A trunk build put only the SDK into a module's node_modules; any other
@novox/mesh-* package its package.json names was left out of the install
and only a group or a member's build brought it, so the forge module's
build failed with "Cannot find module" for its contracts package. Each
such package is now placed at the latest release the registry holds, as
the SDK is; a group's packed package or a member's build still takes
precedence.
Builds: design 47 H H-A-267
A done build whose history read failed was left out of the view for the whole process, so later passes judged from an
incomplete keep set, and every build cost one history read. A build's end is now taken from the build-ended entries of
the builder's log the pass already reads once, and only a build the log lacks is read from its history. While any done
key's end is unread, or the view is not yet taken in, no pass judges: it ends with the finding artifact-builds-unread,
written once and cleared by the next pass that judges. Every change heard while a watch's handover is gathered now goes
into that handover, so a build finished between the handover and its taking in is not lost by the swap (H-A-257).
Builds: design 47 H §H-A-255 (1), §H-A-257, §H-A-258, §H-A-260
A watch asked again after an error handed over the current values on top of the old view, so a key deleted while no
watch ran stayed in the keep set. Each handover is now gathered apart and replaces the view whole once it is in, and
the ends of keys it no longer holds are forgotten.
Builds: design 47 H §H-A-255 (1), §H-A-257
A page read costs the whole bucket per page, so the builder's builds are now a followed view: the watch's first view
at start, then its updates, kept in memory with each build's end read once from its history. A pass reads nothing of
the builds on the bus and judges nothing before the first view is in, so it never sweeps against an empty keep set.
Builds: design 47 H §H-A-250 (2), §H-A-255 (1)
A finished build or a want change started a sweep and an observation each, every one reading the builder's builds key
by key and asking the registry's list again, queued behind one lock: a third of the night's bus flood. Now an event
marks the one pending pass, a pass running when marked runs once more, the builds are read in pages of the state with
each build's end from the page, and a list the registry seats cannot answer ends the pass with artifact-list-unread,
written once and cleared by a pass that reads it.
Builds: design 47 H §H-A-250 (1), (2), (3)
The router held the bus module's destroy-level asks with the bare words
"no channel can carry it now", so the asker could not tell why. holdWhy
now names the destroy level and that no channel carries a destroy answer
yet. The bus module's tool says "offered, channel not yet named" for an
open record with no tickets, never "offered on :".
Builds: design 47 H H-A-237 (3), H-A-239 (2)
The bus module's drop and rebuild asks were at the destroy level, which no
channel carries yet, so the router held every one unshown until it expired
while the tool answered "the operator is asked". The ask is now at the
approve level for the caller's expires_minutes (5 to 1430, default 60,
refused outside), the keeper accepts an approve-level choice for these two
acts and bounds a request by its ask's own expiry, and the tool makes one
read of the router's record after the publish and says offered, held or
refused as it finds it; the request read shows that standing too.
Builds: design 47 H H-A-237 (1) (2), H-A-239 (1) (2)
A directory of a trees listing's files the commit does not hold was
thrown as a plain error and answered with the kind failed, so the
indexer took it for a forge that failed, never ran its one-directory
fallback (issue 1585), and published nothing of the repository.
commitTrees throws DirectoryNotInTree for it, and the seat's failure
answers it with the kind not-found, the error naming the directory.
Builds: design 47 H H-A-232 (2), H-A-234; issue 1585
Review round 1. keepOld took the old plan's objects whole, so a state's `<state>-objects` store
the grants part made beside a bucket the keeper made alone would have been compared, grown and,
on a recreate, deleted by the keeper: it now keeps only the old plan's objects the made record
already holds. CapSum skipped every kept entry as counted in the old sections' caps, which count
objects, logs and streams but never a state: it now skips only a kept objects, log or stream
entry, so a kept state's bucket stays in the sum.
At the delivery the plan digest changes once for every module with old sections (most modules,
since state counts): each plans again once, quietly, as its entries are compared with what the
bus holds and nothing is created or said where they agree.
Builds: design 47 H H-A-229, H-A-230, H-A-231; G §547
The keeper read only a manifest's new-form asks, so the entries it had made for the artifact
registry and the forge, now declared in their old sections, stood as orphans with live data: a
rebuild refused, a drop offered. The keeper now reads every manifest's old sections (state,
objects, logs, streams) by the SDK's naming per section, as the grants part composes them. An
entry it made under one of those bus names is no orphan, is compared and grown to the declared
cap and rebuilt on request; it creates none of them and says no created for one, and counts a
kept one's cap once, in the old sections' sum.
