Conflict in mesh-knowledge main.go, the propose tool's description: #609 (79c2fc7f)
names ideas and the research-active-at-most limit, A6 names a research effort's
partial new version; both sentences kept, A6's rule count (31).
Round 2 of issue 1970's review: a research effort has no clauses, so holds never serves one; Served on an
idea is tested in TestAnIdeaIsListedAndNeverServedAsTruth.
Claude-Session: https://claude.ai/code/session_01JG73KG7AfSdh6FP8VrBqCn
The test of the five ends holds every key of a body to the seat definition's holder events, so the router and the contract cannot drift.
Issue 1968.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
knowledge_overview answers the page's model as data (reads); knowledge_overview_reread asks one source's verb again by hand, no other, and
records the hand act (acts). A source marked from its verb is asked again alone from 10 seconds, doubling to 5 minutes, its first answer clearing
the mark; one marked from its subscription asks nothing while silent and the verb once on its next event. The module's health carries the page's
last rebuild and says it is degraded while a source is unanswered, staying healthy.
Issue 1968.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
The module's bundle (the one process with the bus) serves the page on the port its manifest declares under listens, from: machine,
and requires a route that reaches it. No form, no script, no verb behind it; any method but GET is 405. A source that does not
answer is said on the page, never shown as an empty list.
Issue 1968.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
The model is built once at start from the issue tracker's milestones and list, the delivery groups and the operator channel's asks, and
kept in memory from their events (group, transition, merged; offered, closed; opened, moved, milestone-set, milestone-moved), asking no verb
again; the records come from the ledger and the seven days are counted when the answer is given. The manifest invokes groups and asks,
consumes those events, has no reads and no decided family.
Issue 1968.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
One integration test: the module's states are JetStream buckets, its events are published under their message ids to a stream
that drops a repeat; it moves its records, proposes, is reviewed and accepts a record, loses an event between an operation and its
send, restarts, sends the lost event exactly once, builds its index and checks the database within 40 seconds, and is healthy. The
test found that a restart probed the ledger number by number and stopped at a gap of 50, finding 3 of 14 records: the store now
keeps a key held.<kind>.<number> in the state findings for each record it creates and reads them at start (the old probing, with a
gap of 1000, only for a ledger written before the keys).
Issue 1966.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
Unhealthy, naming the cause, for a ledger or the verdicts' stream that does not answer, a head behind its history, an index behind
the ledger and a stream at its cap; the idle drafts, the parts waiting on the judge and the records the form-only check refused
are counted and leave it healthy. A part waiting more than 30 minutes is asked of the operator once, cleared when it is judged.
Issue 1966.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
A record imported with file-named sections gets plain titles, and its links to files become references of the one form, in one
version written after the move by the tool knowledge_form_only: marked form, listing the names changed, accepted without a
judgement because a check proves each section's text with its links replaced by their labels is the same and each new link names
the record the old one named. The old names stay as aliases get takes; the parts' reviews carry over; a record waiting for its
review is skipped.
Issue 1966.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
knowledge_head, _operations, _section, _draft, _parts, _asked, _fill, _rebuild_index, _review_again, _integrity,
_move and _move_check, beside the health, highest and marks tools; the manifest lists them and a test compares them with the code.
Issue 1966.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
A movableStore serves the database until the switch key says the ledger. The move copies every version of every record as an
import operation (the old author kept apart, marked imported; the ledger's author the module's account), refuses writes for its
length, checks record by record, and sets the switch only on a clean check. From then on the ledger is served and the database is
written beside it, behind or not (database.sequence, database.behind); the check runs at every start before the drop, and the
search index is built then.
Issue 1966.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
ask and holds on the ledger answer from a full-text index of the heads and the current versions' text and clauses; it is
built at start (index not built is the answer meanwhile), a new version replaces its record's documents, a reserved,
withdrawn or retracted record has none. A new pure Go dependency, bleve.
Issue 1966.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
An event of a change carries the record, the version it follows and the event as its message id; the last version sent per
record is a key of the state sent; at start the events of the operations after it are sent again under the same ids, so a
stop between an operation and its event loses none and a second start sends none.
