Audit & Resource: Claude Code’s Uncached Tool Overhead and the Context Shuffle (90% Cost Reduction)

[alxsuv]

The Observation

While auditing the claude-code CLI request bodies (v2.1.126), I identified two patterns leading to significant token waste:

1. Uncached Tool Overhead: Static tool catalogs (often ~24k tokens) are frequently shipped without cache_control headers or with missing ttl fields, defaulting to 5-minute expires. This triggers a 1.25x cache-write surcharge on subsequent turns.
2. The "Context Shuffle": A more recent shift where static project metadata is moved between message indices and injected with <system-reminder> scaffolding, effectively "salting" the cache to ensure a miss.

The Fix: Pino

I’ve built Pino, a ~500-line, zero-dependency Node.js reverse proxy. It surgically re-anchors shuffled context to Msg 0, injects missing cache control headers, and trims historical scaffolding while preserving the model's reasoning integrity.

Real-world numbers (Opus, 2-turn session):

• Native CLI: ~$2.60
• With Pino: ~$0.34

Request

I’d love to get this into the awesome-claude-code list as a resource for engineers who are currently watching their API credits evaporate.

I’m also curious if anyone else has observed similar "shuffle" patterns in their logs?

[MitchellCEOminongsentinel]

hey im mitch first its a real pleasure to be online with guys and gals with your capabilities so thankyou . im experiancing difficultities making my we site public its . minongsentinel.github.io