Skip to content

feat: scaffold @meshtastic/sdk + signals/sqlocal persistence migration#1050

Open
danditomaso wants to merge 61 commits into
meshtastic:mainfrom
danditomaso:sdk-migration
Open

feat: scaffold @meshtastic/sdk + signals/sqlocal persistence migration#1050
danditomaso wants to merge 61 commits into
meshtastic:mainfrom
danditomaso:sdk-migration

Conversation

@danditomaso
Copy link
Copy Markdown
Collaborator

Summary

Lays the foundation for migrating off @meshtastic/core + Zustand + IndexedDB onto a domain-driven SDK (@meshtastic/sdk) with signals-backed reactive state, OPFS-backed SQLite persistence (@meshtastic/sdk-storage-sqlocal), and React bindings (@meshtastic/sdk-react).

10 commits, intentionally non-destructive: web still builds and all 294 web tests pass. Slice-by-slice migrations of UI consumers happen in follow-up PRs (see TESTING.md and the migration plan).

What's in the branch

Commit Effect
1. feat(sdk): scaffold @meshtastic/sdk + @meshtastic/sdk-react Full DDD scaffolding. 9 feature slices (device/chat/nodes/channels/config/telemetry/position/traceroute/files), shared kernel (MeshClient, Transport, EventBus, Queue, Xmodem, signals, packet codec, tslog factory), Phase-A shim re-exporting legacy MeshDevice/Types/Utils.
2. refactor(web): import @meshtastic/sdk instead of @meshtastic/core 74-file mechanical import swap. Web runs on the shim path.
3. feat(sdk): add MeshRegistry + multi-client React providers MeshRegistry (Map<ConnectionId, MeshClient>), MeshRegistryProvider, useActiveClient, useClientById, useMeshRegistry. Existing hooks fall back through registry.
4. refactor(sdk-react): rename useDevice → useMeshDevice Avoids collision with web's existing useDevice Zustand hook.
5. feat(web): mount MeshRegistryProvider at app root App-wide registry singleton wired in index.tsx.
6. feat(web,sdk): register per-connection MeshClient in registry MeshDevice.meshClient getter + registry.register/unregister. useConnections adopts the shim's inner client.
7. feat(sdk): add MessageRepository port + InMemoryMessageRepository Per-slice persistence port. ChatClient takes { repository?, retention?, initialLoadLimit? }, lazy-hydrates per conversation, writes through on each inbound message.
8. feat(sdk-storage-sqlocal): SQLite WASM persistence adapters New workspace package. Drizzle schema (messages/nodes/telemetry/_schema). SqlocalMessageRepository, MultiTabCoordinator, createSqlocalDb, createMemoryDb (sql.js for tests).
9. feat(web): persist chat history via @meshtastic/sdk-storage-sqlocal Web now opens an OPFS DB on first connect; chat slice writes through to SQLite. Retention 1000 msgs/bucket or 90 days. Vite worker format set to ES.
10. test: testing strategy + 19 new SDK / storage / hook tests TESTING.md (six-tier strategy + audit + gates). New slice tests for nodes/channels/config/telemetry/position. Chat persistence round-trip. Schema migrations. Cross-tab BroadcastChannel. Browser-mode harness for real OPFS round-trip.

Architecture decisions locked

  • Replace @meshtastic/core entirely, with shim covering the migration window.
  • Signals via @preact/signals-core; Zustand stays only for non-SDK UI state (theme, sidebar, dialogs).
  • better-result Result<T, E> for new application use-cases; legacy ports keep throwing.
  • Persistence: sqlocal (OPFS SQLite WASM) + Drizzle ORM, per-slice repository ports, lazy pagination instead of bulk rehydrate (fixes the 1000-message reload bug).
  • Multi-tab coordination: Web Locks API + BroadcastChannel; first-tab-wins write lock, broadcast-driven invalidation in readers.
  • ConfigEditor replaces the existing changeRegistry (lands with config slice migration). Baseline = device truth; working = UI edits; commit diffs sections; disconnect discards in-flight edits.
  • Multi-device aware — every storage table has device_id; one DB per origin.
  • No Claude / Anthropic / AI-tooling references anywhere in the repo.

Test counts

Package Tests Status
@meshtastic/sdk 36
@meshtastic/sdk-react 8
@meshtastic/sdk-storage-sqlocal 12 ✅ (Node); browser-mode harness wired
meshtastic-web 294 ✅ unchanged

pnpm -r build is green across all packages including production web Vite bundle.

Out of scope (queued for follow-up PRs)

Test plan

  • Reviewer reads TESTING.md and confirms the six-tier strategy + per-package gates make sense.
  • pnpm -r test runs clean on a fresh checkout.
  • pnpm --filter meshtastic-web build produces a working web bundle (sqlocal worker bundles as ES).
  • Manual: connect a real Meshtastic device via HTTP / Bluetooth / Serial, send/receive a message, reload the page, confirm message history hydrates from SQLite via useChat's lazy load (this requires PR missing mqtt settings #6 to be visible to UI; for now history is in OPFS but UI still reads legacy messageStore).
  • Two-tab smoke: open two browser tabs, send a message in tab A, confirm BroadcastChannel notifies tab B (requires PR missing mqtt settings #6 UI wiring to surface).
  • Optional: pnpm --filter @meshtastic/sdk-storage-sqlocal test:browser (needs @vitest/browser + Playwright installed) to validate real OPFS round-trip.

Adds two new packages laying the foundation for a domain-driven migration
away from @meshtastic/core.

packages/sdk
- DDD feature slices: device, chat, nodes, channels, config, telemetry,
  position, traceroute, files. Each with domain/application/infrastructure/state.
- Shared kernel under core/: MeshClient orchestrator, Transport interface
  (byte-compat with existing transport-* packages), EventBus (typed pub/sub),
  packet codec, Queue, Xmodem, signals helpers, tslog factory.
- Signals via @preact/signals-core. Application use-cases return
  Result<T,E> via better-result; legacy ports keep throwing.
- shim/ re-exports the legacy MeshDevice/Types/Utils API so
  packages/web continues to build unchanged.
- createFakeTransport() under @meshtastic/sdk/testing.
- 16 vitest tests incl. end-to-end fake-transport integration.

packages/sdk-react
- MeshProvider + useSignal/useSignalValue/useClient adapters.
- Hooks: useDevice, useConnection, useChat, useNodes, useNode,
  useChannels, useChannel, useConfig, useModuleConfig, useTelemetry,
  usePosition, useTraceroute, useFileTransfer, useFavoriteNode, useIgnoreNode.
- jsdom-backed hook tests.

Root README rewritten with packages table, architecture, and workflow.
Mechanical swap across all 74 @meshtastic/core imports in packages/web/src.
The SDK's shim layer re-exports the legacy MeshDevice/Types/Utils/Protobuf
surface, so no source edits are required beyond the import path. All 294
web vitest tests still pass.

This is Phase B step 1: web now runs on @meshtastic/sdk's MeshDevice shim.
Per-slice store migrations (messageStore/nodeDBStore/deviceStore →
useChat/useNodes/useDevice etc.) land in follow-up commits.
Phase B prep for web migration. Web holds multiple simultaneous device
connections keyed by ConnectionId, so per-slice hook migrations need a
registry-aware provider.

packages/sdk
- MeshRegistry: Map<ConnectionId, MeshClient> with signals for list/active/activeId.
  create()/get()/has()/remove()/setActive().
- First-created client auto-activates.
- remove() disconnects the client and rotates active to another entry.
- 4 vitest cases covering create, auto-activate, duplicate rejection, and remove.

packages/sdk-react
- MeshRegistryProvider + MeshRegistryContext.
- useMeshRegistry, useOptionalMeshRegistry, useActiveClient, useClientById(id).
- useClient now falls back to the registry's active client when no direct
  <MeshProvider> is present, so existing hooks work unchanged under a
  registry-backed app.

No web-facing changes in this commit; used by follow-up slice migrations.
Prevents collision with packages/web's own useDevice() Zustand hook. All
internal exports + tests updated; no behavior change.

Callers migrating off @meshtastic/core should use useMeshDevice() from
@meshtastic/sdk-react going forward.
Introduces a single app-wide MeshRegistry in packages/web/src/core/meshRegistry.ts
and wraps the RouterProvider in <MeshRegistryProvider>. Registry starts empty;
useConnections continues to instantiate the legacy MeshDevice shim. Subsequent
slice migrations will swap useConnections over to registry.create() and move
consumers onto useMeshDevice()/useChat()/etc from @meshtastic/sdk-react.

Adds @meshtastic/sdk-react as a web dependency. No behavior change; web tests
(294) and production build still pass.
packages/sdk
- MeshRegistry.register(id, client): adopts an externally-constructed
  MeshClient. Complements create() for migration paths where a legacy shim
  already owns the client.
- MeshRegistry.unregister(id): drops the mapping without disconnecting, for
  cases where the caller has torn down the transport itself.
- Legacy MeshDevice shim exposes its inner MeshClient as `meshClient` so
  consumers can adopt it into the registry.

packages/web
- useConnections.setupMeshDevice now registers the shim's MeshClient with the
  app-wide meshRegistry and marks it active on connect.
- removeConnection unregisters from the registry.
- Legacy Zustand deviceStore wiring is unchanged; follow-up commits will move
  read paths to useMeshDevice/useChat etc. and remove the duplicated fields.

