Agent Skills (mblode)

---
name: tidy
description: Reviews a local diff, branch diff, or PR with file:line findings in confirmed and plausible tiers, gates any test the diff adds or changes, and in apply mode lands the fixes and diff-scoped simplifications. Use when asked to "review my changes", "code review", "tidy this", "simplify my diff", "deslop this", "structural review", "security audit", or "audit these tests in my PR". For the PR itself use pr-creator; for CI and review threads use pr-babysitter; for UI defects use ui-design; for repo architecture use codebase-architecture; for pruning the whole suite to a target use test-audit.
---

# Tidy

Review the diff, then fix it when asked. One pass produces the report; the same pass, in apply mode, lands the fixes along with the simplifications a clean diff still hides.

- **IS:** review of a local diff, branch diff, PR, or named security scope, returning severity-tiered findings with `file:line` evidence; in apply mode, the smallest complete fixes for those findings plus diff-scoped simplification.
- **IS NOT:** creating PRs (`pr-creator`), CI failures and review threads (`pr-babysitter`), user-facing UX, accessibility, or rendered quality (`ui-design` Audit mode), library or CLI ergonomics (`dx-audit`), architecture briefs and repo-wide guardrails (`codebase-architecture`), reviewing plans (`planning`), suite-wide test pruning to a measured target (`test-audit`).

## Report or apply

**Report** is the default: "review my changes", "code review", a mode phrase below, or a PR number. Done is the report in the format under Output, carrying the range reviewed, each baseline command's last line, and the references read. The working tree stays as it was. Do not edit a file in report mode, even for a one-character fix; the moment the tree changes the user loses the read-only report they asked for.

**Apply** when the user says "tidy", "simplify", "fix", "apply the findings", or asks for the changes to be made. Done is the confirmed findings and supported simplifications in the working tree, the affected checks passing or their failures shown to be baseline, and a summary of what changed, what was left and why, and the check results. Apply mode edits any file the diff touches or a fix requires, and runs the repository's documented lint, type, and test commands without asking: they run locally against the working tree, and landing the changes is what was asked for. It does not commit, push, open PRs, add or upgrade dependencies, run migrations, or write outside the working tree. A report from earlier in the session is input, not authority: apply what current evidence supports and say why anything was not applied.

## Harness precedence

Claude Code bundles `/code-review` (correctness plus cleanups, with `--fix`) and `/security-review`. A typed slash command runs the bundled skill; this one runs when named or when a mode phrase matches, and adds the Structural and Deslop modes, a whole-repo Security audit, the plausible tier the bundled commands filter out, `REVIEW.md` support, and the same report on any harness with git and the file tools. Findings from a bundled review already run this session enter the verdict step as candidates rather than being re-derived.

## Mode dispatch

Pick one lens from the user's wording. `references/severity-rubric.md` loads in every mode; the table adds the mode's own rubric.

| Mode | Triggers | Adds | Scope |
|------|----------|------|-------|
| **Standard** (default) | `/tidy`, "review my changes", "code review", "tidy this" | nothing | Local or branch diff |
| **Structural** | "structural review", "thermo-nuclear review", "deep code quality audit", "harsh maintainability review", "code judo" | `references/structural-quality-rubric.md` | Local or branch diff |
| **Deslop** | "deslop this", "clean up AI code", "remove slop", "review for AI patterns" | `references/ai-slop-patterns.md` | Local or branch diff |
| **Security audit** | "security audit", "find vulnerabilities", "deepsec", "threat model", "audit for security" | `references/security-checklist.md` | Named subsystem or whole repo, regardless of diff |

Conditional loads, in any mode:

- `references/context-errors.md` when the diff was agent-written, or adds, changes, or removes a module, system boundary, guard, or fallback.
- `references/security-checklist.md` for auth, input handling, external APIs, uploads, dependency or lockfile changes, or environment config.
- `references/performance-checklist.md` for fetching, rendering, images, dependencies, or bundle-affecting imports.
- `references/test-gate.md` when the diff adds or changes a test. A diff with no test changes loads nothing extra here.
- `agents/openai.yaml` only for the optional second-opinion pass in step 3. Otherwise it is launcher metadata for external runners and never loads.

## Workflow

1. **Discover target.** Staged and unstaged changes first (`git diff --stat`, `git diff --staged --stat`); if clean, the branch diff against `git merge-base HEAD origin/<default>`. For a PR, `gh pr diff <n>` with the branch checked out. Record the range or ref pair; it goes in the report.
2. **Gather context.** Open the mode references and the conditional loads with the file tools. A filename in this skill is a pointer, not loaded content; if a required read fails, report the coverage gap rather than claiming the rubric was applied. Capture intent from the user's words, the commit messages, and the PR description. Load scoped `AGENTS.md` / `CLAUDE.md` and a root `REVIEW.md` where one exists: both override this skill's defaults when they conflict, so a pattern they mandate is not a finding. Reuse checks already run on this revision; when a candidate needs execution, run the documented command and preserve its exit status.