All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Entries for unreleased work are not written here directly. Each issue drops a fragment in changelog.d/; the fragments are assembled into a version section at release. See that README for the format and for when a change warrants an entry at all.

0.30.0 2026-09-13

0.30.0 is a minor, because what a composite compiles to changes. A composite that declares outcomes now compiles to a state of its own with one <final> per declared name, and the enclosing body routes its completion through a derived on_<name> slot per declared outcome (ADR-0002 C6 and C7). Beside that, a core.on_event capture literal carrying a raw control character is now refused at compile.

Changed

  • A core.on_event capture literal carrying a raw control character other than tab, line feed or carriage return is now refused at compile, on the capture key, because the value reaches an XML attribute raw and XML 1.0 admits no such character there.
  • A composite that declares outcomes now compiles to a state of its own with one <final> per declared name, so an enclosing body can route its completion, and it derives one on_<name> slot per declared outcome for the blocks that run when it finishes that way.
  • A composite declaration whose slots: names a pass-through slot after one of its own declared outcomes is now refused, because the two kinds of slot cannot share a name; rename one of them.

0.29.0 2026-09-13

0.29.0 is a minor, and its one consumer-visible requirement comes first: the predicator floor moves from ~> 9.0 to ~> 9.4.1. That is the version whose string lexer reads a literal back whole, and it is what the release's one behaviour change needs - a core.on_event capture pair's literal source now accepts a string carrying any character, so the interim restriction that kept a literal inside printable ASCII is lifted and a handler may capture a label such as "Café inscrit".

Beside that, this release is about an editor that tells the truth about what it will accept. A literal capture pair draws a read-only row rather than going missing from the form that cannot author it, and editing a core.on_event block no longer drops one. A read-only mount withholds the card controls it refuses and names the arm behind an empty slot. And two composite-facing reports are corrected: a member the palette cannot resolve no longer lends the composite a raisable outcome, and a summary chip that is both translated and over the cap still carries its declared event name.

Added

  • A literal capture pair - ["const", value] - draws a read-only row in the editor's capture section, so a pair the two controls cannot author is visible rather than absent.

Changed

  • A core.on_event capture pair's literal source - ["const", value] - now accepts a string carrying any character. The interim restriction that a literal string stay inside printable ASCII, tab, newline and carriage return is lifted: it was earned only while predicator 9.4.0's string lexer wrote a literal's codepoints back one byte at a time, and 9.4.1 fixed that lexer. A handler may now capture a label such as "Café inscrit" and the datamodel reads it back whole.

  • Dependency floor: predicator moves from ~> 9.0 to ~> 9.4.1. That is the version whose string lexer reads a literal back whole, which is what the lifted restriction above needs; 9.4.0 and below are excluded.

Fixed

  • A read-only editor mount no longer draws the card controls it refuses: "Replace with its steps" and Delete are withheld, the way the palette column and the gap "+" already were.

  • A composite member whose block type the palette cannot resolve no longer contributes the default done outcome to the composite's raisable set, so a declared outcome name is no longer accepted on the strength of a missing palette entry.

  • A summary chip that is both translated and over the presentation cap now carries its declared event name on the chip's title, rather than the clipped chip's own full translated text, so the generated event name is still on the card for a reader debugging against the compiled chart.

  • Editing a core.on_event block no longer drops its literal capture pairs: a pair the form could not draw was previously replaced away by the next change the form posted.

  • A read-only editor mount's empty slot now names the arm it draws the placeholder for, so a screen reader hears which arm is empty instead of hearing the slot header and nothing.

0.28.0 2026-09-12

0.28.0 is a minor, and its two behaviour changes come first. The compiled chart of a document with a failure-classed final moves: the reserved <donedata> param the compiler mints on such a final is now named statifier_persistence:execution_status rather than statifier_persistence:run_status, so a durable host that reads that key needs statifier_persistence >= 0.12, which is the floor for the new name. And a core.on_event capture pair whose source is absent from the event payload no longer writes :undefined - the destination is left unwritten, so a document that captures by a path compiles to a different chart and a stored compiled document has to be recompiled. A document with no failure-classed final, and a handler that captures nothing or captures only literals, compile byte for byte to what they compiled to before.

Beside those, this release is about what a capture and a composite may say for themselves. A capture pair may take a literal source, written ["const", value], so a handler records a value the document states rather than one the event's payload has to carry. A composite may declare its own outcomes, checked at Resolve against what its expansion can raise. And the editor's profile assign takes a run? key that seats no run at all when it is false, rather than merely hiding the run pane.

Added

  • A composite declaration may carry outcomes, a list of outcome names it declares for itself, on both the use form and the data form. Present, the list replaces the expansion root's derived outcomes - each name labelled by the member that raises it - and the compiler checks at Resolve that the expansion can raise every declared name, reporting an :outcome_not_raisable finding against the composite block when it cannot. Absent, nothing changes: a document that writes no outcomes key compiles to the same bytes as before.

  • The editor's profile assign takes a run? key: run?: false mounts an editor that seats no run at all, so there is no run pane, no run marks on the canvas and no Held here column, whatever the host passes in run and run_session. It defaults to true, and it does not touch the marks a host paints itself through active_marks and invoke_mark.

  • A core.on_event capture pair may take a literal source, written ["const", value] in the block document, so a handler records a value the document states rather than one the firing event's payload has to carry; a string source is still the payload path it has always been, and the two are told apart by shape.

Changed

  • The compiled chart of a document with a failure-classed final changes: the reserved <donedata> param the compiler mints on such a final is now named statifier_persistence:execution_status instead of statifier_persistence:run_status. Upgrade a durable host to statifier_persistence >= 0.12, which is the floor for reading the new key; 0.12 reads both keys for one release and 0.13.0 reads only the new one. A document with no failure-classed final compiles byte for byte to what it compiled to before. Both names stay refused for a donedata_type/1 declaration while the old one is still read.

  • A core.on_event capture pair whose source is absent from the event payload no longer writes :undefined; the destination is left unwritten, so "not answered" and "answered with nothing" are different values and a reader tests a captured path by asking whether it is there. A document that captures by a path compiles to a different chart - each such pair's <assign> is now wrapped in an <if> that tests the path - so recompile stored documents; a handler that captures nothing, and a pair whose source is a literal, compile byte-identically. A reader that tested a captured path for :undefined should test it for presence instead.

0.27.0 2026-09-08

0.27.0 is a minor, and its two breaking changes come first. The compiled chart of every document holding an interruptible group moves: a group's two interrupt transitions and the <raise> of the interrupt pair inside its rail now carry that group's own emitted state id, so a stored compiled chart or a pinned chart identity for such a document must be re-captured. The spelling an author configures and the pair a host block type raises do not change - the salt is applied at emit, and a raise outside any rail is emitted unchanged. And StatifierBlocks.Composite.expand/2 answers a tuple, {:ok, {blocks, param_map}} | {:error, reason} instead of the bare {blocks, param_map} pair, and no longer raises; expand!/2 is the raising spelling the compiler's Resolve and the editor call, and a caller that wants the old return and the old raise adds one character.

Beside those, this release is about admission and the card face. Admission now resolves a composite's members through the palette it is asked about - Assignability.kinds/3, slot_accepts/4 and admits?/4 take the palette first - so a host-rooted composite is admitted on the kinds that host declares and a drop is refused exactly where a compile would refuse it. Beside the Edit.Targets.accepted_types/4 sweep there is now a per-target fast path, admits_at?/5 and accepted_types_at/5: one probe and one check at the slot's append gap, which is the call a "+" chooser at a gap should make. Assignability.assignable?/5 takes a strict: true that refuses a pair either side of which resolves to :unknown.

On the card, the summary-chip presentation cap is 32 characters rather than 24, and a chip past it is drawn clipped with an ellipsis and its full text on the chip's title instead of being dropped. A presentation-cap :lint finding is read in the drawer's Findings tab rather than on the card face, and a finding that does draw on a face is laid out inside the card's own box. The card's controls moved into a reserved strip beside the title, so a title that wrapped to two lines is never covered and nothing truncates; "Save as a step" and its marking tray are no longer drawn on a mount that registered no on_collapse. Two class contracts moved with those: .sb-node__chrome > .sb-node__strip > ... for the controls, and .sb-node__chrome > .sb-finding for a face finding.

For a host drawing its own surface, Editor.ConfigForm.config_form/1 takes an event attr defaulting to "config-change" and posts the block id as a hidden input, so one call replaces a hand-written field pair, and Edit.Session.refusal/1 says what a refused gesture refused in one sentence - the same sentence the editor now draws under its own toolbar.

Added

  • StatifierBlocks.Composite.expand!/2 is the raising spelling of the expansion: it keeps the return and the broken-declaration raises expand/2 had, and it is what the compiler's Resolve, the environment walk and this module's own derivations call.

  • A data composite draws the summary chips its use-composite twin draws: one chip per param the config gives a value to, minus those declared hidden?: true.

  • StatifierBlocks.Edit.Session.refusal/1 says what is in a session's last_error as one sentence, so a host driving the same commit funnel draws the same words for the same refusal.

  • StatifierBlocks.Edit.Targets.admits_at?/5 answers whether one block type would be accepted at one {parent_id, slot} target, with one probe and one StatifierBlocks.Assignability.check/5 at the slot's append gap instead of the whole-document walk accepted_types/4 runs per candidate - the call a "+" chooser at a gap should make.

  • StatifierBlocks.Edit.Targets.accepted_types_at/5 asks that question over a candidate list, defaulting to the palette's own types, so a surface with a shortlist pays for the shortlist. accepted_types/4 is unchanged and stays the sweep; the module's moduledoc says which of the three to call.

  • StatifierBlocks.BlockType.summary_refusal_message?/1 says whether a message is one summary_refusal_message/4 wrote, which is how a surface tells a presentation-cap diagnostic from every other :lint finding.

  • StatifierBlocks.Assignability.assignable?/5 takes an options keyword whose strict: true refuses a pair either side of which resolves to :unknown, for the caller that must not admit a value nothing has typed. The host relation is not asked about an unknown side.

  • The option defaults to strict: false, so assignable?/3 and assignable?/4 decide exactly what they decided before.

  • StatifierBlocks.Assignability.kinds/3, slot_accepts/4 and admits?/4 take a palette as their first argument and resolve a composite's members through it, so a composite whose expansion is rooted at a host type is admitted on the kinds that host declares rather than on the [:step] fallback the core-only kinds/2, slot_accepts/3 and admits?/3 reach. The StatifierBlocks.BlockType.io/1 callback is unchanged and gains no palette argument.

  • StatifierBlocks.Editor.ConfigForm.config_form/1 takes an event attr, defaulting to "config-change", which is written to both phx-change and phx-submit, so a host can draw one block's fields under its own handle_event/3 instead of hand-writing the field pair. The block the params are about still arrives as the hidden block-id input the form already posted.

Changed

  • The card's controls moved into a .sb-node__strip element inside .sb-node__chrome. A host stylesheet or test selecting .sb-node__chrome > .sb-node__remove, > .sb-node__expand, > .sb-node__save-step, > .sb-node__fold or > .sb-node__offer should select .sb-node__chrome > .sb-node__strip > ... instead. The classes, the data-reveal contract and the events are unchanged, and the strip is drawn on every card. Within it the controls are in left-to-right order, so a keyboard now reaches Save before Expand.

  • StatifierBlocks.Editor.Canvas.canvas/1, .Slot.slot/1 and .BlockNode.block_node/1 take a collapsible attr, threaded the way expandable is; it is false by default and the editor passes whether its on_collapse assign is a one-arity function.

  • Breaking. StatifierBlocks.Composite.expand/2 now answers {:ok, {blocks, param_map}} | {:error, reason} instead of the bare {blocks, param_map} pair, and it no longer raises on a broken declaration. A caller that wants the old return and the old raise calls expand!/2 and needs no other change; a caller that would rather be told than raised at matches on the tuple.

  • A read-check :type_mismatch whose disagreeing writer is a block a composite minted for its own expansion now names the composite the author placed, by that composite's sentence, instead of naming an id the author cannot see, select or edit. A writer the author did place - including a block dropped into a composite's pass-through slot - is still named by its own id.

  • The finding's reason tuple is unchanged: it still carries the minted id, so a fixture run and the Source tab still say which member of the expansion wrote the entry.

  • StatifierBlocks.Composite.expand/2 blames a member carrying more than one param's value on the first param in declaration order, instead of answering nil for it; a finding re-anchored onto the composite now names a field to open. A member carrying no param value still answers nil.

  • The summary-chip presentation cap is 32 characters, up from 24. Every message that quotes the cap follows the number; --sb-card-width is unchanged, because the cap is a legibility number rather than a measurement of the card.

  • A summary chip past the cap is now drawn clipped, with an ellipsis in its last position and its full text on the chip's title, instead of being dropped. A blank, multiline or non-string chip is still refused. The :lint finding that reports the length stays, and its sentence now ends "so it is drawn clipped".

  • StatifierBlocks.BlockType.badge/1 is unaffected: an over-long badge is still dropped rather than clipped.

  • A presentation-cap :lint finding no longer draws on the card face; it is read in the drawer's Findings tab, and still counts toward a card's findings rollup. Every other finding a card carries draws where it did.

  • A finding that does draw on a card face is laid out inside the card's own box. On a container it previously drew full-width between the card and the slot label below it, where it read as belonging to the slot. A host stylesheet targeting .sb-node > .sb-finding should target .sb-node__chrome > .sb-finding instead.

  • The compiled chart of every document holding an interruptible group changes: a group's two interrupt transitions, and the <raise> of the interrupt pair inside its rail, now carry that group's own emitted state id (statifier_blocks.interrupt.resume.<group state id>), so a stored compiled chart or a pinned chart identity for such a document must be re-captured. The spelling an author configures and the pair a host block type raises do not change: the salt is applied at emit, and a raise outside any rail is emitted unchanged. A railed composite can now be nested inside a resumable group's body without the two rails contending for one event name.

  • StatifierBlocks.ViewModel.overlay_findings/2 enforces the findings shape validate_config/1 declares: a block type answering with anything but a list of {key, message} string pairs is refused with an error naming that type, rather than failing a clause head that named the view model.

  • Kind admission inside check/5, valid_targets/4 and validate/3, and the editor's expansion check, now route through the palette-carrying arities, so a drop is refused exactly where a compile would be refused. A host-rooted composite that was wrongly admitted at a step slot, or wrongly refused at an interrupts slot, now gets the verdict the compiler gives it.

  • config_form/1's target attr is optional and defaults to nil, which renders no phx-target anywhere - the case of a host whose form posts to the LiveView it is mounted in. Every present caller passes a target and is unchanged.

Fixed

  • The assigns map an :expression field hands a host's expression_component carries a debounce key - the field's own debounce value, and nil when the caller named none - so an override can rate-limit the way every other control in the same form does instead of posting per keystroke. Overrides that ignore the key are unchanged.

  • An :expression field served through the expression_component seam is handed the values the host offered for that field ahead of the document's declared paths - one de-duplicated candidates list - instead of the declared paths alone. A host that named values for the field now sees them in its own control, the way the plain input this package renders itself has always shown them; a host that named none sees the declared paths exactly as before.

  • The "Save as a step" control and its marking tray are no longer drawn on a mount that registered no on_collapse callback, and the four events the gesture is made of are answered with the socket unchanged there. The gesture's only outcome is that callback, so a mount without one offered a marking step, a Save, and then silence. "Replace with its steps" is unaffected in both directions: it commits an edit the editor makes itself.

  • The card's controls no longer draw on top of a title that wrapped to two lines. They sit in a reserved control strip beside the title - a grid column of its own, held whether or not the controls are revealed - so a long name wraps beside them and nothing truncates.

  • The edit algebra's config gate refuses a {:update_config, id, config} whose value for a {:type_expr, opts} field is no arm that field admits, so the editor reports it where the edit is made rather than only at compile.

  • An empty slot on a read-only editor mount now draws a non-interactive placeholder. The mark that says "nothing is here yet" was styled on the gap's "+" button, which a read-only mount does not draw, so an empty arm rendered as a slot header and nothing else.

  • A read-only editor mount now refuses the Expand gesture. "Replace with its steps" reached the document and fired on_change on a mount whose profile said read_only?: true, which every other gesture that changes the document already refused.

  • A gesture the editor refuses now draws the refusal as one sentence in the canvas column, under the toolbar. Every refusal already recorded its reason and nothing read it, so a gesture the editor refused looked on screen exactly like a gesture that did nothing.

  • A composite's expansion is compiled at each member type's current version, so a config value a subtree writes reaches the chart as written rather than through that type's migrate_config/2. A subtree naming core.send or core.wait - the two core types past version 1 - used to have its delay or duration read as a stored value at version 1 and silently rewritten; such a value is now the author's to write in the accepted spelling, and one that is not is refused by name. A block the document stores is untouched: it still migrates.

  • StatifierBlocks.Composite.Collapse.replacement/4 gives the composite it inserts the collapsed arrangement's own block id instead of a freshly minted one, so the state ids of the chart a host gets back after committing the swap are the ones the document implies rather than a new UXID each time.

  • use StatifierBlocks.Composite refuses an option it does not recognize, naming the key, instead of dropping it and compiling the composite with the default the declaration was trying to replace.

0.26.0 2026-09-07

0.26.0 is about what a host does around a composite. StatifierBlocks.Composite.Collapse reads a selected arrangement back as the Composite.Data declaration that stands for it, and the editor's "Save as a step" control hands that declaration to a new on_collapse callback - the gesture edits no document and this package persists nothing, because which table a saved step lives in, and whether it is saved at all, is the host's. A composite may now declare pass-through slots, so an author puts children into one instead of only filling in params, and expand/2 splices them into the mapped member's slot with their ids unchanged. A data declaration may carry a "migrations" chain, so a block saved against an older revision of a saved composite resolves instead of being refused. And all nine field types now have a JSON spelling, so a declaration held as data reaches the same field-type set a module composite declares.

Beside that, this release publishes what a host drawing its own surface over a document was otherwise writing privately: Composite.io/2 and outcomes/2 resolved through a palette you supply, a derived summary/1, Edit.Targets.accepted_recipes/4 and recipe_inserts/4 beside the type half, Environment.with_writes/4, ViewModel.transparent?/2, effective_parent/3, end_of_list_target/3 and core_containers/0, ViewModel.overlay_findings/2 over the session's own draft_findings, ViewModel.order_palette_groups/2, a variant: :inline on Editor.Field.field/1, a debounce assign on the editor mount, and use StatifierBlocks.InvokeStep, failure_outcomes: [...].

It is a minor, and no compiled chart bytes move: a document holding a composite with a pass-through slot compiles to the bytes of the same document with that composite expanded in place, exactly as one with no slot does. Two things answer differently from 0.25.0 for a host that does nothing: StatifierBlocks.Palette.new/2 now runs the duplicate-order refusal that from_modules/2 already ran, so a palette built by merging onto core_types/0 raises where it used to build; and a composite's slots/1 answers its declared pass-through slots rather than always [].

Added

  • StatifierBlocks.Environment.with_writes/5 is public: env with one block's own writes applied, member expansion included. The declarations argument carries a default, so with_writes/4 is the spelling a caller writes.

  • StatifierBlocks.Editor.Field.field/1 takes a variant attribute (:block, the default and byte-identical to what it rendered before, or :inline), so a host placing one field inside a sentence of its own gets the same control, the same posted params and a label that is off the screen but still announced, without the editor gaining a layout mode.

  • core.on_event answers sentence/1 - "When card.authz_timed_out, abandon" - so a host list view draws the handler's own line instead of the palette label "On event".

  • StatifierBlocks.Edit.Targets.accepted_recipes/4 answers which of a palette's recipe names would land at a target, the recipe half of accepted_types/4, so a host drawing its own insert picker offers recipes beside block types.

