mara: "should be https://auth.constellation.darkest.space/settings in
profile pic menu" -- the UserMenu link was pointing at the plain
authelia domain root, which lands on the portal rather than the
account settings page. Appends /settings client-side, same base-URL
source as before (GET /api/links Authelia entry).
Two review findings, folded together.
mara: a swarm integrated auto deployed forge always has metrics, so the
toggle is gone. The endpoint follows behindGateway instead, which is the
swarm-integrated shape and the condition the protected location lives
under. Serving it without that location would put it on a listener
openFirewall can expose with nothing in front.
argus: /metrics matched no access_control rule, so default_policy
one_factor governed it. That is any authenticated subject, which today
means any operator and tomorrow any agent. My audience argument covered
the Bearer path only; the same endpoint also accepts CookieSession, and a
cookie carries no audience at all, so the audience was never what stood
between a browser session and this data.
The rule is deny rather than a client-scoped allow because the collector's
client does not exist yet. Authelia refuses a subject naming an
unregistered client, and does so in a preStart validator rather than at
build time, so naming one early yields a green nixos-rebuild and dead
swarm SSO on the next restart. Denying until the client is registered
makes publishing the endpoint safe on its own; registering it is a
one-line change from deny to that allow.
Rule order is load-bearing: authelia takes the first match.
The gateway authenticates scrapers so services do not each grow a static
bearer of their own, but authelia's auth_request endpoint ran its default
strategies, which are cookie-only. A scraper's OAuth2 access token was
refused no matter how it was minted.
CookieSession is listed explicitly because authn_strategies replaces the
defaults rather than extending them. Omitting it evaluates, renders and
starts, and silently ends every operator session on the swarm UI, which
uses this same endpoint.
Unconditional rather than keyed to whichever service is scraped today:
this makes a scheme available, not an authorisation. Authelia refuses a
token carrying no audience for the requested URL and only issues a client
audiences it is registered for, so nothing passes until a client is
registered against a specific URL.
Forgejo can serve prometheus metrics but nothing turned them on, and
turning them on alone would have published them: forgejo serves
`/metrics` on its normal listener and the gateway vhost proxies `/` to
that listener, so the existing catch-all would have carried the endpoint
to anyone. The option is therefore one switch for both halves, and the
`= /metrics` location is an exact match so it outranks that prefix.
Authentication is the gateway's rather than forgejo's own `[metrics]
TOKEN`: a scraper presents an audience-scoped authelia token which nginx
checks via auth_request, so the swarm keeps one identity system instead
of gaining a static bearer per service.
The subrequest deliberately omits the `error_page 401 =302` that
swarm-ui uses. That redirect sends a browser to a login page; a scraper
would follow it and parse HTML as metrics.
Mechanical addition to the existing bulk-action framework
(renderSelectionBar/addBulkButton, swarm.js) -- two per-agent actions
already exist (POST /api/pause/, /api/resume/, see hive-agent-menu.js),
this just gives the selection bar the same all-or-nothing enablement
rule the other six bulk buttons already use (pause enabled only when
none of the selection is already paused, and vice versa for resume).
.btn-pause reuses .badge-paused's yellow so the trigger and the
resulting state pill read as one colour; .btn-resume is green like
.btn-start (both are "go" actions). Docs updated to list both in the
Selection bar reference.
Table gains an optional emptyMessage prop, rendered as a single
full-width row in place of a bare empty tbody; AgentsPage is the first
consumer. Screenshot-verified against a mock server returning an empty
roster.
No sitewide box-sizing: border-box reset exists, so .ui-dialog's
padding + border were adding on top of width: 90vw instead of being
carved out of it -- 401px rendered against a 390px viewport, real
horizontal overflow. Caught via a real 390x844 screenshot per mara's
"phone size?" ask on the PR, not assumed.
Three doc comments elsewhere in the tree still named the old
CreateAgentPage identifier/page framing after the rename in this
branch -- TextField.tsx, Panel.tsx, FormField.css (x3). None of these
files are touched by the rest of the diff, which is exactly how the
staleness happened.
Adds a generic ui/dialog/Dialog primitive (native <dialog>, no
third-party modal lib and no shadow-DOM custom element per the esbuild
gap on those) and wires a "+ agent" button into AgentsPage that opens
the existing create-agent form inside it, content unchanged from its
former life as a standalone /create-agent route/nav item.
