OmniGlyph 1.0.0

[1.0.0] — 2026-07-07

First public release.

The product

• Context-as-image compression proxy: rewrites the bulky parts of each LLM
  request (system prompt, tool docs, old history, large tool outputs) into dense
  1-bit PNG pages before they leave your machine. Local Node server and
  Cloudflare Workers host.
• Exact per-provider billing math (src/core/): Anthropic 28px patches +
  3–4 tokens/block overhead (own sweep, zero residual), OpenAI and Gemini
  formulas audited against official docs. Exported at the package root
  (anthropicImageTokens, resolveAnthropicVisionTier, tier caps).
• Measured production render config: dense 1-bit glyph atlas (no
  anti-aliasing), standard-tier pages — every choice backed by a benchmark
  receipt in benchmarks/*/results/.
• Benchmark harnesses (benchmarks/): billing-sweep (token accounting) and
  density-frontier (read-accuracy frontier across models/densities), re-runnable
  via API, OpenRouter, Claude Code CLI, or through OmniRoute
  (--via-omniroute).
• Refusal retry: SSE/JSON sniffer replays the original request when a model
  refuses the rendered page (kill switch retryRefusalWithOriginal).
• LRU render cache for deterministic pages.
• OmniRoute engine: ships as the omniglyph compression engine in
  OmniRoute (single mode and
  stacked pipeline), with fail-closed gates and image-aware token accounting.

The numbers (all reproducible)

• Sample UI render: 1015 chars → 438×120 PNG, 254 → 84 tokens (66.9% saved).
• Standard page 1568×728 = 1456 image tokens regardless of how much text it holds.
• Claude reads dense 1-bit pages at 100% on production density; Opus 4.8 reads
  77–87% at 10×16.

Negative decisions (measured, not opinions)

• High-res tier is a billing trap: the 1928² page is billed WYSIWYG but the
  encoder does not receive full resolution — both tiers render standard pages.
• GPT-5.5 rejected: 0/60 reads of the dense strip and ~40× completion
  inflation vs text control.
• gpt-4o-mini never imaged (2833/5667 token floor makes it unprofitable).
• Gemini 2.5-flash confabulates instead of abstaining on dense pages
  (0/26) — pending paid-quota retest.