StatifierBlocks

CIHex.pm VersionHex DownloadsHex DocsLicense

Block document model, one-way SCXML compiler, and LiveView editor components for composing Statifier statecharts from typed blocks.

Statecharts are the right execution model for long-running workflows, and SCXML is the right interchange format for them - but neither is something a non-engineer will author by hand. This package is the authoring layer:

Blocks are typed and host-pluggable: a host registers the block types its own domain needs, and the compiler and editor work off that registry rather than off a closed built-in vocabulary.

Installation

def deps do
[
{:statifier_blocks, "~> 0.3"}
]
end

A worked example

A card-processing flow: place a hold, and settle it when the account has the budget for it. Everything below runs - it is the example the suite executes on every build.

1. Write the block types your domain needs. A block type is a behaviour module: a handful of declarations plus one emit/2. These two are invoking leaves, so they share their emission.

defmodule MyApp.Blocks do
@moduledoc "Emission helpers shared by this host's invoking leaves."
alias StatifierBlocks.{Block, Emission}
alias StatifierBlocks.Compiler.Context
alias StatifierBlocks.Core.Emit
@doc "One compound state that starts an `<invoke>` and finishes either way."
def invoke_leaf(%Block{config: config}, %Context{} = context) do
done = Context.done_id(context)
{:ok, running} = Context.role_id(context, "running")
{:ok, invocation} = Context.role_id(context, "invocation")
waiting =
Emit.state(running, nil, [
Emission.element("invoke", [
{"id", invocation},
{"type", Map.get(config, "invoke_type", "")}
]),
Emit.transition(event: "done.invoke." <> invocation, target: done),
Emit.transition(event: "error.execution", target: done)
])
{:ok, Emit.state(context.state_id, running, [waiting, Emit.final(done)])}
end
end
defmodule MyApp.Blocks.Authorize do
@moduledoc "`myapp.authorize`: places a hold on the card."
@behaviour StatifierBlocks.BlockType
@impl true
def current_version, do: 1
@impl true
def slots(_config), do: []
@impl true
def config_schema(_config),
do: [%{key: "invoke_type", type: :string, label: "Invoke", required?: true, default: ""}]
@impl true
def validate_config(_config), do: :ok
@impl true
def io(_config), do: %{kinds: [:step], produces: "myapp.credit_card_txn"}
@impl true
def emit(block, context), do: MyApp.Blocks.invoke_leaf(block, context)
end
defmodule MyApp.Blocks.Capture do
@moduledoc "`myapp.capture`: settles a hold this flow already placed."
@behaviour StatifierBlocks.BlockType
@impl true
def current_version, do: 1
@impl true
def slots(_config), do: []
@impl true
def config_schema(_config),
do: [%{key: "invoke_type", type: :string, label: "Invoke", required?: true, default: ""}]
@impl true
def validate_config(_config), do: :ok
@impl true
def io(_config), do: %{kinds: [:step], consumes: "myapp.credit_card_txn"}
@impl true
def emit(block, context), do: MyApp.Blocks.invoke_leaf(block, context)
end

io/1 is where a type declares how data flows through it. produces and consumes are opaque strings compared for identity, widened only by a relation the host supplies - there is no built-in type lattice.

That relation rides on the palette (Palette.new(types, assignability: MyApp.Blocks.Types)) and it reaches every consumer through that one value. Assignability.validate/3 - what the compiler runs over the whole document - and Edit.Targets.slot_verdicts/3 - what the editor runs once at drag start to mark every droppable slot before the pointer moves - are the same implementation reading the same relation, so widening the host module opens a drop target and clears the matching finding in the same edit. The relation can only widen: identity is checked first, so a host callback can never refuse something the default rule accepts.

Where a seam refuses, Assignability.finding_reason/2 says why in a small vocabulary (:not_assignable, {:fixable_by, block_id}), and Assignability.seam_reasons/3 names the seams that passed only because a block declared nothing (:source_untyped, :target_untyped, :both_untyped) - the way to find the parts of a palette you have not typed yet. The editor stamps a refused slot's reason beside its validity as data-drop-reason.

