`agents.conf` and `gateway.htpasswd` move from /var/lib/hyperhive/gateway
to /var/lib/hive-gateway/conf, alongside the `tls/` the gateway already
kept there.
nginx reads both as an unprivileged user. Under c0re's state dir it could
only reach them by traversing a directory systemd re-declares `0750
hive-core` on every c0re start — so nginx was given `SupplementaryGroups
= [ "hive-core" ]`, which also handed it read access to everything else
group-readable in that tree. The tokens are individually 0600, but the
broker sqlite carries no explicit mode: every message between every agent
was readable by the process whose job is parsing untrusted network input.
Moving the files removes the need and the exposure together. The group is
gone, and its absence is now commented as load-bearing so it doesn't come
back as a fix for a symptom it would recreate.
Also drops this module's `/var/lib/hyperhive` tmpfiles rule. It declared
`0755 root root` and could never win against `StateDirectoryMode`, and a
losing declaration still reads as a guarantee — that is what sent the
first diagnosis of the outage looking for who had changed the mode.
Ordering is unchanged and still the thing that makes a fresh boot work:
tmpfiles runs before services and seeds both files empty-but-valid, nginx
names them (an `include` of a missing file is fatal, not empty), and
content arrives when c0re writes and reloads — which it does on every
topology change, so a boot against the empty seed resolves itself.
Folds in the mode fix: `write` now sets 0644 on the tmp file before the
rename, because a rename carries the source's mode and discards the
destination's, and the tmpfiles rule that declares 0644 is
create-if-absent so it never re-applies.
The gateway's nginx + dnsmasq no longer run in their own nspawn container.
`nix/host-modules/hive-gateway/default.nix` loses the
`containers.hive-gateway` wrapper and everything that existed only to punch
holes in it: `privateNetwork = false`, `CAP_NET_ADMIN`, five bind mounts,
its own `stateVersion`, `networking.firewall.enable = false`,
`networking.resolvconf.enable = false`, and the `hive-gateway-resolv`
path+service pair. 465 -> 303 lines.
The container never bought isolation here. It shared the host netns by
necessity — nginx binds the host's :80/:443, dnsmasq answers on the bridge —
so each of those settings was undoing a boundary the gateway could not
afford in the first place.
Four things made it more than a deletion, none of them visible in the nix
diff:
- The self-signed cert service also imports the hive CA leaf, so removing it
with the container would have left nginx naming a missing cert file, which
it refuses to load at all.
- The nginx reload is a hive-priv verb. It still needs root, but no longer
for the reason its doc gave, and `--machine=` was both transport and
scope — so the unit name is now hard-coded in the helper as the
containment.
- The lifecycle verb named a container that stops existing.
- `journalctl -M hive-gateway` had no machine to enter.
Per the operator's ruling, the operator verb keeps working and agents lose
it. `InfraContainer` answered three questions that used to share an answer;
it now splits into `name()` (identity), `target()` (Container vs HostUnit),
`service_unit()` (the systemd unit), and `agent_restartable()`, which the
MCP restart path checks before the capability so the refusal cannot read as
"ask for infra_admin". `SIBLING_CONTAINERS` drops the gateway — it gates the
requests that name a container as a string — while `FromStr` still accepts
it, because that answers what a name is, not who may act on it. The
dashboard's gateway journal reads host journald filtered to `nginx.service`.
Prose was corrected where it only named a location, and re-argued where the
container was doing security work: a `0666` per-agent socket was safe
because only the gateway container had the directory bind-mounted. There is
no mount now, so the directory permissions are the whole of the access
control — the constraint holds, its mechanism doesn't.
Gate: nix fmt / clippy --all-targets -D warnings / cargo test all clean (710
tests); hivectl-cli.md regenerated from the clap tree. The nix eval was run
in both TLS shapes at this commit: every delta in the rendered
virtualHosts is one of the three intended path moves, dnsmasq settings are
byte-identical, and the absence probe flips true -> false with bindMounts
emptied.
Closes the #3110 split — lib.rs is now just the crate doc comment and
the pub mod list.
journal.rs's new doc comment fixes a pre-existing bug: the old
JournalPriority doc text in lib.rs was actually half Capability's doc
(a leftover from an earlier reorder that moved the code but not the
comment above it).
Two views of the same graph existed: the typed `DagView`/`NodeView`
(`/api/state.rebuild_queue`, the `QueueDag` socket request, and the
`RebuildQueueChanged` payload) and `hive-jobq-wire`'s generic
`GraphNode` (`/api/jobq/graph`, `QueueNodes`). Every consumer has moved
to the generic one, so the typed pair is deleted rather than kept in
agreement with it.
