Summary

Some Codex reasoning models (observed on gpt-5.3-codex-spark) emit tool-call arguments in a single shot on response.output_item.done and skip the response.function_call_arguments.delta / .done pair that the OpenAI Responses API spec requires between output_item.added and output_item.done.

Downstream clients that assemble the final arguments string from the delta events (LiteLLM's Anthropic-Messages adapter, Cursor, etc.) end up with empty arguments — e.g. an Anthropic tool_use with input: {} — which breaks any tool that needs parameters (Bash, file writes, ...).

This is the symmetric counterpart to #3: that PR fixed the same class of issue on the Chat Completions path (via the Vercel AI SDK tool-call handler); this one fixes it on the Responses API path.

What changes

• New normalizeCodexResponsesSseStream transformer in openai-oauth-core/src/sse.ts. Wraps the upstream Codex SSE ReadableStream and, when a function-call output_item.done arrives with non-empty arguments and no prior delta for the same item_id, injects a spec-compliant .delta + .done pair immediately before re-emitting the done event.
• handleResponsesRequest in packages/openai-oauth/src/responses.ts now pipes the upstream streaming body through that transformer instead of passing it through verbatim.
• Exposed the new function from openai-oauth-core's barrel.
• 6 new unit tests in packages/openai-oauth-core/test/sse.test.ts covering:
  • synthesis when Codex skips delta events
  • no duplication when Codex already streams deltas
  • non-function-call output items pass through unchanged
  • multiple function calls tracked independently
  • skipped synthesis when arguments are empty
  • tolerance for non-JSON SSE payloads

Why this is the right layer

openai-oauth publishes /v1/responses as an OpenAI-public endpoint on top of the Codex-internal backend. It already does request-side normalization (stripping temperature/top_p/user, renaming web_search_preview → web_search, ...). The response-side pass-through in handleResponsesRequest assumed the upstream stream was already spec-conforming; for codex-spark that assumption is violated. Fixing it here means every downstream consumer of the proxy — not just LiteLLM — benefits without needing its own workaround.

Test plan

• bun run test — 64 tests pass (26 in openai-oauth-core, incl. 6 new)
• Rebuilt via bun run build and deployed to a local install; restarted openai-oauth; confirmed Claude Code subagents routed through LiteLLM → codex-gpt-5.3-codex-spark now receive populated tool_use.input for Bash calls where previously input: {} appeared and agents looped on InputValidationError: command is missing.
• Recommended reviewer verification:
  • bun run test
  • ANTHROPIC_DEFAULT_HAIKU_MODEL=codex-gpt-5.3-codex-spark claude-codex -p "list /tmp" and check the Bash tool call receives a real command argument
  • Regression: same thing with codex-gpt-5.4 to confirm delta-streaming models behave unchanged