Watch

atlas 30fa54cbc6 feat(swarmctl): add agent create, queueing the swarm-controller creation DAG

`swarmctl agent create <name> --hive <hive>` POSTs `/api/agents` to
swarm-controller over the daemon's unix socket and prints the queued
job's node id.

It deliberately does not wait. The endpoint queues a DAG whose last node
*publishes* a deploy message; the hive's `hive-c0re` then converges on
its own clock, out of the controller's sight. So even a fully settled
graph would not mean the agent is up, and there is nothing this CLI
could wait for that would let it claim otherwise. Printing the id is
exactly what the response says and all of what it says.

Transport is a bare hyper HTTP/1.1 client handshaked onto a tokio
`UnixStream` via `hyper_util::rt::TokioIo` — the same crate family
`hivectl/src/watch.rs` and `hive-agent/src/web_ui/proxy.rs` already use,
all of it already workspace-pinned. The request/response shapes are a
local mirror rather than a shared crate: the controller's own types are
private to its binary and this crate does not link it, the same
separation `hivectl` keeps from `hive-c0re`.

Errors are reduced to one actionable line — the controller answers
RFC 9457 problem+json, so an unknown `--hive` reaches the operator as
the roster of hives that would have worked rather than a body dump.
Response `warnings` are printed when non-empty.

The nix module wraps the binary with `SWARM_CONTROLLER_SOCKET`, read
from the same `socketPath` the daemon binds.

Refs #4399

2026-09-14 19:40:23 +02:00

5.3 KiB

Raw Blame History

swarmctl

Swarm-level operator CLI. Runs as root on the host running swarm-controller, and acts on that host directly.

Distinct from hivectl, which drives one hive's hive-c0re over its admin socket. This crate does not link swarm-controller, for the same reason hivectl does not link hive-c0re.

Why root, and why no socket

The first verb writes authelia's users database. Making that write rootless was examined and rejected — relocating the file only turns a write problem into a read problem. Move users.yml into a directory the controller owns and the controller can write it, but authelia then has to read it across the same boundary in the other direction. Making that work needs either a hand-pinned gid (the container's uids are allocated inside it, at activation — see the uid-assignment issue) or world-readable password hashes. Both are worse than root.

So this binary serves no socket, publishes no HTTP route and has no privileged helper. When a verb has to run as a non-root user or from another host, the answer is a group-gated admin socket, separate from the controller's 0666 gateway-facing one — not a widening of what root does here.

That rules out a way in, not a way out: agent create is a client of the controller's socket, because the work it asks for is a job graph only the controller can queue (see below). Nothing here becomes reachable by that.

One file, two writers

users.yml — authelia's own users database — is read and written directly. There is no second store.

There used to be: a private users.json here, canonical, with users.yml rendered from it, while swarm-authelia-bridge kept its own pair against the same physical file. Two canonical stores for one file is a seam, and it bit — a writer whose own JSON was missing could not tell "nothing here yet" from "someone else's users", and refused to write.

The argument for the split was that it let this crate work without a YAML parser. It didn't: the JSON was read back on every run, so the round-trip was already being paid — the two files differed only in format.

⚠️ The file is round-tripped, so comments and hand-formatting do not survive a write. Values do, and so do keys this binary does not model. See swarm-authelia-bridge/README.md for what both writers must uphold.

Configuration

Every path comes from the nix module that installs the binary, because every one is derived from an option that module owns. They are required rather than defaulted — a default would be an address we hope points at something, and one that resolves cleanly to the wrong place is worse than an error.

variable	what
`SWARMCTL_AUTHELIA_BIN`	the configured authelia; argon2 params must match the verifier's
`SWARMCTL_AUTHELIA_USERS_FILE`	host-side path of the users database
`SWARMCTL_AUTHELIA_MACHINE`	container name, for `systemctl -M`
`SWARMCTL_AUTHELIA_UNIT`	authelia's unit inside that container
`SWARM_CONTROLLER_SOCKET`	the controller's unix socket, for `agent create`

`user add`

# swarmctl user add mara --display-name "Mara" --group admins

The password is generated by authelia (crypto hash generate argon2 --random) and printed once. It is never passed on a command line: /proc/<pid>/cmdline is world-readable, so a password in argv is readable by any local process for the lifetime of the call.

`agent create`

# swarmctl agent create scribe --hive alpha
queued: job node 42
agent "scribe" will be deployed to hive "alpha" once the job graph runs; `swarmctl` does not wait for it

POST /api/agents on the swarm-controller, over its unix socket (--controller-socket, else SWARM_CONTROLLER_SOCKET, which the nix module sets from the daemon's own socketPath). It prints the queued job's node id and stops there.

It deliberately does not wait. The endpoint queues a DAG — SSO identity, forge user, forge repo, repo membership, config-repo seed, then a deploy message — and the last of those publishes: the hive's hive-c0re picks it up and converges on its own clock, out of the controller's sight. So even a fully settled graph would not mean the agent is up, and there is nothing this CLI could wait for that would let it say so honestly. Watch the swarm UI's job view for the rest.

No approval gate, for the same reason nothing else here has one: running this binary already means being root on the controller's host.

--hive is required. It is an address — where the deploy message goes — not an attribute of the agent, so there is no sensible default. The controller checks it against the swarm's hive roster and names the hives that would have worked when it misses.

The request and response shapes are mirrored in src/agent.rs rather than shared: the controller's own types are private to its binary, this crate does not link it, and there is no wire-type crate between them. Two fields out, two in, both ends validating — a drift shows up as a 400 naming the field.

5.3 KiB Raw Blame History