perf(ui): defer diff parsing until tool cards expand

[marius-kilocode]

What Problem This Solves

Expanding an edit, write, or apply_patch tool card is the most expensive tool-card render in the transcript: each card parses its patch with Pierre and renders a shadow-DOM diff. These cards also did work while collapsed, so a transcript with many of them paid parse cost before the user opened anything.

Why This Change Was Made

Two causes:

• edit and write built their diff view with createMemo. Solid evaluates a memo eagerly on render, and the always-rendered trigger read it to decide whether to show "Open in Diff Viewer". The patch was therefore parsed even when the card stayed collapsed. The apply_patch trigger parsed every file the same way. These are now cheap presence and @@ checks, and the view is only parsed when the body mounts or the viewer opens.
• The deferred mount scheduler mounted one tool body per animation frame. An expanded transcript with many diff cards needed one frame per card before everything was visible. Bodies now mount within a per-frame time budget, so cheap bodies mount together and an expensive body ends that frame. The frame is re-armed in finally so a throwing body cannot stall later mounts.

User Impact

• Collapsed edit, write, and apply_patch cards no longer parse patches, so a transcript with many collapsed diff cards does less work.
• Expanded diff cards fill in faster because multiple bodies mount per frame instead of one.
• Behavior is unchanged: expanded cards render the same diff, and "Open in Diff Viewer" still works for edit, write, and apply_patch.

Evidence

Storybook harness (20-card transcript, 420x720, Chromium, 7 loads per arm, paired before/after):

Tool	Expanded settleMs before	after	Change
edit	39.07	14.21	-63.6%
write	24.02	12.125	-49.5%
apply_patch	53.815	26.855	-50.1%

Collapsed Pierre parses in the same harness: edit 1 to 0, write 1 to 0, apply_patch 2 to 0.

Real VS Code (isolated Extension Development Host, Agent Manager, 20-card sessions, instrumented normalize):

• Collapsed session switching: 20 parses to 0 parses, with 0 diffs-container in both arms.
• CPU profile over three switches: ScriptDuration 0.1556s to 0.1436s (-7.7%), transient Nodes 1052 to 314 (-70%).
• Expanded edit, write, and apply_patch render identically, and "Open in Diff Viewer" opens the preview.

Checks: bun run typecheck and bun run lint in packages/kilo-vscode; the opencode annotation checker passes.

Limitations

• The VS Code session-switch wall-clock comparison is floor-bound: the message list virtualizes and mounts about five cards, so timing was not a reliable signal. The Storybook harness reproduces the expanded-card render cost.
• The harness reuses one patch per tool, which understates per-card parse cost.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (3 files)

• .changeset/defer-diff-card-render.md
• packages/kilo-ui/src/components/message-part.tsx
• packages/ui/src/components/basic-tool.tsx

---

_{Reviewed by grok-4.6 · Input: 192.5K · Output: 14.5K · Cached: 383.4K}

_{Review guidance: REVIEW.md from base branch main}