Extension Git @piex-dev/review

review — First-Class Code Review for Agents

TL;DR

/review turns "paste a diff for AI to look at" into a repeatable capability: clear scope, noise-controlled, one shared engine for human and model.

Overview

"Have AI look at this code" stumbles on: where the diff comes from, too much noise, unstable prompts, split human/model entry points. @piex-dev/review consolidates it: humans use /review, the model uses the review tool, the core is one git-diff collection + noise filter + structured prompt. A lightweight omp review: keeps the diff engine and modes, drops multi-agent parallel review and TUI overlay.

How it works

Review = pick scope + clean noise + feed structured context. Core insight: review quality depends on whether the diff fed to the model is clean — half the tokens being package-lock.json is waste. 5 modes map to real dev actions: uncommitted / staged / vs base branch (with fetch) / commit / custom.

flow

pick mode → git diff/show → parse unified diff
→ drop noise (lock/min/build/vendor) → build Markdown prompt
→ hand to current session model for findings

Human and model share the engine: /review uses a UI menu, the review tool uses params; both end up in parseDiff + buildReviewPrompt.

Usage

Install

bash

pi install npm:@piex-dev/review

Source: extensions/review

Prerequisites

The current directory is a git repo; git is on the machine.

Usage

Human: /review lists the mode menu. Model: call the review tool with action (diff/staged/commit/branch). After selection a prompt is generated and sent back via pi.sendUserMessage(..., { deliverAs: "followUp" }) for the current model to act as reviewer.

Verify

bash

pi -e ./extensions/review/src/review.ts -p "what is 1+1" --no-session

Implementation

~350 lines, single file. Noise filter (EXCLUDED_PATTERNS) auto-excludes lock/build/vendor/generated/media-binary; excluded files still appear in an "Excluded" section (path + reason + line count) so the model knows things were filtered without ingesting them. Diff parsing chunks by diff --git counting +/-. Prompt generation: skips embedding full diff when too large (>50k chars or >20 files, suggests read); unified instructions grade findings critical/warning/info with file+line.

omp	piex review
multi-agent parallel review	current agent reviews directly
TUI overlay	text output
Diff + noise filter	✅
Multiple modes	✅ (light set)
Structured prompt	✅ streamlined

Design notes

Project	Mechanism	piex choice
oh-my-pi review	command + tool; multi-agent parallel + TUI overlay; diff engine + noise filter + structured prompt	Adopted: diff engine, noise filter, multi-mode, shared prompt. Not adopted: multi-agent parallel (needs subagent), TUI overlay
OpenCode /review	built-in command, review uncommitted or specified commit/branch	Borrowed: interactive menu, unified prompt template, multi-source review

Core trade-off: keep the diff engine and noise filter intact (feed the model clean context), drop parallelism and visual overlay.

Changelog

Roadmap

Direction	Plan
Single model, single thread	near-term optional multi-pass (critical then style); long-term migrate to omp multi-agent when subagent is ready
Weak findings↔code binding	output findings JSON; link diagnostics, route runtime suspects to dap
Naive base-branch & noise rules	project-level `.reviewignore`/pathspec; configurable remote & default branch
Result is chat text	attach a machine-readable block beyond Markdown; serve local workflow first, then gating
Extreme diffs	parser covers real PR edge cases (rename/binary/submodule)

Version history

Version	Date	Changes
0.1.1	2026-07-19	Initial release: diff engine (`parseDiff`) + noise filter (`EXCLUDED_PATTERNS`); 5 modes; `buildReviewPrompt` structured prompt (oversized diff not inlined); shared engine for human/model; omp lightweight (no multi-agent parallel or TUI overlay)

Source Markdown: docs/packages/review.md