5 parallel AI review passes, grounded in your diff and converged across rounds. Every dismissal makes the next review sharper.
uv tool install git+https://github.com/fabioesposito/superseded
Requires Python 3.14+ and one provider API key.
Then set one provider key: superseded review --provider openai (or anthropic, qwen) switches provider:
Most tools treat every review like the first. Superseded learns from every decision and tailors each pass to your code.
Unresolved critical findings carry forward. Fixed ones clear. Duplicate comments never get re-posted.
Every finding quotes the changed code. Pre-existing issues are tagged, not blamed on the PR.
Security, correctness, performance, style, and architecture, each with its own prompt, running in parallel.
Review, verify, and converge. Dismissed noise fades; unresolved blockers stay visible until they are fixed.
5 AI passes plus a verify pass analyze your diff with domain-specific prompts.
Dismiss noise or resolve a fixed thread. Both are tracked, separately.
SQLite stores human decisions. Verifier drops never train the rules.
Next round carries unresolved findings, suppresses repeats, and caps the loop.
Each pass gets a targeted prompt and runs concurrently. A verify pass and deterministic grounding filter the merged output.
Injection, auth, secrets, crypto, supply chain
Logic errors, edge cases, error handling
N+1 queries, memory, caching, complexity
Naming, conventions, readability, patterns
Coupling, boundaries, abstractions, tech debt
CLI for local dev. GitHub Actions for CI. Server mode for multi-repo. Grounded findings and converging rounds everywhere.
Inline comments with PRE-EXISTING, UNVERIFIED, and STATIC tags. Carried blockers stay visible by title and location.
DeepSeek, OpenAI, Anthropic, or Qwen: one API key, one --provider flag. No CLI agents or sandboxes.
Composite Action POSTs to your review server. No agents or secrets in CI. The server does the work.
Self-hosted GitHub App. Multiple repos, webhook-driven, configurable concurrency.
JSON for piping, markdown for docs, terminal table for quick scanning.
Auto-detects your linters (ruff, mypy, eslint, tsc, gitleaks, and more) and suppresses advisory duplicates.
Every finding must quote the changed code. Unverified advisory findings are dropped; pre-existing issues never block.
Unresolved critical findings carry forward. Duplicates are suppressed. max_review_rounds stops endless loops.
feedback --stats shows human dismissals vs verifier drops. superseded eval scores a 20-case corpus.
Install and run your first review.