@ai-matrx/records-ui 0.30.0 → 0.40.0

package/CHANGELOG.md

@@ -1,5 +1,318 @@
 # Changelog — @ai-matrx/records-ui
 
+## 0.40.0
+
+**Pressing "Settings" on a table crashed the whole page, and had since the
+enrichment panel was added.**
+
+Lane LAND-SCREENS pressed it in a production build of the app, signed in as the
+admin who owns the table, and the route went to the application's global error
+boundary with React #310 — *"Rendered more hooks than during the previous render."*
+Everything on that panel went with it: rename the table, add a column, take one
+off, reorder, delete the table, and the enrichment panel itself, which lives behind
+a column's own row and was therefore unreachable in the shipped build.
+
+**The cause, and why no suite saw it.** `TableSettings` declared `useState` for the
+enrichment panel BELOW three early returns, one of which fires on the FIRST render
+of every real mount: a host that lets the store decide rights binds no `rights`
+port, so `useTableRights(table.data)` answers `NO_RIGHTS` until `useTable` and
+`custom.my_level` come back. Render one ran eight hooks and returned the "you may
+not change the structure" line; render two, a moment later, ran nine. Every test in
+this package binds `rights: () => tableRightsAt("admin")`, which never takes that
+return — so the panel was only ever rendered with nine hooks and the order never
+changed.
+
+`src/TableSettings.hooks.test.tsx` mounts the panel with NO rights port, which is
+the seat the product actually uses, and fails with that exact error on the previous
+hook order.
+
+Also: `src/form-publish.test.tsx` had two TS7006 implicit-`any` parameters that had
+`tsc --noEmit` red on this package on `main`. Typed; the package typechecks clean.
+
+## 0.39.0
+
+**The kanban view becomes the board, with the store judging every drag.**
+
+`PipelineBoard` renders THROUGH the view that already exists — `ViewSwitcher`'s
+kanban layout — rather than beside it. It draws the very same rows the view holds;
+what it adds is the store's judgement.
+
+* **Drag between columns.** A drop asks `custom.pipeline_transition_refusal` first
+  and either writes the move or leaves the card exactly where it was, carrying the
+  store's own sentence, verbatim. Nothing is written on a refusal, so there is no
+  snap-back to animate — the card never moved.
+* **A stage's demands, prompted on the way in.** When the refusal names the columns
+  a stage needs, they appear on the card with a field each and one button, so the
+  person finishes the move instead of being sent away from it. Until every demand is
+  answered there is no button to press, rather than one that looks live.
+* **Column headings carry the store's numbers** — the count, the chosen measure's
+  total, and the column's limit — over every record this person may see, not over
+  the page the screen happened to load. Which column is totalled is a fact about the
+  VIEW, saved on it like its colours.
+* **A work-in-progress limit shows on the column it belongs to**, and a column over
+  its limit says so in a sentence rather than only turning red.
+* **A viewer's card has no drag handle at all.** Absent rather than dead.
+
+A table that is not a pipeline draws the plain kanban exactly as before; the
+decision is the Table's own `stage_field`, asked once, never guessed from the shape
+of a Choice column.
+
+
+
+
+## 0.38.0 — 2026-09-20
+
+**0.37.0's gate went red on the tarball, not on this package.** The packed
+consumer imports `@ai-matrx/records` from npm, and the published copy did not
+yet carry `openCaptureQueue` — a door the offline capture screens here already
+call. That is the tarball canary doing exactly its job: green in the monorepo,
+absent for an npm consumer. `@ai-matrx/records` 0.26.0 carries it, and this
+release is the first that can resolve against it.
+
+
+## 0.37.0 — 2026-09-20
+
+**Three unhandled rejections were failing the release gate while every test
+passed.** `demo/fixtures/doors.json` had no recorded answer for
+`custom.enrich_cells`, a door the Enrich screens started calling — and the
+recorded-doors fake deliberately REFUSES to invent one, because an empty answer
+there would be a green test about nothing. That refusal arrived as an unhandled
+rejection, so vitest reported `12 passed / 79 passed` and exited non-zero, and
+0.36.0 never reached npm. The fixtures are re-recorded against the live store;
+the suite is now 12 files / 79 tests with no errors.
+
+
+## 0.37.0 — 2026-09-20
+
+**The capture sheet, run on a phone, with or without a signal (PRODUCTS row 15).**
+
+* `CaptureRun` — a DECLARED capture sheet, one question at a time, big targets,