No behavior change visible to users. Web tests (294) + SDK tests (20) still pass.
Lays the persistence groundwork for the chat slice migration.

- MessageRepository port defines paginated reads (loadRecent, loadBefore),
  atomic writes (append, appendBatch), state updates, and retention pruning.
  Conversation keyed by ConversationKey tagged union (channel | direct peer).
- RetentionPolicy: maxPerBucket + olderThanMs knobs. Consumer decides.
- InMemoryMessageRepository ships with SDK as default + test fixture.
- ChatClient accepts { repository?, retention?, initialLoadLimit? }.
  Lazy-hydrates per conversation on first subscribe; writes through on every
  inbound message; prunes after append when retention policy is set.
- ChatClient.loadOlder(conv, before, limit) for pagination UI.
- ChatStore.prepend() for older-first inserts.

Tests: 5 new cases for InMemoryMessageRepository (paginate, update state,
retention). 25 SDK tests total, all green. Web build unchanged.

Paves the way for @meshtastic/sdk-storage-sqlocal to implement this port
against SQLite/OPFS in a follow-up.
New workspace package implementing @meshtastic/sdk repository ports against
sqlocal (SQLite WASM + OPFS) with Drizzle-typed queries.

Schema (single DB, multi-device aware via device_id column)
- messages: chat history. Indexes on (device_id, conversation_key, rx_time)
  for fast pagination and on (device_id, state) for pending lookups.
- nodes: NodeDB snapshot per device (stub schema, repo lands with PR #7).
- telemetry: per-node ring buffer of readings (stub).
- _schema: migration version table.
- Hand-written DDL migrations in src/schema/migrations.ts; applied at boot.

createSqlocalDb({ databasePath })
- Opens OPFS DB, applies migrations, returns Drizzle client typed against the
  schema.
- Single instance per origin; sqlocal serializes writes via OPFS file locks.

SqlocalMessageRepository
- Implements MessageRepository: paginated loadRecent / loadBefore, append /
  appendBatch with onConflictDoNothing, updateState, prune (maxPerBucket via
  windowed DELETE + olderThanMs).
- Optional MultiTabCoordinator broadcasts messages-changed events on append.

MultiTabCoordinator
- BroadcastChannel pub/sub for cross-tab change notifications (no-ops if API
  unavailable, e.g. Node).
- acquireLock() wraps navigator.locks.request with fall-through for non-browser
  contexts.

Testing
- src/testing/createMemoryDb.ts: in-memory sql.js + Drizzle, same surface as
  the sqlocal connection. Lets repository tests run on Node CI.
- 6 SqlocalMessageRepository tests (pagination, retention, multi-device
  isolation) + 2 MultiTabCoordinator tests pass on sql.js.

Notes
- @meshtastic/sdk pkg.json now points types at ./mod.ts so workspace consumers
  resolve types directly from source. Production publish path needs a separate
  follow-up to emit a stable mod.d.ts.
- Storage pkg ships ESM only; dts disabled until tsdown's hashed-name emit
  is reconciled with mod.d.ts resolution. Workspace consumption already gets
  full types from source.
Plumbs the SDK chat slice through the OPFS-backed SQLite repository so
inbound and outbound text messages survive page reloads, with retention
capped at 1000 messages per conversation or 90 days, whichever hits first.

- packages/sdk legacy MeshDevice shim now accepts MeshClientOptions
  (chat, configId, logger). Backwards-compatible with the old
  `new MeshDevice(transport, configIdNumber)` form.
- packages/web/src/core/sdkStorage.ts: lazy singletons for the shared
  SqlocalDb and the cross-tab MultiTabCoordinator. The DB opens on first
  call so test runs that never connect stay headless-safe.
- useConnections.setupMeshDevice is now async, awaits getStorageDb, and
  passes a SqlocalMessageRepository scoped to the connection id. Falls
  back to the SDK's InMemoryMessageRepository if sqlocal init fails (no
  OPFS support, etc.).
- Vite worker format set to "es" because sqlocal's worker is ES-module
  and rolldown rejects iife with code-splitting.
- COOP/COEP dev headers were already in vite.config.ts; no further
  changes required for OPFS.

Web tests (294) and production build still green.

This is the runtime payoff of PR #5: a fresh page load populates chat from
SQLite via lazy pagination instead of rehydrating 1000 messages into memory.
The legacy Zustand messageStore is still in place for now; PR #6 (chat
slice migration) will retire it and switch UI components to useChat.
Adds TESTING.md documenting six tiers (unit / slice / client / storage /
hook / E2E), per-package coverage gates, and the audit of where we
currently sit.

Slice tests in @meshtastic/sdk
- NodeMapper proto round-trip; NodesClient list signal updates.
- ChannelsClient indexes by channel number.
- ConfigClient merges Config + ModuleConfig variants.
- TelemetryClient latest + history per node.
- PositionClient byNode + list.
- ChatClient persistence: hydrate on first subscribe, paginate via
  loadOlder, persist inbound messages through the repository. Fixes a
  reverse-iteration bug in loadOlder discovered by the new test.

Hook tests in @meshtastic/sdk-react
- New tests/hooks.registry.test.tsx covers useMeshDevice, useNodes,
  useNode, useChannels under <MeshRegistryProvider>, plus an active-
  client switch round-trip.

Storage tests in @meshtastic/sdk-storage-sqlocal
- migrations.test.ts validates v1 DDL creates messages/nodes/telemetry/
  _schema and indexes; CREATE IF NOT EXISTS is idempotent.
- MultiTabCoordinator.broadcast.test.ts proves cross-tab BroadcastChannel
  delivery between two coordinators in the same process.
- New vitest.browser.config.ts + tests/sqlocal-opfs.browser.test.ts run
  under @vitest/browser (Playwright provider) for real OPFS round-trip
  verification. Wired as `pnpm test:browser`. Browser test files end in
  `.browser.test.ts` and are excluded from the Node runner.

E2E / firmware-simulator tier (TESTING.md §"E2E / simulator") is scoped
for a follow-up — needs CI Docker for meshtasticd.

Totals after this change:
- @meshtastic/sdk: 36 tests (was 25)
- @meshtastic/sdk-react: 8 tests (was 2)
- @meshtastic/sdk-storage-sqlocal: 12 tests (was 8)
- meshtastic-web: 294 tests (unchanged)
@vercel
Copy link
Copy Markdown

vercel Bot commented Apr 25, 2026

@danditomaso is attempting to deploy a commit to the Meshtastic Team on Vercel.

A member of the Team first needs to authorize it.

Adds an adapter hook that bridges sdk-react's `useChat` and `useDirectChat`
to the legacy `Message` shape MessagesPage / ChannelChat / MessageItem
expect. The page now renders messages directly from the OPFS-backed
SqlocalMessageRepository — page reload hydrates lazily (last 50 per
conversation) instead of pulling 1000+ rows from IndexedDB into memory.

packages/sdk-react
- useChat() gains loadOlder(before, limit) for paginated backfill.
- New useDirectChat(peer) hook covering DM conversations.

packages/web
- src/core/hooks/useChatLegacy.ts: maps SDK `Message` → legacy
  `messageStore/types.ts` Message shape, including state translation
  (SDK Pending → legacy Waiting).
- MessagesPage flips broadcast and direct chat reads to useChatLegacy.
  getMessages call sites removed; setMessageState retained on the legacy
  store for outbound bookkeeping until the next migration step.

Drafts, unread counts, activeChat / chatType continue to live in the
Zustand messageStore — they are UI-only state and stay where they are
per the locked architecture decision. The legacy store's saveMessage,
getMessages, setMessageState, and persistence path remain in place for
now; PR cleanup follow-up will retire them once outbound state and
"delete all messages" flows are switched to the SDK.

Web test suite: 294 still green. Production Vite build clean.
Switches the send-text path from `connection?.sendText(...)` (legacy
MeshDevice) to `client.chat.send({ text, destination, channel })` on the
active MeshClient pulled from MeshRegistry via useActiveClient().

Side effects:
- Outbound message state (Ack / Failed) is now driven entirely by the
  SDK chat slice via routing-packet subscriptions; the manual
  setMessageState calls are removed.
- The legacy Zustand `useMessages().setMessageState` and `MessageState`
  imports are no longer used by MessagesPage. Drafts and unread counts
  still live in the Zustand store.
- `getMyNode` is no longer needed in this page (was only used to label
  the now-removed direct-message state updates).

