Operate across hives

Start with the same-version fleet recipe, host tools and authority contract. Select the actual named hives and a matching macOS or Linux client installation. Either host can read the selected hives; their QEMU, Pi and native provider profiles still determine which actions are available. Use coh fleet --help and read-only status, lease-summary and pressure before any mutation. A missing hive, stale timestamp or failed read is unknown, not healthy or idle. If this skill is copied outside the repository, use the matching installed guides or pin those links to the selected source revision; load companion skills separately. Check the client version and each hive’s selected profile separately. If one host or provider is incompatible, keep that row unknown and offer the matching host bundle or a supported provider configuration for that hive; do not copy another hive’s authority or capability into it.

Produce a useful fleet view

For each hive, record target image/profile and connection, observation time and freshness, Worker readiness, scheduler pressure, active leases, last relevant job and strongest blocker. Keep QEMU, physical Pi and retained replay observations in their own evidence classes. Compare pressure using each hive’s published bounds; do not add quotas as if they were one global pool. Host-side fan-in may improve visibility, but it does not transfer authority from one Queen to another.

If the user asks for a cross-hive action, name the source and destination, which hive admits each step, the exact delegated subject/ticket, limit, expiry and fallback. Preview the route and check each hive’s current state. Submit only the specifically authorised leg. Preserve source and destination IDs and receipts separately; a relay ACK cannot certify a downstream native effect or exactly-once execution. After a disconnect, inspect the original identity at both ends before considering recovery. Do not use a broad shared credential to fill a missing local grant.

Return a fleet table with one row per hive, source and timestamp, pressure, readiness, active work and uncertainty. For an action, add the route, both identities, local admission and native result, independently verified outcome and unresolved leg. Use the evidence skill to compare before/after packs or create a federation incident case. Factory, logistics and other sector applications still require their own live provider and safety evidence; playbook names alone are design patterns.