Demo the lossless codec (compress): 79%/88% fewer, lossless, fact kept — replaces the shrink-allocator demo

README.md

@@ -1,19 +1,22 @@
 ---
 title: Foveance
-emoji: ✂️
+emoji: 🗜️
 colorFrom: indigo
-colorTo: yellow
+colorTo: green
 sdk: gradio
 app_file: app.py
 pinned: true
 license: apache-2.0
-short_description: Shrink your LLM context 60%+ at the same accuracy
+short_description: Lossless token codec for LLM context, up to 82% fewer
 ---
 
-# Foveance — shrink your LLM context, live
+# Foveance — a lossless token codec for LLM context, live
 
-Interactive demo of [Foveance](https://github.com/aimaghsoodi/foveance): cut your LLM token bill
-60%+ without changing your code or your answers. Paste a conversation, pick a budget, and watch it
-compress. Runs offline — no API key.
+Interactive demo of [Foveance](https://github.com/aimaghsoodi/foveance): cut LLM input tokens up to
+**82% — losslessly**. Agent context is full of repetition (the same listings, tool outputs and stack
+traces on every turn); like `gzip` finds repeated bytes, Foveance finds text that already appeared
+and replaces it with a compact back-reference, removing tokens without removing information. Nothing
+is summarised or dropped — every fact survives (`unpack(pack(x)) == x`). Paste a `messages` list and
+watch it compress. Runs offline — no API key, no model call.
 
-`pip install foveance`
+`pip install foveance` · `foveance wrap --codec --agentic-codec claude`