tweak: add experimental chatMaxRetries to config

[dfavato]

This PR tries to give users some power to work around AI_RetryError.

Some related issues this is aiming are:

• Implement exponential back-off when hitting rate limits
• Retry error
• AI_RetryError: Failed after 4 attempts. Last error: Provider returned error (Qwen3 Coder 480B A35B Instruct (
• Feature Request: Request delay setting #1592

This PR is also related to feat: Add retry mechanism for Anthropic overloaded errors however, in #1027 the retry backoff is re-implemented and I'm not sure this is the right approach.

Exponential backoff, and Retry-After behavior is already implemented at ai.util.retry-with-exponential-backoff and it is used in the streamText function here.

I could not find handles in the ai package where we could expose the backoff parameters (delay, maximumDelay, factor, etc..).

If someone can point me to where I should update the documentation, I can do that.

Here's a log from a session where I changed my config.chatMaxRetries to 5.

ERROR 2025-08-20T14:34:51 +14ms service=tui message=AI_RetryError: Failed after 6 attempts. Last error: Provider returned error name=UnknownError Server error

[thdxr]

can you put this flag under experimental ? will merge it after

[dfavato]

Changed.

Here's a log from a session where I changed my config experimental.chatMaxRetries to 1.

ERROR 2025-08-22T14:37:07 +4ms service=tui message=AI_RetryError: Failed after 2 attempts. Last error: Provider returned error name=UnknownError Server error

And here's a log from a session where I used the default config.

ERROR 2025-08-22T14:48:18 +2ms service=tui message=AI_RetryError: Failed after 4 attempts. Last error: Provider returned error name=UnknownError Server error

[outp1]

Why has this become stale? We cannot use OpenAI models, and we need a solution (

[dfavato]

Seems like the team chose to just increase the maxRetries and not expose the config to the users
https://github.com/sst/opencode/blob/dev/packages/opencode/src/session/compaction.ts#L126
https://github.com/sst/opencode/blob/dev/packages/opencode/src/session/prompt.ts#L266

Let me know if you want me to update this PR with the conflict resolution, if not you can close this.

[Sewer56]

I'd be interested nonetheless, because a part of me is curious to see how things would go if I left the agent overnight.
(I've never tried this sort of thing, but even if the result isn't perfect, having something being done when you're not around sounds like a win)

Caveat is, if I leave it overnight, and it gets rate limited, or servers are overloaded, etc. I wouldn't want it to stop; I'd rather have it keep trying as long as it can.

So I could set the limit to like 999 to never stop trying.

[Sewer56]

Ideally exponential backoff would be perfect with this, but I'm not sure if the dev team is interested on that for the time being.

[dfavato]

Exponential backoff, and Retry-After behavior is already implemented at ai.util.retry-with-exponential-backoff and it is used in the streamText.

So this PR would only let users increase the number of retries tried with exponential backoff.

[Sewer56]

Might be worth closing the issue for exponential backoff then.
I was actually unaware this was already implemented.
Seems it was implemented before I started using OpenCode

I still think it'd be useful nonetheless.
On bad times, I've seen Anthropic servers fail for multiple minutes at a time.

[dfavato]

Solved the conflicts just in case.

[rekram1-node]

Im gonna resolve the merge conflicts and merge this pr

[Sewer56]

Ayo merged!