Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
56 lines (45 loc) · 2.53 KB

ARCHITECTURE.md

File metadata and controls

56 lines (45 loc) · 2.53 KB
Raw

Architecture

One-page map of the codebase.

Request pipeline

client (Claude Code / any SDK)
  │  ANTHROPIC_BASE_URL=http://127.0.0.1:47821
  ▼
src/node.ts · src/worker.ts        hosts (node:http / Cloudflare Workers)
  ▼
src/core/proxy.ts                  single Web-standard fetch handler:
  │                                routing, auth passthrough, count_tokens
  │                                counterfactual, usage/telemetry events
  ▼
src/core/transform.ts              THE pipeline (Anthropic path):
  │   1. parse body, resolve vision tier from model
  │   2. profitability gate — exact image cost vs text cost
  │   3. convert: static slab · large tool_results · collapsed history
  │   4. splice back preserving the client's cache_control anchors
  ▼
upstream API (api.anthropic.com / api.openai.com)

Billing (exact, measured)

module provider model
src/core/anthropic-vision.ts Anthropic 28 px patches + 4/block, per-tier resize caps; page geometry (both tiers render the standard 1568×728 page — the high-res tier is a billing trap, see BENCHMARKS)
src/core/gpt-model-profiles.ts + openai.ts OpenAI per-model patch/tile regimes, detail per profile, strip geometry
src/core/gemini-model-profiles.ts Google tile formula (floor(min/1.5) crop unit) + media_resolution flat costs

Rendering

  • src/core/render.ts — text → PNG via a baked glyph atlas (Spleen 5×8 + Unifont fallback), reflow with ↵ newline sentinels, 1-bit atlas in production (measured better than AA on Fable).
  • src/core/render-cache.ts — LRU memoization of deterministic renders (static slab + frozen history chunks re-render on every request otherwise).
  • src/core/history.ts — collapses old turns into append-only frozen image chunks that stay byte-identical so prompt caching keeps hitting.
  • src/core/png.ts — minimal deterministic PNG encoder (no native deps).

Guard rails

  • Model allowlist (src/core/applicability.ts): only models that passed the reading benchmark get imaged; everything else passes through byte-identical.
  • Byte-exact values (SHAs, ids) ride as text in a fact sheet next to the image (src/core/factsheet.ts); recoverable originals via emitRecoverable.
  • Native typed tools (type !== 'custom') are never rewritten (400 guard).

Benchmarks & receipts

benchmarks/ holds the two harnesses that produced every number in the README — see benchmarks/README.md.