  • StatifierBlocks.Edit.Targets.recipe_inserts/4 answers the command list that inserts a named recipe's arrangement at a target, or the refusal, without committing anything.

  • ViewModel.transparent?/2, effective_parent/3 and end_of_list_target/3 answer where a row sits and where an append lands once a host has flattened its transparent containers, taking the transparent type names as an argument.

  • ViewModel.core_containers/0 names the three core containers a host most often flattens, as the documented default for those readers rather than a built-in policy.

  • StatifierBlocks.Edit.Session carries a draft_findings map: the per-field findings each refused change_config/3 was handed, keyed by block id.

  • StatifierBlocks.ViewModel.overlay_findings/2 routes those findings onto a node's form fields, with a finding naming no field landing in form.unrouted - the findings half of overlay_draft/2.

  • StatifierBlocks.Composite.io/2 and outcomes/2 answer a composite block's io and outcomes with every member resolved through a palette you supply, so a composite rooted at a host block type is exact wherever a palette is in hand.

  • A composite derives summary/1: one card chip per declared param that is not hidden?: true, drawn as "<label>: <value>" and overridable.

  • StatifierBlocks.BlockType.outcome_names/1 answers the names in an outcome list a caller already holds.

  • use StatifierBlocks.InvokeStep, failure_outcomes: [...] declares a step family's failure class once at the use site, so a host whose steps all fail the same way no longer writes failure_outcomes/1 on every member.

  • ViewModel.order_palette_groups/2 puts palette groups in a reading order a caller names, keeping the groups it did not name after them by name.

  • StatifierBlocks.Editor takes a debounce assign and passes it to the inspector's config form, so a host that mounts the editor can say how often its controls post without composing ConfigForm.config_form/1 itself.

  • A StatifierBlocks.Composite.Data declaration may carry a "migrations" list: ordered rename / drop / default steps, each keyed by the type_version it migrates from. migrate_config/3 walks every step at or above the stored version and below the declaration's "version" in one call, so a block saved against an older revision of a saved composite resolves instead of being refused.

  • The whole chain is validated at declaration/1, entry-build time: a step of the wrong shape, a step carrying none of the three parts, a key named by both "drop" and "default", a duplicate, out-of-order or gapped "from", a chain not ending at "version" - 1, and a step naming a key the declaration's own params do not account for are each refused before the entry reaches a palette.

  • A composite may declare pass-through slots: use StatifierBlocks.Composite, slots: [%{name: ..., to: {local_id, inner_slot}}] on a module composite and a declaration-level "slots" key on a Composite.Data row. The composite's card draws an interior for each declared slot, and expand/2 splices the children the author put there into the mapped slot of the mapped member, with their ids unchanged.

  • StatifierBlocks.Composite.pass_through/2 answers a composite block's declared slots resolved to the minted id of the member each maps into.

  • StatifierBlocks.Composite.mapping_errors/2 answers the three refusals a pass-through mapping can earn against its own subtree: an unknown local id, an inner slot the subtree does not write, and a mapped inner slot the subtree also fills. A module composite raises them at its first expansion; a data composite answers them from declaration/1.

  • StatifierBlocks.Composite.Collapse.propose/3 reads a selected arrangement back as the Composite.Data declaration that stands for it, without naming it - naming the type stays the host's act.

  • StatifierBlocks.Composite.Collapse.replacement/4 answers the compound that swaps an arrangement for a composite of the name a host registered; the host commits it, and nothing in the package does.

  • The editor's "Save as a step" control on the selected card, with a marking tray for the values the saved step should ask for, handing the proposal to a new on_collapse callback. The gesture edits no document and the package persists nothing.

  • All nine field types now have a JSON spelling in a Composite.Data declaration: select, path, list and type_expr are the type's name plus an optional "options" key.

Changed

  • A surface drawing a refused config draft reads the findings from the session instead of calling validate_config/1 again to re-derive them.

  • StatifierBlocks.Assignability.produces/4 and the editor's core.on_event event candidates read a composite's io and outcomes through the palette they already hold rather than through the core-only fallback. The StatifierBlocks.BlockType.io/1 and outcomes/1 callbacks are unchanged and keep that fallback.

  • StatifierBlocks.Palette.new/2 refuses two entries of one palette-browser group that declare the same order, raising the ArgumentError naming both entries that from_modules/2 already raised, so a palette built by merging onto core_types/0 meets the same refusal; the fix is to renumber one of the two entries, and the moduledoc's "Ordering a group" states the convention.

  • Both builders now admit one pair at a single order: a composite's types entry beside that same composite's own derived <Module>.Recipe, whose palette_entry/0 is the block type's. Registering both stays the host's choice to show two entries; from_modules/2 used to refuse it.

  • The palette column draws its groups in the order a profile's palette_groups list gives them, rather than by name; a mount that names no list draws them by name as before.

  • A block stored below a declaration's earliest migration step still answers {:error, {:no_migration_from, from}}, and so does every stored version of a declaration that writes no "migrations" key - the previous behaviour, unchanged. A module composite is untouched: use StatifierBlocks.Composite gains no migrations: option and still writes migrate_config/2 itself.

  • A composite's slots/1 answers its declared pass-through slots instead of always [], and io/1's slot_accepts answers each declared slot at the mapped inner slot's own accepted kinds.

  • The environment walk descends a declared pass-through slot's children at the mapped inner position - the environment they would have seen had the expansion been placed by hand - and reaches nothing through a slot key the declaration does not declare. A read that fails inside such a child is reported against the child, not against the composite.

  • The editor's config form draws StatifierBlocks.ViewModel.shown_fields/1's list, so a host surface drawing its own form reads the same filter the package's own form reads rather than a second copy of it.

Fixed

  • The editor's drop-check preview no longer under-reports a member mismatch the drop would introduce. It applied a candidate's writes without the member expansion the walk runs, so a read of record.member was answered with an advisory before the drop and an error the moment it landed. The preview and the walk are now one codepath, and they agree.

  • A read-only editor mount no longer draws the "+" button on the gaps between blocks; the gaps themselves stay, so an empty arm still reads as one.

  • The editor's Expand gesture refuses a composite whose declaration is too broken to expand - an empty or non-block subtree/1, a duplicated, blk_-prefixed or __-minting local id, or a pass-through slot mapped at something the subtree does not hold - naming the declaration error, instead of raising out of the author's LiveView. The compiler already answered the same case with a :composite_expansion_failed finding.

0.25.0 2026-09-07

0.25.0 is about composites: a block type derived from params plus a pure subtree, so an arrangement a host writes over and over becomes one thing an author puts down and one thing they read back. use StatifierBlocks.Composite declares one from a module, and StatifierBlocks.Composite.Data derives one from a JSON-shaped declaration, so a host whose users save block types can register a block type held as data - which is what a {module, state} palette entry is for. The compiler replaces a composite with its expansion at the Resolve stage, so a document holding one compiles to bytes identical to the same document with that composite expanded in place, and the editor's "Replace with its steps" performs exactly that expansion as one compound edit a single undo puts back.

Beside that, this release publishes what a host drawing its own surface over a document was otherwise writing privately: StatifierBlocks.ViewModel's readers, StatifierBlocks.Edit.Session as a commit funnel with no socket in it, StatifierBlocks.Edit.Targets.accepted_types/4, Recipe.members/2 with a delete path that offers the whole arrangement, a selected_id a host may write, field_candidates honoured on path and expression fields, and a debounce attr on the config form.

It is a minor, and no compiled chart bytes move. Two things answer differently from 0.24.0 for a host that does nothing: StatifierBlocks.Palette.from_modules/2 now refuses a palette whose browser group holds two entries declaring the same order, which is a mount-time raise for a host that carried a duplicate and built at 0.24.0; and StatifierBlocks.Palette.manifest/1 now answers nil as a type's version where it used to raise.

Added

  • StatifierBlocks.ViewModel answers the readers a host writing its own surface over a document would otherwise write privately: find_node/2, parent_of/2, positions/1, sentence/1, shown_fields/1, fields_for/2, overlay_draft/2 and drafted_field/2.

  • StatifierBlocks.Document.committed_config/2 and effective_config/2 and /3 answer what the document holds for a block, and what an unaccepted draft says instead.

  • The editor's field-list-add and field-list-remove events carry block-id beside key, so a host handling them knows which block the gesture came from.

  • A palette entry may now be {module, state} beside a bare module, so a host whose users save block types can register a declaration held as data. StatifierBlocks.Palette.call/4 is the one seam every callback on an entry goes through: it prepends the state, asks about the arity a stateful module actually exports, and answers a caller-supplied default when the entry declares nothing. StatifierBlocks.Palette.declares?/3 answers declaredness alone for the two presentation branches that need it rather than a value.

  • StatifierBlocks.Composite.Data derives a composite block type from a JSON-shaped declaration - params, a subtree template with one "$param" placeholder arm and one "$literal" escape, and an optional sentence and palette entry - through the same StatifierBlocks.Composite.expand/2 a use StatifierBlocks.Composite module goes through. A data composite and the module composite of the same shape expand to the same blocks and compile to the same bytes. Composite.Data.declaration/1 refuses a malformed declaration before it reaches a palette, which is the last moment a callback can still be pure and total.

  • StatifierBlocks.Recipe gains an optional members/2 callback: asked about a block, a recipe answers the ids of the arrangement that block belongs to, or []. A recipe that does not implement it is unchanged.

  • The core "deadline" recipe implements members/2, recognising its core.send and core.on_event pair structurally - a hand-built pair is claimed, a pair renamed apart is not, and no block outside the enclosing group is ever named.

  • selected_id is a documented input assign. A host that draws a selection surface of its own - an outline pane, a plan view - moves the editor's selection by passing the id through send_update/3, and hears the result back on on_select like any other selection. It is honoured only on an update that carries it, so a re-render the host made for a reason of its own leaves the author's selection where it was, and an id the open document does not hold clears the selection rather than naming a block that is not there. A selection is not a document edit: no command, nothing serialized, and nothing on the undo stack.

  • The editor replaces a selected composite block with the blocks it stands for. The control sits on the composite's own card, reads "Replace with its steps", and commits the removal and the expansion's inserts as one compound edit: one undo puts the composite back whole, and the document it writes is the one StatifierBlocks.Composite.expand/2 answers, so the chart compiles to the same bytes on both sides of the gesture.

  • The gesture is refused, and nothing is written, when the slot the composite sits in will not admit the blocks that come out of it.

  • StatifierBlocks.Edit.Session is the commit funnel as a value: commit/2, change_config/3, step/2, update_list/4 and apply_gesture/2 over a document, a history and the drafts, with no socket in them. Each answers {:ok, session} or {:error, session}, and the tag says whether the document moved - which is when a host stores or notifies. Holding a refused config as that block's draft, and dropping every draft across an undo, are the package's decisions rather than each surface's.

  • StatifierBlocks.Edit.Targets.accepted_types/4 answers which of a palette's block types fit a {parent_id, slot} target, and probe/2 builds the block it asks about: Palette.new_block/2's, with the type's palette_entry/0 default_config merged over it. A surface filtering a palette without that merge answers differently, for any type whose read depends on its config, than the canvas's own drag stamp does.

  • StatifierBlocks.Assignability.context/1 builds the assignability context from a host's :datamodel, so the drop check, the environment walk and a datamodel view are handed the same one.

  • ConfigForm.config_form/1 and Field.field/1 take a debounce attr, written as phx-debounce onto every control the form draws, so a host that persists what the form posts can ask for something slower than one write per keystroke. It takes what LiveView takes - milliseconds, or :blur - and defaults to no attribute, which is what every existing caller already renders. Controls drawn by an expression_component override are that component's own and are not covered.

  • A :composite_expansion_failed finding at the :resolve stage reports a composite whose declaration cannot be expanded, in place of the raise.

  • A host's field_candidates list is honoured on {:path, opts} and :expression fields as well as :string ones: either spelling draws a <datalist> the input is bound to, ahead of the declared datamodel paths, and the value stays typed by the control.

  • use StatifierBlocks.Composite declares a block type from params plus a pure subtree/1, deriving config_schema/1, slots/1, io/1, outcomes/1, current_version/0, sentence/1, palette_entry/0 and a raising emit/2, and leaving only sentence/1, palette_entry/0 and validate_config/1 overridable.

  • StatifierBlocks.Composite.expand/2 answers the blocks a composite stands for, with ids minted deterministically from the composite block's own id, together with the param each expanded block is blamed on.

  • A composite's declaration derives a StatifierBlocks.Recipe at <Module>.Recipe whose insert/2 puts down one composite block.

  • StatifierBlocks.Environment.read_signatures/3 and write_signatures/3 answer a composite with the union of its expansion's reads and writes, taken at the composite's one position in the document.

Changed

  • StatifierBlocks.Palette.fetch/2 and resolve/2 answer the entry as stored: neither normalizes a bare module into a pair nor unwraps a pair into its module, so a host that registered no stateful entry can be handed none and its existing {:ok, module} matches still match. manifest/1, new_block/2 and from_modules/2 read a stateful entry through the seam, including from_modules/2's duplicate-order check - a stateful entry collides with a bare one at the same order and is refused, rather than silently dropping out of the check.

  • StatifierBlocks.Composite.expand/2 and composite?/1 take a palette entry where they took a module. The existing module-only calls are unchanged.

  • Deleting a block a recipe claims now offers to remove the whole arrangement: the card's delete control counts what the next click removes and a Keep sits beside it. Taking the offer commits one compound, so one undo puts the arrangement back whole. Deleting a block no recipe claims is unchanged.

  • The compiler replaces a composite block with its expansion at the Resolve stage, so a document holding a composite compiles to bytes identical to the same document with that composite expanded in place.

  • A finding raised inside an expansion is reported against the composite block, carrying the key of the param that produced it or no key at all, and the provenance map still owns every span by the expanded block that emitted it.

  • nil is refused as the default: of a hidden?: true field for every field type, not only {:type_expr, opts}.

  • Breaking for a palette with a duplicate order. StatifierBlocks.Palette.from_modules/2 now raises ArgumentError when two entries of one palette-browser group declare the same order, naming both. A host whose palette carried such a pair built at 0.24.0 and does not build here: renumber one of the two. The check skips an entry whose module is not loaded, exports no palette_entry/0, or declares no order.

  • StatifierBlocks.Palette.manifest/1 answers nil as the version of a type whose module cannot be loaded or exports no current_version/0, where it used to raise, because it now reads every entry through StatifierBlocks.Palette.call/4.

Fixed

  • core.invoke and the core deadline recipe no longer both declare order: 7, so the Structure group of a palette browser has a stable order; Palette.from_modules/2 refuses two entries of one group that declare the same order, naming both.

Note

  • StatifierBlocks.Composite.Data fixes no migration key, so bumping a declaration's "version" with stored blocks refuses them through ADR-0007's injected migrate_config/2, unchanged and deliberately not papered over.

0.24.0 2026-09-07

0.24.0 is about drawing less of the editor, and saying more about what is drawn. StatifierBlocks.Editor takes an optional profile assign naming which drawer tabs, inspector tabs, palette groups and toolbar chips a mount draws, with read_only?: true rendering a document that offers no way to change it. A config field declaration gains hidden? and readonly?, so a type can hold a value no form shows, or draw one as text beside its label. Beside that: block types may declare sentence/1, one line of prose for a config, which StatifierBlocks.ViewModel.Node carries beside the author's title; StatifierBlocks.ViewModel.outline/1 answers the whole document in reading order as {node, depth, kind}; StatifierBlocks.Palette.new_block/2 builds a block from a palette entry, so a host view no longer reimplements what the editor's insert does; and the Datamodel tab draws the values each declared path enumerates.

It is a minor, and no compiled chart bytes move. Two changes can refuse a document or a block type that was accepted at 0.23.0: a config field declared without a default: key is now refused at compile whatever its field type, where before only a {:path, opts} field was; and the typed environment now puts an entry at every member beneath a record, shape or inline-shape write, so a nested read that previously met the nothing-is-known advisory is now checked against the written type.

Added

  • A config field declaration may carry hidden?: true, which keeps the field out of every form while the compiler, the Source tab and validate_config/1 see its value entirely unchanged.

  • A config field declaration may carry readonly?: true, which renders the field as its value beside its label rather than as an input.

  • StatifierBlocks.ViewModel.Field carries both flags, so a host drawing its own surface filters on the same two booleans the package's own config form reads.

  • StatifierBlocks.Editor takes an optional profile assign naming which of its surfaces a mount draws: %{drawer_tabs:, inspector_tabs:, palette_groups:, toolbar:, read_only?:}, every key optional and every list either :all or the ids it names. A mount that passes no profile renders exactly what it rendered before, and there is no arrangement of the map, %{} included, that removes a surface a host did not name.

  • An id a profile lists that the package cannot resolve is dropped and the mount renders; a key whose value is neither a list nor :all resolves to that key's default. There is no validate_profile/1 and a profile is never checked against the shell's ids at declaration.

  • profile: %{read_only?: true} renders the document without offering any way to change it: no palette column, no drag hook on the canvas, config fields and declaration rows drawn as values rather than controls, Undo and Redo hidden rather than disabled, and every gesture that would reach the document answered with the socket unchanged, so on_change never fires. Selection, on_select and every findings surface are unchanged, and a document is never refused for being read-only.

  • StatifierBlocks.Shell.drawer_tabs/1 and StatifierBlocks.Shell.inspector_tabs/1 answer the package's tabs a profile leaves, in the shell's own order; StatifierBlocks.Shell.drawer_view/1 takes an optional :profile key and filters the host's contributed tabs by the same list.

  • docs/profiles.md is the host-facing guide: the default, the ids each list draws from, the drop rule, a minimal mount and a read-only one.

  • The editor's Datamodel tab draws a Values column beside each declared path: the one_of enumeration the ADR-0006 document declares there, cut at eight values with the remainder counted ("+3 more"). It reports the declaration and not a host's value_candidates override, so a reader can answer "where did this picklist come from" from the table rather than by opening a condition. A path that declares no enumeration draws an empty cell.

  • A write signature whose type is a record, a shape, or an inline shape now puts an entry in the environment at every member beneath the path as well as at the path itself, recursively and to any depth, so a later block reading a nested path is checked rather than given the nothing-is-known advisory.

  • Block types may declare an optional sentence/1 callback, which answers the block as one line of prose for a given config. It is read through StatifierBlocks.BlockType.sentence/2, which is total: a type that declares none, and one whose callback raises, throws, exits or answers a non-string, a blank string or a multiline one, all read as the type's label. Unlike a chip it carries no length cap.

  • use StatifierBlocks.BlockType injects an overridable sentence/1 answering the type's own palette label, so a type that overrides nothing is indistinguishable from one that declares no sentence/1 at all.

  • StatifierBlocks.ViewModel.Node carries sentence: the type's own sentence where it declares a usable one, else the author's title, else the type's label falling back to the type name. A block whose type the palette cannot resolve carries the type name, so a list view never draws a blank line.

  • StatifierBlocks.ViewModel.outline/1 returns the document in reading order: one {node, depth, kind} per block, pre-order, kind in :step | :arm | :rail | :tray. It is pure and reads only the view model. Every block appears exactly once - arms, rails, trays and the drafts shelf are kinds and positions, never omissions - and depth is block nesting depth, so a slot never consumes a level.

  • core.wait, core.branch, core.subchart, core.foreach, core.parallel, core.send and core.assign declare sentences of their own. Every other core.* type answers its label, which is what it answered before.

Cards are unchanged: sentence is not a chip, is never capped, and nothing new is drawn on a block's card or in the palette browser.

  • StatifierBlocks.Palette.new_block/2 builds a block of a named type from a palette - the type's config_schema/1 defaults as the config and its current_version/0 as the stored version - so a host view that inserts from a palette no longer has to reimplement what the editor's insert does.

Changed

  • A config field declared without a default: key is now refused at compile whatever its field type; before, only a {:path, opts} field was. A block type that omitted the key on another type must add it - the declaration was never well formed, and the value it produced was nil.

