Date: 2026-08-03
Deployed a root-cause fix for AI personas dwelling in maintenance loops when the human player uses receptive, minimal, nonverbal, slow-paced, or Proceed input.
- Persona decision context now treats learned behavior as presentation and response guidance, not permission to withhold AI initiative.
- Legacy unscoped observations, player-control metrics, narration-only preferences, and derived boundary-conflict notes are excluded from persona decision context.
- New behavior observations require typed guidance scope.
- Actor-private agendas retain a bounded history of completed milestones and reset lifecycle metadata for a genuinely new objective.
- The existing private action review now returns typed agenda progress. Maintenance or repetition is removed before shared resolution; unavailable review preserves safe Say/Act but does not advance agenda state.
- Agenda state remains private, conflict-gated, and atomically committed.
- Feature commit:
d9495a2f
- Main merge:
b8c05fc5636e0f240c5a1ac705a8e3c92e5dea26
- Image:
gitea.app.lan/eventhorizon/eventhorizon:sha-b8c05fc5636e
cargo fmt --all -- --check- Workspace tests: 403 backend, 11 frontend, doc tests successful
- Requirements JSONL parsed successfully
- Diff checks clean
- Gitea frontend workflow successful in 2m25s
- Gitea build-and-deploy workflow successful in 7m55s
- Deployment rollout completed successfully
- 2/2 replicas ready and updated
- Zero pod restarts after rollout
/healthz: status=ok/readyz: PostgreSQL and Qdrant ready- Affected game session remained unchanged at turn 45 with safety state
running; no synthetic roleplay turn was generated.