offbit-reflow — Python SDK for Reflow

Reflow is a modular flow-based programming runtime built on the actor model. Graphs are declarative DAGs: each node is an actor with named in/out ports, edges route messages, and a network executor runs the whole thing with bounded backpressure and a tracing stream. It ships a standard library of ~300 actors covering data, media, GPU rendering, animation, I/O, and optional ML / CV — plus the hooks to register your own.

This package is the official Python SDK. It wraps the runtime via pyo3 and exposes idiomatic Python classes that mirror the Node / Go SDKs one-for-one.

pip install offbit-reflow

from offbit_reflow import Actor, Network, Message

Quick start

from offbit_reflow import Actor, Network, Message

class Doubler(Actor):
    component = "doubler"
    inports = ["in"]
    outports = ["out"]

    def run(self, ctx):
        n = ctx.inputs["in"]["data"]
        ctx.done({"out": Message.integer(n * 2)})

class Log(Actor):
    component = "log"
    inports = ["in"]
    outports = []

    def run(self, ctx):
        print("got:", ctx.inputs["in"])
        ctx.done()

net = Network()
net.register_actor("tpl_doubler", Doubler())
net.register_actor("tpl_log", Log())

net.add_node("a", "tpl_doubler")
net.add_node("b", "tpl_log")
net.add_connection("a", "out", "b", "in")
net.add_initial("a", "in", {"type": "Integer", "data": 21})

net.start()
# ... later:
net.shutdown()

Authoring actors

Subclass Actor. Class-level attributes declare ports and await semantics; the instance run(ctx) method is the per-tick body:

class Sum(Actor):
    component = "sum"
    inports = ["a", "b"]
    outports = ["sum"]
    await_all_inports = True

    def run(self, ctx):
        a = ctx.inputs["a"]["data"]
        b = ctx.inputs["b"]["data"]
        ctx.done({"sum": Message.integer(a + b)})

Inside run(ctx):

Exactly one of done / fail must be called per tick. If run raises, the SDK calls fail with the exception's message.

Multi-graph composition

Merge N GraphExport dicts into a single runnable graph:

from offbit_reflow import compose_graphs, Graph, Network

composed = compose_graphs({
    "graphs": [left_export, right_export],   # dicts
    "connections": [
        {"from": {"process": "gsrc/src",   "port": "out"},
         "to":   {"process": "gsink/sink", "port": "in"}},
    ],
    "shared_resources": [],
    "properties": {"name": "pipeline"},
    "case_sensitive": False,
})

g = Graph.from_json(composed)
net = Network.from_graph(g)

Standard component catalog

The wheel ships the pure-Rust + av-core slice of reflow_components — roughly 270 templates covering animation, flow control, math, vector, 2D graphics, asset DB, scene graph, HTTP integration, stream ops, DSP, and procedural generation. Heavy optional palettes (GPU, ML, browser automation, video encoding, window events, ~6,700 API-service wrappers) are not bundled and install as actor packs.

from offbit_reflow import template_actor, template_list

net.register_actor("tpl_http_request", template_actor("tpl_http_request"))
print([tid for tid in template_list() if tid.startswith("tpl_math_")])

Full catalog reference: docs/components/standard-library.md.

Actor packs

Packs are .rflpack bundles that publish additional templates into this SDK at runtime. template_actor(id) and template_list() transparently include pack-supplied templates after load.

import offbit_reflow as reflow

# Peek before committing.
print(reflow.inspect_pack("./reflow.pack.ml-0.2.0.rflpack"))

# Load (idempotent).
reflow.load_pack("./reflow.pack.ml-0.2.0.rflpack")

# Pack-owned templates now resolve normally.
net.register_actor("tpl_ml_run_inference",
                   reflow.template_actor("tpl_ml_run_inference"))

print(reflow.list_packs())
print(reflow.pack_abi_version())

First-party packs live under sdk/packs/:

Where to get .rflpack files

First-party bundles ship as assets on every GitHub Release whose tag starts with pack-v. Pack and SDK builds must come from the same release wave (matching REFLOW_PACK_ABI_VERSION) — see the pack ↔ SDK compatibility matrix for the supported pairings. Each release ships two flavours of every pack:

VER=0.2.0
# Slim variant for the host you're running on (Apple Silicon shown).
curl -LO https://github.com/offbit-ai/reflow/releases/download/pack-v$VER/reflow.pack.ml-$VER-aarch64-apple-darwin.rflpack

# Or the full bundle if you don't know the deployment target ahead of time.
curl -LO https://github.com/offbit-ai/reflow/releases/download/pack-v$VER/reflow.pack.ml-$VER.rflpack

Triples published per pack are listed in sdk/packs/README.md.

load_pack() accepts either flavour identically — it picks the binary that matches the runtime triple at load time.

To slim a downloaded full bundle yourself, install the reflow_pack_cli crate and run:

reflow-pack strip reflow.pack.ml-0.2.0.rflpack
# → reflow.pack.ml-0.2.0-<host-triple>.rflpack

Third-party packs are distributed however their author chooses (PyPI data files, GitHub Releases, internal registry) — any local file path works with load_pack().

ABI lockstep. A pack is pinned to the SDK release it was built against. Pick the pack-v* release whose version matches your offbit-reflow; rebuild from source (sdk/packs/README.md) if you need a pack for a different SDK version.

Subgraphs

from offbit_reflow import SubgraphBuilder

sub = SubgraphBuilder(graph_export_json)   # dict or parsed object
sub.register_actor("my_custom", MyCustom())
sub.fill_from_catalog()                    # resolve bundled components
sg = sub.build()
net.register_actor("tpl_sub", sg)

Streams

Producer side:

from offbit_reflow import Stream

s = Stream.create(buffer_size=64, content_type="image/jpeg")
s.send_bytes(frame1)
s.send_bytes(frame2)
s.end()
ctx.done({"out": s.into_message()})

Consumer side:

rdr = ctx.inputs["frames"].take_stream()
while True:
    f = rdr.recv(500)
    if f["kind"] == "data":
        handle(f["data"])
    elif f["kind"] == "end":
        break
    elif f["kind"] in ("closed", "timeout"):
        break
    elif f["kind"] == "error":
        raise RuntimeError(f["error"])

Events

events = net.events()
while True:
    evt = events.recv(timeout_ms=200)
    if evt is None:
        continue
    print(evt.get("_type"), evt)

Subscribe before net.start() so no events are missed.

Building locally

cd sdk/python
python -m venv .venv && source .venv/bin/activate
pip install maturin pytest
maturin develop
pytest -q

Releasing

Releases are built and published by CI — see .github/workflows/publish-python.yml. Tag a commit with python-v<version> (e.g. python-v0.2.0) and the workflow builds wheels for every supported triple (linux x86_64/aarch64, macOS x86_64/aarch64, windows x64), plus an sdist, verifies metadata, smoke-tests the wheel on each host, and uploads everything to PyPI.

Publishing currently uses an API token stored as the PYPI_API_TOKEN repository secret. Migration to PyPI trusted publishing (OIDC) is a one-line swap once the first release is live.

License

MIT OR Apache-2.0.