2. Compose the document. Your two types, arranged by the core.* vocabulary this package ships. Twelve types: the containers that arrange other blocks (core.sequence, core.group, core.branch, core.parallel, core.resumable_group), and the leaves that do a structural thing on their own (core.wait, core.on_event, core.invoke, core.subchart, core.send, core.raise, core.assign). None of them knows a domain - core.invokenames an invoke type for the host to run and never runs one, and core.subchart names another chart the same way. StatifierBlocks.Core carries the table of all twelve with their slots. In a running system an editor writes this tree; it is ordinary data either way.

alias StatifierBlocks.{Block, Compiler, Document, Palette, Provenance}
document =
Document.new(
Block.new("core.sequence",
id: "blk_root",
slots: %{
"body" => [
Block.new("myapp.authorize",
id: "blk_authorize",
config: %{"invoke_type" => "myapp:authorize"}
),
Block.new("core.branch",
id: "blk_approved",
config: %{
"arms" => [%{"slot" => "arm_approved", "cond" => "budget_remaining > amount"}]
},
slots: %{
"arm_approved" => [
Block.new("myapp.capture",
id: "blk_capture",
config: %{"invoke_type" => "myapp:capture"}
)
]
}
)
]
}
),
id: "bdoc_card_capture"
)

3. Build a palette and compile. A palette is a plain value - a type_name => module map you build for one operation and pass explicitly. It is deliberately not application config and not a named process, so two tenants in one runtime never step on each other's block types.

palette =
Palette.new(
Map.merge(Palette.core_types(), %{
"myapp.authorize" => MyApp.Blocks.Authorize,
"myapp.capture" => MyApp.Blocks.Capture
})
)
{:ok, compiled} = Compiler.compile(document, palette)

Compiler.compile/3 is a total function of {document, palette}: no process state, no clock, no IO. It returns {:ok, %StatifierBlocks.Compiled{}} or {:error, findings} - never a raise, never a partial success. The artifact carries the generated bytes, the provenance map, a compilation record joining document identity to chart identity, and the invoke types the chart names:

compiled.invoke_types
#=> ["myapp:authorize", "myapp:capture"]

The SCXML it produced is a chart Statifier runs as-is - one compound state per block, completion signalled by done.state:

<scxml initial="s_blk_root" name="bdoc_card_capture" version="1.0" xmlns="...">
<state id="s_blk_root" initial="s_blk_authorize">
<transition event="done.state.s_blk_authorize" target="s_blk_approved" type="internal"/>
<transition event="done.state.s_blk_approved" target="s_blk_root__o_done" type="internal"/>
<state id="s_blk_authorize" initial="s_blk_authorize__running">
<state id="s_blk_authorize__running">
<invoke id="s_blk_authorize__invocation" type="myapp:authorize"/>
...

4. Point a running position back at a block. That is what the provenance map is for. Hand it the active state ids of a live session and it answers with the blocks the session is inside - which is how an editor highlights the step a run is on, and how a chart-level finding routes back to the config field somebody typed it into.

active_state_ids = Map.keys(compiled.provenance.by_state_id)
blocks_in_play =
compiled.provenance
|> Provenance.owners_of_states(active_state_ids)
|> Enum.map(& &1.block_id)
|> Enum.uniq()
|> Enum.sort()
#=> ["blk_approved", "blk_authorize", "blk_capture", "blk_root"]

For a fixed {document canonical bytes, palette, compiler version} the generated SCXML is byte-identical on every machine and every run, and compiled.record carries all three - so a host can skip a recompile on an unchanged triple. The guarantee is not reversible: identical SCXML does not mean an unchanged document, because metadata is not compiled.

The package's two full worked examples - this card-processing flow and a signup wizard with A/B testing (myapp:signup, variants, conversion events) - live in test/support/document_fixtures.ex and are stored as canonical bytes under test/fixtures/documents/. Between them they reach the whole core.* vocabulary.

