Code Hygiene

A narrow, calibrated review pass — not a general bug hunt. lint catches syntactic violations, test-coverage catches untested lines, and a normal correctness-focused review catches bugs — none of those structurally catch an abstraction that got duplicated instead of reused, a feature that validates but silently does nothing, or a comment that admits a gap nobody closed. This skill's job is specifically the class of smell that only shows up when you read code against this repo's own architectural mandates (CLAUDE.md) and ask "does this actually do what it claims to," not just "does this compile and pass its own tests."

This skill reports; it does not redesign. A real architectural smell usually needs a design decision, not a mechanical patch — report it clearly enough that a human (or a follow-up planning pass) can make that call. See "Auto-fix policy" below for the one narrow exception.

Scope: patch-aware by default

Default to the same scoping convention as lint/test-coverage: review only the files/packages touched by git diff origin/main...HEAD --name-only (or the currently open PR's diff). This never goes hunting through the other 340+ packages the current patch didn't touch. Unlike lint (which wraps a real atmos fix lint command with an actual --new-from-rev flag), this skill has no underlying CLI command — there's no literal flag to pass, only the phrasing below.

Full-repo mode (explicit only)

Only when a human explicitly asks for a full sweep (e.g. "audit the whole repo", "full code hygiene sweep" — never inferred, never run from fix-all's automated cycle). A whole-repo pass is expensive and belongs in an on-demand invocation, not an hourly loop.

Dedup: skip a re-run against an unchanged diff

Before reviewing, hash git diff origin/main...HEAD (the diff content itself — this skill's dedup key is "has the patch changed," not an external event set) and compare against .claude/state/code-hygiene/<branch-slug>.json's last-recorded hash. The .claude/state/<name>/ <branch-slug>.json file shape is borrowed from .claude/hooks/security-remediate-trigger.sh — the one existing precedent for this pattern in this repo, not an established multi-skill convention — adapted here to hash the diff instead of a set of alert IDs, since what's being deduped against is different (a review cycle re-running on unchanged code, not a repeated external notification). If the hash matches, skip re-reviewing and reuse the cached findings array from that state file — report it as a one-line no-op citing the cached result count, don't burn tokens re-deriving the same answer from an unchanged patch. If the hash differs (or no state file exists yet), do the full review below, then write {hash, findings, ts} back to the state file after reporting.

The checklist

For each touched file/package, check for:

1. Duplicated "shared" abstraction. Near-identical logic (same loop/branch structure, same algorithm) appears two or more times inside a package whose entire purpose is to be the shared implementation — e.g. two separate tree-walk-and-hash functions in a package that only needs one, parameterized by what varies between the call sites.

2. Missing sentinel errors. A new or heavily-touched package returns errors but has zero errors.New/sentinel var declarations backing them — every error site is an ad-hoc fmt.Errorf("... %w", err) string literal instead of a wrapped static sentinel, violating CLAUDE.md's "All errors MUST be wrapped using static errors" mandate.

3. A generic that discards its own type. func f[T A | B](...) whose body immediately does a runtime type-switch/type-assert on T (switch any(x).(type) { case A: ...; case B: ... }) on every call path — the generic buys nothing over just accepting any or, better, unifying A and B into one real type. This is a strong "these two types should be one type" signal, not a legitimate use of generics.

4. Self-aware suppression of a mandated rule. A //nolint, //revive:disable, #nosec, or similar suppression comment that turns off a rule CLAUDE.md explicitly mandates (Options Pattern, cyclomatic complexity, file-length limits, sentinel errors) — especially one whose own comment text admits the violation ("too many params", "TODO: refactor", "temporary").

5. Business logic in the wrong layer. A cmd/*.go file containing loops, external API/network calls (git, GitHub, HTTP), or non-trivial branching beyond flag parsing and dispatch — that logic belongs in pkg/. Also flag any new file added under internal/exec/ — this repo has a standing direction to stop growing that package; new abstractions belong under pkg/.

6. Admitted-but-unshipped gap. A TODO/FIXME/"not yet"/"doesn't support X yet" comment on a code path reachable from a documented, non-experimental command or config field, where the gap itself is not reflected anywhere a user would see it (--help, Docusaurus docs, error message) as "not supported."

7. Fake/stub feature — validates but doesn't work. A config field or CLI flag that passes schema/flag validation but whose only runtime behavior is a hardcoded no-op or an unconditional error — the schema is lying about what's usable today.

