Module eval 

The evaluation cycle: find due, claim, invoke, ingest, start, release.

One invocation performs the whole cycle and exits. Nothing stays resident. Losing a claim race is a no-op exit, not an error — that is the steady state when several nodes share a heartbeat.

Why correctness does not rest on the lease

A lease deduplicates “who is working right now”. But the thing that actually needs deduplicating — a firing of trigger T for item I — has a natural identity, so the run id is derived from it. areev run start already refuses a duplicate run id, so two evaluators racing produce one run and one recorded skip, with no lease consulted at all: no duration to guess, no fencing token to compare, no clock-skew window.

The lease then does one narrower job: stopping two nodes from both making the same expensive, side-effecting connector call. Losing it can cost a wasted API call; it can never cause a missed or duplicated firing. That is the risk profile a lease should carry.

Structs

EvalOptions

Knobs for one evaluation pass.

EvalReport

What one evaluation pass did. Shaped like loop run’s report for symmetry.

Evaluator

Evaluates triggers against one memory.

TriggerStatus

One trigger considered, for --dry-run and status.

Enums

StartResult

What starting a run did.

Constants

TRIGGER_NS

Namespace for trigger firing records. Alongside the runtime’s own agent:harness, so operational journalling stays out of user namespaces.

Traits

RunStarter

Starts the workflow a trigger is bound to.

Functions

run_id_for

The run id for one firing of one item.

trigger_name

The declaration’s human label, if it has one.