@ai-matrx/records-ui 0.6.0 → 0.9.0

package/CHANGELOG.md

@@ -1,5 +1,132 @@
 # Changelog — @ai-matrx/records-ui
 
+## 0.9.0
+
+**Sharing reaches the record and table screens (SCR-N-5).** `ShareControl` puts a
+Share button on the record peek and on the table page, and hands the subject to
+the host through a new `share` port on `<RecordsUiProvider>`.
+
+This package ships NO share dialog, deliberately, for the same reason it ships no
+chat: AI Matrx has exactly one sharing surface, and a second one here would give
+the platform two answers to "who can see this". With no `share` port bound there
+is NO BUTTON — not a greyed one, not one that opens an empty panel — and
+`NO_SHARE_REASON` names the port for anything that asks why.
+
+A Table is a record in this store, so one control, one subject shape and one
+dialog serve both screens; `kind` only changes the noun in the sentences.
+
+New exports: `ShareControl`, `useCanShare`, `shareUnavailableReason`,
+`NO_SHARE_REASON`, and the `ShareSubject` type.
+
+### Consumer action
+
+Bind `share` on your `<RecordsMount host={{ … }}>` with the app's own sharing
+dialog, e.g. in matrx-frontend:
+
+```tsx
+import { recordStoreShare } from "@/features/sharing/components/RecordStoreShareSurface";
+
+<RecordsMount host={{ Link, density: "condensed", share: recordStoreShare }} …>
+```
+
+Leave it out and nothing changes — the screens render exactly as they did in
+0.8.0, with no Share button.
+
+
+## 0.8.0
+
+**The worst finding of the 19 September verdict: "Add field" opened a proper
+panel and then refused everything.** Number and Date were refused as undeclared,
+Person was refused as a Formula, a choice list demanded a table the panel could
+not name, and before any of those the Key box was blank, unmarked and required.
+A table made with the product's own button had exactly one column.
+
+- **The panel collects an INTENTION, and the store decides the rest.**
+  `FieldEditor` writes through `custom.field_declare` (`useFieldMutation`) and no
+  longer assembles a Field document of its own. It never sets a behaviour, a
+  format, a unit, a relation target, a relation maximum, a compute_on or a
+  pattern Rule — every one of those was a second place for the thirteen types to
+  be wrong.
+- **Sixteen things a column can hold, in a menu under four headings**, each with
+  one line saying what it is FOR. `fieldTypes.ts` is new;
+  `parityTypesWithNoExplanation()` fails the suite the moment the store ships a
+  fourteenth parity type nobody has written a sentence for.
+- **A choice list is defined where a person defines it** — the words go in a box
+  and the store keeps them as a Table (FLD-5 / FLD-6 untouched). Nobody is asked
+  to build a table before they can have a dropdown.
+- **A Person field offers people.** It resolves to a relation at the kernel
+  Person Table, decided by the store.
+- **The Key box is gone.** The key is derived from the name and shown, never
+  demanded.
+- **A field is edited and removed from the same panel**, and removal names the
+  consequence before the second press. `TableSettings` retires through the field
+  door rather than deleting the record, so the TABLE stops declaring the column
+  too; the three unsafe removals (the title field, the last field, one another
+  field reads through) are refused by name.
+- **Ids stopped being drawn as values.** `RecordLabelProvider` resolves a choice
+  or a relation id to the name of the record it points at, once per Field, for
+  every read-only surface — the Kanban board's column headings were bare
+  36-character identifiers.
+- **`3` in `units` no longer renders as `units3`.** `editorKindFor` called any
+  range field with a unit longer than one character a currency; a currency is a
+  range whose FORMAT is currency, which is what the store's own `parity_type`
+  says.
+- **One word per type everywhere** — the settings row says what the panel
+  offered, and the store's "made of" sentences (`REC-31: a File record reached
+  through a relation`, `evaluated by custom.rule_eval`) are out of the UI.
+- The person picker draws a name and an email as two facts, not one run-on
+  string.
+
+### Consumer action
+
+`renderValue` still works and still draws an unresolved id as "Loading…". A host
+drawing values outside `RecordsMount` should render `<RecordValue>` (or wrap in
+`RecordLabelProvider`) so ids resolve to names.
+
+
+## 0.7.0
+
+What the first headless walk of 0.6.0 through a real browser found, in the same
+hour — every one of them on a table a person had just made.
+
+- **A new table takes an empty record.** `DEFAULT_FIELDS` made `title`
+  REQUIRED, so "New record" — which writes an empty record and opens its first
+  cell — was refused (`REC-51: Title is required`) on every table the app makes.
+  A title is what makes a record a chip (REC-2), not a promise it can never be
+  written without; an organization that wants it demanded says so in the field
+  editor and the store then enforces it for everyone.
+- **`TableSettings` can delete the table.** SCR-3 claims rename, reorder,
+  retype and delete; the panel only ever deleted a FIELD, so a table made in the
+  product could not be got rid of from anywhere in it. It is the same soft
+  delete every record gets, through the same door, and the consequence is named
+  before the second press. `onDeleted` lets the host leave the page, and
+  `TablePage` passes its `onLeave`.
+- **Nobody conflicts with themselves any more.** Three separate paths wrote a
+  record twice against one version and were — correctly — refused `PT409` by the
+  store, so the screen showed a person a conflict with nobody but themselves:
+  Tab fired the cell's key handler AND its blur handler; two cells of one record
+  were written in parallel; and `useRecordVersion` never moved after a save, so
+  switching a saved view's layout twice, or saving a form twice, conflicted with
+  the save before it. One commit per opening, one write at a time per record,
+  and every successful write hands its new version back.
+- **The demo harness renders like the product.** It imported the design
+  system's structural sheet and nothing else, so every page was unstyled HTML in
+  a real browser while `pnpm test` was green. It now runs Tailwind over this
+  package's own source, and the address bar takes `?demo=` and `?table=` so a
+  walk can make a throwaway table, work in it and delete it again.
+
+### Consumer action
+
+**`useRecordVersion` now answers `{ version, note }` rather than a number.** A
+screen that writes the same record more than once must hand each successful
+write's new version back through `note(...)`, or the next write will be refused
+as a conflict with the person who just made the last one. Nothing else changed:
+`version` is the same value the hook used to return, and `null` still means "no
+version was read", which is the store's ordinary last-writer-wins write.
+
+A host that shows `TableSettings` on a page about one table should pass
+`onDeleted`, or the person is left on a page about a table that is gone.
+
 ## 0.6.0
 
 Editing where the value is shown, and four other things an independent walk of