Investigate Gemini 3 flash repetition/looping in browser tasks

[enyst]

Summary:

• Gemini 3 Flash Preview sometimes enters a long repetition loop during browser tasks (e.g., Microsoft login account picker), repeating phrases like "Wait, I'll try to click on index 3090" hundreds of times.
• Responses hit max_output_tokens (~65k), take ~5 minutes, and cost ~$0.20.
• Suspected to be Gemini API behavior; Claude (Haiku/Sonnet/Opus) with MCP Playwright does not show this.

Questions to investigate:

1. Does the agent SDK have built-in repetition/loop detection or loop breakers that should catch this?
2. Are there recommended frequency_penalty / repetition_penalty settings for Gemini (via LiteLLM) to mitigate this?
3. Can streaming in the OpenHands UI be enabled to detect/abort long single-response loops?
4. How often does this occur in CI/evals (see .github/workflows/integration-runner.yml) and can we make it reproducible?

Context:

• User uses OpenHands Agent-SDK via OpenHands UI with VNC browser for complex sites (e.g., Outlook).
• Issue appears "not rare" for them.
• Benchmark shows Gemini 3 Flash performing well, but reliability in complex UI is questioned.

References:

• Discussion mentions integration-runner workflow:

  software-agent-sdk/.github/workflows/integration-runner.yml

  Line 99 in 8296a7f

  llm-config:

• Benchmark: https://index.openhands.dev/home