Removes the "new agent" top-level nav entry and the /create-agent
route entirely -- creation now only reachable from the roster that
gets populated by it. CreateAgentPage.tsx/css renamed to
CreateAgentForm.tsx/css to match its new role as a mounted component
rather than a page.
Screenshot-verified the dialog open/closed states against a mock
server.
Adds an "/api/whoami" same-origin nginx proxy to authelia's own
GET /api/user/info (session-cookie authenticated, no swarm-controller
code needed) and a new UserMenu header component: a generated initials
avatar (first letter of display name, coloured from the same seven
base16 chromatic slots the nav accent already cycles through) opening a
popover with the signed-in name, a link to authelia settings, and log
out — both reusing the existing "Authelia" entry from GET /api/links
rather than a second source of the domain.
Per mara's call on the open avatar-mechanism question: initials now,
a real uploaded photo (authelia's settings UI implies pics are
settable) is an explicit future item, not blocking this.
Dashboard 2 of the set the operator asked for: how Claude is being used
rather than which agent is using it, so the axes are model, effort, token
type and query source. The only deliberate overlap with the agents page
is the cost/token headline.
Every panel was run against the live store before this landed. Three
panels were dropped rather than shipped, because their label has exactly
one live value today and a page of single-bar charts is the same silent
failure as an empty one.
Turn count and turn length are PROXIES and say so in their descriptions:
nothing exports turn stats, so a session record stands in for a turn,
which holds because each turn runs a new claude process.
New AgentsPage at /agents: fetches GET /api/agents (roster names) and
GET /api/config-prs (bulk config-PR status) and merges them into one
table, one row per agent. Reuses the existing Panel/Table/StatusChip/
RefreshIntervalPicker components exactly as HivesPage does — the roster
page and the config-PR panel turned out to be the same page rather than
two separate pieces of UI.
Adds a nav entry (green accent, the next unused base16 chromatic slot)
between hives and new agent.
The operator asked for this when the dashboard was first reviewed and it was
deferred, not declined: the container metric family carried no hive label, so
selecting a hive emptied every container panel while the "All" default hid the
problem completely -- a `.*` matcher matches series where the label is absent.
That family now carries hive and swarm, verified against the live store rather
than inferred from the fix having merged: a hive selection returns the same 7
agents as the All default, at every window out to 168h, with a nonexistent
hive returning zero.
The agent list chains off the hive selection, so picking a hive narrows the
agent dropdown rather than leaving entries in it that resolve to nothing.
Grafana served no dashboards: only datasources were provisioned, while
the module header already claimed dashboards were. This ships the agents
dashboard as a file provider and makes that sentence true.
The datasource uid is bound once and substituted into the dashboard at
build time. Committing the literal would make the dashboard a second
speller of a name the datasource already owns, and the drift failure is
silent -- panels render empty rather than erroring.
The shipped copy drops the `DS` datasource variable: it exists so an
operator can pick a store on manual import, and a provisioned dashboard
must not ask.
`EnsureAgentIdentity` answered `AlreadyExists` whether it had added the
agent marker to an existing subject or done nothing at all, and the
controller discarded the answer outright. So the one case worth telling
a human about — a subject that was NOT an agent a moment ago — could not
survive the socket, let alone reach a log.
`Healed` is a variant rather than a field on `AlreadyExists` because a
field is ignorable: adding a variant makes every existing match fail to
compile until its author decides what a heal means. That is the property
the old shape lacked.
The store cannot tell a pre-marker agent identity from a human operator
account created without a group — both are `groups: []`. So this is
either the intended migration or an agent joining a person's live SSO
account, and only the caller has the context to tell them apart.
Gated by state/gate-3549-heal.sh (8 arms + mutation): the mutation
collapses Healed back and reddens the discriminating arm while leaving
the anti-noise arm green. The W' control asserts exactly one warn in the
whole run, so a build that warned on every routine ensure would fail.
A bare String in a crate whose whole suite types validated names as
hive_types::Ident, so a consumer had to re-derive at the boundary what every
sibling field gets checked for free.
Ident is strictly narrower than the store's charset — [a-z0-9-] against the
[A-Za-z0-9._-] the users database allows, because that file holds humans too.
