ADR 0050 Phase A. Rename the licence's `provider` field to `vendor` — the
inventory already uses "provider" for which node answers a brokered provision,
and one word must not carry two facts — and route the licence layer's sealing
and delivery through a per-vendor adapter selected by that field.
The rename touches the Go struct/params/SQL in internal/licences, the operator
CLI, and the schema: 0001 (the consolidated schema) now creates the column as
`vendor`; a new guarded 0002 renames it on a database that predates the change,
and is a no-op on a fresh one.
The adapter (internal/licences/adapters) has a `shape` and the two verbs a
static-key vendor needs — accept (the generic anonymous-box seal) and deliver
(the sealed blob unchanged). refresh/identity/usage are named as optional
capability interfaces so the refreshable-grant seam exists before its code.
A registry maps vendor→shape (anthropic→static-key for now, with a Phase-B
TODO to swap it to refreshable-grant); an unknown vendor is refused clearly.
Behaviour is unchanged from the operator's view except the field name.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
A module may declare a short `slug`; the mesh derives mesh_<node>_<slug|name> and
refuses at assignment (naming the slug as the remedy) when it would still overflow —
identityLimit is now 20, an S3 access key's, the tightest of the backends a login
reaches (04-ISSUES/010). The slug rides the grant so the provider derives the same
login the consumer does, even across nodes. CheckIdentity is now wired, in grantsFor.
Also, the minted secret shrinks to 40 chars (30 bytes) from 43: an S3 secret key is
8-40, the same fit-the-tightest-backend rule on the credential's other half.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
LavinMQ refuses a non-administrator declaring a queue with a dead-letter
exchange, so a scoped module cannot make its own. EnsureModuleQueue declares
<node>.<module>.events with its DLX as the mesh, and 'module issue' does so
for a consuming module — the runtime then passively checks it rather than
declaring. Verified against a real broker: the scoped account binds and
consumes the pre-declared queue.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
'module issue <module> --node <m>' looks up the module's emits/consumes from
the catalogue, ensures the bus exchanges exist, creates its scoped account
(CreateModuleAccount), and seals an amqps {url,fingerprint} to the node as the
module's broker own-secret — the same delivery as 'builder issue', now generic.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
The report carries the digest of the declaration it applied (mesh-host
8211d8b), and the mesh stores it beside the outcome. `reported` rows in
the status JSON now say `current`: whether the machine's last word
names the declaration last sent.
Not derivable from the timestamps beside it, which is why they were
not enough: an apply begun under the previous declaration reports
after the next send — newer, and still about the old words. The lab
lost exactly that race between one test's closing push and the next
test's opening one.
Empty digests — every host from before reports carried one — read as
not current, which errs toward waiting rather than toward asserting on
files that are not there yet.
"Not waiting" says the declaration is current, not that the machine
finished applying it: the sent digest is recorded at send. So a test
that pushed, saw waiting clear, and asked the machine what it was
running found containers that did not exist yet — the certificate fix
made compositions stable, and the settling that used to fail first had
been hiding the gap behind it.
The mesh already held the missing half: every machine's last report,
with its time. It just was not in the JSON. `reported` now sets each
machine's last word beside when the current declaration went to it, and
"has it caught up" becomes a comparison of two timestamps the mesh
recorded itself — a report newer than the send means the machine acted
on what was sent; older means it is still working, which waiting alone
cannot distinguish.
Confining allocation to the push path took `plan` with it, and `plan`
belongs on the other side: it is a person asking what a push would do to
one named machine, so the port it shows and the secret it seals must be
the ones a push would use. Both are kept once chosen, so showing
numbers a later push would replace answers a question nobody asked.
Caught by the lab: composing the real modules stopped producing
postgres's sealed superuser, because nothing had minted it and the
read-only path correctly declined to.
The line is not question versus command. It is a person asking once
about one machine, against the mesh asking continuously about all of
them — the second is what hung, and the second is what reads.
object-store.json and minio.json described the same thing: same image,
same provision at the same scope, same provisioner. Not two
implementations a person could choose between — one module written
twice. Assigning both to a node would have collided on `s3-bucket`.
It exists because it was written first, to pair with photos.json for the
README's worked edge, and minio.json was the fuller version of the same
module written later. Nobody removed the first.
The pair test keeps its point and now reads the surviving one. Checked
across the rest: this was the only duplicate.
`status` hung. It composes a declaration for every node to answer *is
this machine running what I would send it*, and composing one assigns
each module a machine port — so the question wrote to the database, and
wrote to the same rows as the machine it was asking about.
`port_assignment` is unique on (node, machine). Two transactions
inserting the same port do not race, they queue: the second waits on the
index until the first commits. A status polled every two seconds while a
node applies is two writers on those rows, and the poll stopped
returning rather than returning something wrong — which is the better
failure of the two, and still a failure.
The latent version of this was there before anything polled: two
compositions running at once could both allocate.
So allocation belongs to the send path alone. The mesh chooses a port
when it commits to sending one; every other caller reads what was
chosen. A module with nothing assigned has never been sent, which is
precisely what "waiting" means — the read needs no number to be right
about that, and inventing one would make the answer worse.
Named rather than passed as a bare bool: at three call sites, `true` and
`false` say nothing about which of these two things is meant.
Checked by the lab, which now polls status throughout an apply.
The lab caught this: a module declaring a port and running no container
had its rule set opened on 20000 while its service sat on 9101. The
firewall reported success and blocked the thing it was told to admit,
which is the precise failure the filtering comment warns about, arrived
at from the other side.
Assignment was applied to every declared port. But a container's mapping
is the thing that translates, and where there is none the software binds
what it binds — the mesh choosing a number does not move the service, it
only makes the mesh wrong about where it is.
The declaration side already knew this: publishedOn rewrites container
ports and nothing else. Filtering did not, so the two disagreed about
the same fact. MachineSide is now the one derivation both follow.
