fix: strip x-anthropic-billing-header to fix vLLM prefix caching

[mtparet]

Summary

• Strip x-anthropic-billing-header from system prompts in filterIdentity to fix vLLM prefix caching
• Add unit tests for the identity filter

Problem

Claude Code injects x-anthropic-billing-header: cc_version=...; cch=<hash> into the system prompt on every turn. The cch (conversation context hash) changes per turn, making the system prompt tokens differ between requests. This breaks vLLM's prefix cache for the entire ~31k token prompt — reducing cache hits from 99% to 0.15% (only 48 tokens cached).

Root Cause Analysis

1. Sent 4 identical "test" messages via claudish → MiniMax M2.5 on vLLM
2. VictoriaMetrics showed prefix cache metrics were healthy (~50% cluster-wide)
3. But per-request cached_tokens was stuck at 48 out of ~31k
4. Dumped full request payloads and diffed between turns
5. Found the only difference: cch=8ae40 → cch=4e20a → cch=ad80e in the system prompt

Fix

Added .replace(/x-anthropic-billing-header:[^\n]*\n?/g, "") as the first filter in filterIdentity(). This is safe because filterIdentity only runs for non-Anthropic providers (OpenAI-compat, Gemini, LiteLLM, OpenRouter, local) — never for native Anthropic passthrough where the header is meaningful.

Validation

Before fix: cached_tokens = 48 / 31k on every turn (0.15%)
After fix: cached_tokens = 31,728 / 31,749 by 4th turn (99.9%)

Turn	Before Fix	After Fix
1st	48 (0.15%)	32 (cold)
2nd	48 (0.15%)	24,864 (79%)
3rd	48 (0.15%)	31,296 (99%)
4th	48 (0.15%)	31,728 (99.9%)

Test plan

• bun test identity-filter.test.ts — 9 tests pass
• bun run build — compiles clean
• Manual: run claudish with --debug, send "test" twice, verify cached_tokens grows

🤖 Generated with Claude Code

[mtparet]

cc @erudenko is the CI failure expected ?