Skip to content

armgit / macro

Public Imported from GitHub
0

Macro is a unified workspace for teams: email, chat, docs, tasks, agents, calls, and CRM — @-linked together with shared AI memory.

2.3 KB
Raw Blame

Quick Access scale measurement

Run the existing standalone Vite frontend, then from apps/web:

URL=https://localhost:3009 OUT=/tmp/command-scale \
PLAYWRIGHT_CHROMIUM_EXECUTABLE_PATH=/path/to/google-chrome-stable \
node src/features/command/browser-test/measure.mjs

No login or backend is required. Each size uses a new browser context and a random cache namespace, with 1k/5k/20k synthetic notes written through the real worker host. Closing Chrome releases the contexts. Seed fields are complete for the history projection; missing materialized rows fail the run. Page errors also fail the run. Freeze source and restart the owned frontend between revisions; do not combine samples interrupted by dependency reloads.

At 4× CDP CPU throttle the runner records:

  • Synthetic candidate construction and cache seed time separately.
  • Cold search plus selected-record materialization, then 54 warm searches.
  • 126 local filter + final Cmd+K ranking calls over 18 short prefixes.
  • A separate CPU profile of those prefixes, outside timed samples.
  • Main-thread heap before/after forced GC and outgoing worker/port message kinds.

JSON and CPU profiles go to OUT. Heap deltas include the fixture's candidate objects and main-thread host, not worker/WASM memory or isolated app-index memory. Construction time is not QuickAccessSource activation. Message counts cover web worker transport, not desktop native IPC. JS CPU timings are not input-to-paint. The profile and synthetic candidate stress test intentionally exercise a full local corpus: the cache-enabled app normally ranks a bounded selection in JS.

Compare revisions with identical corpora, throttle, and queries. For causal scorer comparisons, counterbalance revision order and hold wall-clock timestamps fixed; compare complete result ordering outside the timed section. Mixed entities, aliases, comma queries, dynamic labels, people priority and boosts are covered by freshSort.test.ts and rank-command-search-items.test.ts.

The engine's quick-access-v1 policy and JS NAME_MATCH_RANKING are different. These timings do not establish ranking parity or justify replacing final contextual ranking with the engine's existing order. The app already searches the complete cache there, materializes up to 500 hits, then combines local commands and people.