Web tests (294) still green; production Vite build clean.
Moves per-conversation draft text out of the legacy Zustand messageStore
into the SDK chat slice so drafts share the same persistence + signal
machinery as messages. MessageInput now binds directly to the SDK; the
legacy `useMessages().getDraft/setDraft/clearDraft` API is no longer
called from the input.

packages/sdk
- New DraftRepository port + InMemoryDraftRepository default.
- ChatClient.drafts namespace: get(key) returns a ReadonlySignal<string>;
  set/clear keyed by ConversationKey.
- ChatClient.send auto-clears the draft for the resolved conversation on
  success (parity with prior Zustand clearDraft-on-send behavior).
- Lazy hydrate from the DraftRepository on first read of a conversation.
- 3 new tests in ChatClient.drafts.test.ts.

packages/sdk-react
- New useDraft(conversation) hook returning { text, setText, clear }.

packages/sdk-storage-sqlocal
- New `drafts` table (device_id + conversation_key composite PK, text,
  updated_at). Drizzle schema in src/schema/drafts.ts.
- Migration v2 in src/schema/migrations.ts creates the table on first
  open of an existing v1 DB.
- SqlocalDraftRepository implementing DraftRepository: load/save/clear/
  loadAll, scoped by device_id, upsert on conflict, delete on empty save.
- 6 new tests covering save/load round-trip, empty-string deletion,
  upsert, multi-device isolation, loadAll.

packages/web
- useConnections wires the SqlocalDraftRepository alongside the existing
  SqlocalMessageRepository per registered MeshClient.
- MessageInput accepts `conversation: ConversationKey` instead of the
  prior `to: Types.Destination` — fixes the legacy bug where every
  broadcast channel shared a single draft slot.
- MessagesPage passes the appropriate ConversationKey for direct/
  broadcast chats.
- MessageInput.test.tsx rewritten to mock useDraft from sdk-react.

Test counts: sdk 39 (+3), sdk-storage-sqlocal 18 (+6), web 294 unchanged.
Production Vite build clean.
The SDK chat slice now persists every inbound/outbound text packet via the
SqlocalMessageRepository wired in useConnections, so the legacy Zustand
saveMessage path in subscriptions.ts was writing to a store no UI code
reads from.

- subscriptions.ts: removed saveMessage call + PacketToMessageDTO usage.
  Unread-count increments retained as-is (cross-cutting concern, migrates
  in a separate commit).
- subscribeAll's messageStore parameter retained as `_messageStore` for
  callsite stability while the rest of the legacy store is being retired.
- Deleted packages/web/src/core/dto/PacketToMessageDTO.ts (no remaining
  consumers; SDK has its own MessageMapper at
  packages/sdk/src/features/chat/infrastructure/MessageMapper.ts).

Web tests (294) still green; production build clean.

Out of scope, queued:
- useConnections refactor (the hook is overdue for cleanup)
- Strip dead methods from the Zustand messageStore (saveMessage,
  getMessages, setMessageState, getDraft, setDraft, clearDraft +
  Zustand-persist + IDB wrapper). Requires a follow-up sweep of the
  remaining test files that mock those methods.
- Migrate unread counts to the SDK (cross-cutting between chat + nodes).
…ssagesDialog

Rounds out the chat slice with destructive operations so the UI dialog no
longer needs the legacy Zustand deleteAllMessages.

packages/sdk
- MessageRepository port gains clearConversation(key); InMemory and
  Sqlocal adapters implement it (SQL: scoped DELETE by conversation).
- ChatStore gains clearBucket(key) + clearAll().
- ChatClient.clearConversation(conv) and ChatClient.clearAll(): empty the
  in-memory store, drop the hydrated-marker so a future subscribe
  re-fetches fresh, then delete from the repository. Repository failures
  are swallowed — UI must not get stuck behind a write error.

packages/sdk-storage-sqlocal
- SqlocalMessageRepository.clearConversation: DELETE FROM messages WHERE
  (device_id, conversation_key) match.

packages/web
- DeleteMessagesDialog swaps useMessages().deleteAllMessages for
  useActiveClient()?.chat.clearAll(). No-active-client path is a no-op
  but still closes the dialog.
- Test file updated: mocks useActiveClient; new case covers
  no-active-client safety.

Totals: sdk 39 tests unchanged (clearConversation tested transitively via
DeleteMessagesDialog; adapter-specific test queued for follow-up),
sdk-storage-sqlocal 18 unchanged, web 295 (+1).
…istence

Adds SDK-side node persistence so a fresh page load rehydrates the mesh
NodeDB from disk before any device packets arrive. Web's existing Zustand
nodeDBStore continues to work unchanged in parallel — UI consumer
migration to useNodes/useNode lands in follow-up commits.

packages/sdk
- New NodesRepository port: loadAll / get / upsert / upsertBatch / remove /
  clear. InMemoryNodesRepository ships as the default.
- NodesClient takes optional { repository }, hydrates on construction,
  writes through on every onNodeInfoPacket, and keeps live signal +
  persistence in lockstep. remove/reset clear the repository alongside the
  store + drive the legacy admin message.
- MeshClientOptions exposes a `nodes` slot mirroring the existing `chat`
  slot.

packages/sdk-storage-sqlocal
- New SqlocalNodesRepository implementing the port. user / position /
  deviceMetrics serialized as base64-encoded protobuf bytes (stable
  across schema additions). Subpath export at "@meshtastic/sdk-storage-
  sqlocal/nodes".
- 6 vitest cases covering upsert + loadAll round-trip, overwrite, proto
  round-trip preserves user fields, remove, clear, multi-device isolation.

packages/web
- useConnections opens a SqlocalNodesRepository alongside the chat /
  draft repos and passes it to the new MeshDevice constructor.

Test counts: sdk 39 (unchanged — repo round-trip exercised in storage
adapter tests), sdk-storage-sqlocal 24 (+6), web 295 unchanged. Build
clean.
…r hook

Lays the groundwork for migrating web's 31 nodeDB consumers off the
Zustand `useNodeDB().getNodes/getNode` API onto SDK signals, without a
big-bang rewrite.

packages/sdk
- Node domain entity gains channel, viaMqtt, hopsAway, isKeyManuallyVerified
  fields. NodeMapper.fromProto now copies these from the inbound
  Protobuf.Mesh.NodeInfo. Required so downstream UIs that show "X hops
  away" / "via MQTT" / "encryption verified" can read SDK nodes without
  losing fidelity.

packages/web
- New core/hooks/useNodesLegacy.ts: useNodesLegacy() returns
  Protobuf.Mesh.NodeInfo[] derived from SDK signals; useNodeLegacy(num)
  returns a single NodeInfo. Components migrate one at a time by swapping
  useNodeDB().getNodes / getNode call sites for these hooks; templates
  stay unchanged because the result shape matches.

No consumer rewrites in this commit — that is per-component follow-up
work to keep diffs reviewable. Web tests (295) still green; production
build clean. SDK 39 / sdk-storage-sqlocal 24 unchanged.
First consumer migration off the legacy Zustand nodeDBStore onto
SDK-managed nodes. Pages/Nodes/index.tsx now pulls the full node array
through useNodesLegacy() — which subscribes to the SDK NodesClient
signal underneath — and applies the existing nodeFilter predicate
client-side via useMemo.

Behavior parity:
- Same Protobuf.Mesh.NodeInfo shape rendered in the table (the legacy
  adapter ensures channel / viaMqtt / hopsAway / publicKey survive).
- Same debounce semantics — only the underlying source changed.
- hasNodeError + nodeErrors continue to come from the Zustand
  nodeDBStore until PKI-error tracking is migrated to the SDK in a
  follow-up commit (validation logic still in
  packages/web/src/core/stores/nodeDBStore/nodeValidation.ts).

The legacy nodeDBStore still populates from subscriptions.ts, so this
rewrite is reversible and runs alongside the SDK source. Web tests
(295) still green; production Vite build clean.
Sweeps the most-touched UI paths off useNodeDB().getNode/getMyNode/getNodes
and onto the useNodesLegacy / useNodeLegacy / useMyNodeLegacy adapters
introduced earlier. Behaviour parity preserved — the adapter returns the
same Protobuf.Mesh.NodeInfo shape components already render.

Migrated:
- components/Sidebar.tsx — myNode + node count from SDK signals.
- components/CommandPalette/index.tsx — node lookup for connection labels.
- components/UI/Avatar.tsx — single-node lookup.
- components/Dialog/RemoveNodeDialog.tsx — selected node display.
- components/Dialog/PKIBackupDialog.tsx — myNode for download/print headers.
- components/Dialog/LocationResponseDialog.tsx — sender lookup.
- components/Dialog/TracerouteResponseDialog.tsx — endpoint lookups.
- components/Dialog/NodeDetailsDialog.tsx — selected node detail.
- components/PageComponents/Settings/User.tsx — myNode for owner edit.
- components/PageComponents/Settings/Position.tsx — myNode for current position.
- components/PageComponents/Messages/TraceRoute.tsx — hop name lookup.
- components/PageComponents/Messages/MessageItem.tsx — myNode (suspending) +
  message author. The polling Suspense fallback now retriggers on the SDK
  signal instead of polling the Zustand store directly.
- components/PageComponents/Map/Popups/WaypointDetail.tsx — locked-to node.
- pages/Messages.tsx — sidebar node list + selected peer lookup.
- pages/Map/index.tsx — full filtered list + myNode for fitting. Drops the
  now-dead NODEDB_DEBOUNCE_MS constant since the SDK signal layer handles
  re-render coalescing internally.
- TraceRoute.test.tsx mocks updated to mock useNodesLegacy instead of
  useNodeDB.