Config fields and where their values live

A block type's config_schema/1 declares the fields the editor renders for it. A field's key is its identity: the DOM id, the form param name, and what a {:config, block_id, key} finding anchors to. Where the value is stored is a second, separate question, and a field answers it with an optional value_path - a list of keys and list indexes from the config root down to the value it edits.

Most fields need no path: key alone addresses config[key]. Some cannot use one. core.branch keys a condition field by the arm's slot name, because that is what a finding has to name, while the condition itself is stored inside the ordered "arms" list:

alias StatifierBlocks.{BlockType, Core}
config = %{"arms" => [%{"slot" => "arm_approved", "cond" => "budget_remaining > amount"}]}
[field] = Core.Branch.config_schema(config)
field.key
#=> "arm_approved"
BlockType.value_path(field)
#=> ["arms", 0, "cond"]
BlockType.fetch_value(config, BlockType.value_path(field))
#=> {:ok, "budget_remaining > amount"}
BlockType.put_value(config, BlockType.value_path(field), "amount <= 5000")
#=> %{"arms" => [%{"cond" => "amount <= 5000", "slot" => "arm_approved"}]}

value_path/1 answers [key] for a declaration that declares no path, so a caller never branches on which case it has. fetch_value/2 is total and answers :error for a path that does not resolve; put_value/3 writes the last segment whether or not a value was already there - an arm with no condition yet is exactly the one an author is about to type into - but never invents an intermediate map or list a block type did not write. A host block type that stores a value somewhere other than a top-level key declares the path the same way.

Registering your own block types

The core.* vocabulary is structural on purpose: it knows sequencing, branching, waiting and parallelism, and nothing about anyone's domain. A card-processing host adds the steps its own product has by writing a module per step and handing the editor an explicit list of them where the editor is mounted. There is no global registry, no application-configuration lookup, and no discovery pass that finds every module implementing the behaviour - each of those would make two tenants in one runtime share a vocabulary that is supposed to be per palette.

defmodule MyApp.Blocks.RiskHold do
@moduledoc "myapp.risk_hold: parks an authorization until a reviewer clears it."
@behaviour StatifierBlocks.BlockType
alias StatifierBlocks.Compiler.Context
alias StatifierBlocks.Core.Emit
@impl true
def current_version, do: 1
@impl true
def slots(_config), do: []
@impl true
def config_schema(_config),
do: [
%{
key: "queue",
type: :string,
label: "Review queue",
required?: true,
default: "fraud"
}
]
@impl true
def validate_config(config) do
case Map.get(config, "queue") do
queue when is_binary(queue) and queue != "" -> :ok
_missing -> {:error, [{"queue", "name the queue a reviewer picks this up from"}]}
end
end
@impl true
def palette_entry,
do: %{
label: "Risk hold",
group: "Payments",
description: "Parks the authorization until a reviewer clears it.",
badge: "manual review",
accent_token: "--sb-accent-risk"
}
@impl true
def emit(_block, context) do
done = Context.done_id(context)
with {:ok, holding} <- Context.role_id(context, "holding") do
waiting =
Emit.state(holding, nil, [
Emit.transition(event: "myapp.risk.cleared", target: done)
])
{:ok, Emit.state(context.state_id, holding, [waiting, Emit.final(done)])}
end
end
end
palette =
StatifierBlocks.Palette.from_modules(
[{"myapp.risk_hold", MyApp.Blocks.RiskHold}],
core: true
)
{:ok, risk_hold} = StatifierBlocks.Palette.fetch(palette, "myapp.risk_hold")
Map.has_key?(palette.types, "core.sequence")
#=> true
StatifierBlocks.BlockType.badge(risk_hold.palette_entry())
#=> "manual review"
StatifierBlocks.ViewModel.accent_token(risk_hold.palette_entry())
#=> "--sb-accent-risk"

from_modules/2 is a value constructor and nothing more - the palette it returns is passed into the editor, the compiler and validation explicitly, the same way Palette.new/2 and Palette.core/0 are. The list is ordered and later entries win, so core: true puts the core vocabulary underneath and a host that deliberately swaps in its own core.wait writes it after. The registration carries the type name as well as the module because the document names a type by string and the palette resolves the string: the mapping is the host's fact, which is what lets one module serve two names in two tenants' palettes.

