DueCare package inventory

Generated during the 2026-05-10 readiness cleanup pass and reconciled on 2026-07-27. The workspace currently contains 18 buildable Python packages under packages/, all sharing the duecare namespace where applicable.

Inventory

Package	Version	Wheel package root	Console scripts	Extras	Purpose
duecare-llm	0.1.0	src/duecare	duecare, forge	all, dev	Meta package that pulls in DueCare components and exposes CLI entry points.
duecare-llm-agents	0.1.0	src/duecare	none	trainer	Agent swarm layer.
duecare-llm-benchmark	0.1.0	src/duecare	none	none	Bundled benchmark test sets, scoring, and aggregation.
duecare-llm-chat	0.17.0	src/duecare	none	none	Gemma 4 workbench, harness UI, sample bundles, knowledge-file I/O, and the reusable portability contract consumed by the Kaggle notebooks.
duecare-llm-cli	0.1.0	src/duecare	duecare	none	Current reliable CLI package for duecare init, duecare demo-stage, and duecare serve.
duecare-llm-core	0.1.0	src/duecare	none	none	Core contracts, schemas, enums, registries, provenance, and observability.
duecare-llm-domains	0.1.0	src/duecare	none	none	Domain pack system with lightweight bundled data; the full 74,640-prompt trafficking corpus lives in repo config/Kaggle assets.
duecare-llm-engine	0.1.0	src/duecare	none	otel	Pipeline engine wrapper.
duecare-llm-evidence-db	0.1.0	src/duecare	none	postgres, all	Evidence database backends.
duecare-llm-kit	0.1.0	src/duecare	duecare-kit-report, duecare-kit-corpus	viz, nlp, all	Reusable indicator engine, visualization helpers, HTML reports, and corpus exports.
duecare-llm-models	0.1.0	src/duecare	none	transformers, unsloth, llama-cpp, ollama, openai, anthropic, google, hf-endpoint, all	Model adapters.
duecare-llm-nl2sql	0.1.0	src/duecare	none	none	Natural-language-to-SQL translator.
duecare-llm-publishing	0.1.0	src/duecare	none	hf-hub, kaggle, all	Publication helpers for HF Hub, Kaggle, reports, and model cards.
duecare-llm-research-tools	0.1.0	src/duecare	none	http	External research tool wrappers with PII filtering.
duecare-llm-server	0.1.2	src/duecare	none	observability, otel	FastAPI server and demo UI surface.
duecare-llm-tasks	0.1.0	src/duecare	none	anonymization, embedding	Capability tests.
duecare-llm-training	0.1.0	src/duecare	none	clustering, unsloth	Synthetic labeling, active learning, dataset assembly, and fine-tune kickoff.
duecare-llm-workflows	0.1.0	src/duecare	none	none	Workflow DAG orchestration.

Current install and registry truth

None of the 18 distributions is currently published on PyPI (verified against the public PyPI JSON API on 2026-07-27). Bare commands such as pip install duecare-llm-kit describe the intended post-release interface; they are not a working registry path today.

For a development checkout, install the complete workspace from source:

uv sync --all-packages
uv run duecare init
uv run duecare demo-stage
uv run duecare serve --port 8080

For a release-like local install, build all 18 wheels and tell pip to prefer that local directory while resolving third-party dependencies normally:

python scripts/build_all_wheels.py --clean
python -m pip install --find-links dist duecare-llm-cli

The meta-package duecare-llm remains the desired one-command distribution story. It was smoke-tested in an isolated virtualenv from the locally built wheels for duecare --help, duecare domains list, and an end-to-end duecare run rapid_probe --target-model local_smoke --domain trafficking workflow against a local OpenAI-compatible fake backend. Real Gemma/Ollama/API runs still require the corresponding target-model backend and credentials/model files.

The duecare-llm-cli path was smoke-tested in an isolated virtualenv from the locally built wheels: install, duecare --help, duecare init, and duecare demo-stage passed.

Readiness notes

• scripts/build_all_wheels.py now includes all 18 package directories in its default build order.
• .github/workflows/pypi-publish.yml is the sole registry publisher. Generic v* tags do not publish packages, manual runs cannot target production PyPI, and a package-NAME-vMAJOR.MINOR.PATCH tag fails closed unless it selects one exact name/version row in configs/duecare/package_release.toml.
• Packages use independent SemVer. Manual release-candidate runs build all manifest rows in canonical dependency order by default, while a production tag builds and publishes only its selected package.
• duecare-llm-chat intentionally remains on an independent harness cadence; the current notebook portability contract is 0.17.0, exposed through duecare.chat.portability and GET /api/portability, so synchronized 0.1.0 infrastructure package tags are not confused with the chat wheel version.
• The May readiness receipt covered the original 17 packages. The added duecare-llm-kit has its own clean wheel/sdist and isolated-install checks; rerun the 18-package wheel build before a whole-workspace candidate claim.
• scripts/build_all_wheels.py now verifies critical domain-pack files in the duecare-llm-domains wheel and fails on missing or duplicated entries.
• duecare-llm-models lazy-loads the optional Ollama HTTP dependency so importing duecare.models does not require httpx unless the Ollama adapter is used; the ollama extra now installs httpx explicitly.
• Final release readiness still needs a release-grade clean-environment build/install run before claiming PyPI readiness; a fully offline/no-index install also needs a complete third-party dependency wheelhouse.
• duecare-llm-chat remains 0.17.0, duecare-llm-server remains 0.1.2, and the other distributions remain 0.1.0. Those versions are reconciled by the independent-release manifest instead of being treated as a blocker.