Struct TokenBudget 

pub struct TokenBudget {

    pub max_context_tokens: u32,
    pub max_output_tokens: u32,
    pub strategy: BudgetStrategy,
    pub safety_margin: u32,
    pub compression_trigger_percent: u8,
    pub compression_target_percent: u8,
    pub working_reserve_tokens: u32,
    pub fallback_trigger_percent: u8,
    pub prompt_cache_min_tool_output_chars: u32,
    pub prompt_cache_head_chars: u32,
    pub prompt_cache_tail_chars: u32,
    pub prompt_cache_recent_user_turns: u8,
    pub prompt_cache_recent_tool_chains: u8,
    pub max_tool_output_tokens: u32,
}

Token budget configuration for a conversation.

Fields

max_context_tokens: u32

Maximum context window size for the model (input + output)

max_output_tokens: u32

Maximum tokens reserved for model output

strategy: BudgetStrategy

Budget enforcement strategy

safety_margin: u32

Safety margin for tokenizer estimation errors

compression_trigger_percent: u8

Proactive compression trigger threshold as a percentage of context window tokens.

Legacy field — only used when working_reserve_tokens == 0.

compression_target_percent: u8

Compression target threshold as a percentage of context window tokens.

working_reserve_tokens: u32

Fixed number of tokens reserved for model reasoning and output.

Compression triggers when (max_context_tokens - working_reserve_tokens) is exceeded. This provides consistent reasoning space regardless of model context window size (Claude Code uses ~50K).

When working_reserve_tokens == 0, falls back to percentage-based triggering via compression_trigger_percent for backward compatibility.

fallback_trigger_percent: u8

Fallback percentage trigger when the context window is too small for the fixed reserve (i.e. max_context_tokens < working_reserve_tokens * 2).

prompt_cache_min_tool_output_chars: u32

Minimum tool output character length required before prompt-side cache compaction.

prompt_cache_head_chars: u32

Leading excerpt length (chars) kept in cached tool output summaries.

prompt_cache_tail_chars: u32

Trailing excerpt length (chars) kept in cached tool output summaries.

prompt_cache_recent_user_turns: u8

Number of latest user turns protected from prompt-side cache compaction.

prompt_cache_recent_tool_chains: u8

Number of latest tool call chains protected from prompt-side cache compaction.

max_tool_output_tokens: u32

Maximum tokens allowed per single tool output. 0 = use byte-based fallback only.

Implementations

impl TokenBudget

pub fn new( max_context_tokens: u32, max_output_tokens: u32, strategy: BudgetStrategy, ) -> TokenBudget

pub fn with_safety_margin( max_context_tokens: u32, max_output_tokens: u32, strategy: BudgetStrategy, safety_margin: u32, ) -> TokenBudget

pub fn compression_trigger_context_tokens(&self) -> u32

pub fn compression_target_context_tokens(&self) -> u32

pub fn for_model(max_context_tokens: u32) -> TokenBudget

Trait Implementations

impl Clone for TokenBudget

fn clone(&self) -> TokenBudget

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for TokenBudget

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

Formats the value using the given formatter.

impl Default for TokenBudget

fn default() -> TokenBudget

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for TokenBudget

fn deserialize<__D>( __deserializer: __D, ) -> Result<TokenBudget, <__D as Deserializer<'de>>::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for TokenBudget

fn serialize<__S>( &self, __serializer: __S, ) -> Result<<__S as Serializer>::Ok, <__S as Serializer>::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.