Skip to main content
Hooks can only add context to the window; they can never rewrite the history Claude re-sends every turn. To genuinely reduce tokens, DietCode sits in the request path.
This starts a small proxy on 127.0.0.1 (ephemeral port, torn down when you quit), points Claude Code at it via ANTHROPIC_BASE_URL, and launches claude normally.

​
What the proxy sends

​
Why it works

Progressive

Most turns make zero ScaleDown calls. One call folds the oldest turns into the summary only when summary plus recent turns passes SCALEDOWN_PROXY_COMPACT_THRESHOLD.

Replaces native compaction

The forwarded payload stays small, so Claude’s own auto-compaction effectively never fires. ScaleDown owns compaction instead.

Reversible

Compacted turns are stored locally. The summary tells Claude to call sd_retrieve("<id>") when it needs an exact detail.

Cache-safe

System prompt and tools are untouched, so Anthropic’s prompt cache stays warm.

​
Run the proxy yourself

Why not a PreCompact hook? It can only append to the window, so injecting a summary there costs tokens instead of saving them. The real work happens in the proxy, where DietCode controls the request payload.