VARIANCE AUTHORITY
BrowseOrient a coding agent in a codebase

Golden paths

Orient a coding agent in a codebase

An agent dropped into an unfamiliar codebase guesses names, re-implements packages that are already installed, imports across the layering the code has, and reviews a diff without knowing which lines any test watches. This path gives it the readings that replace each guess, in the order it needs them: how to ask, where to start, what it may build on, how the code is layered, and what the change did.

Variance Authority answers from your checkout, its installed packages and the record of the last run. This page orders the steps and links the page that owns each one.

1. Give the agent a way to ask

Install the CLI as a devDependency and the agent asks with variance ask; an agent that speaks MCP gets a second package. With no question, ask lists every question it can answer and works before anything has run. Questions about the source answer from the checkout alone; questions about what the tests did need a first run to read.

Choose an agent workflow

2. Find where to start

Orientation turns a question about the codebase into an address the agent can read: a source name, a file, an import neighbourhood or a captured UI subject. It says where to read next, not that the code there is right.

Find where to start

3. Build on what is already installed

Before it writes code against a library, the agent asks which packages the file can already import, which installed one does the job, and the signature at the version in node_modules.

Use the packages your project already installed

4. Respect the layers the imports build

Every package gets a layer number from the longest chain of imports beneath it, derived rather than declared. Before the agent adds an import, it reads which way the layers run; variance layers in a pull request lists which packages changed number.

See the structure your imports build

5. Review the change against what the code did

A diff cannot show whether the changed lines are watched. The execution record can: it names the cases that covered each changed region, so the agent knows what a passing suite actually checked before it forms an opinion.

Review a change against what the code did, and Review a pull request for the whole review in order.