8. Implemented but never wired up. An exported function with real test coverage but zero non-test callers anywhere in the repo, especially one whose name implies it backs a user-facing command that doesn't actually expose it.

9. Documentation/code mismatch, either direction. Check every command, flag, or config field this patch adds, changes, or removes against the docs that describe it (Docusaurus pages under website/docs/, docs/prd/*.md, agent-skill reference docs, --help text) for all four failure modes:

• Docs describe a feature that isn't actually implemented — aspirational/promised documentation for something that doesn't exist yet, or no longer exists, in the code.
• Docs describe how something used to work — accurate once, but a later change in this patch (or a recent one) shifted the real behavior and nobody updated the prose describing it.
• Missing documentation — a new, shipped, non-experimental command/flag/config field with no corresponding doc update at all, per CLAUDE.md's own "All new commands/flags/parameters MUST have Docusaurus documentation" and "Update all schemas... when adding config options" mandates.
• Implemented but undocumented — the same failure as above, viewed from the code side: a capability that's fully built and reachable, but a user reading the docs would never learn it exists.

What NOT to flag (avoid false positives)

• A stub that's honestly documented as unavailable (--help text, Docusaurus docs, or an intentionally loud error naming exactly what's missing and why) — the smell is a silent or misleadingly-validated stub, not a disclosed one.
• Legitimate generics whose body does real compile-time-dispatched work, not a runtime type-switch on every path.
• Pre-existing code outside the current patch, in default (patch-aware) mode — that's what an explicit full-repo sweep is for.
• Errors already using this repo's sentinel pattern correctly (errors/errors.go or a package-local errors.go mirroring it).
• TODO/stub markers inside test files, example fixtures, or explicitly experimental (ATMOS_EXPERIMENTAL-gated) commands — those are expected to be incomplete by design.
• A doc gap on code outside the current patch, in default (patch-aware) mode — flag it only if this patch touched the feature or its docs; a pre-existing, unrelated doc gap belongs to an explicit full-repo sweep, not a routine cycle.
• Internal/unexported helpers, and code under internal/, that were never meant to have user-facing docs in the first place — item 9 is about user-facing surface area, not every function needing a comment.

Reporting

Use the ReportFindings tool, most-severe first (empty array if nothing survives). For each finding, cite the exact file/line and which checklist item (1-9 above) it matches, plus a concrete failure scenario (what a user configures/calls, and what silently goes wrong). Mark CONFIRMED when the code unambiguously matches a checklist item, PLAUSIBLE when it's a suspicious pattern that needs a human's architectural judgment call either way — don't force a PLAUSIBLE into CONFIRMED just to simplify the report.

Auto-fix policy

Only one class of finding is safe to fix without a human decision: item 2 (missing sentinel errors), and only when the fix is a pure mechanical conversion — replace a dynamic fmt.Errorf/errors.New string with a declared sentinel + %w wrap, preserving the exact original message text and error chain. Nothing else on the checklist gets auto-fixed. Items 1, 3, 5, 6, 7, and 8 are architectural judgment calls by nature — report them, don't guess at a fix. Item 9 (doc/code mismatch) is never auto-fixed either, even though it can look mechanical: writing accurate docs means following this repo's Docusaurus template and house style, not just filling a gap with the first plausible sentence — report it and let a human (or a dedicated doc pass) write the actual content. Item 4 (nolint suppression) is judgment-adjacent: report it, and only remove the suppression yourself if you also fix the underlying violation in the same pass; never just delete a //nolint comment and leave the violation live.

Related

• fix-all skill — invokes this skill at its own step 8 as part of the hourly/on-demand PR-readiness cycle; an unresolved finding blocks the "ready for final review" state the same way an unfixed lint finding or an invalid-but-unresolved CodeRabbit thread does.
• lint skill — catches mechanical/syntactic violations golangci-lint can express as a rule; this skill catches the semantic/architectural ones that require reading intent, which a linter structurally cannot do.
• code-review (native skill) — general bug/security/performance/simplification review, not scoped to this repo's own CLAUDE.md architectural mandates. Use code-review ultra instead of this skill when you want a deep, adversarially-verified, multi-agent pass across a whole PR.
• .claude/hooks/security-remediate-trigger.sh — the precedent this skill's dedup-by-hash pattern is modeled on, applied inline here instead of as a separate hook.