Step 1 of removing the localhost fallbacks: make the renderer emit the
value it already knows, so the option stops being a second, disagreeing
source of truth.
These options existed but nothing ever set them, so every agent fell
back to their localhost:<port> defaults while the real value reached
the container only as an env var. The two are consumed at different
times — the option is baked into scripts at build time (tea-login's
FORGE_URL), the env var is read at runtime — so which answer a given
code path gets depends on which one it happens to read.
Emitting them here follows the shape the otel block already uses: host
state becomes build-time agent module config. It is the precondition
for deleting the defaults, which is the actual fix: a loopback address
is only correct when the callee shares the caller's netns, and the
forge and homeserver are moving to swarm level, possibly onto other
hosts.
An absent var emits nothing rather than a guess. Once the defaults are
gone that surfaces as an eval failure, which is the point — better a
build that stops than an agent quietly talking to a port on the wrong
machine.
The emit is a pure helper rather than an inline loop so it can be
tested without process env. The first version of the test set env vars
and rendered the whole flake; it failed because the parallel runner
raced it against the existing env-mutating test, not because of any
defect. Testing the pure function has no such hazard, and the
render-level variant is kept #[ignore]d with that reason recorded.
Adds hyperhive_rev to the dashboard's /api/state StateSnapshot,
resolved via the same current_flake_rev helper get_agent_meta's
per-agent hyperhive_rev already uses. home.js renders it next to the
existing hive-identity line, truncated to the last 12 chars with the
full value in title=, hidden when the flake ref isn't a local path pin.
Requested by annika (infra.run) via dmatrix, hyperhive#2854.
run_worker only polled shutdown.changed() inside the select! arm, which the
claim-and-spawn branch skips via continue whenever there's ready work. Under
a sustained stream of ready claims (a boot sweep across agents is the
realistic case) exit was deferred until the queue happened to drain instead
of being observed promptly.
check *shutdown.borrow() explicitly at the top of every loop iteration
instead of relying solely on the select! arm, so a busy loop still sees
shutdown promptly.
also documented why dropping pending Queued DAGs on shutdown is safe: every
NodeKind is idempotent-convergent, which is a property of the node set, not
of the queue, and isn't enforced by the type system. flagged the side-effect
tails (EmitRebuilt, ResolveApproval) as the ones closest to the edge.
fixeshyperhive/hyperhive#2848
`NodeId` has been globally unique across DAGs since #2801, so the dag id
carried no information the node id didn't. The parameter was already
underscore-prefixed as unused, but still populated by `claim_ready`, carried
through the scheduler's mpsc on every `Claim`, and passed at the call site —
three layers of plumbing feeding a dead argument.
Removing it surfaced four more dead things it had been keeping alive:
`settle_approval_tail`, `settle_rebuild_tail` and `drain_meta_syncs` each took
a dag id they only forwarded to `complete_node`, and one `submit` binding was
never read. Those are deleted rather than underscore-prefixed — prefixing is
what let the original argument survive this long.
`Claim.dag_id` stays: it has live consumers in the tracing spans,
`append_subgraph`'s container guard, `Ctx` for the build-log link,
`first_error`, and the approval-deploy context.
`hive_host_sock::jobs::State` was a hand-maintained copy of
`hive_jobq::State` — five variants spelled the same in both, kept in sync
by whoever remembered. Adding `Skipped` last week meant adding it twice.
The wire crate now re-exports the scheduler's enum and `to_wire_state` is
gone.
Two states that were hidden now reach clients. `to_wire_state` renamed
`Pending` to `Queued` and folded `Finishing` into `Running`, so the
dashboard could not distinguish a node waiting on its dependencies from
one whose own work is done while its sub-nodes still run. Both are now
visible, and consumers say which they mean.
Every consumer had to move with it, and only the Rust ones said so: the
exhaustive matches in `hivectl` and `DagView::rollup_state` failed to
compile, while the dashboard's fourteen string comparisons would have
gone quietly wrong — a `finishing` node no longer counting as running,
a `pending` node no longer as queued.
The frontend also builds CSS class names out of the state string
(`rqe-` + state, `rqe-node-` + state) and keys its glyph map on it, all
lowercase. Those go through a `stateSlug` helper now; comparisons use the
wire spelling, presentation lowercases. Without that split every queue
entry and node chip would have silently lost its styling.
Dropping the `State as JobState` alias in hive-c0re falls out of this:
the alias only existed to tell two `State` types apart, and there is one
now.
`dag_of` and `dag_first_error` had shrunk to a single delegating call once
the walks moved into `hive-jobq`; their callers say what they mean without
the hop.
`subtree` was worse than redundant. It collected the descendant ids into a
`Vec` and both callers then looked each node up again by id — `dag_view`
needed a `let … else { continue }` for a lookup that could not fail.
Iterating `descendants()` hands back the node directly, so the round-trip
and the re-lookup both go.
