| Filename | Latest commit message | Latest commit date |
|---|---|---|
d6f94e5247
An agent's subagent daemon publishes each subagent's output as terminal
rows on `$SWARM.term.<agent>.sub.<subagent>`, as the agent, into a
per-agent stream it creates itself; swarm-controller lists an agent's
subagents from that stream's subjects and relays one subagent's rows as
SSE; the swarm UI lists them under the agent's terminal preview and
reuses AgentTermPreview, full-screen tab included, with no input.
- swarm-nats.nix: the agent token may also publish
`$SWARM.term.{agent}.sub.>` and `$JS.API.STREAM.CREATE|INFO` on
`term-sub-{agent}`, and nothing else of JetStream. A module-eval arm
pins the agent-token grant as an exact list.
- mcp.nix: hive-subagent-daemon loads the agent's store identity
(`hive-agent-bao-cert/-key/-server-ca`, the ones hive-agent loads)
whenever the agent has a store, not only on the opencode preset. The
agent's own queue secret lives in the store, so this is the credential
the harness connects with.
- hive-subagent-mcp: `swarm_term` reads the agent's queue secret under
that identity, connects with the agent token, opens or creates
`term-sub-<agent>` (max_age 24h), and publishes classified rows from
the sink every subagent line already passes through. The sink only
queues (bounded, drop-and-count); a missing store, refused credential,
failed stream create or failed publish is a log line.
- The stream-json classifier (`stream_enrich`) and the `TermMsg` row
types plus `fit` move from the hive-agent binary into hive-sh4re, so
the subagent daemon publishes the rows AgentTermPreview already
renders. hive-agent keeps its LiveEvent classifier on top.
- swarm-controller: `GET /api/agents/{name}/subagents` and
`GET /api/agents/{name}/subagents/{subagent}/term/stream`.
- docs/swarm: what the UI shows and what the queue carries.
Closes #4827
|
||
| .. | ||
| src | ||
| Cargo.toml | ||
| README.md | ||
swarm-nats-auth
The auth-callout responder for the swarm's NATS queue — the half that lets the server say yes.
nix/host-modules/swarm-nats.nix configures nats-server with an
auth_callout block. That block with no responder is the fail-closed
state, and it is the measured one rather than the obvious one: on the pinned
nats-server, both authorization { } and authorization { users: [] } answer
PONG to an anonymous client, while an auth_callout block sets
auth_required and refuses every client no responder has approved. So the
module ships the final config from the start and this crate only adds the
ability to approve. No interim hole is ever opened.
Why a crate and not more config
The reply is a signed NATS user JWT. Signing needs the account nkey seed and the JWT framing, which is program work — which is why the container and its config landed first and this arrived separately, rather than the pair being one change.
The invariant that is easy to break
⛔ issuer_account must be absent from the issued user token. It is an
operator-mode field. The module renders server-config mode (accounts { AUTH, APP }, no operator), where its mere presence makes the server refuse
the client — Error non operator mode account "AUTH": attempted to use issuer_account — while the responder cheerfully reports granted=true. The
account is named by the claims' aud instead.
nats_jwt::Token::new_user always sets it, so reaching for that constructor
reintroduces the bug. nats-jwt is a dev-dependency: it cannot express
aud on either token, and its role here is as the encoder's test oracle,
not part of the path that runs.
Rules the code follows
- A denial is a signed response carrying an error, never silence. The server cannot tell an absent responder from a refusing one, so staying quiet turns every rejection into a timeout and hides an outage inside what looks like ordinary denials.
- Anything that is not an explicit
{"active": true}denies — including an introspection call that could not be made. The failure modes of an HTTP call are exactly the conditions under which an attacker would most like this to fall open. - Every credential is a path, never a value. A value in nix config lands
in the world-readable store; a value in
argvis readable via/proc/<pid>/cmdline, which is0444. Paths are not secrets, so passing them as flags is fine. - The introspection timeout is pinned below the server's
authorization. timeout, with a test asserting the relation — a responder that answers after the server gave up is indistinguishable from one that never answered.
What the tests do and do not cover
Unit tests cover the JWT framing, including byte-equality against nats-jwt
on the one shape that crate models. They are not sufficient on their own:
this crate's shape is decided by a server that parses what it emits, so the
change was also driven against a real nats-server — every refusal repeated
with the responder live, because a responder that says yes to everyone
passes "a client can connect" perfectly. That harness lives outside this repo;
the PR that added this crate links it.
⚠️ Its introspection endpoint is a stub. Those runs prove this crate's own behaviour and nothing about the real authelia integration, which needs a deployment.