Repository navigation
handle LLM overload errors (e.g. {"type": "overloaded_error", "message": "Overloaded"}) #833
Description
When using opencode with Anthropic, I occasionally hit this response from the model:
{"type": "overloaded_error", "message": "Overloaded"}This is the error Anthropic throws when it has capacity issues. Right now, opencode surfaces the error and halts. Manually typing continue will usually succeed on the next try, which suggests the failure is transient and recoverable.
Proposed behavior
Rather than require manual input, opencode could intercept this response and retry the request automatically. A minimal first pass might:
• Catch overloaded_error responses
• Retry the failed step after a short delay
• Limit retries to a sane default (e.g. 2–3 attempts) to avoid blocking or infinite loops
This would eliminate the need for manual intervention and improve flow, especially for longer sessions.
Next steps
If this direction makes sense, I can put up a PR that adds basic retry logic around the model call. Happy to scope it to whatever level of configurability or error handling you want to support longer term.
Just, let me know.
Activity
rwese commentedon Jul 15, 2025
a general handling of
529 The service is overloadedwould be a good idea, with dynamic wait, incremental wait or similar.having the option to set a 'max retry count', 'max delay' would be also great.
adamdotdevin commentedon Jul 15, 2025
we should better support 529, before submitting a PR can you look into AI SDK (v5) 529 handling? do they not support retries on these already?
rwese commentedon Jul 15, 2025
I searched, in various ways, found nothing which would handle 529 errors, but I might be missing something.
https://github.com/vercel/ai/blob/v5/content/docs/07-reference/05-ai-sdk-errors/ai-api-call-error.mdx
aisflat439 commentedon Jul 16, 2025
I'm also experiencing this and mentioned it in discord. I'll add the link here.
https://discord.com/channels/1391832426048651334/1391832428024430645/1395048739936276670
carlesandres commentedon Jul 16, 2025
It would be nice to have
/retrycommand to make the "continue" behaviour more obvious. I imagine there can always be backend failures that aren't yet handled properly byopencodeor AI-SDK.alexeypolitov commentedon Aug 17, 2025
@thdxr look's like this related by new issue? before, when I got "overloaded", i could write like "retry" and it continue execution. for now it restart from scratch
https://opencode.ai/s/O5qCRoYM
nesl247 commentedon Nov 25, 2025
Running into this all the time lately.
rekram1-node commentedon Nov 25, 2025
Doing it
rekram1-node commentedon Nov 25, 2025
9eabbe2
nesl247 commentedon Nov 25, 2025
Wow that was fast. Thank you very much!
Merge pull request anomalyco#833 from shreyas-lyzr/feat/gitagent-format
fix: route shared bundles through blueprints flow (anomalyco#833)