Struct ContextPartitions 

pub struct ContextPartitions {
    pub system: Partition,
    pub knowledge: Partition,
    pub task_state: TaskState,
    pub signals: Vec<String>,
    pub history: Partition,
}

Three-partition context model aligned with LLM API slots:

Slot 1 — Identity (system): who the agent is; role, rules, constraints. Maps to: Anthropic system[0] cache_control, OpenAI system role. Never changes within a run.

Slot 2 — Knowledge (knowledge): what the agent knows; memory retrievals, skill definitions, artifacts. Low-frequency changes. Maps to: Anthropic system[1] cache_control.

Slot 3 — State (task_state + signals): what the agent is doing right now. task_state = goal/plan/progress (structured). signals = runtime events (rollback notes, interrupts). Maps to: messages[0] user turn, rebuilt every call.

Slot 4 — History (history): what the agent has done; conversation turns, tool calls and results. Compression pipeline target. Maps to: messages[1..N].

Fields

system: Partition

knowledge: Partition

task_state: TaskState

signals: Vec<String>

Runtime signals injected into the current turn (rollback notes, interrupts). Cleared after each render — signals are ephemeral per-turn events.

history: Partition

Implementations

impl ContextPartitions

pub fn new(_config: &ContextConfig) -> Self

pub fn total_tokens(&self, engine: &ContextTokenEngine) -> u32

Total token count across all slots. task_state tokens are measured from its rendered compact form.

Trait Implementations

impl Default for ContextPartitions

fn default() -> Self

Returns the “default value” for a type.