Summary

Add a stream config flag to fallback provider entries so custom endpoints that don't support SSE streaming can opt into non-streaming API calls.

Problem: Custom fallback providers that return plain JSON instead of SSE get zero chunks from the OpenAI SDK's streaming parser. This causes valid responses to be silently discarded as "empty," triggering spurious retry loops and fallback cascades. See #21522 for details.

Solution: Read a per-provider stream field from fallback config. When stream: false, set _disable_streaming for that provider. When advancing to a provider that supports streaming, clear the flag — streaming state no longer sticks across the fallback chain.

Changes

Example usage

fallback_providers:
  - provider: openai
    model: gpt-4o
  - provider: custom
    model: my-local-model
    base_url: http://my-proxy:8080/v1
    key_env: MY_PROXY_KEY
    stream: false    # This provider returns JSON, not SSE

Design decisions

• Minimal change: 12 lines of production code — reads a single config field and sets/clears an existing flag
• Backward compatible: stream defaults to true when omitted — zero behavioral change for existing configs
• Handles YAML string coercion: "false", "0", "no", "off" all treated as false (YAML parsers vary)
• State restoration: Each fallback activation explicitly sets or clears _disable_streaming based on its own config, so a stream: false provider doesn't poison subsequent streaming-capable providers

Closes #21522

Test plan

• python -m pytest tests/run_agent/test_fallback_streaming.py -v — 7 tests, all passing
• python -m pytest tests/run_agent/test_provider_fallback.py tests/run_agent/test_fallback_model.py -v — 53 existing tests, all passing (no regressions)
• Verified defaults match previous behavior (stream: true when omitted)
• Manual test with a non-SSE custom provider endpoint