A Healthy Service Does Not Mean the Task May Continue: An OpenHands Liveness Failure and Agent Recovery Boundaries
A 200 response proves only that some endpoint still answers. It does not prove that the session is making progress, and it certainly does not prove that the current task is still eligible to continue. Reliable agent recovery has to re-establish who is authorized to recover, whether the task is in a legal executable state, whether the current execution right is still valid, and whether the upstream facts required for downstream release are actually satisfied.
CodeFlowMu is a locally run multi-agent collaboration system that uses tasks, roles, gates, reports, and approvals to organize agent work into an execution chain that can be traced, recovered, and verified. We preserved one real recovery record from 2026-08-25. It did not become a tidy “the system self-healed” story, because the record contains a more valuable discrepancy: QA's physical release timestamp precedes the two upstream terminal-report timestamps.
That does not prove the gate was bypassed. It also does not allow us to claim that the dependency was proven satisfied. It exposes a different problem: timestamp order is not causal proof.
