Smithery is now a part of Arcade.dev! Read more in our announcement here
Scan a whole codebase for incoherences and clean it up safely: debug files, temporary files, orphaned or never-used files, dead or obsolete code, unused dependencies, one-shot scripts that were run...
Cleanup is deletion, and deletion is the one edit you cannot always see the consequences of. So the rule is: detect broadly, verify each candidate, show the evidence, and let the user decide before anything is removed. Git history is the safety net; it only works if the work is committed in small, separate steps.
Check git status is clean (or note what is dirty and leave it alone). Work on a branch. Run the
project's build and tests now and record the result. "Still green" at the end is only meaningful
against a known-green start. If the baseline is red, say so and clean around it.
Collect, do not delete. Categories, with how to find each (commands per ecosystem in
references/detectors.md):
| Category | Signals |
|---|---|
| Debug leftovers | Debug dumps, screenshots, scratch notebooks, *.orig/*.bak, committed logs, stray print/console.log/debugger |
| Temporary files | tmp, scratch, editor swap files, generated output committed by accident |
| Orphaned files | Nothing references the file by name and it is not an entry point |
| Dead / obsolete code | Detector findings: unused exports, unreachable branches, flags that can no longer vary |
| Unused dependencies | Declared but never imported; or imported but undeclared |
| One-shot files | Migration, backfill, fix or import scripts run once. Signals: one commit in history, a name like migrate_/fix_/backfill/one_off, no scheduler, CI, package script or doc references it |
| Duplication | Near-identical blocks or files that change together |
| Tracked but ignored | git ls-files -ci --exclude-standard |
| Untracked or ignored junk | git clean -nd (untracked) and git clean -ndX (ignored) are dry runs. Git cannot restore these files, so deleting them is permanent |
| Stray lines and exports | Leftover print/console.log/debugger, exports nothing imports. These are edits, not file deletions |
A detector is a lead, not a verdict. Before a candidate reaches the user, check what the tool cannot see:
main fields, config-driven loaders).import(variable), reflection, getattr, string-built paths.Fix false positives by configuring the detector (declare the entry point) before using ignore lists. An ignore hides a result; it does not resolve it.
No detector installed, or no network? Do it by hand: git grep -lwF <name> per candidate, scan
imports and requires, and check scheduler, CI and container files. Do not let npx or pip install quietly download a tool; say what you would run and let the user decide.
Treat a dynamic-loading chain as one unit. A plugin loaded by a loader that nothing calls is orphaned together with its loader; deleting half of it leaves a broken reference.
Give each candidate a confidence: high (zero references anywhere, clear temp/debug pattern), medium (zero static references, but dynamic loading is plausible), low (detector only). Only high-confidence items are proposed for deletion without extra discussion; the rest are questions. History signals (single commit, age) are weak in young or small repos where every file has one commit: use them to break ties, not as evidence. References from outside the repo (other projects, deploy hosts) are invisible from here, so when that is plausible, ask.
Present a table: path, category, evidence, confidence, risk. Group by category. Wait for the
answer. Then delete with git rm, one commit per category with a message that names what
went and why, so any single group can be reverted. Never run a detector's auto-fix mode that
removes files.
Two cases need extra care. Untracked or ignored files cannot be restored from git: show the
git clean -nd output and get explicit approval first. Stray lines and unused exports go in
their own table as edits; an export of a package other projects consume stays unless the user
confirms nothing outside depends on it.
Remove unused ones (after step 3's confirmation). Then check freshness:
references/detectors.md).Apply KISS and DRY only to code the cleanup already put in play; unrelated refactors belong in a separate request.
After the cleanup, the workflow changed, so these two files are likely wrong. Show the changes
as a diff and apply them once approved; keep CLAUDE.md in the language it is already written in.
git rm --cached anything already tracked that
should not be. If a secret was ever committed, revoking it is step one; ignoring it is not a fix.Re-run the baseline build and tests and the detectors. Report: what was removed (counts per category), what was kept and why, what was deferred to the user, dependency changes, and the before/after of detector findings. Say plainly if anything is still red.