`JobQueue::cancel` decided whether a DAG could be cancelled by reading node
run-state, walked the subtree, judged per node whether that node had asked
to observe cancellation, and re-ran the container's roll-up. Every one of
those is a fact the scheduler owns; core was reaching across the boundary
to compute them.
`Scheduler::cancel_node` now takes the whole subtree: cancelling a node
cancels the work under it, since a group is abandoned by abandoning its
root. The existing method generalises rather than gaining a sibling — it
had one production caller, which this replaces.
The gate runs over the work *under* the node, not the node itself: a group
root's state is its subtree's roll-up rather than a step that ran, so a
container is `Finishing` and never `Pending`, and gating on it would refuse
every cancel. A node with no children is its own work, which keeps the
previous single-node behaviour.
`observes_cancellation` moves in with it — it reads a node's declared edges
and knows nothing about what the payload means.
Core keeps the one genuinely domain-specific step, resolving a wire
`dag_id` to its container node, and is three lines otherwise.
Returning `false` for an unknown id gave the same answer as a node that is
merely still running, so a caller polling a stale id would wait forever for
a state that can never arrive. `Option<bool>` makes the two cases separate,
matching `node()`'s convention that `None` means the id isn't in the graph.
hive-c0re hand-rolled four traversals over a graph it doesn't own, because
`Graph` exposed only `node()` and `nodes()`. They are generic — nothing in
them knows what a hyperhive DAG is — so they move to `hive-jobq` and core
delegates.
`Graph` gains `root_of`, `descendants`, `roots`, `is_settled` and
`first_error`; they reuse the private `is_descendant` the crate already had
for its dep-scope rule. `subtree` gets faster on the way: core walked every
node's whole parent chain to the root for every node in the graph, where
`is_descendant` stops as soon as it sees the ancestor.
`first_error` deliberately looks for the first `Failed` descendant that
*carries* an error rather than the first `Failed` one. A node that rolled
its failure up from a child holds no error of its own and sorts before that
child, so the simpler version reports `None` for the common case and the
dashboard loses the reason. The distinction has its own test.
`dag_is_terminal` is deleted rather than moved: it was already a plain
`state.is_terminal()` read, and its three call sites now ask the graph.
A node ruled out by its own dependency edges settles `Skipped` host-side,
but the wire folded it into `Cancelled` and `dag_view` filtered it out
entirely, so a client never saw which branch a run didn't take. Post-#2785
that is not a rare shape: every approval DAG has two not-taken tails and
every rebuild has one, on the happy path as much as on failure.
`State` gains `Skipped`, and it counts as terminal — the wait loops in
hivectl's progress display and the daemon's dag-settled check decide
"finished" with `all(is_terminal)`, so omitting it would hang them on
essentially every DAG.
`dag_view` now emits skipped nodes but no longer lets them keep a DAG
alive. Serialization and completion were the same expression: a DAG left
the snapshot because its nodes had all been filtered away. Keeping skipped
nodes on the wire under that rule would pin every finished deploy in the
queue view forever, so the completion test is now its own flag.
`rollupState` in the dashboard gains the matching arm. It has no `done`
case — `done` is inferred by falling off the wire — so its trailing
`return 'queued'` catches anything it doesn't recognise, and a green
deploy would have read as permanently queued the moment the backend
started emitting the new state. The Rust and JS roll-ups have silently
disagreed before; they are edited together here and say so.
A host restart brings hive-c0re up and the startup sweep rebuilds every
stale agent. Until now that stop was mechanical: `StopForUpdate` hung
straight off `Prebuild`, so an agent that was mid-turn when the host went
down had its turn cut off rather than finished.
The sweep now builds the same `Signal` -> `Drain` -> `StopForUpdate` chain
a graceful `hivectl restart` already uses, reusing the existing nodes and
`GRACEFUL_STOP_TIMEOUT` unchanged. Cost is bounded: the per-agent drains
overlap, so the sweep waits one timeout in total rather than one per agent.
`Signal` parents the rest of the stop instead of sitting beside it. All
three of `Signal` / `Drain` / `StopForUpdate` declare the agent lease, and
a resource is held across its holder's whole subtree — as siblings each
would take the lease separately, leaving a window between them for another
DAG to claim the agent mid-bounce.
Scope is the boot sweep alone: a manual rebuild, a meta-update cascade
child and a deploy all still stop mechanically, and a test pins that shape.
`rebuild_nodes` takes a `RebuildOpts` struct rather than a second
positional `bool`, which two adjacent flags would have made easy to swap at
a call site. Its callers no longer hard-code the subgraph's length either:
the `EmitRebuilt` tails and `FinalizeDeploy` used literal indices that
silently encoded "this builder emits exactly six nodes with `Reconcile`
last", which a variable-length subgraph turns into a wrong-node edge rather
than a compile error. They read the index off the emitted list now.
Review follow-ups on #2785.
