Model Retry [model_retry]

Complete config.toml using the defaults:

[model_retry]
standard_retries = 3
overload_retries = 5
backoff_ceiling_ms = 60000

An omitted table inherits; an authored table replaces it completely, with defaults for omitted fields. Unknown fields fail. See config.toml for layer precedence.

Controls retries on the current model before the engine advances through its fallback chain. Values count retries after the initial attempt. Zero disables retries for that class, while any negative value retries indefinitely until the request succeeds or the run is cancelled.

Key	Type	Default	Description
standard_retries	signed 64-bit integer	3	Retry budget for ordinary retryable model errors: one initial attempt plus three retries gives up to four total attempts by default.
overload_retries	signed 64-bit integer	5	Separate retry budget for overload errors and retryable HTTP 429/503 responses.
backoff_ceiling_ms	integer	60000	Positive local ceiling, in milliseconds, for exponential backoff after jitter.

Retries use exponential backoff from one second with 25% jitter, clamped to backoff_ceiling_ms and a positive minimum of one millisecond. A provider Retry-After is authoritative when longer and may exceed that local ceiling without a cap. Cancellation still interrupts the wait.

The retry budget applies no matter how far an attempt got: an ordinary retryable failure that arrives mid-stream, after partial output was already produced, still consumes standard_retries on the same model before the loop falls back. The abandoned attempt's partial deltas are discarded (the request is rebuilt from committed history), so provider prompt caches remain warm for the retry.