[loganj]

🤖

Summary

• Typing in a composer in a large channel (#general has about 2,500 members) lagged on every keystroke, in both the channel composer and thread replies. On each key, the app rebuilt the full list of people you could @mention, even when you were not mentioning anyone.
• Now ordinary typing does no mention work. Typing an @ query looks up matching names in a sorted index instead of scanning every member.
• Mention behavior is the same: which people appear, their order, and what Space selects do not change.

Measurements

Local headless Chromium, dev build, against real #general data. Typing 43 characters into the #general composer:

	Before	After
Time per character	29–34 ms	21 ms
CPU in mention code	~330–385 ms	~6 ms

A small channel takes 22 ms per character, so channel size no longer adds cost. Typing @logan in #general dropped from 118 ms to 62 ms per character; mention CPU went from 260 ms to 69 ms, most of it the one-time index build on the first @.

Details

• Closed chooser does nothing. The saved-teams part of the mention chooser built every member's candidate row on each render to check team eligibility, then threw it away while the chooser was closed. It now builds eligibility only for team rows an open chooser shows. The chooser button also no longer joins every member key into a string while closed.
• Candidates are cached by their sources. mentionCandidates caches its rows per session and channel. The cache key is the immutable snapshots it reads: the channel row, profiles, agent choices, archives, roster and directory page. Each directory array has its own entry, so the chooser and its team rows do not evict each other, and only the latest few channels and directory pages keep rows. The rows rebuild only when one of those changes, not on each keystroke. Labels are cached by the same rows plus the draft's selected recipients, compared by key and name, because the draft recreates that list on every edit. Every empty directory page now shares one array. The build itself is now a single keyed pass. Before, it did a find, filter or some scan over agents and directory people for each member.
• Query matching uses an index. The new matchingMentions keeps a sorted list of each choice's name starts: the folded name and every later word start, as the matching rules in src/bundled/mentions/README.md §3 define. A binary search returns only the choices that can match, at a cost of O(log n + matches). Ranking and Space selection then apply the unchanged rules to that subset. A test runs every portable ranking and Space fixture through the index and checks that the results equal the full-list results.
• Composer chips. The rich composer computed chip names over every channel member on each keystroke. It now recomputes them only when the draft's recipients, the members, or naming change.
• Directory pages change only what they add. Most directory hits in a large channel are members whose profile already names them. They change no row, alias, key or label, so directoryChanges leaves them out, and the rest are compared by content, not by array. A page of members rebuilds nothing. Repeated pages and the chooser's merged page lists with the same outsiders reuse one build. A test checks that the rows are equal with and without the people it leaves out.
• New regression tests: a closed chooser reads no member profile while its composer re-renders (the old code read 600). An inline @ query that remounts on each keystroke, with a saved team shown, a new recipient array and settled empty directory searches, rebuilds no name starts (the old code rebuilt 600). Pages of members rebuild nothing, and a page with an outsider rebuilds once while it stays the same (the old code rebuilt 600).

Not changed:

• A multi-word inline @ query still checks all names once per keystroke to decide whether it is a search or prose.
• A directory page that brings in people outside the channel (or renames an agent that is not a member) still rebuilds candidates, labels and the index over all members, once per distinct set of such people. Labels must be resolved over the whole choice set: the naming policy is replaceable and its contract compares every name in the set. Directory names can change member aliases and the namesake qualifiers of member labels, so splitting local rows from directory rows needs its own change.

🤖 Generated with Claude Code