Corrects ADR 0070, written an hour earlier, which had the control plane consuming
the catalogue in order to compose a declaration. That was written before the two
graphs had been told apart and creates a dependency that need not exist: a
catalogue that is down would leave the control plane unable to compose the thing
that would repair it.
The catalogue links module-versions to each other and does not know nodes exist.
The control plane links module-versions to nodes, and holds capabilities and
claims. They meet only when something is installed, and everything between them
travels as events over the broker, on durable queues, so nothing is lost when a
receiver is away.
Build order is not computed anywhere. The builder never consults the graph and
builds what it is asked for; the catalogue asks for the next build after the
previous registration, so ordering holds by construction. Its rule is a condition
rather than a schedule — rebuild once everything a module was built against is
current — which covers a chain and a diamond alike.
Left open: whether an upgrade is applied or merely noticed, and whether a module
on several machines upgrades on all at once.
Claude-Session: https://claude.ai/code/session_01D6qtiYU3P9jk3pnAXyAFyx
ADR 0070 has the init builder clone the source and does not say from where, and
ADR 0067 had rejected building at genesis partly because the forge runs on the
mesh being rebuilt. That objection binds only when those are the same mesh, which
is true exactly once.
So genesis clones from a mesh by name, and if that mesh is lost the name moves to
another that holds a copy — recovery is a name pointing elsewhere rather than a
backup being restored, and every installation adds somewhere it could point.
It names a commit and checks what it got, because the forge a mesh installs from
is the trust anchor for everything that mesh will run. On 2026-09-11 that forge
was running a cryptominer and tampering with git operations in flight; nothing was
altered, but a mesh installing during those hours could not have known that.
What relationship a mesh keeps afterwards is left open and named, so that whoever
writes the init builder does not settle it by accident: a snapshot and then
independence, or a continuing upstream for core modules.
Claude-Session: https://claude.ai/code/session_01D6qtiYU3P9jk3pnAXyAFyx
The graph had no owner: what modules are, what they require, what they claim and
what they are built against all sat in the control plane because that is where it
was written first. The control plane's own test says otherwise — anything a single
machine could answer alone is not its work, and what a module needs requires no
knowledge of any node.
So the catalogue becomes a core module beside the control plane and the builder,
owning the graph and serving tools over it. The control plane consumes it, which
is the opposite of what the tiers suggest and is therefore stated rather than
inferred.
That makes the catalogue a fourth thing that cannot arrive through the ordinary
path, so the claim written this morning that the list was closed at three is
corrected. All four are answered by one mechanism instead: the installer carries
an init builder and the core modules are built on the machine, so what is carried
is a builder rather than a result and nothing is left without a route.
Two things are open and named rather than assumed: where the init builder clones
from, given the forge normally runs on the mesh it would be rebuilding, and what
it publishes into, given the registry is installed later in the order today.
Claude-Session: https://claude.ai/code/session_01D6qtiYU3P9jk3pnAXyAFyx
A build machine was refused the build queue, and this was raised as a gap in what
a manifest can express. It is not: `builder issue` creates exactly that account,
three lines from the code being read at the time.
Kept rather than deleted, for the one real thing in it — the wrong verb succeeds
and reports success, producing an account that authenticates and can do nothing,
so the failure surfaces a layer away as a permissions error that reads like a
missing feature.
Claude-Session: https://claude.ai/code/session_01D6qtiYU3P9jk3pnAXyAFyx
Raising a scenario occupies the machine and the person who started it, and
running in the background against a working copy is worse than waiting: a run
reads that copy as it goes, so editing while it runs yields a result describing a
state that never existed.
Proposed rather than accepted. The load-bearing part is the restriction — a
request names a bed and a commit and nothing else — because a request that could
say what to install and where would make the lab a second way of installing a
mesh, which is the arrangement that just cost a year of late-found faults.
Claude-Session: https://claude.ai/code/session_01D6qtiYU3P9jk3pnAXyAFyx
The design said a module's manifest sits at a repository's root, full stop, which
means one repository per module. Nothing that exists is shaped that way: the
catalogue holds sixty-seven modules one to a directory, no code repository has a
manifest at its root, and the system being replaced has always built a module
from a repository and a path.
So the builder could be asked to build nothing that exists — pointed at the
catalogue it finds no manifest, pointed at a module's source it finds none
either. Recorded as a decision because it moves the core modules' manifests
beside their source, and corrects the design that said otherwise.
Also corrects, in the same document, how the three things the build loop cannot
produce actually arrive. They were written as though all three were carried in.
Only the control plane is: the registry is pulled from the public internet, and
the builder has no route at all — which is now stated as the open one rather than
implied to be solved.
Claude-Session: https://claude.ai/code/session_01D6qtiYU3P9jk3pnAXyAFyx
The one complete account of standing a mesh up was an integration test, and a
fixture is free to invent what it needs — which is how a registry that exists in
no production hid two faults for as long as the lab existed.