`DepWhen` wraps `BitFlags<TerminalState>` instead of a hand-rolled `u8`,
so the bit manipulation belongs to the library and `TerminalState` gains
its flag value from `#[bitflags]` rather than a `bit()` match anyone
could get wrong. `of`/`accepts`/`is_empty` become one-liners over it.
Serialization is written out by hand rather than derived: clippy's
`unsafe_derive_deserialize` fires on deriving over a type with unsafe
internals, and the honest fix is to say what the wire form is. It is now
the list of accepted outcomes — `["done","failed"]` — which reads better
than a bitmask and survives the bits being renumbered.
Also drops the `pub use` re-export of `DepWhen` / `TerminalState` from
hive-c0re's `model`. It existed so that `use super::model::…` kept
compiling, which is a shim for one consumer's convenience; the sites
import from `hive_jobq` directly now.
And removes comments narrating what the code used to be. Git holds that.
Splits what was one `Cancelled` outcome into two, because they were two
different facts wearing one name:
- `Skipped` — the node's own edges ruled it out. Expected; the failure
branch of a run that succeeded is `Skipped`. A parent's roll-up
**ignores** it.
- `Cancelled` — the work was dropped before it could start. Still
not-success for the roll-up, as before.
Without that split, branching on outcome defeats itself: exactly one
branch is always ruled out, `any_child_failed` counted it, and every DAG
containing a branch would have rolled up failed no matter how the run
went. Caught in review before it was written, not after.
`AFTER_ANY` becomes `{Done, Failed, Skipped}` — "anything except the work
being dropped". That is what it always meant; it only swept in
cancellation because cancellation wasn't distinguishable from
elimination. Audited every user rather than assuming, which is how the
one regression in my own proposal surfaced: `{Done, Failed}` would have
refused to run rebuild's recovery `Reconcile` after a failed `MetaSync`
(that eliminates `Prebuild`, so the tail's dep is `Skipped`, not
`Failed`) and left the container down.
With that, the templates stop computing outcomes and let the graph pick:
- `ResolveApproval { approval_id, outcome }` — one tail per outcome, each
edged to accept only its own, so exactly one is ever runnable.
- `EmitRebuilt { agent, ok }` — a pair. `ok` is not derived, it is which
of the two the graph let run.
Edges are conjunctive, so "any of these roots failed" is not directly
sayable. The composition: the success branch is `AFTER_OK` on every root
(so it is itself eliminated the moment one doesn't succeed), and the
failure branch keys off *that* elimination. The failure branch also
waits on every root — without it, a failed `Prebuild` eliminates the
success branch immediately and the failure would be announced while the
recovery `Reconcile` was still running. The tests caught that one.
Deletes, all of them #2770's host-side debt:
- `Claim.deps`, `DepOutcome`, `Claim::deps_state`, `Claim::deps_error`
and the dep-snapshotting loop in `claim_ready`. Executors read their
own variant now; nothing inspects anything.
- `NodeKind::is_tail()` and the `cancel` exemption built on it. Sparing
is derived from the edges: `cancel` keeps a node iff one of its edges
accepts `Cancelled`. An approval tail names it and survives to resolve
the row; `Reconcile` doesn't and is cancelled with the rest. My earlier
claim that this couldn't dissolve was only true while `AFTER_ANY`
accepted cancellation.
`resolve_approval_dag` / `deploy_terminal_tag` now take `TerminalState`
rather than the wire `State`, so both matches are exhaustive instead of
ending in a catch-all.
Skipped nodes are filtered off the wire alongside `Done` ones. That costs
some dashboard detail on a failed rebuild — which steps were skipped —
and the tests say so with a pointer to the follow-up. Surfacing them as
`Cancelled` instead would be worse: the client roll-up ranks `Cancelled`
above `Running`, so a successful DAG with a not-taken branch would read
as cancelled.
`DepWhen` was two named cases, so every new combination wanted a new
variant. It is now a set over the terminal outcomes: a `u8` bitset
newtype, no dependency, with `AFTER_OK` / `AFTER_ANY` kept as the two
constants the templates actually use. "Run regardless" is all outcomes,
"anything that isn't a failure" is `{Done, Cancelled}`, a compensation
branch is `{Failed}` — closed under combination, so it never needs
another variant.
`TerminalState` is its own type rather than a subset of `State`, so an
edge cannot name `Pending` / `Running` / `Finishing`. Those are
meaningless in a dependency and are better unrepresentable than
validated against. The empty set is the one thing that can't be typed
away — nothing satisfies it, so `validate` rejects it next to the cycle
check.
Two consequences worth calling out:
- `cascade_cancel` collapses to one rule: a pending node is doomed once
any edge it names can no longer be satisfied. The hardcoded `AfterOk`
special case is gone, and a weak-edged node survives its dependency's
cancellation because of its own edge rather than by exemption.
- The cascade now runs on **any** terminal outcome, `Done` included.
With sets, success rules dependents out just as failure does — a
`{Failed}` branch is unsatisfiable the moment its dependency succeeds,
and leaving it `Pending` would wedge the subtree non-terminal forever.