It also fixes a second case nobody had hit yet: a mapping the manifest
wrote itself, like the mail system's 7080:80. That is passed through
untouched when composing, so assigning it a machine port would have
opened a rule on a port the container does not publish. Either side of
such a mapping now names it, and the host side is the answer — a module
may read `listens` as what its software binds or as what the machine
exposes, and both readings want the same number.
Recorded either way, assigned or not: the map means where this module's
port is on this machine, and every reader needs that answer regardless
of who chose it.
Tests bite — making it always assignable reproduces the lab failure.
novox/hq ADR 0038. A module cannot choose a port: it is written once and
assigned anywhere, so any number it picks is a guess about a machine it
has never seen. A database module met the mesh's own store on 5432 and
was told, by a container runtime three layers down, that the port was
already allocated.
The number used to appear three times in every module — the rule set,
what a consumer is told, and what the runtime publishes — agreeing only
because one person wrote all three. Now it appears once, in `listens`,
and the other two are derived: the container publishes `20000:5432`, the
consumer is told 20000, and the rule set opens 20000.
An assignment is made once and kept, as a credential is. A port that
moved on every declaration would restart both ends each time and hand a
consumer a number that was true when it was read.
Ports the protocol fixes — mail on 25, submission on 587, DNS on 53 —
say so, and are then claims: one holder per machine, and the second is
refused by name at assignment. That is the mechanism the mesh already
has for what is singular on a machine, pointed at ports.
A mapping written the long way is left exactly as it is. Some things
must be pinned by hand, and quietly overruling somebody who wrote both
halves would be worse than not offering the short form.
Still open, and known: the substrate is not a module, so the mesh has
never heard of its own store and cannot yet assign around it. That is
what 028 will still be about after this.
novox/hq 04-ISSUES/022. A credential was keyed by provision, consumer
node and provider node, so "who is asking" was answered by naming a
host. The node this mesh exists to take over runs eight modules against
one database server.
The symptom had two halves and only one was loud. The provider refused,
naming the modules and explaining they would share one credential, which
reads as a decision rather than a limit. The consumer did not refuse: it
resolved cleanly, wrote one module's credential file and left the others
absent — a service that starts and cannot authenticate, with nothing
saying why. That is 021 again on a different axis.
Three modules wanting one database produced one need, carrying whichever
module mentioned it first, because the resolution walk is a work-list
over names. The fan-out now happens in one place, after the walk. The
record path already did this correctly and said why: a consumer here is
a module on a machine. It is the same rule.
Downstream: the secret's key gains the consuming module, the grant file
is named after both halves, needs are matched by provision and module
rather than provision alone, and the provisioners name the role and the
access key after the module. The refusal in ContributionsTo is gone
because there is nothing left to refuse.
Worth stating plainly: without that refusal, gitea's login would have
opened keycloak's database. From the provisioner's side it created
exactly what it was asked to create.
Existing secrets are discarded rather than backfilled. They cannot say
which module they were for, and a secret is remade and delivered to both
ends on the next push — so this costs one rotation and invents nothing.
Also guards the role name against PostgreSQL's 63-byte truncation, which
is a notice rather than an error and would reintroduce exactly this
collision at a length nobody tests.
Three faults injected — the fan-out removed, needs matched by name
alone, the grant file named after the machine — each caught.
novox/hq ADR 0035: one implementation, several surfaces, and a surface
holds no decisions. The act of assigning — including that an assignment
which does not resolve is kept and still refused — moved into acts.go,
and the command line now calls it too. Two surfaces, one refusal, in the
same words.
It will not run without --issuer, and refuses at start rather than per
request so it is found by whoever ran it rather than by whoever finds
it. There is no flag that removes the check.
The authenticator is honest about what it is: no token can be verified
until an identity provider exists, because that is a module and none is
running, so every request is refused and told that the command line
still works. A surface that functioned without authentication would be
one somebody left running — and the board this stands behind is
published on a public name.
Four refusals, four tests. The last one first asserted "not 200", which
passed because a request with no database fails at the store anyway — it
proved nothing about whether the input was checked. It now asserts the
specific refusal, and bites when the check is removed.
The command refused every real invocation. Go's flag package stops
parsing at the first non-flag argument, so with the positionals first —
the order that reads correctly — `--from -` stayed among them and the
count check rejected it.
The host's own parser carries a note about this exact fault, and the
version it describes is worse: there a flag somebody passed was silently
ignored and the command succeeded anyway. This one at least refused.
The tests did not catch it because every case in them was a rejection.
The command was broken in the only way that matters — it refused what it
is for — and the suite was green. The lab found it at the first call.
Two tests now: the helper, and the command itself with a --from naming a
file that is not there, so the complaint must be about the file rather
than about usage. The second exists because injecting against the first
stayed silent: testing the helper alone left the command free to ignore
it entirely.
The entry point for adopting something already running, and the half
that was missing. The store has carried the distinction since the
beginning — a module secret records whether it was `made` or `accepted`,
and refuses to invent a replacement for the second — and
AcceptSecretForModule existed, with exactly one caller: the broker
account issued to a build machine. Nothing else could write one.
Without it every module secret is generated, which against a database
that already exists puts 32 random bytes where a working credential was.
The machine applies it, reports success, and whatever reads it fails to
authenticate somewhere else entirely, with the mesh insisting the secret
was delivered — which it was.
The value is read from a file or from standard input, never from an
argument: a value on the command line is in the shell's history and in
the process list. Same path a model-access key already takes, and no new
dependency — the first version reached for x/term and the existing one
needed nothing.
Sealed on the way in, plaintext discarded, and not printed back. The
only difference from a generated secret is where the value came from.
Two rules with a test each, and the second is the one that would have
been got wrong: only the line ending is removed, never surrounding
space. Trimming both ends is the obvious thing and would deliver a
password chosen with a leading space as a different password, silently.
