[FEATURE] Token cost transparency and user control for inline Visualizer

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Summary

Claude's inline Visualizer generates SVG/HTML widgets as part of the model's output tokens. This is a fantastic feature, but it introduces a 3-12x token cost multiplier that users cannot see, control, or opt out of.

Problem

Token cost per response type

| Response Type | Output Tokens | Multiplier |
|---|---|---|
| Plain text | ~300 | 1x (baseline) |
| Text + simple SVG | ~1,200 | 4x |
| Text + interactive HTML | ~3,500 | 10-12x |
| Multi-visual | ~5,000+ | 15-17x |

Compounding factors

1. Output explosion: Each visual costs 3-12x more output tokens
2. Context pollution: Rendered widget code persists in conversation history as input tokens (~10-20k dead tokens by turn 15)
3. No user control: Model proactively generates visuals; no per-message or global opt-out

Impact on message budgets (200k token budget)

| Visual response % | Messages available | Reduction |
|---|---|---|
| 0% (text only) | ~182 | baseline |
| 25% | ~105 | −42% |
| 50% | ~74 | −59% |
| 75% | ~57 | −69% |

Proposed Solution

Proposed Solutions

High priority

• Visual richness slider — Setting with levels: Off / Minimal (static SVG) / Standard (SVG + styled HTML) / Full (interactive JS widgets). Sits alongside web search and artifacts toggles.
• Context pruning — Replace rendered widget_code in conversation history with [rendered: {title} | {type}] placeholder. Saves 1-4k input tokens per past visual, per turn.
• Per-message override — Toggle in input field: "respond without visuals"

Medium priority

• Token transparency — Per-message token count on hover. Budget meter. Visual-heavy response indicator.
• Widget template system — Model emits {template: "flowchart", data: {...}} (~100 tokens) instead of raw SVG (~1,500 tokens). Client renders from pre-built templates. 80-95% reduction.

Long-term

• Separate visual token accounting — Don't count widget markup against conversation budget

Additional Context

This analysis was developed collaboratively with Claude in a conversation that itself demonstrated the problem — the visual-heavy conversation consumed approximately 4-5x more tokens than an equivalent text-only discussion.

The Visualizer is one of the best LLM UI innovations I've seen. This request is about preserving its value while giving users informed choice.

Alternative Solutions

_No response_

Priority

High - Significant impact on productivity

Feature Category

Interactive mode (TUI)

Use Case Example

_No response_

Additional Context

_No response_