NodesLayer.tsx, RemoveNodeDialog (removeNode), ResetNodeDbDialog, and
RefreshKeysDialog still depend on hasNodeError / removeNode /
removeAllNodes on the legacy nodeDB store — those move when PKI-error
tracking and the admin-message paths are migrated to the SDK in the
unread/cleanup follow-up.

Web tests: 295 still green; production Vite build clean. No SDK changes
in this commit.
The "Legacy" suffix on the node-adapter hooks reads as if the hook
itself is deprecated, when in fact the hook is the bridge: it converts
SDK Node domain entities into Protobuf.Mesh.NodeInfo so consumer
templates that destructure proto fields keep working during the
migration.

- useNodesLegacy → useNodesAsProto
- useNodeLegacy → useNodeAsProto
- useMyNodeLegacy → useMyNodeAsProto
- File renamed to packages/web/src/core/hooks/useNodesAsProto.ts.

All 16 call sites updated. Test mock in
packages/web/src/components/PageComponents/Messages/TraceRoute.test.tsx
updated to the new module path.

Behaviour unchanged. 295 tests still green; production Vite build clean.

(useChatLegacy follows the same pattern but maps to a hand-written web
type, not a proto — it stays as-is for now and is queued for renaming
once the broader chat-store cleanup lands.)
The favorite/ignore admin-message paths now go through the SDK NodesClient
instead of the legacy Zustand sendAdminMessage + manual proto build.

- useFavoriteNode: meshClient.nodes.favorite(nodeNum) /
  meshClient.nodes.unfavorite(nodeNum). The SDK already has
  FavoriteNodeUseCase + AdminMessageSender behind these APIs.
- useIgnoreNode: meshClient.nodes.ignore / .unignore.
- Both still mirror the optimistic flag flip into the legacy nodeDB store
  via updateFavorite / updateIgnore until that store is fully retired —
  same TODO as before, "wait for ack before flipping".

Tests rewritten to mock @meshtastic/sdk-react via vi.hoisted (so the
factory captures the spy without hitting the vi.mock hoist trap):
- assert SDK favorite/unfavorite/ignore/unignore calls
- assert legacy store mirror still fires
- assert toast + longName fallback behaviour preserved
- assert no-op when the node isn't in the SDK store yet (parity with
  the prior "getNode returned undefined" guard)

Web tests: 295 still green; production Vite build clean.
- ResetNodeDbDialog: connection.resetNodes() → meshClient.nodes.reset();
  on success it now also calls meshClient.chat.clearAll() instead of the
  legacy useMessages().deleteAllMessages(). PKI-error tracking and the
  in-memory nodeDB still get cleared via removeAllNodeErrors /
  removeAllNodes since those subsystems have not yet migrated. Test
  rewritten to mock useActiveClient via vi.hoisted.
- RefreshKeysDialog: drops useNodeDB().getNode in favour of
  useNodeAsProto for the missing-key node display. Test wraps render in
  a MeshRegistryProvider with an empty registry so the adapter resolves
  cleanly with no active client.

Adapter hardening (useNodesAsProto.ts)
- Switched off useNodes() / useMeshDevice() (which both throw outside a
  MeshProvider/MeshRegistryProvider with an active client) onto
  useActiveClient() + a no-op signal fallback. The hooks now return [] /
  undefined when there is no active client instead of throwing, which
  fixes RefreshKeysDialog's "no error → render null" path under tests
  that don't connect a device.

Web tests: 295 still green; production Vite build clean.
Moves PKI mismatch / duplicate-key validation and routing-error tracking
out of the legacy Zustand nodeDBStore and into the SDK NodesClient so
every consumer reads the same source of truth. Validates Android's gap
analysis: Meshtastic-Android does not track per-node PKI errors at all,
so the SDK must own this concern client-side.

packages/sdk
- New NodeError + NodeErrorType (Routing_Error | "MISMATCH_PKI" |
  "DUPLICATE_PKI"). Mirrors the previous web-only types.
- New nodeValidation infrastructure mapper (pure): byte-equal compare on
  publicKey, returns ValidatedNodeInfo { accepted?, error? }. Ports the
  detection rules verbatim from
  packages/web/src/core/stores/nodeDBStore/nodeValidation.ts.
- NodesClient gains an errors signal + errorFor / hasError / setError /
  clearError / clearAllErrors API. handleIncoming runs validation before
  upserting; conflicts skip the store write and record an error. A clean
  refresh of a previously-flagged node clears the error.
- handleRoutingPacket records PKI_UNKNOWN_PUBKEY / PKI_FAILED /
  PKI_SEND_FAIL_PUBLIC_KEY / NO_CHANNEL against packet.from.
- 5 new vitest cases covering MISMATCH_PKI, DUPLICATE_PKI, error-clear-on-
  refresh, routing-error capture, clearError / clearAllErrors. SDK total:
  44 tests, all green.

packages/sdk-react
- New useNodeErrors / useNodeError(num) / useHasNodeError(num) hooks.
  All resolve through useActiveClient() and fall back to empty when no
  client is present.

packages/web
- pages/Nodes/index.tsx, pages/Messages.tsx, components/PageComponents/
  Map/Layers/NodesLayer.tsx swap useNodeDB().hasNodeError for the SDK
  useNodeErrors hook + a memoised Set lookup.
- components/Dialog/RefreshKeysDialog/RefreshKeysDialog.tsx reads the
  current error from useNodeError(activeChat); useRefreshKeysDialog.ts
  drives meshClient.nodes.clearError + meshClient.nodes.remove instead
  of the legacy store. The companion test still passes through the
  empty-MeshRegistry render path because useNodeError tolerates a
  missing active client.
- core/subscriptions.ts no longer calls nodeDB.setNodeError on PKI /
  no-channel routing packets — the SDK records those itself. The dialog
  open trigger stays here.
- pages/Nodes/index.tsx drops the now-dead NODEDB_DEBOUNCE_MS constant.

Web tests: 295 still green; production Vite build clean. The legacy
nodeDB still runs its internal validation but its error map is now a
dead end — removal queued in the plan-file follow-up alongside the rest
of the nodeDB cleanup.
Now that the SDK NodesClient owns public-key validation and per-node
error tracking, the legacy Zustand mirrors are dead code.

packages/web/src/core/stores/nodeDBStore/index.ts
- Drops the nodeErrors map + setNodeError / getNodeError / hasNodeError /
  clearNodeError / removeAllNodeErrors methods from the NodeDB interface
  and factory.
- Drops the validateIncomingNode call inside addNode and inside
  setNodeNum's merge path; legacy mirror is now a straight last-write-
  wins shallow merge. The SDK runs validation independently against its
  own snapshot.
- Drops the nodeErrors entries from the persisted partialize shape.

packages/web/src/core/stores/nodeDBStore/types.ts
- Removes NodeError + NodeErrorType. ProcessPacketParams stays.

packages/web/src/core/stores/nodeDBStore/nodeValidation.ts — deleted
(SDK ports it at packages/sdk/src/features/nodes/infrastructure/
nodeValidation.ts and exposes the verdict via NodesClient).

packages/web/src/core/stores/index.ts — drop the dead NodeErrorType
re-export.

packages/web/src/core/stores/nodeDBStore/nodeDBStore.test.tsx
- Removes tests covering the migrated PKI behaviour (errors map,
  MISMATCH, DUPLICATE, "unions nodeErrors") — equivalent coverage now
  lives in packages/sdk/src/features/nodes/NodesClient.errors.test.ts.
- Trims the surviving merge-semantics tests to the simpler last-write-
  wins shape.
- The "selector re-renders" test swaps the deleted setNodeError mutation
  for an updateFavorite call to keep the slice-stability assertion alive.
- The "addNodeDB instance identity" assertion relaxes from .toBe to
  content equality — immer's pruneStaleNodes path can reseat the entry,
  but the registered DB's id stays stable.

components/Dialog/ResetNodeDbDialog/ResetNodeDbDialog.tsx — calls
meshClient.nodes.clearAllErrors() instead of the legacy
removeAllNodeErrors. Test mocks updated.

components/Dialog/RefreshKeysDialog/* — already migrated to SDK errors
in the previous commit; no further changes here.

Test counts: web 290 (was 295 — 5 PKI-tracking tests retired in
favour of 5 equivalent SDK tests). Production Vite build clean.
                 favourite-flag flips; legacy mirror retired

The SDK NodesClient now subscribes to onUserPacket / onPositionPacket /
onMeshPacket so user-record updates, GPS updates, and lastHeard / snr
refreshes flow into the signal-backed store + repository directly. The
favourite / ignored toggles also flip the local flag once the admin
message resolves successfully, so the UI no longer needs to mirror the
state into a parallel Zustand store.

packages/sdk
- NodesClient: new private patch(num, partial) helper that shallow-
  merges into the existing entry (or seeds a placeholder Node) and
  upserts to the repository in one shot. Used by:
  - onUserPacket → patch user
  - onPositionPacket → patch position
  - onMeshPacket → patch lastHeard / snr (replaces the legacy nodeDB
    processPacket flow)
  - favorite / unfavorite / ignore / unignore — flag flip on Result.ok
- NodesClient.reset({ keepMyNode? }) — preserves the local node entry
  when requested. Mirrors the previous removeAllNodes(true) semantics
  the ResetNodeDb dialog relied on.

packages/web
- core/subscriptions.ts: drops the onUserPacket / onPositionPacket /
  onNodeInfoPacket / onMeshPacket → legacy nodeDB write paths. The
  unused nodeDB parameter is renamed `_nodeDB` for callsite stability;
  it will disappear when the store itself is deleted in a follow-up.
- core/hooks/useFavoriteNode + useIgnoreNode: drop the legacy
  updateFavorite / updateIgnore mirror — the SDK flips the flag on
  success.
- components/Dialog/RemoveNodeDialog: now calls meshClient.nodes.remove
  exclusively; no legacy fall-through.
- components/Dialog/ResetNodeDbDialog: calls
  meshClient.nodes.reset({ keepMyNode: true }) and meshClient.chat.clearAll().
  No legacy nodeDB calls.
- Test mocks for useFavoriteNode / useIgnoreNode / ResetNodeDbDialog
  pruned to match.

