Intermittent connection failures during Berget API requests

[tcsizmadia]

The OpenCode Berget AI auth plugin can fail a request when transient network or gateway errors occur between OpenCode and Berget AI API services.

Observed failure modes include:

• Connection resets / socket hang-ups
• Connection timeouts
• Temporary DNS or upstream connectivity failures
• 502, 503, or 504 responses
• 429 rate-limit responses

Today, model discovery, token refresh, and authenticated inference requests generally fail immediately when these occur. This can disconnect or interrupt an otherwise valid OpenCode session even when the service recovers moments later.

[tcsizmadia]

Hi Berget AI Team,

I've created a PR: #35 - implementing a retry mechanism with a small jitter, so clients don't hammer your server with simultaneous retries.

Before I finalize it, I have a few questions to make sure the changes in the PR work as intended:

1. Chat/completions: Are repeated POST requests safe to retry after a connection failure, timeout, 429, or 5xx? If supported, does Berget AI accept an idempotency key, and which header/format should clients use?
2. Token refresh: If a refresh request succeeds server-side but the client loses the response, can the same refresh token be retried safely? Does a successful refresh always rotate and invalidate the previous refresh token?
3. Retry guidance: Which status codes and transport failures should clients retry? Does Berget AI emit Retry-After for 429 or 503, and should clients honor it?
4. Streaming: Are there recommended connection/header/idle timeouts for streamed chat responses? We want to avoid aborting healthy long-running streams.