Issue 241 withdrew seven consumers in one pass; 0.1.10 refuses a file it
cannot read, and a file read whole that names nobody still withdrew
everybody. A pass that would withdraw more than withdrawAtOnce consumers
(1), or more than withdrawFraction (0.5) of those held, now withdraws
nothing and says so; while the mesh goes on not asking, one is let go every
releaseEveryMs (an hour). A consumer asked for again is kept.
An unreadable, unparsable or unrecognised contributions file was read as "no consumer asks", and the
reconcile pass then removed every consumer it had applied. On 2026-10-04 one such read made the
postgres provider drop every database on the control node. Such a pass now applies and removes
nothing and says why, once; only a file read with a given list withdraws a consumer, and every
removal is logged before it is made.
The harness says 'secret not readable yet' when it cannot open a consumer's
sealed credential, which is true and ordinary on the first pass. It said it
4330 times in three hours on the control machine while two consumers were
never provisioned, because a permission the host will never grant reads
exactly like a file that has not arrived.
After about a minute it says so plainly instead, once and then rarely, and
names what to look at: who owns the file and who this process runs as. The
ownership itself is the controller's to fix.
0.1.9.
state(name) / stdio.State(name): get, put, delete, keys and watch over the
runtime's mesh/state verbs. A watch's handler is registered under the id it
is asked with, so its current values — which arrive before the watch is
answered — and its changes reach it and no other watch.
Provision gains `derived`; the harness keeps it per applied login so remove()
is told it too, and includes it in the change hash — a provider that renames
what it derives has given the consumer a different resource.
Version 0.1.7. Rebased onto main after the bundles refactor; the record this
cites was written as 0188 and renumbered to 0201, because ADR 0188 on main is
now the bundles record.
The stdio channel is the bundle's bus: subscribe asks the runtime (mesh/subscribe), which binds the
module's durable consumer and hands each event over as mesh/event — answered once every matching
handler has taken it, as an error when one failed so the runtime does not acknowledge it; request
asks the runtime (mesh/ask). The session starts as the stdio module loads when the runtime launched
the process, because a handler module subscribes at import, before the launcher says what it
serves; initialize and tools/list are answered once it has. topicMatches as the runtime's. 0.1.6.
A launched bundle holds no bus connection. The stdio loop binds a broker whose publish asks the
runtime (mesh/publish) and resolves when the runtime has published; asking, answering and
subscribing are refused by name. Tool calls run beside the loop, so a tool awaiting the runtime's
answer to its emit does not deadlock the line that carries it. 0.1.5.
serveRegisteredOverStdio lists the registrations of the module named by MESH_SERVED_MODULE by
their own names and every other registration as <seat>.<verb>. Before, the first registration
was taken for the module's own, so a bundle registering its seat first listed the seat's verbs as
its own tools. Contributors are given the process's environment, the bundle's own. 0.1.4.
A node's runtime serves many modules in one process; a module's composed paths and words must
reach its own tools and no other's. The runtime knows which bundle made which registration, so
it says which environment each contributor gets. collectTools(env) is unchanged. 0.1.3.
The node's runtime launches a tools bundle as a process and speaks MCP over stdio to it, so a tool
may be written in any language and the transport stays in the runtime (ADR 0039's refusal, kept).
Each SDK is the loop and the tool type and nothing else: `@novox/mesh-sdk/stdio` (serveStdio, and
serveRegisteredOverStdio for a bundle that already registers the in-process way), python/, go/,
rust/ and c/ beside it, each with an example answering `greet` and the seat verb `node-lamp.on`.
Skeletons, by the operator's direction: the bare minimum one bundle per language needs to be built
and answer, proven here by each example answering tools/list and tools/call on stdio.
A lost consumer whose re-apply fails is retried at the next check with
its count unchanged, instead of waiting out a backoff meant for adapters
whose create and holds disagree. A check that fails for one consumer no
longer stops checking the consumers after it; only a timeout does.
A check that hangs no longer stalls every consumer: it times out after
30s and counts as could-not-ask, and the rest of that pass is not asked.