Both were briefly untested for different reasons — the trimming lived
where no test could reach it, and then a -run filter matched neither
test. Extracted, and injected against the whole suite.
It sat beside `secrets` — where a *provision's* credential lands on a consumer.
Both were name-to-path, both held something secret, and the names
distinguished them not at all. Reaching for the wrong one parsed cleanly and
failed somewhere else entirely, which is the shape of fault this whole design
exists to prevent, sitting in the manifest format.
The axis that separates them is not how secret they are — both are — but
whose. `secrets` is keyed by the provision it is for and belongs to a
relationship with another machine. `own-secrets` is keyed by a name the module
chose and belongs to nobody else.
A manifest using the old name is told the new one rather than refused with
"unknown field": whoever wrote it knew what they meant, and the mesh knows what
it is called now. An invented key is still refused as one rather than guessed
at.
Found by auditing the 19 manifest fields for whether any could be mistaken for
another. This was the only pair that could — and while checking it, a second
instance of the same collision turned up one layer down: `Manifest.Needs` and
`Resolution.Needs` were different concepts sharing a name in Go. The rename
separates those too.
Working out what a machine should be reaches the identity context for its
certificate and the licence context for its model access. Both were opened —
and waited on — inside functions called for every node in a push. Two machines
hid it. Fifty would be fifty connect-and-wait cycles for data that does not
change while the push runs.
So a command holds what it has open, and passes it. Each context is opened on
first use rather than up front, because most commands need one and paying to
reach three would be the same waste from the other side.
The contexts stay separate, which is the point: this is one struct holding
three connections to three databases, not one connection to a shared one. No
context reaches another's store, and each still holds only its own credential
(novox/hq ADR 0008).
A pure move again — the gate is green before and after, and no test changed.
2,769 lines and 59 functions, holding command parsing, store opening,
resolution, the board, rotation, licences, builds and status rendering.
Nothing in it was wrong. It grew because appending was always the cheapest next
step, and no single edit was the one that should have been a new file.
That is exactly how novox/hq ADR 0001 records `hal/sdk` reaching 155 files and
34,636 lines — "containing code from every context", with each addition
avoiding a cycle and none of them the mistake. This is the same shape at 8% of
the size, which is why it is worth doing now rather than noting.
Eight files, along boundaries that already existed: what a machine is; the
private network; the catalogue; working out what one machine should be; sending
it; builds; the three questions; and reaching each context's store. main.go
keeps what a main is for — parsing arguments and dispatching.
A pure move. No behaviour changed, no test changed, and the gate is green
before and after — which is the only thing that makes a refactor this size
safe to do in one commit.
Services are named under the machine they run on — postgres.novox.internal,
plex.ace.internal. The first label is the service and the rest is the node, so
what has to resolve is anything under a node's name. What routes it once it
arrives is a proxy's concern and stays separate.
A hosts file cannot do that. It answers exact names, and a wildcard there would
mean writing down every service in advance — which is the enumeration the
arrangement exists to avoid. novox/hq 08-connectivity named this exact case as
the trigger for needing a resolver rather than a file, and it is the first
thing to meet it.
The mesh writes the data and runs no daemon. A resolver is third-party
software, and third-party software runs on the mesh rather than being of it
(ADR 0001): the mesh has no business shipping one, choosing which one, or
knowing its configuration language. What only the mesh can know is which
machines exist and where they are. A module that runs a resolver requires what
this provides and reads one file, so swapping the daemon changes that module
and nothing here.
Separate from names rather than part of them: a machine with no container
runtime can still have a hosts file, and folding them together would take exact
names away from a machine that cannot run a daemon in order to give it a
wildcard it cannot use either.
A machine with no address is left out. A wildcard pointing at nothing is worse
than no wildcard — every name under it resolves and then hangs, where an
unresolvable name fails at once and says which name it was.
Internal names are written to the machine's hosts file, which serves the
machine and not what the machine runs: a container gets its own hosts file
holding only its own hostname. So every name the mesh wrote was invisible to
the majority of things that need one — and on the machine it always worked,
which is exactly what made it easy to miss.
It was hit for real in the lab, and worked around by resolving the address on
the machine and passing it in. That workaround is now removed, and its absence
is the assertion.
A file rather than a resolver, which is the decision the mesh already made
about names and this extends rather than overturns: it works on every runtime,
needs no package and has no failure mode of its own. The stated trigger for a
resolver — names that are not one-per-node, service names, wildcards — is
still not met.
Given by the mesh, not chosen by a module: a module that listed the machines
would go stale the day one joins, and one that did not would be a module whose
containers cannot reach anything by name. A container that named its own keeps
them and gets the mesh's beside them.
Only containers, and not the ones on the machine's own network: a runtime
refuses to write a hosts file for those, and a file or a service given the
field is a declaration the host refuses outright — so getting it wrong breaks
the whole machine for something that was never about names.
Adding "not running what the mesh would send it" to `status` and not to the
board would have left two answers to one question with a person in front of
each — which is the single thing this page's design forbids, introduced by the
change that was supposed to make the question answerable.
The published JSON carries it as well, so the page, the command and anything
built against either say the same thing from the same read. Additive, because
that shape is hard to change once anything is built against it.
Never told stays separate from out of date on the page as it is everywhere
else: same remedy, and nobody has ever asked that machine to be anything.
It meant "failed or refused". So a machine that applied cleanly and whose
declaration has since changed was not behind — and novox/hq ADR 0010's
question, did my change go out?, was answerable exactly for the machines that
broke. For every machine that worked, the answer was silence whether the change
had gone out or not, which is the thing replacing a pipeline was supposed not
to cost.
The mesh now records a digest of what it last sent each machine. A digest
rather than the declaration: it can compute what a machine should be at any
moment, and keeping a copy would be a second account of it able to disagree
with the first. What cannot be recomputed is what was actually sent.
