perf(home): a streamed talk token wakes no whole-message Home view

Merged
LA
2d ago
+207-23

Summary

HOT PATH AUDIT area F, row H1 (AUDIT.md). Home shows each Chief reply whole by default, so a streamed talk token changes nothing on screen. Yet each token:

  • rebuilt and compared the live run of every conversation (HomeLiveBook.shown), then assigned the whole live dictionary;
  • woke every open Home view's store observation, and each view's inputs (they carried the reply text) differed, so each ran an O(transcript) projection pass on the main thread.

The fix:

  • publishLive updates only the token's conversation (live[c], streams[c], new runShapes[c]).
  • HomeRunShape: what a whole-message view draws of the running replies (who writes, the activity line, the thought headline), never the text. Whole-message views observe and compare it, so a talk token does not wake them. Typing dots, the activity line and the headline still wake them.
  • Streaming mode (HomeReplyDisplay.streaming) reads the text as before.

New perf-hot path id home-stream-token (scripts/cmux-next/perf-hot-benches.json).

Testing

Red test first: HomeWholeMessageTokenCostTests (MessagesLab). It streams 1,500 tokens of the Chief's live run over a 2,000-message history with 1 and 4 views open. It expects 0 store-observation wakes and 0 projection passes, and prints HOMETOKEN lines (ms per token on the main thread).

Observed / Not run

  • Observed: compile-check 21869a1f80d8 (red head 7e48dc27c46) and 5d2989681d5e (fix head f3fc5899d39) passed.
  • Not run yet: the red step at 572c1cde03f and the green step at f3fc5899d39 (package-test-lane.sh suite Packages/Shared/CmuxMessagesLab HomeWholeMessageTokenCostTests). The branch waits for a Mac lane line (STATUS 9221). The existing Home streaming suites (HomeStreamingPerfTests, HomeStreamStressTests, HomeMidStreamSendTests, DragWhileStreamingTests) are not run yet either.

Changelog

  • Changed: a Chief reply that streams no longer slows typing and scrolling in Home.

No capture: no visible change; only main-thread work per token goes down.

Autofix with [code]smith
Open in [code]smith · Need help on this PR? Tag @codesmith-bot with what you need.

Summary by cubic

Stops a streamed talk token from slowing typing and scrolling in Home. Home shows each Chief reply whole by default, so a token changes nothing on screen — yet before, every token rebuilt and compared the live state of every conversation, woke each open view's store observation, and ran an O(transcript) projection pass on the main thread.

Changes

  • publishLive now updates only the token's conversation (live[c], streams[c], new runShapes[c]).
  • Adds HomeRunShape (who writes, the activity line, the thought headline — never the text); whole-message views observe it, so a talk token leaves them asleep while typing dots, the activity line and the headline still wake them. Streaming mode reads the text as before.
  • Empty shapes are stored as no entry, so thought tokens before a run has any reply text wake no view either.
  • Adds perf-hot bench home-stream-token and HomeWholeMessageTokenCostTests (1,500 tokens, 2,000-message history, 1 and 4 views) expecting 0 wakes and 0 projection passes, plus a test for thought-before-reply tokens.