Companion repositories: reusable-component inventory

This inventory covers pi-agent-kit and LLM4ChemE, both inspected read-only in October 2026.

Licensing

Both repositories carry a proprietary Trade Secret License, while core is AGPL-3.0-or-later. A component can enter core only in one of two ways:

1. Relicensing. Its owners relicense the component under AGPL-3.0-or-later (or a compatible license) and contribute it. Third-party data has its own terms: LLM4ChemE's offline_databases/{polymers,solvents} each carry their own LICENSE and SOURCE.md, and those terms decide separately.
2. Staying external. The component remains where it is. A project, or a downstream application, depends on it next to core. core never imports it.

LLM4ChemE is in category 1. It is owned by the same group as core, and its owner has cleared its code and skills for migration into core under core's license, AGPL-3.0-or-later (the license at the time of clearing and again now). Third-party data that moves keeps its own LICENSE and SOURCE.md files, and a dataset moves only if its license permits redistribution in an AGPL-3.0-or-later package. What moves, where, and what stays is tracked in llm4cheme-migration.md; the LLM4ChemE table below is the original inventory. After M2 every row is either in core or recorded there as not migrated; the solvent and polymer datasets stay in LLM4ChemE until their upstream licenses are recorded.

pi-agent-kit stays in category 2. It is not cleared; nothing from it is copied into core. The "Target" column of its table only records where a component would belong after relicensing.

pi-agent-kit

The package is still template-ML-repo, a project template whose only runtime dependency is requests. It has 2 test files and CI (ruff, pyright, pytest).

Component	What it does	Deps	Tests	Target	Blockers
tools/property_retrieval/online_retrieval.py	PubChem lookup by name: formula, molar mass, SMILES, InChIKey, descriptors	requests (network)	yes	domains/chemistry (property sources) as an integrations/pubchem adapter	license; network access must stay optional and cached
skills/property-retrieval/ (SKILL.md, scripts/batch_lookup.py)	agent skill and batch script around the tool above	requests	yes	domains/chemistry/skills/	license; the script should call the shared function, not keep its own copy
CI, Dockerfile, README template	project scaffolding	none	n/a	none	nothing to migrate

LLM4ChemE

The package is llm4cheme. Its dependencies are langchain/langgraph, OpenAI/Ollama/Mistral clients, thermo, rdkit, Pint, scipy and pandas. It pins pyDEXPI-dev by commit. There are about 113 test files, and CI runs ruff, pyright and pytest, with testcontainers for Postgres.

