PR #3758 - ReviewStage/luke

Stacked on #3726. This is the larger part of the Neon Realtime lag Charles reported: the task's window sat at about 1.1 GB and used 10–30% CPU while idle.

Cause

While an agent is writing, its tab receives a transcript page up to twice a second. Each page touches only the reply still being written, but every page:

re-rendered every message in the transcript and re-parsed every message's parts;
re-rendered the task's page (through onHeld), and that re-parsed every section of the plan with react-markdown, even while the plan tab was hidden.

Change

Transcript messages are memoized (SteadyAgentMessage and sameMessageDraw in agent-tab.tsx). A message re-renders only when the message object, its position, or its doors change. applyMessagesPage already keeps unchanged messages as the same objects.
- The window rebuilds cards and chat on every render, so they no longer travel inside each message's props. They go through a LiveDoors context that only the rows that use them read: delegation rows, target rows and chat notices. A memoized message therefore still shows the current cards.
- The onOpen and onOpenTarget handlers keep the same identity across renders (useSteady).
Context values stay stable while their content is unchanged: the message place, the Mac owner, the mentioned servers, and the steering and undelivered lines. Whether the tab offers Fork now has its own boolean context, so a new ForkDoors object doesn't re-render every message.
agentParts is cached per message object in a WeakMap. Messages are never changed in place.
MarkdownMessage is memoized. The plan keeps one stable task toggler per section, which reads the plan as it is at the moment of the press. A section whose text didn't change isn't parsed again.

Measured

Benchmark in the test DOM (jsdom, React dev build): 300-message transcript, 20 pages each updating the newest reply.

Per page: about 265 ms before, about 77 ms after. With 10 messages it is about 31 ms, so most of what is left is fixed cost.
Re-rendering the agent tab from its parent with new cards, presses and fork doors now re-renders no messages.

The benchmark isn't committed. I haven't measured on Charles's Mac yet.

Tests

New agent-tab-kept.test.ts, driven through AgentTab. A message that no page changed shows the current card after the window re-renders with a renamed agent. Its worker's row also reflects the worker's status from the newest page. The test fails if rows read cards from the memoized props instead of LiveDoors.
plan-document-editor.test.ts: ticking a box in a section whose text didn't change, after another section changed, saves against the current plan. The test fails if the toggler holds on to the first plan.
vitest run apps/desktop/src/renderer/tasks apps/desktop/src/renderer/desktop: 1595 tests pass. Lint, typecheck, knip and discarded-effect also pass.

Not done

Passing the held transcript to the task page on every page (onHeld). The task page still re-renders, but what that re-render triggered (the messages and the plan) is now skipped. Moving the held transcripts into a store would touch about ten files, and I'd rather measure on the real task first.
Memory. Every transcript opened stays in memory while the task is open. The transcript is also not virtualized.

Made with Luke