Review business logic, not diffs.

Coding agents produce more change than any human can read — review is the bottleneck now. codepremise has your agents map every piece of business logic to the code that implements it. You review what actually changed in behavior instead of every single line, and your agents stop causing side effects and regressions, because every line is mapped to why..

github.com/acme/shop/pull/42

See live demo →

Start now

Install the codepremise kit in your repo

$curl -fsSL https://codepremise.dev/install.sh | sh

Existing project? Tell your agent to build the initial map — it mines your past agent conversations and git history, then asks you only what it couldn’t find:

›Bootstrap the codepremise map: mine our past agent conversations and git history, then interview me to fill the rest.

Open any pull request by swapping the domain: github.com → codepremise.dev

One map, two readers

Plain-language claims on one side, your code on the other — every claim pins the exact lines that implement it, verified by hash. Humans read the left; agents follow the lines.

.codepremise/ — what & why

flow

What happens at checkout

a flow’s steps are edges to other claims — its two steps are the claims nested below

rule

Carts are validated before anything else

why: pricing invalid carts corrupted orders (incident CART-77)

validate.ts:5–19

rule✓ test-backed

Returning customers get 10% off

why: applied before tax — after-tax broke reporting (bug #482)

discount.ts:12–24 · discount.test.ts

corner caserefines ↑unverified

…but never on top of a coupon

why: discounts must not stack (finance audit)

discount.ts:26–31

rule⚠ shared code

Orders over €500 need a fraud check

why: chargeback spike on high-value orders (incident FRAUD-112)

discount.ts:18–24 — the same lines the discount rule claims

your code

src/checkout/validate.ts

5export function validateCart(cart) {6  return RULES.every(r => r(cart));

src/checkout/discount.ts 12–24 18–24 shared 26–31

12if (!customer.hasPreviousOrders)13  return total;18const risk = riskScore(order);19if (risk > 0.8 && total > 500) {20  flagForFraudReview(order);24}26// discounts never stack (June audit)27if (order.couponCode) return total;30return total * 0.9;

tests/discount.test.ts

40it("returning customer gets 10% off before tax")

agents write

The map builds itself

A skill for Claude Code, Codex and Cursor makes your agents keep a .codepremise/ folder of claims while they work — amended when behavior changes, re-affirmed on pure refactors. Hooks catch what they forget.

you review

Claims, not hunks

A PR reads as what changed in behavior: new rules, amended rules, removed rules — and the existing behavior your change puts at risk. Approve when every claim is checked.

agents read back

Bugs stay fixed

Before an agent edits code, it asks the map what those lines can affect. The why is explicit, avoiding side-effects and regressions.

Works with the agents you already use

One command installs the skill, hooks and validator for Claude Code, Codex and Cursor. The map is an open format in your repo — no lock-in, readable by any tool.

$curl -fsSL https://codepremise.dev/install.sh | sh

What exactly gets installed →