Claude Code Routing

What it does

• Three pieces, plain Claude Code: code-write (Haiku writes boilerplate to disk), bulk-read (files in, dense bullets out, Claude's context untouched), and a 350-line full-read block hook
• Settings: BLOCK_MIN_LINES (default 350), CHEAP_MODEL (default haiku), CLAUDE_BIN if claude is not on PATH; install.sh backs up settings.json before touching it
• Reproducible benchmark: four scenarios on psf/requests with a runner and analysis in benchmark/, one command to rerun (benchmark/run.sh opus, about $6-7)
• Negative results published too: the hook never fired across the four tasks and bulk-read was never called — Claude self-routed once the scripts existed

When to use it

Cost-conscious heavy Claude Code users who want routing with receipts — including the honest finding that routing pays on boilerplate (type stubs), not on judgment work like writing tests against an existing style.

Getting started

git clone https://github.com/ToolMonsters/claude-code-routing
cd claude-code-routing && ./install.sh   # copies scripts, hook and skills into ~/.claude (backs up settings.json first)
# then restart Claude Code

# reproduce the benchmark (clones psf/requests into benchmark/work):
benchmark/run.sh opus

Things to know

MIT. Not affiliated with Spotify. One run per benchmark cell — treat small gaps as noise, the author's own words. Ranged reads (offset/limit) pass the hook regardless of size; in S2 Claude slipped 545 lines through that door. The cheap model's spend is billed in separate claude -p calls, so it does not show in /cost — the scripts print it to stderr. Do not route debugging, architecture or anything subtle to the cheap model; Spotify's own account includes a worker missing a thread-safety bug Claude caught in seconds.