Without a judge cache the replays still ran mesh-lab's checkout in the workspace
with the container runtime's socket, so a self-judging pull request could write
a test into them (hq issue 709, review 12303). They now run from a copy in a
container volume of the check's own, made before anything of the change runs,
with a fresh home and throwaway caches, removed after the check.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
mesh/merge-gate head 0214b45469ec merged into trunk 76304d9db817 — pass: builds build-agent, mesh-controller → ace, g14, novox, shanks; no bus step; every …
The replays hold the container runtime's socket, yet took their HOME, Go
environment, build cache and module cache from the workspace a pull request's
own script writes, so a planted -toolexec hook ran as root on the build node
(hq issue 709). They now run from a copy of mesh-lab made before anything of
the change runs, with a fresh home and the judge cache's caches, as a judge is
built (ADR 0358 §6).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
mesh/delivery-group group fix/704-release-branch-judge rejected: its order contradicts itself: novox/mesh-catalog@47ef2644aca1 and novox/mesh-controller@2bee29…
mesh/merge-gate head 2bee290eea56 merged into trunk 76304d9db817 — pass: builds build-agent, mesh-controller → ace, g14, novox, shanks; no bus step; every …
Review of #329: a judge's home made with a temporary directory inside the build
agent is resolved on the host by the container runtime, leaving a cold cache
behind on every check, and a kept judge was lost on every redeploy. Judge builds
now use the judge cache the build agent declares (MESH_JUDGE_CACHE), mounted into
judge builds only: one warm Go cache, the kept judges, and a fresh home per check
removed after it (hq ADR 0358 §6, issue 704, issue 482).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review of #329: a judge built with HOME in the workspace runs git under a
configuration a pull request's script can plant, so a hook could rewrite the
judge's sources before they compile and the result be kept under an honest key
(issue 482, issue 704). Every judge now builds with its own HOME, configuration
and Go cache outside the workspace, no git configuration and no VCS stamping.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review of #329: mesh-lab's replays run with the container runtime's socket, so
a same-named branch must never be what runs there; the shared Go cache and HOME
are writable by a pull request's own script, so the judge kept for a head must
not be built from them (issue 704, issue 482). A forge answer naming no branch
is an error, and a trunk is the branch a repository follows, not a name.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
mesh/merge-gate head ea7d91bff190 merged into trunk 535c7641ac41 — pass: builds build-agent, mesh-controller → ace, g14, novox, shanks; no bus step; every …
ADR 0358 version 4 asks the verdict to say whether mesh-host's release-branch
validator or the vendored one judged, and checks that a release branch's own
pull request into its trunk is judged as any trunk pull request (issue 704).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
mesh/merge-gate head 4a75cc51dda6 merged into trunk 535c7641ac41 — pass: builds build-agent, mesh-controller → ace, g14, novox, shanks; no bus step; every …
A release branch could never carry a change that needs another repository's
release-branch change: its pull requests were judged by the running controller
and against the other repositories' main (hq issue 704, ADR 0358).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
mesh/merge-gate head 060f2a9ca5fd merged into trunk 04730b54a8f2 — pass: builds build-agent, mesh-controller → ace, g14, novox, shanks; no bus step; every …
The gate and the repository's check mount the workspace read-write, so a
build from that tree could hold what the tests wrote under a key that
describes the merge. A check now clones the merge outside the workspace
before anything of the change runs, and builds from that. A check keeps,
and a walk takes, only a build an earlier check of the same pull request
at the same head made: another pull request's build, or an earlier
head's, is never registered. The walk reads the verdicts once
(novox/hq issue 703, review of #328).
mesh/merge-gate head 65508d9df9f4 merged into trunk ca3f43c400ba — pass: builds build-agent, mesh-controller → ace, g14, novox, shanks; no bus step; every …
mesh/repo-check head 65508d9df9f4 merged into trunk ca3f43c400ba — pass: THE CHANGE ALTERS ITS OWN CHECK (merge-check.sh): main's version judged it; the ch…
Store tests skip without a store and no agent may start one, so each failure
cost a 15-30 minute review round. The try verb asks the build seat for the
repository's merge-check.sh on the ref merged into the trunk, as a check in
the build queue, and reads its outcome from the build's log: no status on
the forge, nothing registered, nothing built for a merge.
mesh/delivery-group group feat/350-one-build-per-change checking: composed check out of date: novox/mesh-controller@cd49b50c4075 was judged against trunk bff5b…
A merge built every moved module a second time although its check had judged
the same tree (issue 649). The check now builds into the scratch namespace
under a reuse key (source fingerprint plus the builder's own digest), and the
walk of the trunk commit copies a build of the same key by digest instead of
building, saying per module why when it cannot (novox/hq ADR 0350).
