Skip to content

Without Cilium, a profile's own endpoints are admitted as the installation's are: a name on another port as the public rule's addresses on that port, an address as that one address on its port. One at a loopback or link-local address is admitted by nothing, and is what is still refused

Decision 752 · accepted · 2026-10-02 · issue #268

749 refused an LLM gateway on 8443, or an MCP server on the office network, where the profile was set up, so an installation without Cilium that ran its gateway in-house could not use it. - What is admitted. The worker NetworkPolicy gives each of a profile's endpoints the public rule does not reach a rule of its own, as 724 does OpenBao and the object store: the LLM endpoint, each MCP server (its own, and its channel's bundle's through the mcpServers the plane projects) and each egress.fqdns entry, whose port is written host:port. An entry with none is reached where a name is, on 443 and 80, an address as itself on those. An IPv6 address, which 749 refused because the public rule has no v6 block, is its /128. A public address on 443 gets its own rule too, as 724's do; a name on 443 and 80 needs nothing and is unchanged. No CRD change. - The cost is 724's. A name on 8443 opens 8443 to every public address for that profile's workers, which is as much as a NetworkPolicy can say about a name. A name that resolves to a private address is still not reached and still not said, since what it resolves to is not known where it is typed; the docs say to give it as its address or use Cilium. - Loopback and link-local stay refused, 127.0.0.0/8, ::1, 169.254.0.0/16 and fe80::/10, and a v4 address in v6 spelling (::ffff:169.254.169.254) is the v4 address it names, where the connection goes. From a pod, loopback is the pod itself, and link-local is the node's, where a cloud's metadata service answers: a rule would open that to the pod for one profile's endpoint. These are what admin.profile.put, the editor, a bundle's publish and EndpointUnreachable name now, and the message says why and to give the endpoint as a pod reaches it. - One judgement. Troupe.WorkerProfile.Reach judges each endpoint once, into the rules that admit it (admitted/1, through admission/2, which the operator also uses for the platform's endpoints) or the sentence that refuses it (unreachable/1), and the public rule's ports and ranges are read from it. So the NetworkPolicy, the refusal and the condition cannot disagree about an endpoint. - The git hosts and an MCP identity's token endpoint are still not judged, as #268 does not name them. With Cilium nothing changes.

Governs: apps/troupe_operator/lib/troupe/operator/reconciler.ex, apps/troupe_operator/lib/troupe/operator/resources.ex, apps/troupe_plane/lib/troupe/plane/reach.ex, apps/troupe_protocol/lib/troupe/worker_profile/reach.ex