Dynamic Attributes — plan and progress record

The durable plan for pattern.md: the phases as planned, what each delivered, the second-pass audit and what it changed. Branch claude/cloud/dynamic-attributes; development mode, so nothing here is merged or deployed.

Phases

Phase	Scope	Status
Research	Internal inventory and external survey → research.md	Done 2026-11-05
A — core contract	`#lib/dynamic-attributes/`: JSON bounds, limits, kind registry and seven core kinds, codec with upcasters, engine operations, diagnostics, three-way merge, store contract, text projection	Done
B — renderer	`#lib/components/composites/attributes/`: editor state (undo, coalescing, refusals), `AttributeEditor`, `AttributeRow`, one control per kind, lazy-control support, keyboard reorder, mobile layout, `attributes_*` i18n (en, de, ru)	Done
C — persistence	Spreadsheet autosave lifted into `#lib/state/document-*` (generic, plus `rebase`); `desk.record` detail table; `createRecordFile` / `updateRecordByFileId` (parse → lock → version → revision); API GET/PUT/POST with 409/413/422	Done
D — desk integration	`RecordPanel` (menus, AI context, bus refresh, flush on leave), panel registry, About help, Explorer New Record, open by file type, unsaved work at sign-out, assistant read path and retrieval projection	Done
E — showcase	`/showcases/dynamic-attributes`: two writers over one browser store, offline switch, reset, stored-JSON view, import box with hostile samples, an application kind (`app.rating`)	Done
F — library and docs	This directory, `pattern-library/registry.json` entry `dynamic-attributes`, naming vocabulary, desk/spreadsheet/execution docs, rebuilt pattern pages and MCP excerpts	Done
Tests and gates	Unit (core, editor state, autosave, drafts), PGlite (record mutations), HTTP contract, browser drive of the showcase, benchmark	Done — see Verification

Verification

Unit and contract tests: codec.test.ts, operations.test.ts, merge.test.ts, stores.test.ts, attribute-editor.svelte.test.ts, document-autosave.test.ts (with three rebase cases), spreadsheet-drafts.test.ts, record-mutations.pglite.test.ts, api/desk/files/[id]/server.test.ts (record cases).
Browser (cloud dev server, Playwright, 2026-11-05): a fresh record opens with one empty Name row; text, number, rating and a text→number change with the kept previous value; writer A saves, writer B edits without reloading → conflict → Merge both → saved with both edits; reload restores the stored record and offers the unsaved draft for recovery; the import box refuses the prototype-pollution sample (invalid_shape, reserved key) and accepts the valid one, whose unknown kind renders read-only. No console errors.
bun run attributes:bench — figures in pattern.md.
Not run in the cloud: db:push against Neon (the desk.record table and the file_revision.document column exist only after a push), the RAG ingest of these docs, and the desk panel against a real signed-in session (the API and the mutations are covered by the contract and PGlite tests).

Second-pass architecture audit

Done after E, reading the branch as a reviewer would. Findings and outcomes:

Finding	Outcome
The pattern-library record `versioned-document-session` and three docs still named the moved `spreadsheet-autosave.ts`.	Repointed to `#lib/state/document-autosave.ts` / `document-drafts.ts`; pages and excerpts rebuilt.
`canonicalJson` was declared twice (here and in asset production), caught by the naming gate.	One implementation in `#lib/utils/canonical-json.ts`, imported by both.
The editor lived in a feature-rank component directory, so the desk panel and the showcase importing it broke component layering (architecture gate).	Moved to `#lib/components/composites/attributes/`: it imports only primitives, which is what a composite is.
The bot's desk layout listed spreadsheet and markdown panels' file type but not records'.	Record panels report `fileType: 'record'`.
Sessions were held in deep `$state`, which proxies the session object and its engine.	`$state.raw` in the panel and the showcase; snapshots are replaced, never mutated.
Type selects had no accessible name (the `Select` primitive took no `aria-label`).	`Select` takes `id` and `aria-label`; every type select is named; the add bar puts the type before Add.
The proposal undo path has no record branch.	Left as is on purpose: `StepRecoveryPoint` types revision points as spreadsheet or markdown only, because the assistant has no tool that writes a record. Adding a write tool must add the undo branch with it.
Every operation re-serializes the record to check the byte cap (O(n) per keystroke).	Kept: 0.35 ms at the 200-attribute limit. An incremental size count would add state that can drift from the truth.

Decisions that need Stas

None block the branch. Two are worth a look when reviewing: whether the assistant should get a record write tool (it would need the proposal-undo branch above), and whether record is the name the desk's Explorer should show for this file type.