What that removes, beyond the types: the `QueueDag` request and
`HostResponse::dags`; `Queue::snapshot`; `dag_view`, `visible_dags`,
`shown_on_wire`, `dag_finished_at` and `containers`; and the
`rebuild_queue` field on `/api/state`. `RebuildQueueChanged` keeps its
seq and loses its payload — nothing read it, and shipping the graph
both on an event and on an endpoint is the duplication this issue is
about. It stays an event rather than becoming a poll because
push-on-change is what every other live surface here does.
Two behaviours came out simpler for a structural reason. `await_dags`
needed two rules — settled means "gone from the snapshot" *or* "present
with every node terminal" — because the typed view evicted finished
groups; the generic view doesn't, so pending is just "some node isn't
terminal". And `state_of` in the tests no longer derives a roll-up at
all: a group root's own state is the scheduler's answer.
That second one found a bug. `cancelled_dag_still_runs_its_approval
tail` asserted the group reads `Cancelled` while the tail it exists to
protect was still pending — `rollup_state` flattened the surviving
child away and called the group settled. The root reads `Finishing`,
which is what the scheduler documents: own logic done, children still
running. The test now asserts that, with the reasoning inline so it
doesn't get "fixed" back.
Kept: `Source`, `State`, `PermPayload` and the `NodeId` alias in
`hive-host-sock::jobs` — shared vocabulary, still used by hivectl.
The wire types were in hive-host-sock, which is the host *socket* crate — so
anything living there is core-shaped by construction, and the projection had
quietly grown two core dependencies to match: it selected roots by matching
NodeKind::Dag, and rendered payloads through free functions in hive-c0re that
nothing obliged a second host to write.
hive-jobq is the wrong home too. That crate is the scheduler — logic — and
folding presentation in means every consumer of it carries a JSON vocabulary
it may never serve.
So: a new hive-jobq-wire. A host implements WireNode for its payload N and
WireResource for its resource name R; GraphWire::wire_snapshot is
blanket-implemented for Graph<N, R> when both hold, and for nothing else. A
payload that has never said how it displays has no way onto the wire.
wire_snapshot takes the roots to serve rather than reading Graph::roots
itself. Nothing is ever removed from a Graph, so retention is a policy only
the host can hold; hive-c0re passes visible_roots(), which is the existing
MAX_HISTORY_DAGS bound selected structurally (a root is a node with no
parent) instead of by node kind.
`jobs::NodeView` can only ever display hive-c0re's queue. Five of its
fields are domain knowledge: `approval_id` is only ever on a
`DeployWindow`, `inputs` only on a `MetaLock`, `build_log_id` only on the
nix-heavy kinds, `agent` is derived from the payload, and `kind` is a
payload tag consumers branch on. A component built against that shape
cannot render a second jobq.
`graph::GraphNode` is `hive_jobq::Node` with both generics erased: the
crate's own field set, with everything domain-specific in one opaque
`payload.data` slot the consumer renders without branching on. That is
the crate boundary made visible — hive-jobq owns structure, its host owns
meaning — and it is the same split #2957 drew inside the code.
Two details that are easy to get wrong and are pinned by tests:
`GraphDep::Node` carries `accepts` as the **set** of terminal outcomes,
not a strong/weak flag. A template emits its tails as a pair edged on the
same upstream node, and the only thing telling them apart is which
outcomes each accepts; collapsing that renders two structurally different
nodes identically.
There is **no roll-up field**. A group root ships as an ordinary node
with `parent: None`, and its own `state` is its subtree's answer —
`Finishing` means "own logic done, children still running", the terminal
states are the rolled-up outcome. A separate field would be a lossier
copy: `DagView::rollup_state` flattens `Running` and `Finishing` into
one, which is exactly the distinction a viewer wants.
`State` and `TerminalState` are re-exported from `hive-jobq` rather than
redeclared, so they cannot drift from the scheduler that produces them.
Follow-up from #2955 (mara: 'make core able to give agent a todo').
First migration slice: Spawned was pure FYI-check-when-convenient
material, not something needing an immediate turn.
Coordinator::push_todo/push_todo_submitter do a best-effort live dial
of the target agent's own hive-agent-sock (hive_host_sock::agent_todo_
socket), sending the exact UpsertTodo request in-container producers
(matrix/bash/forge-notify) already send. Push, not queue: agent
offline (socket absent) or dial failure is a silent no-op, no retry,
no fallback delivery -- matches mara's 'not available if offline'
call exactly.
HelperEvent::Spawned removed entirely (enum variant + all 3 call
sites migrated: handle_spawn's two arms, finish_approval's Spawn
approval-kind arm) rather than kept alongside a translation layer --
per mara's correction on the first design attempt, migrating the
producer means deleting the old path, not bridging it.
Verified: cargo build/clippy/test -p hive-c0re -p hive-host-sock
-p hive-sh4re clean (318 tests), nix fmt clean.