A check judged the head alone, so two pull requests that each passed could
break the trunk together (node-engine #66 on 2026-10-09), and a recheck after
the trunk moved judged the same head again. The build seat now merges the head
into the trunk's head in its own clone, never pushed, fails a head that does not
merge cleanly naming the files, and the verdict names the head, the trunk
commit and the commits read beside it, for the forge and mesh-delivery to weigh
against the trunk's next move (novox/hq ADR 0323, issue 341).
mesh/merge-gate pass: builds build-agent, mesh-controller → ace, g14, novox, shanks; no bus step; every machine composes with the change as it did without …
A build wrote .npmrc into its source tree and never removed it, and its clone
recorded the URL's userinfo; a later check's container mounts the workspace as
HOME. The .npmrc and the tree now go when the build ends, clones record their
URL without userinfo, and a check first removes any .npmrc an older builder left.
mesh/merge-gate pass: builds build-agent, mesh-controller → ace, g14, novox, shanks; no bus step; every machine composes with the change as it did without …
The toolchain container mounts the workspace as HOME, so the credential kept
there, and a clone's recorded userinfo, were readable by any pull request; its
output is kept on the bus for days. The credential now lives in a private
directory outside the workspace only while cloning, clones record their URL
without userinfo, and every line said to the build's log and the verdict
passes a redactor copied from the journal tool (sharing it: issue 471).
mesh/merge-gate pass: builds build-agent, mesh-controller → ace, g14, novox, shanks; no bus step; every machine composes with the change as it did without …
A long run's logs pushed the --- FAIL lines out of the last 200 lines the
summary was named from, so a failed check could not say which test failed.
Pick the lines that name a failure as the output streams, keep them whole at
the end of the verdict's report, and say the whole output in the build's log.
mesh/merge-gate pass: builds build-agent, mesh-controller, route-proxy → ace, g14, novox, shanks; no bus step; every machine composes with the change as it…
The gate passes only the manifests a change touches, so a module that starts using the operator channel
was refused for a declaration it could not see. Over every manifest it stays a refusal.
mesh/merge-gate pass: builds build-agent, mesh-controller, route-proxy → ace, g14, novox, shanks; no bus step; every machine composes with the change as it…
A change could delete or gut its merge-check.sh and skip its own tests, and a
branch cut before the script escaped them. Where the base branch holds one, its
script, declared parts and toolchain run against the change's tree, and the
verdict says when the change lacks or alters the check (novox/hq issue 310).
mesh/merge-gate pass: builds build-agent, mesh-controller, route-proxy → ace, g14, novox, shanks; no bus step; every machine composes with the change as it…
On the control node the store-bound packages of the controller's suite ran
five to seven times slower than on any other holder, and every controller
check that landed there ran past the suite's thirty minutes: a throwaway
store flushing to a disk the mesh's own store, bus and forge keep busy.
A store that lives for one check needs no crash safety.
A holder recreated mid-check left the check's store and bus running, and
the redelivery went to another machine, so nothing removed them: eleven
pairs across four machines. A starting holder has taken nothing, so every
container labelled with an ask of the seat is an earlier holder's.
mesh/merge-gate pass: builds build-agent, mesh-controller, route-proxy → ace, g14, novox, shanks; no bus step; every machine composes with the change as it…
mesh/delivery covered: a later merge that contains it was delivered: novox/mesh-controller@6faf701656fe (merged as c714d077 into main, walk plan-17914198…
mesh/merge-gate pass: builds build-agent, mesh-controller, route-proxy → ace, g14, novox, shanks; no bus step; every machine composes with the change as it…
mesh/delivery stopped: covered: its merge cdbbd9ec4bd0 is contained in cec797e996b2 which is delivered (walk plan-1791499975016484055); closed on the ope…
mesh-lab's merge-check.sh runs in the TypeScript toolchain, which holds no Go
compiler, so its replays register was never compiled before a merge and a
broken one would have merged green. A script now declares a further part with
'# mesh-check-also: <toolchain> <script>'; the build seat runs it in that
toolchain's own container, and mesh/repo-check passes only when every part does.