Written from what the installer does, not what it should do: genesis and joining
are separate moments, the lab runs the installer rather than describing
installing, and three things that are not true yet are named rather than glossed,
including one rule nothing checks.
Claude-Session: https://claude.ai/code/session_01D6qtiYU3P9jk3pnAXyAFyx
Found by deleting the lab's registry and giving the machines a real path out:
public images fetched, the operator's own could not be fetched at all. There is
no provision for a registry credential, no manifest field, and no step in
enrolment that establishes one.
It applies to the mesh's own store too, which today asks for nothing — a
decision that has never been written down as one, and so reads as an absence.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Both carried 'model access', which is not one of the six the reading order
knows, so neither had a place in it. 0050 — the record they extend, on the same
subject — is 'what runs on it'.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Nine references now name the digest their tag resolved to. What closed is the
immediate fault; the open questions stand, because a digest in a repository is
wrong the moment anybody rebuilds — which is the reason the design wants the
repository to name artifacts and the mesh to hold digests.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
0066 was proven on a four-node bed before it was ratified: one node setting
moved an entire domain, a routed name resolved inside the mesh and was issued a
certificate by the internal authority, and TLS verified against that authority
with no override. 08-connectivity rests on it and could not while it was
proposed.
0067 records what deleting the lab's registry exposed — that pinning quietly
required a registry before the thing that lets a mesh have a registry could
start — and the pivot that resolves it.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Both carried a topic outside the six the index knows, so neither had a place to
be read in — 0067 had none at all. Both are 'the tiers', beside 0036 (bootstrap
ends at a usable mesh) and 0007 (connectivity), which is what they extend.
0067 cited 0041 for tier 0's property; on this trunk 0041 is events, and the
record it meant is 0005. A citation that resolves to the wrong record reads as
corroboration, which is worse than a dead link.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
The doc's own rule is that the lab verifies capability by outcome, never by
reading a setting. A path out is exactly that kind of claim — a route and a
policy can both read correctly while nothing gets through — so it is asserted by
fetching something, in the table with the rest.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
A container runtime on the same workstation sets the kernel's forwarding policy
to drop, and the virtualisation daemon's own accept rules do not override it —
both are consulted and a drop anywhere is the answer. The machines then get
addresses and resolve names, because the daemon's resolver is on the bridge, and
discard every packet to anything real.
The sharpest 'available is not adequate' yet: nothing is misconfigured, nothing
logs, and it presents as every image pull hanging. A workstation that runs
containers is the ordinary case, so it is a prerequisite — verified by reaching
something, never by reading a setting.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
029 says a module providing the artifact store may not build artifacts. The
pivot shows that is the narrow case: it may not require anything the store is
needed to deliver. A route-label migration gave the registry a public name and
a route requirement, and at genesis nothing provides a route — nor can anything,
since the routing stack needs images and images need the store.
The same cycle through a door the existing wording did not cover, so the rule is
widened where the bootstrap decision states it, with a check that would catch the
next one where it is written.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
There is no installer. The complete account of standing up a mesh is an
integration test in the lab, and a fixture may invent what it needs — this one
raised a registry no production has and rewrote every image reference through
it, concealing both 039 and the fact that a first node outside the lab had no
bootstrap path at all. The bed being green said nothing about whether a mesh
could be installed.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Nine images across seven modules name a tag, not a digest. ADR 0006 forbids it
and the host refuses it by name — and the refusal has never fired in a bed,
because the lab pushed every image into its own registry and rewrote every
reference to the digest it had just assigned. The harness was supplying the
property under test. Found by deleting the harness.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
The control plane's image is built from source and pushed nowhere, so it has no
manifest digest; a registry assigns those. Pinning therefore required a registry
before the thing that lets a mesh have a registry could start — a dependency the
rule created by accident. The lab hid it by raising a disposable registry no
production has.
Proposed, not accepted.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
0056 is 'the authority is the control plane, not a database', drafted on the
in-progress record chain this branch was cut from before those three records
landed. Two files would have collided at merge, which is the kind of thing that
is cheap now and confusing later. The code written against it still says 0056
and is corrected separately.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
The real defect is same-node consumers announced the declared port instead of
the assigned/published one; the loopback observation was a stale pre-0038 build.
Fixed in mesh-control fix/same-node-provider-announced-port (c147a26) with a
regression test.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
A from:mesh provider is announced (per 018's fix) at the node's private-network
name, but its port is published bound to loopback, so consumers dialing the
announced <node>.internal:port reach nothing. Diagnosed from the lab: DB
consumers that require the database at startup crash-loop; the provider is
healthy on loopback only.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
A route contribution carries a label; the node carries its public domain; the
mesh composes <label>.<public-domain> and holds no name map. A granted route is
published into internal resolution so anything in-mesh (notably an internal ACME
authority) can resolve and reach it. That authority certifies routed names by the
same path a public one would, differing only in issuer and trusted root.
