Add global search across Markdown Files (#117)

[zcuric]

Closes #117.

Adds an MVP global search palette to Hubble Desktop: press Cmd+P (or File → Go to File…) to find a Markdown File by name, by path, or by a phrase inside it, then Enter to open it.

The issue asked for the indexing decision to be made as part of the work, so this PR contains both the spec that records it and the implementation that follows it.

The decision: naive on-demand search, no index

Recorded in specs/gh-117/TECH.md. Three reasons, in descending order of how costly they'd be to get wrong:

• A desktop index would fight ADR-0008. An index needs invalidation, and the only cheap invalidation signal is a recursive watcher — precisely what ADR-0008 rules out. Instead, content search reads the candidate paths the sidebar snapshot already holds in the renderer, so main never re-walks the folder. Search sees exactly what the sidebar sees, goes stale the same way, and refreshes on the same focus trigger. No second view of the folder, no cache to invalidate, no watcher.
• Web wouldn't benefit from one. The Convex files table already stores content inline and getFilesByWorkspace returns full rows — which is why loadPath in apps/www does a .find() rather than a fetch. Every note body is already in client memory. A .searchIndex() would build a server-side index to search data the client already has.
• Pagefind is the wrong shape. It's a build-time indexer over built HTML. Hubble has no build step over user content, and that content changes on every keystroke. Worth stealing its ranking and excerpt presentation, not its architecture.

What changed

• packages/ui/src/lib/fuzzy.ts — path-aware fuzzy scoring (exact > prefix > substring > subsequence, matching the existing command menus) plus matchRanges so the palette can emphasize where a query matched. A basename hit outranks a parent-folder-only hit.
• packages/ui/src/components/GlobalSearchPalette.tsx — shared cmdk palette. Name matching is synchronous; content search is injected as a searchContents prop, debounced 150 ms and gated at 3 characters. A file matching both ways appears once, under Files.
• apps/desktop/src/lib/searchContent.ts — literal, case-insensitive line matcher with windowed excerpts. Deliberately not a regex: user input would need escaping anyway, and a regex dialect is a promise we don't want to make for an MVP search box.
• apps/desktop/electron/main.ts — desktop:search-file-contents IPC. Concurrency-pooled reads, assertGranted on every path, 2 MiB per-file cap, 50-file result cap surfaced in the UI rather than silently truncated. A superseded search is abandoned by comparing request ids between files, since an AbortSignal can't cross IPC.
• Desktop wiring: CmdOrCtrl+P in the existing keymatch chain plus a File menu item, both open-only and idempotent (this is what CmdOrCtrl+Shift+O already does). Selecting a result goes through the existing loadPath chokepoint, so search results participate in document history for free.

Performance

specs/gh-117/TECH.md originally carried an estimate; it now carries a measurement. Over 2,000 markdown files totalling 16 MiB, concurrency 8, warm cache:

query	matching files	median
no match (full scan, worst case)	0	44 ms
rare needle	20	49 ms
common needle	capped at 50	1 ms

The worst case is a full scan at ~45 ms, comfortably inside the 150 ms debounce. A common needle is faster, not slower, because the result cap short-circuits the walk almost immediately. If a much larger corpus ever misses this budget, the escalation is a stat-validated content cache in main — strictly cheaper than an index, and still no watcher.

Two base-ui bugs found while testing

Both share one root cause — Dialog.Popup stays mounted after close — and neither was caught by tests. Both were found by driving the running app over CDP.

1. Ghost clicks. At opacity: 0 the popup still hit-tests. After opening the palette even once, a real click in the middle of the editor landed on an invisible result row and opened whichever file sat under the cursor (reproduced: a click meant for the text navigated to Project Ideas). Fixed with data-[closed]:pointer-events-none on the popup and the backdrop.
2. Dead second open. autoFocus fires only on mount, so the second Cmd+P left the caret in the editor: the palette rendered but swallowed every keystroke and Enter selected nothing. Fixed with base-ui's initialFocus={inputRef} on Dialog.Popup, which runs on every open. Note that a manual focus() in an effect is not sufficient — it races base-ui's own focus management and lands non-deterministically (it passed once, then failed on the next run).

packages/ui/src/primitives/modal.tsx has both latent behaviors — I confirmed the Settings modal also stays mounted and captures clicks. It's harmless there today because it's a small centered dialog, but it's the same trap waiting for the next full-width overlay. Left for a separate change and recorded in the spec's risks.

Testing

• Unit (22 new): fuzzy.test.ts covers the score ladder, separator normalization, basename preference, and matchRanges offsets including the subsequence case. searchContent.test.ts covers line numbers, excerpt windowing at line start/end/mid-line, offset correctness after trimming indentation, CRLF, the per-file match cap, and blank queries.
• Full suites: 64 UI + 73 desktop tests pass. pnpm check and pnpm build:desktop are green.
• End-to-end in the running app (via CDP against the dev build): opened with Cmd+P; mtgnts fuzzy-matched meeting-notes-2026-07-02.md with the subsequence characters emphasized; pagefind produced a name hit plus a content hit in indexing-strategies.md at line 3, correctly deduped; always-on recursive produced a content-only hit; Enter navigated in every case. Three consecutive open → type → Enter cycles verified after the focus fix. Ghost-click regression verified: with the palette closed, a real click at the editor centre hits the editor, not a result row. Checked in both light and dark themes.

Not in this slice

• Web. apps/www is untouched. Its content is already client-side, so it's a small addition behind the same injected searchContents prop.
• Jump to the matching line. Selecting a content result opens the file at its top. Mapping a source offset to a ProseMirror position is a real problem and would roughly double this change; it becomes near-free once source mode (Add source mode (raw markdown editing) to the editor #142) lands.

Reviewer notes

The two base-ui bugs above have no regression test — they're behavioral and only reproduce in a real Electron renderer. A future refactor of the palette can silently reintroduce them. Happy to add a harness if you'd like one, though it needs a running app rather than jsdom.

How it works in the app

Screenshare.-.2026-07-09.9_02_20.PM.mp4

[vercel]

@zcuric is attempting to deploy a commit to the bholmesdev's projects Team on Vercel.

A member of the Team first needs to authorize it.

[bholmesdev]

This is a seriously excellent PR. Thank you for making it! I also appreciate the breakdown of performance in the fuzzy search algorithm in the product and tech specs. I think this implementation is good to merge.

I just made some tweaks to the UI to make it feel a bit more friendly, like using sans instead of monospace font for content, stripping out markdown characters, and removing some of the extra hints on how to interact with the search bar; I decided to remove that text and just show the keyboard legend. This is in line with how Notion's command palette works.

[zcuric]

@bholmesdev thanks! I really appreciate it. I used Fable to shape it up. Hope I can contribute mode of these kinds of PRs.