Long live OpenSysML by HuiJun · Pull Request #56 · Open-MBEE/OpenSysML

HuiJun · 2026-08-18T20:36:56Z

We have renamed the project to OpenSysML, reflecting our commitment to keep this implementation available to the community.

OpenSysML 0.1.0

Everything cut under 0.1.0 in CHANGELOG.md. The release is the project's rename to
OpenSysML plus the first batch of work that makes a model decide, run and render
rather than only parse and resolve.

What this is

A SysML v2 / KerML 1.1 implementation in Go: hand-written lexer and parser, immutable AST with
semantic side tables, lazy memoized name resolution, tiered validation, lowered ActionGraph /
StateGraph execution IR, and four front ends over one core — the sysml CLI/REPL, the
sysml-lsp language server (with a VS Code extension), the sysml-grpc service, and the
opensysml Python client generated against it.

Highlights

Objects run the behavior their type exhibits. Materializing an object starts the
exhibit state machines and performed actions its type declares: one execution, event queue
and feature-value set per object, entry/do/exit/effect bodies reading and writing that
object's values, cross-object send reaching the addressed object, and a defined startup and
quiescence with typed budget errors instead of a hang. %invoke <object> <op> performs an
operation of the object's type; %state <object> attaches the debugger to the object's own
machine.
An SMT solver decides what a model's conditions permit (experimental, solver optional):
%check for satisfiability with an assignment, %explain for a minimal unsat core with
file:line:col per conflicting condition, %solve for a witness that keeps fixed what already
is, %configure for the variant selections an assertion permits, and %optimize for the
objectives an analysis def states — lexicographic, with every optimum verified rather than
reported on trust. What a backend can do is modelled as probed capabilities, so a missing
feature is a typed refusal naming backend, feature and operation; z3 and cvc5 are both gated in
CI, with and without a solver present.
A view renders. %render <name> [form] and sysml -render turn a view's exposed set into a
tree, interconnection diagram, state machine, action flow or table, as text, Mermaid or Markdown.
State and action renderings read the same lowered graphs the runtime executes, so a diagram
cannot drift from what runs. Rendering is a read: no object is materialized and a debugging
session keeps stepping across it.
A model is edited through the source it was parsed from. ApplyEdits (and model.edit() in
Python) rewrites the bytes of the parsed source, leaving every untouched byte identical, then
re-parses and re-analyzes before handing the result back — all edits of a request or none.
Behavior converts to RDF (experimental): action and state bodies now have metaclasses and
round-trip byte-identically through Turtle; 102 of 120 example models convert, up from 71, and
the remaining 18 are refused with the construct named rather than partly written.
Semantics and notation fixes worth calling out: a nested value body governs over an
inherited value (KerML §7.3.4.5); require/assume bodies resolve and type-check to any depth
in their own scope; binding connectors propagate in both directions with typed conflict and
cycle errors; bind accepts qualified and chained ends; named control nodes register their
names; feature-chain resolution is linear in chain length; and an element reached through two
scope trees is exposed and rendered once, since a symbol's identity is now the declaration it
was built from.
The exp/ln/log/atan2 inconsistency is gone. They are declared by the non-normative
OpenSysMLMathFunctions extension, so a bare call is now a typed error naming the import that
makes it legal, instead of being reported unresolved and evaluated anyway.

Breaking changes

The rename is a clean break — every name has exactly one spelling from this release on:

Was	Is
`github.com/Open-MBEE/Systemica`	`github.com/Open-MBEE/OpenSysML` (old path resolves only for `v0.0.x`)
`pysysml` (PyPI, import)	`opensysml` 0.3.0; entry point `opensysml-generate`, state dir `~/.opensysml`, tag `opensysml-v*`
`SYSTEMICA_*` env vars	`OPENSYSML_*`
`urn:systemica:sysml:`	`urn:opensysml:sysml:` (still the `sysx:` prefix)
`SystemicaMathFunctions`	`OpenSysMLMathFunctions`
VS Code `systemica.*` settings	`opensysml.*`, extension `opensysml-sysml`
archives `systemica-<os>-<arch>`	`opensysml-<os>-<arch>`, formula `Open-MBEE/tap/opensysml`

The pre-release "slot" vocabulary is removed rather than aliased, since it never shipped:
%slots → %features; Instance.slots and SlotValue are gone from the wire with field
number 3 and the name slots reserved, so neither is reused; the Python client keeps only
features / raw_features / get_feature / FeatureValueError / feature_value*, and requires
the service's feature_values capability before handing back an object, so a pre-0.1.0 service is
named instead of answering with an object that appears to hold nothing. Diagnostic text changed
with the vocabulary (feature value …, do action steps); exit statuses did not.

Not in scope / known limitations

Stated in full under Known limitations in the 0.1.0 changelog entry. The ones that shape
expectations: constraint solving and RDF conversion are experimental (RDF omits expressions
and has no demonstrated triplestore round trip); %optimize needs z3 specifically; an edit sets a
value or renames a declaration and refuses create/delete/rename-of-referenced; only an action
member is executable via %invoke; a package-owned binding does not propagate; and a rendering is
tool-defined output, since SysML v2 §10.2 leaves rendering to the tool.

Verification

gofmt -l . empty, go build ./..., go vet ./..., go test -race ./... and make lint
(staticcheck + gosec) green; the Python suite green against a service built from this source; the
stdlib conformance gate clean; the OMG training corpus at its pinned 98/100 baseline; SMT suites
run under both z3 and cvc5 as well as with no solver present; scripts/check-doc-links.py reports
no broken link. Compliance is tracked rule by rule, with honest status flags, in
docs/project/spec-compliance.md.

…l-metaclasses * feat(export): map behavioral nodes to RDF metaclasses An action or state body converts instead of being refused: the initial and final node, perform, send, accept, terminate, assign, the fork/join/merge/ decision control nodes, while/loop/for, if/else, and the state machine's states, substates, regions, entry/do/exit, defer, pseudostates and transitions, each with a notation to RDF to notation round trip. Shapes that are not mapped, including a succession that does not name both of its ends, are still reported with the node named. Co-Authored-By: jason.han <jason.han@jpl.nasa.gov> * docs(export): record the behavioral RDF mapping and the new conversion rate 102 of the 120 models under examples/ now convert, up from 71. RDF stays experimental: no triplestore round trip has been run. Co-Authored-By: jason.han <jason.han@jpl.nasa.gov> * fix(export): carry metadata annotations and implicit kinds through RDF A metadata annotation was converted as an empty string and lost, and a feature that wrote no kind keyword came back with the kind's canonical keyword written out. Both are now carried, and the annotation shapes that cannot be written back are reported. Corrects the experimental notice, which still said behavioral bodies are refused. Co-Authored-By: jason.han <jason.han@jpl.nasa.gov> * fix(export): publish the else marker as a sysx term The flag helper only routes an allow-listed name to the extension namespace, so the else marker of a branch was emitted as sysml:isElse, a term the OMG vocabulary does not define, while the decoder and the docs both name sysx:isElse. It is now written directly, as every other behavioral flag in the file is, and pinned by a test that reads the branch back from the marker alone. Co-Authored-By: jason.han <jason.han@jpl.nasa.gov> --------- Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Co-authored-by: jason.han <jason.han@jpl.nasa.gov>