Every agent name is a legal Ident by construction (creation parses one before
the job is queued), so the narrowing costs real agents nothing. What it does
mean is that a human hand-added to the agent group cannot be described as an
agent: the reader omits that row and logs it, rather than failing the whole
roster or quietly shrinking the answer.
The REQUEST keeps its String. That side carries what a caller asked for, and
validating it server-side is what lets a bad name be refused with a message
instead of failing to deserialise.
The test named for pinning the external tag covered all but the newest
variant, and two doc comments still counted three outcomes where there are
four.
Folded the missing variant in rather than editing the number: a test that
claims to settle the wire shape and quietly omits one is worse than a
narrower one, because the next variant gets added with nothing to point its
author here. The counts are gone rather than corrected — prose that counts
its subject is falsified by every addition and is invisible to a grep for
the vocabulary that changed.
GET /api/agents, beside the POST that creates one. The identity store is
the roster rather than a source to assemble one from, so this is a read
with nothing to merge or reconcile.
It says nothing about health, deliberately. A roster is the set other views
are complete against — it is what makes "this agent has never reported"
expressible, and that only survives while the declared set and the reported
set stay apart.
The two verbs treat a missing bridge differently and the asymmetry is the
design: POST queues a job that fails loud when claimed, GET has nowhere to
defer to and returns 503. Answering [] there would render a store nobody
could read as a swarm with no agents.
The users database holds humans and agents in one namespace and nothing in
it said which was which, so a roster read had no predicate to read with.
Marks positively, at creation. The alternative — everyone who is not an
operator — fails in the direction that matters: an account created without
a group is an operator who cannot log in, a mistake the swarm UI docs
already warn about, and it would have rendered as an agent. The group is a
constant for the same reason the operator group it mirrors is one.
An identity that predates the marker gains it when agent creation runs
again, which is already the agreed migration for those; AlreadyExists
therefore reports that the subject was there, not that nothing was written.
ListAgentIdentities reads through this process because the store is owned
by a uid swarm-controller does not have — the same reason the write goes
through here — and answers with names alone, never the digests it sits next
to.
The wire crate's module doc and its README both said the bridge "owns both
users.json (canonical) and rendering users.yml internally". That store was
removed — swarm-authelia-bridge/src/store.rs is explicit that users.yml *is*
the store — so both sentences described a file that does not exist, in the
one place a reader goes to learn what the API is for.
The paragraph's actual argument (per-operation rather than
wholesale-replace) is untouched; only the mechanism it cites was wrong.
An agent name was checked for shape and never for meaning, so a name the
system already answers to could be taken. The list is derived, not invented:
operator/manager/root are broker recipients, and forge/reminder/schedule/
system/todo are message senders.
Senders are in it for a reason that is not cosmetic. The harness switches on
the sender name, so an agent called todo would have its messages read as
todo-wakes and one called system as helper events. That is a parse, not a
display quirk.
⚠️ The check is deliberately its own function rather than part of
validate_username, which runs over every user on every write — including
humans this bridge did not create. Folding it in would mean a store already
containing an operator named 'operator' could never be written again: a rule
about what an agent may be NAMED would have become a rule about what the file
may CONTAIN. A test pins that.
This is the static half only. The dynamic half — a name already taken by
another subject — waits on the heal-vs-refuse question open on #3550.
The scraper shipped with no targets, so nothing exercised it. This is its
first user, and the one with the least new surface: victoriametrics
publishes prometheus metrics on the listener it already serves queries on,
so there is no exporter, no extra port and no new reach — the collector's
otlphttp exporter already writes to that same loopback address.
Declared from this module rather than the collector's, per the option's own
rule: an entry exists only where the service that named it runs.
Nothing in this deployment read a Prometheus endpoint, so enabling
/metrics on a managed service added an endpoint and no data. The swarm
collector receives OTLP pushes and does not scrape; VictoriaMetrics
stores what is pushed and has no scrape config. The pipeline was entirely
push-based and every such service is pull-based.
Adds a prometheus receiver, a resource/swarm processor and a
metrics/swarm pipeline alongside the per-hive ones.
