Crate areev_run_core 

areev-run-core — the pure scheduler of areev run

Sans-IO by construction (governed-agents proposal §4): this crate cannot observe the world. It exports step() — a pure function from (plan, env, state, events) to (commands, state) — and the vocabulary around it. The driver crate (areev-run) owns the store, the clock, the executors, and the journal; it materializes journal lookups into EventIns and performs Commands. Replay IS this function: feed the journaled events back in and assert the same commands come out.

Purity is enforced mechanically, not aspirationally: the dependency tree carries no clock/rand/net crate (CI checks cargo tree), and this crate’s clippy.toml disallows SystemTime::now/Instant::now. Scheduler state uses Vec/BTreeMap exclusively — hash-iteration order is the classic silent replay-divergence source and has no business here.

Re-exports

pub use error::BudgetAxis;

pub use error::Result;

pub use error::RunError;

pub use plan::PlanEdge;

pub use plan::PlanGraph;

pub use state::DecideInFlight;

pub use state::EdgeRes;

pub use state::FoldInFlight;

pub use state::NodeState;

pub use state::PendingAsk;

pub use state::Phase;

pub use state::SchedulerState;

pub use state::Spent;

pub use step::bound_tool_content;

pub use step::flow_key;

pub use step::step;

pub use step::DecideEnv;

pub use step::StepEnv;

pub use step::StepOutcome;

pub use step::DEFAULT_MAX_EFFECTS_PER_ATTEMPT;

pub use types::Ask;

pub use types::Budgets;

pub use types::Command;

pub use types::DecisionRecord;

pub use types::EdgeOutcome;

pub use types::EffectKind;

pub use types::EffectOutcome;

pub use types::EventIn;

pub use types::DynamicDispatch;

pub use types::DynamicTool;

pub use types::FailCause;

pub use types::FoldRecord;

pub use types::JournalKey;

pub use types::LlmNodeConfig;

pub use types::NodeExecutor;

pub use types::OfferedTool;

pub use types::RunOutcome;

pub use types::DECIDE_TOOL;

pub use types::DECIDE_URI;

pub use types::INBOX;

pub use types::PARKED_ASKS;

Modules

cond

The frozen v1 condition grammar (governed-agents §6.4):

error

The RUN error domain (governed-agents §6.9). Codes are append-only from day one, same rule as every other domain (ERROR_CODES.md): never renumber, never reuse. The code is the leading token of Display, and code() returns it — both pinned by tests below and by the workspace uniqueness/format tests.

plan

The validated plan graph — load-time validation V1–V6’s pure parts (governed-agents §6.1) over an OMS §8.4 Workflow grain, plus the derived structure the scheduler walks: indexed nodes and edges, adjacency in declaration order, and the Tarjan condensation V5 needs.

state

SchedulerState — everything a checkpoint carries (§5.4). Rebuilt-from on resume with nothing inferred from wall-clock or environment; every collection is a Vec or BTreeMap so iteration order is content order, never hash order (§4’s determinism rule).

step

The sans-IO step function (governed-agents §4): apply the driver’s events, progress the BSP machine as far as pure reasoning allows, return the commands the driver must perform. Same plan + same env + same events in the same order ⇒ byte-identical state and commands — that is the property the DST harness and the three CI gates exercise.

types

The step-function protocol: journal keys, events in, commands out (governed-agents §5.1/§6). The driver owns every effect; this vocabulary is how the pure core describes what it wants done and learns what happened.

Constants

SCHEDULER_EPOCH

The scheduler’s replay generation (#288).