fix(mothership): stop chat perf decay from permanently-animated streamed messages

[TheodoreSpeaks]

Summary

Chat perf decay (the main fix)

• Chat got progressively laggier over a long session because every streamed message stayed in Streamdown's animated pipeline forever — one span per character, tens of thousands of permanent DOM nodes in the mounted transcript. Refresh fixed it because reloaded transcripts render static.
• Settled messages now remount to a span-free tree 300ms after the reveal drains. Remount via `key` is required: streamdown 2.5 memoizes its default element components on className + source position, so flipping `isAnimating` off leaves stale span DOM in place (verified with a standalone repro). The settled instance keeps the streaming parser (`mode`), so the swap trades byte-identical pixels — no flash on replies with unbalanced markdown.
• Mid-stream: `sep: 'char'` → `'word'` (~5x fewer spans rebuilt per reveal tick; visually identical with `stagger: 0`), and fades stop past 6k revealed chars in one segment so walls of text can't swamp the frame budget.
• Unified the plain-text and special-tag render branches: most replies gain a trailing `` tag (suggested follow-ups) at the very end, and the old branch switch re-parented Streamdown at that moment — remounting it and re-fading the whole message from transparent (the long-standing "flash at the conclusion", predates this PR).
• Animation latches reset when a reused ChatContent instance receives replaced (non-append) content, so regenerated turns still animate; latches held in `useState` per the updated render-phase hook rules.
• Measured mid-stream on a single giant paragraph: p95 frame time 49.9ms → 9.4ms, long tasks 1256ms → 0 per 12s window; at ~10k chars: p95 83ms → 10ms. Spans drain to 0 after every turn; swap is pixel-stable (0 height/text delta).

Text-selection follow-up (regression from #5389 that this branch merged in)

• The global solid-brand + forced-white `::selection` rendered black-on-blue in the chat input and every workflow subblock editor — those draw text in a mirror overlay the forced white can never reach (and Chromium ignores `::selection` color in form controls anyway).
• Kept one global rule, made it lighter across the board per Waleed's suggestion: `::selection { background-color: var(--selection-muted) }` (`#1a5cf647` light / `#4b83f759` dark), no text-color override so selected text keeps its own paint on every surface. Landing keeps a consistent, non-dimming branded selection; verified on landing, chat prose, and the overlay inputs in both themes.

Type of Change

• Bug fix

Testing

Tested manually in the live app (streamed multi-turn chats; measured rAF frame deltas + long tasks before/after; DOM-diffed the settle swap; verified 0 animated spans after settle, no completion flash, no re-fade when follow-ups arrive; selection checked on landing/chat/inputs in light + dark). `vitest` message-content + smooth-text + file-viewer suites pass, `bun run lint`, `bun run check:api-validation:strict` pass.

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
docs	Skipped		Jul 4, 2026 7:43pm

[cursor]

PR Summary

Medium Risk
Chat rendering lifecycle is substantially more complex (remount timing, latch resets, unified tree); regressions could show as completion flashes, missing fades, or wrong static/streaming state on regeneration.

Overview
Fixes progressive chat lag in long sessions by tearing down Streamdown’s per-token animation DOM after each reply finishes, and by cheaper mid-stream fades. Settled messages remount Streamdown (~300ms after reveal) with a stream → settled key so animation spans are dropped while the streaming markdown parser stays latched (avoids re-parse flash). Mid-stream, reveal animation uses word instead of char separation, and fades stop after ~6k revealed chars in one segment.

Also unifies plain-text and special-tag rendering so a trailing <options> block no longer re-parents markdown and re-fades the whole message. Latches reset when content is replaced (not appended) on a reused row instance.

Global CSS: ::selection uses new translucent --selection-muted instead of solid brand + forced white, so selection stays readable in mirrored chat/workflow inputs and form controls.

^{Reviewed by Cursor Bugbot for commit 5edddda. Bugbot is set up for automated code reviews on this repo. Configure here.}

[greptile-apps]

Greptile Summary

This PR fixes a progressive performance degradation in long chat sessions caused by Streamdown's animated pipeline accumulating thousands of permanent <span data-sd-animate> DOM nodes (one per character) for every streamed message. A page refresh fixed it because reloaded history renders in static mode, while live-streamed messages stayed in the animated pipeline forever.

• Animated→static lifecycle: Adds animationDrained state that flips 300ms after the reveal completes, then remounts Streamdown via key ('stream' → 'settled') to shed the per-character spans — intentionally using the same streaming parser (parserTree stays latched) to avoid a flash from re-parsing unbalanced markdown through a different pipeline.
• Mid-stream span reduction: Switches sep: 'char' to sep: 'word' (≈5× fewer spans rebuilt per tick), and introduces a 6 000-char fadeCutoff latch that stops new fade-in spans from being created on very long paragraphs while leaving the streaming parser running — measured p95 frame time dropping from 49.9 ms to 9.4 ms on a single large paragraph.
• Unified render path: Collapses the previous two-branch return (with-/without-special-tags) into one, preventing Streamdown from re-mounting (and re-fading the whole message) when a trailing <options> block appears at completion.

Confidence Score: 5/5

Safe to merge — well-scoped rendering optimisation with careful state transitions and no regressions on the existing test suite.

All three state latches (streamedThisSession, animationDrained, fadeCutoff) are one-way and correctly gated: the drain timer fires only after reveal settles, the key swap remounts Streamdown without touching the streaming parser, and the content-replacement reset is guarded by animationDrained so it cannot fire mid-stream. The unified render path removes the structural branch that caused re-mount flashes. No logic paths loop or leave state inconsistent.

No files require special attention.

Important Files Changed

Filename	Overview
apps/sim/app/workspace/[workspaceId]/home/components/message-content/components/chat-content/chat-content.tsx	Core streaming lifecycle refactor — adds three one-way state latches (streamedThisSession, animationDrained, fadeCutoff), a drain timer, and a sep+cutoff mid-stream optimisation; all transitions are carefully gated and the comments match the code. Unified render path eliminates the two-branch return that caused re-mount flashes.
apps/sim/app/_styles/globals.css	Adds --selection-muted CSS variable (translucent brand blue) and switches ::selection to use it without a forced color override; fixes broken selection rendering in form controls and transparent-textarea chat inputs. The edit touches a globally-scoped pseudo-element that cannot live in a component stylesheet.

_{Reviews (2): Last reviewed commit: "improvement(styling): one lighter transl..." | Re-trigger Greptile}

[cursor]

Stale fade cutoff after replace

Medium Severity

When displayContent is replaced while animationDrained is still false, latch reset skips fadeCutoffRef even though the instance can be reused for a new logical message. If the prior message had crossed FADE_MAX_REVEALED_CHARS, the new stream keeps fadeActive false for the rest of that mount, so reveal fades never run for the replacement content.

 

^{Reviewed by Cursor Bugbot for commit 3961a9c. Configure here.}

[cursor]

Whitespace-only text not rendered

Low Severity

Plain replies without special tags now always go through flushMarkdown, which only enqueues inline groups when pendingMarkdown.trim() is truthy. Whitespace-only text still becomes a text segment via parseSpecialTags, but it never produces an inline group, so Streamdown is omitted and the message body renders empty.

 

^{Reviewed by Cursor Bugbot for commit 3961a9c. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

There are 3 total unresolved issues (including 2 from previous reviews).

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 56f391b. Configure here.}

[TheodoreSpeaks]

@greptile review