Trait RetryObserver 

pub trait RetryObserver: Send + Sync {
    // Required methods
    fn on_retry(&self, attempt: u32, delay: Duration, reason: &str);
    fn on_exhausted(&self, attempts: u32, elapsed: Duration, reason: &str);
}

A sink for retry activity on a request that is being re-attempted.

The reporting site — the LLM completion adapter’s retry loop — records that it is backing off without knowing where the notice lands. The agent HTTP handler implements it by pushing an AgentEvent::SystemWarning into the SSE stream it already owns, so a waiting user sees “retrying” rather than a frozen cursor; a one-shot CLI path with nothing to notify passes no observer at all, making the calls no-ops.

This mirrors CacheMetricsSink — the same optional-upward-reporting seam, so both are wired into the adapter the same way and neither couples it to a transport.

Implementations are called from inside the request path and must not block: a slow observer delays the retry it is describing.

Required Methods

fn on_retry(&self, attempt: u32, delay: Duration, reason: &str)

A retry has been scheduled. attempt counts completed attempts, so it is 1 on the first retry. delay is how long the caller is about to wait, and reason describes the failure being retried.

fn on_exhausted(&self, attempts: u32, elapsed: Duration, reason: &str)

The sequence gave up. reason describes which limit was reached — see GiveUpReason::as_str.

Implementors