The pipeline is separate because that is the ruling, not for tidiness:
every resource/<hive> processor UPSERTS a hive key, so a scraped swarm
sample routed through one would acquire the single label a swarm-level
service must not have. Keeping it out makes the absence structural rather
than something to remember to strip, the same way hive stays a property
of which authenticated receiver accepted a push.
Targets come from an option each service fills in from its own module,
under its own enable, rather than a list assembled here. That is what
puts the scraper and the target on the same host by construction: an
entry exists only where the service that named it runs. Co-location is
true of the all-local deployment and is not a guarantee, and that is
exactly the case where assuming it is invisible.
All three additions MERGE with the per-hive attrsets rather than
replacing them. A plain assignment would drop every hive's receiver,
processor and pipeline and still render a config the collector starts
cleanly on.
Empty target set emits no receiver, no processor and no pipeline — an
enabled scraper with nothing to scrape is the inert configuration this
issue is about, and the target set ships empty here because the targets
themselves are separate issues.
Formatting verified with nix fmt. The evaluation gate is not written yet;
nothing here has been evaluated against a fixture.
The function's own doc comment still said one subject per hive, four
lines above an inline comment correctly saying one publish for all of
them. Left behind by the single-writer rework.
Swept the rest of the touched files for the old design's vocabulary
rather than fixing only the line argus quoted: every other per-hive
mention is about webhook registration or the callout policy's
extra-subject template, both of which are still per-hive and still
accurate.
Review call: 46 lines of documentation around a single constant, part of
it already stale. The worst paragraph explained why the earlier per-hive
shape had been justified wrongly — history of a design that never
shipped, written into the file within an hour of that design being
dropped. A file is not a changelog; why it was wrong belongs in the PR.
The constant moves to lib.rs beside the status bucket name, keeping only
the rationale that stays true: three crates must agree on the string, and
the one that agrees hardest speaks neither jetstream nor kv, which is why
it cannot sit behind a feature gate.
status earns a module of its own because it holds a bucket name AND the
functions that open it. This held a constant.
Review call: the event was addressed per hive — `$SWARM.events.<hive>.knowledge`,
published in a loop over the roster, granted through a wildcard. It does not
need to be. The payload is empty and the event means the same thing to every
hive, so one publish to one subject delivers exactly what N publishes to N
subjects did, and core NATS already fans out to whoever is subscribed. A hive
that was down misses it either way and reconciles on its next periodic pull.
That deletes rather than reshuffles: the roster loop, the wildcard, and the
shared subject-building function whose entire purpose was keeping the grant and
the publish from drifting apart. With one literal there is nothing to disagree
about.
The per-hive shape was justified by the callout policy's rule that an extra
subject must contain the hive name. That rule governs `extra_hive_subjects` —
what a HIVE may publish. This subject lives in the controller's reader grant,
which the rule does not constrain, so a real rule was carried across into a
decision it had no authority over.
Knowledge becomes its own category rather than a leaf under a general event
namespace, since a namespace shaped for events that do not exist yet is a
decision made before there is anything to decide from. The empty config-PR match
arm goes with it: an arm with no body claims this is where the deploy path is
handled, and it is not.
The deny test stays and matters more, not less: with one shared subject a forged
event would reach the whole swarm where a per-hive one reached a single hive.
The test's doc named knowledge::ensure_webhook as one of two hive-side
reapers. That function is gone; its replacement matches the full URL and
cannot touch another hive's hook.
The assertion arm stays. The hazard is not what this repository's source
says, it is what is deployed — a hive still running the previous version
reaps by suffix until it is upgraded, so the arm guards the transition
rather than a current code path. Recorded when to drop it.
A webhook has exactly one target URL, so every hive registering one
against the shared internal/knowledge repository was last-writer-wins
rather than idempotent: all but the most recent silently stopped
receiving deliveries. The swarm controller holds the single registration
and now addresses an event to each hive over the queue instead.
This is a migration, not a deletion. Not registering any more fixes
nothing on a hive that has already run — the hook it created persists on
the forge, so the contention would survive on exactly the deployments
that have it while fresh installs looked fixed. The hive that created a
hook removes it.
It removes only its OWN, matched on the full URL rather than the
/webhook/knowledge suffix. A hook with that suffix and a different base
belongs to another hive, possibly one not yet upgraded, and deleting it
would break that hive's knowledge sync until it caught up. Reaping a
neighbour's registration is the behaviour being removed here; doing it
while fixing it would only invert the direction.
