Domain packs

Domain knowledge lives in packs; the core never imports one. Each pack registers through the agent_fabric.domains entry point (build_registry(discover=True) loads every installed pack) or explicitly (build_registry([stat_fabric.register])).

Pack	Import	Components	Pipelines	Demo agents
statistics	stat_fabric	summary, linear_model, anova, time_series, pca, clustering	exploratory_analysis, experiment_analysis, full_study	examples/research_team
lakehouse	lake_fabric	source_inspect, medallion_plan, airflow_dag_render, dag_check	ingest_to_lakehouse, medallion_design	packages/lakehouse/config
coworker	coworker_fabric	repo_index, context_select, context_pack, code_review, draft_edits, patch_propose	improve_code, propose_patch	packages/coworker/config
text-pack	text_pack	text_stats, keywords, summarize_text	document_digest, document_summary	examples/research_team

• statistics: formulas are built from validated column names; no model-written code runs. Also provides the stats_rules planner (no LLM).
• lakehouse: generates Airflow DAGs that load CSV/JSON/XLSX/API/MCP sources into Trino/Iceberg bronze, silver and gold tables. Generated code is never trusted twice: static template, validated and quoted identifiers, credentials only in Airflow connections, dag_check before use.
• coworker: picks code context within a token budget (measured by labelled evals), reviews it and validates patch proposals. It never writes to disk and only reads allowlisted roots (COWORKER_ALLOWED_ROOTS).
• text-pack: small and dependency-free; the proof that the core is domain-agnostic and a template for new packs.
• notes (examples/notes): not a package, a 100% Markdown domain made of runtime: prompt skills.

Per-package API: agent-fabric · statistics · lakehouse · coworker · text-pack. To write your own pack see Extending.