Recorded after the send, not before — a digest kept for something that failed
to send would make the machine look current for a declaration it never
received.
Never told stays separate from out of date. The remedy is the same push and the
situations are not alike: nobody has ever asked that machine to be anything.
And a machine the mesh could not work out is not reported as waiting, because
saying so would invent a comparison — that is `plan`'s answer to give.
`status` says it and `push --behind` sends it, or the flag would know something
the person reading the status does not.
novox/hq ADR 0009: a capability's presence gates an assignment and its detail
carries a value — seat: card1-DP-1, an architecture, an amount of memory. So
'can this run here' and 'what should it be configured as' are one fact read two
ways, and the mesh was keeping the first read and discarding the second.
The reason an absent capability is absent went the same way, which is the case
a person most needs: 'this machine has no container runtime' is the answer and
'docker is not installed' is why, and only the machine knows why.
`node show` says it back. Never reported and reported nothing stay different
things there — one machine has not run the host, the other ran it and can do
nothing, and those send a person to different places.
novox/hq 03-DESIGN/01-to-be/11-a-board.md, built. The board being replaced is
one service reading every context's database directly — ADR 0008 violated by
the one component with a reason to violate it. The cost is not hypothetical: a
boundary nothing may cross can move, and one thing crossing it is enough to
freeze it. A board that reads the provisioning tables breaks when provisioning
changes them, and the change then gets weighed against the board.
So the three questions are read once, by one function, for all three ways of
saying them — a person's status, its JSON, and this page. Three
implementations of "which machine is not doing what it was told" would be three
chances to disagree.
Refused and failed stay distinct all the way to the page: refused means the
machine is exactly as it was and what is wrong is in what was sent; failed
means it is in a state nobody declared. Different places to fix, so one word
for both would send half the readers to the wrong one.
It stores nothing, changes nothing, and every action it might offer already
exists as a command. A board that cannot reach the mesh says so rather than
rendering an empty page — an empty page says "nothing is wrong" in the one
situation where nobody can know that.
One test earns its place twice: a machine's own words are the whole reason the
page is useful and the one thing on it nobody in this repository wrote, so they
are shown and are not markup.
A status that names what is wrong and not what to do about it makes somebody go
and find the command — and the command is the whole point of having noticed.
The hint existed on one of the two paths that print this.
Its own store, its own test database, the same shape every other context has.
Five properties: a key with nobody to seal it to is refused rather than kept
readably; a key is sealed once per holder and the blobs differ because they are
sealed to different machines; a holder recorded afterwards has none and the
existing ones keep theirs; releasing a consumer takes its key; and a licence
nobody recorded is refused by name.
The last was the only one whose message mattered and whose message was not
checked — the database's own foreign-key error is true and mentions a
constraint, which sends somebody to read a schema instead of typing the name
they meant.
Partial sealing now says how far it got. The person holding the key is the only
one who can finish, and running it again knowing what it will do is different
from running it hoping.
novox/hq ADR 0010 replaced a pipeline with a comparison, and named the risk:
losing the question "did my change go out?". The mesh could already answer
which modules are behind their source — and then a person read that list and
retyped each repository, which is a person being the loop, and the loop is the
thing the pipeline was doing before it was taken away.
The mirror of `push --behind`, with the same argument and the same refusal to
combine the two forms: naming a repository and asking which need building are
different requests.
One failing does not stop the others, for the same reason one broken module no
longer blocks a machine's whole declaration: a mesh where one bad repository
holds back nine good ones is a mesh where nobody dares add the tenth.
Each is built from its own recorded ref rather than the commit the mesh
happened to notice — pinning to that would quietly turn a tracked branch into
a pin.
novox/hq ADR 0024, gaps 1 and 2. The user's stated requirement, and the first
thing here that no machine can answer: a hosted model is on nobody's node and
is reached over the public internet, so the rule that refuses two ends sharing
no private network must not apply to it.
A licence is a named thing and the name is the operator's — *the personal
account*, *the organisation's* — because the whole point is saying which one a
given consumer uses, and an anonymous credential hanging off a provider cannot
be said. Many to many, so deliberately not a claim: two machines sharing an
account is ordinary rather than a collision.
Gap 2 is the missing verb, *accept*: take a value somebody supplied, seal it to
each holder, discard the plaintext. With the consequence stated rather than
hidden — a holder recorded after the key was supplied has no key and the mesh
cannot make one, so it is refused by name with the remedy, not silently handed
an empty file.
Refusal is felt, as the record warns: a mesh holding three ways to reach a
model refuses every consumer that has not chosen. So the refusal names the
candidates and the exact command. Being right is not the same as being usable.
Gaps 3 and 4 — a consumer that is not a machine, and switching as a reaction
rather than a declaration — remain gaps. Half-building them would put a
conditional in the declaration language, which is what ADR 0024 says plainly to
avoid.
Its own context, with its own store and its own credential: a licence is a
different aggregate from anything inventory owns, and it refers to nodes by
name because that is what crossing a context boundary may carry.
novox/hq 08-connectivity §3, built. The mirror of a database grant: there the
consumer supplies a name and receives credentials; here it supplies a target
and receives a name. Nothing new in the vocabulary — a route is a provision
like any other.
One field was missing and it is the one that matters for anything reaching
back: a contribution now carries where the mesh says that machine is. A reverse
proxy is told to send traffic to a consumer and has to open a connection, so
without it every provider implementing a provision would have to know how the
mesh names machines — a convention leaking into every module.
The proxy itself is an example, not part of the control plane: the contract is
the file, not this program. It replaces its table whole rather than merging,
because the file is the whole truth about who has a route and merging would
keep serving a name whose module was unassigned — the stale-route fault
08-connectivity lists as open, reintroduced one level down. A name it does not
serve is refused by saying which it does: a route withdrawn and a name that
never existed are different things.