mesh/merge-gate pass: builds build-agent, mesh-controller, route-proxy → ace, g14, novox, shanks; no bus step; every machine composes with the change as it…
An ask redelivered after the build agent stopped mid-check (the rollout it was checking updated it)
found its own throwaway store under its name, and the check said it could not run (mesh-controller#105,
build-1791331512096605198).
mesh/merge-gate pass: builds build-agent, mesh-controller, route-proxy → ace, g14, novox, shanks; no bus step; every machine composes with the change as it…
A check asked by the controller before its own build of the issue 285 fix was registered was judged by
the controller the mesh ran then, which passed 0 of 4 composing. The judge is the running controller by
design, so the rule is read where the verdict is taken too: from the machines the verdict lists.
mesh/merge-gate pass: builds build-agent, mesh-controller, route-proxy → ace, g14, novox, shanks; no bus step; every machine composes with the change as it…
The gate composed 0 of 4 machines with the change and without, and passed every change: the store it
raised held each module's bus credential but no account for it (issue 203's refusal), no outward links
(so no filter could be composed), and refused settings the mesh holds. Now the account is minted with
its credential, the facts carry each machine's outward links (a stand-in for an older snapshot), the
mesh's layers are kept as held, and a withheld path keeps a path's shape. A machine the mesh composes
that the gate cannot raise makes the verdict an error, never a pass; the verdict alone is on stdout.
A merge-check.sh that passed on an agent's machine failed on the build seat: a newer gofmt, siblings at
a feature branch, another user. `mesh-controller check-here` runs builder.Check with the ask the
controller would make, from facts that now name the toolchains and the refs cloned beside; a failed
script is said by what failed. (novox/hq issues 282, 283)
While the mesh-delivery seat has a holder on record, a merge that moves no
core module opens its walk and asks nothing until mesh-delivery or a person
says go; nothing of it is registered before its turn, so no other send
carries it. The controller keeps the planner, the gate, sending and the
walk, and gains the verbs the owner asks with: delivery-plan, -order,
-check (a group composed as one future state), deliver, delivery-stop,
delivery-walks; every walk kept is said as plan-moved.
de-spiegel's and link2pay's manifests already fail the module check on main; without
comparing against the base branch every pull request touching them would fail the gate
for a fault none of them made. The gate's own rule: what was already so is said.
Every pull request the forge announces is mapped onto the mesh's module graph by the
merge handler's rule (issue 278): touching a module — or adding one — runs the gate
(mesh/merge-gate), its judge chosen by the graph (the controller judges itself, the
node-engine by its validator); a repository of the mesh that touches none runs only its
own merge-check.sh (mesh/repo-check), a warning when it has none. Nothing is left pending:
a repository outside the mesh touching nothing is told so as a pass.
The gate moves out of the per-repository scripts into the build seat, so a script is the
repository's own tests and declares its toolchain (go or typescript). The controller's
manifest names every verb of its seat again (ADR 0132), held by a test.
Every check the mesh had ran after a merge, on a machine: a manifest the node-engine refused
(236), an identity a real machine's name made too long (263). merge-gate raises the mesh as the
facts snapshot says it is and the mesh with the change, each in a throwaway store through the
controller's own records, composes every machine twice and validates it with the node-engine's
validator, and fails what the change breaks, naming the machine's roles and the module - plus a
manifest the judging controller cannot read, a consumer left out of its grant, a module removed
while a machine runs it, a new module the node-engine would refuse; it warns on a wide rebuild.
The forge's new head of a pull request becomes a check the controller asks of the build seat:
the head and, beside it, the controller the mesh runs, the catalogue, the host and the lab; a
throwaway store and bus of the versions the mesh runs; the repository's merge-check.sh in the
mesh's Go toolchain with no container runtime socket; then mesh-lab's replays. The verdict is
said as checked, an error never a pass, and nothing is recorded or registered.
mesh-adminandClaude Opus 5.5
jschoubben
jochen