npm ci
npm test # node:test, no model calls
npm run test:browser # the dashboard in Chrome, against the dev server
npm run typecheck # CLI and web
npm run lint # ESLint, including the layer boundaries below
npm run format # Prettier (format:check in CI)
npm run build:web # → dist/web, served by agentdiff view and the hosted site
npm run bake-example -- <compare dir> <name> "<title>" "<what changed>" # add a recording to the siteci.yml runs the lint, format check, typecheck, tests and build on Node 22 and 24, and the browser tests in Chrome. The dashboard imports the CLI's own src/compare/signals.ts, so the page and the report can't disagree.
The sms-desk-example and booking-example workflows record the SMS and booking examples against the real models; they are started by hand. This repo also runs the Action on example/ for PRs labelled run-agentdiff.
src/ reads in pipeline order. Each layer imports only the ones below it, and eslint.config.js enforces that:
cli.ts, commands/ |
Parse flags, call the pipeline, print. |
pipeline/ |
runComparison (the seven stages), rebuildReport, runTests — the API the CLI, CI and dashboard workers share. Reports through a Reporter; never prints. |
config/ suite/ agent/ grade/ compare/ explain/ report/ discovery/ github/ |
One folder per step. agent/ is the only code that knows agent frameworks (Mastra, and the adapters for the OpenAI Agents SDK and plain functions). |
storage/ |
The .agentdiff/ Workspace (beside the config) and the files inside each comparison. |
lib/ |
Thin wrappers: model providers, git, canonical JSON, errors. |
dashboard/ |
agentdiff view: server, jobs, project settings, Braintrust. |
The browser-safe modules in compare/ and the */types.ts schemas are also bundled into web/, so they stay free of Node APIs.