Test counts: sdk 44 (unchanged), sdk-react 8, sdk-storage-sqlocal 24,
web 290. Production Vite build clean.

Remaining surface on the legacy nodeDB: addNode / addUser / addPosition /
processPacket / setNodeError-equivalent are now unused; the store retains
only updateFavorite / updateIgnore (no callers) and the per-device
plumbing required by the deviceContext hooks. Full deletion of
nodeDBStore queued in the plan-file follow-up.
useMyNodeAsProto already replaces getMyNode here; the import was left
behind during the nodeDB mirror retirement.
…ource of truth

The Zustand nodeDBStore had no remaining writers (the mirror in
subscriptions.ts was retired in 005d000) and no remaining readers
outside of bookkeeping (addNodeDB / setNodeNum / removeNodeDB calls
that fed nothing). Drop the whole store along with its persistence
shim and the persistNodeDB feature flag.

- Delete packages/web/src/core/stores/nodeDBStore/.
- Drop the _nodeDB parameter from subscribeAll; it was unused.
- Drop addNodeDB / useNodeDBStore wiring from useConnections; the
  meshDeviceId reuse check now keys off the device store instead.
- useNewNodeNum no longer pokes the nodeDB.
- FactoryResetDeviceDialog drops its removeNodeDB call (and the test
  drops the matching expectation).
- Strip persistNodeDB / VITE_PERSIST_NODE_DB from featureFlags +
  dev-overrides.
- Remove the useNodeDB / useNodeDBStore re-exports + the
  bindStoreToDevice wiring from core/stores/index.ts.
The persisted message Zustand store had no remaining consumers — chat
history, drafts, and message state all live on the SDK ChatClient
(SqlocalMessageRepository). The messageStore Zustand surface
(saveMessage, getMessages, setMessageState, getDraft, setDraft,
clearDraft, deleteAllMessages, clearMessageByMessageId, plus the
addMessageStore / removeMessageStore / getMessageStore / setNodeNum
plumbing) is now fully unused.

Collapse messageStore/index.ts to just the MessageState / MessageType
enums + the legacy `Message` shape that the useChatLegacy adapter and a
couple of message components still consume. Delete the Zustand
implementation and its 32-test suite.

Other knock-on cleanups in this commit:
- Drop the _messageStore param from subscribeAll (unused after the
  saveMessage-from-subscriptions retirement).
- Drop addMessageStore wiring from useConnections, removeMessageStore
  from FactoryResetDeviceDialog, setNodeNum from useNewNodeNum.
- Drop the message branch from the router context (no readers).
- RefreshKeysDialog stops keying off `useMessages().activeChat` (which
  was permanently 0 — a dead handle that quietly broke key-refresh
  UX). Use the SDK NodesClient's first error directly via
  `useNodeErrors()[0]`. The dialog manager already opens the dialog on
  PKI_UNKNOWN_PUBKEY, so picking the first error matches the intended
  flow.
The "legacy" suffix on its own read like the hook itself was deprecated;
the actual purpose is "use SDK chat history but project it into the
legacy Message shape." Rename for clarity. Companion type names follow
suit (UseChatAsLegacyMessages{Broadcast,Direct,Params}).
…dirty state

Phase 3c-2 of PR #9. With every web settings page already migrated, the
deviceStore changeRegistry plumbing has no remaining writers or readers
and is deleted.

ConfigEditor (sdk):
- New queueAdminMessage(message) for one-off side flows. Queue drains
  inside commit() between beginEdit and commitEdit. Used by
  Position.tsx for setFixedPosition; future side flows (e.g. SetTime)
  reuse the same queue.
- Disconnect / commit success now also reset the admin-message queue.
- isDirty includes a queued admin-message check.

Web Settings/index.tsx:
- handleSave is now a one-liner: editor.commit(). All the legacy
  get*Changes / get*ChangeCount / clearAllChanges / connection.setConfig
  loops + the post-commit deviceStore mirror writes are gone.
- handleReset calls editor.reset() instead of clearAllChanges().
- Save-button gating reads editor.isDirty + RHF dirty.
- Section change-count badges read directly from editor.dirtyRadioSections /
  editor.dirtyModuleSections / editor.dirtyChannels lengths.

Tab dirty-dot indicators (RadioConfig.tsx / DeviceConfig.tsx /
ModuleConfig.tsx): rewired from hasConfigChange/hasModuleConfigChange/
hasChannelChange/hasUserChange to editor signals (dirtyRadioSections,
dirtyModuleSections, dirtyChannels, isOwnerDirty).

Position.tsx: queueAdminMessage call now goes through editor — pending
fixed-position coordinates ride along with the rest of the editor's
dirty state and ship under the same beginEdit/commitEdit window.

deviceStore.ts:
- Drops the entire changeRegistry surface from the Device interface and
  factory: setChange / removeChange / hasChange / getChange /
  clearAllChanges / hasConfigChange / hasModuleConfigChange /
  hasChannelChange / hasUserChange / getConfigChangeCount /
  getModuleConfigChangeCount / getChannelChangeCount /
  getAdminMessageChangeCount / getAllConfigChanges /
  getAllModuleConfigChanges / getAllChannelChanges / queueAdminMessage /
  getAllQueuedAdminMessages, plus the changeRegistry field.
- getEffectiveConfig / getEffectiveModuleConfig collapse to plain
  device.config[variant] / device.moduleConfig[variant] passthroughs
  (kept as a thin compatibility shim for residual callers; the
  changeRegistry merge is gone).
- types.ts: ValidConfigType / ValidModuleConfigType derive directly
  from LocalConfig / LocalModuleConfig keys (was imported from the
  deleted changeRegistry.ts).
- The 252-line changeRegistry.ts file is deleted; ~250 lines also drop
  out of the deviceStore factory.
- deviceStore.test.ts loses the change-registry describe block (3
  tests). 237 web tests still passing (was 240; -3 dead tests).

Net: -1066 / +136 lines.
…TelemetryRepository (PR #10)

Telemetry slice gains the same persistence shape as chat / nodes: a
repository port on the SDK side, an in-memory default, an OPFS-backed
SQLite adapter on the storage package side, and lazy hydration into the
in-memory store.

SDK (@meshtastic/sdk):
- New TelemetryRepository port (loadRecent, loadBefore, append,
  appendBatch, prune, clearNode, clear) + TelemetryRetentionPolicy
  (maxPerNode, olderThanMs).
- New InMemoryTelemetryRepository — default when no adapter is wired.
- TelemetryClient grows TelemetryClientOptions { repository?, retention? }.
  Each incoming onTelemetryPacket appends to both the in-memory store
  and the repository, then the configured retention policy prunes the
  repository.
- latest(nodeNum) and history(nodeNum) lazy-hydrate the in-memory store
  from the repository on first subscribe (HYDRATE_LIMIT = 256). loadBefore
  passes through to the repository for paged reads.
- clearNode / clear methods exposed on the client.
- MeshClient gains options.telemetry which is passed through to
  TelemetryClient.

@meshtastic/sdk-storage-sqlocal:
- New SqlocalTelemetryRepository under ./telemetry subpath. Drizzle-typed
  insert/select/delete against the existing telemetry table; payload is
  stored as base64 of the proto bytes (per-kind schema lookup) so the
  wire shape is the source of truth across schema additions.
- prune({ maxPerNode }) trims with a per-node offset query. prune({
  olderThanMs }) deletes by ts cutoff.
- Six new tests cover round-trip, ascending order, deviceId scoping,
  per-node retention, age-based prune, and proto payload preservation.
- Re-exports added to mod.ts + ./telemetry subpath in package.json.

web:
- useConnections wires SqlocalTelemetryRepository per connection with
  retention { maxPerNode: 500, olderThanMs: 30 days }.

Tests: SDK 57 (was 49 — +8 new), storage 30 (was 24 — +6), web 237.
Build clean.
ChatClient gains a `chat.unread` namespace mirroring `chat.drafts`:
- byKey: ReadonlySignal<Map<conversation-key-string, number>>
- total: ReadonlySignal<number>
- count(key): ReadonlySignal<number>
- markRead(key): void

Increments happen in the existing onMessagePacket subscriber when
packet.from !== client.myNodeNum (so the outbound echo of our own send
doesn't bump anything). Uses the same ConversationKey shape the rest
of the chat slice already keys by, so direct messages keyed by peer
and broadcasts by channel can never collide.

