| name | kanso-refactor |
| description | Use when the user asks to clean up, tighten, simplify, de-bloat, de-slop, or refactor code. Also use when acting on findings from /kanso-audit or /kanso-nuclear. Behaviour-preserving only. |
| argument-hint | [scope: diff|path|audit-report|current-file] |
| disable-model-invocation | true |
| allowed-tools | Bash(git *) Bash(rg *) Bash(find *) |
kanso-refactor
Behaviour-preserving cleanup. Removes dilution, tightens code, deletes cruft. Never changes behaviour. If a change would alter behaviour, the skill surfaces it as a question instead of making the change.
The principles from kanso-principles apply. The anti-dilution taxonomy is the primary target.
Always run inline
This skill runs in the calling chat — including when invoked as the follow-up to /kanso-audit. Never dispatch via the Agent or Task tool, never run as a subagent, never split work into a parallel runner whose output lands in a side window. Edits, summaries, and any clarifying questions all happen in the user's current transcript.
The hard rule
Refactoring must not alter behaviour. Behaviour changes must avoid refactoring.
Mixing the two obscures intent, breaks cherry-picking, and ruins git bisect. If a task requires both, do them in separate commits, each with a single purpose.
Before making any change, ask: would this change produce different outputs for any input the existing tests or usage patterns cover? If yes, it's not a refactor. Stop and ask the user.
Cases where the boundary is tested:
- Removing a defensive try/catch that swallowed errors silently → behaviour change. The caller now sees the error. Ask.
- Inlining a filler variable → refactor. Identical behaviour.
- Replacing a manual loop with
itertools.chain → refactor only if the iterator semantics are identical. Double-check laziness.
- Removing dead code that was unreachable → refactor.
- Removing dead code that was reachable but had no test coverage → behaviour change. Ask.
- Collapsing a premature abstraction → refactor if call sites are updated atomically. Verify.
- Renaming a variable → refactor if scope is local. If the name is exported, it's a behaviour change (API change). Ask.
When in doubt, treat it as a behaviour change and ask.
Resolve the scope
$ARGUMENTS is one of:
diff → the working tree changes (default)- a path like
src/billing/invoice.ts → that file or directory
audit-report → the user will paste or reference a prior /kanso-audit report; act on its Tier 1 and Tier 2 refactor-category findingscurrent-file → whatever file is in focus
If no scope is given and there's no obvious focus, ask.
Calibrate before changing
Get the lay of the land in one turn before reading anything:
- Working tree state:
!git status --short
- What's actually changed:
!git diff --stat
- Recent commits for voice:
!git log -n 10 --oneline
Then read surrounding code. Refactor decisions must fit the repo's existing style, not impose a new one. Check:
- Naming conventions (
get_* vs fetch_*, camelCase vs snake_case)
- Error handling patterns (exceptions vs result types vs error returns)
- Comment density (some codebases genuinely warrant more comments)
- Formatter and linter configs — anything they enforce is not your job
If the repo has a test suite, note how to run it — the verify step below will use it.
The refactor targets
Attack in this order. Stop when the signal-to-cost ratio drops.
High value
- Delete dead code. Unreachable branches, unused imports, unused parameters, unused variables, commented-out code. Git remembers.
- Delete tautological comments. Comments that restate the code. Keep comments that carry business context, constraint reasoning, or non-obvious tradeoffs.
- Delete step-marker comments.
// Step 1:, // Now do X, // Add retry as requested. Always artefacts.
- Inline filler variables.
const result = x(); return result; → return x();. Keep the variable only if it adds a meaningful name or is referenced more than once.
- Collapse defensive theatre. Remove nested try/catch blocks that return None on failure. Surface the failure case to the user as a question: "This swallows errors from
db.query(). Should those propagate, be logged, or be wrapped?"
Medium value
- Rename zero-entropy identifiers.
userDataProcessingResult → something that says what it is. Only if the scope is local; exported names are API changes.
- Replace vanilla reimplementations. Manual
flatten → itertools.chain. Hand-rolled groupBy → lodash.groupBy if already a dependency. Only if the replacement is genuinely equivalent.
- Collapse premature abstractions. One-implementation factory, one-consumer interface, one-subclass base class. Inline them. Verify no external caller relies on the abstraction.
Lower value
- Tighten over-long functions. Extract a helper only if the extraction is used or makes the caller clearer. Don't extract for extraction's sake.