  • A hidden?: true field whose default: is its type's empty value is refused at compile: it can carry nothing and no form can ever give it a value. A hidden :boolean defaulting to false is legal, false being a decided value rather than an absence.

  • A form ignores any posted value for a hidden? or readonly? field, so a crafted payload cannot reach a key the form withheld.

  • A document may stop validating where a block reads a nested path the record written above it types differently; the fix is either the reading block's expects or the written record.

0.23.0 2026-09-06

0.23.0 makes a block type's typed surface something an author writes rather than only names. StatifierBlocks.BlockType's field-type set gains {:type_expr, opts}: a field holding the name of a type the datamodel document declares, an inline unnamed shape written member by member, or nothing at all, with one shared check behind it and an inspector control that draws both arms. core.map's collect_type and core.on_event's payload are the first two fields to take it, and the statifier_datamodel requirement moves to ~> 0.4, which is where a type expression admits an inline shape. Beside that: the typed environment now seeds the declared path types the datamodel document carries, so a read at a declared path meets the declared type rather than an advisory; a compile refusal carries the Config and Structure stages' findings together, so a mis-typed field on one card no longer hides an unsatisfied read on another; a root block type declares a typed child summary that a child_use compile emits as <final> params; drop-time typing fires from a palette entry's new default_config rather than from config_schema/1's defaults alone; a :type_mismatch finding carries the config_key of the field whose read declared the path; and StatifierBlocks.Palette.manifest/1 gives a host one assertion to pin its palette with.

It is a minor, and compiled bytes move for one kind of document: one compiled under child_use: true whose root block type declares a donedata_type/1, whose top-level <final> elements now carry one <param> per declared field after the two the compiler mints. Every other document compiles to the bytes it compiled to at 0.22.0, and each entry below that could have moved one says so in its own words. Two changes are worth reading before upgrading, neither of them a byte: a stored document that validated may refuse once its host supplies a datamodel, because the environment now seeds declared path types and a read that found nothing before now meets a declared one; and a host that compiles with terminate:, child_use:, known_invoke_types: or datamodel: must pass its own compile_options to the editor, or the Source tab, the fixture verdicts and the Run pane's marks are each about a chart nobody ran.

Added

  • A palette entry may declare default_config, the config the editor's insert probe is built with over config_schema/1's own defaults, so a block type whose read is declared on a path field says which path it will read before anyone has configured one.

  • StatifierBlocks.BlockType's closed field-type set gains a ninth member, {:type_expr, opts}: a field holding the name of a type the datamodel document declares, or an inline unnamed shape as a list of "name" / "type" / "required?" objects, or nothing at all. opts carries arms: (which of the two the field admits, both by default) and allow_empty?: (whether an empty value is admitted, true by default).

  • StatifierBlocks.BlockType.type_expr_findings/2 is the one shared check behind that field type, consulted by both the compiler's :config stage and the editor's view model: a value that is not an arm the declaration admits is a :config finding carrying the field's key, and an undeclared type name is not one.

  • The editor draws a {:type_expr, opts} field in the inspector's Config tab: a text input bound to a <datalist> of the document's declared type names for the name arm, an ordered member-list form for the inline arm whose member types are the same control recursing, and a toggle when the field admits both. Switching arms replaces the value rather than translating it, and a value the control cannot read is drawn raw with its finding beneath it.

  • StatifierBlocks.Environment's type expression admits an inline unnamed shape, {:shape, members}, and StatifierBlocks.Environment.inline_shape/1 reads a stored member list into one. The read check reaches it through StatifierDatamodel.Types.satisfies/3 unchanged - this package defines no second one.

  • StatifierBlocks.Environment's context accepts :skip_blocks, a set of block ids whose declared writes the walk leaves out. Absent, as it is for every editor query, nothing is skipped.

  • StatifierBlocks.Palette.manifest/1 returns a palette as a sorted list of {name, version} entries - a block type beside its current_version/0, a recipe beside :recipe - so a host pins what its palette carries in one assertion that names the entry which moved instead of a count that does not.

  • A worked example of the failure propagation 0.22.0 introduced, which the 0.22.0 Notes stated as a rule without showing a run. A core.sequence body holds a myapp:authorize invoke followed by a myapp:capture invoke, and authorize ends on its error outcome. With the authorize invoke's own on_error slot filled - handling is declared by the failing block, never by the sequence around it - the slot's child runs, the invoke ends on error, the sequence advances, and the run ends at the document's ordinary completion final with capture having run. With on_error empty and the document compiled under child_use: true or terminate: true, the invoke's error final is emitted anyway, the root's transition on done.outcome.<authorize state id>.error is selected before the sequence's own done.state, and the run ends at the shared top-level failed final carrying the reserved statifier_persistence:run_status param with the value failed - capture never runs. Compiled under neither option there is no root catch, nothing selects the outcome event, and the sequence advances to capture exactly as it does on success. ADR-0002's amendment of 2026-09-06, section 4 ("The nested-to-root propagation rule") and section 6 ("What this costs a host, and what it does not"), carry the full walk-through.

  • StatifierBlocks.BlockType declares an optional thirteenth callback, donedata_type/1: what a document's <donedata> carries when it is compiled for use as a child, as a list of %{name, path, type} fields typed in statifier_datamodel's vocabulary. StatifierBlocks.BlockType.donedata_type/2 is the resolver every consumer reads it through, and it is total - a type that does not export the callback, or exports one returning something else, declares nothing.

  • Compiling with child_use: true, each top-level <final> now carries one <param name="<name>" expr="<path>"/> per entry the root block type declares, in declaration order, after the outcome param and the reserved statifier_persistence:run_status param. A root type that declares nothing compiles to exactly the bytes it compiled to before, and a terminate compile emits no declared field at all.

  • A declared field name that collides with either of the two names the compiler mints, or that is not a bare lowercase identifier, is refused at compile with an :invalid_donedata_field Emit finding against the root block, rather than silently shadowing a param.

  • core.map gains an optional seventh config field, collect_type: the type name a collected answer's "donedata" carries, read through StatifierDatamodel.Types.parse/2 against the parent document's declarations. It is a type rather than a path, it produces no bytes, and it has no findings of its own. An absent or empty one means what every stored document means today.

  • StatifierBlocks.BlockType.agrees?/3 answers whether a child's declaration covers what a parent's collect_type expects of it, as StatifierDatamodel.Types' own read check and in that package's own reason vocabulary. The check is dormant: it is not a compile finding and changes no compiled byte, and it answers only where both documents are in hand.

Changed

  • The environment now seeds the declared path types the datamodel document carries: before the walk begins it holds an entry at every path the document declares, at the type it declares there, and a type a block writes replaces it from that position on. A read at a declared path that found nothing before, and was an :info advisory, now meets the declared type - so a stored document that validated may refuse once its host supplies a datamodel, and the fix is either the block's declaration or the document's. A caller that supplies no datamodel is unaffected in every particular.

  • A seeded entry's writer is :declaration, which StatifierBlocks.Assignability's :type_mismatch reports as its upstream ref. Such a refusal carries no {:fixable_by, block_id} reason - there is no block whose declaration an author would change - and its message says the datamodel document declares the type rather than naming a block.

  • core.map's collect writes {:list, <the ADR-0009 envelope>} rather than {:list, :unknown}: one collected element is a shape of index, status, an optional donedata typed by the block's own collect_type, and an optional failure. A block after a core.map learns this whether or not anything was declared, and no compiled byte moves.

  • The statifier_datamodel requirement rises to ~> 0.4, which is where a type expression admits an inline shape. Raising the floor alone also widens what a document projects: an entry whose type names a declaration contributes that declaration's fields as declared paths beneath it.

  • core.map's collect_type and core.on_event's payload are {:type_expr, opts} fields admitting both arms, rather than :string fields carrying a type name. A stored string is the name arm: the same bytes, the same resolution, the same findings, and no compiled byte moves.

  • core.map's collect_type may now be written inline - a list of "name" / "type" / "required?" objects - and the inline shape becomes the "donedata" member of the type collect writes, one level inside the collected element's envelope.

  • core.on_event's payload may now be written inline, and the refusal it buys reads that arm the same way: a capture pair whose source path names a member the inline payload does not carry is the same :config finding on the capture key, walking below the first segment through a member typed by a declared name or by a nested inline shape.

  • StatifierBlocks.BlockType.agrees?/3 reads an inline collect_type as the shape it writes rather than as nothing. The check stays dormant in every other respect.

  • core.on_event's validate_config/1 no longer reports a finding of its own for payload. What a value of a {:type_expr, opts} field may be is StatifierBlocks.BlockType.type_expr_findings/2's one check, which the compiler's :config stage and the editor both already consult, so bytes that are neither arm are refused once rather than twice on the same key.

  • compile_options is required of any host that compiles with terminate:, child_use:, known_invoke_types: or datamodel:. Leaving it unset is not an error and never was: the editor compiles a different chart than the host does, and a run against that chart is silently unmarked, the Source tab lists a chart nobody runs, and the fixture verdicts are about neither. Pass the host's own option list.

  • StatifierBlocks.SourceView.build/3 and StatifierBlocks.Runtime.FixtureRuns.run/4 forward their whole opts to the compiler rather than :declare alone, less the one key each keeps for itself (:previous and :view_model).

  • A compile refusal now carries the findings of the Config and Structure stages together, instead of the first of the two that failed. A mis-typed field on one card no longer hides an unsatisfied read on another, so an author sees both in one refusal rather than one per round trip. Refusal semantics are unchanged: such a document still does not compile, and every stage after Structure still stops the pipeline at its first failure.

  • A block whose config the Config stage refused is skipped by id in the Structure stage: it reports no structure finding of its own, and its declared writes leave no entry in the typed environment, because a write signature is read off the config that was refused. The walk continues past it - its siblings and its children are checked exactly as before.

  • The palette an armed "+" opens now filters recipes as well as block types. A recipe whose arrangement cannot land at the armed position - insert/2 refuses it there, or the commands it answers with reach outside ADR-0005 clause 3C's bound - is absent from the list rather than offered and then refused at the click, which wrote nothing and said nothing. The pick still runs both checks, so a stale pick is refused exactly as before.

Fixed

  • Dragging or picking a block type whose read is declared on a config field now greys the slots that would refuse it. The probe carried only the schema's defaults, and a path field defaulting to "" names no path, so such a type declared no read at insert time and every slot accepted it.

  • A block type that declares a datamodel-path field without a default: key is now refused at compile, as a :config finding naming the field, instead of raising a FunctionClauseError out of the view model when something rendered the block. The view model reads a declaration's default: permissively, so a field declared without one renders with no default rather than crashing the build.

  • The editor's Source tab and its fixture runs now compile with the host's compile_options, as the provenance recompile already did. All three compiles pass one option list, so the listing, the fixture verdicts and the Run pane's marks are all about the chart the host actually compiled.

  • A root block declaring an outcome named failed is now refused at compile with a :config finding naming the field the name was written in, rather than reaching Statifier as a duplicate state id reported against the package. The name is the one the shared final for an unhandled failure below the root already mints; rename the outcome. A block below the root is unaffected, and no other document's compiled bytes move.

  • The editor toolbar's Fit active now works over a run the editor seated itself: with nothing selected it fits and reveals the innermost block the run's marks name, where before it was enabled only by the host's own active_marks list.

  • A :type_mismatch compiler finding now carries the config_key of the field whose read signature declared the path, so an editor can put a refused read on the control the author has to change. Two path fields reading the same path were indistinguishable in a finding, which named the path only.

0.22.0 2026-09-06

0.22.0 makes failure a first-class outcome across the block vocabulary, and widens what a document can declare. core.invoke now declares the two outcomes it has always emitted, done and error, and classes error as a failure; every host type built on StatifierBlocks.InvokeStep classes its own error the same way by default, and opts out with failure_outcomes/1 returning []. core.invoke, core.map and core.subchart emit their error outcome's <final> whether or not the failure slot is occupied, and a failure-classed outcome a block below the document root leaves unhandled now reaches the document's own ending, through one shared top-level <final> carrying the reserved statifier_persistence:run_status donedata param. Beside that, core.on_event takes an optional payload declaration naming a type the datamodel document declares, and a capture pair reading a member that payload does not carry is refused at compile; core.map's collect accepts any datamodel path rather than only a bare identifier; and the editor takes a compile_options assign, so the recompile behind the Run pane's marks resolves against the chart the run is a run of.

It is a minor, and a minor with notes. Compiled bytes move for three kinds of document: one containing a core.invoke, core.map or core.subchart whose failure slot is empty; one with a failure-classed outcome left unhandled below its root, compiled under child_use: true or terminate: true; and one containing a host type built on InvokeStep. Each of those is a new chart revision. Every other document compiles to the bytes it compiled to at 0.21.0. ADR-0002's amendment of 2026-09-06, section 6, works through what this costs a host and what it does not. The statifier_datamodel floor also moves to ~> 0.3; a host on 0.1 or 0.2 updates it with the rest of the dependency tree.

Added

  • core.on_event takes an optional payload config key: the name of a type the datamodel document declares, saying what _event.data carries for the event that handler names. With one declared, a capture pair whose source path reads a member the payload does not carry is refused at compile - one :config finding on the capture key, naming the pairs and the payload - so the interpreter's unbound marker is never written for a captured path on a typed document. A handler with no payload, a payload naming a type the document does not declare, and a compile with no :datamodel are all unchanged: no new finding, and the same compiled bytes. payload itself emits nothing.

  • The editor takes a compile_options assign: the option list the host compiled the open document with, forwarded to the recompile behind the Run pane's marks so a run of a terminate: true chart resolves against the chart it is a run of.

  • A failure-classed outcome a block below the document root leaves unhandled now reaches the document's own ending: under child_use: true or terminate: true the compiler emits one shared top-level <final> carrying the reserved statifier_persistence:run_status donedata param, and one transition into it per unhandled pair.

  • core.invoke declares the two outcomes it has always emitted, done and error, so a parent may wire done.outcome.<state id>.error and the editor offers both on the outcome-event candidate list.

Changed

  • core.map's collect accepts any datamodel path, not only a bare lowercase identifier, so a fan-out can assemble its answers at cards.batch and not only at batch. All four fields this package writes an <assign location="..."> from now read one grammar and one refusal wording.

  • core.invoke classes its error outcome as a failure, and so does every host type built on use StatifierBlocks.InvokeStep. A host whose error is routine defines failure_outcomes/1 returning [] beside its outcomes/1 to keep the old classing.

  • core.invoke, core.map and core.subchart emit their error outcome's <final> whether or not the failure slot is occupied; with the slot empty the failure transition targets that final directly instead of being selected by nothing. A document containing one of the three with an empty failure slot, one with an unhandled failure below its root under the two compile options, or a host InvokeStep type, compiles to different bytes and so is a new chart revision; every other document is byte-identical.

  • The statifier_datamodel floor moves to ~> 0.3, so a datamodel entry whose type names a declaration contributes the declaration's fields as declared paths beneath its own: they are offered as expression candidates, they carry their declared types in the Datamodel tab, and a block that writes one gets no undeclared-path advisory, exactly as an inlined object entry's fields do. A host on statifier_datamodel 0.1 or 0.2 updates it with the rest of the dependency tree.

  • The Datamodel tab renders an entry typed by a declaration as the name it names, rather than raising on the {:declared, name} that release added.

0.21.0 2026-09-06

0.21.0 makes the editor a debugger and widens what a block can say. The drawer gains a Source tab - the compiled SCXML with every run of bytes traced back to the block that emitted it, and a click on a config-emitted span opening the Config tab on the field it came from - and the canvas takes a seat in a run pane, with statifier-ui's status, scrubber and event log around it and a send control that pushes a block type's fixture events straight into a live session. core.branch declares a third slot, undecided, for an arm whose condition cannot be decided (ADR-0012); a block type may class one of its outcomes as a failure, and the compiler stamps a reserved <donedata> param on that outcome's final so a durable stepper can tell a chart that finished badly from one that merely finished (ADR-0008's accepted amendment); a host can state a rule about a whole document through validate_document/1 and the palette's new :validators list; core.invoke's and StatifierBlocks.InvokeStep's assign_to take any datamodel path and declare the path they write; and core.map names what a child run sees its item and its position under, through item_as and index_as.

It is a minor, and nothing stored is invalidated by it: a core.branch that leaves undecided empty, and every type that classes no outcome as a failure, compile to the bytes they compiled to at 0.20.0. statifier_ui is now an optional dependency at ~> 0.9, the release that added the state-id reads StatifierBlocks.Runtime.Marks.from_trace/2 composes; statifier_datamodel stays at ~> 0.1.

Added

  • core.branch declares a third slot, undecided, labelled "Cannot be decided". A branch that puts children in it emits one extra transition, after every arm and before otherwise, taken when an arm's condition could not be decided - a comparison predicator answers with its undefined sentinel rather than true or false, such as a path missing under a bound datamodel root, or operands whose types do not match.

  • The drawer has a Source tab: the compiled SCXML as numbered lines, with the block that emitted each run of bytes, and a click on any run selecting that block.

  • StatifierBlocks.SourceView builds that listing from a document and a palette, without the editor.

  • The editor takes a run_session - a Statifier.Session.server() - and the Run pane draws a send control over it: one button per event in the selected block's type's fixture sample, sending straight into the session through statifier-ui's EventInjection.

  • The send control is enabled only over a live stream with a session supplied; over a persisted run, or with no session, every button renders disabled with a one-line note, and a send never writes to the document.

  • StatifierBlocks.Runtime.Marks.from_trace/2 turns a statifier-ui trace read model and a provenance map into the run marks the editor canvas already accepts, so a host observing a run marks the blocks a configuration is inside without naming them itself.

  • A block type may class one of the outcomes it declares as a failure through the new optional failure_outcomes/1 callback, and the compiler emits a reserved <donedata> param - statifier_persistence:run_status with the value failed - on that outcome's top-level <final> under both the :child_use and the :terminate compile options, so a durable stepper can tell that a chart finished badly rather than merely finished. core.map and core.subchart class their error outcome; every other type classes nothing.

  • A host can state a rule about a whole document, not just about one block's config: StatifierBlocks.DocumentValidator is the behaviour, validate_document/1 its one callback, and a module implementing it goes in the palette's new validators list (StatifierBlocks.Palette.new/2's :validators option, defaulting to []).

  • A validator says where and what - {anchor, message} or {anchor, message, severity: ...} on decision 11's existing anchors - and the package stamps the source :lint and defaults the severity to :warning. The findings render in the editor's Findings drawer tab and count toward StatifierBlocks.Editor.findings_count/3 like every other finding.

  • core.map declares item_as and index_as - the names a child run sees its item and its position under, defaulting to item and to no position name - and carries both into its <invoke> beside the list's path, so a child recipe reads what the author named rather than whatever the fan-out handler chose.

  • Clicking a config-emitted span in the Source tab opens the Config tab and focuses the field it came from; a span with no owning field still selects the block as before.

  • The editor takes a run - statifier-ui's StatifierUI.Live.State, live or persisted - and seats the canvas in a run pane: statifier-ui's status and scrubber above it, its event log below, and no Mermaid diagram. Scrubbing or clicking a log entry moves the canvas's run marks, and clicking a log entry also selects the block whose state handled that step.

  • While a run is seated the Datamodel drawer tab's "what is known here" table shows what the run held at each path, beside the type the position declares.

  • StatifierBlocks.Runtime.Handled.block/3 answers which block's state handled one macrostep of a run, and StatifierBlocks.Runtime.RunValues.at/1 what a run held at its selection; both are pure and neither needs statifier-ui to be present.

  • StatifierBlocks.Runtime.Selection.scrub/2 and select/2 move a run's selection, which is what the pane's two events do.

Changed

  • A core.branch that leaves undecided empty compiles to exactly the bytes it compiled to at 0.20.0, so every stored document is unaffected until its author wires the slot.