The invariant novox/hq ADR 0001 records as unowned, and it was measurably
false in HAL: a provision documented as never rotating minted a new password on
every adoption and updated only the provider's row. Consumers on three nodes
held dead credentials for two days while the mesh reported success. Nothing
enumerated who held the old one.
Three things make that impossible here. The holders are a set the mesh can name
— each pair has its own credential, so rotating one consumer touches one role
and the affected list is a query rather than an assumption. Both ends are
pushed by this command rather than a later one, because leaving the sending to
whoever remembered is the fault exactly. And it is all-or-nothing: if any
affected machine cannot be resolved, nothing is sent and the old credential
keeps working, which is a mesh that has not rotated rather than one that has
half-rotated.
The window is stated rather than hidden: a role's password changes on the
provider and the file changes on the consumer, and they cannot be simultaneous.
The provisioner now takes its superuser password from the file the mesh wrote,
which is how the mesh delivers one. Passing it through the environment needed a
person in the middle of the one path that exists so there is not one — and put
a superuser password where `docker inspect` prints it.
The builder hashed the certificate and compared a bare digest against a
fingerprint written as `sha256:` followed by 64 hex characters. It could never
match — and it failed as "this is not the broker this builder was told about",
which is the one thing this check exists to report truthfully. A check that
cries wolf on every correct broker is worse than no check, because the first
thing anybody does is remove it.
The error now prints what was expected beside what arrived, the way the host's
has always done: without both, the message describes a mismatch nobody can
confirm.
And the pin check has its own test, driven against a real TLS handshake — it
accepts the certificate whose fingerprint the mesh wrote and refuses another.
A pin only ever exercised through a live broker is a pin nothing tests.
The credential was a URL and nothing else, so the builder verified the broker
the ordinary way — against public roots. A mesh's broker presents a certificate
of the mesh's own, which is in no trust store anywhere, so the connection could
only ever succeed against a broker somebody else vouches for. It failed at TLS
with an error about an unknown authority rather than about a missing pin, and
the container sat there running: up, credential on disk, connected to nothing.
So the sealed credential now carries the URL and the broker's fingerprint —
the same two facts a node's token carries, for the same reason, delivered out
of band relative to the thing being trusted. The builder pins it: the standard
chain check is replaced rather than removed, and what replaces it is stricter,
accepting one certificate instead of every certificate a public authority
would sign.
A file holding only a URL still works, for a builder somebody runs by hand
against a broker with an ordinary certificate.
A rule nobody derives is a rule somebody keeps in step by hand, and five HAL
manifests carry a `scope:` key that reads as a restriction and restricts
nothing. Both halves are closed here.
Manifests are parsed strictly. An unknown key is refused, which is the
discipline the host's declaration parser has always had; `scope:` survived
because nothing rejected it.
A module says what it listens on and who may reach it, and saying from where is
required — a rule with no source is open, and must say so rather than appear to
restrict something. The mesh gathers every assigned module's ports, widens
where two overlap, names every module that wanted each one, and renders one
nftables file per node. What no module declared is closed.
Three things it deliberately does not do: it writes no forward policy, because
what a machine routes is the container runtime's business and dropping there
stops every container on the node; it never flushes the whole ruleset, only
its own table; and it carries no command to load itself, because the link may
not carry an action. A service declares `restart-on` the file instead, which is
the shape that rule leaves.
Also fixes a fault the lab found: certificateFor asked where every node is
without the catalogue, so nothing resolved, every machine looked like it was on
no private network, and every certificate the mesh was asked for was refused
with a reason that was not true. Asking that question without the catalogue is
now refused rather than answered wrongly.
08-connectivity keeps two authorities apart on purpose: a public one for
names the outside world reaches, and the mesh's own for names only the
mesh knows. Nothing implemented the second, so anything between machines
was plaintext or trust-on-first-use — which the design refuses everywhere
else.
A node now generates a fourth key at enrolment and reports the public
half. A fourth, because a key used for two purposes is one rotation away
from breaking the other: the identity key signs messages to the mesh and
would do for TLS, and reusing it would mean rotating a node's identity
every time its certificate is replaced.
**Nothing secret travels and nothing is sealed.** A certificate authority
says "this name belongs to the holder of this key", so the mesh signs a
public half it cannot use, and the certificate it issues is public. A
module asks for one and is given the certificate and, if it wants,
the mesh's own — the private key is a path to a file the machine already
has, the same arrangement the private network's key uses.
Asserted by verifying rather than inspecting, because a certificate that
parses and does not chain fails at the moment something connects:
- what the mesh issues verifies against the mesh, for the name asked for
- the name is in the subject alternative names, since a certificate
carrying it only in the common name is refused by every modern client
- it certifies the key the node generated and no other
- another mesh's certificate does not verify, which is the whole point of
two authorities being separate
- the authority cannot sign another authority — one that could is one
that can be delegated without anybody deciding to
- two control planes starting together agree on one authority, or a mesh
has certificates half its machines refuse
Certificates last ten years, which is a choice: a short life needs
something to renew it, and a renewal that fails silently is a mesh that
stops trusting itself on a date nobody wrote down. What makes one
replaceable is that the mesh reissues on demand, not that it expires.
A board reads through interfaces and holds nothing. Everything it needs
is already answered — as text, for people, which is not something a page
can read.
`--json` rather than a serving API, because nothing needs one yet:
whatever serves a board runs the command, and the constraint holds either
way — the board never touches a context's store. An API is the larger
thing and should wait until something asks for it.
Both forms are gathered from the same reads before either says anything,
so they answer the same questions rather than being two implementations
that can drift. That was not true of the first version: the JSON printed
after the text, because the branch was too late.
