fix(chatgpt): forward prompt cache key

[steveonjava]

TLDR

Problem this solves:

• ChatGPT Responses requests silently drop prompt_cache_key
• Session-scoped prompt caching cannot reach ChatGPT Codex

How it solves it:

• Forward prompt_cache_key through the ChatGPT provider allowlist
• Retain filtering of unsupported context_management

User Flow

Before: a developer sends a ChatGPT Responses request with a cache key, but the backend never receives it

1. They send POST https://litellm-domain/v1/responses with model chatgpt/gpt-5.6-sol and "prompt_cache_key": "session_123"
2. LiteLLM forwards a ChatGPT Codex request without prompt_cache_key
3. The request cannot use the caller's intended prompt-cache scope

After: the same request forwards the caller's cache key

1. They send the same POST https://litellm-domain/v1/responses with "prompt_cache_key": "session_123"
2. LiteLLM forwards the ChatGPT Codex request with prompt_cache_key unchanged
3. The backend receives the caller's intended prompt-cache scope

Relevant issues

• Related to bug: ChatGPT Codex rejects unsupported parameters (user, temperature, context_management, etc.) #21193 and Fix/21193 chatgpt codex unsupported params #21209
• Restores the field previously included by unmerged Add ChatGPT proxy credential support and managed Codex responses websocket support #24983
• Complements feat(chatgpt): add GPT-5.5 and GPT-5.6 models #35720 and fix(anthropic): map metadata.user_id to prompt_cache_key on the /v1/messages bridge #37623
• Does not forward context_management; live subscription proof was unavailable

Linear ticket

Pre-Submission checklist

• I have added meaningful tests
• The handful of test files covering my change pass locally
• My PR passes all required CI/CD checks (e.g., lint, schema.d.ts sync check, etc.)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review

Screenshots / Proof of Fix

Before (6d47468)

1. Run uv run pytest tests/test_litellm/llms/chatgpt/responses/test_chatgpt_responses_transformation.py::TestChatGPTResponsesAPITransformation::test_chatgpt_drops_unsupported_responses_params -q -o 'addopts=' after asserting prompt_cache_key is preserved.
2. Result: 2 failures, both KeyError: 'prompt_cache_key'.

After (025438c)

1. Run uv run pytest tests/test_litellm/llms/chatgpt/responses/test_chatgpt_responses_transformation.py -q -o 'addopts='.
2. Result: 19 passed, 1 existing dependency warning.
3. Run uv run pytest tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py -q -o 'addopts='.
4. Result: 63 passed, 2 existing warnings.

Type

🐛 Bug Fix

Caveats (if any)

• GPT-5.6 Sol/Luna/Terra context-management probes received identical 401 responses
• The stored ChatGPT subscription tokens were expired and refresh was unauthorized
• Keep context_management filtered pending authenticated backend acceptance proof
• Codex's documented /responses/compact path remains the likely compaction integration

Final Attestation

• The test covers allowlist preservation while retaining existing unsupported-field filtering

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing steveonjava:litellm_prompt_cache_key (682ffbf) with litellm_internal_staging (e07a712)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (cde1344) during the generation of this report, so e07a712 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