usage limit resets.

Turn CLAUDE.md rules into hard blocks: protected paths, no new report files, banned commands, and rules re-pinned every N prompts and after compaction.

Keep a copy of every Claude Code session that the 30-day cleanup cannot delete, as JSONL and readable markdown, with message times in the transcript and /history to list and search old sessions.

Fingerprint the system prompt, tool descriptions, model and effort every session, and toast when something changed between versions, including changes the changelog leaves out.

Claude cannot say "done" while the tests fail: a Stop gate runs your test command after every turn that changed files, and existing test files are locked so they cannot be patched green.

Compaction without amnesia: the summarizer is told to keep your goal, constraints, decisions and paths verbatim, a card with your goal, prompts, pinned notes, edited files and open todos follows the summary, and the full transcript before compaction is archived as markdown.

Stop "You're absolutely right!": a tone rule in the system prompt, praise openers stripped from what the model reads back, and /tics to count the ones that still get through.