gemini retry support

[spinegar]

I seem to be hitting a quota with Gemini Flash 2.0. When this occurs, it would be great if the MCP server could implement a retry strategy using the retryDelay property in the error message.

Stagehand[browserbase_session_ad367ce1-ed23-4d00-8f70-3dce12e4de12_1762539928814_973009f6-15c1-4597-b6d4-0207019abe38]: Error during Google AI chat completion: {"error":{"code":429,"message":"You exceeded your current quota, please check your plan and billing details. For more information on this error, head to: https://ai.google.dev/gemini-api/docs/rate-limits. To monitor your current usage, head to: https://ai.dev/usage?tab=rate-limit. \n* Quota exceeded for metric: generativelanguage.googleapis.com/generate_content_paid_tier_input_token_count, limit: 4000000\nPlease retry in 47.434392449s.","status":"RESOURCE_EXHAUSTED","details":[{"@type":"type.googleapis.com/google.rpc.Help","links":[{"description":"Learn more about Gemini API quotas","url":"https://ai.google.dev/gemini-api/docs/rate-limits"}]},{"@type":"type.googleapis.com/google.rpc.QuotaFailure","violations":[{"quotaMetric":"generativelanguage.googleapis.com/generate_content_paid_tier_input_token_count","quotaId":"GenerateContentPaidTierInputTokensPerModelPerMinute","quotaDimensions":{"location":"global","model":"gemini-2.0-flash"},"quotaValue":"4000000"}]},{"@type":"type.googleapis.com/google.rpc.RetryInfo","retryDelay":"47s"}]}}

[yawbtng]

Hit this today in a real session when using the Browserbase MCP to explore a landing page (stora.sh — ~14 pages, mix of extract and observe calls). Free-tier Gemini Flash Lite quota tripped at ~20 calls and the raw 429 leaks straight through to the MCP client, which is a rough first-run experience on the default model.

The good news: Gemini's 429 response already carries the answer. The payload includes a RetryInfo field with an explicit retryDelay:

{
  "@type": "type.googleapis.com/google.rpc.RetryInfo",
  "retryDelay": "47s"
}

Proposed fix:

• Wrap the LLM call in a retry handler that detects 429 + RESOURCE_EXHAUSTED
• Parse RetryInfo.retryDelay and back off for that duration (capped at a sane max, e.g. 60s)
• Limit total retries (default 2–3, env-configurable)
• On final failure, surface a friendlier error than the raw provider payload and point at --modelName / --modelApiKey as the escape hatch

Happy to take this if it's unclaimed. Have one merged in the org (browserbase/stagehand#2190) and a few others open. Can have a PR up this week with a test.