fletchr-studio

The flowgraph engine behind the fletchr studio GUI: a graph document model for source → transformer → sink flowgraphs, a block palette built by introspecting the live fletchr registries, graph validation, and Python code generation with round-trip reopen.

The package also ships the studio itself: a FastAPI host (fletchr_studio.server) serving a React Flow canvas UI, with sessions, sampled/full graph runs, and per-node table previews. See docs/design/studio.md at the workspace root for the design record and phased plan.

Running the GUI

fletchr-studio            # serves http://127.0.0.1:8410

Drag blocks from the palette, connect them (incompatible ports are rejected), configure params in the inspector, then Preview Run — sources are sampled and sinks are skipped, so a preview never overwrites real outputs. Click any node to see its table preview. Full Run executes everything, sinks included. Generate shows the emitted script; Save writes it; Open reopens a generated .py (hand edits are flagged as drift).

Rebranding for meta-packages

A downstream protocol wrapper can present the studio as its own tool. Ship a console script that launches it with a Branding:

# acme_decoder/studio.py
import acme_decoder.plugins  # noqa: F401 - registers blocks via entry points
from fletchr_studio import Branding
from fletchr_studio.server import main


def cli() -> None:
    main(branding=Branding(name="ACME Decoder Studio", favicon="path/to/icon.svg"))

[project.scripts]
acme-studio = "acme_decoder.studio:cli"

The name flows into the browser tab, the top bar, and the FastAPI title; the favicon replaces the default. Generated scripts still record fletchr-studio <version> as their generator — that's provenance, not presentation. The wrapper's transformers, readers, and writers appear in the palette automatically via the normal entry-point plugin discovery.

The canvas bundle builds into src/fletchr_studio/static/ and ships in both the wheel and the sdist. Packaging is guarded by hatch-jupyter-builder: building a dist with the bundle already present needs no Node (skip-if-exists); with the bundle missing it runs npm install + npm run build itself; and with neither bundle nor Node the build fails rather than producing a headless dist (ensured-targets). Editable installs only warn, so Python-only contributors without Node get a working headless dev server. Escape hatch for deliberate headless builds: SKIP_JUPYTER_BUILDER=1.

To rebuild the bundle by hand you need Node 18+:

cd frontend
npm install
npm run build     # type-checks and outputs to ../src/fletchr_studio/static
npm run dev       # dev server with /api proxied to a running fletchr-studio

What it does

from fletchr_studio import GraphDoc, generate_code, load_graph

doc = GraphDoc()
doc.add_node("capture", "source", "read_file", path="capture.arrow")
doc.add_node("sel", "transform", "Subframe", columns="1-32")
doc.add_node("inv", "transform", "Invert")
doc.add_node("out", "sink", "write_file", path="frames.parquet")
doc.add_edge("capture", "sel")
doc.add_edge("sel", "inv")
doc.add_edge("inv", "out")

code = generate_code(doc)

generate_code validates the graph (unknown blocks, port type mismatches, cycles, missing params — reusing the same subclass-aware compatibility rules Pipeline enforces) and emits a plain Python script depending only on the fletchr packages:

"""Flowgraph generated by fletchr-studio 0.1.0."""

import argparse

from fletchr_core import read_file, write_file
from fletchr_core.transform import Invert, Subframe


def main(capture_path='capture.arrow', out_path='frames.parquet'):
    capture = read_file(capture_path)
    inv = (Subframe(columns='1-32') | Invert())(capture)
    write_file(out_path, inv)


def _cli():
    parser = argparse.ArgumentParser(description="Flowgraph generated by fletchr-studio.")
    parser.add_argument("--capture", dest="capture_path", default='capture.arrow', help="source 'capture' path")
    parser.add_argument("--out", dest="out_path", default='frames.parquet', help="sink 'out' path")
    main(**vars(parser.parse_args()))


if __name__ == "__main__":
    _cli()


__fletchr_graph__ = {...}

Source and sink paths are lifted into main()'s signature and an argparse CLI, so the artifact is parameterizable without editing: python flow.py --capture other.bits, or import it and call main(capture_path=...) from a loop.

The embedded __fletchr_graph__ literal is the graph document — load_graph("flow.py") reopens it, and a stored SHA-256 of the code section detects hand edits (drift) without parsing arbitrary Python. Linear chains compile to | pipelines; fan-out becomes named intermediates; merge blocks (stack, merge) compile to plain calls.

The palette (build_palette()) discovers every registered transformer — including plugin-contributed ones — plus file source/sink blocks and the merge blocks, with parameter schemas introspected from attrs fields or __init__ signatures and port types from apply annotations.