docs: Design for agent-relayed console proxy

This commit is contained in:
2026-07-29 12:16:45 +01:00
parent ed4c39650c
commit 6d047e25ab
@@ -0,0 +1,201 @@
# Agent-relayed console proxy
Date: 2026-07-29
Status: approved, not yet implemented
## Problem
`consoleTunnel` builds guacamole parameters from `srv.IPAddress` and hands them
to guacd, which then dials the target itself. On a self-hosted deployment the
control plane and the managed servers share a network, so that works. On Vantage
Cloud they do not: guacd runs on the cloud host and the customer's server is on
an RFC1918 address behind their NAT. Every cloud console session to a private
address fails, for SSH, RDP and VNC alike.
Agents already hold an outbound gRPC connection to the control plane. The fix is
to carry the console's TCP bytes over that existing path rather than asking guacd
to route somewhere it cannot reach.
## Decisions
**Self-relay only.** The agent relays to its own host and nowhere else. It is
never told a hostname; the host is hardcoded to `127.0.0.1` on the agent side and
only the port comes from the server. A jump-host mode (reaching agentless devices
through a neighbouring agent) was rejected: it would give an agent the power to
dial arbitrary addresses on the customer's LAN, and the console today can only
target servers that run an agent anyway.
**A dedicated bidirectional RPC, one stream per TCP connection.** Multiplexing
console bytes onto the existing `CommandStream` was rejected — that stream
already carries control commands and workflow stdout, and an RDP framebuffer
would introduce head-of-line blocking against key sync and step output. A
separate stream also gets connection lifetime, flow control and close semantics
for free instead of needing a hand-rolled connection-ID demux.
**Always proxy, both deployments.** Direct dial is deleted rather than kept as a
self-hosted fast path or a fallback. One code path means one tested code path,
and the cloud path is the one no developer can reproduce locally. A
try-direct-then-fall-back design was rejected outright: it puts a timeout in
front of every private-network session and makes "which path did this session
use" unanswerable from the audit log.
The cost is that the console now requires a live agent, where a self-hosted
deployment could previously reach a server whose agent was down. In practice an
offline agent almost always means an offline host, and the failure is now an
immediate, explicit refusal instead of a hang.
## Architecture
Three parties rendezvous on a single `proxy_id`. Neither guacd nor the agent
changes which direction it dials: guacd still makes an outbound TCP connection,
the agent still only connects outbound to the control plane.
```
consoleTunnel (server)
1. proxy.Open(instance, server_id, port) -> proxy_id + ephemeral listener :N
2. push OpenProxyCmd{proxy_id, port} down the existing CommandStream
3. agent dials 127.0.0.1:port locally, then opens ProxyStream and sends
ProxyOpen{server_id, agent_token, proxy_id}
4. guacd dials PROXY_ADVERTISE_HOST:N (the params it was handed in step 1)
5. registry holds both halves -> io.Copy in both directions
6. either side EOFs -> close listener, close stream, drop the registry entry
```
Steps 3 and 4 race, so a registry entry has two slots and starts piping when the
second one arrives. Both waits share a single 10 second deadline; expiry closes
everything and frees the entry.
The agent dials locally *before* opening the stream, so a refused connection
arrives as an explicit `ProxyClose{reason}` rather than as a hang.
`BuildGuacParams` stops reading `srv.IPAddress` and takes the relay host and port
instead. `IPAddress` remains in use for display and for monitors.
## Wire protocol
Additive only; no existing message changes shape.
```protobuf
rpc ProxyStream(stream ProxyClientMsg) returns (stream ProxyServerMsg);
message OpenProxyCmd { // ServerCommand oneof field 8
string proxy_id = 1;
uint32 port = 2;
}
message ProxyClientMsg {
oneof payload {
ProxyOpen open = 1; // first message only
bytes data = 2;
ProxyClose close = 3;
}
}
message ProxyOpen { string server_id = 1; string agent_token = 2; string proxy_id = 3; }
message ProxyServerMsg { oneof payload { bytes data = 1; ProxyClose close = 2; } }
message ProxyClose { string reason = 1; }
```
## Security
**The agent only ever dials `127.0.0.1`.** The port is the only field it takes
from the server; the host is hardcoded agent-side. A compromised control plane
cannot use an agent to reach anything else on the customer's network. This is the
strongest property in the design and the reason self-relay was chosen.
**`proxy_id` is 32 random bytes, single-use and scoped.** On `ProxyOpen` the
server checks three things together: the agent token hash matches that
`server_id`, the `proxy_id` exists in the registry, and the entry's `server_id`
and `instance_id` match the authenticated agent. Any mismatch closes the stream
without revealing which check failed.
