Chapter 9 · agent-neuron · Lab
Effects
A side effect a newer input never cancels: both messages go out, a crash is recorded as unknown, and a source reconciles it.
One side effect, three updates
A build monitor sends a message about a build, and the neuron that sends it is declared an effect. Three updates arrive for one key. The second replaces the first while its message is already being sent: nothing is cut, both messages go out, the first answer is recorded but not forwarded. The third loses its connection after sending, so its call is recorded as unknown, and reconcile, the source the network names for exactly that, is told where to look. Then the same network runs with the effect removed, and the first message is cut and never sent.
The rule it proves
An effect is declared, not inferred (SPEC 11): a started call is never cut, its answer is recorded with forwarded: false when the input moved on, and a crash, hang or timeout is unknown rather than an error, because nobody can tell whether it happened (11.4). What really happened reaches the network only through a source that is told, and every event streams as it happens (9.1).
Run it
npm run demo:notifyWhat it is not
The engine reconciles nothing on its own and keeps no retry policy: something outside declares the source and injects the clue, and an effect inside a composed neuron is simply refused (11.5).
