Open source · MIT licensed

Code review that
learns from your team

5 parallel AI review passes. Every dismissed finding makes the next review sharper. Code quality compounds over time.

$ uv tool install git+https://github.com/fabioesposito/superseded

Requires Python 3.14+ and at least one AI CLI (claude-code, opencode, or codex).

Low High Week 1 Week 2 Week 3 Week 4 Week 5+ Review quality

The review that
stops repeating itself

Most tools treat every review like the first. Superseded learns from every decision and tailors each pass to your code.

No more repetitive findings

Dismiss a finding once and it's gone. Superseded remembers so the same noise never resurfaces.

Feedback memory

Every dismissal sharpens the next review. Your team's judgment compounds into a higher-quality signal.

Specialized passes

Security, correctness, performance, style, and architecture — each with its own prompt, running in parallel.

Every review raises the bar

Dismiss a finding once. Superseded remembers. The next review avoids that noise. Code quality compounds with every PR.

1

Review

5 AI passes analyze your diff in parallel with domain-specific prompts.

2

Dismiss

Mark a finding as noise with one click. Add context if you want.

3

Learn

SQLite memory stores the decision and reasoning, scoped to your repo.

4

Improve

Next review injects dismissed findings as negative context. Noise drops.

Fewer false positives over time Dismissed findings stop resurfacing as memory accumulates

5 specialized passes.
One parallel run.

Each pass gets a targeted prompt and runs concurrently. Findings merge and deduplicate automatically.

SECURITY

Security

Injection, auth, secrets, crypto, supply chain

CORRECTNESS

Correctness

Logic errors, edge cases, error handling

PERFORMANCE

Performance

N+1 queries, memory, caching, complexity

STYLE

Style

Naming, conventions, readability, patterns

ARCHITECTURE

Architecture

Coupling, boundaries, abstractions, tech debt

Every review runs
inside its own vault

Each agent executes inside an ephemeral microVM — a private filesystem, a private kernel, a hard boundary. Choose Docker Sandboxes (sbx) on Linux/KVM, or smolvm (embedded smolmachines SDK) on macOS, Linux, or Windows. Your host is never touched, and nothing leaks between reviews.

host machine
microVM · one per review
AI agentclaude-code · opencode · codex
PR checkoutread-only workspace
⦿ provider keys injected per-review · never on disk inside the VM
01

Per-review microVM

A fresh sandbox spins up for every PR. No shared state, no leftovers.

02

Host untouched

Private filesystem and kernel. The agent cannot reach your machine.

03

Two backends

Docker Sandboxes (sbx, Linux/KVM) or smolvm (smolmachines SDK, macOS/Linux/Windows).

04

Zero leakage

The sandbox is destroyed after the review. Nothing persists.

Built for real workflows

CLI for local dev. GitHub Actions for CI. Server mode for multi-repo. Memory that improves over time.

GitHub integration

Post findings as inline PR review comments. Critical issues request changes. Uses gh CLI.

Pluggable agents

Use Claude Code, OpenCode, or Codex. Choose per-review or configure a default.

GitHub Action

Composite Action POSTs to your review server. No agents or secrets in CI — the server does the work.

Server mode

Self-hosted GitHub App. Multiple repos, webhook-driven, configurable concurrency.

Structured output

JSON for piping, markdown for docs, terminal table for quick scanning.

Static analysis

Auto-detects linters (ruff, mypy, eslint, tsc, bandit, gitleaks, go vet) and injects signals before AI review.

Agent skill

Run superseded skill install once and Claude Code, OpenCode, and Codex invoke superseded directly — no more “I don’t recognize that tool” detours.

Start reviewing in 30 seconds

Install and run your first review.

$ uv tool install git+https://github.com/fabioesposito/superseded
$ superseded review --diff HEAD~1..HEAD