That is a hang, not a wrong answer, so it is the load-bearing half of
this commit.
Edges are conjunctive, so "any of these N failed" is not directly
sayable. The composition that works is in the tests: the success branch
depends `AFTER_OK` on every root, so it is itself cancelled the moment
one of them doesn't succeed, and the failure branch hangs off *that*
with `{Cancelled}`. Exactly one of the two runs.
Also deletes hive-c0re's duplicate `DepWhen` enum and the
`to_crate_when` translation beside it. The copy bought nothing and had
to be widened in lockstep with the crate's edge model — it is the
in-between layer #2772 exists to remove, and it is what broke the build
when the crate's spelling changed.
All 34 jobq tests pass, including the four new ones covering both
directions of a failure-only branch, weak-edge survival of a cancelled
dependency, and the aggregator composition.
Review catch from argus on #2770: with the terminal hook gone, cancel
spares the DAG's tail node so it can report the cancellation — which
leaves that node `Pending` until the scheduler's next pass.
`post_rebuild_queue_cancel` emits its snapshot synchronously, and
`rollup_state` ranked `Queued` above `Cancelled`, so the operator who
just cancelled a DAG saw it go back to **Queued**: "the cancel didn't
take". Asserted in `cancel_clears_queued_dag`, which failed before this.
Fixing it surfaced an older disagreement. `builds.js::rollupState` has
always been `failed > cancelled > running > queued`; the Rust was
`failed > running > queued > cancelled`. The two had silently diverged
under a doc-comment claiming they agree. Harmless until now only because
cancelled nodes never coexisted with live ones — a spared tail running
over cancelled work would have rendered `Running` host-side and
`Cancelled` in the dashboard.
The JS was the correct side, so the Rust moves to match it exactly:
`Failed > Cancelled > Running > Queued > Done`. No frontend change.
The queue carried a per-DAG `HookKind` that fired an inline side effect
from outside the graph when a container rolled up terminal. mara asked
three times why this could not be an ordinary node; the answer in the
code was a doc-comment claiming a node could not work, and it was wrong.
`DepWhen::AfterAny` already existed with two live users, and a weak edge
is satisfied by a `Cancelled` dep, so a tail node runs on success,
failure and cancel alike. What was genuinely missing was smaller than a
hook: a node had no way to learn how the work it followed ended.
So: `Claim` now carries `deps: Vec<DepOutcome>`, snapshotted at claim
time from the graph the scheduler already holds (no `hive-jobq` change).
`Claim::deps_state()` / `deps_error()` roll that up, and two new kinds
consume it — `ResolveApproval { approval_id }` and `EmitRebuilt { agent }`.
Templates append one as a group-root with `AfterAny` edges onto the DAG's
other group roots; a root's state is its subtree's roll-up, so that
covers every node without fanning out to each of them.
Deleted: `HookKind`, `DagSpec.hook`, `NodeKind::Dag.hook`, `DagMeta.hook`,
`TerminalDag`, `terminal_dag()`, `terminal_summary()`, `dag_agents()`,
`dag_rollup()`, `fire_terminal_hook()`, `run_terminal_hook()`,
`emit_rebuilt()`. `complete_node` returns `()`.
Load-bearing details:
- `JobQueue::cancel` spares tail nodes instead of cancelling the whole
subtree, and returns `bool`. Without this a cancelled approval DAG
would dangle its approval forever — the hazard `tests.rs` already
named. The spared tail's deps are `Cancelled`, which satisfies its weak
edge, so the scheduler claims it and it resolves the row as cancelled.
`hive-jobq` anticipated exactly this: `cancel_node`'s doc already says
to settle afterwards so "a weak-edge terminal node observing the
cancellation" can advance.
- The existing `complete(container)` call after cancelling is kept and is
deliberately a no-op when a tail was spared (a non-terminal child parks
the container back in `Finishing`), so power ops still settle
synchronously with no branch.
- `DeployTail` is NOT `is_tail()`: it does real compensating work, and a
cancelled DAG has nothing to compensate.
- `exec::failure_reason` falls back to `first_error(dag_id)` because a
group root that rolled up `Failed` from a child carries no error of its
own — without it every tail-reported failure would lose its reason.
- `EmitRebuilt` is per agent, so a multi-agent DAG reports each agent's
own outcome rather than painting all of them with the DAG roll-up.
- `ResolveApproval` is agentless: the approval row already names its
agent, and that is also what lets one tail close a multi-agent DAG.
Transients-derived-from-running-nodes and the frontend's node-kind
strings stay out of this change; they touch iris's slice and review
better next to their own diff.
mara on PR #2769: "make the default null instead of special casing """.
`claude_code_path` was a `String` whose empty value meant "no host-level
pin". That is a sentinel doing an `Option`'s job — the same shape argus
and mara already rejected on #2755's weights, and the same
empty-field cruft mara called out on #2756.