The push side modelled a store per peer hive: a --peer argument, a
swarm.peers.<domain>.snapshotStorePort option, and a swarm_peers module
whose entire job was answering "which peer". A swarm has exactly one
store, so none of that had anything to select between.
The receiver already proved it. It keys destination directories by
agent, not by sending hive, precisely so an agent that migrates keeps
one unbroken incremental chain -- which only makes sense if every hive
pushes to the same place. Per-hive stores would split the chain in two,
the case that keying exists to prevent.
So the destination moves to services.hyperhive.swarm.snapshotStore,
rendered into HYPERHIVE_SNAPSHOT_STORE, and swarm_peers is deleted
rather than adapted. address has no default because it is a
deployment fact this host cannot derive; port defaults because it is a
convention both ends read from the same option docs. An unset or empty
address fails naming the option instead of connecting somewhere
arbitrary, and a test asserts the message suggests no value.
Adds the caller the fd-passing machinery existed for: hivectl agent
<name> subvol snapshot push --peer <hive> resolves the peer, connects
to its snapshot store, writes the agent header, and hands the connected
socket to hive-priv, which runs btrfs send straight into it.
The split keeps the root helper ignorant. Everything that involves
knowing where a peer is, what the wire protocol looks like, and which
hive to trust happens in the unprivileged daemon; hive-priv only ever
receives an already-open descriptor. Once btrfs send starts, neither
process is in the data path, so a multi-gigabyte transfer costs no
per-byte work and survives a hive-c0re restart.
call_with_fd takes the descriptor by value and closes it as soon as the
kernel has it. A socket stays open until every copy closes, so holding
one back would leave the receiver waiting for an EOF that never comes:
btrfs receive blocks and this side reports success for a transfer the
peer never committed. Ownership makes that unrepresentable.
The peer's store port is a new swarm.peers.<domain>.snapshotStorePort
option rather than a constant matching the module default. A pushing
hive cannot read the receiver's configuration, so assuming 51821 would
push at a port nobody promised to listen on; absent, the push fails
naming the option. swarm_peers parses the mesh address the host module
has always rendered into HYPERHIVE_PEERS but nothing read.
The DagView / NodeView / Source / State / PermPayload types only ever
travel on the host admin socket and the dashboard channels hive-c0re
serves off the same snapshot; their whole consumer set is hive-c0re,
hivectl and the socket protocol crate itself. Living in hive-sh4re made
the five other crates that depend on it carry job-queue types they never
name.
Pure move: git mv of the module plus the import sweep, no type changes.
hive-sh4re keeps its own chrono (wire_time still needs it).
The agents root is 0700 and owned by the daemon's user, so hivectl's
client-side existence guard hit EACCES on traversal for anyone not root.
It reported that as "this command needs root; re-run with sudo", which
turned three verbs' pre-flight check into a permission error about the
wrong thing: `choom`, `subvol upgrade` and `subvol snapshot create` all
failed at the guard rather than at whatever they actually needed.
The daemon runs as the owning user and already answers this question for
its own provisioning paths, so expose it on the host socket as
`AgentExists` and have hivectl ask. Operators reach that socket through
the `hive-admin` group, so the guard now works without sudo.
`choom` still needs root for `machinectl shell` — we ship no polkit rule
granting those actions — so it now checks the effective uid and says so
directly instead of failing later inside systemd's authorisation.
Both groups only ever act on a single managed agent's state dir, so
they belong in the `agents` namespace rather than as top-level verbs
next to `forge` / `matrix` / `wg`.
Renames `quota limit` -> `quota set`: the enclosing group already
carries the noun, so the bare verb matches the flat `set-parent` /
`set-limits` spelling without stuttering, and it removes the
`set-limits` (cpu/mem) vs `quota limit` (disk) ambiguity. Adds a
cross-pointer from `set-limits` to `agents quota`.
Handlers stay in their own modules; `run_agents` gains the reparenting
glue. Regenerates docs/tools/hivectl-cli.md.
Refs #2724
The hive applies one `agentCpuQuota` / `agentMemoryMax` to every
container. That's the right default and the wrong ceiling: a build-heavy
agent needs headroom the other twelve don't, and raising the hive-wide
value to suit it hands that headroom to everyone.
Adds a per-agent override, persisted host-side and resolved per-field
against the hive defaults.
Follows the existing `meta/*.json` pattern (`capabilities.json`,
`tool-groups.json`): a host-side map read by `hive-c0re`, staged and
committed in the meta repo so every change lands in the audit trail.
```json
{ "sock": { "cpu_quota": "400%", "memory_max": "8G" } }
```
Fallback is **per field**, not per agent: an entry with only
`memory_max` leaves that agent on the hive-wide CPU quota. Absent file,
absent agent and absent field all resolve to the hive default, so the
feature is inert until someone opts an agent in.