Four properties, each asserted and each confirmed to fail when removed:
- refused and failed stay distinct all the way out. They are fixed in
different places, so one word for both sends half a page's readers to
the wrong one — and how much DID apply is carried, since "three of
eight" and "none of eight" are different machines
- a machine that never spoke carries no time at all, rather than a zero
one that any page would format as a date in 1970
- nothing is null. A page distinguishing "no machines are wrong" from
"this field is missing" has to handle both, and null is the one that
gets forgotten
- no field is named like a secret. Everything here comes from records
that hold no readable one, but a shape a page is built against is
exactly where one would eventually be added for convenience
`status` says which machines are not doing what they were told, and
nothing acted on it: a machine that refused or failed stayed wrong until
somebody ran push again naming it.
`push --behind` sends only to machines whose last report was not a clean
apply. A command rather than a timer, deliberately: a scheduler is then a
scheduler over this, where building the scheduler first would have meant
two paths to one act with nothing to compare them against.
Naming a machine and asking which machines need one are different
requests, so `push <node> --behind` is refused rather than guessed. With
nothing behind it says so, because "nothing needed one" and "this did not
run" must never look the same. A machine failing the same way for six
hours is pushed to anyway and said about — refusing would leave no way to
retry after fixing the cause, and this is a command somebody ran.
Proven in the lab: a machine is broken with a package that does not
exist, `push --behind` names it and not the machine that is fine, the
module is corrected, and the machine recovers without anybody naming it.
And the builder can be told where to publish rather than configured. A
builder that is a module requires an artifact store, and the mesh writes
it the same binding any consumer of any provision gets. A binding with no
address is refused rather than falling back to anything — that would
publish to a store on the wrong machine and be found out much later. The
variable remains for a builder run by a person, which is how it is still
run while being developed.
Two things, both found by trying to write a real postgres module and
discovering it could not be said.
A database has a superuser password, a broker an administrator, a
registry an account. None of them is *for* anybody — they are not the
credential a consumer is given, and the mechanism that hands those out
has a consumer in the middle of it. So a module may declare what it needs
and where to put it, and the mesh generates one per node, seals it, and
reads it no more than it reads any other.
Per node, deliberately: a module running on three machines has three
passwords. One in the manifest instead would put the same secret on every
machine that ever runs it, in a file anybody can read, for ever. Made
once and kept, or a running database would be handed a password it was
not started with; remade when the machine's sealing key changes, like
everything else sealed here.
A need declared and not made is refused rather than skipped, because a
module whose own credential is silently absent starts, fails to
authenticate, and the reason is three layers from the machine reporting
it.
And the provisioner can watch. That is what lets it be a module rather
than a binary somebody places: run once, it needs invoking after every
declaration by a timer or a unit wired to a file; watching, it is an
ordinary long-running service the host already supervises. It polls
rather than watching the filesystem, because the host writes atomically —
the file is replaced, so a watch on the path stops seeing anything after
the first replacement, and a watcher that silently stops working is worse
than a poll. Credentials are compared by digest and never held: this runs
for as long as the machine is up.
A node reports back after applying a declaration: it worked, some of it
failed, or the whole thing was refused. A refusal or a failure moved
last_seen and the reason went to a log line — so "which machine is not
doing what it was told" had no answer the next morning, which is the
question a mesh exists to answer.
Refused and failed are kept as different things, because they are
different situations with different remedies: refused means the machine
is exactly as it was and what is wrong is in what was sent; failed means
it is in a state nobody declared and what is wrong is on the machine. One
word for both would make the record say less than the node did.
One row per node, replaced. The question is the machine's current state —
"this failed an hour ago and then succeeded" is not a machine anybody
needs to look at, and a table of every report would bury the ones that
matter under the ones that do not.
`status` now answers three questions in the order somebody asks them: is
anything broken, is anything not answering, is anything out of date. The
first has consequences now, the third is a plan for later, and a status
leading with the third would bury the first. A machine that has never
spoken is reported as quiet rather than as broken — new, switched off and
unreachable are not the same as tried and could not.
The mapping from a report to an outcome had no test at all, which the
injection caught: it is the code deciding which of those situations a
machine is in. It has four now, including that a partial report never
becomes the account of what the machine holds — the fault that destroyed
a substrate once.
The builder was documented as holding its own broker credential and
nothing else, and nothing issued one — so in practice it used whatever it
was handed, which was the broker's administrative account. A program
documented as holding its own credential and given somebody else's is
worse than one with no story at all.
`builder issue <name>` creates an account that may read the build queue
and write to the mesh exchange. Not a node account: a build machine is
not a node, and a node's queue carries its declarations.
Two faults found by running it, both about the answer path:
- the reply queue was left for the broker to name, and the account was
scoped to `amq.gen-*` — one broker's convention. The builder built,
could not answer, and the connection closed. Reply queues are named
here now, deterministically.
- the answer then went via the DEFAULT exchange, where permission is
granted per exchange rather than per queue. A builder allowed to use it
could publish into any node's queue, which is the privilege a build
machine most obviously should not have. Answers go through the mesh
exchange, which it already may use, and an asker binds its reply queue
to the same key and filters by correlation.
Verified against a real broker: a builder cannot consume a node's queue
and cannot publish to the default exchange. That check nearly reported
the opposite — an unconfirmed publish is asynchronous, so the refusal
arrives as a channel close afterwards and a naive test sees success. With
publisher confirms it is immediate. A negative security assertion made
against an asynchronous call is not an assertion.
Redelivery was observed working while fixing this: builders that died
before answering left their work on the queue, and the next builder did
all of it.
Also: the queue and exchange names exist in both `broker` and `link`,
because `link` imports `broker`. A test in an external package keeps them
agreeing — a builder scoped to a queue nothing publishes to takes no work
and says nothing about why.
A build result was answered to whoever asked and kept nowhere. So "when
did this last build", "why did it fail" and "which machine built what is
running" had no answer, and a build nobody was waiting for was reported
into the void — which is the same as not reporting it.
Failures are recorded too, and that is the point rather than a detail: a
failed build that leaves no trace is indistinguishable from one nobody
asked for, and the difference is the whole of whether somebody should be
looking at something. A build that never learned what it was building
keeps the repository, because that is what a person goes and looks at.
Recording is idempotent on the correlation id, because a result can
arrive twice — as the answer to whoever asked, and on the exchange when
nobody was. Two rows would show one build as two, and which is real is
not answerable afterwards.
The serving control plane now binds `built` as well, so results from
builds it did not ask for are kept. It refuses them loudly when it has
nowhere to put them rather than dropping them, so the broker's own
counters show something arriving that nothing handles.
`builds [<module>]` reads it: what happened lately across the mesh, or
what has happened to one module — the first asked after something goes
wrong, the second when deciding whether to trust something.
What was published is kept with the build, so a digest traces back to
what made it without holding the manifest twice in a place that can
disagree with the first.
A build is work, not state. Everything else the control plane sends a
node is a declaration — this is what you should be — reconciled forever.
A build happens once and is finished. Putting it in a declaration would
mean rebuilding on every reconcile, or a declaration carrying "and I
already did this", which is state about an event rather than about a
machine.
So it travels on its own queue and the answer comes back correlated. One
queue, so several build machines share the work and each request is done
exactly once — which a per-machine routing key would not give.
mesh-builder is the program a build machine runs. Not the control plane,
which must not run commands on a machine; not the host, which would then
need a container runtime and git everywhere to do something almost no
machine will ever do. It holds its own broker credential and nothing
else.
Three properties that are decisions:
- a request is acknowledged only once the answer is away, so a builder
that dies mid-build leaves the work for another machine rather than
losing it with nobody ever hearing why
- one build at a time. Five at once against one runtime finishes all five
slower than it would have finished the first, and the queue is what
shares work between machines
- a failure is a RESULT. A build that fails silently is
indistinguishable from a builder that is not running, and those want
different responses
And `module list` is a catalogue: what exists, at which version, built
from which commit or handed over by hand or shipped with the control
plane, whether it is behind its source, and which machines run it. All of
that was recorded from the first build and none of it was shown, so "is
this current?" could only be answered by reading the database.
Proven against a real broker, registry and store: the mesh asked, a
builder consumed, built, published, answered; the manifest was recorded
with its commit; the source moved and the catalogue said "behind";
rebuilding caught it up with a new digest because the content changed.
One store, and it is the registry the bootstrap already pulls from. An
OCI registry is a content-addressed blob store that also understands
images: PUT a blob and it is retrievable at /v2/<name>/blobs/sha256:… for
ever, by digest. An archive is a content-addressed blob.
A second store beside it was considered and is the right answer for
objects that are mutable, need per-reader access, or are not build output
— somebody's uploads, a backup, a thing with a lifecycle. None of that
describes a digest-pinned archive, and running a second service to hold
one kind of immutable blob is two things to run, two to back up, and two
ways for an artifact to be missing. Overturnable by reading: the manifest
carries a URL and a digest, and neither says what served it.
`build <repository>` clones, reads module.json, builds what it declares,
publishes, and records the manifest with the commit it came from. It is a
command rather than something the control plane does on its own, because
building runs things on a machine and what the control plane may send a
machine is bounded by the declaration language. This is the shape the
builder module takes when it is given work over the broker.
Proven end to end on a real repository and a real registry: a shell
module with a package, a user and a dotfile archive built, published,
fetched back at the digest it declared, rebuilt to the same digest, and
its manifest accepted by the host's own parser — including `user` and
`archive`, which did not exist this morning.
A tag is never accepted as a pin, and a blob already stored is not sent
again — it is named by its content, so re-uploading asks the registry to
store what it already has under the name it already has.
Found by raising a mesh end to end. The broker account a joining node
authenticates as is named after the node, and exists before that machine
has been told anything — so the node has to know its name before the mesh
can tell it. Without it, enrolment fails at the broker with an empty
username, which says nothing about why.
Not a secret, and the issuer already knows it. The wire-format test now
covers it, so a rename on either side fails in both repositories rather
than at enrolment on a real machine.
Contributions were node-local, so a mesh-scoped provider — the one case
that most needs them — never heard from its consumers. A database was
given a password and no idea what to create it for.
Cross-node consumers now reach the provider's `receives` file, merged in
with the ones on its own machine: from the provider's side they are the
same thing, and a provider that had to read two lists would read one of
them. Each names the file its credential is in rather than carrying it,
because the mesh discarded the value and could not put it there. The
readable half therefore stays readable.
And examples/postgres-provisioner, which is the last step: it reads what
the host wrote and makes PostgreSQL accept it. Explicitly not part of the
control plane — the control plane decides and never touches a machine.
This runs on the machine and touches it, and a real one ships with the
module that ships PostgreSQL. It lives here because this is where the
contract is defined, written as something that runs so it can be read.
It reconciles rather than applying a change, because it is never told
what changed. Three things that follow, and each is a fault somebody has
shipped:
- the password is set every time, not only on creation, or a rotation
reports success and changes nothing
- what it made and nobody asks for any more is revoked, or a departed
consumer keeps a working login for ever
- what it did not make is left alone, or it cannot be run on a database
that predates it
Proven in the lab against a real PostgreSQL, each assertion confirmed to
fail with the behaviour removed. The suite is in mesh-lab, which also
records the two ways the test itself was wrong first.
HAL keeps env vars in the registry, encrypted at rest. Its own tooling
records what that bought and what it did not. `secret_locate` matches by
value rather than by name — because the same password sits in
mesh_provisions, in module_env, in each node's .env in plain text, and
inside every connection string composed from it, and its documentation
says those URL copies "are often the only copies actually in use". And a
query against the encrypted column returns zero rows and proves nothing,
so auditing moved to the decrypted copies on the nodes.
Two faults there, and encryption at rest addresses neither: the control
plane can read what it stores, so a copy of the database is a copy of
every credential; and one secret has many homes with nothing tracking
them.
So here the mesh generates a password, seals it to each end with keys
those nodes generated, stores both blobs, and discards the plaintext. It
cannot read what it holds. Neither can the broker relaying it. And
nothing is composed centrally — a connection string is assembled on the
machine that needs one — so no copy is ever minted in a shape nothing
tracks. `Compromise of a node is compromise of that node` (ADR 0004) is
now true of secrets, not only of identity.
Two files rather than one, because the mesh cannot compose a document
containing a value it discarded: `binds` carries the readable facts,
`secrets` carries the credential alone. The readable half stays readable
in the declaration; the secret half changes only when the secret does,
which makes restart-on precise. The provider gets a directory, one file
per consumer, for the same reason.
It is made once and kept — regenerating per declaration would restart
both ends on every push, and the password a provider was told to create
would never be the one its consumer was given. It is remade when either
end's sealing key changes, and both ends learn the new one in the same
push, so there is no window where half the mesh holds a dead credential.
Two tests found passing for the wrong reason, both caught because their
injection came back clean:
- the provider's copy was asserted non-empty, which reads the same
whichever column is selected. It now opens the blob with the
provider's own key.
- RotateSecret deleted and re-created; the re-create was dead, because
the next read makes one anyway. Removed, and a second path to the same
act is how two ends come to disagree.
And one real fault: three places built a declaration, and the one behind
`--json` predated credentials, so it silently produced a declaration
missing them — a difference between what `plan` showed and what anything
reading `--json` got. There is one path now.
Knowing that a machine needs the anchor's database is useless to the
program that needs it unless the program is told. It knew; nothing was
written anywhere it could read.
Two fields, mirroring contributes/receives in the other direction:
serves: {database: {port: 5432, driver: postgres}} on the provider
binds: {database: /etc/app/database.json} on the consumer
The provider says what a consumer needs to know; the mesh adds the half
only it has — which machine, and what that machine is called on the
private network. The file says, in itself, that it carries no credential
and why. A missing field looks like a bug; a stated absence looks like a
boundary.
Binding something answered on this machine writes nothing. A file saying
"it is on this node" is a fact nobody needs and one more thing to keep
true.
And two machines that share no private network are refused rather than
wired together. An app here and a database there with no path between
them is a mesh that reports itself configured and does not work — the
failure surfaces as a connection timing out, which is the slowest place
to find it. This is checkable now only because the network became
something a machine is given rather than something it has by having an
address.
One fault, found by running it: working out who is on the private network
resolved the mesh, and resolving the mesh asks who is on the private
network. It hung for two minutes. The comment above the function said not
to do that and the function did it anyway; it now resolves each node
locally, which is the right answer to the question regardless — whether a
machine is on the network depends on what it was assigned, not on what it
takes from others.
Two different things were both written `requires`. A shell, a display
server and a private network have to be on the machine that needs them.
A database does not — it runs somewhere and is reached over the network.
Both were answered the same way, so requiring a database installed
PostgreSQL on every machine that ran a web application.
What a module provides now carries a scope, the same idea claims already
use, written short in the ordinary case:
"provides": ["shell"]
"provides": [{"name": "database", "scope": "mesh"}]
A mesh-scoped requirement is answered by finding the node already running
it — never by installing it here. Choosing a machine to put a database on
is a decision with consequences, and nothing resolving a web application
should make it silently. With nothing anywhere it refuses and says which
module to assign; with two it refuses and says how to choose.
Choosing is `pin <node> <provision> <from>`, kept per node because that
is the granularity the choice has. A pin at a machine that does not
provide it refuses rather than falling back — a fallback would quietly
move somebody's data. One provider does not overrule a pin either.
Resolving a node now needs to know what the others offer, and working
that out needs them resolved, so it is two passes: the first answers only
what each node offers, the second answers everything. Nothing is ever
declared from the first.
A node's plan says what it takes from elsewhere. It is the only part of a
set that stops working when a different machine goes away, and nothing
else in that output would have said so. It is also where a credential
will hang once there is a mechanism for handing one back.
One test found passing for the wrong reason: it read pins through a join
on the provider, which hides a dangling row whether or not it was cleaned
up. It counts rows now, and bites when the cascade is removed.
`requires` said a thing must be there. It never said what to do with it,
so a web application requiring a reverse proxy had nowhere to put "this
name, this port". The two modules that needed it most went round the
outside and opened a connection to the control plane's database, which is
why every node holds a credential to it permanently.
Two fields close it:
contributes: {reverse-proxy: {host: board, port: 8080}}
receives: {reverse-proxy: /etc/traefik/dynamic/mesh.json}
The control plane collects every contribution on a node and writes them
to the path the provider named, ordered by module so the file does not
churn. Contributing to something is requiring it — asking to be published
means a publisher must exist, and a module that had to say both would
eventually say one.
The control plane does not know what a reverse proxy is and does not
write one's configuration. It delivers facts; the module turns them into
whatever it runs. That is why swapping the proxy touches nothing that
publishes through it, and why the host needs no new vocabulary — a
received file is a file.
Settings reach a contribution the same way they reach a file, because a
hostname is exactly what differs between one mesh and the next.
Two things found by running it:
- the file had a `//` header, so it said "do not edit" to a person and
failed to parse for the program meant to read it. The note is inside
the document now.
- a provider with no consumers gets an empty file rather than none. It
cannot otherwise tell "nothing asked for me" from "the mesh never
wrote it", and those want different responses.
Also `plan <node> --json`, which is how the declaration gets handed to
the host's own parser.
jschoubben