fix(litellm): bill codex/ChatGPT OAuth path at cache-read rate

[PurpleCHOIms]

Problem

The codex_chatgpt handler (ChatGPT/Codex OAuth backend, auth/gpt-*) dropped the Responses-API prompt-cache hit count from its usage object, so those calls were cost-calculated at full input rate — the OAuth GPT path showed 0% cache in LiteLLM_SpendLogs while the Claude OAuth path (auth/claude-*) correctly showed 88–90% cache-read. On a real 3-engagement sample, GPT-5.5(codex) was ~20–23% of engagement cost, entirely uncached in the billing figures.

Fix (2 places)

1. config/codex_chatgpt_handler.py — surface usage.input_tokens_details.cached_tokens as OpenAI-style prompt_tokens_details.cached_tokens in both the completion and streaming final-usage paths (mirrors claude_code_handler surfacing cache_read_input_tokens). Absent field → 0 → no-op.
2. config/litellm_dynamic_config.py — add cache_read_input_token_cost to the auth/gpt* shadow pricing (GPT-5 generation cached input = input × 0.1, i.e. 90% off; verified vs OpenAI pricing: gpt-5.5 $5→$0.50/1M, gpt-5.4 $2.50→$0.25/1M). LiteLLM only discounts cached tokens when this cost is present in model_info.
3. test — test_codex_chatgpt_handler_cache.py: cached_tokens surfaced (present / absent / zero), hermetic module loader. 277 passed, ruff clean.

Why draft — verification pending

The change is defensive (harmless no-op if the backend omits the field) and unit-tested, but not yet verified against a live codex /responses payload (local codex refresh token was rotated by prod → refresh_token_reused, can't authenticate locally). External evidence that the field is present: OpenAI Responses API spec includes input_tokens_details.cached_tokens; Codex CLI keeps stable prefixes for caching; promptfoo #7546 shows the Codex SDK returns cached_input_tokens.

To finish: either (a) codex login + confirm a real payload carries input_tokens_details.cached_tokens, or (b) after deploy, confirm LiteLLM_SpendLogs shows cache_read > 0 for auth/gpt rows.

Deploy note (SaaS)

This is baked into the litellm image config → applying to prod needs the litellm image rebuilt + image_tag_litellm bumped + redeploy. No effect until then. Commercial (metered openai/gpt-* API) path caches automatically regardless; this fix is for the current OAuth dogfooding period's cost accuracy.

🤖 Generated with Claude Code

[PurpleCHOIms]

검증 결과 닫습니다. 2026-07-05 fresh codex login + 직접 /responses 호출로 확인: 동일 7,556-tok 프롬프트 3회 + prompt_cache_key까지 써도 cached_tokens=0. 근본 원인은 우리 코드가 아니라 ChatGPT-구독 백엔드(chatgpt.com/backend-api/codex)의 캐싱 한계 — store:false 강제(store:true→400) + OpenAI 공식 이슈 openai/codex#5556(ChatGPT-login 모드 캐시 ~1:1 붕괴 vs API-key 모드 20:1). 이 수정은 correct하나 백엔드가 항상 0을 주므로 OAuth 경로에서 inert(비용 변화 0). 실효 레버는 상용 metered openai/ API 전환(자동 캐싱 정상) 또는 GPT-5.5 비중 축소. 백엔드 동작이 바뀌거나 API 경로로 갈 때 재개 가능.