A trigger's notify_url is refused at every address a worker's endpoint is, read from Reach's one list, and anywhere in 100.64.0.0/10, by literal or by a name that resolves there, and is posted to the address checked; every IPv6 spelling of a v4 address is judged as it¶
Decision 866 · accepted · 2026-10-11 · issue #588
- One list.
Troupe.Plane.Triggers.Notifyno longer keeps its own clauses for loopback, the unspecified address and link-local. It asksTroupe.WorkerProfile.Reach.refusal/1, now public, which reads the list a worker's endpoints are refused from (@refused, Decisions 758 and 856) and takes an address as:inetgives one as well as a host as written. The list stays in the protocol app, where the operator and the plane can both read it; the next metadata address is added there once and refused in both. - The plane's own: multicast and the whole shared address space. Multicast stays the
plane's alone, as before.
100.64.0.0/10is refused whole,is not in the shared address space, which is not what 856 chose for a worker: there an address in it is admitted when given as that address, because a hosted object store or an installation's gateway may answer there. The plane differs because nothing holds its egress and because anotify_urlis posted to on a name's word at send. A webhook receiver at an address in that range (a CGNAT overlay or a tailnet) is refused, and is given at a private or a public address instead. The private ranges stay open to it, as they were; the egress allowlist governs those. - Posted to the address checked. The request now goes to the first address the check
passed, a v4 one when there is one, as the connection would have chosen: the URL names
that address and the connection asks DNS nothing. The name goes as the
Hostheader, and as Mint's:hostname, which is the TLS server name and the name the certificate is checked against. A literal in an older spelling (1684301000) is written as the address it was read as. - Every spelling. The check parses an address as a connection would, so the
spellings 856's tests use (
FD00:EC2:0:0::254,::ffff:100.100.100.200) and older v4 ones (1684301000,0144.0144.0144.0310) are refused as the address they name. At send every address the name answers is checked, so one bad answer among good ones refuses it. - IPv6 spellings of a v4 address, for workers too.
Reach.ipv4/1reads the v4 address in v4-mapped (::ffff:a.b.c.d), IPv4-compatible (::a.b.c.d, but::and::1, which are their own), NAT64's well-known prefix (64:ff9b::a.b.c.d) and 6to4 (2002:aabb:ccdd::/48), andrefusal/1judges that address. A profile's endpoint and anotify_urlwritten so are refused as the address is. None of those prefixes overlaps a refused IPv6 block, so nothing is lost by judging the v4 address alone. A rule that admits an endpoint is still written for the address as given: only a v4-mapped one is the v4 address on the wire, and the others are reached as IPv6, through a translator or a relay. - The Cilium deny takes NAT64's spelling, not the others. 856 denies by address
because a
toFQDNsrule admits a name at whatever address its DNS answer carries. In a cluster with NAT64, a DNS64 resolver answers a name whose only address is169.254.169.254with64:ff9b::a9fe:a9fe, and a translator carries the pod's packets on to the v4 address. RFC 6052 has a translator drop a well-known-prefix address naming a non-global one, which these all are, but nothing here can know the translator does. So the deny adds the v4 blocks in that prefix,64:ff9b::a9fe:0/112and64:ff9b::6464:64c8/128, computed from the same list. 6to4 and IPv4-compatible addresses are not added: no DNS64 answers with them, a pod has no route to a 6to4 relay (RFC 7526 retired them) and none for IPv4-compatible addresses, so a name leading there reaches nothing. A network-specific NAT64 prefix (64:ff9b:1::/48or a provider's own) is not known here and is not covered.
Governs: apps/troupe_plane/lib/troupe/plane/triggers/notify.ex, apps/troupe_protocol/lib/troupe/worker_profile/reach.ex