  • statifier_ui is now an optional dependency at ~> 0.9, the release that added the state-id reads from_trace/2 composes; hosts on 0.8 upgrade the package to use the new module and are otherwise unaffected.

  • A document whose root block is a core.map or a core.subchart, compiled with child_use: true or terminate: true, gains the reserved param on its error final. Its content hash changes with it, so it is a different chart revision under statifier-ex ADR-0052 - the same one-time choice opting into terminate already is. Every other document compiles to the bytes it compiled to.

  • A palette entry's singleton declaration is now derived through the same document-rule path a validator's findings take, and runs first among them. Its findings are unchanged: still :config, still :error, still anchored at the root.

  • A core.map compiled before this release gains an item_as param carrying the default name; a stored config that never had the key still validates and needs no migration.

  • core.invoke's assign_to is declared as a datamodel path field, so the editor offers the host's declared paths as candidates on it and a path it does not declare draws the usual advisory rather than a refusal.

  • core.invoke's and StatifierBlocks.InvokeStep's assign_to accept any datamodel path, dotted or not, in place of a bare lowercase identifier - the rule core.assign and core.subchart already accept for the same <assign> element. core.map's collect is unchanged.

Fixed

  • A path core.invoke writes is visible to the environment: it was emitted without being declared, so nothing downstream saw the write.

0.20.0 2026-09-06

0.20.0 is about types. The datamodel document's declarations become something the whole package reads: StatifierBlocks.Environment carries the path-to-type map at any position in a document with one pre-order walk, every core.* block type declares what it reads and writes at datamodel paths, and the data-flow check is defined over that environment rather than over the block immediately before this one. The path/type index itself moved out to the statifier_datamodel package, taken from Hex at ~> 0.1; the drawer's Datamodel tab answers what is known at the selected block and lists the declared records and shapes; and the condition editor is handed the declared kinds, so a clause on a path declared integer offers the numeric operators.

It is a minor with two breaking edges to read before upgrading. StatifierBlocks.Predicates.Datamodel is gone: its functions are StatifierDatamodel.Index's, under the same names, so a caller swaps the module name and nothing else. And a type expression spelled exactly unknown now reads as the permissive :unknown rather than as an opaque expression that happens to be spelled that way - it only ever admits, so nothing that passed before is refused now. The signature changes under Changed keep their arities; {:type_mismatch, ...} gains a sixth member naming the datamodel path the read was checked at.

Added

  • The kinds a datamodel document declares reach the condition editor. With a document supplied, StatifierBlocks.Datamodel.path_types/1 projects it to the expression language's own value kinds and the editor hands that map to statifier-ui's expression editor, so a clause on a path declared integer offers the numeric operators and one on a date path the date set, rather than whichever set its current source happens to imply. It draws a control and decides nothing: the operator the source carries is still offered, the value in it is still kept, a disagreement renders as an advisory beside the clause, and a path the document does not declare renders exactly as it did. A host that supplies its own expression_component is handed the same map.

  • A :type_mismatch finding names a declared type by its label. A pair of names the datamodel document's types key declares reads as the two human names an author recognises; a pair of opaque spellings a host carries reads exactly as it did before there were declarations. The rule is one function, StatifierBlocks.Environment.type_label/2, so the finding and the editor cannot disagree about what a type is called.

  • The drawer's Datamodel tab answers what is known here: the paths the environment holds at the selected block's position, with their types, as computed by the pre-order walk at that exact position - before the block's own writes land. Nothing selected and nothing known are two different states and the panel says which it is.

  • The same tab lists the declared records and shapes, each with its ordered fields, their types and their required marks, through StatifierBlocks.Datamodel.declared_types/1. An author told that a record does not cover a shape can now read what that shape requires without leaving the editor.

  • Write signatures across the core.* vocabulary, so the environment survives one. core.assign writes its path and core.subchart its assign_to as known-but-untyped; core.map writes collect as {:list, :unknown}, so the block after a fan-out knows it is looking at a list; core.on_event writes one path per capture pair on the interrupt path; core.wait, core.send, core.raise and core.await write nothing and leave the environment exactly as it reached them; and a container hands its children's writes out through the per-path merge.

  • StatifierBlocks.Environment - the datamodel path to type map at any position in a document, carried by one pre-order walk. at/3 answers it, subject_path/2 names the path a document's subject lives at, read_signatures/3 and write_signatures/3 say what a block declares there, and type_of/2 and satisfies/3 read a declaration's spelling against the datamodel document's own type declarations.

  • palette_entry/0 gains an optional subject key: the datamodel path a document's subject lives at, read from the document's entry block.

  • StatifierBlocks.Assignability.context/0 gains an optional :datamodel key, and StatifierBlocks.Compiler.compile/3 threads its existing :datamodel option into it, so the read check can consult the document's record and shape declarations.

  • The reason vocabulary gains {:shape_not_satisfied, missing}: the environment holds a record at the path, the read expects a shape, and missing names the required fields the record does not cover.

  • A {:path, opts} field declares what it reads and writes at its path: expects: T is a read signature the environment at the block's position must satisfy, and writes: T puts T there for every block after it. Both keys are optional, and a field carrying neither behaves exactly as it did - a write of :unknown, known but untyped, that refuses nothing.

  • field_candidates, the values a host offers for one field, keyed {type_name, field_key}. A :string field with a closed list - [{value, label}] - renders a <select>; an open one - {:open, [{value, label}]} - renders the text input with a <datalist>; no list renders the input the field already had. It draws a control and decides nothing: validate_config/1 is still the only authority on a value, and a stored value a closed list does not offer is drawn as its own option rather than silently rewritten. It is an editor assign and a StatifierBlocks.Compiler.compile/3 option, where a value outside a closed list is a warning on the compiled artifact and never an error.

  • core.on_event's capture has an authoring surface: a repeated two-control row, one row per pair - the datamodel path written beside the path read inside the firing event's payload - with the source control offered the block type's own fixtures/0 payload for the configured event. There is always one blank row at the end, which is what adds a pair; clearing both controls of a row removes one. The key was authored through the document before this and had no control at all.

  • A capture's target paths reach the declared-path advisory, anchored on the capture key. They are datamodel paths that no field declaration names, so the pass covering every other datamodel path could not see them.

Changed

  • The optional statifier_ui dependency's floor moves to ~> 0.8, which is the release that takes :path_types. The dependency is optional in the same sense it always was: a tree without it renders the plain source input, and nothing adds it for you.

  • statifier_datamodel is a Hex dependency at ~> 0.1 rather than a pinned git revision, so a host resolves this package's dependencies from Hex alone.

  • StatifierBlocks.Datamodel reads the datamodel document through statifier_datamodel rather than through an index of its own. declared_paths/1, candidates/3, candidates_under/2, value_candidates/2 and declared_view/3 keep their signatures and their behaviour; what changed is where the projection lives.

  • The declared type set gains date alongside the eight it already carried, because the re-homed record widened it. A document that used no date entry is unaffected.

  • Breaking. The data-flow check is no longer a question about the block before this one. A block declares what it reads and writes at datamodel paths, and check/5, valid_targets/4, validate/3, inbound_type/4 and seam_reason/4 are defined over the environment at a position. They keep their arities. io/1's consumes and produces keep their meaning as sugar: consumes is a read at the document's subject path and produces is a write there. A palette that declares no subject on its entry block's palette entry has no subject path, so that sugar declares nothing and the document validates exactly as an untyped one always did.

  • Breaking. {:type_mismatch, block_id, ref, held, expected} gains a sixth member, the datamodel path the read was checked at. A block may carry several read signatures on several paths, so a message that says which two types disagreed without saying where is one an author cannot act on. {:kind_not_admitted, ...} is unchanged.

  • Breaking. A type expression spelled exactly "unknown" now reads as the permissive :unknown rather than as an opaque expression that happens to be spelled that way. It only ever admits: an opaque "unknown" compared by identity was already satisfied against another "unknown", so nothing that passed before is refused now.

  • The refusal a {:type_mismatch, ...} names is the block whose write signature put the type at the path, found by name rather than by adjacency. A refusal at index 0 of a slot that used to answer :not_assignable - because there was no previous sibling to name - now answers {:fixable_by, block_id} when a block upstream of the container did declare the type.

  • The read check itself is StatifierDatamodel.Types.satisfies/3: unknown, then identity, then a record covering a shape's required set. The palette's host relation still runs, and now runs last, after that coverage step - so the floor a host cannot lower is higher than it was, and a host that was widening records into shapes by hand can delete that half of its module.

  • A core.branch's arms and a core.parallel's lanes no longer blank everything downstream. What leaves a container is the per-path merge: a path every arm holds at one type keeps it, and only a path the arms disagree about drops to :unknown.

  • core.subchart's assign_to accepts a datamodel path, not only a bare lowercase identifier: any non-empty string with no whitespace in it, which is exactly what core.assign accepts for the path it writes. The validation and the emission widen together. It is a widening and nothing else - a bare identifier is a one-segment path, so every document that validated before still validates and compiles to the same bytes - and it settles the field offering dotted candidates its own validation refused. The identical refusal on an <assign> location elsewhere in the vocabulary is untouched.

Removed

  • StatifierBlocks.Predicates.Datamodel is gone. The datamodel document's path/type index moved to the statifier_datamodel package as StatifierDatamodel.Index, which carries the same functions under the same names (index/1, declared_paths/1, sensitive_paths/1, datamodel/1, entries/1, fetch/2, type/2, declared?/2, under/2); a caller swaps the module name and nothing else.

Fixed

  • The editor's drop check consults the datamodel document the editor already holds, instead of asking the data-flow question with an empty context. The coverage step - a record satisfying a shape by covering its required set - could not run there, so a placement the compiler accepts was drawn as refused. StatifierBlocks.Edit.Targets.droppable_slots/3, droppable_slots_for/3 and slot_verdicts/3 each take an optional context as a fourth argument; called with three, every one of them behaves exactly as it did.

0.19.0 2026-09-05

0.19.0 is about what a chart does with the world outside it. core.map (ADR-0009) runs another chart once per item of a datamodel list and waits for the whole batch, compiling to exactly one <invoke> whose bytes do not scale with the list; core.await holds until a named event arrives, with an optional cancel-scoped deadline; and core.on_event gains a capture map that writes values out of the firing event's payload into the datamodel. The editor learns the datamodel's shape alongside them - a {:path, opts} field type with declared-path candidates, event-name candidates read out of the enclosing body, and a chart_outcomes assign the host fills so a subchart's declared outcomes can be checked against the chart's own. The drawer's tab strip answers the WAI-ARIA arrow keys.

Added

  • core.on_event takes an optional capture map: each pair writes one value out of the firing event's payload into the datamodel, the key naming the datamodel path written and the value the path inside _event.data it is read from. The pairs compile to one <assign> each on the handler's own transition, ahead of the raise that carries the outcome, ordered by their datamodel paths so a document compiles to one byte sequence. A handler without the key, or with an empty map, compiles exactly as it did before the key existed. A source path the payload does not carry is written as the interpreter's explicit unbound marker rather than nil, so a consumer of a captured path tests for that marker. The key has no editor control yet - the field-type set has no member describing a map - so it is authored through the document.

  • {:path, opts} is the eighth value in a block type's closed field-type set. A field declared with it holds a path into the host's datamodel, and the editor reaches its control by the type alone: a text input bound to a <datalist> of the declared paths, and the :info advisory for a path the datamodel does not declare, anchored on the field. opts carries no defined key today, so a declaration writes type: {:path, %{}}. It suggests and never constrains - free text stays valid and validate_config/1 remains the only authority on what a value may be.

  • The editor offers value candidates on core.on_event's event field: the completion events the blocks in the handler's enclosing body raise, written as the generated done.outcome.<state id>.<outcome> names the compiler mints and labelled by each block's own card label and outcome. The body is read through the enclosing type's slot_accepts declaration - a slot that admits ADR-0003's :step kind - so a host group is offered on the same terms a core.group is, and only blocks that implement outcomes/1 contribute. The field is still a plain :string: the list is a <datalist> that suggests and never constrains, a free-typed event name validates exactly as it did, and a body with nothing to offer draws the plain input rather than an empty picker.

  • core.map, the durable fan-out block type (ADR-0009): it runs another chart once per item of a datamodel list, all at once, and waits for the whole batch. The palette calls it "For every item, run a chart" - it sits beside core.foreach, which runs the blocks inside it one item at a time, and is not a mode of it. Four fields: items (the datamodel path holding the list), chart (the document id run for each item), collect (where the assembled answer is written, optional), and on (all, the default, or first_error). items and collect are {:path, opts} fields, so the editor offers the host's declared datamodel paths on both; collect accepts what core.subchart's assign_to accepts, refused with the same wording. Any on outside the two permitted values is a config finding at authoring time, which is what reserves the word quorum for its own walk.

  • It compiles to exactly one <invoke> of the constant type "statifier_blocks:map", carrying the block's own id, the document id as src, and the four values as literal <param> children. The compiled bytes do not scale with the length of the list and cannot: the list is a runtime value the compiler never sees, so the params carry the path and the host's registered handler is what resolves it and starts the runs. Nothing about the size of a batch is validated here - a bound on it, if one is needed, is the fan-out runtime's to enforce and to refuse against. The type is deliberately a different string from "statifier_blocks:subchart": a host that wired a single-child handler has not thereby wired a fan-out handler, and StatifierBlocks.Compiler.InvokeTypes reports that gap at deploy time.

  • Two fixed outcomes, done and error, each with an optional slot (on_done, on_error), because N children report N outcomes and there is no branch target to be had by joining them - the per-child answers go to collect instead, and an author who wants to branch on them reads that list with a core.branch after the block. The collected answer is written once, on the success transition, in the shape core.subchart already uses.

  • The core vocabulary is now seventeen types, and core.map is registered in StatifierBlocks.Palette.core_types/0 beside the rest.

  • core.await joins the core vocabulary: an in-flow leaf that holds until a named event arrives, with an optional deadline. It is valid wherever core.wait is, takes event (required) and timeout (an optional duration), and declares two outcomes, received and timed_out, so a parent can route the two ways it can end. A configured timeout arms a cancel-scoped delayed send, so leaving the await early - because the event arrived, because an interrupt fired, because a group was abandoned - leaves no timer behind. Both outcomes are declared whether or not a deadline is stored; with no timeout, no timer is armed and no timed_out final is emitted.

  • A chart_outcomes editor assign, %{document id => [outcome]}: what the host says each of the documents it compiles with :child_use finishes with. A core.subchart names a chart by document id and declares its outcomes by hand - a block type cannot read the document it references - so the assign is where that knowledge enters. Two things read it: the outcomes field offers the named chart's finals as a <datalist>, and StatifierBlocks.ViewModel.outcome_findings/3 reports a disagreement between that list and what the block declares, anchored on the outcomes key and also reachable through StatifierBlocks.Editor.findings_count/3's new :chart_outcomes option.

  • The disagreement is a :warning, never an error: the document compiles either way, and what a mismatch costs is a conditioned done.invoke transition that can never match. A chart the map does not name produces nothing at all - unknown is not disagreement - and so does an entry holding an empty list, so %{} (the default) leaves the editor drawing exactly what it drew before. The field is still a plain :string: the list suggests and never constrains, and a free-typed name validates as it did.

  • The drawer's tab strip answers the WAI-ARIA tablist arrow keys. With the strip's one Tab stop focused, Left and Right move one tab and wrap, Home and End go to the ends, and the tab moved to is both selected and focused - so a tab past the clipped edge at the narrow breakpoint is reachable from the keyboard and is scrolled into view when it is reached. Host-contributed tabs are walked alongside the package's own, in the order the strip draws them. The movement is server-side and adds no JavaScript hook: the strip reports the key through phx-keydown and the editor picks the neighbour out of the same tab list a pointer pick reads.

Changed

  • core.subchart's assign_to is declared {:path, %{}} rather than :string, so it offers the declared datamodel paths as candidates and gets the undeclared-path advisory. What it accepts is unchanged - a bare lowercase identifier, which is a one-segment path - and it compiles to the same <assign location=...> it always did.
  • StatifierBlocks.BlockType.datamodel_path?/1 answers true for a {:path, opts} field by construction. The datamodel_path?: true key is not withdrawn: it keeps its meaning, its control and its advisory, a declaration carrying both spellings says one thing twice rather than contradicting itself, and every declaration written before the type behaves exactly as it did.

Fixed

  • The palette's count line no longer counts a recipe as a block type. It reports the two kinds apart - "16 block types, 1 recipe" - and when a slot's acceptance set is narrowing the list it counts types alone, naming the recipes in a clause of their own ("2 of 16 block types fit here; 1 recipe also listed"), because a recipe is not a block type and no set of type names can answer whether one fits.

0.18.0 2026-09-05

0.18.0 lets a palette put down more than one block at a time. A palette may name recipes beside block types - the core palette ships one, "deadline", whose single pick writes the core.send / core.on_event pair that spells a clock interrupt - and the edit vocabulary grows a compound, so an arrangement like that undoes in one gesture. Alongside it a palette entry may declare how many blocks of its type a document holds, the two core timer types rewrite a stored duration written in the retired spelling as the block resolves, and the editor toolbar's Fit active follows the run marks a host paints.

Added

  • A palette may name recipes beside block types: arrangements an author picks the way they pick a type. A recipe is a module implementing StatifierBlocks.Recipe - insert/2, handed the armed position and the document, answering with the commands that build the arrangement, and palette_entry/0, which draws it in the palette browser exactly as a type draws. Register recipes with the new :recipes option on StatifierBlocks.Palette.new/2 (a name => module map) or on from_modules/2 (an ordered list, later entries winning), and resolve one with fetch_recipe/2. Recipe names and type names are two namespaces, not one: a recipe named "deadline" and a block type named "deadline" do not collide. A recipe may target the armed position and any slot of the block enclosing it, and nothing above that.

  • The core palette registers one recipe, "deadline" (StatifierBlocks.Palette.core_recipes/0). One pick puts down the pair that spells a clock interrupt: a core.send carrying a generated deadline event and a delay, at the head of the enclosing group's body, and a core.on_event naming the same event on that group's interrupts rail. The pair compiles clean before the author types anything. Picked at a position whose enclosing block has no interrupts rail, the gesture is refused and nothing is written.

  • StatifierBlocks.Edit.t() admits a composition, {:compound, [t()]}, carrying a non-empty list of the five commands. Edit.apply/2 applies its members left to right and returns the compound of their inverses in reverse order, so StatifierBlocks.Edit.History remembers it as one undo entry: one gesture in, one gesture out. A member that refuses refuses the whole compound, with that member's own error term and no document at all. The set of edits is still five - a compound's leaves are drawn from it, and a list that is empty or holds a compound of its own is refused rather than flattened.

  • A palette entry may declare singleton: :head | :anywhere, saying how many blocks of its type a document may hold and, for :head, that the one it holds is first in the root's first declared slot. A document that does not comply draws one :config finding per violating type, anchored on the root block, so a host gets "exactly one of this, at the top" without writing a validator of its own. Read it with StatifierBlocks.BlockType.singleton/1. Nothing is inserted, removed or moved on the author's behalf - the editor says what is wrong and the author acts. An entry that omits the key, or declares a value this package cannot read, is unconstrained exactly as it is today.

Changed

  • StatifierBlocks.Editor.Toolbar.toolbar/1's :selected? attribute is now :fittable?, because a selection is no longer the only thing that gives Fit active something to fit. Rename the attribute at any direct call site; hosts rendering the editor pass nothing here.

  • A StatifierBlocks.ViewModel.PaletteGroup entry now carries :kind (:type or :recipe) and :name, the name in whichever of the palette's two maps it came from. :type_name is unchanged on a type entry and absent on a recipe entry, which has no type name at all.

Fixed

  • The editor toolbar's Fit active follows the run marks a host paints: it is enabled whenever there is a selection or an active mark, and with nothing selected it fits and reveals the first marked block.