The predecessor did reap by suffix, to clear loopback hooks left by an
older single-hive layout. That was safe when a hive was alone on its
forge and is not safe now. The hive-side registrars also acted as reapers
of hooks under their own path, which is why the swarm hook lives under
/webhook/forge/; removing this registrar removes that reaper too.
Intended, and stated because no reviewer would infer it from the diff.
The receive endpoint goes with it. A live HMAC-verified
/webhook/knowledge that nothing can legitimately reach would tell the
next reader that this is how a hive learns about knowledge changes.
Docs move in the same commit: docs/swarm/README.md said two hooks exist
per swarm-wide repo and neither should be deleted, which is now true for
agent-configs and wrong for internal/knowledge — a half-correct
description being worse than an uncorrected one.
A hive learned the knowledge repository had changed only by registering
its own forge webhook. This subscribes to the per-hive subject the
controller now publishes on and calls the pull this daemon already runs
at boot.
Shares the hive's ONE queue connection rather than opening a second: a
second connect would double the auth-callout traffic against authelia and
give the two paths independent reconnect state, so one could be serving
while the other was still down. Same argument as the controller side.
No payload is read, because there is none to read — the webhook handler
this replaces took two fields from Forgejo and used neither, then ran
`git pull`, which re-derives everything from the repository.
At-most-once, and that is not a regression: a webhook delivery to a hive
that is down is lost identically today, and the boot pull covers it.
JetStream would require this end to publish to
`$JS.API.CONSUMER.CREATE.<stream>`, which the callout policy does not
grant, so durability would cost grants on both sides to remove a failure
the boot pull already handles.
⚠️ Documented at the call site rather than left implicit: a refused
subscription is indistinguishable from a quiet one, because NATS reports
authorization violations asynchronously on the connection. If hives stop
hearing events, the server log is the thing that knows.
futures-util comes from the workspace (same version swarm-controller
already uses), not a new dependency version.
The controller verified a knowledge delivery, logged it, and returned OK.
Nothing downstream ever heard about it, so a hive learned the repository
had changed only by registering its own webhook — which is the
last-writer-wins contention this issue is about.
The event carries no payload. The hive-side handler this replaces reads
two fields from Forgejo's push webhook and uses neither — both are
filters — then runs `git pull`, which re-derives everything from the
repository. What crosses the queue is an edge trigger, and fields would
invent a contract nobody reads.
One subject per hive, so the callout policy can express "this hive may
hear its own events" at all; a subject with no hive component is the same
subject for every hive.
`ConfigPr` deliveries are deliberately not forwarded. A hive does not
want to hear that a config PR opened — it wants to be told when to
rebuild from main, which the controller decides after a merge rather than
by relaying this delivery. That is deploy coordination's job, and the
empty arm is there so the omission reads as scoped rather than forgotten.
Fails soft: a missed announcement costs a hive stale knowledge until its
next boot pull, which is the same cost as a webhook delivery to a hive
that was down — what this replaces. A permission failure cannot be
observed at the call site (a NATS authorization violation is reported
asynchronously on the connection, reaching a client as a timeout or as
nothing), so the doc says the flush proves only that the bytes left this
process and points at the server log.
The controller could not emit an event at all: a reader's grant is
`reader_subjects()`, which is `$JS.API.*` only, so a publish to any event
subject would be refused — and a NATS refusal reaches the client as a
timeout, so the visible symptom would have been a hive that never hears
about a change, with nothing in any log naming a permission.
Adds `swarm_queue_client::events`, following `status::BUCKET`: three
crates must agree on these strings (the controller publishes, a hive
subscribes, the callout responder decides whether the publish is
permitted), and a literal repeated across crates is an agreement nothing
checks. The responder speaks neither jetstream nor kv, so the module is
unconditional and carries no NATS types, exactly as the bucket name is.
The grant takes the wildcard form from the same function the publisher
calls, so the two cannot drift; a separate wildcard constant would have
re-created the disagreement this module exists to prevent.
