Specialist agent operations
All 28 contracts become bounded model-backed jobs.
The deployed runtime imports the recovered specialist registry, snapshots allowlisted input packets, queues tasks in SQLite and executes authenticated model calls through a bounded worker pool. Results retain validated structured findings and private drafts. Separate reviewers examine the batch without self-approval. A timer detects meaningful source changes; retries and stale leases are bounded.
Verification: Runtime boundary tests and real model executions are recorded independently. The live panel shows actual task states and review outcomes. A completed job means the model execution finished; its verdict can still require work or human judgment.
Remaining: The agents do not autonomously mutate game saves, award resources, contact players, publish findings, or alter DNS. They provide evidence and drafts within their role scopes. Long-term host recovery and measured agent quality require continued operation.