The three presentation declarations

A palette entry may also say how the editor should draw the type, and three of those keys are worth calling out because a host reaches for them immediately:

KeyWhat it declaresAbsent means
accent_tokenthe NAME of a --sb-* custom property, never a colourthe editor's own accent
badgea short chip for the card headerno chip
join_labela one-argument function of config, phrasing the join marker under a side-by-side arrangementthe editor's own word

All three are read through a total normalizer that refuses rather than repairs: a badge longer than 24 characters is dropped, not clipped, and one carrying a newline is dropped rather than collapsed to a space, because a truncated chip reads as a bug in the editor where a missing one reads as the declaration it is. An accent that is not an anchored --sb-* name never reaches a style attribute. A join_label is host code on the layout path, so it is a pure function of its argument and it is called inside a rescue - a type with a bug in it gets an ordinary join marker rather than taking the canvas down.

Embedding the editor

The editor ships in this package, and a host that never renders anything must not pay for it. phoenix_live_view is therefore an optional dependency, and every module under StatifierBlocks.Editor.* is compiled behind a presence guard: an authoring API that compiles documents in a background job, a test suite that exercises validation, a migration script - none of them drag in Phoenix, and none of them compile a line of editor code.

A host that wants the editor already has LiveView, since there is nowhere else to put the editor, so it adds nothing to mix.exs. It does three things:

1. Import the hook. The package's entire client-side surface is one hook. Add the package to assets/package.json:

{ "dependencies": { "statifier_blocks": "file:../deps/statifier_blocks" } }

and register it in app.js:

import { StatifierBlocksDrag } from "statifier_blocks";
let liveSocket = new LiveSocket("/live", Socket, {
hooks: { StatifierBlocksDrag },
});

2. Import the stylesheet. It is structural CSS only - the column layout, the drag affordances, the finding treatments - with no visual opinion and no framework:

@import "../../deps/statifier_blocks/assets/css/statifier_blocks.css";

3. Render the component.

<.live_component
module={StatifierBlocks.Editor}
id="editor"
document={@document}
palette={@palette}
on_change={&save_draft/1}
/>

Optional assigns: findings (yours, merged with the ones the view model derives), icon (a function component that turns an icon name into markup), expression_component (an override for :expression fields), theme, and class.

Icons. You do not have to pass icon. The package ships StatifierBlocks.Editor.Icons, a small set of inline SVGs for the names the core block types declare - no font, no CDN, nothing to register in your asset pipeline - and the editor uses it when you pass nothing. Every glyph paints with currentColor and fills its tile, so the two tokens the tile reads (--sb-block-accent and --sb-block-accent-tint) are all a theme has to touch. See docs/theming.md.

Pass icon when you have an icon set of your own, and it wins on every tile - the canvas cards and the palette rows alike. It is a component taking name and class, and the name is what the block type declared:

<.live_component
module={StatifierBlocks.Editor}
id="editor"
document={@document}
palette={@palette}
icon={&icon/1}
/>
# A heroicons-style component: the name in, your markup out. The core types
# name heroicons ("clock", "bars-3", "arrow-path", ...), so a host already
# using them resolves every one by prefixing.
attr :name, :string, required: true
attr :class, :string, default: nil
def icon(assigns) do
~H"""
<span class={[@class, "hero-" <> @name]} aria-hidden="true" />
"""
end

Two rules the seam keeps. A block type declares a name, never markup, so nothing a palette entry carries is injected into the editor's render tree. And a block type that declares no icon at all gets no tile rather than an empty one, in the shipped set and in yours: your component is never called with a nil name.

Underneath the component is a pure command algebra - StatifierBlocks.Edit (insert, remove, move, update config, each with its inverse) over StatifierBlocks.ViewModel - with no UI framework dependency at all. A host that wants to drive document edits from something other than this editor uses those directly.

