The import wizard turns an architecture.review.yaml file into a diagram. That file is produced by pointing a coding agent at a workspace config — it runs the@archatlas/llm-importer pipeline end-to-end and writes the review file deterministically, correlating analysis artifacts with no model call of its own. Run it locally first, then come back here to load the result.
Create import.yaml pointing at your repositories and where the analysis artifacts and output should live:
version: '2.0'
output:
directory: ./architecture-output
repositories:
- path: /path/to/service-a
name: Service A
- path: /path/to/service-b
name: Service BThe importer core never talks to a model. Ask any coding agent (Claude Code, Cursor, Copilot, Codex, Windsurf, and more — against whatever model you've configured it to use, local or hosted) to run the plugins/repo-analysis skill/plugin (an AGENTS.md procedure) against import.yaml. One request runs the whole pipeline itself: it gathers context, analyzes every listed repository (the one step touching a model), and then runs import — deterministic cross-repository correlation, no model call — to write architecture.review.yaml and architecture.arch.json into output.directory. No separate commands to run by hand. See plugins/repo-analysis/README.md for setup.
architecture.review.yaml (and architecture.arch.json) are now in your output.directory. The .review.yaml is the file you'll upload in the next step.
Click “Next” to load the review file and start classifying elements and reviewing connections.