**The listener is the exposed surface and is narrowed four ways.** It binds an
ephemeral port; it lives at most 10 seconds unclaimed; it accepts exactly one
connection and closes immediately afterwards; and the accepted connection's
remote address must resolve to a host named in `GUACD_ADDR`. Without that last
check, any other container on the Docker network could claim the session during
the window.
**Agent-offline is refused early.** `consoleConnect` checks
`srv.Status == "active"` and returns 409 `agent_offline`, rather than letting the
browser open a WebSocket that dies on a deadline.
**Audit.** `console.opened` gains the relay port and `proxy_id`. A relay that
expires or is refused writes `console.proxy_failed` with a reason, so a failed
console session stops being invisible.
Credentials are unchanged. Private keys and RDP passwords travel from the server
to guacd inside the guacamole handshake and never reach the agent. The SSH and
RDP sessions are negotiated end-to-end between guacd and the target daemon, so
the agent relays bytes it cannot read.
## Components
New, server:
| Unit | Responsibility |
| --- | --- |
| `server/internal/proxy/registry.go` | `Open`, `AttachAgent`, `AttachTCP`, expiry sweep. Pure state — no net, no gRPC, testable alone |
| `server/internal/proxy/session.go` | One relay: listener, deadline, the `io.Copy` pair, teardown-once |
| `server/internal/grpc/proxystream.go` | The `ProxyStream` handler: authenticate, then hand the stream to the registry. No relay logic of its own |
New, agent:
| Unit | Responsibility |
| --- | --- |
| `agent/internal/proxy/proxy.go` | `Open(ctx, client, proxyID, port)` — dial loopback, open the stream, pump bytes. No build tags; Linux and Windows share it |
Changed:
- `proto/vantage/v1/vantage.proto`, and both generated pb trees
- `server/internal/services/console.go``BuildGuacParams(srv, relayHost, relayPort, …)`
- `server/internal/api/console.go` — offline pre-check in `consoleConnect`; open the relay before the guacd handshake in `consoleTunnel` and close it in `OnDisconnect`
- `agent/internal/sync/sync.go` — handle `OpenProxyCmd`, one goroutine per proxy
- `deploy/docker-compose.yml`, `deploy/docker-compose.site.yml``PROXY_ADVERTISE_HOST=server`
Two new optional environment variables on the server: `PROXY_ADVERTISE_HOST`
(default `server`, the name guacd resolves the control plane by) and
`PROXY_LISTEN_HOST` (default `0.0.0.0`).
Nothing new is opened on the customer's firewall — the relay rides the agent's
existing outbound gRPC connection.
`docsite/docs/reference/ports-and-networking.md` and
`docsite/docs/vantage/browser-console.md` must say so, and must state the new
requirement that the agent be online.
A secondary benefit beyond cloud: a VNC or RDP service bound only to `127.0.0.1`
is now reachable, where a direct dial from guacd never could be.
## Failure modes
| Failure | Behaviour |
| --- | --- |
| Agent offline at connect | 409 `agent_offline` from `consoleConnect`, before any WebSocket is opened |
| Agent never opens the stream | 10s deadline; listener closed; `console.proxy_failed{reason:"agent_timeout"}`; WebSocket closed with a message the UI surfaces |
| Local dial refused (daemon down, wrong port) | `ProxyClose{reason}` relayed up as the same audit event, reason `dial_refused` |
| guacd never dials | Same deadline path, reason `guacd_timeout` |
| Bad token, unknown or foreign `proxy_id` | Stream closed with no detail leaked; `console.proxy_failed{reason:"rejected"}` |
| Agent process dies mid-session | Stream EOF, relay torn down, console shows a disconnect |
| CommandStream reconnects mid-session | No effect on live sessions — the relay is on its own stream. Only a new `OpenProxyCmd` needs the control stream |
Teardown is guarded by `sync.Once` on both sides: both `io.Copy` goroutines
finish, and whichever finishes second must not double-close.
## Testing
Written test-first.
- `server/internal/proxy/registry_test.go` — the two halves pair in either
order; expiry frees the entry; a second claim on a used `proxy_id` is
rejected; a mismatched `instance_id` is rejected. No network.
- `server/internal/proxy/session_test.go` — two `net.Pipe` halves; bytes flow
both ways; EOF in each direction tears down; double-close is safe.
- `server/internal/grpc/proxystream_test.go` — the authentication matrix: valid,
wrong token, unknown `proxy_id`, `proxy_id` belonging to another instance.
- `agent/internal/proxy` — a refused dial emits `ProxyClose`; the happy path
echoes bytes.
- End-to-end in `server`: a fake agent plus a `net.Listen` echo server, asserting
bytes traverse listener → registry → stream → echo and back. This is the test
that would have caught the original bug.
Manual verification, in this order: self-hosted SSH (proves no regression),
cloud SSH to a private-network host, cloud RDP to a Windows agent.