- Align with repo conventions. Normalise mixed
get/fetch/load patterns. But only within the touched scope — don't drift into unrelated files.
Never targets
- Formatting that a formatter handles
- Style preferences not codified anywhere in the repo
- "Modernisation" that a linter could do
- Anything that changes a public API
- Anything that changes a test's pass/fail outcome
Working process
- Identify candidates. List what would change, grouped by target category above.
- Confirm the scope with the user if the list is large. "I've found 23 things to clean up in this file. Want all of them, or just Tier 1?" Don't bulk-edit silently.
- Make changes one logical group at a time. Each group should be commit-able independently (even though this skill doesn't commit).
- Preserve voice. Match the author's existing code style, indentation habits, and comment density within the repo.
- Verify mentally. For each change, walk through: does this produce the same output for the same input? If not, roll back.
- Verify. Run the project's verification command (see below) and capture the result. A refactor that breaks the build isn't a refactor — it's a regression in disguise.
- Report what was done. Summary at the end: what was deleted, what was inlined, what was renamed, with file:line references. Include the verify result. Flag anything deferred as a question for the user.
Verify
A refactor is only behaviour-preserving if you can show it. Run the project's verification command after edits land and paste the exit code.
In Claude Code, kanso's PostToolUse hook already lint-checks each edited file as it lands. In a harness without edit hooks, that safety net did not run — the verification here is the only check, so never skip it there.
Discover the command
In priority order:
AGENTS.md / CLAUDE.md — explicit commands. Use these first.package.json scripts — test, typecheck, lint, check.pyproject.toml, tox.ini, Makefile — pytest, make test, make check, ruff, mypy.go.mod — go test ./..., go vet ./....Cargo.toml — cargo test, cargo check, cargo clippy.
Pick the narrowest command that covers the touched files. Typecheck plus the relevant test file is usually enough; a full suite is fine when the refactor is broad.
The command may prompt for permission on first run — that's expected. Don't reroute around it.
Report the result
On pass:
✓ Verified — <command>
exit 0
<last meaningful lines>
On fail:
✗ Verification failed — <command>
exit <n>
<failing output>
Failure means the refactor altered behaviour. Roll back the edit and surface the surprise to the user — don't iterate silently to make the failure go away.
When no command exists
⚠ No verification command found.
Checked: AGENTS.md, package.json, pyproject.toml, Makefile.
The refactor is applied but unverified — review the diff manually before committing.
Never silently skip. Either verify, or say you didn't and why.
Output format
After making edits, produce a summary like:
## Refactor summary
**Scope:** <what was refactored>
**Files touched:** <list>
### Changes made
- Deleted 14 tautological comments across `src/billing/*`
- Inlined 6 filler variables in `invoice.ts`
- Collapsed `UserProcessorFactory` into `processUser` (only one implementation existed)
- Removed dead import: `./legacy/unused-helper`
### Deferred — needs your decision
- `src/auth/session.ts:42` wraps `db.query()` in a try/catch that returns null on failure. Removing this would change behaviour. Should the error propagate, be logged, or be wrapped in a typed error?
- `src/api/user.ts:89` contains a `fetch_user` function while the rest of the module uses `get_*`. Renaming would be an API change if this is exported. Is it safe to rename?
### Verification
✓ Verified — `<command>`
exit 0
<last meaningful lines of output>
### Next steps
- Review the diff before committing
- Use `/kanso-commit` to stage and message the changes
Framing
- Never use
git commit directly. That's /kanso-commit's job. This skill produces a clean working tree, nothing more.
- Never delete tests to make them pass. If a test breaks, the refactor altered behaviour. Roll back.
- Don't refactor and add features in the same session. If the user asks for both, do the refactor first, commit it, then move on.
- Don't touch files outside the requested scope unless a change is mechanically required (e.g. a callsite update for a rename). Scope creep is a leading indicator of a broken refactor.
Failure modes to avoid
- Silently changing behaviour because the change "felt safe". Ask.
- Aggressive renaming that breaks imports elsewhere. Grep for usages first.
- Deleting comments that looked tautological but carried business context ("Legal requirement: 18+"). Read the comment fully before deleting.
- Refactoring a file the user didn't ask about because it was open in the session.
- Producing a giant changeset that can't be reviewed as a single logical refactor. Split.