Unlike the other meta files this one is **not** injected into the
container — a limit is something done *to* an agent, not something it
reads about itself.
```
hivectl agents set-limits sock --cpu-quota 400% --memory-max 8G
hivectl agents set-limits sock --reset
```
Values are validated before they're persisted: they go into a systemd
drop-in verbatim, and a typo there makes the unit fail to *start* —
turning a fat-fingered quota into a container that won't come back.
The command is declarative: each call replaces the agent's whole entry.
That makes a forgotten flag a silent revert, so a bare `set-limits
<name>` is rejected at the clap layer and clearing needs an explicit
`--reset`.
`ContainerView` gains `cpu_quota` / `memory_max`, both always populated:
there's no "unset" state to render, only "same as everyone else". They
reflect what the drop-in *says* — what the next start will enforce — not
a live cgroup reading.
The write goes through `meta::commit_resource_limits` rather than the
bare setter, so it's staged and committed under `META_LOCK`. Writing
without committing would leave the meta working tree dirty for the next
`prepare_deploy` to trip over.
Docs: `persistence.md` (the new meta file, and why it isn't injected),
`tools/hivectl.md` (the prose guide), `tools/hivectl-cli.md`
(regenerated clap dump).
Closes: internal/requests issue 25
A paused agent keeps its container, its claude session and its
dashboard/todo servers up, but stops driving turns. Messages queue
unacked and are drained on resume.
The whole protocol is a single marker file, `<harness>/paused`. That
directory is already a bind-mount shared between host and container, so
both sides just stat the same path: the harness reads it to decide
whether to drive a turn, hive-c0re reads it to render the badge and
writes/removes it for `hivectl pause|resume`. No new wire protocol, no
container round-trip, and it is sticky across restarts by construction.
Not calling `recv_next` while paused *is* the queueing semantic, so
there is no fencing to get wrong: reminders buffer in their unbounded
channel, the todo `Notify` permit coalesces, and a `request_next_turn`
that raced the pause survives because the gate sits above
`self_continue.take()`.
Graceful stop is handled host-side rather than in the harness: a paused
agent provably has no turn in flight, so `run_signal` skips the fence
entirely instead of eating the full `GRACEFUL_STOP_TIMEOUT` waiting for
a checkpoint turn that will never run.
`paused` is reported on `ContainerView` / `AgentStatusRow` for the
dashboard, orthogonal to `running` and reported for stopped containers
too.
Closes: hyperhive/hyperhive issue 2271
With the meta-update cascade (#2476) and startup sweep (#2450) folded
into single DAGs that grow per-agent subgraphs via append_subgraph,
nothing links parent/child DAGs anymore — parent_id is dead.
hive-c0re: drop parent_id from Dag/DagSpec (+ the DagView copy); delete
append_children and cancel_children (no callers); simplify trim_history
(no more terminal-parent-with-live-children guard — a one-big-DAG is
terminal only when its whole graph settles); drop the rebuild() parent_id
param; QueueDag returns just the polled DAG (no fan-out children to
gather). hive-sh4re: drop the DagView.parent_id wire field.
frontend: a multi-step op is one DAG now, so renderRebuildQueue drops the
childrenOf/orphans cross-DAG grouping and renders each entry flat; its
per-agent subgraphs render as nodes within the one row (split by deps).
Removed the dead rqe-child style + isChild plumbing.
Docs + the child-DAG queue tests updated/removed to match.
mara's review on #2436: no submit-await-submit composition, even
server-side. Adds Template::GracefulRestart (Signal -> Drain ->
StopForUpdate -> Reconcile, wanted=Up) mirroring how Restart already
does StopForUpdate -> Reconcile, plus submit::graceful_restart and
templates::graceful_restart. handle_restart_scoped now submits exactly
one DAG per agent up front for both the graceful and non-graceful
case -- no await_dags in the loop anymore.
hivectl restart --agent NAME previously composed stop() then start() as
two separate client-side daemon calls glued by CLI-process control flow
— not one DAG, and a dropped hivectl connection mid-restart (ssh drop,
Ctrl-C) left the agent stopped with no automatic follow-up. mara flagged
this as the first target for the 'dagify hivectl commands' issue.
New HostRequest::RestartScoped{scope, graceful} handles it server-side:
each targeted agent now rides exactly one atomic Restart-template DAG
(same one hivectl agents restart / restart-all already use) in the
common non-graceful case. --graceful has no single-DAG template yet, so
it submits the graceful-stop DAGs, awaits them server-side, then submits
the start DAGs — still one daemon call end to end, just not yet a single
DAG (noted as a follow-up). Infra containers restart synchronously as
before (no lease/DAG concept for them).
CLI-side restart() now just makes the one call + waits, same output
shape as before via render_lifecycle.