Counts are in-memory only — persistence will come later if needed
(would require tracking a lastReadAt timestamp per conversation in the
message repository and computing unread on hydrate).

sdk-react: new useTotalUnread / useUnreadCount(key) / useUnreadByKey
hooks. All fall back to safe empty values when no client is active.

Web migration:
- Sidebar reads useTotalUnread instead of summing
  deviceStore.unreadCounts.
- MessagesPage: per-channel SidebarButton counts read from
  useUnreadByKey()'s `channel:N` keys; per-direct-message counts from
  `direct:N` keys. Click handlers call meshClient.chat.unread.markRead
  with the typed ConversationKey instead of resetUnread.
- subscriptions.ts: drops onMessagePacket -> incrementUnread mirror
  (SDK now owns it). myNodeNum local goes away too — nothing else
  consumed it.
- deviceStore: removes unreadCounts field + incrementUnread /
  resetUnread / getUnreadCount / getAllUnreadCount methods. The mock
  + test block in deviceStore.test.ts also drop their unread coverage.
- deviceStore.mock.ts: also strips the dead changeRegistry / setChange
  / hasConfigChange / etc. methods that survived from the earlier
  changeRegistry deletion.

Tests: SDK 61 -> 65 (+4 ChatClient.unread.test), sdk-react 8, web 236.
Build clean.
useConnections went from 661 lines down to 264 by extracting three
single-responsibility helpers under packages/web/src/core/connections/.
The hook is now pure orchestration; transport / heartbeat / SDK-client
/ status-probe concerns live in their own files.

New modules:
- core/connections/heartbeat.ts (45 LOC): startConfigHeartbeat (5s),
  startMaintenanceHeartbeat (5min), stopHeartbeat. Owns the heartbeats
  Map; replaces the inline interval bookkeeping. Both helpers stop the
  prior heartbeat first so callers don't have to.
- core/connections/sdkClient.ts (61 LOC): buildMeshDevice(connId,
  deviceId, transport) opens the OPFS DB, builds the four sqlocal
  repositories (chat / draft / nodes / telemetry), and constructs the
  MeshDevice with the canonical retention defaults (chat 90d / 1k per
  bucket; telemetry 30d / 500 per node). Falls back to in-memory when
  sqlocal is unavailable.
- core/connections/transports.ts (236 LOC): per-transport openTransport
  factory (HTTP reachability check, BT permission re-acquisition with
  optional prompt, Serial port lookup with close-then-reopen),
  probeConnection for refreshStatuses, closeTransport for cleanup.
  Discriminates over conn.type so useConnections doesn't carry the
  switch logic.

useConnections (now 264 LOC):
- Owns the cachedTransports + configSubscriptions maps and the
  Zustand selectors.
- teardown(id, conn) consolidates the heartbeat + config-sub +
  meshDevice.disconnect + transport-close cleanup that used to be
  duplicated across removeConnection / disconnect.
- connect calls openTransport and forwards the resulting transport +
  cached BT/Serial handle into setupMeshDevice. The BT
  gattserverdisconnected listener is wired here (it needs the connId).
- refreshStatuses simplifies to a filter + Promise.all over
  probeConnection.
- syncConnectionStatuses unchanged.

Behavior is unchanged. No new tests — the existing web suite (236)
still passes; build clean.
Web is a deployable SPA, not a library — no @meshtastic scope, no
exports map, no npm/JSR publish target, no other workspace member
depends on it. Conventional pnpm/Turbo/Nx layout puts deployables
under apps/ and libraries under packages/. Doing the move now (after
PR #9 / #10 / unread / useConnections refactor land but before Phase
C ships) so the lib-only packages/* directory remains stable as we
delete packages/core.

Mechanics:
- pnpm-workspace.yaml: add `apps/*` to packages.
- git mv packages/web apps/web (entire directory tree).
- vite.config.ts uses process.cwd(), so no internal path edits needed.
- pnpm filter commands still resolve by package name (`pnpm --filter
  meshtastic-web run build`) — no script changes.

External-path consumers updated:
- README.md table row.
- tsconfig.json reference.
- 7 GH workflows (pr.yml, nightly.yml, release-web.yml, the three
  crowdin-*.yml, release-packages.yml). The
  packages/release-packages.yml `packages/* | grep -v packages/web`
  filter is now redundant since web isn't in packages/* anymore — it
  collapses to plain `ls -d packages/*`.

Verified: web build clean, web 236 / sdk 65 / sdk-react 8 / storage
30 tests all pass.
…meshtastic/sdk, bump majors

The legacy `@meshtastic/core` package is gone. The six transport-*
packages and the web app no longer depend on it; everything routes
through `@meshtastic/sdk`.

Transport packages (transport-deno, transport-http, transport-node,
transport-node-serial, transport-web-bluetooth, transport-web-serial):
- package.json deps swap `@meshtastic/core: workspace:*` for
  `@meshtastic/sdk: workspace:*`.
- src/transport.ts + src/transport.test.ts imports point at
  `@meshtastic/sdk` (the `Types` and `Utils` namespaces are still
  exported from the SDK so source code is otherwise unchanged).
- README.md examples updated to import from `@meshtastic/sdk`.

apps/web:
- Drops the leftover `@meshtastic/core` workspace dep; nothing in
  apps/web/src imports from it.

@meshtastic/sdk:
- README description loses the "Replaces @meshtastic/core" suffix —
  there's nothing left to replace.
- Three legacy shim files keep their re-exports but their docstrings
  drop the "Phase-A shim, removed in Phase C" framing. The MeshDevice
  facade survives because the web app's `connection.factoryResetDevice()`
  / `connection.reboot()` / etc. callsites still go through it; new
  consumers should reach into `client.config` / `client.chat` / etc.
- New scripts/rename-dts.mjs is a postbuild step that renames tsdown's
  hashed entry-point dts outputs (`mod-<hash>.d.ts` etc.) back to their
  canonical names so package.json `types` and downstream dts-bundlers
  can find them. Internal chunk dts files (e.g. `Transport-<hash>.d.ts`)
  are intentionally NOT renamed because mod.d.ts imports them by path.
- build:npm runs tsdown then the rename script.

Version bumps (signal: now require @meshtastic/sdk):
- @meshtastic/sdk             0.1.0 -> 1.0.0
- @meshtastic/sdk-react       0.1.0 -> 1.0.0
- @meshtastic/sdk-storage-sqlocal 0.1.0 -> 1.0.0
- @meshtastic/transport-http       0.2.5 -> 1.0.0
- @meshtastic/transport-web-serial 0.2.5 -> 1.0.0
- @meshtastic/transport-web-bluetooth 0.1.5 -> 1.0.0
- @meshtastic/transport-deno       0.1.1 -> 1.0.0
- @meshtastic/transport-node       0.0.2 -> 1.0.0
- @meshtastic/transport-node-serial 0.0.2 -> 1.0.0

Workflows: nothing referenced packages/core directly (only the
release-packages.yml glob which already collapsed in the apps/web
move). README packages table loses the legacy `packages/core` row.

Verified: web build clean, all four package suites green
(web 236, sdk 65, sdk-react 8, storage 30).

The transport packages' dist build is currently blocked by a tsdown
cross-chunk dts resolution glitch where the `Types` namespace import
can't follow `mod.d.ts`'s reference to the internal `Transport-<hash>`
chunk. Runtime is fine — only the published-types path is affected,
and transport packages publish via their own release flow. Logged as
follow-up.
…to "configuring" before DB open; await transport disconnect

Three reconnect-flow bugs:

1. probeConnection returned "configured" for serial / bluetooth when
   the browser had stored permission for the device — but permission
   ≠ a configured Meshtastic device. The card showed "connected"
   and a Disconnect button before the user had ever clicked Connect.
   Probe now returns "online" (= "available, click to connect"),
   matching the HTTP path. "configured" is reserved for the
   onConfigComplete callback.

2. setupMeshDevice flipped status from "connecting" → "configuring"
   AFTER awaiting buildMeshDevice, which opens the OPFS DB. If the
   DB open stalled (multi-tab contention, slow first-time init), the
   card sat at "connecting" indefinitely. Move the status flip ahead
   of the persistence await — UI shows "configuring" while
   persistence is spinning up.

3. teardown's `device?.connection?.disconnect()` returned a Promise
   that was never awaited, so the underlying transport's port.close()
   could race the next port.open() on a fast disconnect → reconnect.
   Make teardown async, await disconnect, propagate via async
   removeConnection / disconnect callers.

Plus: connect() catch block logs the underlying error before turning
it into a status update, so the actual failure shows up in console
even when the toast text is generic.
…e capability gate

Two backlog UX nits.

DynamicForm:
- FormGroup gains an optional `footer?: ReactNode` slot that renders
  after the field list. Lets pages drop arbitrary JSX (action buttons,
  notes, etc.) into a card without subclassing the form.

