The most important runtime scenario. Shows what happens when the user runs bausteinsicht sync.

Shows the continuous sync loop triggered by file changes.

Shows how an LLM agent uses CLI commands to modify the architecture.

Shows how an existing model is brought into Bausteinsicht (bausteinsicht import <file> --from structurizr|likec4|xmi).

The result is immediately syncable with bausteinsicht sync. The importer never touches draw.io — it only produces the model. The XMI path additionally buffers the whole file in memory and converts Windows-1252 to UTF-8 before parsing (peak RSS ~2× the raw file size, 200 MB read cap enforced) — a materially different cost profile from the other two formats, though the CLI-level flow shown above is the same for all three.

The REPL keeps the model in memory and, on save, chooses how to persist so JSONC comments survive when possible. This is the key data-integrity flow.

Shows how bausteinsicht stale flags forgotten elements using git history.

When run outside a git repository the git lookups degrade gracefully (no date), and detection falls back to the view/relationship membership checks.

The contract requires at least one error/recovery scenario. This shows why a failed or interrupted sync never corrupts the working files.

The non-element add subcommands — add relationship, add view, add specification (with its own add specification element and add specification relationship children, for extending the specification itself rather than the model), and add-from-pattern — cover relationships between existing elements, views, specification entries, and pattern-based bulk generation. bausteinsicht add-from-pattern doesn’t have its own list subcommand; the sibling add pattern list command (a separate group under add, not under add-from-pattern) lists the patterns available to apply. All are pure model edits, same shape as add element (see LLM-Driven Modification above).

Versioned model captures: snapshot save, snapshot list, snapshot diff, snapshot restore, and snapshot delete. snapshot restore is destructive (overwrites the live model), so it’s the one step here worth flagging as such.

diff compares two model states; changelog walks git history calling diff at each relevant commit to build a human-readable architecture changelog.

The three export commands — export, export-diagram, and export-table — read the model (and, for export, the synced draw.io document) and render it to an external format — the output-side counterpart to import. None of them write back to the model or draw.io.

Read-only architecture-quality checks, all following the same shape as Stale Detection (load model, analyze, report) — shown together since the flow is structurally identical across all four.

validate runs the same check other commands run internally before writing (see REPL Save, LLM-Driven Modification above), but as its own entry point — typically in CI or a pre-commit hook, checked before any file is touched.

Cross-cutting setup/maintenance commands, not part of the daily edit-sync loop: multi-model workspaces (workspace merge, workspace validate, workspace list), auto-layout, template scaffolding, and JSON Schema regeneration (schema generate) for IDE support.

Two interactive-group commands outside the REPL itself: metric heatmap overlays on diagrams (overlay apply, overlay remove, overlay list), and ADR (Architecture Decision Record) scaffolding (adr list, adr show).