So it is `Option<String>` end to end:
- host module: `claudeCodePath` evaluates to `null` when
`claudeCodePackage` is unset, so `serve.json` carries JSON `null`
rather than `""`.
- `Coordinator` + `HiveEnv`: `Option<String>`, defaulting to `None`.
- `render_flake`/`render_flake_with_lookup`: `Option<&str>`, and the
emission is an `if let Some(path)` instead of an `is_empty()` guard.
- agent module: `hyperhive.claudeCodePath` is `nullOr str`, default
`null`.
Behaviour is unchanged in both directions; only the way "unset" is
spelled moves. The `builtins.hasContext` assertion still guards the
pinned case (short-circuited by the null check, so an unpinned hive
never evaluates it).
16/16 `meta::` tests, clippy clean, `nix fmt` no-op, `nix build .#docs`
green.
Agents run whatever `claude-code` the meta flake's `nixpkgs` resolves
to, and that is normally a release channel. This one package moves fast
enough that stable trails unstable by weeks — 26.05 is on 2.1.187 while
unstable carries 2.1.220 — and an agent cannot fix it for itself: it
only ever sees the single nixpkgs hive-c0re injects, so an `agent.nix`
has no other tree to reach for.
New host option `services.hyperhive.c0re.claudeCodePackage` takes the
package directly and rides the existing `hyperhiveDocs` threading path —
serveConfigJson -> HiveEnv -> render_flake — to reach each agent as
`hyperhive.claudeCodePath`. Null (the default) is today's behaviour.
What travels is the store *path*, as a plain string literal, not a flake
input: containers share the host's `/nix/store`, so the build is already
reachable inside them with its whole closure and has nothing to travel.
An input would be worse than useless — a `path:/nix/store/<pkg>` input
is re-copied as a reference-less `-source`, which strips exactly the
closure the binary needs.
The catch is that a path written into a generated flake is text, so
nothing in the container's closure keeps the binary alive. The host does
that instead, and gets it for free: the package is interpolated into
`/etc/hyperhive/serve.json`, `builtins.toJSON` preserves string context,
so the /etc entry references it and the system closure gc-roots it for
as long as that generation is the one the agents were rendered from. An
assertion pins that property, because losing the context is invisible at
eval and at deploy — it would surface only as every agent failing to
spawn `claude` whenever the next gc ran.
Container side wraps the path in a symlink farm rather than putting it
on PATH directly: `systemd.services.<name>.path` and
`environment.systemPackages` both coerce a store-path *string* through
`lib.toDerivation`, i.e. `builtins.storePath`, which pure evaluation
rejects. Interpolating the path into a builder is just text and
evaluates anywhere. `claude-code` drops out of systemPackages when a
pin is set, so there is exactly one claude in the container.
Refs #2693
`Template` was a DAG-level enum that three different things read back
out: `terminal_hook()` mapped it to a side effect, the retention pass
bucketed history by it, and a tracing field printed it. None of those
needed a *label* — they needed the two facts the label happened to
encode. So the enum was a lossy stand-in for intent, and every new DAG
shape had to pick the variant whose inferred behaviour matched, whether
or not the name fit (`reparent` rode `MetaUpdate` for exactly this
reason, with a 10-line comment apologising for it).
Replace the inference with a declaration: `DagSpec.hook:
Option<HookKind>`. Only the builder assembling a DAG knows why it did
so, so only the builder can say what should happen when it settles.
`run_terminal_hook` becomes a field read, and `reparent`'s apology
becomes `hook: None`.
Hook assignment is byte-identical to the old precedence rule
(`approval_id.is_some()` wins, then `Rebuild | PermChange`), checked
site by site; `meta_update` is the only builder with a variable
approval id and so the only remaining conditional.
Retention loses the per-template bucket with the enum that keyed it.
The dashboard renders one recent-builds list, so one flat newest-first
cap (`MAX_HISTORY_DAGS`) bounds it. `HISTORY_GRACE_SECS` goes too — it
existed to stop a burst of same-template DAGs evicting each other
inside one poll interval, which is not a failure mode a flat cap has.
That takes `snapshot_capped()` and the `snapshot_no_grace()` test hook
with it.
The queue is runtime-only (empty graph on boot), so the serde changes
carry no migration risk.
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).
Encoding "not configured" as weight 0 worked (the writer omitted the
line) but the type lied: 0 is not a legal cgroup v2 weight, and every
reader had to know the sentinel. Use Option<u32> end to end instead —
wire type, priv_client, HiveEnv, drop-in writer — so "unset" is a state
of the type rather than a magic value.
The nix options become nullOr, keeping their default of 80; null now
expresses "leave the setting out of the drop-in entirely" declaratively,
which is the useful shape on a host whose IO scheduler ignores
io.weight anyway.
Backward compat is unchanged: the fields stay #[serde(default)], so a
request from an older hive-c0re deserialises to None and reproduces the
pre-weights drop-in byte for byte. The test that pins that now passes
None instead of 0.
