autoreview

AutoReview

A rule file is a suggestion. This turns it into a verdict on every commit.

Per-file architecture gates for your coding agent. Write rules in plain English Markdown; a reviewer verifies each file against the rules that match it, on every commit. Matching runs locally; the reviewer runs on a local Ollama or any API model.

---

I built this after getting annoyed at my agent skipping audit logging on a payment mutation for the third time. CLAUDE.md said to emit audit events. Agent read it. Agent skipped it. Tests passed. I only caught it because I happened to diff that specific file.

A rules file is a suggestion. There are no consequences for ignoring it. This is the reviewer that turns it into a rule.

Quickstart

Three steps, in order:

1. Install Ollama (or skip to step 2 if you're using a paid API). Grab it from https://ollama.ai, then ollama serve & in a terminal.
2. Install the plugin in Claude Code: /plugin install autoreview. Nothing happens to your repo yet — plugin install alone is inert.
3. Scaffold AutoReview in your repo: /autoreview:setup (or just ask the agent to "set up autoreview with ollama and install the pre-commit hook"). The skill probes for Ollama, asks which model to use, and runs init.mjs to create .autoreview/, install the git pre-commit hook, and ship one example rule.

Three things to know before your first commit:

• severity: error rules block. A [reject] or [error] (including provider unreachable) on a severity: error rule causes exit 1. Rules default to severity: error. Mark rules severity: warning to warn without blocking.
• Nothing runs until step 3. Only the setup skill drops the hook and scaffolds config. If you see no verdicts on commit, setup hasn't run.
• What leaves your machine. Ollama keeps everything local. Paid providers (Anthropic/OpenAI/Google/openai-compat) receive the full file content plus the matching rule body on each call. Trigger matching runs locally — files that match no rule never leave the box.

To add your first rule, just tell the agent: "add a rule that forbids console.log in production code" — it'll walk you through the 7-step wizard and save the rule at .autoreview/rules/.

After setup, the agent reads templates/agent-rules.md (copied into your repo at install time) and uses that as its operating manual for AutoReview skills. You don't need to memorize anything; just talk to the agent.

The problem

You wrote rules in CLAUDE.md. Your agent applies maybe 70% of them. The rest it "optimizes away" because it decided they're noise. You tell it again, it does better for a while. Next session, same thing.

Tests pass. Lint passes. But the handler skipped audit logging, called a service it shouldn't, used Date.now() in a deterministic module. You find out in a PR with 50 changed files. Or you don't.

How AutoReview is different from a regular Claude Code skill

A normal skill is static text injected into the agent's context. "Here's how this library works, use it when relevant." Passive. Informational. The agent can read it and ignore it, same as CLAUDE.md.

AutoReview is not that. It runs an actual reviewer LLM against the code after it's written. The verdict is concrete. Pass or fail per rule. If a rule rejects, the commit blocks.

The skill surface just wires it in. The enforcement happens in the reviewer loop.

Regular skill	AutoReview
Text in the agent's context	LLM reviewer against the file
Agent decides if it matters	Verdict per rule
No verification	Verified on every commit
"Hint"	Gate

How AutoReview is different from Yggdrasil

AutoReview and Yggdrasil share the same reviewer loop. AutoReview stays per-file with a trigger-matched Markdown rule. Yggdrasil adds a graph of components, flows, and cross-file aspects on top.

	Yggdrasil	AutoReview
Scope	Cross-file, graph-aware	Per-file only
Setup	Map your codebase	Write a Markdown rule
CI	Hash-based incremental verify	Pre-commit hook or validate
Distribution	npm package with CLI	Claude Code plugin with CLI
Deps	Node + a bunch	Node, zero npm deps

Use AutoReview when you want one rule on one file. Reach for Yggdrasil when rules need to reason across files.

What a rule looks like