+  the camera and the microphone through `accept`/`capture` (which is what opens
+  them on a phone rather than a file browser), and the durable queue from
+  `@ai-matrx/records@0.25.0` underneath. The whole sheet is fetched ONCE by
+  `custom.capture_open` and everything after that happens on the phone: a
+  validation that needs a round trip does not exist in a basement. It never says
+  *saved* when it means *queued* — the thank-you screen says which, the waiting
+  count is on screen the whole time, and "last synced" is a real moment read from
+  the queue. A member who may only look is told in words on the FIRST screen,
+  because `may_capture` comes from the door with a sentence attached.
+* `CaptureSheet` gains `sheetId` and delegates to `CaptureRun` when it is given.
+  One component, two modes: with a sheet it runs what somebody DECLARED (the crew
+  case — twelve people, one round, the same questions); without one it stays the
+  ad-hoc capture that guesses which Field a reading goes in, which is right for a
+  one-off and wrong for a crew. No second capture component.
+
+## 0.36.0
+
+**"Not between ten at night and seven in the morning" — the two settings the notify
+editor could not offer, and the summary you can ask to see.**
+
+`NotifyRuleEditor` could say WHICH view and, out of a two-word vocabulary, roughly
+how often. There was no way to say "Monday at eight" and no way at all to say when
+NOT to be told, so a subscription that texts on a new lead texted at 03:00. It now
+asks for the cadence (from the store's own catalogue — a picker can never offer a
+word the runner does not act on), the channel, the schedule, and both ends of a
+quiet-hours block, and it says plainly that a send inside those hours is HELD until
+they end rather than dropped.
+
+It also stops assembling the Rule itself and writes through `subscriptionDeclare`.
+The store is the one place that knows which cadences the runner honours and which
+quiet hours it can parse, and it refuses the rest by name — so the sentence a person
+reads on a refusal is the store's own.
+
+`SubscriptionsPanel` gained the two answers anybody looking at a notifications list
+actually wants: **when the next summary is due** (computed by the same store
+functions the runner uses, so the screen and the runner cannot drift — and absent
+rather than wrong on a muted subscription) and when it last told you anything. Plus
+**"Send me a preview now"**, which renders the summary it would send at this moment
+with what arrived, what left and what changed NAMED. It assembles only: nothing is
+delivered, nothing is recorded and the watermark does not move, so pressing it never
+costs somebody a text and never robs the real Monday summary of a week of news.
+
+Needs `@ai-matrx/records` ≥ 0.26.0.
+
+
+## 0.36.0 — 2026-09-20
+
+**0.35.0 never reached npm: the demo harness asked the store for ten thousand
+rows in one page.**
+
+`custom.read_records` serves at most a thousand rows in one page and now REFUSES
+a bigger ask outright (`22023`, `PAGE-1`) rather than quietly serving fewer —
+which is the right behaviour and exactly the kind of honesty this package is
+built on. The harness's reset asked for 10,000, so it read nothing, threw, and
+took the whole release gate down with it. It pages now, a thousand at a time,
+until a short page says it is done. The ceiling is the store's to raise
+(`custom/page_size_ceiling`); a caller's job is to page.
+
+Proven against the LIVE store with the demo server up: `demo/live-smoke.test.tsx`
+green, 1 passed.
+
+Also in this release, from 0.34.0/0.35.0 and this lane:
+
+- **"Who changed this?" is drawn for a person who cannot edit the cell.**
+  `columnForField` chose between the cell shell and a plain value on whether
+  there was an EDIT SESSION, and the grid passes `null` there whenever the table
+  is not this person's to write — so on a viewer, a portal and every read-only
+  embed the control vanished, while `gridEditing.tsx` promised it in its own
+  comments. `GridCell` now takes `editing: GridEditing | null`; a null session
+  means no editor, and nothing else.
+- The tarball canary names `HistoryPanel`, `CommentThread` and
+  `FieldHistoryPanel`, so a screen that is green in the monorepo and absent from
+  an npm consumer's tarball is caught at the gate.
+
+
+## 0.35.0
+
+**Per-record AI enrichment: the badge on every agent-owned cell, and the panel behind it.**
+
+`EnrichBadge` is drawn beside every value a model owns, always — not on hover and
+not behind a toggle, because a value a machine guessed and a value a person typed
