local-review

AI-powered local code review for git repos. Supports Claude, Codex, Gemini, OpenCode, and Copilot as review backends.

Catches bugs, security issues, and logic errors before code hits a PR. Runs as a pre-push hook or standalone CLI.

Developer setup (one-time)

git clone git@github.com:84codes/local-review.git ~/.local/share/local-review
~/.local/share/local-review/local-review setup

This symlinks local-review into ~/.local/bin and installs /local-review and /local-review-fix slash commands into ~/.claude/commands/ (Claude Code only). Make sure ~/.local/bin is on your PATH.

To update:

local-review update

Repo setup

From the target repo:

local-review install

Creates config files and a pre-push hook wrapper. Won't overwrite existing files (use --force to replace).

What gets created

File	Purpose
.git/hooks/pre-push	Hook wrapper (gitignored, per-developer)
.review-config	Config: CRITERIA=<path>, MODEL=, and optional REVIEWER=
.claude/review-criteria.md	Review criteria (only if no existing criteria file found)

The installer auto-detects existing criteria files in this order: .claude/review-criteria.md, .github/copilot-instructions.md, AGENTS.md, GEMINI.md, CLAUDE.md, .claude/CLAUDE.md. If one is found, it records the path in .review-config. If none exist, it creates .claude/review-criteria.md with sensible defaults.

The hook wrapper lives in .git/hooks/, which is gitignored by default. Each developer who wants the pre-push hook runs local-review install in the repo. The wrapper includes a guard: if local-review isn't on PATH, the hook exits 0 and the push proceeds normally.

Hook manager integration

The installer detects existing hook managers:

Manager	What happens
husky (.husky/ exists)	Creates/appends .husky/pre-push with guarded local-review hook call
overcommit (.overcommit.yml exists)	Adds PrePush: LocalReview to .overcommit.yml
None	Creates .git/hooks/pre-push (gitignored, per-developer)

Usage

Running without a subcommand defaults to review:

local-review                              # review local changes vs origin/main
local-review 42                           # review PR #42
local-review --reviewer=codex             # force a specific backend
local-review --model=claude-opus-4-6      # override model
local-review --criteria=my-criteria.md    # use custom criteria file
local-review --raw                        # raw markdown output (no pager)

Pre-push hook

Triggers automatically on git push. Asks before running. If issues are found, you can:

1. Push anyway
2. Abort
3. Address interactively (pipes findings into the selected backend for follow-up)

Inside a Claude Code session, the hook emits the diff to stderr so Claude Code picks it up for interactive review.

Review and fix loop

Automatically review, fix, and re-review until clean:

local-review --fix              # review-fix loop (max 5 iterations)
local-review --fix --max=3      # limit iterations
local-review --fix 42           # checkout and fix a PR

Each iteration: review the diff, fix findings, re-review. Stops when "No issues found" or max iterations reached. Requires an agentic backend (claude, codex, gemini, or opencode).

Slash commands (Claude Code / Copilot)

local-review setup installs slash commands into ~/.claude/commands/. These work in Claude Code and GitHub Copilot sessions.

/local-review          # review local changes (full codebase access)
/local-review 42       # review PR #42
/local-review-fix      # review-fix loop (auto-accept, no confirmation)
/local-review-fix 42   # checkout and fix a PR

The in-session commands are better than the CLI path: Claude Code can read source files to verify findings and make surgical edits.

Design: one review voice everywhere

Claude Code ships with an official /code-review plugin (and the pr-review-toolkit). These are great for deep, multi-agent PR reviews with confidence scoring and deduplication. But they solve a different problem.

local-review gives you one review voice across every surface: pre-push hooks, CLI, and in-session slash commands. Same criteria, same quality bar, same behavior. The official plugins have their own opinions baked in (5 parallel agents, their own scoring rubric, CLAUDE.md-focused). You can't tell them "only report things you'd block a PR for" or feed them your repo's review criteria.

They coexist well:

• local-review for the everyday flow: hooks, quick CLI reviews, /local-review in-session
• /code-review for deep-dive reviews when you want the full multi-agent treatment

Backend selection

Supports claude, codex, gemini, opencode, and copilot as backends. Selection order:

1. CLI flag: local-review --reviewer=codex
2. Environment variable: REVIEWER=copilot local-review
3. Config file: Create .review-config in repo root with REVIEWER=claude
4. Auto-detect: Uses the first installed CLI found (claude > codex > gemini > opencode > copilot).

Backends

Backend	CLI	Non-interactive invocation	Can fix (--fix)?
claude	claude	claude -p + model from .review-config	Yes
codex	codex	codex exec -	Yes
gemini	gemini	gemini -p -	Yes
opencode	opencode	opencode -p -q	Yes
copilot	copilot / gh copilot	Pipes to stdin	No (review only)

Configuration

Setting	How to configure
Review backend	--reviewer=claude, REVIEWER=claude in .review-config or env var
Model	--model=claude-opus-4-6, or MODEL= in .review-config (claude, codex, gemini)
Review criteria	--criteria=FILE, or CRITERIA=<path> in .review-config (built-in defaults when absent)
Base branch	REVIEW_BASE_BRANCH=develop env var (default: main)