lh-code-health v1.1A

Immutable. This exact content never changes and is served at /api/v1/blob/b186fe409e3b1498.

---
name: lh-code-health
description: >
  Technical debt judgment: what to improve in this diff, when a shortcut is
  taken on the record, which debt to pay first. Trigger: load when deciding
  whether an improvement belongs in this diff or in a ticket, when tempted by
  a shortcut, or when ranking debt. Clean code itself needs no skill.
license: Apache-2.0
metadata:
  author: jrodrigopuca
  version: "1.1"
  local-agents-id: "sk-09fb64"
---

## Critical Patterns

### 1. Clean code = the next reader understands it without you

That's the whole definition, and the clean-code canon (names, one reason to
change, guard clauses, no dead code) follows from it without needing a list
here. Two diagnostics are worth keeping because they catch design, not style:
if naming something is HARD, the design is usually wrong — the naming
difficulty is the finding, don't paper over it; and a comment that explains
WHAT the next line does is an apology for unclear code, while one that states
a constraint the code can't express (why NOT the obvious approach, an
invariant, the decision's link) is the only kind that earns its place.

### 2. The boy scout rule — scoped to the diff

Leave every file you touch slightly better: a clearer name, an extracted
function, a deleted dead branch. But "better" stays INSIDE the task's blast
radius — improving lines you're already changing is hygiene; rewriting the
module you passed through is scope creep (flag it, per the
[peer contract](../../AGENTS.md)). The test: would the reviewer see the cleanup
as part of this change, or as a second change hiding in the diff? When it is
its own change, it gets its own commit inside the same PR — that is how this
rule and "never refactor and change behavior in one commit" (#6) coexist: same
review, separate diffs.

### 3. Debt is a loan, not a sin — but only when it's taken consciously

There are two kinds of technical debt and only one is legitimate:

| Kind | How it happens | Verdict |
|------|---------------|---------|
| Deliberate | "We ship the simple version now because X; the cost is Y" — said out loud, recorded | Legitimate tool |
| Accidental rot | Shortcuts nobody chose, accumulated silently | Just decay |

When taking deliberate debt: say it in the report, and leave a trace at the
site — `// DEBT: <what's owed, why deferred, what triggers repayment>` (plus a
ticket if the team tracks them). Debt with no record converts to rot the moment
you forget it.

### 4. Prioritize debt by interest rate, not by ugliness

The debt worth paying first is where **change frequency × failure cost** is
highest. Ugly code that nobody touches and never fails charges almost no
interest — leave it alone; refactoring it is aesthetics, not engineering. The
mediocre code in the file the team edits every week is what's actually taxing
you. Check `git log --format=%H -- <path> | wc -l` before campaigning to
refactor something.

### 5. Smell → response table

Not every smell triggers a refactor; each has its threshold. A threshold is a
signal to LOOK, not a trigger to act: four parameters on a constructor that
mirrors a record are fine, and two genuinely independent toggles can be the
clearest API. Crossing one means "check"; the response column is what to do
when the check confirms the smell.

| Smell | Threshold | Response |
|-------|-----------|----------|
| Duplication | Third occurrence (rule of three) | Extract — two instances is often coincidence, not pattern |
| Boolean flag params | Second flag | Split the function or use composition |
| Long param list | 4+ | Group into an object that names the concept |
| Deep nesting | 3+ levels | Guard clauses, early returns, extract |
| God function/module | You can't name it honestly without "and" | Split by responsibility |
| Speculative abstraction | Abstraction with one consumer "for the future" | Inline it — YAGNI applies to abstractions too |

Premature abstraction IS debt: the wrong abstraction costs more than
duplication, because it's harder to see and harder to unwind.

### 6. Refactor with a net, never mid-flight

- Never refactor and change behavior in the same commit — reviewer can't tell
  which diff lines are "same behavior, better shape" and which are the feature.
- Green tests before, green tests after; if the code has no tests, the refactor
  starts by adding the characterization test that pins current behavior.
- Small mechanical steps (rename → extract → move), each one shippable. A
  refactor you can't pause halfway is a rewrite wearing a refactor's name.

## Resources

- Sibling skills: [lh-pairing](../lh-pairing/SKILL.md) (how to raise debt findings in
  review), [lh-react-next](../lh-react-next/SKILL.md) and
  [lh-fullstack-boundaries](../lh-fullstack-boundaries/SKILL.md) (what "good shape"
  means in each layer)
- For debt whose repayment is an expensive-to-reverse decision:
  [architect/lh-tradeoffs](../../../architect/skills/lh-tradeoffs/SKILL.md)