What the mounted component holds

The document you pass in, an undo history over it, the current selection, and a drafts map of config edits the validation gate has not accepted yet. A draft is never in the document and never on the undo stack: a form whose config has not been accepted names the fields that are outstanding and offers "Discard edits", because a draft was never a command and so cannot be undone.

There is deliberately no datamodel assign. A field that names a datamodel path - core.assign's path, a core.invoke param - is checked for shape and nothing more, because this package does not own the datamodel path grammar and holds no declaration to check a path against. A host that knows its own datamodel checks paths itself and hands the result in through findings.

The host seams that exist today

Everything a host can say about how its own types behave and look is a declaration on a value it already passes in - the palette, the palette entry, the theme - rather than a callback the editor calls back into:

SeamDeclared onWhat it does
:assignabilityPalette.new/2 (also from_modules/2)the host's widening relation for "may this block land in this slot" - both gates, kind admission and data flow, run against the palette the caller passed (ADR-0003 decision 6)
accent_tokenpalette entrythe NAME of a --sb-* property, stamped on that type's cards and palette rows
badgepalette entrya short chip for the card header
join_labelpalette entrya one-argument function of config, phrasing the join marker under a side-by-side arrangement
slot_outcome_keypalette entrynames the config key the blocks in one slot carry their outcome under, so a renderer routes an interrupt rule's escape without branching on a type name; it reaches the view model as Slot.outcome_key and the resolved value as Node.outcome
--sb-* tokensthe theme assign, or your own CSSevery colour, space, radius and drag treatment - see docs/theming.md
compile findingsfindings assignStatifierBlocks.Finding.from_compiler/2 adapts a compiler finding into the shape the editor renders, so a compile result routes back to the field somebody typed it into

The metadata readers are total and refuse rather than repair: a badge that is blank, carries a newline, or runs past 24 characters is dropped rather than clipped, an accent that is not an anchored --sb-* name never reaches a style attribute, and a join_label that raises degrades to the editor's own word. Assignability answers with reason-carrying refusals (sb-ue7, in flight).

Routing a compile pass into the findings pane is two calls:

{lint_findings, _refused} =
StatifierBlocks.Finding.from_compiler_all(compiled.warnings)

from_compiler_all/2 returns the findings it could anchor and, separately, the ones it refused with the reason - a finding that names no block has nowhere in the editor to land, and dropping it silently would be the wrong answer. Pass the anchored ones as the findings assign, or straight into StatifierBlocks.ViewModel.build/3 if you are driving the view model yourself.

Not yet

Honest about the edges, so you do not go looking for these:

Theming

Every class the package emits is prefixed sb-, and every color, space, radius and drag treatment is a --sb-* custom property with a default. Set them through the theme assign, or in your own CSS against the prefix:

<.live_component
module={StatifierBlocks.Editor}
id="editor"
theme={%{"--sb-accent" => "var(--brand-500)", "--sb-radius" => "10px"}}
...
/>

Enough that a host can make the editor look like its own product without forking it, and not so much that the package acquires a theming DSL.

docs/theming.md is the full guide: the three tiers the surface is organised into, why --sb-color-scheme is not optional, how a block type gets an identity of its own by naming a token, and a complete host theme you can copy. The rule it holds itself to is that a theme sets --sb-* properties and writes no other declaration - and that example is read out of the document and audited in the gate, so it is checked rather than promised.

What stays yours

Which palette entries a tenant may use, who may edit or publish a document, where it is stored, and what publishing means. The editor is also a single-session component: it surfaces the revision it loaded so you can do optimistic concurrency on save, and it does not merge or resolve anything.

Design records

The contracts this package is built out of are written down as ADRs in docs/adr/: the document schema (0001), the block-type behaviour (0002), host-pluggable assignability (0003), the compiler and its provenance map (0004), and the editor architecture (0005). A module's docs cite the decision it implements; when the two disagree, the record is the contract and the code is the bug.

License

MIT - see LICENSE.