Struct TokenBudgetConfig 

pub struct TokenBudgetConfig {
    pub work_budget: usize,
    pub reminder_threshold: usize,
    pub reminder_template: String,
    pub fallback_prompt: String,
    pub fallback_buffer: usize,
    pub guidance_message: String,
    pub seed_message: String,
    pub min_work_multiple: f64,
    pub min_absolute_work_room: usize,
    pub user_trail_max_tokens: usize,
}

Configuration for the token-budget window strategy.

Work-room semantics: work_budget is the conversation space ABOVE the window’s fixed base (system prompt + per-window boilerplate), not the total window size. The reset point is base + work_budget + buffer, so any positive budget produces a window that fits — the “budget below system-prompt overhead” failure class is structurally impossible. Use TokenBudgetConfig::with_work_budget to keep the reminder/buffer thresholds proportional when overriding the default.

Fields

work_budget: usize

Work-room budget: tokens of conversation space above the fixed window base. Corresponds to “how much room the model has to work with” — the base is added by the framework.

reminder_threshold: usize

Reminder threshold: inject the reminder when remaining work room ≤ this value.

reminder_template: String

Reminder message template. {n_remaining} is replaced with estimated remaining work-room tokens.

fallback_prompt: String

Fallback prompt injected when the work budget is exhausted (window about to reset).

fallback_buffer: usize

Extra work-room overshoot tolerated past work_budget before the reset fires. Allows the model one more turn to save state after the fallback.

guidance_message: String

Guidance message injected once per window start (explains the private context-management tools).

seed_message: String

User seed message injected as the LAST message of every new window. Satisfies the engine’s compaction-output contract (at least one non-System message — System maps to the top-level system parameter, and an all-System window is an HTTP 400) and gives the model a turn to reconstruct state.

min_work_multiple: f64

Minimum viable work budget as a multiple of the base overhead. Below this no single turn of work fits in the room (one tool result overflows it) and the session thrashes. Construction clamps the work budget up to max(min_work_multiple × base, min_absolute_work_room) and scales the reminder/buffer bands to match.

min_absolute_work_room: usize

Absolute floor for the work room, independent of base size. Tool results don’t shrink with the system prompt — an app with a tiny prompt still needs room for real tool output, so the viability floor is whichever is larger: the proportional bound or this constant.

user_trail_max_tokens: usize

Token cap for the mechanical user-message trail carried into a fresh window. User messages are the task trail — facts, not model output — so they are re-injected verbatim (newest-first under this cap, oldest truncated) instead of asking the model to recover them from history (cf. codex build_compacted_history).

Implementations

impl TokenBudgetConfig

pub fn with_work_budget(work_budget: usize) -> TokenBudgetConfig

Config with proportional reminder/buffer thresholds for a custom work budget (reminder = 20%, buffer = 10% of work_budget).

pub fn hard_limit(&self, base_overhead: usize) -> usize

Hard reset threshold in absolute estimated tokens: base_overhead + work_budget + fallback_buffer.

Trait Implementations

impl Clone for TokenBudgetConfig

fn clone(&self) -> TokenBudgetConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for TokenBudgetConfig

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

Formats the value using the given formatter.

impl Default for TokenBudgetConfig

fn default() -> TokenBudgetConfig

Returns the “default value” for a type.