| Filename | Latest commit message | Latest commit date |
|---|---|---|
b68b74d687
A new `swarm-builder` service, in its own container on the host that sets
`deploy.builder.enable` (defaulted on by `allSwarmServices` where the host
has the builder's store leaf, like the review classifier).
When an `agent-configs/<agent>` PR is opened, reopened or synchronized,
swarm-controller queues a `build_agent` jobq node. The node sets a
`swarm-builder` commit status to pending on the PR head, sends
`{agent, rev, pr}` to the builder over NATS request-reply
(`$SWARM.builder.build`), and writes success (store path), failure (last
line nix printed) or error (no builder / no reply) on the head. The node
fails when the build does, so the Jobs page matches the PR.
The builder builds nix/builder/check-flake's `check` with every input
overridden: hyperhive and nixpkgs to the host's own
`c0re.hyperhiveFlake`/`c0re.nixpkgsFlake` (what a hive there renders into
its meta flake), and the agent to the PR head on the forge. Evaluation is
pure and runs in the container; builds go to the host nix daemon as an
untrusted client. The private config repo is fetched with the token of a
new read-only `builder` forge account (in the `hives` team), which the
controller mints into `swarm/services/builder/forge/token`; git gets it
from a credential helper reading the child's environment, so it is in no
argument or URL.
Credentials: a `builder` store leaf minted by glue-bao-tls on the store
host, a `swarm-builder` reader policy (forge token + its own OIDC client
secret), an authelia machine client, and a `--builder-client` grant in the
queue's auth-callout responder. The controller gains a grant on the token
path and the secret publisher a deny on it.
This is the agent-config build check atlas scoped on #5081 (slice 3):
nothing is pushed to a cache, nothing is written to KV, and a red status
gates no deploy. The check composes agent-base + the config without the
values a hive adds (CA, docs, claude-code path, ports, input follows), so
it can disagree with a hive's own build.
|
||
| .. | ||
| src | ||
| Cargo.toml | ||
| README.md | ||
swarm-secret-client
Reading and writing a swarm credential in the secret store, over vaultrs.
The HTTP is that crate's job. What this one owns is the agreements both
ends of the store have to state identically: where a credential lives, which
field its bytes are in, and how this deployment's environment becomes a
logged-in client.
The surrounding picture — which secret is minted where, and why delivery is a
copy rather than a bind mount — is docs/swarm/secrets.md.
Why a crate and not a module per binary
The store has two Rust ends and they are peers: the swarm controller writes a credential, a hive reads it. Neither is senior to the other, so a path formatted at each call site is an agreement with no owner — it holds right up until one side is edited alone, and then it fails as a missing key rather than as a mismatch.
The third end is what settles it. nix/agent-modules/forge-token.nix reads
the store with bao kv get -field=value. That reader is a shell line in
a nix module: it cannot be renamed by the same refactor as a Rust struct, and
no Rust test reaches it. So the field name is pinned by a test against the
serialised literal rather than left to the struct definition.
The identity is a certificate, and the role is the hive's name
Authentication is the store's cert auth method.
nix/host-modules/glue-bao-tls.nix mints the client certificate with its CN
set to the hive's name, because a cert-auth role matches on the CN. So
cert_role is not a free choice for the caller: a hive passes its own name,
and the policy attached to that role is what scopes what it may read.
The listener's tls_require_and_verify_client_cert is a different thing and
not a substitute. It decides who may open a connection; it says nothing about
who the connection belongs to, and a store with the option set and no cert
mount configured refuses every login made here. That refusal is what
Error::Vault out of connect means.
Configuration
Settings::from_env reads BAO_ADDR, BAO_CLIENT_CERT, BAO_CLIENT_KEY,
and the optional BAO_CACERT.
The BAO_ spellings are read explicitly rather than left to vaultrs. Its
own defaults look for VAULT_ADDR / VAULT_CLIENT_CERT / VAULT_CLIENT_KEY,
which no unit in this tree sets. Falling through to them builds a client with
no identity at all, and that surfaces as a TLS handshake failure — a place
that names neither the variable nor the reason.
Empty is as absent as unset. systemd renders an unset nix option as
Environment=BAO_CACERT=, so empty is the shape a missing value arrives in.
The certificate and key are paths, not values, and are read at connect time — same rule as every other credential in this tree, for the same reason: a value in a nix expression is rendered into the world-readable store.
Reading the environment is separate from connecting (Settings::from_lookup)
because every one of those failures is a misconfiguration an operator has to
read an error about, and none of them needs a reachable store to happen.
Names that arrive from elsewhere
matrix::account_path is fallible, which for a string formatter needs saying:
its segments are an agent name from the topology and an account name from that
agent's own config. A / turns one agent's segment into another agent's
directory and .. walks out of the prefix entirely, so the charset it accepts
is deliberately narrower than what the store would.
One module per kind of secret
client moves whatever type a caller names; it decodes nothing itself. What a
stored object holds is stated in the module that also builds its path —
matrix today, and a second kind of swarm secret gets a module beside it.
The split is deliberate. A single shared struct that grows one field per consumer ends up carrying, on every path, a field only one path's reader has ever heard of; and the two things a kind of secret must pin — where it lives and what is in it — are one agreement that reads worse split across modules.
What this crate does not do
It has no opinion on what a caller may read. That is the policy attached to the cert role, and it lives in the store.
It holds the token minted at login and renews nothing. A handle is built per credential, so the login is the cheap part of a rare operation — a caller that wanted to keep one alive across a token's lifetime would need more than this.