Builds: design 47 H H-A-229, H-A-230, H-A-231
A one-line comment in ~/.fehbg (the files archive) moves feh's fingerprint and
is readable on every machine that holds feh once it is delivered, with no hand act.
Design 47 H §5 runbook step 7: "Prove it. One single-module change (a comment in
a module with lint, unit and composition steps only), merged to main through a
merge call (merge-pull, design 47 H §734), followed from take-in.taken through
pipeline-ended to the node-engine's running and the gate's ended record, with no
hand act".
G §370: "The version reaches every artifact's build": the delivered files
archive carries the version the pipeline gave it.
Builds: design 47 H §5 step 7, G §370
A requirement of the checked repository at a real version was fetched from the Go seat while the
indexer's fingerprint takes the tree, so the build and its key disagreed; and "same repository" was
read from the owner name novox/ written into the code. The fetch now takes every requirement of its
own repository from the fetched tree before anything is downloaded, at whatever version go.mod
names, keeps those directories in the narrowed tree, and judges "same repository" by the fetch's
repository as the ask names it.
Builds: design 47 H S31, S3, L8 (H-pins-spec-and-gaps section 8 item 6); H-A-204, H-A-202, H-A-198
A package artifact's go.mod may require a package of its repository that
requires another of it at the zero version; the narrowing kept only the
first level, so the build of the second found no module of that path. The
entry's paths keep the manifest's form; the Go build module reads each kept
package's go.mod in turn, each directory once, and keeps what it requires.
Builds: design 47 H H-A-202, H-A-198
Given no pin, the module took the version of another package of the same
repository, or the latest, and built against versions never published. A
requirement of another repository is now taken at the pin the ask names for
its own path, else the build fails naming the path; a same-repository
requirement is still built from the fetched tree.
Builds: design 47 H H-A-192, H-A-198, G §370
disk-load's Go package diskreadings is imported by mesh-grants, whose build
waits for a publish no release made: an imported mesh package is a declared
artifact of its module, so every release publishes it. The gate imports the
assigner's asks, published in the assigner's series 0.3, so it names that
series in its packages.
Builds: design 47 H H-A-186, H-A-188, H-A-198, G §370
Review round 1: versionsAnswer's choice of digest per seat had no test.
TestTheVerbsSpeakTheSeatsContract now holds it, with a store whose blob
and manifest digests differ. OCI.Blob refuses a manifest that keeps no
single blob, naming its layers, and versions says that in the journal:
the contract's answer has no field for why.
Notes for H2: OCI.Held (oci.go) compares the tag's manifest digest with
a bundle's or an archive's blob digest, the same kind as this fault;
versions answers kind "" since the tool reads a kind its input does not
take.
Builds: design 47 G A-430a (G-22-lead-answers-2), design 51 *One contract*
The node-engine resolves a want's version to an archive's digest through
oci-artifact-registry's verb versions and fetches /blobs/<digest> by it.
The holder answered the tag's manifest digest, so shanks asked
claude-code/files' manifest as a blob and got 404 (claude-code 0.1.17).
On the artifact seat versions now answers the manifest's one blob, the
digest the reference and the event published carry; the image seat
keeps the manifest digest.
Builds: design 47 G A-430a (G-22-lead-answers-2), design 51 *One contract* (the seats' events carry the digest), the seat's contract Artifact and Listed.artifact
Review round 1. The node build module runs the seat generator's check at its unit level
(steprun.OutOfStep), so its narrowed workspace keeps what H-A-136 and H-A-137 name: every
generated-from directory of the module, and for a root module the seats/ directory and the
paths of every artifact naming generated-from. Narrowed bare, a step of one mesh-contracts
package named the generated files left out as out of step. The fetch and the steps verb now
narrow the same way, and the steps verb reads its narrowed tree as the levels do.
narrowTree and generatedFromDirs are copied from mesh-build-go without its Go part, not moved
into mesh-delivery-sdk's steprun: the move is a change to two more repositories and a group
delivery for some sixty lines, in a fix that unblocks the build steps now; the copy says so.
packageDir and narrowedAway share one path helper.
Builds: design 47 H H-A-136, H-A-137 (hq design/47-H1-answers-59), ADR 0499 clause 4
A step of one TypeScript package of a module with several (mesh-contracts' root:
ts-meshbroker beside ts-meshrecovery, pipeline main.3 of 2026-11-20) fetches a workspace
narrowed to that package's path. The fetch and the lint and test levels read every package
artifact of the manifest, and one whose path the narrowing left out fell back to the
module's directory, which holds no lock: every TypeScript step of mesh-contracts ended
"could not run: the package in . has no package-lock.json". Such a package is now left
out of the fetch's install and of the levels' targets; an unnarrowed tree is read as before.