  • A document saved before the duration pivot opens clean instead of showing a refusal on every timer: core.wait and core.send are at type_version 2 and rewrite a stored duration written in the older spelling into the one the field reads. The rewrite happens in memory as the block resolves and nothing is written back, so persisting it stays the host's decision.

  • The drawer's tab strip scrolls sideways at the narrow breakpoint instead of wrapping to a second row, so the bar stays one row tall however many tabs a host adds, and a fade marks the end the strip is clipped at.

0.17.0 2026-09-05

0.17.0 makes a stored duration mean one thing. A :duration field reads the duration strings Predicator.Duration parses - 30s, 1h30m, 2d, 3d8h - and refuses every other spelling, which is what finally lets 500ms and 1.5s through: the intermediate canonical form that stood between them and the engine is gone, and with it two public functions. The editor surfaces grow alongside it - a Fixtures tab on the inspector, a fixture-derived hint beside a condition field, datamodel-derived value candidates, done-event chips drawn as the block they name, and an on_select callback for a host panel that follows the canvas.

Added

  • The inspector carries a fourth tab, Fixtures, about the selected block: one row per fixture row the block owns, with the slot the row expects, the slot the compiled chart took, and the verdict. It carries a count chip in the Findings tab's style, reading the selected block's row count. StatifierBlocks.Shell.inspector_tabs/0 now answers four tabs, with :fixtures last, and StatifierBlocks.Shell.inspector_tab/1 resolves the new name.

    The rows are the same StatifierBlocks.Runtime.FixtureRuns result the drawer's Fixtures tab renders - no second execution path - filtered to the selected block. Runs now refresh when the inspector's tab is picked and when the selection moves, as well as on the drawer's own tab, so switching blocks cannot leave the previous block's verdicts on screen.

    The drawer's Fixtures tab is unchanged: same rows, same verdicts, same count, same place in the tab order. With no selection, no fixtures source, or no rows for the block, the new tab reads its own empty state for that case; a document mid-edit reads as mid-edit rather than as a fixture failure.

  • The editor accepts an on_select function and calls it with a %{id:, type:, label:} descriptor for each new selection, or nil when nothing is selected, so a host panel can follow the canvas.

  • Sub-second and fractional durations are expressible in a :duration field for the first time: 500ms and 1.5s both validate and both compile to a delay the engine resolves. The recogniser that stood between them and the engine is gone.

  • A condition's value picker now offers what the datamodel declares. A path whose ADR-0006 entry carries one_of gets those values by default, with no value_candidates map supplied; a host entry for a path replaces the derived list for that path and leaves every other path's default in place. StatifierBlocks.Datamodel.value_candidates/2 is the derivation, and nothing validates against the list - one_of stays a hint.

  • StatifierBlocks.ViewModel.summary_chip_titles/1 gives the raw text behind each summary chip, nil where the chip is drawn as its type declared it.

  • StatifierBlocks.Compiler.StateId.undone_event/1 inverts a generated done.state or done.outcome event name back to the block it names, or answers :error for a name that does not invert unambiguously.

  • A condition field now draws a hint beside its control, derived from the selected block's fixture rows. The exemplar is what the block's first fixture row in declaration order binds to the path the condition names, and every distinct value that path takes across the block's rows is listed on the hint's title. StatifierBlocks.Shell.fixture_hint/3 is the derivation, over the fixtures the editor already holds.

    It is a hint and never an option: nothing reaches a picker, nothing is merged with one_of or with a host's value_candidates, and no value it shows can be selected. A block with no fixture rows, or a document with no fixtures source, renders exactly as it did before.

Changed

  • Breaking. A :duration field reads one grammar: the duration strings Predicator.Duration parses, such as 30s, 1h30m, 2d and 3d8h. Any other spelling is a format finding.

    Migration: a core.wait duration or a core.send delay stored in ISO-8601 - PT30S, PT1H30M, P1D - no longer validates. Rewrite the value in the unit spelling above; PT1H30M becomes 1h30m, P1DT6H becomes 1d6h, and the compiled chart is byte-identical either way.

  • StatifierBlocks.Core.Duration.to_delay/1 now takes the normalised duration parse/1 returns rather than a canonical string.

  • A summary chip whose text has the shape of a generated done-event name is drawn as the named block's label and the outcome, with the raw name on the chip's title attribute, so a card no longer shows the compiler's spelling of a fact the author stated (ADR-0005 decision 10w).

  • The presentation cap measures the drawn chip rather than the generated one, so the cap lint no longer names a string the author cannot shorten. A translated chip that is still over the cap is refused exactly as before.

  • summary/2, summary_refusals/2 and summary_refusal_message/3 take an optional trailing map of block labels. A call that passes none behaves exactly as it did: without labels nothing is translated.

Removed

  • StatifierBlocks.Core.Duration.to_iso/1 and StatifierBlocks.Core.Duration.predicator?/1. With one grammar in and one attribute out there is no intermediate form to canonicalise to, and no narrower predicate to hold beside duration?/1. Callers that compiled a stored value should use parse/1 and then to_delay/1.

0.16.0 2026-09-04

0.16.0 fills the expression seam. A condition's :expression field now renders statifier-ui's expression editor - picklists of field, operator and value over the source it can round-trip, and the plain source input over the rest - so a signup-wizard condition like step in ['payment', 'review'] is composed rather than typed, and a host can offer the values each datamodel path accepts through a new value_candidates assign. statifier_ui joins phoenix_live_view as an optional dependency: with it absent, nothing raises and an :expression renders exactly what it rendered before.

Added

  • An :expression field renders statifier-ui's expression editor - picklists of field, operator and value over the source it can round-trip, a text input over the rest - when statifier_ui is on the load path, and the plain source input when it is not.
  • A value_candidates editor assign, %{path => [%{label:, value:} | binary]}, carrying the values a host offers per datamodel path through to the expression control.

Changed

  • statifier_ui is a new optional dependency, resolved the way phoenix_live_view already is: absent, nothing raises and an :expression renders exactly what it rendered before.

0.15.0 2026-09-02

0.15.0 is about seeing what a document actually declares and what its fixtures actually do. The editor's drawer gains two more tabs: Fixtures, which drives every attached fixture row through the compiled chart and reports the outcome it took against the one it expected, and Datamodel, a read-only grid of every declared path with the surfaces that declared it. Compiling a core.resumable_group whose body opens with a delayed send now warns that no deadline is re-armed after the first resume. The rest is polish on the drawer and the truth table: keyboard reach, wrapping, pinned case names, and refusals that read as sentences.

Added

  • The editor's drawer gains a fourth tab, Fixtures, that drives every attached fixture row through the compiled chart and reports each one's expected outcome against the one actually taken.
  • Compiling a core.resumable_group whose body opens with a delayed core.send and whose interrupts rail carries a resume handler now produces an advisory warning: no deadline is re-armed after the first resume, and the two escapes are arming the deadline outside the group or using a core.group.
  • The editor's drawer gains a fifth tab, Datamodel: a read-only grid of every declared datamodel path, with the surfaces that declared it and the type, scope and label the ADR-0006 projection carries. It shows exactly the vocabulary the undeclared-path advisories are decided against, and StatifierBlocks.Datamodel.declared_view/3 is that projection on its own.

Fixed

  • The drawer's tab strip is now a single Tab stop instead of one per tab, matching the inspector's strip: Tab moves from the canvas to the active drawer tab and then on to the height slider, rather than stepping through all six tabs on the way.
  • A drawer tab keeps its label and its count on one line at every width; a bar too narrow for all six tabs now wraps whole tabs onto a second row instead of breaking each tab into a three-line stack.
  • A folded container carrying findings no longer draws its findings count underneath the fold control: the card reserves the corner the fold occupies whenever it renders a badge.
  • A truth table wider than the drawer can now be scrolled from the keyboard. The table's scroll region is a Tab stop and a labelled region named for the table, so the arrow keys, Home and End reach the columns past the right edge; previously every cell in it was text and nothing in the panel was focusable, which left those columns reachable by pointer alone.
  • A mismatched truth-table cell now says what the disagreement was - "expected yes, selected no" beside the status word and on the cell's tooltip - instead of only the word mismatch, and a row's note renders under the row name instead of being dropped.
  • A truth table's case column stays pinned to the left edge while the table is scrolled sideways, so a status word keeps the row name (and its note) beside it instead of losing its row on any table wider than the drawer.
  • A truth-table row whose bindings fail to build a context now reads as a sentence naming the binding and what went wrong, instead of the inspected error tuple; an unrecognised reason still shows its term.

0.14.0 2026-09-01

0.14.0 is about a document declaring its own data, and about a subchart that outlives the process running it. core.subchart gains a durable runtime: the child runs as its own persisted run and re-enters the parent through the driver's own door when it finishes. The editor's drawer gains a third tab, Declarations, so the datamodel roots a document depends on are authored in the document rather than supplied only by the host, and every :expression field offers the declared paths as suggestions without constraining what an author may type.

Added

  • StatifierBlocks.Runtime.DurableSubchart runs a core.subchart child as its own persisted statifier_persistence run: dispatch_fun/1 builds the fun the driver's :dispatch option takes, and the child's completion re-enters the parent through the driver's own door.
  • The editor's drawer carries a third tab, Declarations, where an author adds, edits, reorders and removes the document's own datamodel roots (ADR-0001 decision 11's entries: id, expr, description).
  • StatifierBlocks.Edit gains a fifth command, {:set_datamodel, entries}, which replaces the document's whole declaration list and whose inverse is the list that was there before, so a declaration edit undoes and redoes like any other. It refuses a list StatifierBlocks.Document.validate/1 would refuse, in the same {:malformed_envelope, {:datamodel, reason}} family.
  • StatifierBlocks.Declarations, the pure list-to-list arithmetic behind the panel: add/1, remove/2, move/3, put/4, change/3, count/1 and refusal/1.
  • StatifierBlocks.Datamodel.candidates/3 returns the declared datamodel paths, unioned from the host's datamodel, the compile call's :declare roots and the document's own datamodel key, sorted and deduplicated.
  • StatifierBlocks.Datamodel.candidates_under/2 narrows those candidates to a prefix, in the datamodel document's own order.
  • An :expression config field offers the declared paths as a <datalist>, suggesting without constraining; supplying no datamodel renders the plain input unchanged.

Changed

  • The expression_component override now receives a :candidates key alongside :field, :id, :name and :value. Existing overrides read the keys they know and are unaffected.

Fixed

  • An arranged container whose slot ends in a drafts shelf drew its rejoin edge from the shelf's own outlet, asserting flow out of the one card nothing enters and nothing leaves. The rejoin now leaves the last step in the slot, and a slot holding nothing but a shelf rejoins from its header, the same as an empty one.

0.13.0 2026-09-01

0.13.0 is a small release about what an editor shows first. The drafts shelf 0.12.0 introduced opened expanded, so a document with parked work opened showing its shelf rather than its flow; now a stocked shelf opens folded and the flow is what the author sees. Nothing about the shelf's placement rules, its compile output or the fold control itself moves.

Changed

  • The editor opens a non-empty drafts shelf folded, so a document with parked work opens showing its flow rather than its shelf. An empty shelf opens as before, the fold control is unchanged, and unfolding lasts for the session.

0.12.0 2026-09-01

0.12.0 is the release where a document can hold work that is not finished yet. core.drafts is the document's shelf - one per document, a direct child of the root, holding fragments an author has built but not placed - and it compiles to nothing, so shelving work never moves a byte of the emitted SCXML. core.placeholder is the other half: an in-flow marker for a step left unwritten on purpose, compiling to a state that completes on entry so a preview walks straight through the gap. Both are visible to a host rather than silent - misplacement and duplication are Structure errors, and a non-empty shelf or a placeholder raises an Emit warning on a compile that still succeeds. The editor draws the shelf as a tray: a detached row at the foot of the canvas with no boundary box and no connectors, because parked work is not part of the flow.

Added

  • core.drafts, the document's shelf: a container whose one body slot holds fragments an author has built but not placed. It is admitted as a direct child of the root block's body and nowhere else, a document carries at most one, and it compiles to nothing at all - a document with work on its shelf, the same document with the shelf emptied, and the same document with no shelf produce byte-identical SCXML.
  • core.placeholder, an in-flow leaf marking a step left unwritten on purpose. One optional note field, and it compiles to a state that completes on entry, so a preview of a half-built workflow walks straight through the gap.
  • Two Structure-stage errors: :drafts_block_misplaced and :duplicate_drafts_block, the second naming the second and every later shelf in document order rather than the first.
  • Two Emit-stage warnings on a compile that succeeds: :draft_blocks_present, once per document on a non-empty shelf, and :placeholder_block, once per marker, carrying the author's note. What a host does with either - a publish gate, say - is the host's.
  • slot_style: :tray, a fourth value for palette_entry/0's slot_style map. A tray is a detached shelf: no boundary box, and no connector into it, out of it, or between one fragment and the next. StatifierBlocks.ViewModel gains tray?/1, shelf?/1, flow_children/1 and shelf_children/1.
  • StatifierBlocks.Shelf, the module owning the shelf's placement rules and the type-name predicates other layers ask about it.

0.11.0 2026-08-31

0.11.0 is the release where declaring a block type stops being boilerplate. use StatifierBlocks.BlockType fills in every callback but emit/2, and use StatifierBlocks.InvokeStep goes further for the shape that keeps recurring - a leaf step naming one host invoke type - so such a block type is a handful of options rather than a module of callbacks. Beside it, the statifier_blocks:subchart invoke type gains a canonical handler: use StatifierBlocks.Runtime.Subchart generates the in-memory Statifier.Session implementation a host would otherwise write by hand, leaving the document resolver as the host's seam and putting a stated reason on every refusal.

Added

  • use StatifierBlocks.BlockType declares the behaviour and injects an overridable default for every callback but emit/2.
  • use StatifierBlocks.InvokeStep declares a leaf step that names one host invoke type, filling in every callback from an invoke type, an optional produced type, extra config fields, and palette keys.
  • use StatifierBlocks.Runtime.Subchart declares the canonical statifier_blocks:subchart invoke handler for the in-memory Statifier.Session case: the host supplies a document resolver and a palette, and the generated module resolves the chart a core.subchart names, compiles it as a child, and starts it as a child session. Refusals surface on error.communication.invoke with a stated reason - unknown_document, child_compile_findings, or cycle_refused.
  • StatifierBlocks.Runtime.Subchart.handlers/1 builds the :invoke_handlers map for Statifier.Session.start_link/2.

0.10.0 2026-08-31

0.10.0 is the release where a document declares its own data. A block document carries a top-level datamodel key naming the <data> roots its own guards and assigns read, so a document compiles on its own rather than depending on whichever host happens to declare them; the compile call's :declare option stays beside it as the host-wins surface, and the editor's undeclared-path advisory reads both. Around that, core.on_event takes a cond, so an interrupt rule fires only when its event arrives and the condition holds; the done outcome of an invoke run mark gets theme tokens of its own instead of borrowing the accent blue it was indistinguishable from; and the canvas stops falling short of the tree it holds.

Added

  • The block document gains a top-level datamodel key: a list of StatifierBlocks.Document.DatamodelEntry structs, each an id plus an optional expr and an optional description, naming the <data> roots the document's own guards and assigns read. Document.new/2 takes a :datamodel option (default []); a document declaring none compiles and encodes byte-identically to one built before this key existed. The key is part of the document's canonical bytes, so it participates in content_hash/1 and in compile determinism.
  • Compiler.compile/3's :declare compile option now leads a second declaration surface rather than being the only one: the compile call's roots emit first, the document's own datamodel roots follow, and block-declared roots follow those, all in one <datamodel>. A root both the compile call and the document declare is host-wins: the compile call's declaration is emitted, the document's is dropped, and the compile succeeds with a :shadowed_document_root warning on Compiled.warnings rather than refusing. A document root colliding with a block-declared root is still refused as :duplicate_binding.
  • core.on_event takes an optional cond: an :expression config field that becomes the guard on the handler's transition, so an interrupt rule fires only when its event arrives and the condition holds. A handler whose cond is absent, empty, or whitespace emits exactly the bytes it emitted before the key existed, so the key is additive over every document authored without it. The condition is passed through to predicator verbatim - validate_config/1 only asks that the stored value be a string - and the transition carries "cond" as its attribution key, so an upstream expression error lands on the field the author typed into. Recorded as ADR-0002's 2026-08-31 note.
  • StatifierBlocks.Editor takes a declare assign, the {id, expr} roots the host will pass StatifierBlocks.Compiler.compile/3 as :declare, defaulting to []. Editor.findings_count/3 takes the matching :declare option, so the host's number and the drawer's stay the same number. The document's own roots need no option - they are read off the document.
  • StatifierBlocks.Datamodel.declared_roots/1, the total normalizer for a root set, beside declared_paths/1.
  • Two tier-2 theme tokens, --sb-run-done and --sb-run-done-bg, for the done outcome of an invoke run mark. A host that themes the editor now sets the "came back" colour directly instead of inheriting whatever it set --sb-info to.

Changed

  • Decoding a document now refuses an envelope object carrying a key outside id, revision, root, schema_version, metadata, datamodel, and refuses a datamodel entry carrying a key outside id, expr, description, or carrying an explicit JSON null for expr or description - matching the round-trip discipline already applied to unrecognized block keys.

  • The family's two worked examples - the ADR-0001 card authorization document and the signup wizard - now declare the <data> roots their own guards read through ADR-0001 decision 11's document datamodel key, rather than leaving them to whichever host compiles them. Each declares exactly what it reads and no more: budget_remaining and amount for the first, variant for the second, none of them carrying an expr, since all three are per-run values a host seeds or a step assigns. Either document now compiles on its own, with no :declare option, without its guard raising error.execution over a root nothing declared.

  • Both documents' canonical bytes move with the key, and so do their document hashes and the worked example's pinned chart identity. The identity move was verified rather than accepted: deleting the one <datamodel> element from the new bytes reproduces the previously pinned hash exactly, so nothing else in either emission changed. Only the test fixtures ship these documents; no packaged code moved.

  • The editor's undeclared-path advisory now reads three declaration sources rather than one. A config field a block type annotated datamodel_path?: true is declared when the host's datamodel holds its path, when the document's own datamodel key declares its root, or when the compile call's :declare option declares its root. A declared root covers every path beneath it; the datamodel's own paths are still matched whole. ADR-0005's decision 11 is amended as 11k-11m, taking the open question ADR-0001's 2026-08-31 amendment left to it in its clause 11g.

  • The check's precondition widens to match: it runs when a datamodel was supplied or when either surface declares a root, and still produces nothing at all when nothing anywhere was declared. A host that passes no datamodel and no roots sees exactly what it saw before; a document that declares its own roots now lints its own paths with no host involved.

  • An invoke mark whose outcome is done is drawn in --sb-run-done rather than --sb-info. In the shipped light palette --sb-info and --sb-accent are the same blue, so a block that was active and a block that had come back differed only in treatment - a halo outside the border against a border and a fill - and read as one state at a glance. They are now different hues in every theme the package ships or documents.

    This is a visual change for a host that did not override --sb-info, and a host that had retuned --sb-info specifically to colour its done marks should move that value to --sb-run-done. Nothing else reads the new tokens, and the error outcome still takes --sb-error.

    The complete host theme in docs/theming.md restates both tokens, as the theme audit requires of every colour token, and the accessibility discipline is unchanged: --sb-run-done is held to the 3:1 non-text threshold as a border that carries information, and --sb-run-done-bg is recorded as translucent and held to no ratio.