Issue 1966.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
One judgement per part with no tools, at most two at once; a section is read with its frame (the clauses and records it
names and the digest of its Context), so a usual revision judges only what changed. The latest approval of each part is
kept in the head and carried to the next version; a part that could not be judged keeps the others' verdicts and is the
only one asked again; reviewed is said once per version and each verdict as part-reviewed on the stream reviews. A part
in flight is a key of the state asked. A version is accepted when every part approves; findings name their part.
Issue 1966.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
A new rule, section-form (31 now), refuses from the day the records move a section whose title is a file name, a number
prefix, a slash or over 200 bytes, a title twice in a record, and a text that holds a local path, a file and line, or a
link to a file. Only a section that is new or changed is judged, so a record imported with file-named sections keeps
them until its form-only version; write refuses at once.
Issue 1966.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
A draft is a key of the state drafts (author, fields, sections as title, digest and the sequence of their write-section);
write puts the text into the record's history at once, submit makes the draft a version through the rules propose is
judged by and writes no section again. Only the author writes; take-over hands a draft, a proposal or a reserved number
to the caller with the reason in the next version's message. A new version of an effort starts from the served sections
and keeps its title, date and deciders (the rows of the commit folded in, now run over the ledger too). A tool call is
named to the ledger so the runtime stamps the caller it proved.
Issue 1966.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
Closing a long research effort (its status and became links) meant retyping
its whole text through propose. A research effort's new version that gives no
sections now keeps the served version's, and its title, date and deciders when
none are given; the version so composed is judged and reviewed as any other.
A decision, or an effort with no served version, still needs its text.
Issue 1966.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
ledgerStore implements Store on the tool runner's ledger calls: a record is the instance decision/0290 or research/081,
a Save is the operation named by what changed, a section's text is written once in its write-section and named by
digest and sequence in the head, a version is read back from the history, the whole-set findings and the told marks
are keys of the state findings. Every test that runs over each store now also runs over the ledger, played with the
manifest's own types.
The bundle's pipe to the tool runner needs the SDK's raw request (request to A1); until then it answers loudly.
Issue 1966.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
The manifest keeps postgres-database until the drop and asks, in the typed form, the ledger records (a type for each
kind, the operations of the move), the states drafts, asked, findings and sent, and the stream reviews (kept for good,
refusing when full, granted by the trusted value readers). It consumes claude-code's judged and says the eleven new
events. The mesh-delivery indexer's judge accepts it (checked against main).
Issue 1966.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
mesh/delivery-group group feat/sdk1 rejected: composed together: the checker's verdict at the members' heads (6b0f5a2e9397): integration: meshbuildgo/cmd/mesh-…
mesh/check failed: integration: meshbuildgo/cmd/mesh-build-go… | 68e7a2ddbfb9 on ee75d4327ce5 | judged by the candidate's mesh-build-go at 773f299aed81
mesh/delivery-group group feat/1126-removal rejected: composed together: the checker's verdict at the members' heads (a523066f1791): integration: meshbuildgo/c…
mesh/check failed: integration: meshbuildgo/cmd/mesh-build-go… | 7f8701637f9d on 32c4c52261c4 | judged by the candidate's mesh-build-go at b82f7f676986
A test that slept a fixed time and then asserted what a goroutine, process or
watcher did failed under the checker's load. Each now waits for its positive
evidence (a started baseline, a call counted, a state reached) within 30 s, and
the concurrency-limit tests meet at a barrier instead of overlapping by sleep.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
mesh/delivery-group group feat/judge-once rejected: composed together: the checker's verdict at the members' heads (5c42ba1fb5f3): integration: artifactstoreto…
mesh/check failed: integration: artifactstoretools/cmd/artifa… | 3e08301b27ea on 4b16c193f0fb | judged by the candidate's mesh-build-go at 8dd51540f73a
mesh/check being carried to 5d50b952d4cb on 65fad0ae1c31 | from a68b39f7b512
A record proposed before a retraction and accepted after it marked the
retracted clause changed or the retracted record superseded, serving it again;
and returning a drafted record numbered its clauses anew, bringing retracted
ones back in force or making a waiting Approve mark another clause. accept now
leaves what is retracted alone, Return refuses a record with a retraction, and
retract waits until a record's drafting is done.