Component	What it does	Deps	LLM-coupled	Tests	Target	Blockers
src/domain/, src/application/, src/infrastructure/ (agents, routers, LLM factories, YAML graph factory, skill loader), src/cli.py, main.py, configs/	multi-agent framework and runtime	langchain, langgraph, LLM SDKs	yes	yes	stays external (agent runtime; core stays usable without one)	LLM runtime; license
src/tools/dexpi_process/	deterministic edits of a DEXPI ProcessModel	pyDEXPI	no	yes	none	uses pyDEXPI's private editor layer (its process.* commands), which core must not use; core edits files directly
src/llm_tools/dexpi_process/	LLM tool wrappers and argument schemas for the above	pyDEXPI, langchain	yes	yes	none	private editor layer; LLM-coupled
src/workflows/dexpi_process/ (graph state, document, edit summary, validation report, serialization)	graph state around a ProcessModel; edit summaries through pyDEXPI diff	pyDEXPI, langgraph	yes	yes	none	process_model_document.py and validation_report.py reference a private pyDEXPI extension package; editor layer; LLM state
skills/general-dexpi-operations, materials, process-steps-and-streams (with references/)	DEXPI ProcessModel skills and class references	none	skill text	yes (references checked against pyDEXPI)	none; core's process skill is YAML-first (process-design)	skill text written for the editor tools; general-dexpi-operations refers to a private pyDEXPI extension package; license
src/tools/chemical_property/, src/databases/online_databases/ (PubChem, NIST WebBook, OSHA, thermo, UNIFAC retrieval), cactus.py, rdkit_tools.py	unified property lookup over online sources, thermo and RDKit	thermo, rdkit, requests; scrapy and selenium (extra)	mostly no	yes	domains/chemistry + integrations/ (one adapter per source, optional extra)	license; network and scraping; overlaps with core's pinned thermo adapter (process/props.py); OSHA data is safety-relevant
src/databases/offline_databases/, src/tools/offline_databases/	local solvent and polymer databases (SQLite/Postgres) with query tools	SQLAlchemy, sqlmodel, psycopg	tools yes	yes	stays external, or integrations/ behind an extra	license plus third-party data licenses; database server dependency
src/tools/phase_calculations/, wrapped_phase_tool.py, wrapped_thermo_tools.py, skills/phase-calculation	flash and phase calculations	thermo, Pint	wrappers yes	yes	domains/process (thermodynamics), next to process/props.py	license; must share core's pinned thermo method rather than add a second one
src/tools/thermo_model_selection/, skills/thermodynamic-model-selection	rule-based choice of thermodynamic model (knowledge base and inference)	none	engine no	yes	domains/chemistry or domains/process (property method choice)	license; the knowledge base's provenance must be stated
src/tools/separation_selection/, skills/separation-method-selection, src/workflows/separation_method/	separation method selection	none	partly	yes	domains/process (separations)	license; tied to evaluation datasets
src/tools/reaction_kinetics/, src/databases/{nist_kinetics,rmg_kinetics,open_reaction_db}.py, src/schemas/reaction_kinetics.py, kinetics_output_schema.py, src/workflows/reaction_kinetics/	kinetics retrieval and filtering (NIST, RMG, Open Reaction Database)	rdkit, requests	partly	yes	domains/chemistry (kinetics) + integrations/	license; network sources
src/tools/heat_integration/, src/schemas/heat_integration.py	pinch analysis	PyPinch, Pint	tools yes	not seen	domains/process (heat integration)	license; no tests found
src/tools/unit_conversion.py, src/schemas/quantity.py	unit conversion and quantity schema	Pint	no	not seen	would inform a foundation units convention; core keeps its own unit tables in process.model for now	license; adding Pint to foundation must be decided deliberately
services/rmg_service/, services/grappa_service/, services/hanna_service/ and their clients (tools_grappa.py, tools_hanna.py)	FastAPI services around RMG, a GNN property model and an activity-coefficient model	RMG, ML runtimes, FastAPI, Docker	client wrappers	yes / partly	integrations/ clients only; the services stay external	license; heavy runtimes
src/evaluation/, src/data/*.json (benchmark datasets), examples/ notebooks	evaluation harness, benchmark data, tutorials	langsmith, openai	yes	yes	stays external (paper and research)	research-specific
scripts/generate_pydexpi_reference.py, example_dexpi_bfd.py	generate reference docs from pyDEXPI classes; example	pyDEXPI	partly	n/a	none (core has its own scripts/build_snippets.py)	license

No component in either repository covers requirements, control, spatial or (apart from the OSHA lookup) safety.

Research and paper projects

A paper project is not migrated into core. It keeps its own repository, with its manuscript, experiments, datasets and results, and depends on the framework:

• It pins cheme-as-code in its pyproject.toml (a git URL with a tag or commit, as core itself does for pyDEXPI), plus the extras it needs (props, serve, png).
• It imports domain functions directly, for example from chemeascode.domains.process.model import load, balances. It uses cheme init only when it is itself a plant repository.
• It records the cheme-as-code version and the git commit of its inputs alongside its results, using the same provenance fields core writes (generator_version, source_sha256, commit sha), so that results can be traced.
• A capability that proves reusable is proposed to core as an ordinary contribution, under AGPL, with tests. The rest stays in the paper repository.