Repository navigation
v3.4.1
One fix, but a subtle one about who owns the retry budget.
#987 taught the model transport to stop retrying permanently-failed requests (403 exits immediately instead of walking the whole ladder) — but the tier layer borrowed the transport's own retry allowlist as its definition of permanence. Any status the transport doesn't itself retry fell through, including transient gateway errors like Cloudflare's 524, which meant a configured ai-primary-retries budget was silently ignored on exactly the errors a retry budget exists for.
- The tier owns its permanence rule (#1007, closes #1006): only
400/401/403/404get exactly one tier attempt; everything else —429,5xxincluding524, network and timeout errors — consumes the configured tier retry budget before fallback.501and507are pinned tier-retryable, and tests assert524stays out of the transport's own set so the two policies can't silently re-merge. - Unchanged:
429/500behavior, and thenon-retryableannotation still fires only on the four permanent statuses.
Full changelog: v3.4.0...v3.4.1