fix(llm): Java OpenAI provider 400s on gpt-5 reasoning models with tools (spring-ai has no Responses path)

[dhyansraj]

Summary

Cross-runtime parity follow-up to #1334. The Java native OpenAI handler (src/runtime/java/mcp-mesh-spring-ai/src/main/java/io/mcpmesh/ai/handlers/OpenAiHandler.java) uses spring-ai OpenAiChatModel / OpenAiApi, which target /v1/chat/completions only. So a Java LLM provider on a gpt-5 reasoning model (e.g. gpt-5.6-terra) with function tools 400s identically to the pre-#1334 Python path ("Function tools with reasoning_effort are not supported ... use /v1/responses").

Why it's separate from #1334

#1334 fixed the Python provider by routing reasoning+tools to the OpenAI Responses API. spring-ai has no drop-in Responses client, so the Java fix is a larger effort (a custom Responses call path, or waiting for upstream spring-ai support). TypeScript is unaffected — @ai-sdk/openai v3 already defaults to the Responses API.

Interim guidance

Because mesh delegation is provider-managed (the provider's runtime makes the OpenAI call), a Java consumer can already use reasoning+tools today by pointing at a Python (post-#1334) or TypeScript provider. Only a Java-as-provider on a reasoning model with tools is blocked.

Acceptance

• Java OpenAI provider on a gpt-5 reasoning model completes an agentic tool loop (reasoning preserved) without a 400.
• gpt-4o / gpt-5 chat variants unchanged (stay on chat.completions).
• Live smoke parity with the Python fix(llm): OpenAI gpt-5 reasoning models 400 with function tools (native path is chat.completions-only) #1334 both-models test.

Relates to #1334.

[dhyansraj]

Investigation (2026-07): spring-ai has no OpenAI Responses API yet — deferring.

No released spring-ai version — including 2.0.0 GA (2026-06-12) — supports the OpenAI Responses API; it only does Chat Completions. Native support is tracked upstream in spring-projects/spring-ai#2962 (OPEN, milestone 2.1.x).

Hand-rolling a Java Responses call that bypasses spring-ai would be throwaway once 2.1.x lands it natively. Decision: defer this issue until spring-ai 2.1.x, then implement Java parity via the native spring-ai Responses API.

Interim (unchanged): mesh delegation is provider-managed, so Java consumers can use reasoning+tools today via a Python (post-#1334) or TS provider. Only a Java-as-provider on a gpt-5 reasoning model with tools is blocked.

Blocked on: spring-projects/spring-ai#2962.