[marius-kilocode]

What Problem This Solves

Streamed assistant text in the VS Code sidebar stepped visibly instead of tracking the stream. The webview throttled text and reasoning parts to a fixed 100 ms (TEXT_RENDER_THROTTLE_MS), so live text could repaint at most ~10 times per second, while the host scheduler already delivered deltas on a 16 ms focused lane. The TUI renders at roughly frame cadence, so VS Code felt slower for the same model output.

Why This Change Was Made

createThrottledValue now accepts a reactive interval. Text and reasoning parts pass ~one frame (16 ms) while the part is streaming and fall back to the existing 100 ms once it settles. Completed history keeps the slow cadence, so scrolling old messages stays cheap. The default interval is unchanged, so other callers such as shell-rolling-results.tsx are unaffected.

This also stays safe with many concurrent agents. The host already paces each lane: focused 16 ms, visible 50 ms, background 150-400 ms with per-agent backoff. Background sessions stream slower than 100 ms, so their render cadence is unchanged. The webview now matches the focused lane instead of overriding it with 100 ms.

User Impact

Streamed text and reasoning appear smoothly as tokens arrive instead of in 100 ms steps. Total response time is unchanged, since it is model-bound. Background Agent Manager sessions are unaffected.

Evidence

Measured in an isolated VS Code instance with a real model and the compiled webview bundle verified to contain the change. Same prompt and viewport, 3 runs each, sampling rendered text per animation frame.

Inter-update render gap:

Metric	Before	After	Change
Median	100 ms	41 ms	-59%
p25	100 ms	25 ms	-75%
Minimum	94 ms	15 ms	-84%
Visible updates/sec	8.5	15.0	+77%

Before, every gap was >=86 ms (0% under 75 ms). After, about 40% of gaps are <=33 ms. The p90 gap stays near 100 ms because token arrival now sets the ceiling, not the render throttle.

No CPU regression in Chromium metric deltas over the stream: ScriptDuration 0.73-1.00 s before vs 0.87-0.93 s after; RecalcStyleDuration and LayoutDuration within run-to-run noise. Zero webview console errors; final answer rendered correctly with headings, lists, and code blocks.

Validation

• bun test --conditions=browser ./src/components/ from packages/kilo-ui: 72 pass, 0 fail (includes 3 new cadence tests).
• Manual: streamed a normal reply in the isolated VS Code sidebar, confirmed smooth text, correct markdown, and no console errors.

Limitations

• p90 still tracks token arrival, so bursty providers can still gap.
• Measured for one focused stream, not for many simultaneously visible Agent Manager sessions; those use the host's 50 ms visible lane.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (2 files)

• packages/kilo-ui/src/components/tool-utils.test.ts
• packages/kilo-ui/src/components/tool-utils.ts

Previous Review Summary (commit 8d5cae6)

Current summary above is authoritative. Previous snapshots are kept for context only.

Previous review (commit 8d5cae6)

Status: 1 Issue Found | Recommendation: Address before merge

Overview

Severity	Count
CRITICAL	0
WARNING	0
SUGGESTION	1

Issue Details (click to expand)

SUGGESTION

File	Line	Issue
packages/kilo-ui/src/components/tool-utils.ts	26	Switching to the idle interval can delay the last streamed tokens

Files Reviewed (4 files)

• .changeset/streaming-text-frame-cadence.md - 0 issues
• packages/kilo-ui/src/components/message-part.tsx - 0 issues
• packages/kilo-ui/src/components/tool-utils.test.ts - 0 issues
• packages/kilo-ui/src/components/tool-utils.ts - 1 issue

Fix these issues in Kilo Cloud

---

_{Reviewed by grok-4.6 · Input: 111.4K · Output: 9.6K · Cached: 70.1K}

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