shared-filesystem-agent-coordinationlisted

# Shared Filesystem Agent Coordination Use shared storage as a message transport, not as a concurrently edited chat document. The default topology is a star: - one coordinator Agent owns global decisions, shared artifacts, and requests; - each node runs a campaign-scoped token-free dispatcher that launches one bounded Agent invocation per request; - workers execute bounded local requests and return structured evidence; - worker-to-worker needs return to the coordinator for routing. This creates command-and-report collaboration, not unrestricted peer chat. Keep durable control artifacts separate by purpose. The protected task contract is immutable; campaign identity is stable across retries; the active attempt and fencing epoch are dynamic; each request and result is immutable and schema-versioned. Existing project runbooks and progress documents may provide human context, but they are not mandatory bus state. Prompts should name the canonical contract, authority, and requested action instead of copying stale attempt history or re-encoding the protocol in prose. ## Assign control ownership Use Goal mode only for the coordinator's bounded objective. Worker execution is stateless by default: a dispatcher launches a fresh ephemeral Agent for one request, with shared records and repository handoffs as its complete context. This removes TUI, thread, transcript, and conversation-lifetime dependencies. Use exact-thread resume only as an explicit compatibility adapter when pri