Struct MaxTokensRecovery 

pub struct MaxTokensRecovery {
    pub max_attempts: u8,
    pub scaling: TokenScaling,
    pub ceiling: Option<u32>,
}

Recovery strategy when the provider returns StopReason::MaxTokens.

On hit, the loop discards the truncated assistant turn (the MessageStart/MessageEnd events for it still fired — listeners correlate via the new AgentEvent::OutputTokensEscalation) and re-streams with a higher cap.

Bounded by max_attempts per turn. Hits the ceiling if set. Fixed ladders run out by definition once attempts >= caps.len().

Fields

max_attempts: u8

Hard upper bound on retries within a single turn. The loop emits at most max_attempts escalation events per turn; the attempts + 1th call simply uses the ladder’s last cap and the result is accepted regardless.

scaling: TokenScaling

How to derive the next cap from the previous one.

ceiling: Option<u32>

Hard upper bound on the cap itself. None means no ceiling (relies on max_attempts to bound the spend). Recovery stops short when the next computed cap would equal or fall below the previous one (no progress).

Implementations

impl MaxTokensRecovery

pub fn doubling() -> Self

Default: 3 retries with doubling, no ceiling. Meant as the “least-config option” for callers who just want the recovery without tuning. Real deployments usually pin a ceiling matching their model’s hard max.

pub fn next_cap(&self, prev_cap: u32, attempt_zero_indexed: u8) -> Option<u32>

Compute the cap for retry attempt attempt_zero_indexed (0 = the first retry, after the original turn). Returns None when the ladder cannot make further progress (Fixed exhausted, ceiling reached at the previous step).

Trait Implementations

impl Clone for MaxTokensRecovery

fn clone(&self) -> MaxTokensRecovery

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for MaxTokensRecovery

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.