Feature: configurable failover reasons

[a-bali]

As per the docs, failover currently only happens in case of 5xx and 429 errors. OpenCode seems to return 400 in case of temporary upstream issues (see request & log below); would it be possible to extend failover reasons or make them configurable?

{"time":"2026-07-28T06:06:38.074537266Z","level":"WARN","msg":"request failed","type":"invalid_request_error","status":400,"message":"Error from provider (Console Go): Upstream request failed","provider":"open
code_go","code":"invalid_request_error","method":"POST","path":"/v1/chat/completions","request_id":"c97141be-c4ed-4f2b-aa69-c17ffe6a62f4"}
{"time":"2026-07-28T06:06:38.075040719Z","level":"INFO","msg":"REQUEST","method":"POST","uri":"/v1/chat/completions","status":400,"latency":"553.201393ms","host":"ai","bytes_in":"1485","bytes_out":157,"
user_agent":"OpenAI/Python 2.24.0","remote_ip":"10.244.0.37","request_id":"c97141be-c4ed-4f2b-aa69-c17ffe6a62f4"}
{"time":"2026-07-28T06:06:38.669617633Z","level":"WARN","msg":"request failed","type":"invalid_request_error","status":400,"message":"Error from provider (Console Go): Upstream request failed","provider":"open
code_go","code":"invalid_request_error","method":"POST","path":"/v1/chat/completions","request_id":"eb834987-9e89-419d-bf04-da47a6d7044b"}
{"time":"2026-07-28T06:06:38.670858768Z","level":"INFO","msg":"REQUEST","method":"POST","uri":"/v1/chat/completions","status":400,"latency":"578.032008ms","host":"ai","bytes_in":"1240","bytes_out":157,"
user_agent":"OpenAI/Python 2.24.0","remote_ip":"10.244.0.37","request_id":"eb834987-9e89-419d-bf04-da47a6d7044b"}

Note that the actual model targeted by the virtual model does have failover targets set:

[SantiagoDePolonia]

@a-bali Thank you for reporting.

This should be fixed and released - v0.1.63 . Let me know how it works for you.

You are almost right that "... failover currently only happens in case of 5xx and 429 errors". There is also a couple of conditions not related to the status code (as visible on the screenshot). I've added a new one to fix your case.

Those exceptions exists because - as you have noticed with opencode go - the providers APIs are not perfect (sometimes return inappropriate HTTP status codes). With this fix we can somehow make it working for them :)

[a-bali]

Following up on this, it works for the original case, but is there a reason for not allowing extension of failover reasons? For example, opencode now came up with 403 to notify about an opt-in necessity. Honestly, I would prefer GoModel to just fail over in any case.

[SantiagoDePolonia]

I've reopened it and will think about it soon.