Position:
- New `UseBrowserLocationButton` component lives inside the Device GPS
  card via the new `footer` slot. Calls navigator.geolocation, writes
  lat/lng/altitude into the form via `useFormContext` (the
  PositionValidation form is wrapped in FormProvider by DynamicForm).
- Truncates lat/lng to 7 decimals to match the proto's max precision.
- Gracefully no-ops when navigator.geolocation is unavailable
  (insecure context).
- Failed reads surface as a toast carrying the GeolocationPositionError
  message; busy state disables the button + flips the label.
- New i18n keys: position.useBrowserLocation.{label,busy,failed}.

Telemetry:
- `device_telemetry_enabled` is only writable on firmware ≥ v2.7.12
  (mirrors Android's `Capabilities.canToggleTelemetryEnabled`). Adds a
  small `firmwareAtLeast` semver helper and reads
  `device.metadata.get(0)?.firmwareVersion`. The toggle is hidden on
  older firmware so we don't push a value the device will silently
  drop.
- Unparseable / unknown firmware versions default to "show the toggle"
  (rather than hide), so we don't accidentally hide the control on
  devices we can't classify.

Build clean, 236 web tests still pass.
…ils namespace

Previously transport-* package dts builds failed with "Missing export" because
rolldown's dts bundler couldn't follow tsdown's cross-chunk re-exports out of
@meshtastic/sdk. Three things land here:

1. mod.ts now re-exports `fromDeviceStream`, `toDeviceStream`, `Queue`, and
   `Xmodem` directly. Transport packages were importing these via the legacy
   `Utils` namespace; namespace re-exports are exactly what rolldown chokes
   on. Transport packages (http, deno, node, node-serial, web-bluetooth,
   web-serial) and shared transportContract.ts now use direct imports —
   `import { Transport, DeviceOutput, ... } from "@meshtastic/sdk"`.

2. package.json `exports` split: `types` points at `./dist/<entry>.d.ts`
   (built dts) so tsdown's downstream rolldown reads a self-contained file;
   `default` keeps source paths so vitest + workspace-internal imports still
   resolve to .ts source.

3. Postbuild `rename-dts.mjs` now (a) inlines internal chunk dts files
   (Transport-<hash>.d.ts) into each entry's dts, rewriting the letter
   aliases tsdown produces; (b) strips `type` modifiers from re-exports
   so rolldown accepts the bindings as both value- and type-shape;
   (c) declares the synthetic `Types`/`Utils` namespaces explicitly,
   replacing the broken `<wrapper>_d_exports as Types` aliases tsdown
   emits for `export * as` patterns.

vitest.config.ts gains `apps/*` so the workspace move keeps test discovery
working for the apps/web project.
- Drop bogus `dts.isolatedDeclarations` field from SDK (not a real tsdown
  option; tsdown was silently ignoring it).
- Make every package explicit about the perf-relevant flags rather than
  relying on defaults that drift between tsdown versions: target=esnext
  (no syntax downleveling), sourcemap=false, minify=false, treeshake=true,
  report=false, splitting=false.
- Modest wall-clock wins: SDK 1020→752ms, transports ~720→~600ms.

`isolatedDeclarations: true` in tsconfig would unlock the bigger oxc-dts
fast path but the codebase isn't ID-clean yet — left as future work.
Sweep covering several unrelated cohorts:

- Delete dead `NewDeviceDialog.tsx` (commented-out in App.tsx, imports point at
  non-existent `@components/PageComponents/Connect/*`).
- Add `better-result` to sdk-react deps so `useChat`/`useDirectChat`/
  `useFavoriteNode`/`useIgnoreNode` resolve.
- Fix `test-utils.tsx`: route at `routeTree.gen.ts` was never generated;
  re-export the existing `routeTree` constant from `routes.tsx` and pass
  the missing router context + DeviceWrapper deviceId.
- Pre-existing `noUncheckedIndexedAccess` violations in `pskSchema.test.ts`,
  `x25519.ts`, `Table/index.tsx` + test, `useCopyToClipboard` (Timeout vs
  number) — non-null asserts where the index is guaranteed by the
  surrounding code, plus a defensive `if (!cell) return 0` in the table sort.
- Immer `WritableDraft<T>` mismatches against SDK's `MeshDevice`/`Device`
  classes — cast through `Draft<X>` where the inferred recursive draft type
  can't represent SDK class internals.
- `deviceStore.mock` updated for the `connectionPhase`/`connectionId` +
  setters added when the connection lifecycle was extracted.
- `HeatmapLayer` accepts the `isVisible` prop everyone else's layer accepts.
- `SNRTooltipProps.to` widened to `string | undefined` (single-node hover
  has no `to`).
- Zod resolver returns `Record<string, never>` for the error-path values
  shape that react-hook-form expects.
- Mock `NodeInfo` in `useFilterNode.test.ts` gains the `isMuted` proto
  field that landed upstream.
…nnect failure

`NetworkError: Connection attempt failed` from `device.gatt.connect()` is the
most common transient when reconnecting BT — OS stack hiccup, device just
woke from sleep, etc. Adding one retry with a 750ms delay clears most of
those. Persistent failures get a wrapped error message naming the actual
likely causes (out of range, powered off, paired with phone or another
tab) instead of leaving the raw DOMException for the user to interpret.
… typed error

Per the DDD layering — transport-level concerns belong in the transport
package, not the web app's connection orchestrator. Moved the GATT
`NetworkError: Connection attempt failed` retry-once + the user-facing
error wrapping out of `apps/web/src/core/connections/transports.ts` and
into `TransportWebBluetooth.prepareConnection`.

New `BluetoothConnectError` carries:
- `kind: "transient" | "unavailable" | "missing-service"` so callers can
  decide whether to suggest retry vs re-pair vs firmware upgrade.
- `userMessage` — human-readable, actionable string ready for UI.

The web app's `transports.ts` is back to a thin wrapper that just calls
`TransportWebBluetooth.createFromDevice` and surfaces whatever error it
throws. `BluetoothConnectError` is re-exported through the connections
module so the Connections page can `instanceof`-check without depending
on the transport package directly.
Mirrors Meshtastic-Android's `regionUnset` flow. A freshly-flashed device
boots with `Config.LoRa.region == UNSET` and won't transmit at all until
the user picks one. Web now surfaces this:

- `ConfigClient.isRegionUnset: ReadonlySignal<boolean>` — computed from
  the radio config signal, false until the first LoRa packet arrives so
  consumers don't flash the prompt during the connect handshake.
- `useIsRegionUnset()` hook in sdk-react, re-exported from the package's
  public surface.
- `RegionSetupReminder` mounts inside the connected-device branch of
  App.tsx. While the SDK reports `isRegionUnset == true`, a persistent
  toast offers a "Set region" CTA that deep-links into the LoRa tab
  (`/settings/radio`). Toast auto-dismisses the moment a real region is
  committed.

Predicate matches Android exactly (single check, no owner/name/setupComplete
heuristics). Lives in the SDK config slice — UI is a thin consumer.

3 new SDK tests; SDK 64 / sdk-react 8 / web 236.
… connect errors

Pre-flight close + open-with-backoff lived in apps/web's transport
orchestrator — wrong layer per DDD. Lifted into TransportWebSerial
itself.

Changes:
- `TransportWebSerial.createFromPort` (and `create`) now return
  `Result<TransportWebSerial, SerialConnectError>`. No throwing on
  expected failures; callers branch on the Result.
- New `preparePort` private static that:
  * force-closes the port if `readable` / `writable` are non-null
  * opens with up to 4 attempts (250 / 500 / 750 ms backoff) on
    `InvalidStateError` — common during USB re-enumeration
  * tags persistent failures `kind: "in-use"` (other tab / app
    holding the port) vs `"busy"` (transient / unknown) vs
    `"unavailable"` (no streams after open)
- `SerialConnectError` carries `kind` + `userMessage` ready for UI.
- `apps/web/src/core/connections/transports.ts` drops its old
  conditional `port.close()` + 100ms sleep; just calls
  `createFromPort` and unwraps the Result. `SerialConnectError` is
  re-exported alongside `BluetoothConnectError` so the Connections
  page can pattern-match.
- `AnyTransport` definition simplified — was deriving via
  `Awaited<ReturnType<...>>` which broke once `createFromPort`
  returned a Result.
- Test FakeSerialPort updated to mirror the real spec: `readable`
  /`writable` go null on `close()` and re-create on `open()`.
- 3 new port-hygiene tests cover the close-on-prior-open path,
  the backoff-retry success path, and the persistent-`in-use`
  Err path. Suite: 5 → 8.

Matches the better-result preference for new application/use-case
code. BT transport stays on the throwing `BluetoothConnectError`
shape for now — separate refactor if we want to converge.
createLogger now reads `localStorage["mesh-debug"]==="1"` (browser) or
`MESH_DEBUG=1` (node) on each invocation. When set, default minLevel
drops to debug; otherwise stays at info. Existing callers
(MeshClient + co.) are unaffected.

Logs added (all `[name]`-prefixed via tslog) covering the connect
lifecycle the user just hit:

- TransportWebSerial.preparePort: enter, close-needed branch,
  close result, each open() attempt, retry decisions, final error.
- TransportWebSerial constructor: pipe wiring, toDevice pipe abort
  vs reject path.
- TransportWebSerial read loop: start, done, throw (with
  closingByUser flag), reader-lock release.
- TransportWebSerial.disconnect: each cleanup step (abort, pipe
  settled, fromDevice cancel, connection.close).
- transports.ts openSerial: cached-port flag, getPorts count,
  picker invocation, resolved-port stream state, createFromPort
  Result outcome (kind + userMessage on Err).
- useConnections.connect / setupMeshDevice / teardown: phase
  transitions, configure() resolve / reject, onConfigComplete fire,
  BT gattserverdisconnected, transport.disconnect rejection.