`CPUQuota=`/`MemoryMax=` are hard caps: they throttle an agent even when
the host is idle, so they are the wrong tool for "be polite under
contention". The cgroup v2 relative shares are, and neither was wired.
Adds `services.hyperhive.{agentCpuWeight,agentIoWeight}` (1..=10000,
default 80) threaded through the existing drop-in path: HiveEnv ->
write_dropins -> WriteResourceLimits -> hyperhive-limits.conf, next to
the caps already there. Hive-wide only, as the operator scoped it on the
issue: no per-agent override, no resource-limits.json field, no
dashboard form.
The default of 80 is below the kernel's 100, so agent containers yield
to everything *not* on this drop-in path -- host services and the infra
containers (hive-ci, hive-forge, hive-gateway, hive-matrix). It does not
rank agents against each other; they all carry the same weight.
`WriteResourceLimits` gains two `#[serde(default)]` fields, and the
writer treats weight 0 as "not configured" and omits the line, so an
older hive-c0re talking to a newer hive-priv still produces the exact
pre-weights drop-in. The body is extracted into `limits_dropin_body` so
that is covered by a test rather than asserted by eye.
`Coordinator::set_paused` wrote the marker directly with `std::fs::write`
from hive-c0re, which runs as the unprivileged `hive-core` user. The
agent's harness dir is chowned to the agent user on every container boot
(`user.nix`'s activation chown), mode 0755 — so hive-core can stat the
marker but gets EACCES creating or unlinking it. Pause therefore only
ever worked on an agent that had never booted; the read side works
because a stat needs traverse, not write, which is why the paused pill
and `is_paused` looked healthy.
Route both directions through hive-priv, the root helper that already
owns the other writes into agent-owned directories:
- `PrivRequest::SetAgentPaused { agent_name, paused }`, with the marker
filename constant moved to hive-priv-sock. That is the narrowest crate
all three sides share (hive-priv deliberately does not depend on
hive-sh4re, which re-exports it for the in-container resolver). A
private copy on any one side would break pause silently, since every
reader just sees "no marker".
- `write_agent_state_file` generalised to `write_agent_dir_file`, taking
the target directory: `state/` and `harness/` are both agent-owned,
which is the same reason both need root.
- resume unlinks via `remove_file`, which acts on the leaf and never
follows a symlink — an agent could otherwise plant a link at the
marker path and have root delete an arbitrary file.
`Coordinator::set_paused` becomes an async round-trip; its three call
sites were already async. Both directions stay idempotent because the
dashboard toggle and `hivectl pause|resume` fire without reading the
current state first.
hive-c0re, hive-claude, hive-forge and hivectl each ship a README.md but
never declared it in their package manifest, so cargo/docs.rs metadata
did not pick it up. Every other workspace member already sets the field;
this closes the gap left after the README backfill.
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 revert hook is dead by construction, so it can only ever be wrong.
DAG state `Cancelled` has exactly one producer: `JobQueue::cancel`, which
refuses unless every work node is still `Pending`. A cancel *cascade*
(some node failed, downstream cancelled) rolls up `Failed` instead —
`dag_rollup` short-circuits on any failed subtree node. So on a DAG that
reaches `Cancelled`, no node ever executed: the `SetWanted` head provably
never ran and `wanted` still reads whatever the operator last set it to.
There is therefore nothing to revert, and `revert_intent` did not revert
anything — it wrote `Wanted::from_running(observed)`, i.e. the agent's
*observed* state, over an intent the DAG never touched. Harmless when
observed already matched, silent corruption otherwise: cancel a queued
start for an agent that is down but `wanted = Up` (crashed, or caught
mid-bounce) and the intent flips to `Offline`, leaving it
deliberately-stopped as far as reconcile and crash-watch are concerned.
The hook made sense when `set_wanted` was a pre-submit side effect
written before the DAG ran; moving it into the DAG as a node left the
hook vestigial.
Drop `HookKind::RevertIntent`, `revert_intent`, and the power-op arm of
`terminal_hook` — start / stop / graceful-stop now settle with no
terminal hook, same as restart always did. The test asserts the general
statement across restart/stop/start x graceful x running: stop and start
carry a `SetWanted` head, and cancelling them still fires no hook.
`terminal_hook` mapped `Restart` / `GracefulRestart` to `RevertIntent`, but
a restart never writes `wanted` — `restart_chain` deliberately has no
`SetWanted` head, so the tail `Reconcile` converges to the agent's existing
intent and a deliberately-stopped agent isn't forced up by a hive-wide
restart.
`revert_intent` writes `Wanted::from_running(observed)` unconditionally on a
cancelled DAG. So for an agent that is `wanted = Up` but currently down
(crashed, or caught behind another queued op), submitting a restart and then
cancelling it writes `wanted = Offline` — reverting an intent the DAG never
touched, to a value nobody asked for. Reconcile and crash-watch both then
read the agent as deliberately-stopped and leave it down.
It's invisible for a running agent, since `from_running(true)` equals the
intent already on file, which is why it went unnoticed. `cancel` only
succeeds while every node is still `Pending`, so the reachable window is
exactly "queued restart + observed != intent" — precisely when someone
restarts and then thinks better of it.
Drop both restart templates from the `RevertIntent` arm; they fall through
to no terminal hook, which is correct for a DAG that writes no intent.
Document the invariant on `HookKind::RevertIntent` and on `revert_intent`
itself: the hook writes *observed* state, so dispatching it for a template
with no `SetWanted` head doesn't restore an intent, it invents one.
Test covers all four restart shapes (graceful x running), asserting both
that the spec carries no `SetWanted` and that a cancelled restart dispatches
no hook, with a contrast arm pinning stop's revert in place.
Fixeshyperhive/hyperhive#2710
Adds CPU/memory cap columns and an inline edit form to the container-load
table in /core.html, backed by a new POST /api/resource-limits/{name}
dashboard endpoint.
## Backend (hive-c0re)
lifecycle_ops.rs — new post_resource_limits handler:
- Parses ResourceLimitsForm { cpu_quota, memory_max } (both optional; empty
string = clear override, fall back to hive-wide default).
- Validates each non-empty value via resource_limits::validate_cpu_quota /
validate_memory_max — returns 422 UNPROCESSABLE_ENTITY with a human-
readable message on invalid input so the dashboard can surface it inline.
- Calls meta::commit_resource_limits (staged git write under META_LOCK, same
as hivectl set-limits).
- Re-applies the drop-in immediately via lifecycle::write_dropins so the new
ceilings take effect on the next container start without waiting for a
rebuild.
- Triggers rescan_containers_and_emit so ContainerView.cpu_quota/memory_max
update via SSE without waiting for the next periodic sweep.
dashboard/mod.rs — registers the route:
POST /api/resource-limits/{name}
## Frontend (core.js + system-sections.css)
core.js:
- containersState derived from /api/state snapshot alongside tombstonesState
— supplies configured cpu_quota/memory_max to the LOAD table.
- lastLoadRows stash lets SSE-triggered re-renders call renderContainerLoad
without waiting for the next 5s poll.
- renderContainerLoad: adds cpu cap / mem cap columns (muted; tooltip
'configured ceiling — takes effect on next start') sourced from
ContainerView, plus a per-row S3T toggle button that expands an inline
edit form with cpu_quota / memory_max text inputs and a S4V3 button.
The edit form shows a restart hint, surfaces validation errors inline, and
collapses on success.
- container_state_changed SSE handler: updates containersState in place and
re-renders the LOAD table so the cap columns flip immediately after a save.
system-sections.css:
- CSS for the new cap columns (.cload-cap-th, .cload-cap) and inline edit
form (.cload-edit-row, .cload-edit-form, .cload-edit-label, etc.).
- Remove dead .rqe-step rule (step sub-step label retired from the wire in
'job_queue: retire the now-off-wire step sub-step label').
A config deploy could silently discard committed agent config. sock's
icon commit carried a full proposal/approved/building/deployed tag set
yet was not an ancestor of `main` — genuinely deployed, then dropped.
Two gaps compounded.
`prepare_applied_target` is documented as "fast-forward applied/main to
target", but `git_update_ref` is `git update-ref <ref> <target>` with no
old-value guard: an unconditional force move. Anything reachable from the
old `main` but not from `target` leaves the branch without a word.
And nothing checked that it *was* a fast-forward. `run_deploy_merge_verify`
asserts exactly one thing about history — that the live PR head still
equals the reviewed sha. That is a drift gate on the *head*; it says
nothing about the *base*. A PR opened from a stale base passes it
unchanged and then rewinds `main` when it lands.
Adds, in the order they run:
- an ancestry gate as step 3 of MergeVerify — the reviewed head must
descend from `applied/main`, else bail before the irreversible merge.
It sits after the fetch (the commit has to be local to check
reachability) and before the eval, so it stays inside the region where
nothing is mutated and the node is still safely retryable.
- `git_update_ref_cas`, used for the `applied/main` move.
`git update-ref <ref> <new> <old>` refuses, and leaves the ref alone,
when the current value is not `old`. The ancestry gate only proves the
target is safe against the `main` observed *then*; the CAS is what
keeps that proof true *now*. `run_deploy_apply` already reads
`prev_main` to park the rollback ref, so that value is threaded in —
re-reading it inside the callee would reintroduce the race.
`git_is_ancestor` returns `Ok(false)` for exit 1 rather than treating
"not an ancestor" as a failure. Its doc comment notes this is not the
"did this branch land upstream" question: a squash-merge rewrites the
commit, so `--is-ancestor` correctly answers false for a branch whose
contents were merged. Different question, same command.
Tests cover both directions of the ancestry check, and that a stale CAS
both errors *and* leaves the ref where it was — a guard that fails while
still moving the ref would be worse than none.
Not covered here, deliberately: the non-PR apply path also writes `main`
and wants the same treatment. Kept separate to stay reviewable.
The `step` label was taken off the wire in #2661, when each deploy phase
became a first-class DAG node. Since then it has been written but never
read: `NodeRuntime` derives only `Debug, Default, Clone` — no serde — so
the field could not reach any client, and the only reads of it were the
dedup checks inside its own setters. This deletes the machinery.
Removed:
- `NodeRuntime.step`, `set_step`, `set_step_running`, and the
`rt.step = None` clear in `complete_node`. `NodeRuntime` keeps its
remaining `build_log_id` field (deliberately still a struct — collapsing
it to a bare `Option<i64>` would churn every call site for no gain).
- `Ctx::step` and its ~15 call sites in `job_queue/exec.rs`. `Ctx` itself
stays: it is the build-log sink, which `run_prebuild` and `run_swap`
still use.
- `Coordinator::set_queue_step` and its 11 callers in `actions.rs`.
- `JobQueue::running_node_of`, reachable only from `set_queue_step`.
- `swap_update`'s `on_step` parameter and its one body call.
- The `set_step_only_on_running_and_signals_change` test.
Dropping the calls orphaned parameters, which are removed with their call
sites: `ctx` on ten executors that used it only as a step sink, and
`queue_entry_id` on `run_deploy_merge_verify` / `run_deploy_apply` /
`run_finalize_deploy` plus both `coord` and `queue_entry_id` on
`prepare_applied_target`. `run_deploy_tail` KEEPS its `queue_entry_id` —
that one has a genuine surviving use (the build-log link in the failure
comment posted to the PR).
One behavioural change, called out so it is not mistaken for a dropped
dashboard refresh: `Ctx::step` and `set_queue_step` each emitted a
`rebuild_queue_changed` snapshot when the label changed, and those
emissions go away with them. This is safe — the snapshot payload has no
step field, so those pushes carried nothing a client could observe. Real
state transitions still emit from the scheduler's claim and completion
paths, from `submit`, and from the three `actions.rs` sites. Net effect is
strictly fewer redundant SSE pushes.
Docs: `docs/coordinator.md` still listed `step` as a `NodeView` wire field
and `docs/web-ui/dashboard.md` documented a cyan `↳ <step>` sub-line under
each queue row. Neither has existed since #2661 — both corrected here, plus
the `job_queue/model.rs` module doc.
Not touched: `frontend/packages/dashboard/src/system-sections.css` has a
dead `.rqe-step` rule with no JS referencing it. Left for the frontend
owner rather than deleted here.
Closes: #2664
`container_view::build_all` renders every container on every scan, and
each agent's row resolved its limits through `effective()`, which reads
and parses `resource-limits.json` from disk. That is one file read per
agent per scan of a file that is identical for all of them.
Split the resolution in two: `effective_from` takes an already-loaded
map, and `effective` keeps the read-then-resolve shape for the
single-agent callers (`write_dropins`, which runs once per spawn and
rebuild and has no map to hand). `build_all` now loads the map once at
the top — the same treatment `topology::read()` already gets there —
and calls `effective_from` per agent.
No behaviour change: the fallback matrix lives in `resolve`, which both
paths still go through, and its tests are untouched.
The now-redundant `limits_for` is gone; `effective_from` covers its one
caller.
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
When ContainerView.paused is true, show a clickable yellow `⏸ paused`
badge on the agent card that POSTs to the new /api/resume/{name} endpoint
to un-park the turn loop. The badge doubles as the resume button so the
state is self-documenting and one click to fix.
The agent action menu gains ⏸ P4US3 (when not paused) and ▶ R3SUM3
(when paused), orthogonal to the running/stopped start/stop actions.
On the backend, /api/pause/{name} and /api/resume/{name} POST routes
wire to Coordinator::set_paused and trigger an immediate rescan so the
badge flips via the existing SSE ContainerUpdate without polling.
Depends on the ContainerView.paused field and Coordinator::set_paused
added in the parent PR.
Add NodeView::parent to the wire (hive-sh4re + hive-c0re dag_view), then
render the recursive parent/child tree in the dashboard build queue instead
of the previous flat chain/fan-out layout.
Wire change (hive-sh4re, hive-c0re):
- NodeView gains parent: Option<NodeId> (skip_serializing_if = None)
- dag_view() projects node.parent, filtering out the Dag container id
(top-level work nodes become parent: None on the wire)
Frontend (builds.js):
- Replace nodeComponents + splitFanOut with buildNodeTree (uses parent
edges directly) + topoSort helper (orders siblings by deps)
- renderTreeNode walks the tree depth-first, rendering indented rows
with └─/├─ connectors and agent label per chip
- Flat chains and fan-out heuristics are gone; structure comes straight
from the scheduler's parent axis
Closes: none (parent issue tracked in forge)
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