Records the decision as proposed and amends connectivity SS2/SS3/SS5 plus its Open
list with the lab findings behind it.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Extends ADR 0050: model-access, a provision, may be answered by a NODE that
hosts a model (Ollama/vLLM serving an OpenAI-compatible endpoint) as well as by
a licence record. A node-answer delivers an endpoint (base URL + model, and a
key only if the server wants one), rides the ordinary serves/binds path, and
uses no adapter; the resolver already prefers a local model over a licence. One
provision, two answers — the mesh's own model sits behind the same interface as
a vendor's. Names the mesh-scope-alongside-licence limitation and its fix.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Closes the usage-tracking half ADR 0050 left open. One row shape (0050's), recorded at two
consumer grains: the holding module (licence-level, e.g. Anthropic utilization%) and the agent
session (per-session tokens/cost, since a session IS a consumer per ADR 0026). Produced by the
vendor adapter; read on a schedule (0053); recorded as events the audit-logger keeps (0041/0042)
plus a queryable usage context store (0008); usage is not a credential and is recorded in the
clear. Reuses the session, schedule, event, and store the mesh already has. Accepted per the
user's choice to build the full feature incl. per-session token/cost.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
The recurring twin of run-once (0052): a schedule modifier on the container shape, reusing
its security bound (no new host action/shape, strictly less than an action) and reversing its
gating rule — a scheduled step runs after convergence, does not gate the apply, and a failed
run is logged, not fatal. Unblocks kometa's sync and pollers. Accepted per direction to build
the primitive now.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
The run-once lifecycle primitive: run-once:true on the existing container shape,
gated by declaration order + exit 0, idempotent by digest. No new host shape, no
arbitrary command — strictly less powerful than an action. Resolves issue 037.
Verified sound and faithful (ADR 0005/0047); status proposed->accepted.
A module can declare state but not a step that runs. mosquitto must seed its
dynsec admin into dynamic-security.json before the broker starts, or the plugin
aborts; the database providers need the same for first-boot migrations and
health gates (04-ISSUES/037). The old event-hook engine that did this was
powerful and flaky; this is the narrowest sound mechanism instead.
A run-once step is an ordinary container marked `run-once: true`: the host runs
it to completion, requires exit 0, and gates the apply on it — so what the
declaration places after it (the broker) starts only once it has finished.
Gating is by declaration order, not a resolved dependency (ADR 0005); the
completion marker is the recorded declaration digest (ADR 0018), so a re-apply
does not re-run it unless the declaration changed. No new host shape and no
arbitrary host command: strictly less powerful than an `action`.
Points 04-ISSUES/037 fixed-by/amended-design at the record; index regenerated;
records.py and index.py pass.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Your decision, ratified: the media library (and shared/pre-existing data) is
operator-owned and external; a module declares access, not ownership; the host
mounts but owns nothing (no create/chown/reconcile/remove); an absent accessed
path is refused clearly; several accessors co-resolve. Status proposed->accepted;
index regenerated (records + index checks pass). Implementation lives on the code
branches (mesh-control/catalog/host), held for merge after the convergence fix.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Resolves 04-ISSUES/036: eight media modules each declared the shared
library and download directories as their own resources, and the
resolver's duplicate-owner refusal — right in general — would refuse the
stack's only sensible assignment the first time two landed on one node.
The decision, from the operator: shared, pre-existing data is
operator-owned and external. The mesh does not create, chown, reconcile
or remove it. A module declares it needs access to such a path (read or
read-write); the host mounts it and owns nothing. Several modules
accessing one path is normal — the duplicate-path refusal is about
ownership, not use. An accessed path absent at apply is refused clearly,
not created. Extends ADR 0030: the third case the host had no word for,
what it neither made nor configured and must not touch.
Point 036's fixed-by/amended-design at the record; mark it located in
mesh-control, mesh-catalog and mesh-host. Regenerate the decision index.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Turn the completed vendor-agnostic analysis into HQ design. The model-access
provision stays one vendor-blind interface (extends 0024/0027); the
vendor-specific lifecycle moves into a per-vendor adapter keyed by the licence's
`vendor` field, mirroring registrar-scoped public-dns providers (0044), named at
the consumer's real coupling per 0040.
The crux is the sealing-vs-central-rotation carve-out: for refreshable-grant
vendors only, the manager node holds the refresh token encrypted at rest (a
bounded, declared exception), access tokens sealed per holder, refresh stripped
on delivery. Static-key vendors keep full sealing.
Amend 03-DESIGN/01-to-be/14-model-access.md with the adapter generalisation as a
proposed section (prose + diagram, no code); regenerate the decision index.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Init's 003 was already resolved with its own consolidated attribution
(03-DESIGN/01-to-be/08-connectivity.md); the re-homing overwrote it with the
session's ADR-0045 firewall attribution. Init is canonical and the firewall
decision is recorded in the ported ADR 0045 regardless, so 003 is restored
untouched. No initialization record is modified by this reconciliation.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
The real work lived on initialization (consolidated decisions 0001-0038, the
fuller issue set 001-031, the control-plane/substrate/node-lifecycle/delivery
design, research 011/012, the checks tooling). main had diverged onto a stale
base and only carried this session's genuinely-new work. This merge makes
initialization's tree canonical on main; this session's 11 new ADRs and 6 new
issues are re-homed on top in the following commits. initialization is recorded
as a parent so its history is preserved in main's ancestry.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Surfaced converting the catalogue: a module can declare things that exist
(dir/file/network/container) but not a step that runs at a point in its
lifecycle. mosquitto's dynsec admin client must be seeded before first start;
the DB providers have nowhere for a migration or health-gate; it is the timing
face of issue 011. Framed as a missing module capability, not a defect. Records
the prior-art event hooks and their real warning — powerful but complex and
flaky — so the resolution avoids rebuilding that. Ends in open questions
(run-once resource vs general lifecycle hook, where the code runs, idempotency).
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Numbers 008 and 009 were taken on main after this branch was opened
(008-provider-runtime-has-no-seal-key, 009-runtime-config-change-does-not-restart,
both merged). Renumber the seed-file-wipe and shared-directory issues to the next
free numbers so merging records two more issues rather than duplicating two.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Records the branching-and-merging workflow for code changes across the mesh
repos, written against a failure it names: branches and MRs opened per unit of
thought, treated as done when opened not merged, and named differently per repo,
so they pile up unmerged — one session left sixteen to consolidate by hand. The
rule is one feat/<slug> shared across every repo a feature touches, isolated in
.work/<slug>/<repo> worktrees off main, pushed and opened as one MR per repo only
when the whole feature is done, then merged promptly. Adds the ground-rule
pointer in AGENTS.md and the row in the process overview.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Brings the independent ADR branches (0044-0052) onto one branch so hq lands as a
single MR, and ratifies the five that were still proposed — 0017, and 0049-0052,
which are implemented and green in the lab. With 0053/0054 already accepted here, the
whole ADR chain 0044-0054 is accepted on this branch.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
ADR 0054 accepted with option E (a declared slug). Issue 010 resolved: the login fits
via the slug, and the minted secret shrinks to 40 chars for S3's secret-key limit —
both halves of an S3 credential now fit the tightest backend.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Implementing A (bound the identity at 20) revealed the readable budget is node+module
<= 14 chars — so tight that the catalogue's own test names (workstation+keycloak, 25)
compact to an opaque hash. B's fallback would fire for the common case, not the rare
overflow, inverting A+B into mostly-opaque identities. Option E — an optional short
slug a module/node declares, preferred over the cleaned name — is the escape hatch B
wanted to be without the opacity: legible because a person chose it, and it makes an
early refusal palatable (refuse on the slug field, not the machine's name). B dropped;
E recommended over a bound of 20, composing with C later if needed.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Sketches the options for issue 010: the mesh's identityLimit (63, postgres's) is not
the shortest among the backends the derived name reaches — S3's is 20 — so CheckIdentity
lets an over-long access key through and minio fails at provision time. Options: bound
by the true minimum and refuse at assignment (recommended, with a compact fallback held
in reserve), per-interface bounds, or a provider-generated identity (rejected — breaks
"the mesh says the identity once"). Links issue 010 to it.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Found doing the per-backend provider e2e: redis and postgres accept the mesh's `as`
(mesh_<node>_<module>) verbatim, but minio's S3 access key is capped at 20 chars and
`as` is 22, so the provisioner cannot create the service account. `as` is doing two
jobs — a stable identity the two ends agree on, and a literal identifier a backend
must accept — and those are not always the same string.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
ADR 0053 accepted; adds the scope boundary the umami rework surfaced (credential
provisions vs data provisions — analytics' generated siteId return is left to a
separate decision) and records the lab proof. Issue 008 marked resolved: the sdk
harness and the four adapters are reworked, the symmetric seal removed, and
provider-uses-mesh-credential is green.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Issue 008's trace confirmed the premise in control-plane code: the mesh already
mints one password per consumer/provider pair and delivers the provider its copy
(SecretFor/SecretsFrom/grantsFor -> Grant.Sealed; the receives contribution carries
As + Secret). The provisioner's symmetric seal is an orphaned, contradictory second
model. ADR 0053 corrects the provider contract in one place (the sdk harness):
providers create the resource with the mesh-supplied login and password and drop
seal/key/return entirely. Reframe 008 as contract-first (every provider, not four).
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
A cross-repo trace showed nothing writes the provisioner's grant-request files,
nothing reads its sealed credentials, and no consumer unseals — while the mesh
already mints and delivers provider/consumer credentials asymmetrically with no
shared key. The fix is to drop the symmetric seal and have providers consume the
mesh-minted password, a breaking provider-contract change that wants an ADR.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Found rolling the runtime out to the catalogue: a module's runtime reads its
settings-merged config file once at start, but a container is only recreated on a
spec change, and file content is not part of the spec. So updating settings
re-renders the file and nothing re-reads it — ADR 0051's "on the fly" holds only
for config set before first start. Services have restart-on; containers do not.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Found building the module-runtime vertical slice: a provider's provisioner
requires a seal key it has no way to receive, and the consumer no way to obtain
the matching one. The runtime cannot come up as delivered.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
The runtime-model gap the review found. A module with tools or events runs one
container — the tool runtime carrying its code — holding the one scoped account
ADR 0048 gave it. A node-wide runtime can't: it would hold the union of every
module's permissions, the isolation 0048 draws. So per-module: one module, one
process, one account. Tools served per key (serve.<tool>) so a caller names a
tool and only its module answers (superseding a shared tools.invoke); events in
the same process under the same account; the runtime image is the tool runtime
plus the module's code (the audit-logger's shape, made the rule). A plain
service module runs no such process. A provider's provisioner is a runtime too —
which is why a provisioner that emits must carry a broker credential or not emit.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
A module is assigned to a node (there is no mesh assignment; 'mesh' is a scope).
The manifest is what the module IS, plus defaults; the configurable values are
settings, carried by the assignment — per-node or mesh-wide, applied at
resolution, changeable live (what a meshboard edits). Extends settings from a
config file's content to the manifest fields marked settable: foremost
listens.from (postgres from:mesh by default, from:anywhere per node — the
firewall follows), and a provider's own config (a registrar's zone/domain/
ingress). Static config in a manifest is config in the wrong place: it cannot
vary per node and cannot change without a rebuild.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
The manifest refuses unknown keys (DisallowUnknownFields), 'from' is the field
that scopes a port and it is rendered to nftables (AsNftables), and the firewall
module applies the rule set. The chain from a declared scope to a dropped packet
is closed. Amended-design: ADR 0050.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
0049: a public name is provisioned like any capability — a module requires
public-dns and contributes its host; a neutral interface answered by
registrar-scoped providers (cloudflare-dns, route53-dns) that create/remove
the record pointing the name at the mesh's public ingress. Pairs with route
(the proxy) and a public cert (the proxy's ACME).
0050: answers the firewall question. The firewall is NOT a provider like the
proxy — it is a machine's own filter, derived by the host as the sum of what
its modules declare they listen on, with 'from' the whole of public-vs-internal.
Enforced both ways, unknown keys refused — closing 04-ISSUES/003. A public
service is exposed through the proxy (listens from:mesh + requires route), not
by opening its own port.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Events (0046) and their wire (0047) left open how a module reaches the
broker. The code has no generic module broker-account: only node and
builder scopes exist, so emits/consumes are enforced by nothing — a
manifest declaring a scope the broker does not draw (04-ISSUES/003).
Decides: on assign, a module gets a broker account whose permissions ARE
the manifest — read on mesh.events + its own queue bound to consumes;
write to mesh.events under module.<self>.* only; nothing else. Consuming
'#' is a deliberate, auditable grant. The account is what makes the
declaration a rule the broker enforces, not a comment.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
The wire contract ADR 0046 left open: two topic exchanges (mesh.events,
mesh.rpc, kept apart so # is a clean audit); the routing key as the event
type namespaced by origin (module.*, mesh.*, node.*); metadata in AMQP
headers (required x-event-id/x-source/x-node/x-time/content-type; optional
x-causation-id/x-schema; unknown x- headers ignored) with the body only the
payload; persistent messages; per-consumer durable dead-lettered queues
with prefetch; at-least-once with idempotent consumers (no false exactly-
once). The precedent is ADR 0043 for declarations.
Supersedes the sdk's first cut (metadata in body -> headers); that and the
queue config are code to align in mesh-sdk and mesh-tools.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
A module emits and consumes events, both declared (emits/consumes),
parallel to provides/requires. Events are 1:many, broadcast, credential-
free — no provisioner, just the broker's topic routing — so most inter-
module reaction should be an event, not a provision. Every event carries
source/node/time so it is auditable; the audit logger is just a module
consuming '#', no privilege. A consumes for an event nothing emits is a
dangling edge and refused, like requires. One per-node runtime serves
tools, provisioning and events alike.
Extends ADR 0045; builds on ADR 0001 (the broker) and 0044 (emit/on are
stable sdk surface; the binding and runtime are not).
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
A module is one self-contained piece of software the mesh installs and
manages; the software is its identity, and capabilities/seats/provisions
are the relationships between modules, not what a module is. Records the
three relationships (shared seat, exclusive seat, provide/require), that
interfaces are mesh-owned and providers adapt to them, and the naming
rule: draw the interface at the consumer's real coupling — neutral where
the coupling is thin (analytics), protocol-scoped where the consumer
speaks a protocol (postgres/mssql/mongodb), never false genericity.
Supersedes 0017 (domain grouping — wrong axis), refines 0002, generalises
0027's protocol-not-product rule.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Supersedes ADR 0030's 'types, not behaviour' line for mesh-sdk. The
boundary is change-frequency, not kind: the SDK holds the stable spine
(tool-serving harness, messaging/event framework, contracts, core
primitives) and refuses per-module clients, per-module tool code, and
anything volatile — because those are what turned hal/sdk into constant
maintenance and made every edit rebuild every module.
States the rule (frequent AND cascading is the disease), why the root
cause was intra-module feature-sharing leaking into inter-module
coupling, and where per-module shared code lives instead (in the
module — a shared file, or a module-local sdk for the few large ones).
Updates repos.md's canonical mesh-sdk description to match; leaves 0030
untouched (immutable).
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Both from the 2026-09-02 review of the catalogue examples, and both
design gaps rather than defects in a file: a seed file the host
reconciles back to empty over the grants that grew in it, and a module
stack refused co-assignment because six manifests each own the
directories they exist to share. Fixing either in place would have
been picking an answer the records do not yet hold.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Declarations queue and the host applies all of them, oldest first, so a
machine pushed five things in a minute spends five applies becoming the
last one. Correct every step — each declaration is the whole machine —
and wasted in all but the final step.
Only visible since a report names its declaration: the reports arriving
were about ever-older ones, while timestamp comparisons used to happen
to pass. The fix is consumption order, not the queue; the open question
is what a superseded declaration's report should say, because silence
reads as disobedience and "applied" would be a lie.
The five declarations cover relations; a module is more than its
relations. Add the facet-by-facet coverage table so the effort cannot
conclude while tools, verification and contributions are unplaced —
and weigh each candidate gap rather than adopting it: contributions
probably dissolve into declared resources, mandatory verifiers risk
trivial ones, and the tool surface is the one facet with no home.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Composing a declaration signed the machine's certificate anew each
time, and a signing carries a fresh random serial — so what the mesh
would send differed from what it had sent by one byte, for ever, and
every machine carrying a certificate stood eternally waiting.
Third find of the same rule: issued once and kept. The port had it, the
secret had it, the certificate composed fresh on every asking — and the
keeping column had existed since the serving-key migration, written by
nothing, the same shape ReleasePorts was found in.
Found by keeping the scenario standing and diffing two plans seconds
apart: one line, where four theories had none.
The registry module names its image by digest, the way the bundle names
the three a first node starts from, and goes in as a manifest. The lab
now walks that path and passes.
A manifest that provides the store and also builds artifacts is refused
at parse. The lab had been passing only because its Docker Hub stand-in
quietly received the push — a prop covering for the thing under test.
Installing the module that provides `artifact-store` requires something
that provides `artifact-store`: its image is mirrored in, mirroring
publishes to the store, and the builder refuses to run without one.
Never seen, because the lab always has a registry standing before the
mesh asks for one, and so does any mesh built on a machine that already
had one.
What it blocks is larger than a registry. The store holds images and
packed archives both — it is the module catalogue in artefact form — so
until it exists a mesh can run only what its bundle already carries.
The substrate record already answers it. It asks of each candidate
whether it can grant itself the thing it provides: the store cannot
create its own database, the broker cannot create its own virtual host,
and the registry cannot grant itself a repository. So the registry
module names its image and is never built.
With a limit worth saying out loud rather than discovering: a module
providing the store may not build artifacts of its own, a UI or a tool
server included, because there is nowhere to put them until it runs.
Such a registry is two modules.
Enforcing "declare what you mount" refuses the builder, which mounts the
container runtime's socket. That socket is not the builder's data: it
exists already, the machine owns it, and declaring it as one of the
module's directories would be a lie the host would act on.
Two kinds of mount are spelled identically today — the directory my data
lives in, and a machine facility I was granted. Until a manifest can say
which, the fourteen declared mounts are right by coincidence, which is
what this issue was opened about.
Recorded rather than decided: separating them is new vocabulary, and
inventing it to turn a check green is how a mechanism nobody chose ends
up load-bearing.
The fourteen undeclared mounts are declared. More to the point, a
manifest that does not declare one is now refused: they were right by
coincidence, and a checklist nothing enforces is a checklist that is
true until the next commit.
Refused in the control plane, because the machine cannot tell the
difference — asked to mount a path that does not exist, it makes the
directory, which is a thing it is perfectly able to do.
A module now says its port once, in `listens`, and the container's
mapping, the rule set and what a consumer is told are all derived from
one assignment. The three hand-written copies that agreed only because
one person wrote them are gone.
The half that made this an issue rather than an inconvenience was that
the substrate is not a module: nothing in the mesh had heard of its own
store, so it handed a database module the port the store already had.
The machine now says what it carries, and the mesh assigns around it.
Ports the protocol fixes became claims, which needed no new mechanism —
the mesh already had one for what is singular on a machine.
Left open, and unchanged by any of this: whether a module should publish
to the machine at all. Assignment makes publishing safe without making
it necessary.
Jochen's call, and the right one: a module cannot choose a port well,
because it is written once and assigned anywhere. Any number it picks is
a guess about a machine it has never seen, and two modules guessing the
same number is not a mistake either of them made.
Writing it up turned up something the issue had missed. The same number
appears three times in every module — the rule set, what a consumer is
told, and what the runtime publishes — and nothing checks that they
agree. They agree today because one person wrote all three. A module
whose `serves` said one thing and whose container published another
would resolve, compose, apply, and hand every consumer a port that
answers nothing.
So the decision is one source with the other two derived, and an
assignment made once and kept, as a credential is.
The part that needed thought is ports that cannot move — mail on 25,
submission on 587. Those become claims, which is what the mesh already
has for what is singular on a machine. Two modules wanting 25 is the
same shape as two wanting the seat, and gets refused by name at
assignment rather than by a container runtime at apply. That makes this
mostly a matter of pointing an existing mechanism at ports.
Left open: whether a module should publish to the machine at all.
Assignment makes publishing safe without making it necessary.
Found by fixing 027 and pushing again. The declaration is now accepted
and the database container still cannot start: the mesh's own store
holds 5432 on that machine, and the module publishes 5432.
Nothing catches it because the substrate is not a module. It arrives
from the bundle before there is a mesh to ask, so the control plane has
never heard of the store and does not know it holds a port. Resolution
can compare modules with each other and cannot compare one against what
the mesh is built on.
Nor does it compare modules with each other. A port is exclusive on a
machine in exactly the way a claim is, and the mesh has a mechanism for
that which ports do not use.
It has been met before: the end-to-end test that exercises a real
database publishes 5433 rather than 5432, inline, with nothing saying
why. That is how a constraint becomes folklore.
The open question is bigger than the bug. Whether a module should
publish to the machine at all decides how a consumer reaches it, and
changes what `serves` means.
Found by the forge failing to start. I had put `restart-on` on nine
containers so they would pick up a rotated credential; it belongs to a
service, and the host refused the whole declaration.
Removing it fixes the modules and leaves the reason I reached for it.
The mechanism is written against exactly this, in the host's own words:
a running service does not re-read its configuration, so replace the
file, find it already running, do nothing, and the machine keeps
behaving as before while every check passes. Every word of that applies
to a container, and nearly everything the mesh runs is one.
The cost is concrete. Rotation replaces the file and tells the provider
to accept the new credential. A provider reconciles, so it takes it. A
consumer is usually a container, so it does not — and the two ends hold
different passwords, which is the fault ADR 0001 records costing two
days. The test that proves rotation works uses a consumer that reads the
file on each attempt, so it does not meet this.
Two things a fix has to keep: it stays declared state rather than a
command, because the link may not carry an action; and where an env-file
changed, the honest verb is recreate rather than restart, because a
container's environment is fixed at creation.
Three corrections, two of them to things I wrote today.
026 is the serious one. Four modules mounted fourteen host paths nothing
declared — the mail spool, the databases, the object store's data. The
runtime creates those as root, so owner and mode go unapplied, and the
rule that keeps a directory holding data the mesh did not put there is
written in terms of declared directories. It reached the configuration
and missed the data. The cause was carrying compose files across: a
container shape that can express one gets filled in like one.
025 claimed nothing turns a tag into a digest. That is false, and the
answer was designed and built before I wrote it. A module names an
artifact, not an image, and `kind: upstream` mirrors somebody else's
image into the mesh's own registry, pinned by the digest it lands with.
The two-document split the issue described as the shape of a fix is the
design. Pinning twelve images by hand was treating the symptom, and left
them pointing at a public registry rather than the mesh's.
And the image store was written up as something the mesh does. It is an
ordinary module — considered for the substrate and removed, because the
test is whether the control plane needs it before its first instruction,
not whether it can grant itself one. So somebody's own registry is the
same module as the mesh's.
Recorded against phase 3, because the phase note said running them
needed images stocked and provisioners built — and missed that not one
of them named an image that exists. Sixty-four zeros where a digest
belongs, eighteen times, parsing and resolving perfectly.
The forge now runs: on a database another module provides, with a
password it did not choose and a connection string it could not have
written itself. First of these descriptions to be started rather than
planned, and it exercises the whole of the credential work.
What remains is the mechanism rather than the data. Nothing turns a tag
into a digest as part of the mesh's own work, so it was done by hand —
which is what the issue says a person should not be asked to do. Asking
a registry takes a second and pulls nothing, so the main argument for
leaving it undone is gone.
The refusal landed, and the examples pin images that exist. What is
still missing is the part that makes it unnecessary: nothing in the mesh
turns a tag into a digest, so it was done by hand — which is precisely
what the issue says a person should not be asked to do.
Recorded because the resolution mechanism turned out to be trivial:
asking a registry what a tag points at takes about a second and pulls
nothing. That removes the main argument for leaving this open.
Also records the two faults that fell out of pinning for real. The mail
system named seven repositories that do not exist, because it publishes
to a different registry than the manifest assumed, and one of the seven
had been renamed upstream. Nothing checking only the shape of a
reference could have found either.
The code and its tests went in hours before the record was touched, so
an issue that read `located` had been fixed all along. That is the exact
failure the frontmatter exists to prevent: status is meant to be
answerable from the record rather than by reading the code.
Closed with the commit that did it, and cross-referenced to 022 and 023,
which came out of the same mistaken instinct — treating the machine as
a boundary, then as an identity, then finding a consumer had a password
and no name to present with it.
The module descriptions sit in `examples/` inside the control plane, and
that name has been doing harm: everything there reads as a sketch, and
one shipped naming a container image nothing builds. A directory called
the catalogue would have made "does this work" the obvious question.
The shape of the answer turns on one measurement. Of the 126 modules in
the system being replaced, 47 are software in their own right — the
largest is 182 source files, and a speech-capture module carries a whole
daemon. Another 44 ship helper scripts. Only 35 are a description and
nothing else.
So a catalogue cannot be a folder of manifests, because two thirds of
modules are programs. That splits them four ways, and only two of the
four belong in a catalogue: things the world made that we describe, and
packages with some files. What the mesh is made of stays in the
repositories that build it. What we wrote keeps its description beside
its code, in the same commit, because nothing else can stop the two
drifting.
The mesh's list of modules is a table, not a repository, and it already
records where each module came from and at which commit. Nothing needs
inventing for modules from anywhere; a repository of ours is just the
source we curate.
The check that a description is valid should move to a command on the
control plane's binary. Today a test reaches into the control plane's
internals to parse manifests, and another reads its build file to check
images exist — two jobs tangled. A command would also give the same
check to somebody describing their own application, which is the case
that matters most and has none.
Left open: how a provisioner's image gets published and pinned, and
whether thirty-five install-a-package modules deserve to be modules at
all.
The cause was one line. Machines get a systemd-networkd unit with a
static address, so networkd finishes and reports the link configured.
The registry ran `ip addr add` inline, which leaves networkd waiting to
configure something it was never told about — and
systemd-networkd-wait-online has an infinite timeout.
So network-online.target was never reached and everything ordered after
it never started. On these machines that is Docker, so `docker load`
blocked on a socket whose daemon was queued behind a target that would
never come, and three bounded timeouts stacked to thirty-five minutes.
These machines have no DHCP by design, so that wait was never going to
end.
The hypothesis in this record was wrong and the record now says so.
Stocking had just been changed, so stocking looked guilty; stocking
takes 34 seconds and always did, timed directly before changing
anything.
Fixed with two things that made it cost hours instead of minutes: an
image placement now waits for the runtime and refuses after 120s naming
what systemd is waiting on, and the end-to-end test passes onProgress —
the raise reported every step and the test discarded it, which is why
thirty-five minutes and four minutes of silence looked the same.
The suite then ran to completion, 23 of 24, the one failure a check of
its own flagging a path as a credential because `/` is in the base64
alphabet.
Also recorded: a redirected log lags, because Node block-buffers stdout
to a file. Read as a stall twice, the second time right after the real
fix — where a buffering artifact argues the fix did not work.
Seen twice today. Once mid-run: thirteen passes, then the process ended
with no summary, no failure and no receipt. Once from the start: the
first test ran 35 minutes against a measured 4.5 and was still running
when it was stopped.
Ruled out rather than assumed: not memory (84 GiB free, no OOM), not the
daemon (the stalled machine answered `incus exec` immediately), and not
the changes under test — the anchor VM had no host log and no
containers, so the run never reached placing the host.
What changed just before is that the rebuild went from two artifacts to
six, and every one of them is pushed into the scenario's registry, which
is the step the second stall sat in. Recorded as what changed, not as
the diagnosis.
The reason this is an issue and not a slow test: the suite prints
nothing between starting a scenario and finishing its first test, so
four minutes and thirty-five look identical from outside, and the only
recourse is to guess. That is how a workstation was left unbootable in
August. And a run that ends silently after thirteen passes is a run
somebody may believe.
023 is fixed, so the design faults are gone and one concrete thing is
left: the realm provisioner does not exist. Its manifest named an image
nothing builds and no program backs, which has been removed — a manifest
describing a program nobody wrote is the same mistake as the credential
files that could never be read.
Keycloak's manifest now says what is true today, and the gap is loud: it
no longer claims to provide oidc-client, so a consumer asking for one is
refused by name at plan time instead of resolving cleanly and waiting
for a client nothing will create.
The provisioner should be written against a real Keycloak in the lab
rather than from the API documentation. The object store's took three
corrections that only a running server produced.
Tool servers — 56 modules, over half — were written up as the biggest
missing thing. They are expressible with what exists, and the first
framing was wrong in a way worth keeping: a module provides `tools` and
the session requires them does not work, because a requirement has one
answer and 56 modules offering tools would be 56 answers.
Turned around it fits exactly. The session provides `tool-host`; every
module offering tools requires it and contributes where its tools are.
Many-to-one is what `contributes` has always been, and the session
receives all of them in one file. Verified by resolving it rather than
by reading the code.
It only became possible today: until 022, several modules on one node
requiring the same thing was refused outright. Worth noting because it
means the credential fix bought more than credentials.
What remains is a decision about what a tool server is, which is work
rather than a missing shape.
The entry stays in the list rather than being deleted — a checklist that
quietly loses its biggest item reads as though nobody looked.
jschoubben