v0.5.10 — Retry, Don’t Reject
v0.5.9 spaced out the cognition requests so they wouldn’t stampede. But when the LLM server did hit a rate limit, it was still rejecting the request outright — a dropped thought, gone forever. The agent would silently skip its decision cycle and stand there, briefly lobotomized.
Now the server retries. A 429 isn’t a rejection, it’s a queue. The request backs off, waits, tries again. Combined with the lowered client concurrency cap from this patch, the startup burst that used to hammer the endpoint is gone entirely. Thirty agents initializing at once no longer means thirty simultaneous LLM calls.