Flip on:
  localStorage.setItem("mesh-debug", "1"); location.reload();
The connect path silently hung at `buildMeshDevice` when sqlocal failed
to open. Adding visibility:

- `getStorageDb` logs creation start, success (with elapsed ms), and
  failure (with name + message).
- `buildMeshDevice` races the DB open against a 5s timeout. If sqlocal
  hangs (stale Web Lock from a crashed prior tab, OPFS contention,
  worker hang), the connect path falls through to the SDK's in-memory
  repositories instead of blocking forever — the user gets a working
  session and the warn-level log surfaces the underlying cause.
- `buildMeshDevice` logs entry, DB-ready elapsed, repos-opened, and
  fall-through decisions.
Surfaces what the connection handshake is doing in real time. Per DDD,
the predicate + counters live in the SDK; the web overlay is a thin
consumer of the signal.

SDK
- New `MeshClient.progress: ReadonlySignal<ConnectionProgress>` with the
  state machine `idle → configuring → configured`. Counters
  (config / modules / channels / nodes + boolean myInfo / metadata)
  tally inbound packets each `configure()` cycle.
- `configure()` resets to `configuring` with empty counters.
  `onConfigComplete` flips to `configured`.
- `ConnectionProgress` + `ConnectionProgressCounters` exported from the
  package surface.
- 6 new tests in `MeshClient.progress.test.ts`.

sdk-react
- `useConnectionProgress()` hook reads the active client's signal,
  returns `{ phase: "idle" }` when there is no active client so the
  caller can render unconditionally.

apps/web
- `ConnectingOverlay` redesigned as a terminal-style streaming log:
  - macOS-window-style header (red/yellow/green dots) with a live
    `phase` indicator pulse.
  - Monospace event lines with `→` (active) / `✓` (done) / `•` (info)
    glyphs and a relative-seconds timestamp column.
  - Auto-scrolls to the latest entry; blinking caret on the trailing
    active line.
  - Lines derive from `device.connectionPhase` transitions and per-
    counter bumps on `MeshClient.progress`. Resets on each fresh
    connect cycle.
- Mounted at the top of the device tree (outside the device-conditional
  branch) so it shows during a first-time connect from the Connections
  screen as well as reconnects from inside the app.
- i18n strings under `connections.overlay.log.*`.
…sibility

Two changes:

Visibility fix — overlay now reads `savedConnections` for any entry
whose status is "connecting" or "configuring". The previous code keyed
off `device.connectionPhase` from the active deviceStore entry, but
that entry doesn't exist until partway through `setupMeshDevice` —
the overlay was hidden during transport-open + storage-open (often
the slowest part of the connect flow).

Visual redo (option C — hero card with radial pings):
- Center: transport-type icon (Globe / Bluetooth / Cable) on a
  gradient disc with two concentric ping rings.
- Below: device name + phase label with small spinner.
- 2x2 grid of stat chips (Identity / Metadata / Channels / Nodes)
  that flip from gray (dashed circle) to emerald (check or count)
  as the SDK reports each piece arriving.
- Slate-to-black gradient panel; non-dismissable while active.
- i18n strings under `connections.overlay.*` updated to match.
…escape

Two reasons the overlay could "never disappear":

1. Persisted state from a prior session.
   `savedConnections` is Zustand-persisted, so a prior tab that died
   mid-connect (page reload, hot reload, force-quit) leaves a saved
   connection at status "connecting" / "configuring" forever. On cold
   boot the overlay reads that and stays visible. `onRehydrateStorage`
   now sweeps any persisted "connecting" / "configuring" /
   "disconnecting" entries back to "disconnected" — no live JS code
   could complete those, so they're stale by definition.

2. Live attempt that genuinely hangs (firmware in CLI / bootloader,
   framing out of sync, `onConfigComplete` never arrives).
   After STUCK_THRESHOLD_MS (15s) the overlay surfaces a Cancel button
   + a hint. Cancel calls `useConnections().disconnect(id)` which tears
   down the transport and flips status off "configuring". Resets if a
   new attempt starts.

i18n adds `connections.overlay.stuckHint` + `cancel`.
…pleteId is buffered

Reproed on serial reconnect: the device kept emitting from a prior
session, so the new MeshClient's `transport.fromDevice → decodePacket`
pump processed the buffered `configCompleteId` synchronously inside
the `buildMeshDevice` await — before `setupMeshDevice` had a chance to
subscribe to `onConfigComplete`. SimpleEventDispatcher doesn't replay
to late subscribers, so the event was lost. The connect overlay (which
keys off the saved-connection status) stayed visible forever even
though the rest of the app saw nodes / messages flowing.

Fix: wire two parallel paths to a shared idempotent `markConfigured`
handler:

- `events.onConfigComplete` (fast path, fires on the next live event).
- `MeshClient.device.status` signal — `decodePacket` also flips the
  signal to `DeviceConfigured` on the same packet, and signal
  subscribers don't drop the current value if it's already there.

Plus an initial synchronous check in case the status signal had
already settled before either subscribe ran. `configSubscriptions`
unsubscribes both on teardown.
…me tokens

Reuses the same `--color-text-primary`, `--color-text-secondary`,
`--color-background-primary`, and `--color-link` tokens the rest of
the web UI already binds to (light + dark variants resolve via
`[data-theme=...]`), so the overlay now picks up the active theme
instead of fighting it with a hardcoded slate/sky/indigo palette.

- Hero ring + disc → `bg-link` + `bg-link/{20,30}` ping ripples.
  Icon uses `text-background-primary` so it stays legible on both
  themes (white-on-blue light, dark-on-blue dark).
- Title / phase / hint copy → `text-text-primary` / `text-text-secondary`.
- Stat chips use `green-500` (matches `Button.success`) for the done
  state and `text-secondary`-derived neutrals for idle.
- Cancel button reverts to the default `Button` outline variant — was
  carrying its own slate overrides, no need.
Reproed in MessagesPage on /messages/broadcast/0: typing a message and
pressing Enter cleared the input, but the message never rendered.

Root cause:
- `ChatClient.send` calls `sendText` → `MeshClient.sendPacket(...,
  echoResponse=true)`. The echo branch dispatches `onMeshPacket` only.
- `ChatClient` subscribes to the per-portnum `onMessagePacket`, not the
  raw `onMeshPacket`. So the outbound message is never appended to its
  conversation bucket.
- The Meshtastic firmware does not loop the user's own outbound text
  back via `fromradio`, so there's no inbound echo to fall back on.

Fix:
- `ChatClient.send` now optimistically appends the outbound message
  (state=Pending) to the matching channel/direct bucket and persists
  it via the configured repository as soon as `sendText` resolves Ok.
- The existing `onRoutingPacket` subscriber flips state Pending→Ack /
  Failed when the routing ack arrives keyed by packet id — unchanged.
- `ChatStore.hasMessage(key, id)` added; the inbound `onMessagePacket`
  subscriber and the `send()` path both consult it so a
  belt-and-suspenders firmware echo doesn't double-append.

Tests: ChatClient.send.test.ts — 4 cases (broadcast append, direct
append, echo dedup, routing-ack flips state). Suite 70 → 74 / web
suite untouched at 236.
…render)

Previously the optimistic append happened after `await sendText(...)`,
which itself awaited `queue.wait(id)` — i.e. the firmware ack. On LoRa
that's 1–3 seconds, so the user saw a noticeable delay between hitting
Enter and the bubble appearing.

Move the append before the await. To do that the chat slice needs to
know the packet id synchronously, so:

- `MeshClient.sendPacket` gains an optional `packetId` parameter
  (defaults to `generatePacketId()`, behaviour unchanged for callers
  that don't pass one).
- `SendTextInput.packetId` plumbs through `SendTextUseCase` to
  `sendPacket`.
- `ChatClient.send` generates the id up-front, appends the message
  with `state=Pending` immediately, persists it, then awaits
  `sendText(..., { packetId })`. On Err the optimistic message stays
  visible but flips to `Failed` so the user sees what happened.
  Routing-ack subscriber still flips `Pending → Ack` keyed by id —
  unchanged.

Tests added (suite 74 → 76):
- "appends the optimistic message before the send promise resolves"
  — asserts the bucket contains the message synchronously after the
  send call returns its pending promise.
- "flips outbound state to Failed when sendText returns Err"
  — exercises the empty-text validation path.
@danditomaso danditomaso marked this pull request as ready for review May 27, 2026 01:26
Copilot AI review requested due to automatic review settings May 27, 2026 01:26
@danditomaso
Copy link
Copy Markdown
Collaborator Author

@copilot resolve the merge conflicts in this pull request

Copy link
Copy Markdown
Contributor

Copilot AI left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copilot wasn't able to review this pull request because it exceeds the maximum number of files (300). Try reducing the number of changed files and requesting a review from Copilot again.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants