Reference

Glossary

Babbit uses each of these terms for exactly one thing, here and in the product.

Agent: A model working on your project in a harness, with one job: Product, Code Writing, Verify or QA. See Agents.
agents.md: The file of rules each agent follows, one section per agent, kept in your spec and changed only through a Manage change.
Base snapshot: The saved machine image every workspace starts from, with your repositories and their dependencies in place. Rebuilt when main moves.
Box: A small black box in your code, such as a function, an API handler or a table, with a contract, an owner and a record of every crossing. See What's in the box?.
Brief: Your sentence, plus the details it left out: the asks, the tier, what done means, how it will be verified and when to stop.
Build: The side of Babbit for what your product does. Its changes edit the product spec.
Capture: A screenshot or short video the Verify agent records of the app before and after a change.
Change: One request and everything that follows from it: its brief, proposals, pull requests, evidence and events. See Changes.
Change set: A change that touches several repositories: one spec pull request and one code pull request per repository, landing together.
Check run: A pass or fail result Babbit posts on a pull request head: newdev/shape, newdev/evidence or newdev/set.
Connector: A small package the Code Writing agent writes to talk to an outside service. It declares the hosts it calls and the key or sign-in it needs.
Contract: What goes into a box and what comes out, and in what shape.
Drift: A push to main that no change explains. Babbit reads it, re-runs the scenarios it touches and classifies it. See Keeping spec and code in sync.
Egress proxy: A workspace's only way out to the internet. It swaps secret handles for values, one request at a time.
Evidence: One run of one test at one commit, bound to the hash of the scenario line it proves. See Evidence and verification.
Feature: A capability someone can see from outside the product, named with a verb and a noun. Holds at most three goals.
Flaky: A test that failed and passed at the same commit. Flaky is never verified.
Gate: A pause in a change for something only a person can give: a key, a sign-in, an answer or a decision.
Goal: A quality a feature must have, stated as a rule. A feature has at most three.
Handle: An opaque string that stands in for a secret inside a workspace. Only the egress proxy turns it into a value.
Harness: The program an agent runs in, such as Claude Code or Codex.
Jev: TypeSafe's judging model. It decides between candidates in about a tenth of a second and returns a probability.
Landed partial: A change whose spec merged but one of whose code pull requests was refused.
Line: One goal or scenario as written in the spec. A proposal is an edit to one line.
Manage: The side of Babbit for how your product is built. Its changes edit agent rules and settings. Shown in clay.
Map: The file newdev/map.json in each repository. It names the test that proves each scenario and the boxes each scenario uses.
MCP: Model Context Protocol, the standard way agents call tools. The Babbit MCP is the only way an agent touches your product. See Babbit MCP tools.
Outside agent: An MCP client you run yourself, such as Claude Code on your laptop, connected to Babbit.
Outside writer: An outside agent that writes a change itself: it proposes the lines and, after approval, pushes the code.
Owner: The role that holds every permission on a project.
Proposal: One exact edit to the spec, waiting for a person's approval.
Provenance chip: A small label on anything a model decided, saying where it came from and how sure Babbit was.
Qwen: The small model that writes Babbit's words: briefs, draft scenarios and summaries.
Release: One commit of main per repository and every change in it, moving through preprod to prod. See Tiers, releases and hosting.
Replay: Running last week's prod traffic against a release in preprod, with every outside call answered from recordings.
Run: One recorded path through your product: a person's action, a request or a job. Every run leaves one trace.
Scenario: One example of a feature at work: an action and the outcome someone sees. Each maps to a test.
Shape: The standard layout Babbit keeps every codebase in, enforced by the newdev/shape check.
Skill: Babbit's instructions for one kind of task, with an example beside every rule. Runs pin the version they start with.
Smoke scenario: A scenario that runs against a deployed release to prove the tier works before the release counts.
Spec: The tree of features, goals and scenarios that describes what your product does. The source of truth. See Features, goals and scenarios.
Stop rule: An agent's two limits: the first pauses a run and asks you, the second stops it.
Tier: Where code runs: dev, preprod or prod.
Trace: The record of one run across every service it crossed.
Verified: A scenario whose test passed at its pull request head, in every repository that tests it.
Viewer: The role that reads and comments, and changes nothing.
Workspace: An isolated copy of your repositories on a machine Babbit runs for your project, with its own user and database. Agents work only in workspaces.