Open source

AI code review
that learns

Multi-pass review powered by Claude, GPT, or local models. Runs from your terminal or GitHub Actions. Remembers what you dismiss.

Get Started View on GitHub

$ git clone https://github.com/fabioesposito/superseded && cd superseded && uv sync

Terminal

$ superseded review --pr 42 --agent claude-code

Fetching diff...

Running review with claude-code...

Sev Pass File Line Title

---------------------------------------------------------------------------

🔴 critical security src/auth.py 42 SQL injection in query

🟡 suggestion style src/utils.py 15 Unclear variable naming

Total: 2 findings

critical: 1

suggestion: 1

How it works

Five passes, one command

Each pass sends a targeted prompt to an AI agent in parallel. Findings come back as structured JSON — merged, deduplicated, and sorted by severity.

Diff

Fetch the diff

Pull from a GitHub PR via gh or a local git diff range.

Feedback

Load feedback memory

Query the local SQLite store for previously dismissed findings. Inject them as negative context so the agent avoids repeating patterns you've already flagged.

Review

Run 5 passes in parallel

Each pass focuses on one concern with a domain-specific prompt. All five run concurrently via ThreadPool.

security correctness performance style architecture

Merge

Merge and deduplicate

Combine findings across all passes. Sort by severity. Drop duplicates found by multiple passes.

Output

Output results

Terminal table for quick scanning, JSON for piping, or markdown for docs. Optionally post inline review comments to the GitHub PR.

Features

Built for real workflows

CLI for local dev. GitHub Actions for CI. Memory that improves over time.

Multi-pass review

Five specialized passes with focused prompts. Security, correctness, performance, style, and architecture — each with domain-specific instructions.

Feedback memory

SQLite store tracks which findings humans dismiss. Future reviews avoid repeating dismissed patterns. The tool learns from your team.

GitHub integration

Post findings as inline PR review comments. Critical issues request changes. Suggestions post as comments. Uses gh CLI — no API keys to manage.

Pluggable agents

Use Claude Code, OpenCode, or Codex. Each has its own adapter. Choose per-review or configure as default. Bring your own API keys.

Structured output

JSON for piping to other tools. Markdown for docs. Terminal table for quick scanning. Every finding has severity, file, line, and suggestion.

CI-native

Docker-based GitHub Action. Runs on every PR. Configurable passes and agents. Results posted directly to the PR as review comments.

Agents

Use your favorite AI CLI

Superseded delegates to existing AI CLIs. No API key management — use what you already have configured.

Claude Code

claude -p --bare --model

Auth: ANTHROPIC_API_KEY

OpenCode

opencode run

Auth: ANTHROPIC_API_KEY or provider-specific

Codex

codex exec --json --model

Auth: CODEX_API_KEY

Start reviewing

Install, configure, and run your first review in under a minute.