diffwarden

A small CLI for agent-callable code review.

Diffwarden gives any coding agent one stable review command: it fans a diff out to multiple reviewer engines, cross-checks their findings, and returns a single machine-readable JSON artifact — without holding any API keys of its own.

A real session: two reviewers flag a planted P1, the fix lands, and the same command comes back green with exit 0.

Quick Start

Requires Node >=22.19.0. Developed and tested on macOS and Linux; Windows is untested — the CLI contains Windows-specific handling but it has not been validated.

From any Git checkout, run guided setup. init probes installed executables, SDK packages, and auth signals to discover which reviewer engines this host can already run, then writes a user config — all without running a review or spending model budget:

npx --yes diffwarden@latest init

Then run a real review against your configured reviewers:

npx --yes diffwarden@latest review --target base:main

With a permanent install (below), the same command is just diffwarden review --target base:main.

Reviews bill against your existing engine subscriptions or API keys; Diffwarden holds no keys of its own and never bills you directly. Duration and cost vary by engine, model, and diff size.

For a permanent install, use npm:

npm install --global diffwarden
diffwarden --version

To verify a specific reviewer's runtime, auth, model, and effort settings, or to force the discovery scaffold:

diffwarden init --discover   # force the discovery scaffold
diffwarden doctor --reviewer-set 1
diffwarden doctor --reviewer claude
diffwarden review --target base:main --reviewer claude

Agents should opt into direct text output with --agent:

diffwarden review --target base:main --agent

Development and CI Smoke Tests

The built-in fake reviewer runs a credential-free review, useful for smoke-testing the pipeline in CI or local development without any engine installed:

diffwarden review --target uncommitted --reviewer fake
diffwarden review --target uncommitted --reviewer fake --agent

For local development from a source checkout:

git clone https://github.com/aurokin/diffwarden.git
cd diffwarden
pnpm install
pnpm build
pnpm dev -- review --target uncommitted --reviewer fake

Why Not Your Agent's Built-In Review?

Built-in review commands ask the same model that wrote the code to judge it. Diffwarden provides an independent second opinion across engines: multiple reviewers see the same diff, and their findings are cross-checked and aggregated. The output is a stable machine-readable contract any agent or CI gate can consume, the review path is read-only by design, and Diffwarden never holds credentials — reviewers use the auth you already have. See docs/comparisons.md for the full argument.

Common Commands

diffwarden review --target uncommitted --reviewer fake
diffwarden review --target base:main --reviewer cursor
diffwarden review --target base:main --reviewer claude --model sonnet --effort high
diffwarden review --target base:main --reviewer pi --model anthropic/claude-sonnet-4-5
diffwarden review --target base:main --reviewer droid-cli --model claude-opus-4-7
diffwarden review --target base:main --reviewer-set 2
diffwarden review --target base:main --reviewer cursor --reviewer pi:openrouter-high
diffwarden review --target commit:abc123 --json
diffwarden review --target base:main --reviewer-set 2 --agent
diffwarden review --target base:main --reviewer-set 2 --agent --focus "focus on state management" --focus "focus on localization"
diffwarden review --target base:main --reviewer-set 2 --agent --no-overview --focus "focus on state management"
diffwarden review --target base:main --reviewer-set 2 --report
diffwarden review --target base:main --reviewer-set 2 --fail-on-findings P2
diffwarden review show review.json

Examples above mix flagship engines (claude, codex — fully supported and live-tested) with experimental engines (cursor, pi, droid, and others — functional, best-effort); see Current Status for the tier breakdown.

Verify reviewer runtime, auth, model, and effort settings without reviewing a diff:

diffwarden doctor --reviewer cursor --model composer-2.5
diffwarden doctor --reviewer claude --model sonnet --effort high
diffwarden doctor --reviewer pi --model anthropic/claude-sonnet-4-5

List configured reviewers and reviewer sets without running preflight checks:

diffwarden reviewers list
diffwarden reviewers list --json

Probe the host for usable reviewer engines, then add, edit, or remove reviewers in the user config. Discovery never runs a review or spends model budget; --deep additionally runs adapter preflight:

diffwarden reviewers discover
diffwarden reviewers discover --deep
diffwarden reviewers discover --json
diffwarden reviewers add codex
diffwarden reviewers add claude --transport cli --set 1
diffwarden reviewers edit codex --model gpt-5.1-codex
diffwarden reviewers set add 1 codex
diffwarden reviewers set remove 1 codex
diffwarden reviewers remove codex

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