Configure a degradation ladder
Opt into bounded retries and careful context trimming before ordered fallback.
On this page
Keep the default or opt in
Models without a configured ladder keep the existing router behavior. In Edit chain, you can enable a bounded wait-and-retry step and context trimming before falling through the ordered model chain. Every action is recorded on the receipt.
Wait and retry
A bounded retry uses the same credential and respects reported Retry-After, quota reset and circuit cooldown evidence. It does not cycle account keys to evade a provider rate limit.
Trim only after an explicit context error
Trimming happens after an explicit context-limit rejection. It removes complete older user turns, retaining system and developer instructions, the current user turn and matched tool pairs. It is off by default.
Token saver is a separate control
The token saver compresses supported repetitive tool output before forwarding. It preserves IDs, JSON lexemes and source code. The console reports UTF-8 bytes removed; a measured 20–40% tokenizer saving has not been established.