Fixed

  • A summary chip whose text is a single identifier - signup.reminder_due
    • no longer breaks mid-token onto a second line. The chip keeps one line and clips at its own edge with an ellipsis, which is the treatment the invoke type on the line below it already used. The card's title keeps its wrap: a title is prose an author wrote, not an identifier.
  • The canvas stage is now at least as wide as the tree it holds (min-width: min-content). Its box took the panel's width, while the tree has a floor of its own - nesting paddings, the interrupt channel, a card that will not shrink past its token width - so a panel narrower than that floor left the stage short of its own content, and everything read off that box, the measurement hook's offsetWidth included, was short with it.
  • Connectors.fan_path/3, Connectors.join_path/3 and Connectors.interrupt_path/4 now clamp an ascending edge level, as flow_path/3 already did: the head is raised to the tail's own y rather than routed upward. An arrowhead is oriented along its path, so an ascending arm rendered an arrow pointing back at the block the flow just left - a loop the document does not contain.

0.9.0 2026-08-31

0.9.0 is the release where the editor can watch a run. A host hands the editor the block ids a run has activated and the block it is calling out to, and the editor draws them on the cards themselves, in the accent and finding tokens a theme already retunes. The same host seam widens the drawer: a host contributes its own tabs beside the package's Truth tables and Findings, so a run feed the host is appending to lives in the drawer rather than beside the editor. Around that the chrome gives room back - the inspector folds to a rail the way the palette already does - the config form's fields are dressed from the same tokens as the palette's search box, and connectors stop overshooting the cards they point at.

Added

  • The editor's inspector folds to a rail from a chevron on its own header, the way the palette already does, giving its width back to the canvas.
  • --sb-inspector-collapsed-width, the theme token for how wide that rail is.
  • The editor takes run marks from its host: active_marks, the block ids a run has activated, and invoke_mark, the block a run is calling out to together with how the call came back. Both are ordinary assigns, so a host pushes them with Phoenix.LiveView.send_update/3 and needs no new API; both are held as editor state, so a re-render the host makes for its own reasons does not drop them; and both are cleared when the host opens a different document, because a mark addresses one block.
  • The marks reach the markup as data-run-active, data-run-invoking and - only once a call has come back - data-invoke-outcome on the block's .sb-node, and the stylesheet draws them in tokens a host theme already retunes: the accent family for the active mark, the finding severities for the outcomes. A mark on a folded container stays visible.
  • The editor accepts an optional invoke_types assign, and an invoke_type field then offers those types as a suggestion list; free text stays valid and an unknown type stays a lint rather than a refusal.
  • StatifierBlocks.Connectors.slots_anchor/1, the anchor key for a container's body box. The editor stamps it on the element holding a container's slots, and an interrupt channel is now offset from that box rather than from the container's whole node box.
  • The editor takes drawer tabs from its host: drawer_tabs is a list of %{id:, title:, content:} descriptors, with an optional count:, and each one is drawn beside the package's own Truth tables and Findings tabs and activates the same way. content is a function component, the same seam shape icon and expression_component use, called when its tab is the active one. The descriptors are ordinary editor state, so a host pushes them with Phoenix.LiveView.send_update/3 and a feed the host is appending to redraws as it grows - a tab whose content changes under the host is what the seam exists for.
  • The collapsed strip and the unchosen-tab resolution reach host tabs too: a document with no truth tables and no findings and a running feed opens on the feed rather than on an empty Truth tables 0, and a collapsed drawer names it.
  • StatifierBlocks.Shell.host_tabs/1 is the admission rule, and StatifierBlocks.Shell.drawer_tab/2 resolves a tab name against the package's tabs and the host's together. A host tab named for one of the package's own, or repeating an id already used, is not drawn: the id is stamped into the tab's DOM id and its panel's. No tab name is ever turned into an atom, so a crafted phx-value-tab reaches at worst a tab the host declared.

The package's own tabs, the drawer's five states, its resize and its collapse are unchanged, and a host that contributes no tabs gets the drawer it had.

Changed

  • The config form's fields are dressed the way the palette's search box always was - a token-built box with a border, a radius and the pane's own fill - instead of being left to whatever the host's browser paints a bare <input> or <select> as. The two are now one rule rather than two, so a host retuning --sb-border, --sb-radius-sm or --sb-bg moves every control in the editor together.
  • Placeholder text in those fields takes --sb-fg-subtle, and a disabled field is muted with --sb-disabled-opacity, which is what a disabled sb-button has always used. Boolean fields, which render a checkbox, are excluded from the box by selector and are unchanged.

Fixed

  • Connector arrowheads keep one size at every stroke width and land their tip on the endpoint, rather than being scaled by the stroke of the path that references them and overshooting the card they point at.
  • A flow edge whose head was measured above its tail is drawn level instead of ascending, so no arrowhead points back at the block the flow just left.
  • An interrupt edge clears the container it exits instead of turning down inside it. A container is now as wide as what it holds rather than as wide as the box around it, so the box the routing is measured from means what it encloses.
  • A gap's insertion marker masks the flow line it sits on, and its glyph stays legible when the canvas is zoomed out. The dashed placeholder ring stays on empty slots only.

0.8.0 2026-08-30

0.8.0 is the release where a container folds shut. A container in the editor now carries a fold control on its own card, and folding it does not hide what is inside: a folded container wears a ring badge counting the findings under it, so a problem stays visible from the outside. Alongside that, an editor whose host swaps one document out for another fits the new document the way it fitted the first - the fit attr is spent once per open document rather than once per editor, so a swap now behaves exactly as a mount does. Hosts: see Removed - the --sb-fg-on-accent theme token is gone, and a theme that still sets it can drop the declaration.

Added

  • A container in the editor folds shut from a control on its own card, and a folded container carries a ring badge counting the findings inside it.

Changed

  • The editor's fit attr is now spent once per open document rather than once per editor: a document the host swaps in is armed from the attr passed in that same update and fitted by the next measurement, exactly as at mount. A host re-render carrying the document already open still never re-fits, and :manual or an absent attr still arms nothing.

Removed

  • The --sb-fg-on-accent theme token, which no rule in the stylesheet reads any more. A host theme that sets it can drop the declaration.

0.7.0 2026-08-30

0.7.0 is the release where a finding looks like a finding wherever it is read. One row anatomy - severity word, anchor tail, source chip, message - now draws on the drawer's tab and on both inspector panels, which used to show the bare message, and the two document-level surfaces carry a severity pill row above the list. Alongside that, the 24-character presentation cap on summary chips stops failing silently: a dropped chip raises a :lint warning against its block and summary_refusals/2 says which chip and why, so a card with no second line can be told apart from a type that declared none. And an editor opened with a fit no longer flashes at 100% before snapping to it. Hosts: see Removed - :no_presentation_source leaves Finding.from_compiler/2's refusal type, and the canvas toolbar's dead inserting? attribute is gone.

Added

  • StatifierBlocks.Shell.severity_counts/1 cuts a findings list by severity, in :error, :warning, :info order and omitting any with nothing at them; it sums to findings_count/1 for the same list.
  • StatifierBlocks.Editor.Findings.row/1 and anchor_tag/1 are public, so a host rendering findings of its own gets the editor's row anatomy rather than re-deriving it.
  • StatifierBlocks.BlockType.summary_refusals/2 reports the summary chips the 24-character presentation cap dropped, as {index, reason} with reason in :too_long, :blank, :multiline or :not_a_string, and summary_refusal_message/3 puts one into the words an author reads.

Changed

  • A finding renders the same way everywhere: severity word, anchor tail (config.duration, slot:body, nothing for a block anchor), source chip and message. The inspector's two findings panels showed only the message before.
  • Both document-level findings surfaces - the drawer's Findings tab and the inspector's with nothing selected - carry a severity pill row above the list. The list itself is still grouped by block.
  • The view model raises a :lint warning against a block for every summary chip the presentation cap refused, so a card that draws no second line can be told apart from a type that declared none. A well-formed document gains no findings.

Removed

  • StatifierBlocks.Finding.from_compiler/2 no longer declares the :no_presentation_source refusal in from_compiler_error/0; nothing has produced it since the adapter began mapping unplaced compiler findings to :compile, so a caller that matched on it can drop the clause and keep the {:unanchorable, finding} one.
  • The canvas toolbar's dead inserting? attribute, left declared when 0.6.0 removed its "Cancel insert" button. A host still passing it to StatifierBlocks.Editor.Toolbar gets an undefined-attribute warning; drop the assignment (statifier_blocks#172).

Fixed

  • An editor opened with fit: :width or fit: :active no longer paints its first frame at 100% and then snaps to the fit. While a fit is armed the root carries data-fit-pending and the stylesheet keeps the stage unpainted under it, so the first frame an author sees is the fitted one; a host that never imported the measurement hook, and so never spends the fit, is revealed by a CSS-only fallback half a second in rather than left blank.

0.6.0 2026-08-30

0.6.0 is the release where the editor tells one story about findings. The Findings tab reads the whole document when nothing is selected, the drawer's strip and a host header can finally show the same number, and a container no longer wears a badge on every uncollapsed face. Alongside that: a block type's summary draws as a row of chips rather than one joined string, the palette's search field wears the package's own chrome, and a host can open a document already fitted to the canvas. Hosts: see Removed - source: :arity leaves StatifierBlocks.Finding (pass :assignability) and :compile joins the enum; ViewModel.subtitle/1 now answers only a type label, with the chips behind the new summary_chips/1; .sb-badge is rendered by nothing; and the canvas toolbar's "Cancel insert" is gone, its job done by the palette's Cancel.

Added

  • StatifierBlocks.Finding's source gains :compile, and from_compiler/2 maps any compiler finding its by-stage rule cannot place onto it instead of refusing, so a compile error raised against generated SCXML or against the document envelope can be rendered by the editor.
  • StatifierBlocks.Shell.findings_groups/3 groups a document's findings by the block each is anchored to, unanchored ones last, without dropping any - the grouping behind that panel, and headless like the rest of Shell.
  • The editor takes a fit attr - :manual (the default, unchanged behaviour), :width or :active - so a host can open a document already fitted to the canvas instead of leaving the author to press Fit width on every document. The first measurement performs the fit once; after it the editor behaves exactly as if the author had pressed the button, and an unknown value is refused into :manual.
  • StatifierBlocks.Editor.findings_count/3 returns the number of findings the editor's Findings tab reports for a document, from the same document, palette, findings and datamodel a host already passes the component, so a host header and the drawer cannot show two different numbers.

Changed

  • A block type's summary/1 chips draw as a row of separate chips under the card's title (.sb-node__summary, one .sb-node__chip per chip) instead of one string joined with ", "; the row wraps, and a type declaring no summary draws no row at all.
  • StatifierBlocks.ViewModel.subtitle/1 answers only the type label an author- named card carries, and nil otherwise. Read the chips from the new StatifierBlocks.ViewModel.summary_chips/1 instead of the joined string.
  • The inspector's Findings tab reads the whole document when nothing is selected: the count beside the tab is the document's findings number - the same one the drawer's strip and StatifierBlocks.Editor.findings_count/3 report - and the panel lists those findings grouped by block, each row selecting its block. Findings anchored to a block the document no longer holds get an Unanchored group of their own, since they are inside the count. With a block selected the tab is that block's findings, unchanged. A host styling the panel has three new classes: .sb-inspector__groups, .sb-inspector__group (with data-block-id and data-unanchored) and .sb-inspector__group-title, plus .sb-inspector__group-row on the rows.
  • The palette's search field is drawn by the package - a border, a surface, padding and a radius, all from --sb-* tokens - rather than left to whatever box the host's browser paints inside the pane.
  • A gap's "+" wears the editor's button chrome at rest, so an insertion point reads as a control without being hovered first. Its hover, armed and drag states are unchanged.
  • The finding count badge no longer renders on a container's face. ADR-0005 places it on a collapsed subtree and the editor has no collapse command yet, so a badge on every container read as an error on every card while the counts multiplied up the tree. Every finding is still reachable: the node keeps its subtree rollup in data-findings-count, and the drawer's Findings tab and the inspector both list them. A host styling .sb-badge should know the class is now rendered by nothing.

Removed

  • :arity leaves StatifierBlocks.Finding's source type. Nothing ever produced it: slot arity and undeclared-slot violations arrive through the compiler's :structure stage and have always carried :assignability. Pass :assignability where you passed source: :arity; the finding's {:slot, block_id, slot_name} anchor, and so where it renders, is unchanged.
  • The canvas toolbar's "Cancel insert" button. Leaving an insert is the palette's Cancel, beside the line that names the slot the next pick fills, or Escape.

0.5.0 2026-08-30

0.5.0 is the release where the editor behaves like the spike it was drawn from. It bounds its own height and hands scrolling to the panes; every zoom step and both fits scale the canvas for real; an arranged container's lanes size to their own content, so connectors stop crossing sibling cards; inserting visibly arms the gap the pick will land in, and a palette entry can be dragged onto one; an unresolvable block's card is compact again, with its findings and stored config moved to the inspector; the core block types summarise themselves on a card's second line; nesting depth is banded across the canvas; the plain controls render as buttons; and a host's icon component is rendered as a function component. Hosts: see Changed - .sb-node__raw-config is now .sb-inspector__raw-config, the editor's buttons carry a new sb-button class, and two band tokens (--sb-band-even, --sb-band-odd) join the tier-2 theming surface.

Added

  • A palette entry can be dragged onto a gap on the canvas to insert a block of that type there. The slots that accept the dragged type highlight as soon as the drag starts, exactly as they do when a card is dragged, and the drop produces the same insert a "+" and a pick produce.
  • --sb-editor-height bounds the editor: set it to a length and the panes scroll in their own boxes while the drawer stays pinned at the bottom, instead of the whole document growing the host page. The default is auto, so an editor nobody bounds is unchanged.
  • Two theming tokens for the editor's nesting bands, --sb-band-even and --sb-band-odd (tier 2). Both default to a surface the theme already carries, so a host that restates --sb-bg and --sb-bg-sunken in its own palette bands the canvas without setting either one.
  • Block types may export an optional summary/1, returning nil, a short string, or a list of chips, which the editor draws as a card's second line when the author has not named the block. It is read through StatifierBlocks.BlockType.summary/2, which normalises every shape to a chip list and drops an over-long chip rather than truncating it.
  • core.parallel, core.wait, core.on_event, core.send and core.branch summarise themselves on the card: lane names, timer <duration>, the outcome and the event, the event, and N arms + otherwise.
  • sb-button, one class carrying the editor's button look, so a host restyling the family changes one selector rather than seven.

Changed

  • The measurement hook's payload carries the canvas panel's usable box under a viewport key, read from the element the editor stamps data-sb-anchor="viewport". A host that registers the hooks from the package's default export needs no change; a host that reimplemented the hook against the documented payload should send the new key for the fits to resolve to a number.
  • Every slot on the canvas carries data-sb-depth, its root-relative nesting depth, and the stylesheet paints a full-width band per nesting level, alternating by the depth's parity. Depth 0 is deliberately unbanded, so the canvas keeps its own dotted ground.
  • The interrupt-rules rail has a ground of its own, in the warning family its dashed edge already uses.
  • StatifierBlocks.ViewModel.Node carries a summary field, and ViewModel.subtitle/1 returns it for a block whose title is its type's. A block the author has named still reads its type's label there.
  • An unresolvable block's card reads its type and one short reason; its findings and its stored config moved to the inspector, so the card is the same width as its siblings.
  • The inspector's Block section shows an unresolvable block's stored config as canonical JSON, wrapping mid-token rather than spilling past the pane.
  • The stored-config <pre> moved from the card to the inspector, and its class with it: .sb-node__raw-config is now .sb-inspector__raw-config. A host styling the old class should restyle the new one.

Fixed

  • A slot that refuses the block being dragged no longer accepts a drop when it sits inside a slot that accepts it. The gaps in the refused slot were live targets, and dropping on one put the block in the slot that had said no.
  • Every zoom step in the editor toolbar now scales the canvas, and the panel scrolls the scaled size rather than the unscaled one.
  • Fit width resolves to the largest zoom step at which the chart fits the canvas panel, instead of only recording that the fit was asked for.
  • Fit active resolves to the largest zoom step at which the selected block fits, and scrolls that block to the centre of the panel.
  • A host's icon component is rendered as a function component rather than applied to a bare map, so it receives a tracked assigns map and may use Phoenix.Component helpers such as assign/3 and assign_new/3. A host that worked around the old behavior by adding __changed__ to the assigns itself no longer needs to; the component must still return a ~H template, which it always had to.
  • Connector edges no longer cross sibling cards: an arranged container's lanes size to their own content, so a nested arrangement wider than one lane no longer overflows into the lanes beside it (ADR-0005 decision 10b).
  • Clicking a gap's "+" now visibly arms that gap, and the palette says which slot of which block the pick will land in, with a Cancel beside it and Escape as the way out.
  • A palette pick made with nothing armed says why it did nothing instead of failing silently.
  • A palette narrowed to the types a slot accepts now looks different from a palette that simply holds that many types.
  • The editor's plain controls - Undo, Redo, Cancel insert, the zoom steps, Fit width, Fit active, and a list field's add/remove - render as buttons rather than as whatever the host's browser paints, with a hover, a muted disabled state and an accented pressed state.

0.4.0 2026-08-30

Added

  • The inspector's Config tab is two labelled sections. Block states the selected block's type label, its id and the slot it sits in, and renders with nothing selected too - three rows in the same place, reading as a dash.
  • StatifierBlocks.Shell.slot_label/2 answers which slot a block sits in, by the slot's label rather than its name, with "root" for the document root.
  • A block type may declare a :string config field keyed label, and the editor draws that value as the card's title with the type's own label as a subtitle underneath - so a host's steps read as the names an author gave them without the type ever being hidden.
  • A card carrying invoke_type in its config draws it in mono on a third line, which is the fact an author checks most on a step that calls out to a handler.
  • StatifierBlocks.ViewModel.title/1 and subtitle/1 answer what a card's two name lines say, and ViewModel.Node carries title and invoke_type for a host rendering its own cards.
  • The compiler refuses a core.subchart whose chart names the document the block sits in, as an :emit-stage :self_reference finding against that block: a document cannot run itself. A cycle through two or more documents needs the host's document graph and stays the host resolver's to refuse.
  • StatifierBlocks.Shell.drawer_tabs/0, drawer_tab/1 and drawer_title/1, in the same shape as the inspector's tab helpers - an unknown tab from a crafted phx-value-tab resolves to the first one.
  • .sb-findings__row, __severity, __subject, __label, __id and __message, the row's parts. The severity colour stays on .sb-finding and its severity modifier, so a host that had restyled one severity keeps that styling with no edit. No new custom property.
  • A slot's header shows the condition it is subject to: the expression source, read-only, in a monospaced chip under the slot's name, clipped to one line with the whole of it in the chip's title. A branch's arms on the canvas now say what picks between them instead of only naming themselves.
  • StatifierBlocks.ViewModel.Slot.condition carries that source. It is derived from the container's own :expression config field keyed by the slot's name, read through the field's declared value_path, so a host block type that guards a slot the way core.branch guards an arm gets the same chip without the editor learning either type's name.
  • The palette and the inspector render as framed panes with a header row. The palette's names the pane and carries a chevron that folds it to a rail, giving its width back to the canvas; the inspector's names the pane and states its subject on the right - the selected block type's label, or no selection.
  • --sb-palette-collapsed-width, the width the folded palette narrows to (tier 2, default 2.25rem).
  • The palette carries a count line under its search box: the size of the palette when nothing is filtering it, and how much of it is left plus what is doing the narrowing when something is - a query, or the acceptance set of the slot the palette was opened from.
  • Each group header carries the number of rows currently under it, so a filtered section says how much of it survived the filter.
  • The editor draws the join marker under a container whose slots sit side by side, reading the words the block type's join_label callback returns - core.parallel completing on its first lane says "continue at first" - and falling back to the editor's own word for a type that declares none.
  • StatifierBlocks.ViewModel.Node carries join_label, the normalized words that callback returned for the block's config, or nil when it declared none.
  • The editor's canvas is a named panel: the toolbar is its header row, with a Canvas label, a nested tree chip, one segmented zoom control, and the depth and block-count metrics as right-aligned chips.
  • The canvas sits on a bordered, dotted ground, and --sb-canvas-grid is the tier-2 token a theme sets to change the dots' colour and spacing together.
  • The editor draws a ONE OF pill above a container whose body slots are alternatives and an ALL OF pill above one whose lanes run concurrently, which is the only place that distinction is stated on the canvas.
  • StatifierBlocks.ViewModel.arrangement/1, body_slots/1 and fan_label/1 derive how a container arranges its body slots, shared by the renderer and the connector geometry so the layout and the lines cannot disagree.
  • StatifierBlocks.Connectors.fan_anchor/1 and join_anchor/1 name the two markers a fan leaves from and rejoins at, so an edge is no longer drawn through the words that say what it means.
  • --sb-card-width sets the width of a block's card, which is what keeps the measured connector geometry from collapsing every edge onto one spine.

