Overview

​

What it does

1. Classifies intent and prepends a one-line hint such as [DietCode intent: file_read (87%)], so the agent picks the right tool.
2. Compresses large contexts when you paste a big codebase and ask a retrieval-style question, cutting 50-70% of tokens before the prompt reaches the model.
3. Compresses large tool outputs. ls, grep and git diff/log/status are compacted with no latency; anything still large goes through ScaleDown.
4. Tracks savings live in the Claude Code status line, for example ↓ 125.4K saved · 747 reqs.

your prompt ──▶ [intent classify] ──▶ [needle-in-haystack?] ──▶ compress ──▶ agent
                     │                        │
                 one-line hint          /compress/raw/  (50-70% smaller)