perf(web): cut re-render storms and continuous repaints in bot chats

[leoisadev1]

Stacked on perf/ws-payloads.

Bot chats re-parsed their entire transcript's markdown on every render because the runtime hooks memoized the linked thread ref on a freshly allocated target object, so downstream memoization never bailed out. The landing subscribes to the environment snapshot, meaning this fired on every streaming batch from any bot. Spoken-text derivation re-ran per message per render, message rows were unmemoized inline maps, working avatars ran a requestAnimationFrame loop at display refresh mutating SVG mask attributes, the status shimmer repainted text every vsync, the roster sidebar re-rendered all rows per shell event.

Thread refs are now memoized on their primitive ids. Message rows moved into memoized components with value-aware comparison and content-visibility on rows, chosen over a virtualized list because the conversation scroll area's follow-to-end logic would fight one. Spoken text is cached per reply and invalidated when text changes. The working avatar pose ticks near 30fps and skips unchanged DOM writes, while the shimmer and ultrathink animations use stepped keyframes with reduced-motion guards. New derived atoms map each bot to its latest thread with stable identity, so roster rows subscribe narrowly. A debounced draft store was tried here and dropped in review: every variant opened a cross-tab race, so draft writes stay synchronous. The spoken-text cache instead evicts its oldest entries rather than clearing wholesale, so long chats stay cached, and review also hardened the playback baseline and the avatar eye cache.

Verified with 305 roster and 169 lib tests passing, new tests for the spoken-text cache and its eviction, row memoization, and derived-atom identity, and a clean typecheck. An integrated pass on a fixture environment confirmed the chat renders and actions work. Not visually verified: streaming-state animations, which need a live provider turn.

Created with Claude Fable 5.1 in Claude Code.

🤖 Generated with Claude Code

---

[vercel]

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

Project	Deployment	Actions	Updated
akeru-bot-landing	Ready	Preview	Sep 26, 2026 5:54pm UTC

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Refactors chat message rendering and animation performance.

No outstanding findings block merging.

Summary

The PR reduces rendering and animation work in web bot chats. Drafts continue to save synchronously. No new findings remain.

_{Reviews (7) · Last reviewed commit: "fix(web): restore synchronous draft stor..."}

[greptile-apps]

Comments Outside Diff

These findings sit on lines the diff does not cover, so they could not be posted inline. Each one leaves this list once its file changes.

•  Preserve drafts across tabs apps/web/src/components/roster/botDraftStore.ts:44 ▶

  When two tabs save drafts for different bots at nearly the same time, both can read the shared storage entry before either writes. Each then replaces that entry with its own update, erasing the other bot’s draft. The person loses that draft when they return to it or reload.

[devin-ai-integration]

Devin Review found 4 new potential issues.

🐛 3 issues in files not directly in the diff

🐛 Existing per-bot drafts disappear

When a draft exists under a v2 per-bot key, readBotDraft returns an empty string instead. The composer cannot restore that unsent message after reload.

---

🐛 Concurrent tabs lose unrelated bot drafts

When tabs edit different bots, writeBotDraft can overwrite one tab's draft with the other's stale snapshot. Both edits replace the same storage key, so one draft disappears.

---

⚠️ Every keystroke rewrites all bot drafts

Each composer keystroke calls writeBotDraft, which synchronously parses and rewrites every saved draft. Typing slows as unrelated drafts accumulate.