Changed

  • Configuration's empty state is a box standing where the form stands, saying what selecting a block would let the author do, rather than the one-line sentence the other tabs use for having nothing to read.
  • A required field is marked with the word Required beside its label instead of an asterisk on the end of it - the asterisk needed a legend the editor does not have and is read aloud as "star".
  • The Findings tab's count is a pill in the error hue rather than a tinted rectangle, and it is still the block's own findings, never the subtree's.
  • The delete control on a card is revealed on hover, on keyboard focus and on the selected card, and is hidden at rest. It is still in the DOM and still focusable, so the keyboard path is unchanged.
  • The card title reads as a title rather than as a native button, and the count badge, the subtitle and the invoke line are placed by a grid on the card's chrome.
  • The per-block-type accent stripe is drawn on cards whose type declared an accent_token and on no others. A type that declared nothing keeps a plain card; its icon tile is unchanged.
  • The document-level findings list is the drawer's Findings tab, beside Truth tables, and no longer a block of text under the canvas (operator ruling R4, 2026-08-29, under ADR-0005 ruling 1A: a list of findings is a grid of rows about the whole document). Each row carries the finding's severity, the block it is about - label and id - and the message, and clicking one selects and reveals that block. The inspector's Findings tab is unaffected and stays the selected block's findings (3A), as do the per-card counts.
  • The collapsed drawer strip reports the active tab. An author who has not picked a tab gets the first one that holds something, so a document with findings and no fixtures reads Findings 4 rather than Truth tables 0; once a tab is picked the pick stands. A host swapping the open document resets the pick along with the drawer's open state.
  • StatifierBlocks.Shell.drawer_view/1 accepts :tab, :findings and :orphan_findings, and its result gains tab, tabs, findings and orphans. title and count now describe the active tab rather than the truth tables specifically; the truth-table status values are unchanged.
  • StatifierBlocks.Editor.Findings.findings/1 takes findings, orphans, root and target instead of view_model, and renders the tab's panel rather than a headed section: the tab is the heading.
  • Slot labels are small, uppercased and letter-spaced - the treatment the fan pill and the join marker already carry for chrome that labels a structure. The transform is presentation only: the string a block type declares for a slot is unchanged, and every other reader still sees it as written.
  • Concurrent lanes carry a rule in the block accent across the top of each lane's header, drawn off data-arrangement="lanes". The pill above says ALL OF once; the rule is what carries that distinction down a document taller than one screen, where a set of lanes and a set of branch arms otherwise look alike.
  • An interrupt rail's dashed edge and its heading take the colour the connector layer already draws an interrupt edge in, so the rail and the edge leaving it read as one thing.
  • StatifierBlocks.Editor.PaletteBrowser takes a collapsed attribute, and the editor answers a palette-collapse event with one boolean and no hook, in the same shape as the shell amendment's other gestures. The fold is not reset when the host swaps the open document: it is a preference about the pane rather than state about the document.
  • The narrow arrangement (ADR-0005 ruling 7A) is unchanged. Below a container width of 780 the strip and its sheet are still the palette's whole chrome and the pane header stands down, so the fold has nothing to do there.
  • A container draws a box around its body only when it is a boundary - a container with a slot in the rail partition (ADR-0005 decision 10c, as amended by 10h). Every other container draws none: its own card stays at the head of its body and its children sit under it with the connectors, where a box around each of them turned a deep document into nested rectangles.
  • A boundary's box is the border, the radius and the inset that enclose its body and its rail, rather than the border colour it was before.
  • A container's card carries its own border, its accent stripe and its selection ring, so the block is still a card on the canvas when the box around its subtree is gone. A leaf card is unchanged.
  • A palette row is now a tile, a name, and the type's description on a second line, at a row height that gives the description room to wrap. The tile is a slot rather than an icon: a block type that declared no icon still renders the box, so every name in the list lines up.
  • A row's accent moved from a stripe down its leading edge onto the tile, which is where the card the pick produces carries it. The stamp itself is unchanged - a type that declares no accent token still gets neither the attribute nor the custom property.
  • A core.wait mints its delayed send under the reserved send role, so a chart containing one now compiles to s_<block id>__send where it compiled to s_<block id>__timer.
  • The package's default JavaScript export now carries both hooks, so hooks: { ...StatifierBlocks } registers StatifierBlocksDrag and StatifierBlocksMeasure together - a host that registered only the drag hook got an editor with no connectors and no error explaining it. Both names are still exported individually, and statifier_blocks/measure still resolves for a host that wants measurement alone.
  • The canvas stage renders inside a .sb-canvas-panel element, which is now the scrolling box; #sb-canvas stays the drag hook's element, the stage anchor, and where the theme assign's declarations land.
  • A container with more than one body slot now lays its slots out side by side and fans into them, as a container declaring layout: :columns already did - a branch's arms no longer stack full-width with each fan edge running down through the arm above its target.
  • Columns are a CSS grid taking their natural heights, cards are a fixed width centred in the box they sit in, and a column's header is card-width and centred over the first card it governs.
  • The "+" between two blocks is now the insertion marker: subtle at rest, highlighted on hover, on keyboard focus and for the whole of a drag, and drawn as a placeholder ring in a slot that is still empty. It is the same button with the same events, so nothing about the keyboard path changed.
  • StatifierBlocks.Editor.Slot stamps data-empty, and StatifierBlocks.Editor.BlockNode stamps data-container and data-arrangement, on the markup a host may style against.
  • The collapsed drawer's strip reads as a label and a quantity: a small-caps title with letter-spacing, and the count as a chip carrying a bare number rather than a parenthesised one inline with the title.

Removed

  • The stub exit tick a rail drew below itself in CSS. The exit edge is measured and drawn now, and a fixed-length mark beside it was a second claim about the same thing that pointed somewhere else.

Fixed

  • Leaving the scope around a core.wait cancels the wait's delayed timer, so a wait abandoned before its duration elapses no longer leaves an armed timer behind in a durable host.
  • A core.subchart outcome-routing condition is no longer attributed to the outcomes config field, so a chart finding landing inside it reports fault: :package with no config_key rather than blaming the author for bytes the compiler composed.

0.3.0 2026-08-29

Charts get more shapes to compile into. Campaign 015 adds four emitters to the core.* vocabulary - core.subchart, which runs another chart and routes on the outcome the child reported; core.foreach, a container whose body runs once per item of a datamodel list; scope-correct cancellation for a delayed core.send; and core.parallel's complete: "first", which finishes the block at the first lane's completion and exits the losing lanes. The compiler gains two root-document options, terminate: and declare:, a typed datamodel index that refuses a document reading a path the host declared sensitive, and predicator duration strings wherever a duration is typed. In the editor, connectors graduate out of the spike, a default icon set ships so a host needs no asset pipeline to get one, and the shell is laid out as the arrangement record describes it. This is a minor bump because every block's conventional <final> moves from s_<block>__done to s_<block>__o_done: the compiled bytes of every outcome-bearing chart move with it, so a host that stores compiled charts or provenance maps recompiles them, and a chart-level position saved against the old bytes no longer resolves.

Dependency floor: unchanged - statifier ~> 2.2 and predicator ~> 9.0.

Added

  • A core.subchart block type: a step that runs another chart and waits, routing done.invoke on the outcome the child reported (_event.data.outcome) to one slot per declared outcome, with core.invoke's on_error slot unchanged for a failed invocation.
  • A core.foreach block type: a container whose body runs once for each item of a datamodel list, compiled as a plain SCXML loop - a per-loop cursor and a snapshot of the list taken once on entry, the item_as and index_as bindings re-assigned from that snapshot on each pass, and the body compiled once with an internal loop-back transition. Iteration ends on the out-of-bounds read, snapshot[cursor] === undefined.
  • core.parallel accepts a complete config key choosing when the block is done: "all" (the default, and what a document stored before the key reads as) keeps the shipped rule, and "first" compiles the racing rule - one transition per lane on the <parallel> element itself, taken on that lane's own completion event and targeting the block's done final, so the block finishes at the first lane's completion and the engine exits the losing lanes with their <onexit> content and one CancelInvoke per live invocation.
  • Block types may declare outcomes/1, an optional callback returning ordered {name, label} pairs for the ways a block can finish; a type that does not export it has exactly one outcome, done, and behaves as before.
  • A child summary in the compiler context carries an outcomes field: the child's declared outcomes in declaration order, each with the <final> it compiled to and the completion event a parent wires on. It is never empty.
  • StatifierBlocks.Compiler.compile/3 accepts terminate: true, which emits one top-level <final> per root-block outcome with no <donedata>, so a compiled root document reaches :done when its root block completes; without it a root document never terminates, and passing it together with child_use: true is refused with an :emit finding.
  • StatifierBlocks.Compiler.compile/3 accepts child_use: true, which compiles a document for use as another chart's child: the emission gains one top-level <final> per outcome the root block declares, carrying that outcome name as done data, so a parent session can see which way the child finished.
  • StatifierBlocks.Compiler.compile/3 takes a :declare option - a list of {id, expr} pairs - so a host can declare the <data> roots its root document assigns to and guards on, hoisted ahead of block-declared roots into the chart's single <datamodel>.
  • Compiler.compile/3 takes a :datamodel option and refuses a document that reads a path the host declared sensitive?: true into a position the chart evaluates against the datamodel; with no datamodel supplied nothing is produced.
  • A block type may now contribute declared <data> roots to the chart: StatifierBlocks.Compiler.DeclaredRoots.declare/2 emits a declaration among the block's own children and the compiler lifts every one of them into a single top-level <datamodel>, in document order. A document that declares no roots emits no <datamodel> element, so charts compiled before this change are byte-identical.
  • A new Emit-stage finding, :duplicate_binding: a declared root whose name a block it sits inside already declares is refused against the declaring block and the config field the name was typed into, because early binding makes both roots global and the inner one would silently overwrite the outer.
  • The compiler refuses a block type that declares a malformed or duplicated outcome name with an :invalid_outcome Emit finding, against the block whose type declared it.
  • A chart-stage finding for an expression the author typed carries config_value_span, the byte range of the offending sub-expression within that config value, so an editor can underline the sub-expression rather than the whole field.
  • StatifierBlocks.Predicates.Datamodel indexes a datamodel document - the typed, three-scope declaration sb ADR-0006 defines - into a path/entry index: the type of a path, the entries under a prefix, whether a path is declared, and the record's one total derivation of the declared-path set. The index is advisory; an undeclared path is unknown, never wrong.
  • StatifierBlocks.Datamodel.declared_paths/1 accepts such a document as a fourth shape, alongside nil, a list and a MapSet, and projects it through that derivation. A document declaring no entries normalizes to the empty set - a host claim - rather than to nil, which stays reserved for no datamodel at all.
  • The editor renders the arrangement ADR-0005's shell amendment records: a palette, canvas and inspector across three columns with a full-width drawer row beneath them, at container-query breakpoints of 1280, 1024, 900, 780 and 640.
  • A canvas toolbar with stepped zoom, Fit width, Fit active, and the document's block count and depth.
  • The inspector is tabbed - Config, Findings and Condition - where Findings is the selected block's own and Condition reads the per-arm predicator source.
  • A drawer that is never open-or-gone: collapsed it is a strip carrying a title and the document's table count, and opened on a block with no table it shows an index of the blocks that have one.
  • fixtures, an assign carrying %{block_id => [TruthTable.t()]}, is what the drawer's truth-table tab reads; with none supplied the drawer is still present and reads 0.
  • drawer_height and on_drawer_resize are the host's seam for the drawer's resizable height, which the host remembers per viewer.
  • A :header slot the host fills with the outer header - document identity, the switcher, the theme control, compile and publish - which this package now explicitly does not draw.
  • Below 780 the palette collapses to a strip that opens as a sheet.
  • StatifierBlocks.Shell exposes the shell's arrangement as pure functions - the zoom ladder, the document metrics, the drawer's five states, and which of a block's fields are conditions.
  • Three tier-2 theme tokens: --sb-palette-width, --sb-inspector-width and --sb-drawer-height.
  • The editor draws connectors. Adjacency inside a slot, a container's entry, the fan and rejoin around a container arranged side by side, and a rail's exit are rendered as SVG in the LiveView tree, derived from the document's shape rather than authored.
  • A second JavaScript entry point, statifier_blocks/measure, exporting the StatifierBlocksMeasure LiveView hook. Its whole job is measurement: after a render it reads the boxes the browser laid out for the anchors the server stamped and pushes them, and it issues no commands and mutates no DOM. A host that wants connectors adds one import; a host that does not gets the editor it had before, minus the drawn connectors.
  • StatifierBlocks.Connectors: the connector geometry as pure functions from measured rectangles to SVG path data, outside the Phoenix guard, so a host can route its own connectors and a test can assert them without a browser.
  • Three --sb-* tokens for the connector layer: --sb-edge, --sb-edge-interrupt and --sb-edge-width.
  • StatifierBlocks.Editor.Icons, a default icon set the editor uses when the host passes no icon component. Inline SVG for the eleven names the core block types declare, with no font, no CDN and nothing to register in a host's asset pipeline. Every glyph paints with currentColor and fills its tile, so --sb-block-accent and --sb-block-accent-tint still decide the colour and a per-block-type accent_token still moves a type's tile with its stripe.
  • Palette entries render their icon. The icon assign the editor passes the palette browser was declared and never rendered, so no host could put an icon on a palette row; a type now looks the same in the palette as on the card the pick produces.
  • A slot declaring slot_style: :failure renders in its own vocabulary - a solid error-family edge, its own sb-slot--failure class, and an ordinary flow edge where an interrupt rail draws a dashed escape.
  • StatifierBlocks.ViewModel.exit_edge/1 says which edge vocabulary a slot's exit is drawn in, and the editor stamps it as data-exit-edge.
  • The editor takes an optional datamodel assign - the paths the host declares - and reports a config field whose declared datamodel path is not among them as an :info finding in the findings panel; with no datamodel supplied nothing is produced.
  • Block types may declare a config field with datamodel_path?: true, saying its value is a path into the host's datamodel; core.assign's path field carries it.
  • StatifierBlocks.Core.Parallel.join_label/1, declared on the type's palette entry, so a renderer draws "continue at first" or "continue when all" from the block's config without learning the type's name.
  • StatifierBlocks.DurationInput reads a typed duration for that control, accepting exactly what StatifierBlocks.Core.Duration compiles and naming the limit a refused value hit.

Changed

  • Every block's conventional <final> moves from s_<block>__done to s_<block>__o_done and now raises done.outcome.<state id>.done on entry, so compiled SCXML moves for every document; a host that stores compiled charts or provenance maps recompiles them, and a chart-level position saved against the old bytes no longer resolves.
  • Compiler.compiler_version/0 (and every compilation record's compiler_version) moves to 0.3.0 with the package, per ADR-0004 decision 6, and is the third input to the byte-determinism guarantee: this release is where the outcome-final byte movement is recorded.
  • core.send now emits its <send> with id="<the block's state id>__send", so a delayed send can be named after it is armed.
  • A delayed core.send is now cancelled by its scope: the compiler emits <cancel sendid="..."/> in the <onexit> of the nearest enclosing <state>, so a pending send does not outlive the sequence, group, region or lane that armed it. Charts containing a delayed core.send change bytes; every other chart is unchanged, core.wait timers included.
  • core.wait accepts a predicator duration string (1h30m, 2d, 3d8h) as well as ISO-8601, stores whichever spelling the author typed, and compiles it to the emitted delay attribute; its refusal message names both spellings.
  • core.wait's declared duration default is now the predicator string 1h rather than the ISO-8601 PT1H, so a newly inserted block starts from the spelling an author types. Both spellings stay accepted and each compiles to the same delay attribute, so no chart's emitted SCXML changes and no stored document has to be retyped.
  • A :duration config field renders as one text control taking predicator duration strings, with 30s, 15m, 1h30m, 2d and 3d8h shown beside it; ISO-8601 is still accepted and the author's string is stored verbatim.
  • An empty :duration field omits its config key rather than storing an empty string, so a cleared field and a never-set field are the same value.
  • An icon entry that declares no icon renders no tile, rather than an empty one, and a host's icon component is never called with a nil name.
  • An icon name the shipped set does not have renders a neutral mark with the name in data-icon.
  • A slot_style this editor does not recognize renders as an ordinary body slot instead of reaching the markup unresolved; its children are still rendered, still selectable and still saved.

Removed

  • StatifierBlocks.Editor.Field.units/0, format_duration/2 and parse_duration/1, which served the retired value/unit control. Call StatifierBlocks.Core.Duration.to_iso/1 to canonicalise a stored duration.

Fixed

  • The editor no longer renders a U+25A1 white square in every icon tile when the host passes no icon. Passing one still overrides every tile, on the canvas cards and the palette rows alike.
  • A delayed core.send in the body of a core.group that carries interrupt rules is now cancelled when the group is abandoned. Its <cancel> is emitted in the body region's <onexit> rather than the group's own, and abandoning the group exits the region without exiting the group, so the old placement never fired.

Known limits

  • A core.foreach list holding a nil item iterates to its end rather than stopping at it: === is strict, so only an out-of-bounds read is undefined.
  • Two core.foreach blocks in one document may not bind the same name, even when neither is inside the other; the second is refused with a duplicate-id finding on its item_as field.

0.2.0 2026-08-29

The editor ships. Campaign 014 graduated the authoring spike into the package: StatifierBlocks.Editor renders from assets/ with a documented --sb-* theming surface, a drag marks the slots that accept a block and can say why a slot refused, and a host registers its own block types through Palette.from_modules/2. The core.* vocabulary grows by core.invoke, core.raise, core.assign and core.send; the compiler now evaluates predicator conditions, refuses slot-arity and undeclared-slot violations, and adapts its findings into the shape the editor renders.

Dependency floor: statifier ~> 2.2 (was ~> 2.0); predicator ~> 9.0 is now a direct dependency.

