fix(proxy): strip Claude Code billing header from prompt for non-Anthropic providers

[jsboige]

Summary

Strips the x-anthropic-billing-header: cc_version=...; cch=XXXXX; line that Claude Code injects into the system prompt body when the request is not routed to Anthropic's native API.

Problem

Claude Code injects a billing header into the system field of every /v1/messages request. The cch= token changes on every request, which breaks prefix caching on self-hosted inference engines (vLLM, Ollama, LM Studio) that use strict hash matching for KV cache hits. Each request is treated as a unique prefix, resulting in zero cache hits and wasted GPU compute re-processing the same system prompt repeatedly.

Fix

• In the /v1/messages handler, after resolving the handler and before forwarding, check if the handler is a NativeHandler (Anthropic direct).
• If not native (i.e., OpenRouter, local models, custom endpoints, etc.), strip the billing header line from body.system using a regex.
• Handles both string and array (block) forms of the system field per the Anthropic API spec.
• Zero impact on Anthropic routing — native requests pass through unchanged.

Testing

• Verified diff scope: only packages/cli/src/proxy-server.ts touched (23 insertions, 0 deletions).
• The NativeHandler import already exists in the file — no new dependencies.

Co-Authored-By: Claude Opus 4.7 noreply@anthropic.com

[mtparet]

We have the same fix #101 this is really a must have otherwise claudish is not usable because it cost 10X @erudenko