Skip to content

scaryrawr/anti-slop

v0.2.3MIT

Bundled Anti-Slop checks after edits, plus pre-edit guardrails, without repository lint setup.

GitHub

Anti-Slop Plugin

Anti-Slop gives Copilot durable guidance for writing precise TypeScript and JavaScript without requiring each repository to install a linter, add a build step, or adopt a new dependency.

The skill teaches the reasoning behind Anti-Slop findings and the preferred alternatives. A preToolUse hook checks proposed file edits before they are applied and rejects two high-confidence patterns. After successful edits, a postToolUse hook runs the complete parser-backed Anti-Slop rule set against the resulting file and advises the agent about findings in newly edited code. Neither hook modifies repository configuration.

Prerequisites

  • GitHub Copilot CLI with plugin, hooks, and Agent Skills support.
  • Node.js 20.19+ (or 22.12+) and npm on the machine running Copilot CLI.
  • Registry access the first time the post-edit hook runs, to install pinned Oxlint 1.83.0 into the user's Copilot cache. Subsequent runs are offline.
  • No project dependencies or repository configuration are required.

Installation

copilot plugin marketplace add scaryrawr/scarypilot
copilot plugin install anti-slop@scarypilot

Usage

Once installed, the skill applies while writing, reviewing, or refactoring TypeScript and JavaScript. The hooks run automatically when Copilot proposes apply_patch, create, edit, or str_replace_editor calls on source files. Read-only str_replace_editor views do not run checks; its create, replacement, and line-insertion commands are handled according to their individual arguments. Undo operations cannot be projected beforehand, and post-edit feedback explicitly notes that their changed lines cannot be attributed. The pre-edit hook checks new executable code only, and blocks newly introduced chained assertions (as unknown as) and unknown-only type aliases (type Json = unknown). Multiline forms are included. Comments, strings, JSX text and attributes, generated files, vendor/build output, and files outside the current workspace are excluded. A denial gives the agent a short reason and remediation rather than adding guidance to every tool call.

The post-edit hook uses bundled copies of the 18 generic upstream Anti-Slop rules plus Oxlint's no-accumulating-spread, and enables the five optional Effect rules when the nearest package declares effect. It also flags new generic value is Record<string, unknown> predicates (including renamed is*Record helpers) and is*Record functions with generic object checks. It reads the full file to analyze syntax and context, but reports only diagnostics on the edited lines (and adjacent spacing diagnostics). Output is bounded to four findings plus a count, returned as additionalContext; it does not block or rewrite the successful tool result. A real I/O boundary can legitimately use a guard, so the agent must judge the finding in context. Existing findings do not trigger repeated advice on unrelated edits.

The skill can also be invoked explicitly:

  • "Use anti-slop while implementing this."
  • "Review this change for low-evidence TypeScript patterns."
  • "Unslop this module without adding wrapper guards."
  • "Fix these Anti-Slop findings at their actual boundaries."

The skill directs the agent to:

  1. Preserve precise inferred and schema-derived types.
  2. Parse external data once at its I/O boundary.
  3. Prefer dependency injection over module mocking.
  4. Replace assertion chains and open dictionaries with concrete contracts.
  5. Avoid generic wrappers, any, or trivial guards created only to silence a diagnostic.
  6. Use repository lint tooling when present without onboarding tooling unless the user asks for it.

For repositories without Anti-Slop tooling, the skill also includes a dependency-free, read-only heuristic scanner:

node scripts/scan.mjs [paths...]

The scanner identifies a deliberately small set of high-confidence candidates, supports JSON output, and skips conventional generated and build-output directories. Repeat --ignore DIR to exclude repository-specific output such as lib or packages/api/generated; ambiguous directory names are not ignored by default. It is a review aid, not a replacement for parser-backed linting.

Scope and enforcement

Installing the standalone plugin opts into its hooks across CLI workspaces. Pstack's /deslop remains a separate, general code-cleanup skill and does not install or run Anti-Slop. Disable the standalone plugin with copilot plugin disable anti-slop to stop its hooks and skill. Neither hook intercepts shell-written files or edits from other tools. Copilot cloud agent jobs do not load installed local plugins. Only the two pre-edit patterns above block edits; the complete Oxlint checks are advisory and cannot undo a write. The skill's separate dependency-free scanner covers a smaller heuristic subset for read-only use.

Plugin hook scripts run relative to ${PLUGIN_ROOT} so installation paths need no repository-specific setup. The first post-edit check downloads Oxlint into $COPILOT_HOME/anti-slop/ (or ~/.copilot/anti-slop/), never into the target repository. If it cannot run, the agent receives an explicit failure message rather than a claim that the file is clean. Repositories needing deterministic CI enforcement can separately install or vendor dmmulroy/anti-slop.

Maintaining the bundled rules

After changing the vendored rules in tools/oxlint/anti-slop/, run npm run build:anti-slop and commit the generated dist/ bundles. This builds both the generic and optional Effect rule sets. The runtime installs only the platform-appropriate native Oxlint package; it does not run a repository build or install packages in a user's project.

Validating agent behavior

The hook tests check detection and feedback delivery. The paired behavioral eval in evals/ runs the same scratch task with and without hooks, grades the resulting code, and records whether feedback actually reached the agent. Use multiple repetitions to evaluate recovery, false positive rewrites, and changes in task completion.

Attribution

This guidance is adapted from the rule set and remediation philosophy in dmmulroy/anti-slop, reviewed at commit c44ef22ca116d0ba62a3ff663a0bd13a3f3fa40b. The upstream project is Copyright (c) 2026 Dillon Mulroy and licensed under the MIT License. See THIRD_PARTY_NOTICES.md.

This plugin was imported into this branch from the ScaryPilot source snapshot be2124bd00fd82a9b8274fdc4a1429d6e7034d34; that repository revision is distinct from the upstream Anti-Slop revision above.

Resources