[loganj]

🤖

Summary

• In a very large channel (9,000 members), typing @ opened a mention menu that froze the app for about 117 ms when the member list finished loading. Most of that time was spent ranking names the menu never shows.
• Now the menu does only the ranking work it needs. In sampled browser runs, the longest frame dropped to 50 ms, and no sampled frame was longer than 50 ms. These are measurements of a few runs, not a guaranteed maximum. The order of the suggestions does not change.

Sampled browser measurement profile-roster-scale (local Chromium, 3 runs each, on main 15febc1)	main	this PR
longest sampled frame	116.7 / 116.6 / 116.6 ms	50 / 50 / 50 ms
sampled frames over 50 ms	1 / 1 / 2	0 / 0 / 0
time past one 60 Hz frame	118 / 134 / 134 ms	51 / 68 / 51 ms

Frame times vary with machine load. An earlier version of this PR, on main fe6eaa2, measured 150–167 ms on main; #2328 has since removed the truncated flag, which was part of that cost.

Details

• The shown rows came from a full sort of every match, then the first 50 were kept. rankMentions now takes a limit and keeps the first rows of the same order in one pass (a bounded insertion, about one comparison per name). The menu always passes MENTION_ROWS (50), the most rows it shows.
• The Space check (does Space complete an exact name?) folded every name for accents and case to find names that continue the query. No name continues an empty query, so continuingNames now returns early for one. exactName already returned nothing for an empty query.
• Outside suggestions (people and agents not in the channel) are now ranked only when a row can still be added, not on every render while 50 rows are already shown.
• One shown-row limit: the chooser's 50-row limit is now one constant, MENTION_ROWS, shared by ranking, appending and the inline completion budget. A hook test checks that 51 members show exactly 50 rows; it fails with the ranking limit removed.

The ranking test checks that a limited ranking equals the first rows of the full ranking for 240 mixed choices, 6 queries and several limits, and that an empty query has no continuing or exact name.

Not in this PR

The remaining ~50 ms sampled frame is not ranking. It is name labels (~40 ms), building the candidate list (~30 ms) and signature checks for the last profile page (~25 ms), from a profiled build on the earlier base. A non-empty query also still rebuilds the name index whenever the candidate list changes.

🤖 Generated with Claude Code

[kalvinnchau]

🤖 One regression needs correction before merge: this patch reverses the FI backup-discovery fix from #2344.

[wesbillman]

🤖 Star Lord’s automated source review via Wes’s account, with independent Rocket, Mantis and Groot coverage.

No actionable findings in the five-file change. Source inspection supports preservation of ranking keys and stable ties, append-only chooser rows, full-candidate Space checks and the 50-row truncation threshold. The public description and feature commit text were inspected; the description has no attached images.

Reviewed head: c5319a4. Target base: cbc71c8. PR merge base: 67f6229.

Source-only: no tests, builds, app launches or PR code execution. A read-only hosted-check snapshot reports CI required successful; Windows native validation and browser flake checks were skipped. Reported browser timing improvements were not independently reproduced. Native/human acceptance, rendered accessibility and runtime focus/retry behavior remain unverified; this comment is not approval or merge authorization.

[wesbillman]

🤖 Star Lord’s automated source review via Wes’s account, with independent Rocket, Mantis and Groot coverage.

No new actionable findings in this rebased five-file change. The ranking limit preserves the full sort’s prefix and append-only chooser order; Space resolution still checks the uncapped candidate set. The updated hook and its inline/picker consumers retain live selection checks and error/retry handling. Removing the obsolete truncation helper is consistent with the updated base. The public PR description and feature commit messages were also inspected; there are no attached images.

Reviewed head 2af321a8332f18ab1ce863549fc3767016b13e8c against base 49bdc82b9e29c876e0b709b0c3de263b75e5db3b, with comparison to previously reviewed c5319a4aed541c3d550b255a7dcf582cc9d2c8c1. This follow-up covers the feature’s changed integration, not an independent review of all intervening main-branch changes.

Source-only: no tests, builds, app launches or profiling were performed. Rendered focus/accessibility, human acceptance and the reported frame timings remain unverified; the timing samples identify a different baseline (15febc12e). A one-time hosted-check snapshot still had browser, JavaScript, Rust and measurement jobs running; Windows validation was skipped. This COMMENT is not approval or merge authorization.