Guards

Rough idea in, precise prompt out: /polish or a trailing ?? rewrites your prompt into the input box for you to edit and send.

Blocks catastrophic shell commands (rm -rf ~, curl | sh, force-push to main, DROP DATABASE, format C:) and asks before risky ones (git reset --hard, rm -rf, terraform apply), in Bash and PowerShell.

Keeps work off protected branches: no commit, merge, rebase, cherry-pick, am or push on main/master/release/* (or an explicit push to one), with an optional auto-branch for commits.

Stops accidental truncation of real code: no "// ... rest of the code" placeholders over real code, and a question before a Write or Edit throws away most of a file.

Supply-chain check before installs: npm/pnpm/yarn/bun/npx, pip/uv/poetry/pipx, cargo and gem packages are looked up first; nonexistent names are refused, brand-new, unpopular, deprecated or lookalike ones need your yes.

Protected files and a project boundary: Claude can't edit .env, keys or .git, asks before touching lockfiles and migrations, and asks before writing outside the project.

Notices the commands you keep approving by hand and offers a one-click standing allow rule (never for risky or interpreter-wide commands).

"Look, don't touch": /readonly on blocks edits and every command that is not provably read-only, and tells Claude to investigate and explain instead.

Keeps credentials out of your code, your shell commands and Claude's context: blocks hard-coded secrets in edits, asks before commands with literal keys, redacts secrets from tool results and flags them in prompts.

Opt-in: questions about the codebase (where is, what does, how does, which file, find) go to a read-only haiku scout subagent first; its answer comes with Ask Claude anyway and Good enough, and the tokens saved.

During long turns, haiku writes one line above the prompt every ~30 s on what Claude is doing and what comes next, from a compact log of its tool calls; a one-line recap toast after turns over two minutes.

Opt-in: after a turn that adds or changes functions, sonnet (or haiku) drafts focused unit tests in the background, in the repo's framework and style; Review, Add & run (only that file), or Dismiss; never writes without a press; daily budget cap.

Plain English under every permission prompt: what the command does and how risky it is.

One compact status line: model, context fill, cache hit and warm countdown, session cost, the fullest rate limit, and git branch with changed files.

Every subagent, live: a /subagents pane with status, model, elapsed time, steps, tokens, last tool and answer, plus a band while any run.

Tool latency: times every tool call, shows a live timer for slow ones, and a /stopwatch pane with per-tool averages, p95, max and the slowest calls.

A receipt for every turn: tools, tokens, cache hits, estimated cost and lines changed, on the turn's closing line (or in the transcript), with a /receipts pane.

A /dashboard pane charting this session's turns (prompt tokens by cache kind, hit rate, cost) and your last 14 days of spend.

Forecasts how many turns remain until auto-compaction from smoothed per-turn context growth (tool output vs conversation), and at natural breaks offers Compact now with a focus Haiku drafts.

When Claude re-reads a file that changed since it last read it, it gets only the changed hunks (current line numbers, a few context lines) instead of the whole file.

Huge Bash/PowerShell outputs without the context bill: Claude gets the head and tail, the full text goes to a file it can open.

Stops Claude from re-reading a file it already read in full and that has not changed since; counts the tokens saved.