+are different facts and the difference has to survive onto the screen. Hovering it
+gives the rest: who, when, what was read, how sure the model was, what it nearly
+said instead, and the one control that matters — taking the cell back. `stale`
+rides beside it as a second WORD rather than a colour, because "this is six weeks
+old" is a sentence and a hue has to be explained every time.
+
+`useEnrichCells` fetches a whole page's provenance in ONE `custom.enrich_cells`,
+and `Grid` uses it for the ids it is actually drawing.
+
+`EnrichPanel` is rewritten. It was a per-record "refill these fields" button over a
+port nothing bound; it is now the whole of an enrichment in one screen, in the
+order somebody thinks in: what should go in this column · what it may look at ·
+how often to check it again · what it has cost, per run and per row · SEE IT ON
+FIVE ROWS · and only then, turn it on. Setting one up from a column's own menu
+lands here too, so "Let AI fill" is not a button that leads nowhere.
+
+`TableSettings` gains that menu entry beside Edit and Remove, and marks a column a
+model already fills in.
+
+**BREAKING:** `EnrichPanel` no longer takes `recordId`. It is a panel about a
+COLUMN and its whole Table, which is what "keep this filled in" is about; the old
+signature could only ever refill one record.
+
+The package's own gate now resolves `@ai-matrx/records` to the workspace SOURCE
+rather than the last published tarball — before that, a suite could only exercise
+doors already on npm, so every new door arrived untested through the one gate meant
+to catch it.
+
+## 0.34.0
+## 0.33.0 — 2026-09-20
+
+**"Who changed this?" rendered zero controls for the person most likely to ask
+it.**
+
+`gridEditing.tsx` says in its own comments that the cell's history control is
+offered at VIEWER too — "a person who may not edit a cell is exactly the person
+who wants to know who did". `columnForField` did not agree: it chose between the
+cell shell and a plain value on whether there was an EDIT SESSION, and the grid
+passes `null` there whenever the table is not this person's to write. So on a
+viewer, a portal and every read-only embed the shell vanished and took the
+control with it. The shell is now drawn whenever there is an editor OR a
+question to ask, and `GridCell` takes `editing: GridEditing | null` — a null
+session means no editor, and nothing else. `src/who-changed.test.tsx` covers
+it; its load-bearing clause fails on the old gate.
+
+
+## 0.32.0
+
+**The import wizard, rebuilt on the store's own import doors (SCR-10 / SCR-N-7).**
+
+`ImportWizard` parsed a file and then called `recordWrite` **once per row**. At
+the five thousand rows a real spreadsheet holds that is five thousand HTTP round
+trips; there was no duplicate check, so importing a corrected file made a second
+copy of everything; nothing recorded that a value came from a file; and a
+refusal carried a row NUMBER into a file the screen no longer had, so "row 3,184
+was refused" was information nobody could act on. `onProposeField` handed a
+header to a callback and the platform forgot about it.
+
+The bytes are still parsed here — Postgres cannot unzip a workbook — and
+**everything after the parse is the store's**:
+
+- **The plan, before anything is written.** `importPlan` answers what each
+  column is, reading the parity map AND this organization, and the screen shows
+  the kind, the store's own sentence saying why, and the first values. Every
+  column's destination is a dropdown a person can change, and what they change
+  is what the store is sent.
+- **A duplicate key, chosen on the screen** ("same row means the same …"), with
+  "leave it alone" or "bring it up to date".
+- **Batches of 250** through `importRows`, with a live count, and **one outcome
+  per row**: landed, already here, or refused **with the store's own sentence
+  and the row from the person's own file** beside it. The outcome table filters
+  to refused / already-here and is the click-through SCR-10 asks for.
+- **The same file twice writes nothing.** The file's sha-256 is the run's
+  identity; the screen shows when it ran and what landed, and never sends a row.
+  Where Web Crypto is genuinely absent the screen SAYS it cannot recognise the
+  file again rather than inventing a hash.
+- **Columns this table does not have go to the ONE approvals inbox** by default,
+  and the screen says so on the control itself. "Add them straight away" is
+  offered only to an admin and is **absent, not disabled,** for anybody else —
+  who are not stuck, because the default still files their column.