A consumer still not held after being applied again is checked at
doubling intervals up to an hour, and said loudly, so an adapter whose
create and holds disagree costs one re-apply an hour, not one a minute.
The consumer's password is scrubbed from every error the harness logs.
An optional holds() on the adapter is asked for every applied consumer
every minute; false applies it again. A backend that forgets what it was
given while the provisioner runs (hq issue 120) is healed within a
minute instead of failing its consumers in silence. Unable to ask is not
treated as loss. Adapters without holds() behave as before.
One name per thing, per the HQ glossary: the module/container/image/binary/repo
becomes mesh-controller, the seat the-controller, and the store+broker pair the
foundation (embedded base bundles, default template and example lock renamed with
their go:embed directives). No behaviour change — a pure vocabulary rename.
Claude-Session: https://claude.ai/code/session_01D6qtiYU3P9jk3pnAXyAFyx
Rewrites the provisioner test to the new contract (a contributions file + an
unsealed secret; the adapter is handed the mesh's login and password, and removal
follows the consumer leaving the file) and drops the seal round-trip test with the
primitive it covered. Also fixes two tool tests left stale by the per-key serving
rework (ADR 0052): invoke by module.tool, and refuse one module's duplicate name
(two modules may now share a name). Suite green: 6 pass.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
emit stamps the ADR 0047 headers — x-event-id, x-source, x-node, x-time,
content-type, and optional x-causation-id / x-schema — and publishes the
body as only the domain payload. on() reconstructs the Event from those
headers. Event gains id (the x-event-id a consumer dedups on) plus the
optional causation/schema. EventHeaders joins the contracts spine.
Supersedes the first cut that carried source/node/time in the body.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
The lighter sibling of provisioning — 1:many and broadcast, no credential,
just the broker's topic routing. A thin, audit-ready surface over the
broker's publish/subscribe:
- emit(type, body): publishes an Event carrying who emitted it (MESH_MODULE),
on which node (MESH_NODE) and when (ISO timestamp) — so a listener can
build a real audit trail.
- on(pattern, handler): react to events by topic pattern. The audit logger
is just on("#", ...).
Tested: a module emits; a targeted listener (module.umami.#) hears only its
events, the audit sink (#) hears every module's, and the metadata audit
needs is present.
The declared side — a manifest's emits/consumes, so the mesh knows the
event graph — and the audit-logger module are the next pieces.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Add serveTools(broker) — the tool runtime's core: collect every module's
registered tools, index by name (refusing a duplicate name across two
modules rather than silently shadowing), and answer 'tools.invoke'
requests by running the named tool and returning its result. Plus
listTools() for discovery and Broker.handle() (the server side of
request/reply).
Proven by test: a real async, network-calling tool is registered (as a
module does), served over an in-memory broker, and invoked by name — it
reaches its upstream and returns the computed result. So a module's tools
genuinely serve: register -> collect -> serve -> invoke -> real work ->
result. The per-node runtime process that binds the mesh's real broker
and imports the assigned modules is the thin wrapper over this.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
Per novox/hq ADR 0044/0045: the sdk holds only what rarely changes and
is shared across modules; per-module code (a client, tool impls, a
create-a-resource adapter) lives in the module.
Five areas, real and tested:
- contracts: the runtime shapes module code touches (grant, credential,
a mesh Interface, tool + envelope types) — not the manifest schema,
which the control plane owns.
- provisioner: the reconcile harness every provider shares (watch grants,
create via the module's adapter, seal + write the credential, remove on
withdrawal). A module writes only the adapter.
- tools: registerModuleTools + collectTools — the serving harness; tools
and their client live in the module.
- messaging: the Broker/Envelope/event contract over the mesh broker; the
concrete binding is provided by the hosting runtime.
- primitives: AES-256-GCM seal/unseal, semver, resolved-env access.
Compiles (tsc, NodeNext) and passes tests: sealing round-trip + wrong-key
rejection, semver, tool registration (a thrower is skipped not fatal), and
the provisioner creating then removing a sealed grant.
Claude-Session: https://claude.ai/code/session_01LrgweAeERJYBg88c5cKDzF
jochen
jschoubben
mesh-admin