fix(model/openai): wrap streaming branch in Flux.defer to make retry re-issue HTTP requests

[birdie7761]

Description

Fixes a bug where streaming retry in OpenAIChatModel did not re-issue HTTP requests for retryable errors such as HTTP 429.

Root cause

OpenAIChatModel.doStream0() returned the streaming Flux directly:

  return client.stream(...).map(...).filter(...);

When ModelUtils.applyTimeoutAndRetry() later applied retryWhen, re-subscriptions reused the same publisher. For transports whose stream() is backed by a one-shot future (notably the default JdkHttpTransport), the same failed future was replayed
instead of sending a new HTTP request. As a result, a maxAttempts=3 retry exhausted after a single server hit.

Fix

Wrap the streaming branch in Flux.defer(), matching the existing non-streaming branch, so every retry creates a fresh publisher and re-issues the HTTP request:

  return Flux.defer(() -> client.stream(...).map(...).filter(...))
          .subscribeOn(Schedulers.boundedElastic());

Verification

Added OpenAIChatModelStreamingRetryTest with MockWebServer scenarios:

• Streaming 429 exhausts retries and hits the server maxAttempts times.
• Streaming 429 retries then succeeds on the final attempt.
• Streaming 400 does not retry.
• Streaming 429 retry re-issues requests for JdkHttpTransport, OkHttpTransport, and the default transport.
• Non-streaming 429 retry re-issues requests as a control.

All tests pass locally:

  mvn spotless:apply
  mvn test -pl agentscope-extensions/agentscope-extensions-model/agentscope-extensions-model-openai

Checklist

• [x] Code has been formatted with mvn spotless:apply
• [x] All tests are passing (mvn test)
• [x] Javadoc comments are complete and follow project conventions
• [ ] Related documentation has been updated (not needed for this internal bug fix)
• [x] 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

Fixes a bug where streaming retry in OpenAIChatModel did not re-issue HTTP requests for retryable errors (e.g. HTTP 429). Root cause: the streaming Flux was returned directly, so re-subscriptions replayed the same one-shot failed future instead of creating a new HTTP request.

Findings

• [Info] OpenAIChatModel.java:310 — Wrapping in Flux.defer() + subscribeOn(Schedulers.boundedElastic()) is the correct fix, matching the existing non-streaming branch pattern.
• [Info] OpenAIChatModelStreamingRetryTest.java — Excellent test coverage with MockWebServer: 429 exhaust, 429→success, 400 no-retry, multi-transport verification. The control test for non-streaming is a nice touch.

Verdict

Minimal, correct fix with thorough tests. LGTM.

---

Automated review by github-manager