+- **One word per type, everywhere.** The wizard reads `FIELD_TYPE_CHOICES`, the
+  same list the "Add field" panel offers, so a column is "Money" in both places.
+
+`ImportWizardProps.onProposeField` is **gone**: the offer is now a real change
+waiting in a real queue, not a callback the host had to invent a home for. Both
+existing call sites (`TablesHome`, `TablePage`) already pass only `tableId` and
+`onDone`.
+
+New in the blocking gate: `src/import-wizard.test.tsx`, five clauses against
+door answers recorded from the live store. Its red twin is `demo/import.red.tsx`
+(the screen as it was), which fails on all three arms.
+
+Requires `@ai-matrx/records` >= 0.19.0.
+## 0.31.0
+
+**A record's past is a list of VERSIONS, and an agent names the person it acted for.**
+
+`HistoryPanel` listed each field's LATEST value with the author of the last
+write beside it. Useful, and not a history: it could not say what a price was
+before, when it changed, what else changed with it, or put any of it back. It
+reads `custom.record_history` now — one line per VERSION, newest first, each
+naming who, when, the operation in the word somebody used, and every field that
+moved with its before and its after, the Rule version that produced a derived
+value and the evidence behind it. Keyboard-first: Up and Down walk the versions,
+Enter opens one, R starts a restore.
+
+**The put-it-back action names what will change before it does.** Pressing
+Restore asks the store what the write would do and shows it — *"Price 1,100 →
+1,200, Notes cleared"* — and only the second press writes. It says out loud that
+the write is a new version and nothing in the chain is erased, which is the
+thing people are afraid of about an undo button.
+
+**`FieldHistoryPanel` is new: "who changed this?", asked of a CELL.** Every grid
+cell now carries the question (a quiet control on hover or focus, and `H` from
+the keyboard at viewer too — a person who may not edit a cell is exactly the
+person who wants to know who did), and the answer is that COLUMN's history
+across the whole Table, with "put this one back" on each row.
+
+**`CommentThread` stops keeping its own store.** It declared a package-owned
+system Table called `records_ui_comment` into whichever organization opened the
+screen — the same second store lane FORMS removed for forms and lane DASHBOARDS
+removed for dashboards — and here the rung was wrong in BOTH directions: a
+`commenter` could not comment at all, and anybody with write on the package's
+comment Table could comment on a record they may not open. It reads
+`custom.comment_thread` and writes `custom.comment_write` now, with the people
+NAMED, a comment anchored to one Field, resolve and re-open, and an `@` mention
+picked from the organization's own roster. Whether this person may comment is
+the STORE'S answer, in the same call that answers the thread.
+
+**Two defects the SCREENSHOTS found**, neither visible to a test, a type or a
+door: a version's timestamp rendered `9/20/2026, 7:12:37 AM` and wrapped onto two
+lines, doubling the height of every row — it is a relative time now (`12 minutes
+ago`, `Tue 14:12`, then a date) with the exact moment on hover, which is what
+Linear and Notion do and is also what fits. And the cell's "who changed this?"
+control was `display:none` until hover, so it was unreachable by the keyboard
+and invisible to anything driving the screen; it is `opacity-0` now and appears
+on hover OR focus.
+
+**Consumer action:** `COMMENT_TABLE` is gone — it described a store that no
+longer exists and nothing outside this package used it. `HistoryPanel` gains an
+optional `onAskAboutField`, `CommentThread` an optional `fieldKey`, and `Grid`,
+`ViewSwitcher` and `columnForField` an optional `onAskWhoChanged`; every one is
+additive and absent means the control is not drawn. Needs
+`@ai-matrx/records` >= 0.18.0.
+
 ## 0.30.0
 
 **The notification editor stops asking for the notifier's own reader.**

package/README.md

@@ -1,8 +1,8 @@
 # @ai-matrx/records-ui
 
-THE canonical screens of the AI Matrx record store. Every component here reads
-and writes ONLY through [`@ai-matrx/records`](../records/README.md), renders
-with `@ai-matrx/design-system`'s own primitives and data table, and carries
+The platform's primary record UI components. Every component here reads and
+writes ONLY through [`@ai-matrx/records`](../records/README.md), renders with
+`@ai-matrx/design-system`'s own primitives and data table, and carries
 `@ai-matrx/alchemy`'s menu where a person reaches for one.
 
 There is no bespoke per-feature record UI in this platform. A feature that needs