Implement exponential back-off when hitting rate limits

[BigMitchGit]

OpenAI has some brutal RPM rate limits on tier 1, which will often abruptly stop opencode and requires me to manually resume by sending 'continue' message.

An exponential back-off with auto-retry would keep the agent loop running, it'd just take a bit longer to complete when rate limits are hit.

[daliusd]

If that helps here is error you get when you hit rate limit:

{"type":"error","sequence_number":2,"error":{"type":"tokens","code":"rate_limit_exceeded","message":"Rate limit reached
for gpt-5-mini in organization org-XXXXXXXXXXXXXX on tokens per min (TPM): Limit 200000, Used 124035,
Requested 79017. Please try again in 915ms. Visit https://platform.openai.com/account/rate-limits to learn
more.","param":null}}

[AlexeyYar]

Same here: {"type":"error","sequence_number":2,"error":{"type":"tokens","code":"rate_limit_exceeded","message":"Rate limit reached for gpt-5 in organization org-************* on tokens per min (TPM): Limit 30000, Used 14582, Requested 21045. Please try again in 11.254s. Visit https://platform.openai.com/account/rate-limits to learn more.","param":null}}

[WilliamAGH]

Yeah, Cerebras and Groq are currently unusable with Opencode right now still because of this.

Constant errors like:

AI_RetryError: Failed after 4 attempts. Last error: Too Many Requests

[JC1738]

Same, AI_RetryError: Failed after 4 attempts. Last error: Requests per second limit exceeded - too many requests sent.

cerebras with Qwen 3 Coder 480B

I can't get through the /init without the being blocked by this error, and unfortunately you can't resume the init.

[jsugg-qventus]

Same here. On each message.

[sra]

After these rate limit stops, what do you type to make it continue? I've not had great luck.. Cerebras is awesome until the brick wall per minute limit

[rekram1-node]

We have backoff for ratelimits, and we respect the retry-after type headers so once it is ready to retry it should automatically go through