Added

  • core.raise raises an event for an enclosing group's interrupt rules.
  • core.assign writes a literal to a datamodel path.
  • docs/theming.md: the theming guide - the three tiers of the --sb-* surface, the scheme token, per-block-type accents, and a complete host theme that sets custom properties and nothing else.
  • core.send sends an event, now or after a delay.
  • A block type may declare slot_outcome_key in its palette entry, naming the config key the blocks in one of its slots carry their outcome under, so a renderer can route an interrupt rule's escape without branching on a type name; the declaration reaches the view model as Slot.outcome_key and the resolved value as Node.outcome.
  • StatifierBlocks.BlockType.slot_outcome_key/2 and StatifierBlocks.BlockType.outcome_name/2 read that declaration totally: a malformed declaration or value is refused rather than repaired, and reads as no declared outcome.
  • A block type may declare accent_token, the NAME of a --sb-* property, and the editor stamps it on that type's cards and palette rows. Two rules in the stylesheet read it; no rule and no module names a block type (ADR-0005 amendment 14d, consumption side).
  • StatifierBlocks.Finding.severity_class/1, and :info as a third severity for advisory findings (decision 11, amended 2026-08-29). Nothing emits one yet; only :lint may.
  • A form whose config the gate has not accepted names the fields that are outstanding, says why nothing is stored, and offers "Discard edits". A draft was never a command, so it cannot be undone - it can only be thrown away, and that gesture had nowhere to live.
  • :expression and :duration controls carry a placeholder. A bare :string still carries none: there is nothing a type that wide can suggest.
  • A theme audit test over the stylesheet, failing in both directions: a var(--sb-*) with no declaration, and a declared token no rule reads (14e).
  • StatifierBlocks.SlotValidation, a palette-aware whole-document check for a block's declared slots (:undeclared_slot) and each declared slot's arity (:slot_arity_violated).
  • StatifierBlocks.Predicates evaluates a condition expression against a binding context through predicator, returning a boolean or a tagged error.
  • StatifierBlocks.Predicates.TruthTable builds a checked truth table over fixture rows, applying first-match-wins arm ordering.
  • StatifierBlocks.Finding.from_compiler/2 and from_compiler_all/2 adapt a compiler finding into the presentation shape the editor renders, so a host can route compile findings through ViewModel.build/3.
  • StatifierBlocks.Assignability.seam_reason/4, finding_reason/2 and seam_reasons/3 name why a data-flow seam came out the way it did: :not_assignable and {:fixable_by, block_id} for a refusal, and :source_untyped / :target_untyped / :both_untyped for a seam that passed only because a block declared no type. seam_reasons/3 is how a host finds the parts of its palette it has not typed yet.
  • StatifierBlocks.Assignability.target_verdicts/4 returns every position valid_targets/4 enumerates with its full verdict, and StatifierBlocks.Edit.Targets.slot_verdicts/3 projects those to slots - the accepting ones and the reason each refusing one gives.
  • The editor stamps a refused slot's reason as data-drop-reason beside data-drop, so a hover affordance can explain a refusal with no round-trip and no JavaScript.
  • A core.invoke block type: it names an invoke type for the host to run, sends datamodel values along as <param>s, writes the result where its assign_to names, and takes an optional on_error subtree entered on a permanent invoke failure.
  • StatifierBlocks.Compiler.Context.outcome_id/2 and outcome_event/2, for a block type with more than one way to finish: one <final> per outcome, and the done.outcome.<state id>.<outcome> event a parent wires on.
  • StatifierBlocks.Palette.from_modules/2, the registration API a host uses to contribute its own block types: an ordered, explicit list of {type_name, module} registrations, with core: true to sit on top of the core.* vocabulary. Later entries win. It is still a value - no global registry, no application-configuration lookup, and no discovery pass.
  • A palette entry may declare badge, a short chip for the card header, and join_label, a one-argument function of the block's config phrasing the join marker under a side-by-side arrangement (ADR-0002 amendment B). StatifierBlocks.BlockType.badge/1 and join_label/2 read them.
  • Both readers are total and refuse rather than repair: a chip that is blank, carries a newline or tab, or runs past 24 characters is dropped, not clipped, and a join_label that raises degrades to the editor's own word rather than taking the canvas down.
  • accent_token, badge and join_label are admitted keys of StatifierBlocks.BlockType.palette_entry/0.
  • The README carries a worked host example - a myapp.risk_hold block type registered beside the core vocabulary, with a badge and an accent token - and it is executed on every build rather than trusted.

Changed

  • Compiler.compiler_version/0 (and every compilation record's compiler_version) moves to 0.2.0 with the package, per ADR-0004 decision 6: a record compiled by 0.1.0 identifies itself as such.
  • --sb-drop-ok-border moves from #2f9e5f to #2c945a. The outline that says a slot accepts a drop was 2.93:1 on the sunken surface, under the 3:1 a mark carrying information is held to; the tint follows it.
  • predicator is now a direct dependency (~> 9.0), because StatifierBlocks.Core.Duration calls Predicator.Duration.parse/1. It already resolved transitively through statifier, so the resolved version does not move; naming it records the call.
  • The editor's stylesheet carries a scoped reset, and every selector in it matches the container through :where(.sb-editor) so a component rule always wins (ADR-0005 amendment 14b).
  • --sb-color-scheme is declared and read as color-scheme on the editor's own container, so the parts of a control the browser paints - a <select>'s drop-down, the scrollbars, the caret - follow the theme (14a).
  • The --sb-* surface gains the space, type and shape scales, a third text step, a strong border, status tints, the drag seam's drag-time height, and the canvas sizing constants that were literals in rules.
  • A :secondary and a :failure slot are both placed as attached rails, and a container declaring either is drawn as a boundary box - the rail partition, not the :secondary partition (amendments 10c and 10h).
  • --sb-drop-no-opacity is retired. A drag now marks the slots that accept the block and leaves the rest alone rather than dimming them; the disabled-control opacity it doubled as is --sb-disabled-opacity.
  • The compiler now refuses a document whose slots violate their declared arity or name a slot the block type does not declare, instead of silently dropping those children from the emission.
  • Reasons change no verdict: :unknown stays permissive in both positions, Assignability.validate/3 reports exactly the findings it did before, and neither finding tuple gained a field.
  • slot_style admits a third value, :failure, for a slot whose children are an in-band continuation taken on a bad outcome; core.invoke declares it for on_error.
  • The role namespace beginning o_ is reserved for outcome finals; Context.role_id/2 now refuses such a role with a :reserved_role finding.

Fixed

  • The editor's root rule sets font-family rather than the font shorthand, so --sb-font reaches the editor. font: <family-list> is not a valid shorthand, so the whole declaration was dropped and the editor's text did not inherit the host page's font as the token promised.

0.1.0 2026-08-27

First release: the authoring layer above the statifier statechart engine. A block document is the source of truth - a tree of typed blocks, each with a declared shape - and it compiles one way to SCXML plus a provenance map that points a runtime position back at the block that produced it. Block types are host-pluggable: a host registers the types its own domain needs, and the compiler and the editor work off that registry rather than a closed built-in vocabulary. The core.* structural vocabulary, the compiler, the edit algebra, and the LiveView editor shell all ship here.

The Changed entries below describe the shape of callbacks and metadata as they stand at this first release; there is no earlier published version to have changed from.

Added

  • StatifierBlocks.Document.validate/1 checks a block document's structure: schema version, envelope shape, per-block shape, and document-wide id uniqueness.
  • StatifierBlocks.Document.to_json/1 encodes a document to ADR-0001's deterministic canonical JSON: sorted object keys, no insignificant whitespace, empty slots/config/metadata omitted, no floats.
  • StatifierBlocks.Document.content_hash/1 returns a "sha256:" <> hex document identity over to_json/1's canonical bytes.
  • StatifierBlocks.Document.from_json/1 decodes canonical JSON back into a document, structurally and registry-free: unknown block types decode successfully, and every refusal is one of ADR-0001's typed error arms.
  • StatifierBlocks.BlockType behaviour: the nine-callback authoring-time extension seam (ADR-0002), five required (slots/1, config_schema/1, validate_config/1, current_version/0, emit/2) and four optional (io/1, migrate_config/2, fixtures/0, palette_entry/0).
  • StatifierBlocks.Palette: a caller-supplied type_name => module value (ADR-0002 decision 2), with new/1 to build one and a total fetch/2 that returns {:ok, module} or {:error, {:unknown_block_type, type_name}} and never raises (ADR-0002 decision 3).
  • StatifierBlocks.Palette.resolve/2: resolves a block through the palette and migrates its config in memory when the stored type_version is below the type's current_version/0 (ADR-0002 decision 8). Migration is applied to the returned struct only and never written back to a document; a stored version above current_version/0 hard-errors as {:error, {:block_type_too_new, id, version}} rather than reading best-effort, and a failing or missing migrate_config/2 surfaces as {:error, {:migration_failed, id, reason}}.
  • The core.* structural block types (ADR-0002 decision 10), one StatifierBlocks.BlockType module each: StatifierBlocks.Core.Sequence, .Group, .Branch, .Parallel, .Wait, .ResumableGroup and .OnEvent. core.branch derives one slot and one :expression field per declared arm, core.parallel one slot per declared lane, and each type is the authority on its own config through validate_config/1.
  • StatifierBlocks.Palette.core/0 and StatifierBlocks.Palette.core_types/0: the core vocabulary as a palette, and as the plain type_name => module map a host merges its own entries into.
  • Structural placement through ADR-0003 decision 3 kind tags: core.on_event declares kinds: [:interrupt_handler] and the group types accept only that kind in their interrupts slot, so an interrupt handler is admitted there and refused everywhere else, and an ordinary step is refused there - in both directions, from the declarations alone, with no special-cased rule.
  • fixtures/0 on core.branch (an arm condition evaluated against two datasets) and core.on_event (one example event payload). The bundle shape follows an amendment to ADR-0002 decision 9 that is not yet accepted, and is documented as provisional until it is.
  • StatifierBlocks.Assignability: the one decision function for whether a block may land in a slot, checking structural admission by kind tag and data-flow compatibility by type-expression identity plus an optional host-supplied widening relation (ADR-0003). check/5 decides a single candidate position; valid_targets/4 lists every position a candidate may occupy in a document; validate/3 reports every finding already present in a document; inbound_type/4 and assignable?/3 are the two primitives both are built from.
  • StatifierBlocks.Assignability.Relation: the behaviour a host implements to widen data-flow compatibility beyond exact type-expression identity. A host module can only grow the accepted set, never shrink it.
  • StatifierBlocks.Palette gains an assignability field naming the host's Assignability.Relation module, set via Palette.new(types, assignability: MyApp.Blocks.Types). Defaults to nil, meaning no widening relation is declared; existing calls to Palette.new/1 are unaffected.
  • StatifierBlocks.Compiler: the one-way compile (ADR-0004 decisions 1-4, 6-7). compile/3 is a total function of {document, palette} returning {:ok, %StatifierBlocks.Compiled{}} or {:error, [%StatifierBlocks.Compiler.Finding{}]} - no process state, no clock, no IO, and no arm that raises. The pipeline runs Document, Resolve, Config and Emit, stopping at the first stage that produces errors and reporting every error from that stage.
  • StatifierBlocks.Emission: the structural representation of one SCXML subtree a block type returns from emit/2, with element/3 and the child_ref/1 placeholder the compiler splices its children into.
  • StatifierBlocks.Compiler.Serializer: the deterministic serializer. Attributes sorted, one canonical empty-element form, no incidental whitespace at all. It is identity-bearing code - chart identity hashes source bytes (st-ADR-0052) - and serializer_test.exs now enforces the whitespace sensitivity ADR-0004 decision 6 named and left unenforced.
  • StatifierBlocks.Compiler.StateId: state_id/1, state_id/2, unstate_id/1 and done_event/1. State ids derive from block ids ("s_" <> block_id, "__" <> role for an auxiliary state), so they are unique, invertible and total over generated states.
  • StatifierBlocks.Compiler.Context: what a block type is entitled to know while emitting - its own ids, the document id, its children's summaries (block id, state id, done event) and the role-minting function. No palette, and no child's emitted SCXML.
  • StatifierBlocks.Compiled and StatifierBlocks.CompilationRecord: the artifact, and the join between document identity and chart identity. chart_name carries the document id and chart_version stays nil, so a revision bump or a metadata-only edit still matches the identity a running session holds.
  • StatifierBlocks.Core.Emit: the SCXML shapes the core.* vocabulary compiles to, and the builders a host block type follows to compose with them.
  • StatifierBlocks.Provenance: the map from generated SCXML back to the blocks that produced it (ADR-0004 decision 5). Keyed by state id for highlighting a running session's configuration, and by byte span for routing findings that carry no element reference. owner_at/2, owner_of_state/2, owners_of_states/2, and canonical to_json/1 / from_json/1 so a host can store the map beside the chart.
  • StatifierBlocks.Compiled now carries all five of ADR-0004 decision 1's fields: provenance, invoke_types and warnings join scxml and record.
  • invoke_types publishes the sorted set of invoke types the chart emits, unconditionally, so a host can compare it against its Statifier.Session registration at deploy time (ADR-0004 decision 8).
  • Compiler.compile/3 accepts :known_invoke_types, an opt-in lint that warns - never errors - for every emitted invoke type absent from the set the caller believes will be registered.
  • Compiler.compile/3 accepts :entry_type, ADR-0003 decision 4's caller-supplied context, which the new Structure stage passes to StatifierBlocks.Assignability.validate/3.
  • The compiler now runs a Structure stage (assignability) and a Chart stage (statifier's own pipeline over the generated bytes), and maps every upstream finding back to the block that caused it.
  • StatifierBlocks.Emission.attributed_to/2, from_config/2 and attribute_from_config/3: the hints a block type leaves so a finding lands on the block an author would recognise, and on the config field they typed into.
  • StatifierBlocks.Edit: the editor's command algebra - insert, remove, move, and update-config - as a purely structural, invertible rewrite over a document with no palette involved (ADR-0005). apply/2 applies one command and returns both the new document and the command that undoes it; check_config/3 is the separate config-validity gate one layer up.
  • StatifierBlocks.Edit.History: undo and redo over Edit commands. commit/4 is the one funnel a host calls - it runs Edit.check_config/3 before Edit.apply/2, then pushes the inverse and clears the redo stack, so invalid config never reaches the document on any path, undo and redo included.
  • StatifierBlocks.Edit.Targets: droppable_slots/3 and droppable_slots_for/3, which slots would accept a dragged block, at slot granularity rather than gap granularity, built as a reduction of StatifierBlocks.Assignability.valid_targets/4.
  • StatifierBlocks.Finding: the presentation finding ADR-0005 specifies, anchored to a block, a slot, or a config field so the editor knows where to render it. Distinct from the existing StatifierBlocks.Compiler.Finding, which serves the compile pipeline.
  • StatifierBlocks.ViewModel: the structure the editor actually renders, derived from a document, a palette, and a list of findings. Resolves and normalizes every block's slots, form fields, and palette presentation metadata, and routes every finding to the position that renders it.
  • StatifierBlocks.Editor: the LiveView editor shell (ADR-0005). A LiveComponent a host embeds over a %Document{} and a %Palette{}; it is the only stateful module in the package's rendered half, and everything it does is translate a phx- event into one of StatifierBlocks.Edit's four commands. Drag is two round-trips - one at dragstart to enumerate valid slots, one at drop - with zero per hover, because validity reaches the client as data-drop markup rather than as client-side logic.
  • StatifierBlocks.Editor.Canvas, .BlockNode, .Slot, .ConfigForm, .Field, .PaletteBrowser, .Findings: the function components the shell renders, each independently renderable in a test. BlockNode and Slot recurse into each other, and there is no per-block-type component: a block type's layout and slot_style presentation metadata is the only thing that distinguishes a group from a set of lanes.
  • assets/js/statifier_blocks.js: the package's entire client-side surface, one hook named StatifierBlocksDrag, shipped as source. A host adds "statifier_blocks": "file:../deps/statifier_blocks" to its assets/package.json and imports the hook in app.js; this repository bundles nothing and has no Node toolchain.
  • assets/css/statifier_blocks.css: one stylesheet of structural CSS and no visual opinion beyond it. Every class is prefixed sb-, every color, space, radius and drag treatment is a --sb-* custom property with a default, and every top-level component takes a class attr appended to its own.
  • A headless CI job that resolves the dependency tree with phoenix_live_view absent, compiles it with warnings as errors, and runs the non-LiveView suite
    • the acceptance property that makes the optional dependency's guard trustworthy rather than decorative. STATIFIER_BLOCKS_HEADLESS=1 reproduces it locally without disturbing the ordinary build.
  • StatifierBlocks.BlockType: a config field declaration may now carry an optional value_path, a list of keys and list indexes from the config root down to the value it edits (ADR-0002 decision 7, amended 2026-08-27). A declaration without one behaves exactly as before - its key addresses config[key]. The key remains the field's identity in both cases: the DOM id, the form param name, and what a {:config, block_id, key} finding anchors to.
  • StatifierBlocks.BlockType.value_path/1, fetch_value/2 and put_value/3: the reader and writer that resolve a declaration to a path and then read or write through it. value_path/1 answers [key] for a declaration that declares none, 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.
  • StatifierBlocks.ViewModel.Field carries value_path, and ViewModel.Field.value_path/1 reads it with the same [key] default.

Changed

  • StatifierBlocks.BlockType.io/1's return type is StatifierBlocks.Assignability.io/0 instead of term(). Every core block type already returns a value of this shape; a custom block type implementing io/1 should confirm its return value conforms.
  • All seven core.* block types implement emit/2 for real; the {:error, {:not_implemented, block_id}} placeholder and StatifierBlocks.Core.Config.emit_deferred/1 are gone.
  • StatifierBlocks.BlockType.emit/2 is narrowed from (Block.t(), term()) :: {:ok, term()} | {:error, term()} to (Block.t(), StatifierBlocks.Compiler.Context.t()) :: {:ok, StatifierBlocks.Emission.t()} | {:error, StatifierBlocks.BlockType.emit_error()}. A host block type that was returning something else now has a type to conform to.
  • StatifierBlocks.Compiler.Finding gains path, severity, fault and code. fault is :author when a document edit fixes the finding and :package when it is a bug in this package or a host's block type - which is what lets an editor say "this cannot be fixed here" rather than blaming the author for a generated state id.
  • Findings from every stage come back in document order over blocks rather than in the order a stage happened to collect them.
  • A bad :expression config field now surfaces as an :author finding naming the arm's config key, rather than as an unrouted upstream error.
  • StatifierBlocks.BlockType.palette_entry/0's return type is StatifierBlocks.BlockType.palette_entry/0 instead of map(). Every core block type already returns a value of this shape; a custom block type implementing palette_entry/0 should confirm its return value conforms.
  • phoenix_live_view is a declared optional dependency at ~> 1.0, matching statifier_ui's floor. Every module under StatifierBlocks.Editor.* is compiled behind Code.ensure_loaded?(Phoenix.LiveView), and no module outside that namespace references Phoenix - so a host that only compiles documents adds no Phoenix dependency and compiles no editor code.
  • The hex package's files: list includes assets. The hook and the stylesheet ship as source, and source that is not in the tarball is not public API.
  • Every illustrative example in the package - ADR worked examples, doc examples, and the shipped fixture bundles - uses one of the family's two canonical example domains: credit-card authorization and capture, or a signup wizard with A/B testing.
  • StatifierBlocks.Core.Branch.fixtures/0 ships budget-decision datasets ("approved" / "declined") and the expression "budget_remaining > amount". A host rendering the bundle in a palette panel sees those names.
  • The arm-slot and lane-name validation messages on core.branch and core.parallel name "arm_approved" and "capture" as their exemplars.

Fixed

  • A core.branch arm's condition is now readable and editable in the editor. Core.Branch.config_schema/1 keys one :expression field per arm by the arm's slot name, but the condition is stored at config["arms"][i]["cond"]; the form previously read and wrote the slot name as a top-level config key, so every branch condition rendered empty, no edit to one reached the arm, and a junk config["arm_approved"] accumulated beside it. Each per-arm field now declares value_path: ["arms", i, "cond"], and StatifierBlocks.ViewModel and StatifierBlocks.Editor.ConfigForm read and write through it. i is the arm's index in the stored list rather than its index among the well-formed ones, so a good arm below a malformed one still addresses its own condition while an author is mid-edit.
  • StatifierBlocks.Editor's field-list-add and field-list-remove events read and write the rows through the field's value_path as well, rather than through the top-level key. A key naming no field in the selected block's schema now edits nothing, matching the guard ConfigForm.decode/3 already applied.
  • StatifierBlocks.Assignability.check/5 and valid_targets/4 no longer raise a MatchError when the candidate is the document root. Document.fetch_path/2 answers {:ok, []} for the root, and the vacated-seam check now reads that as what it is - the root occupies no slot, so it leaves no seam behind - instead of calling List.last/1 on the empty path. StatifierBlocks.Edit.Targets.droppable_slots/3 answers [] for the root rather than crashing, so a caller no longer has to guard around it.