Two real bugs found while live-testing the new GET /api/mesh/self
endpoint with two actual jump-host containers (mesh-joined for real,
not mocked):
1. routes/api.js mounted `/` (routes/jump.js, admin-session-gated)
before `/mesh`. Since router.use('/', ...) matches every /api/*
path, EVERY /api/mesh/* request -- including /register, which is
authenticated by a bearer mesh join token, not an admin session --
hit that admin gate first and 401'd before routes/mesh.js ever ran.
Confirmed live: a real gateway-to-gateway /join call failed with a
checkApiToken/LoginFailed error instead of ever reaching /register.
Reordered so /mesh is mounted first.
2. POST /register (the receiving side of a join) persists a `(self)`
registry entry via ensureOwnMeshIndex(), but POST /join (the
initiating side) never did -- so GET /api/mesh/self and the mesh
UI's own-entry handling silently saw nothing on whichever gateway
called /join. Fixed by registering a self-entry there too, using
the exact meshIndex the remote assigned (models/mesh_gateway.js's
register() now accepts an explicit meshIndex instead of always
auto-picking one from the local registry, which has no reason to
agree with what's actually configured on the live wg0 interface).
Verified with two real containers joined over a live network: both
sides now report their own correct mesh IP via GET /api/mesh/self,
and both appear correctly in GET /api/mesh/gateways.
wg_iface.removePeer() previously just did `wg set ... remove` -- the
kernel routes setPeer() adds for a peer's AllowedIPs (since wg itself
only configures crypto-routing, not kernel routes -- see setPeer's own
comment) were never cleaned up, a real TODO flagged in code but never
exercised because nothing removed a mesh peer at all.
- removePeer() now queries the peer's current AllowedIPs (`wg show
<iface> allowed-ips`) BEFORE removing it -- once gone, wg no longer
knows what to clean up -- and issues `ip route del` for each.
- New DELETE /api/mesh/gateways/:id (models/mesh_gateway.js gained
remove()) actually calls removePeer(), so the fix has a real caller;
previously there was no removal path anywhere in the mesh feature at
all. Refuses to remove the local "(self)" entry. Does not reach out
to the remote gateway to remove the reciprocal peer -- that side
needs the same action taken independently.
- Mesh UI: remove button per non-self peer row, using app.messages.confirm
(not native confirm() -- caught by this repo's own no-native-dialogs
test, which failed on first pass and is now green).
Verified for real with a live WireGuard interface in a container: routes
for a peer's AllowedIPs present after setPeer, confirmed gone after
removePeer, while the interface's own local route correctly survives.
The existing WireGuard code (models/wg_site.js, routes/wireguard.js) is the
roaming-client/exit-node feature -- individual peer configs an admin hands
out, not gateway-to-gateway mesh peering. This adds the latter, per
MULTI_SITE_SPEC.md §4: two theta-gateway instances mesh by one calling the
other's POST /api/mesh/register with a join token (minted via
POST /api/mesh/join-tokens, admin-gated); both sides end up with a live
wg0 peer for the other, mesh-indexed per Appendix A's addressing
(172.24.<idx>.0/16 + 10.<idx>.0.0/16, idx 1-254).
- utils/wg_iface.js: brings up the local interface, preferring in-kernel
WireGuard (ip link add type wireguard) and falling back to userspace
wireguard-go when the kernel module isn't available. Both packages
added to the Dockerfile.
- utils/mesh_addressing.js: pure addressing math, unit tested
(test/unit/mesh_addressing.test.js).
- models/mesh_gateway.js: Redis-backed registry of known peer gateways
(same pattern as wg_site.js), assigns + persists mesh indexes.
- utils/mesh_join_token.js: single-use bootstrap credential, same
GETDEL-on-Redis pattern already used on the theta-directory side.
- routes/mesh.js: /join-tokens (admin), /register (bearer token, no
session -- called by a remote gateway), /join (admin, initiates from
this side), /gateways (admin, list).
Verified with a REAL two-container test (not mocked): two independent
containers, each running this actual code, meshed via a live join-token
handshake, brought up real kernel WireGuard interfaces, and passed ICMP
traffic across the resulting encrypted tunnel end to end (0% packet
loss). That test caught a real bug worth calling out: `wg set ... peer
... allowed-ips` only configures WireGuard's own crypto-routing table --
it does NOT add a kernel route for that destination (wg-quick normally
does this as a separate step; we don't use wg-quick). A real encrypted
handshake completed between the two containers with the route missing,
and ping still showed 100% loss until setPeer() was fixed to add the
corresponding `ip route add <allowed-ip> dev <iface>` itself.
wmantly