Builds: design 47 H H-A-136 (a step's workspace holds the artifact's paths; nothing else is fetched), ADR 0499 clause 4
The licence manager says each binding change as binding.changed.<consumer>
(mesh.mod.claude-licence-manager.event.binding.changed.<consumer>), but both
manifests consumed claude-licence-manager.binding.changed, which the grants part
composes into the exact filter mesh.mod.claude-licence-manager.event.binding.changed:
the bus consumer never handed either module one change, and the vault kept only its
first view (2026-11-20: g14 refused a switched licence's token until the first view
was asked again by hand). The consumed pattern now names the further part, `*`.
Builds: design 47 G A-134; design 39 bus-and-switch §17, §19, §22
A login not handed is kept only against a lease that does not succeed it; a lease
of the same account that succeeds it replaces it, said as login superseded. A login
to another account is kept until the manager adopts it.
Builds: ADR 0206, ADR 0209, design 39 §6, H-A-138 (hq design/47-H1-answers-59 175ed320), H-A-139 (hq design/47-H1-answers-59 e359911e), H-A-140 (hq design/47-H1-answers-59 b75d1386)
The manager's refresh keeps the licence's refresh-token expiry, so a grant of the
same lineage with a later access expiry replaces the login; a grant with no
mark, an older lineage or no later access expiry is the older lease and is kept.
Builds: ADR 0206, ADR 0209, design 39 §3 §6, design 14, H-A-138 (hq design/47-H1-answers-59 175ed320), H-A-139 (hq design/47-H1-answers-59 e359911e)
One reason, login kept, on both paths, with a journal line; the not-handed
return skips the binding's generation on purpose.
Builds: ADR 0206, ADR 0209, design 39 §3 §6, design 14, H-A-138 (hq design/47-H1-answers-59 175ed320)
A login made on a node (while the tool runner was stopped, or before the
manager asked claude_code_grant for it) was overwritten by the next apply
of the binding's older grant: Apply took any refresh token in the file as
one already handed, and forced the write. The person's login was lost and
the session asked for a login again.
The node side now keeps the fingerprint of the login claude_code_grant last
sealed (state login-handed, never the token). Apply leaves a file whose
login is not that one untouched (applied false, reason login-not-adopted);
its report makes it the manager's candidate, and the successor the manager
hands back replaces it, as before without a refresh token.
Builds: ADR 0206, ADR 0209; design 39 §6 (the latest login wins, a login
moves its node); design 14 (a node never holds a refresh token).
With the guard gone nothing reads the machines' names, domains and
addresses: remove Machine, the name lookup and the ten-minute timer that
resolved them again, so the tools' table is read only when a followed
state changes. Also remove the now unused cut helper.
The operator asked for the shell guard to go (2026-11-20): the plugin no
longer carries a PreToolUse hook or guard.json, the binary no longer has
a guard verb, and the claude_code_guard tool, its record and the
refusal-ranked order of the instead-of table are gone. The SessionStart
and Stop hooks of the wait list stay, run by the same binary at the same
path, which is also the home-fs helper's. The rendered rule against ssh
to a mesh machine stays as the session's own and no longer claims a
guard enforces it; claude-agent no longer passes the guard's override.
A module at the repository's root naming generated-from "." (mesh-contracts)
kept nothing of it: the narrowing kept "." alone, which prunes everything,
so the seat generator's check of go-meshconfig still found every file out
of step. Such a module now keeps the seats/ directory the generator reads
and the paths of every artifact naming generated-from; nothing else of the
root.
Builds: design 47 H H-A-137
The first pipeline of the clean state (main.3) failed every build of a
mesh-delivery-contracts client: the fetch narrowed the tree to the
artifact's paths (ADR 0499 clause 4), which removed
modules/mesh-delivery-sdk, so its requirement at the zero version found
"no module of that path in the fetched tree". The same narrowing left the
seat generator's out-of-step check without its seat definitions.
The narrowed workspace now also holds every generated-from directory an
artifact of the module names and the directory of every package of the
same repository the artifact's go.mod requires at the zero version, found
by the repository's manifests; the root's manifest and lock files are not
kept when narrowing from the root. The fetch's own Go modules stay those
under the artifact's paths (A-447a), not the added directories.
Builds: design 47 H H-A-136; design 47 G-22 A-447a
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy