A new swarm service, swarm-notify-router, in its own container on the host
that sets deploy.notify-router.enable (defaulted on by allSwarmServices where
the host has the router's store leaf, like the claude login minter).
- swarm-controller registers instance-wide hooks for `issues` +
`issue_assign` and `pull_request` + `pull_request_assign`, updates the
event list of a hook already registered with fewer (leaving its active flag as it is), and keeps each
verified body unchanged in the SWARM_EVENTS stream on
$SWARM.events.forge.<X-Forgejo-Event-Type>, deduped by X-Forgejo-Delivery.
A delivery the stream did not keep answers 503.
- swarm-notify-router drains its durable consumer on that stream, runs its
flows (source -> match -> todo sink, stored as nodes plus edges and checked
to be a tree) and queues each op on $SWARM.router.todo.<agent> in the
work-queue stream ROUTER_TODOS. Built-in flows: assigned, unassigned (same
key, rewritten) and closed (prefix-clears every holder's router todos). A
per-issue snapshot in its notify-router-state bucket holds the last
assignees and the todo holders; only agents in the agent-status bucket get
ops, and an assignee that was not one is left out of the snapshot so the
next assign event retries it. After the first event that must be retried,
the rest of the batch is nakked so redelivery keeps stream order.
- hive-agent creates its own durable consumer todo-<agent> and applies each
op to its todo store as subsystem `router`, then acks; a refused op stops
the rest of its batch the same way. The store's (subsystem, key) dedup
makes a redelivery harmless. New Todos::clear_prefix.
- swarm-nats-auth: a --notify-router-client arm; the controller may publish
$SWARM.events.forge.*; each agent gets three {agent} templates (extended
consumer create whose filter the server checks, MSG.NEXT, ACK).
- Credentials: a notify-router store leaf minted by glue-bao-tls, a
swarm-notify-router bao policy (read on its own queue client secret), an
authelia machine client, the secret publisher entry, a journal forwarder
route. No manual step.
hive-forge-notify is unchanged, so agents also still get forge's own
notifications.
Refs #5571
per #5471, verifiers stop depending on the queue client. the module is
store-free and needs only thiserror, so swarm-agent-verify gains one
dependency and loses its swarm-queue-client edge. consumers import it
from swarm_agent_verify; swarm-logs and swarm-metrics, whose only use of
swarm-queue-client was this module, drop that dependency, as does
swarm-logs-auth.
The agent's state/AGENTS.md stays the truth and its harness the only
writer. The harness mirrors the file into a new `agents-md` KV bucket,
and swarm-controller serves that copy, so a read works while the agent is
stopped. An edit is a unified diff plus the revision it was made against,
sent over the agent's existing control subject; the harness refuses a
stale base, applies the diff with `git apply --check` then `git apply` to
a temp copy, and writes the file only when both pass. Fixes#5519.
Files:
- swarm-queue-client/src/agents_md.rs, lib.rs: the bucket name and
subject, the mirror value, the `ApplyAgentsMd` control command and its
reply. `AgentControl` loses `Copy` because the new variant carries text.
- hive-agent/src/swarm_agents_md.rs, main.rs: the mirror task (start,
10 s re-read, immediate republish after an edit) and the apply.
- hive-agent/src/swarm_input.rs: the control loop answers both commands.
- hive-agent/src/claude_md_watch.rs: its AGENTS.md path is shared.
- hive-agent/Cargo.toml, Cargo.lock: sha1 for the git blob id (already in
the lock), tempfile moves from dev to runtime for the apply directory.
- swarm-controller/src/agents_md.rs, main.rs: the bucket reader and
GET/PUT /api/agents/{name}/agents-md; test AppState literals get the
new field.
- swarm-controller/src/agent_control.rs: the sink hands back raw reply
bytes, so each command decodes its own reply type.
- swarm-controller-api/src/agents_md.rs: the route bodies, shared with
swarmctl.
- swarmctl/src/agents_md.rs, main.rs: `swarmctl agent agents-md get|apply`.
- nix/host-modules/swarm-nats.nix, nix/module-eval/nats-authelia.nix: the
agent's publish grant on its own `$KV.agents-md.<agent>` key.
git needs no new package: agent containers enable programs.git, and the
harness unit's PATH includes /run/current-system/sw.
The agent read the operator-profiles bucket with a keys() listing, an
ordered consumer whose grant (STREAM.INFO, CONSUMER.CREATE) and flow-control
reply had never run against a permissioned queue. The controller now writes
the whole list as one JSON array under one key, only when it changed, and
the agent reads it with one DIRECT.GET the way agent_inference::read_own
reads its key. The agent grant shrinks to that one subject.
swarm-controller reads each member of authelia's operator group's forge
profile every five minutes and keeps one key per operator in the new
operator-profiles bucket. Every agent may list and read that bucket, and
hive-agent renders it into the system prompt at harness start, falling
back to today's prompt when the bucket is empty or doesn't answer within
five seconds.
Fixes#5383
hh-swarm-queue-connect-failed matched "swarm queue connect failed; this
agent publishes nothing upward", which nothing has logged since the
harness started retrying its first connect and exiting (235bc481b). The
rule now matches the error the harness logs when every attempt has
failed, CONNECT_EXHAUSTED in swarm_queue.rs, and its comment and summary
describe the retry-then-exit behaviour.
A failed MintAgentIdentity node logged nothing: the error reached only
the job graph. mint_identity now warns with the agent and the error, and
hh-agent-identity-mint-failed fires on that line.
Refs #4651
mara on #5170: "dont split caps and perms. its all perms. / dont include
the now empty tool groups / nothing set -> default / also add perms to
agent creation screen".
- hive-sh4re: one `Permission` enum (messaging, meta, inbox, execution,
web_tools, manage_root_agent) replaces `ToolGroup` and `Capability`;
`lifecycle`/`forge` are gone. `Permission::default_for` is the default
for an agent the swarm sets nothing for.
- The record is `{ "permissions": [...] }`. Missing key, unset field or
no successful read yet = default; `[]` = none; a failed read after a
good one keeps the last value.
- hive-agent reads its own key every 10s and builds each turn's --tools,
--allowedTools and MCP config (rewritten before every turn) from it;
the subagent daemon reads it on each start/continue. Agents get a
read-only grant on their own key.
- hive-c0re restarts an agent only when its manage_root_agent changes.
The import now merges meta/capabilities.json and meta/tool-groups.json
into one list per agent and deletes both; the "import pending, don't
read" decision is `ImportOutcome::read_allowed`, table-tested.
- swarm-controller: `GET /api/permissions` (names + default) and an
optional `permissions` on `POST /api/agents`, stored before the job is
queued. swarm-ui's create-agent form gets a permissions field.
- Deleted: HIVE_TOOL_GROUPS and its forward to the subagent daemon,
tool_groups.rs, ruth's tool-group seed, the permissions rebuild job
(PermPayload/WritePermFile/perm_change), the dashboard permissions tab
and its stale-entries panel.
Part of #5170.
The per-agent web UI no longer sets the model: the `/model` slash
command, the model picker, `POST /api/model`, `Bus::set_model` and the
`hyperhive-model` writer are gone. The model chip is display-only and
shows where the model comes from (`model_source` in `/api/state`). The
model is set on the swarm (swarm UI agents page, `swarmctl agent
inference set`).
The read side is unchanged: with no swarm value an agent runs its
agent.nix model, else an existing `hyperhive-model` file, else haiku.
Existing files are not touched.
With the picker gone, its model list goes too: `available_models` in
`/api/state`, the ACP model-list filter, and the harness's global
`HIVE_AVAILABLE_MODELS`. The subagent daemon keeps its own
`HIVE_AVAILABLE_MODELS` from mcp.nix, and the `availableModels`
assertion stays.
`/effort` (command, picker, route, file) is unchanged.
Refs #5034
The header on `$SWARM.agent-state.<agent>` went out only when a bus event
changed it, and the core subject keeps nothing for a late subscriber. An
idle agent therefore said nothing to a reader that attached after its
last change: swarm-controller after a restart showed `effective: null`
for that agent until its next turn, which can be hours.
The publish loop now also wakes every 60 s and re-sends the current
header through the same `publish` path (clearing the dedup state first).
Transitions still go out as they happen. Same subject, same grant.
An agent whose queue connection is down now logs `swarm agent state:
publish skipped` once a minute even while idle, so the
`hh-swarm-agent-state-publish-skipped` alert also fires for idle
disconnected agents, not only busy ones.
Refs #5034
swarm-controller keeps each agent's inference settings in a new
`agent-inference` KV bucket, one key per agent, and is its only writer
(`GET`/`PUT`/`DELETE /api/agents/{name}/inference`; `swarmctl agent
inference get|set|clear`). Every field is optional. Each agent reads its
own key every 10 s through a direct-get grant on that key alone, and a
set `model` or `effort` replaces the Bus value, so the next turn runs on
it and the badge follows.
The swarm value is applied without persisting it: `Bus::set_model` and
`set_effort` are split into a persisting setter and a non-persisting
`apply_*`, and the poller uses the latter. Otherwise the swarm value
would overwrite `hyperhive-model`/`hyperhive-effort`, and clearing it, or
a restart, would no longer find the agent's own value. An unset field
changes nothing; a field cleared after being applied goes back to the
local value (`harness_state::local_model`/`local_effort`, the same
resolution `Bus::new` uses), read again at the clear.
`subagent_models` is validated, stored and shown by `get`; no agent
applies it yet.
Refs #5034
swarm-controller keeps every scheduled prompt (id, name, agents, body,
anchor, interval, paused) as one document under one key of a new
`schedule-definitions` KV bucket, so each change is a single
compare-and-set and a prompt exists for all of its agents or none. It
writes each agent's key of a new `agent-schedules` bucket as the prompts
naming that agent, without names or other agents, after every change and
on a one-minute reconcile pass. It is the only writer of both, behind
GET/POST /api/schedules and PATCH/DELETE /api/schedules/{id}.
`swarmctl schedule add|list|edit|pause|resume|rm` drives those routes over
the controller socket.
An agent's own queue credential is granted one more subject, the direct
get of its own key in `agent-schedules`. A KV watch is not granted: its
consumer create subject carries no key, so it would cover every agent's.
hive-agent's new schedule ticker reads that key every 10 seconds and fires
due slots as the same `schedule:<id>` todo hive-c0re's worker pushes. Fire
times are anchor + k*interval; the last fired slot per schedule id is kept
in the state db, so a body edit or list refresh never moves or refires a
slot, downtime fires the latest missed slot once, a restart refires
nothing, a one-shot fires once, and paused slots are skipped.
Refs #4914
AgentSession becomes hive_runtime::AgentRuntime, picked at startup from
the environment. Unset HIVE_RUNTIME keeps the claude backend, built
from the same title, store and PercentPolicy as before; drive_turn,
the 401 retry and the error mapping see the claude errors unchanged.
On acp the harness keeps the session id in the harness dir, answers
permission requests like the claude built-in allow-list (no built-in
shell, web fetch only with web_tools), maps runtime errors to Failed,
and reports an operator /compact as skipped instead of done.
Refs #4391
The harness now reads swarm/agents/<agent>/queue from the store itself,
under the agent's own store certificate, and holds it in memory only.
It reads once before the first connect and again on every reconnect
attempt (async-nats `ConnectOptions::with_auth_callback`), so an agent
whose secret was re-minted reconnects with the new value instead of
being refused until the container restarts.
hive-agent-queue-credential.service, the /run file it wrote, and
HIVE_AGENT_QUEUE_AGENT_SECRET_FILE are gone; queue-identity.nix now
hands hive-agent.service the store address, its certificate paths and
the agent name.
A failed or empty read before the first connect still falls back to
the hive's shared client. Each read is bounded by a 10s timeout, and
retries wait out the existing reconnect backoff (500ms doubling, capped
at 60s).
Closes#4783
The auth callout grants an agent that presents its own queue credential
one more subject, `$KV.agent-icons.<agent>`: its own key in the
agent-icons bucket and no other. The hive's shared agent client is
granted none of the bucket, since every agent on a hive presents it.
hive-agent writes `/etc/hyperhive/icon.svg`, the file its `GET /icon`
serves, to that key once per start, as a JetStream publish straight to
the subject (what `kv::Store::put` sends, minus the bucket lookup), so
the one subject is the whole grant. No icon deletes the key. A failed
write, including one that arrives before the bucket exists, is retried
with backoff until acked. An agent connected with the hive's shared
client publishes nothing.
swarm-controller creates the bucket as soon as its queue connection is
up, instead of on the first icon read, so an agent's write does not
wait for someone to look.
Measured against a local nats-server with a user allowed publish on
`$KV.agent-icons.atlas` only: the write to its own key is stored and
readable, a write to `$KV.agent-icons.argus` is refused (the ack times
out), the DEL marker makes the key read as absent, and a write before
the bucket exists fails with "no responders".
A record written to stdout carries no priority, so journald files the
whole stream at one level and the swarm log store shows `info` whatever
level `tracing` gave it. Under a systemd unit the process's stdout
already *is* the journal, so the fix is to speak the journal protocol
directly and let each record carry its own severity.
New `hive-log` crate holds the one sink chooser, called by `hive-c0re`,
`hive-agent` and `swarm-controller`. It builds the same `EnvFilter`
those binaries always built, then installs exactly one layer — never
both, since a journald layer stacked on the `fmt` layer under a unit
stores every record twice.
The choice is an fstat compare, not a presence test: a child inherits
`$JOURNAL_STREAM` even when its own stdout was redirected elsewhere, so
the variable existing proves nothing. The crate parses `dev:inode` out
of it and compares both numbers against an fstat of stdout, the
descriptor the `fmt` layer writes to by default. No match, unset, or
unparseable takes the `fmt` branch. A journald layer that fails to
construct despite a match falls back to `fmt` and warns through it —
a process must never fail to start because of its logger.
The harness has had its queue coordinates since the credential reached
the container, but nothing used them. This offers each terminal row
upward on `$SWARM.term.<hive>.<agent>`, so a swarm-level terminal can
render an agent without reaching into the hive that hosts it.
It publishes the same `TermMsg` the web UI is handed rather than a
second model of the same events, so a new tool or a reclassified event
changes both surfaces together. It subscribes to the event bus rather
than to the SSE handler: the handler classifies per connected browser,
so hanging this off it would mean an agent nobody is watching publishes
nothing. That also means its own long-lived `ClassifyCtx`, since a
publisher restarting its correlation state would lose the `tool_use` →
name mapping a `tool_result` needs to render.
The hive in the subject is derived from the queue client id, not from
the harness's hive display name. Those come from different sources with
no rule tying them together, and the responder builds its grant from the
client id — so deriving it from the display name yields a publish the
broker refuses, reaching an operator as a terminal that is merely empty.
The prefix and suffix that bracket the hive are the responder's flags,
which the agent is not told; it restates their defaults, and the symptom
of a deployment retuning one without changing this is every publish
refused rather than a wrong subject accepted.
Oversize rows degrade in the publisher. Exceeding `max_payload` is not a
truncation: the server refuses the message and closes the connection, so
an oversize publish costs the row, the connection, and the rows racing
behind it through the reconnect. The body is the only unbounded field —
summaries are already trimmed at classification — so it is the field
spent, and the row keeps its icon, level, summary and coalesce key. A
row that does not fit even then is logged and dropped rather than sent.
The limit is read off the connection, so `8388608` stays spelled once in
the queue's own module; size is measured by serializing, because JSON
escaping separates character count from wire length by an unbounded
factor on exactly the rows already near the limit.
Best-effort throughout: no queue, an unparseable client id and a failed
connect each disable the publisher with one log line, and a failed
publish loses its row and nothing else. The turn loop and the web UI
never block on the queue.
Refs #3805
hive-c0re stats the two files `swarm-bao-queue-agent` lands and forwards
them into every agent container as systemd credentials, and the harness
resolves a `QueueConfig` out of them at boot. Nothing connects yet.
A credential and not a bind mount, and the mode is what forces it: the
secret is root:0600 and the harness runs as the unprivileged agent user,
so a bind would deliver a file that user cannot open. nspawn's
`--load-credential` is read by the container manager as root and
re-exposed under the consuming unit's own `User=`. hive-c0re never reads
the bytes either way, which is just as well — it runs as `hive-core`.
Absent files stay legal and become visible rather than silent: the
publisher lives on the authelia host and mints on its first boot, so
"nothing at that path" is the ordinary early state of a swarm. c0re
forwards nothing and logs why; the harness logs that it has no queue.
The client id comes out of the delivered file rather than being rebuilt
from `hiveName` in nix, which is the agreement `swarm-secret-client`
states. `QueueConfig::from_env` wants it as a value, so the harness reads
the file itself — assigning the variable instead would need
`std::env::set_var` in a process that has already spawned threads.
Refs #3805
Part of hyperhive#1898 (b): every docs subdir should have a top-level
README.md link, achieved by moving/renaming where an existing file
already fits the role.
docs/turn-loop.md already served as the hub + index for the three
sub-pages under turn-loop/ (claude-invocation.md, config.md, mcp.md),
so it moves wholesale rather than leaving a redundant top-level
pointer stub. Fixes every inbound/relative link across the repo
(top-level README.md, CLAUDE.md, docs/persistence.md,
docs/tools/scheduling.md, the sub-pages own back-link, hive-agent
README + doc comments, hive-agent/Cargo.toml, .prettierignore per-file
exemption entry) - grepped the whole tree for both turn-loop.md and
turn-loop/ to find every reference rather than trusting a partial
list.
nix fmt clean, cargo check -p hive-agent clean.
Six places in the tree hand-rolled the same connect / write one JSON
line / read one JSON line back. Two of them — the harness serve loop's
client and the MCP server's — were byte-identical apart from a six-line
wrapper, ~145 lines of literal copy-paste. The other four each
reimplemented a subset, and the subsets had drifted: some named the
socket path in their errors and some did not, one classified transient
against fatal failures and the rest retried nothing at all, two drained
the response and two decoded it.
That duplication was defended when the daemons were split out, on the
grounds that a daemon's socket etiquette should stay visible in the
crate that depends on it. The etiquette genuinely does differ. The code
does not, and five copies is where "each daemon documents its own
etiquette" stops paying for itself.
`hive-sock-client` now owns the transport once, generic over the
request and response types so it is protocol-agnostic: the host-served
control socket and the harness's in-agent socket both use it with their
own wire-type crates. The two real differences become values instead of
forks. Retry is `Retry::RideOutRestart` (2/4/8/16/30s, sized to ride out
a service restart) for callers with no natural retry of their own, or
`Retry::None` for callers already inside a poll loop where the poll
interval is the retry — and the reason each caller picked one is a
comment at the call site rather than a reimplementation. The response is
either decoded (`request`) or half-closed and drained (`notify`, where
the drain exists so the server's write-back doesn't land on a closed
socket). Whether a failure propagates or is logged and swallowed stays
at the call site, because that is the caller's choice and not a property
of the transport.
Errors always name the socket path now, everywhere. That detail is
load-bearing: a permission problem on a socket that reads as "is the
daemon running?" sends the operator to fix the wrong thing.
The transient-against-fatal enum is gone rather than moved. Serialising
happens before the retry loop and deserialising after it, so only
connect, I/O and short-read failures can reach the loop at all — a
deterministic failure is now unretryable by construction instead of by
classification.
