Core developers
Mesh discovery and placement
Finding a Core, finding an agent and asking it to do work each require different evidence. Start with the identity you need and the scope the destination owner approved.
The scoped direct-turn exchange was exercised across two machines with a fixture provider. Discovery and a graph-read grant have separate purposes.
Find a machine or a teammate
Open cleetus-host mesh console and enter account login, then account hosts. The account directory returns registered Host advertisements, which can be stale or offline. For a direct paired connection, select connect HOST_ID, check the destination's certificate fingerprint, enter its one-use invitation and compare the code at both ends. The destination's grant determines which reads are available.
Inside a reviewed local Codex team, the tool call cleetus_list_peers {} returns the current recipients that member may contact. A developer can see a reviewer when the roster permits that direction and the peer has a live reviewed launch. Messages carry an original operation ID. An identical repeat can retrieve the receipt while the sender remains authorized.
Inspect a remote Core
Two managed Cores can separately review an execution-graph route. Both prove control of their enrolled device keys over TLS. The gateway can then read the exact exporting Host and instance using the identifiers shown during review:
cleetus-host execution-graph direct read "$HOST_ID" "$INSTANCE_ID"
For a native client, the gateway exposes POST /v1/inspection/routed-execution-graph on its authenticated private socket. The v1 request names both the gateway and the exporting Core by Host and instance ID. A graph grant permits a bounded read. Placement, provider credentials and team membership have separate owners. Let displayed evidence expire when the route closes or its observation expires.
Place one turn
The direct-turn protocol has a narrow integration path for one Codex turn on an existing destination session. The destination reviewer names the source device and instance, workspace, conversation, model, policy and capacity. The source records the selected destination. The destination checks the offer, prepares the exact request, then rechecks authority at commit. A cross-machine Mac-to-Linux fixture exercised this exchange with a synthetic provider and recovered the original receipt after restart.
The proposed cross-host scheduler would filter eligible offers, score them, reserve capacity at a destination and bind the run to that reservation. Each participating machine would contribute explicit capacity for particular projects and providers. A heartbeat expiring would remove that machine from new placements. Lost replies would still require original receipt lookup before releasing a hold. Production cross-host scheduling remains a design target. Task readiness · Agent teams