test(openai): cover ReAct streamEvents backpressure regression

[guslegend0510]

AgentScope-Java Version

2.0.3-SNAPSHOT

Description

Refs #3055.
Related issue: #2957.
Follow-up to the transport fix in #2963.

The streaming stall described in #3055 matches the request-starvation issue already fixed in OkHttpTransport by #2963.

This PR adds regression coverage through OkHttpTransport, OpenAIChatModel, and ReActAgent.streamEvents():

• Require all 96 text deltas to arrive in order before sending [DONE].
• Cover ordinary consumption and cross-thread incremental demand.
• Verify exactly one model request and one model-call end event.
• Document why the blocking transport uses subscribeOn(Schedulers.boundedElastic(), false).

The existing runtime fix is preserved; this PR adds regression tests and an explanatory comment.

Validation

• Temporarily restoring the old request scheduling makes both new tests time out.
• Both tests pass with the existing fix.
• Full Core suite: 2,323 tests, 9 skipped.
• Full Harness suite: 1,010 tests, 14 skipped.
• Full OpenAI module suite: 557 tests, 12 skipped.
• Zero failures or errors across these suites.
• Spotless, required focused checks, and git diff --check passed.

The reproduction uses a controlled streaming response body; no live model API was called.

Checklist

• Code has been formatted.
• Relevant module test suites pass.
• Comments explain the scheduling constraint.
• No public API or user-facing documentation changes are needed.
• Code is ready for review.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[oss-maintainer]

Summary

Regression coverage for the #3055 streaming stall: a new OpenAIReActAgentStreamingBackpressureTest that drives 96 SSE deltas through OkHttpTransport → OpenAIChatModel → ReActAgent.streamEvents() and asserts every delta is delivered before [DONE], under both eager and incremental (publishOn(parallel(), 1)) demand, plus exactly one model request and one ModelCallEndEvent. The only production change is an explanatory comment on subscribeOn(Schedulers.boundedElastic(), false). Clean, valuable test-first PR — the negative control (restoring the old scheduling makes both tests time out) is exactly the evidence a regression test needs, and the fake-Pipe harness avoids any live model call.

Two small notes, neither blocking:

• [Info] agentscope-extensions-model-openai/src/test/.../OpenAIReActAgentStreamingBackpressureTest.java:78 — the response body is created with a null content type; passing MediaType.parse("text/event-stream") would make the fixture closer to what a real OpenAI-compatible endpoint returns, in case the decoder ever starts inspecting the content type.
• [Info] same file, responseSink is fully buffered before the request is subscribed. That is safe for 96 chunks (~8 KB against the 64 KB pipe) but will block the test thread (rather than fail cleanly) if someone scales expected up past the buffer. A one-line comment on that limit, or writing chunks from a separate task, would keep it future-proof.

CLA: signed ✅

---

Automated review by github-manager-bot