The trust layer for code written by agents. Two frontier models review every change. Only findings they both agree on get posted. Every closure is pinned to the commit SHA that resolved it. The receipts are the artifact. Live receipts: https://www.antfleet.dev/receipts · Architecture: https://www.antfleet.dev/architecture
import { stackedProvider } from "@antfleet/core/providers/stacked";
import { anthropicProvider } from "@antfleet/core/providers/anthropic";
import { openaiProvider } from "@antfleet/core/providers/openai";
const reviewer = stackedProvider({
providers: [anthropicProvider, openaiProvider],
agreement: "unanimous",
});
const result = await reviewer.review(repoRoot, prompt, null);
// result.findings only contains findings BOTH providers clustered onto.- A two-model agreement primitive for code review (
providers/stacked+providers/agreement). - Schema-strict provider adapters for Anthropic (tool-use) and OpenAI
(
response_format: json_schema), plus an OpenRouter adapter that falls back to in-prompt schemas for models that don't honor strict outputs. - A patch-suggestion lane (Patch Agent v1.5,
providers/patch-gate) that ships a unified-diff fix only when both providers independently propose one. - A planted-bug corpus + spike runner (
examples/dogfood/) used to measure whether the agreement filter separates signal from noise. The baseline receipt lives atexamples/dogfood-results/WEEK1-VERDICT.md.
- A GitHub App. The hosted reviewer that posts inline comments + SHA-pinned receipts is a separate surface — see https://www.antfleet.dev for the live service. This package is the primitive it's built on.
- A linter or static analyzer. The providers ask LLMs to find defects; the agreement gate is what makes the output trustable.
- A drop-in replacement for any single model. Unanimous-of-N degrades to the weakest provider's recall — pick your stack accordingly. The dogfood verdict above documents this trade-off honestly.
Upstream credit: the original CLI scaffolding and feature-slicing primitives
that AntFleet was built on top of come from
openclaw/clawpatch (MIT). The code
in this repo is the differentiated layer — provider adapters, the agreement
primitive, the patch-suggestion gate, and the dogfood measurement harness.
Requirements: Node 20+, pnpm, ANTHROPIC_API_KEY, OPENAI_API_KEY.
git clone https://github.com/AntFleet/antfleet-core.git
cd antfleet-core
pnpm install
pnpm test # 74 unit tests, no API calls
export ANTHROPIC_API_KEY=...
export OPENAI_API_KEY=...
pnpm tsx examples/dogfood/spike.ts # one live review, ~30–60sThe spike calls both providers in parallel against the planted-bug corpus
under examples/dogfood/src/, runs the unanimous agreement filter, and
writes a markdown receipt to examples/dogfood-results/spike-<ts>.md.
The hosted reviewer is already posting receipts on real PRs in public agent repositories. Sample PRs the AntFleet reviewer commented on:
Liquid-Protocol-Ops/agent-autonomopoly- More live receipts: https://www.antfleet.dev/receipts
- Web: https://www.antfleet.dev
- Architecture: https://www.antfleet.dev/architecture
- Receipts: https://www.antfleet.dev/receipts
- X / Twitter: @AntFleetDev
MIT — see LICENSE. Built by AntFleet — see AUTHORS.md.