Summary

The Codex /v1/responses backend differs from the public OpenAI Responses API in two ways that matter for clients forwarding through openai-oauth (LiteLLM → Anthropic-format, Cursor, etc.):

1. Codex reasoning models reject user / temperature / top_p. LiteLLM's Anthropic-Messages adapter forwards temperature by default (Claude Code sets it on every request), so every call errors out with:

   The feature "temperature" is not supported. temperature is not supported for reasoning models

   before the backend even starts generating.

2. Codex uses the internal tool name web_search for what OpenAI publicly calls web_search_preview. It also defaults to cached-only mode and requires external_web_access: true for live searches. Without those two adjustments the tool is accepted but runs in cached mode, so the model returns plausible-looking but hallucinated results.

Additionally, when web_search_preview is renamed, a caller-supplied tool_choice that points at a specific function becomes stale (Codex returns Tool choice 'function' not found in 'tools' parameter). Reset it to \"auto\" in that case; \"none\" is preserved intact.

What changes

• normalizeCodexResponsesBody in packages/openai-oauth-core/src/transport.ts:
  • delete normalized.user
  • delete normalized.temperature
  • delete normalized.top_p
  • Map tools[*].type === \"web_search_preview\" → { type: \"web_search\", external_web_access: true, ...rest }
  • If that rename happened and tool_choice !== \"none\", set it to \"auto\"
• 6 new unit tests in packages/openai-oauth-core/test/transport.test.ts covering stripping, renaming, tool_choice handling in both the rename and no-rename paths, and preservation of sibling keys (e.g. search_context_size) on the web_search tool config.

None of these changes affect requests that don't carry the offending params or tools.

Related

• Companion to fix: synthesize function_call_arguments events for one-shot Codex tool calls #6 (response-side synthesis of function_call_arguments events for codex-spark). Together they make the Responses API round-trip via LiteLLM usable for Anthropic-format clients with tool calls and WebSearch.

Test plan

• bun run test — all tests pass (+6 new)
• Recommended reviewer verification:
  • bun run test
  • End-to-end with claude-codex and a WebSearch-capable prompt: confirm the backend performs a live search, e.g. the response mentions today's date or content that post-dates the model's knowledge cutoff.