Universal search

Design and implementation plan

Core owns versioned contracts, provider coordination and a compact grouped search dialog. Apps own queries, data and resource views. Providers include Memory, Bookmarks and Todo.

Verification: core and web tsc --noEmit pass. Targeted ESLint passes. Repo-wide bun run lint reports unrelated unused variables in plugins/nr/db/evaluation-queue.ts (NostrEvent) and plugins/translate/service.ts (settings). A one-off check opened both existing plugin databases read-only and validated search results, next pages, encoded-address round trips and resource views through a fresh capability registry. Each provider returned three hits with a next page; both first-hit views passed the shared schema. No indexes or data were changed. No tests were added/run and no browser session was used. The authenticated UI still needs user verification: header entry, typing, multi-source selection, tree expansion/collapse, keyboard navigation, pagination, opening links and missing/deleted-resource feedback. Host restart is required; no restart signal was created in the active chat.

Contracts

Contracts live in src/capabilities/, using the existing capability registry. resource-view:v1.open accepts a resource type/ID and returns either a provider-owned WebNodeRoot or an explicit not-found result. It never mutates data. Providers may implement resource views independently of search.

Addresses are complete CapabilityResourceRef values for resource-view:v1. Provider IDs use registry identity, not configurable aliases. The canonical browser link is #/resource/v1/<encoded-provider>/<encoded-type>/<encoded-id>; each segment is encoded separately. Core routes exactly to the addressed provider, never substitutes another provider, and uses existing capability missing-provider/error UI. Addresses are scoped to an AppWeaver installation and its active workspace; they do not transfer data between installations.

search:v1.search accepts {query, limit, cursor} and returns {results, nextCursor, notices}. Results contain {id, text, context, address}. Text is plain text. Cursors are provider-owned, query-scoped strings; null starts a search or indicates no next page. Providers must use deterministic ordering, bound database work/results, and return only resources belonging to themselves. Core uses registry metadata for source labels/icons and deduplicates result IDs within each source. It preserves provider ordering rather than comparing scores. Providers opt into this capability as an interactive search service; remote providers must make network-query behavior explicit to their users.

Example result (source metadata is attached separately from the registry):

{
  "results": [{
    "id": "53",
    "text": "Release engineering guide",
    "context": "Engineering · https://example.com/releases",
    "address": {
      "capability": { "name": "resource-view", "version": 1 },
      "providerId": "appweaver-bm-plugin/resource-view/v1",
      "resourceType": "bookmark",
      "resourceId": "53"
    }
  }],
  "nextCursor": null,
  "notices": []
}

Providers register SearchV1 and, independently, ResourceViewV1 with defineCapabilityProvider, declare both under appweaver.capabilities.provides, and set coreApiVersion to ^14.11.1 or newer. Core binds source identity at registration. Every result's resource provider must belong to that same owner. The reusable coordinator is in src/core/search/coordinator.ts and takes an injected transport; the web dialog uses the authenticated generic capability WebSocket path. The coordinator does not depend on browser state or plugin code.

Surfaces and behavior

The authenticated web header opens a compact search dialog from the search icon to the left of /bot status. Discovered sources appear below the search box as toggleable labels (all active by default) for multi-source filtering. Queries are debounced; each active provider progresses independently, with at most four simultaneous calls and a per-call deadline. New queries, workspace/session changes, disconnection and dialog cleanup discard obsolete work. Timeouts stop waiting; they cannot preempt synchronous in-process plugin code. Search does not write timeline history. Results render inside a WebNodeShadowRoot tree, with each source as a collapsible treeItem branch, optional collapsible sub-type treeItem branches when a source returns multiple resourceTypes (such as code and markdown), and each result as a leaf treeItem. Result addresses may optionally specify surface: "modal" | "timeline" (defaulting to "modal"). Selecting a result opens the owning app's view on that surface. Encoded resource links can also be opened directly after authentication.

Memory searches existing committed code/Markdown snapshots in the active workspace, never refreshes indexes, and opens matched files in the timeline via workspace-file-view:v1 (falling back to read-only inspection text when no file location is available). Bookmarks searches saved bookmark fields using bound SQL and opens the /bm list widget in the timeline with the matched bookmark treeItem focused (focused: true). Todo searches todo titles, descriptions and tags using bound SQL and opens the /todo list widget in the timeline with the matched todo treeItem focused (focused: true). See Memory, Bookmarks and Todo.