Claude Code → Responses translation keeps the per-request billing header, so Muse/OpenAI prompt cache never hits #6627

Client or integration

Claude Code

Area

Proxy and routing

Summary

Claude Code requests translated to a Responses upstream get essentially no prompt-cache hits. I saw this with meta-muse (Muse Spark 1.3 Contributor). The translated instructions start with Claude Code's per-request billing line:

x-anthropic-billing-header: cc_version=…; cc_entrypoint=…; cch=…;

cch changes on every request. So the first bytes of the prompt differ every turn, and a prefix cache never matches. Anthropic excludes this line from caching; Muse and OpenAI do not. src/adapters/google.ts already strips it (stripAntigravityBillingHeader). systemToInstructions in src/claude/inbound-content-options.ts does not, and it passes the line through to every translated route. That function is unchanged on dev at 0511f458f1.

The system-derived prompt_cache_key fallback hashes the same system text, so it also rotates per request when a client sends no metadata.user_id.

Reproduction

1. Route a Claude Code subagent (Desktop entrypoint, first-party mode) to meta-muse/muse-spark-1.3-contributor, directly or through a combo.
2. Run a 5-turn task, where each turn is one tool call.
3. With OPENCODEX_CACHE_DEBUG=1, cache-debug.jsonl shows:
   • prompt_cache_key is present and equal on every turn.
   • The tools and the later instruction and message blocks keep stable tags.
   • Only the first instruction block's tag changes on every turn.
   • Upstream cached_tokens is 0 on every turn.
4. Control: send /v1/responses directly with a stable ~15k-token instructions and a fixed prompt_cache_key. Turns 2–3 cache about 99%. That rules out the account, the key and non-ASCII content.

claudeCode.stabilizePromptCache: true does not help, because it only relocates trailing footers.

Verified fix. Strip a leading x-anthropic-billing-header: line in systemToInstructions, matching only at the prompt start, with the same regex as the Google adapter. Apply it to the string itself for a string system, and to the first text block only for an array system, dropping the block if it ends up empty:

function stripClaudeBillingHeader(text: string): string {
  return text.replace(/^x-anthropic-billing-header:[^\n]*\n*/, "");
}

With that patch, the same 5-turn subagent went from 0 cached on every turn to 21,361, 22,129, 22,257 and 22,385 cached on turns 2–5 (about 99%), on two machines. Native Anthropic passthrough doesn't use this translation and is unaffected.

Version

opencodex 2.76.0, Claude Code 2.1.286

Operating system

macOS

Provider and model

meta-muse / muse-spark-1.3-contributor (also reached through a failover combo)

Logs or error output

usage.jsonl, same Muse account:

• Codex (inboundProtocol: responses): 96.5% of input cached, over 63,895 requests.
• Claude Code (inboundProtocol: messages): 0.7% cached, over 15,907 requests, every day since first use.

Redacted configuration

claudeCode.desktopMode: first-party; meta-muse provider adapter: openai-responses, baseUrl: https://api.meta.ai/v1.

Checks

• I searched existing issues and documentation.
• I removed secrets, tokens, account details, request credentials, and personal data.