Vakyartha documentation

1//! vak-context: the context engine (docs/design/68-context-engine.md).
2//!
3//! Everything that decides *what one request carries* for the model in
4//! front of it, as pure functions over the session ledger:
5//!
6//! - [`capacity`] — `CapacityProfile`: the measured capacity of one model on
7//! one provider (window, usable instruction horizon, tokens per char,
8//! prefill rate, cache behaviour), probed at bind time and revised from
9//! every receipt. Nothing in a request is sized from a declared number.
10//! - [`planner`] — `WorkingSetPlanner`: which closed turns ride along at
11//! `Full`, `Card` or `Packet` fidelity for this request, chosen against
12//! the measured budget by `max(recency, relevance, anaphora)`. A turn is
13//! never split; nothing model-visible is cut by a character count.
14//! - [`assemble`] — the request assembler: the byte-stable prefix and its
15//! digest, the per-turn tail attached to the last user message, cache
16//! breakpoints per §10, char accounting that feeds the profile, and the
17//! summariser request behind a compaction packet.
18//!
19//! The ledger itself — `TurnIndex`, `TurnCard`, `derive_with_plan`, the
20//! range-keyed packets and `recall` — lives in `vak-session`: that is the
21//! one rich original every projection is computed from, and this crate
22//! only reads it. The agent loop (`vak-agent`) and the host (`vak-core`)
23//! own the I/O: probing, the summariser call, and writing what came back.
24//! Nothing here performs network or disk I/O.
25
26pub mod assemble;
27pub mod capacity;
28pub mod planner;
29
30pub use assemble::TailInput;
31pub use capacity::{CapacityProfile, ProfileKey};
32pub use planner::{PlanInput, plan, plan_for_session};
33pub use vak_session::{Fidelity, WorkingSetPlan};
34