ACI-070: Fleet Failure & Its Signals

Thesis

A fleet fails in ways a single agent cannot, and each failure has a tell that can be read on the board. Read heartbeats against the clock and against each other, with the hypervisor as the exception; read inbox counts, claims and resume items against what the tree says. Some failures announce themselves and others wait for a reader, so turn each incident into a rule with one durable home, and retire the rule once a mechanism replaces it.

Story

On Intent 3 each pickup stamps the node's heartbeat and status, and intent wb status prints one line per node with its item and message counts. A heartbeat more than a week old marks a node's claims reclaimable, but nothing computes that: it is a reading the nodes make, and a stale hypervisor heartbeat marks nothing (Intent@9311a3fd5, 2026-09-26). The header can be wrong as well as stale. At one boot all four Claude headers in Intent named sessions that no longer existed, because pickup recorded a session id only when it was passed one, and the fix went into my close-out (Intent@8241b3c01, 2026-09-17).