Tests pin that a reader gets the subject and that a hive does NOT — a
hive able to publish here could tell a neighbour the knowledge repo
changed when it had not, which is an unauthenticated write into someone
else's control path. That one asserts on the subject root rather than a
rendered subject, so a future event leaf fails it too instead of passing
because the test only knew about `knowledge`.
Both assertions mutation-tested: removing the grant fails the reader
test, granting a hive the subject fails the denial test, each on its own
assertion line, and the unmutated tree is green.
argus, reviewing, ran the guard against a source path that does not
exist rather than one that is empty. grep writes nothing to stdout in
that case, so `|| true` left the count variable empty and the -eq test
died with "integer expected" instead of reporting.
The unit still failed — cat hits the same missing file and set -e stops
it — but with a generic "no such file" rather than the message naming
which half is absent, which is the only thing this guard is for.
`|| echo 0` on all three counts. The gate gained the arm that was
missing: an absent source, asserted to fail THROUGH the guard rather
than merely to fail.
The guard inspected the assembled file for any certificate. The system
store always holds certificates, so it passed unconditionally — including
in the one case it was written to catch, where the hive CA half
contributed nothing.
That half is the only one that matters here: every name these consumers
verify is issued by our own CA, so a bundle of nothing but public CAs is,
for this purpose, an empty bundle that measures as full. The failure is
silent and total — the unit reports success and every egress TLS call to
a swarm service then fails.
Counts the source on its own before assembling, and checks the result
carries what both halves brought, so a source truncated between the count
and the copy is caught too.
Scope is stated at the guard: it proves the anchor was contributed, not
that it is usable. A consumer reading only the first certificate ignores
it regardless, which is what took the swarm collector down, and no check
on this file can see that. Only a handshake can.
- serverName upgrade note: state that it changed, drop the historical
list of what it changed from/to
- openFirewall upgrade note: remove entirely, change is old enough
that it's no longer worth flagging
serverName/gatewayHost and openFirewall both led with 'Breaking
change'/'has now moved twice' phrasing woven into the primary
explanation, reading as an active/ongoing concern rather than
settled current behaviour with a historical upgrade note attached.
Reframed to state the current default first, then a demarcated
'Upgrading...' section/note for deployments carrying an older
value forward — same pattern swarm/README.md's own upgrade notes
already use. No factual content changed, removed, or added.
Root cause (found via a zoomed pixel-level screenshot check, not
guessed): the indicator span always exists in the DOM starting from
left:0/width:0/transparent, and the very first real position landed
via a CSS-transitioned change from that fallback rather than a snap —
so the underline visibly grew in from nothing over 140ms instead of
being there immediately, reading as entirely absent on a fast/slow
first paint alike depending on timing. Every subsequent navigation was
unaffected (always transitioning between two already-visible states).
Fixed with a one-render-only transition suppression
(indicatorSettledOnce, gated one tick behind the indicator's first
non-null commit) so the first placement snaps instead of animating in,
while every later hop still animates normally. Verified against the
exact repro: a fresh page load at the same short virtual-time-budget
that previously showed no underline now shows it immediately.
The brand/logo text now rides the same accent value the nav underline
uses, including mid-sweep, so the header reads as one accent changing
rather than the underline alone — per mara's review comment.
Content entrance was a scale+overshoot pop; mara didn't like it on
review, swapped for a plain fade. The nav underline now hops through
every nav item it passes over on its way to the new active one
(position and colour together) instead of tweening directly between
the two endpoints, matching what she actually meant by "move through
color wheel" — a->c visibly touches b's colour along the way, not a
smooth continuous hue rotation.
Shell's page content now remounts on every navigation and plays a
scale+fade pop-in entrance. The active nav link's underline is now a
single shared element that slides to its new position instead of
snapping, re-colouring to a per-tab accent — a discrete cycle through
the existing base16 chromatic slots, not a continuous hue rotation.
Both are gated on the data-motion/prefers-reduced-motion plumbing
lib/motion-apply.ts already had wired and waiting for a first real
consumer.
Reuses HivesPage's exact RefreshIntervalPicker/useRefreshInterval
pattern (30s default, off/10s/30s/1m presets, pauses while the tab is
backgrounded). Ticks a refreshToken bump rather than doing its own
fetch — JobqGraph/JobqRollup already accept that prop as their
documented refetch lever, from the dashboard's rebuild-queue view.
