Problem
When using high-frequency LLM calls (e.g., game demos that call llm:choose every tick), rate limit errors from providers cause immediate failures. The extension throws a RuntimeException with no retry logic.
Example error from GPT-5.4-mini:
Rate limit reached for gpt-5.4-mini in organization ... on tokens per min (TPM): Limit 200000, Used 192306, Requested 11670. Please try again in 1.192s.
Current Behavior
BaseHttpProvider.sendChatRequest throws immediately on any HTTP error response, including rate limits.
Proposed Fix
Add retry with exponential backoff in BaseHttpProvider.sendChatRequest:
- Detect rate limit responses (HTTP 429 or error body containing
rate_limit)
- Retry up to 3 times with exponential backoff (1s, 2s, 4s)
- Parse
Retry-After header if present and use that as the wait time
- Only retry on rate limit errors — other errors should still fail immediately
Affected File
src/main/providers/BaseHttpProvider.scala — sendChatRequest method (line ~103)
Workaround
Demos can handle this at the NetLogo level with wait + stop in the error handler:
carefully [
let action llm:choose obs choices
...
] [
if member? "rate_limit" error-message [
wait 2
stop
]
]Context
Discovered while testing the Baba Is AI demo with gpt-5.4-mini. History accumulation + frequent calls hits TPM limits quickly.
Problem
When using high-frequency LLM calls (e.g., game demos that call
llm:chooseevery tick), rate limit errors from providers cause immediate failures. The extension throws aRuntimeExceptionwith no retry logic.Example error from GPT-5.4-mini:
Current Behavior
BaseHttpProvider.sendChatRequestthrows immediately on any HTTP error response, including rate limits.Proposed Fix
Add retry with exponential backoff in
BaseHttpProvider.sendChatRequest:rate_limit)Retry-Afterheader if present and use that as the wait timeAffected File
src/main/providers/BaseHttpProvider.scala—sendChatRequestmethod (line ~103)Workaround
Demos can handle this at the NetLogo level with
wait+stopin the error handler:Context
Discovered while testing the Baba Is AI demo with
gpt-5.4-mini. History accumulation + frequent calls hits TPM limits quickly.