mise-clean

You are running the hygiene sweep — "clean as you go, sweep the line at close." Same mise ethos as everything else: consent-first, non-destructive, snapshot-first. Report first, then act on a yes, one category at a time. Never blind-delete, never git clean -fdx, never rewrite history to "tidy," and never touch a file that might hold a secret (defer to the secret protocol — do not ingest the value). Focus: $ARGUMENTS if given, otherwise sweep it all.

Survey, then present findings grouped, each with a one-line "why it's cruft" and the exact fix:

1. Junk tracked in git that shouldn't be. Build output (DerivedData/, build/, .next/, dist/, target/), dependencies (node_modules/), OS/editor cruft (.DS_Store, *.xcuserstate). Fix = untrack but keep on disk: git rm --cached and add the pattern to .gitignore in the same step (untracking without ignoring isn't durable — a later git add re-tracks it). Removes bloat from the repo without deleting the user's files.
2. Backup & scratch cruft. mise's own .bak-* files and .mise-backup/, a resolved HANDOFF.md, stray *.log, temp/scratch dirs. These are deletable — list them and delete only on a yes.
3. Git detritus. Local branches already merged into the main line (offer to delete), orphaned worktrees (especially leftover mise-cook brigade lanes — git worktree list, then prune the stale ones), stale stashes (report; never auto-drop). Branch and worktree removals are recoverable from the reflog for a while — say so.
4. Untracked files. Separate real-looking work from junk. Never delete untracked files blindly — they may be the user's uncommitted work. List them, sort into "looks like junk" vs "looks like work," and let the user decide per item or per category.

Optional — the CLAUDE.md audit (opt-in; don't run it by default). The four sweeps above are cheap surveys of files that shouldn't be there. This one is different: the file should be there, but its contents rot — and verifying that means reading the code, which is a real context spend. So it runs only when asked (a focus arg like "audit CLAUDE.md" / "check the brain"), or when the user takes you up on the one-line offer at close. Never bundle it into a routine sweep.

It belongs in the hygiene lens because CLAUDE.md is cruft's worst hiding place: it's written once and then read by every future session — by an agent that believes it. A stale line doesn't just sit there like a .DS_Store; it actively misinforms every run that follows, and every surplus line is a tax on every session's context. Same consent contract as the rest of this sweep — fix what you can prove wrong, propose the rest, never blind-delete.

• Stale — names a path, script, flag, or command that changed or vanished. Verify it (ls the path, run or grep the command; don't reason about whether it's probably still true). Verified stale is a fact, not a judgment: fix it and report it in a line.
• Duplicated — restates the README/ARCHITECTURE/docs, and has since drifted from them. The doc is the source of truth: propose replacing it with a one-line pointer.
• Generic — advice a competent model already has ("write tests", "use types"). Pure context tax: propose the cut.
• Unearned — a rule the code doesn't actually follow. Flag it, never cut it — it may be a live intent rather than dead weight.

Weigh every cut by what it costs to be wrong: a wrong trim silently removes a guardrail and nobody notices until it bites; a wrong keep costs some tokens. So the default is keep and flag, and "I don't see why this is here" is a reason to ask, not to delete. Watch the one real failure mode: stripping the odd, hyper-specific gotchas — the highest value-per-token lines in the file — because they look like one-offs out of context. A weird rule is evidence of a scar. Group findings into a few real decisions with the diff and the reason, never a line-by-line checkbox cart. A tight, accurate brain earns a one-sentence "it's current" — not a manufactured trim.