Stream Pacer & Reveal Presets

In generative AI interfaces, token arrival rates are non-uniform. Piping raw network packets directly into the DOM triggers visually jarring spasms.

Token Arrival Jitter in LLM Streaming

Server-side batching and network hops introduce severe burstiness:

• Long initial delays waiting for the First-Token (TTFT).
• Sudden TCP bursts dumping dozens of characters instantaneously.
• Unpredictable multi-second pauses mid-sentence.

Without smoothing buffers, the UI oscillates between frozen pauses and overwhelming text dumps, fatiguing the reader's eyes.

The useStreamPacer Clock State Machine

TuffEx decouples raw network streams from DOM rendering via an internal clock pacer:

useStreamPacer({
  wordMs: () => 24,    // Baseline release cadence: 24ms per word unit
  maxLagMs: () => 600, // Backlog deadline: units queued > 600ms accelerate
  drainMs: () => 320,  // Cleanly drains residual tokens within 320ms on close
  pauseMs: () => 400,  // Declares paused state after 400ms idle; caret breathes
})

• Steady Stream: Tokens emerge at a rhythmic 24ms cadence.
• Catch-up Burst: If network bursts threaten the 600ms latency ceiling, the pacer raises release rates dynamically to match real-time model progress.

Reveal Style Presets (Aurora and Blur)

As words emerge, delicate compositor presets enrich the presentation:

• aurora (460ms): Refractive micro-bloom transitioning from soft chromatic shift to solid ink.
• blur (420ms): Recreates physical ink dispersion, transitioning from filter: blur(2px) to tack-sharp focus within 90ms–120ms.
• hue (180ms): Subtle cool-tone entrance.
• none (0ms): Instantaneous text painting (zero overhead).

Single Caret Handoff Contract

When a rich Markdown payload spans paragraphs, lists, and code blocks: Multiple blinking carets are strictly forbidden! As downstream parts take over generation, upstream parts instantly retract their caret. When the answer completes, the final caret gently dissolves.