It is deliberately a new crate and not part of `hive-agent-sock`. The
`*-sock` crates are pure wire types by convention — `hive-agent-sock`
depends on serde and nothing else — and the two largest copies talk to
the host socket, whose types live in a different crate entirely. A
transport in either wire-type crate would drag tokio into it and point
the wrong way besides.
No wire-format change: same JSON line in, same line out.
The poller was a `tokio::spawn` inside the `hive-agent` serve loop. It
never needed anything from that loop except a socket path, so being
in-process bought nothing and cost two things: a harness restart took
forge notifications down with it, and the whole forge/HTTP dependency
tree was linked into the serve-loop binary.
It is now `hive-forge-notify`, a per-agent daemon with its own systemd
unit, a sibling of `hive-bash-daemon` and `hive-matrix-daemon`. Same
contract as those two: it reaches the harness only by upserting todos on
the in-agent socket, and nowhere else.
The module moves verbatim (`notify.rs`) — the formatters, the activation
gates, the dedupe map and all 33 tests are unchanged. Only the socket
call sites are rewritten, onto a small local `todo_client` rather than
the harness's. That mirrors what both sibling daemons already do, and
the etiquette differs on purpose: the harness's client carries a 60s
backoff schedule sized to ride out a hive-c0re restart, which its
callers need because they have no retry of their own. This poller's two
call sites both sit inside the 30s poll loop and both treat a failure as
"leave the thread unread, try next tick", so the poll interval already
is the retry; a second backoff would only stack sleeps and delay the
rest of the batch.
The unit is `Restart=on-failure`, not `always`. An agent with no forge
account is a supported configuration and the poller reports it by
logging why and exiting 0 — under `always` that clean exit would be a
restart loop on every forge-less agent.
`forgejo-api`, `url` and `time` drop out of `hive-agent`'s dependencies
with the module.
Also corrects docs that outlived the code they described: the persisted
`forge_cursor` field is long gone (forge's own read-state is the durable
record of what has been delivered), but `docs/persistence.md` and the
`harness_state` module docs still documented it as live.
Closes 2718.
The operator has been going through agent dirs by hand with ncdu,
deleting 20+GB target dirs. Agents had no way to know they were the
ones sitting on the space.
New `disk_watch` module in the harness: every 15 minutes it statvfs's
the filesystem backing the agent's state dir and, past 80%, raises a
keyed `disk` todo telling the agent to free space — with the operator's
rules inline: only delete things that are actually big, build output
first, and never delete something still needed, ask for more space
instead.
Over threshold it also walks the agent's own tree (`/agents/<label>`
plus `$HOME`) and names the directories worth looking at, so the todo
says where the bytes actually went rather than just that the disk is
full. The walk is bounded on every axis — entry budget, recursion cap,
report depth — pinned to the state dir's device so it can't wander into
`/nix` or the shared bind mounts, and it does not traverse symlinks. It
reports the deepest oversized directory on each branch, so the agent
gets pointed at `<workspace>/target` rather than at `/agents/<label>`.
Anti-nag is the whole design constraint. The todo is keyed, and the
summary is deliberately stable: the percentage is bucketed to 5 points
and no raw byte counts appear anywhere in it. An unchanged situation
re-upserts as `changed == false` and never fires the wake, so a disk
that has been steady at 89% for a week sits quietly in the loose-ends
list; crossing into a new bucket speaks up once. Dropping back under
the threshold clears the row.
Harness-local by construction, per the operator's call that this gets
no core wiring: hive-c0re cannot push a todo at all (the store and its
wake live inside the container), and running in-process means this
skips even the in-agent socket and calls `Todos::upsert` directly.
Worth recording, since it shaped the scope: btrfs does NOT fold qgroup
limits into statfs. Measured with quota counting enabled and a 20G
limit set on a real subvolume, statvfs returns byte-identical whole-FS
numbers for that subvolume, an ordinary agent dir, and the root. So
this watches host-FS pressure, which is valid before and after the
planned subvolume migration; per-agent quota awareness would need the
limit handed to the agent explicitly.