The operator prefers removing a wrong decision to stacking a correcting
record on it, and withdraw took only reserved or proposed records. retract
asks the operator through the router like a record that needs the operator,
and changes nothing until the Approve; a retracted record or clause no longer
holds in ask, holds, get, links and progress, history keeps every version, and
the whole-set rules still read a retracted record's links so a retraction alone
raises no finding. withdraw with retraction true takes a waiting one back.
mesh/check checking b5475cfa5c3d on 395a37dddab9 | novox/mesh-catalog#571 in group fix/held-module-and-assigner-contract
mesh-healers slept 2 s before its first subscribe to the keeper's conditions; its whole test delivered
condition-raised once the lease was read, often before that, and the event was dropped: "never: the would-act
written" in about two runs of three on main, failing the group check. Eight modules had the same shape.
Quick fix: sleep after a failed subscribe in mesh-healers alone.
Thorough fix (chosen, the kind is in eight modules): every loop that asks the runtime for events or a watch
waits for stdio.Served() and subscribes at once, sleeping only after a refusal (mesh-healers, mesh-config,
mesh-mail, mesh-issues, mesh-provisioner, mesh-watcher, mesh-delivery-gate); mesh-knowledge, whose pinned
mesh-sdk has no Served, subscribes first and sleeps after a refusal. The healers' whole test delivers the
event only once the bundle has subscribed it, as the runtime's consumer holds it until then.
47 packages carried the marker and were never kept: about 2000 tests per level ran on every check, whatever it
changed. The marker now lists, as paths or globs from the repository's root (? before a path whose absence is read),
what the package's tests read outside it; the digest hashes those files with the package's own non-Go files. A marker
that lists nothing, names an unclean path, or a line that matches nothing keeps today's never-kept digest.
The 46 markers list their reads; checks/deps and checks/trunk lose theirs (their tests read only temporary
directories). Digest form 6.
Quick fix of issue 1936, chosen by the operator 2026-10-26 21:56; the thorough fix (the Go test log's files kept with
each verdict) stays on the issue.
mesh/check checking 76e4ec4dc680 on 26908a89506f | novox/mesh-catalog#506 in group feat/1781-milestone-records
mesh-issues keeps each design's milestone in the state milestones (history 5), says milestone-set and
milestone-moved, and serves milestones, milestone-set and milestone-move. An issue's milestone is its own link
of kind milestone, so at most one milestone per issue is checked inside the issue's own write; kind commit
links a fixing commit. The triage links no milestone, pull request or commit. import with milestones=N seeds
a design's table once. mesh-knowledge's progress reads issue-tracker.milestones and its table parse is gone.
Squashed from four commits so the branch history does not carry a compiled bundle that was committed by
mistake (6 MB); the path is now in .gitignore.
mesh/check pass 87fe2c45cfed on 26908a89506f | carried from 27a7757b970e, hq 3a5c4c7c4d9c | unit 2091 run, 2327 kept, 6 skipped | integration 2112 run…
A resolved issue whose delivery is held or failed is stuck, one merged and not delivered
is moving; a row the tracker did not answer, an answer that does not read and linked
issues cut by the list's limit are said as warnings instead of passing silently.
A lead's five-minute report and 'is design 47 nearing its end' were joined by hand from
the design file, the tracker and the deliveries. One verb now works the state out
(ADR 0424 §1), minimal for every five minutes, detailed for the report's tree.
Milestones are read from the design's table until the tracker's milestone records
(issue 1781) are delivered.
mesh/merge-gate head 684e8801c420 merged into trunk 3395af96bc34 — pass: builds mesh-knowledge → novox; no bus step; every machine composes with the change…
mesh/repo-check head 684e8801c420 merged into trunk 3395af96bc34 — pass: its merge-check.sh passed, each part in its toolchain: merge-check.sh (go), merge-…
WithdrawRecord only ended the waiting ask locally; the router kept it open, so
the operator still saw Approve/Decline for a withdrawn record and an answer to
it was dropped silently. It now also publishes a cancel on the router's
subject once the withdraw is saved, as the controller's asker does; a cancel
the router cannot be told of is logged and the withdraw still stands. The
module's existing "uses": ["operator-channel"] already grants this; no manifest
change needed, since the ask and cancel subjects are both under that seat.
A dropped proposal or a stale reservation stayed proposed forever: there was
no way to say the operator decided against it. knowledge-base.withdraw takes
a reserved or proposed record to withdrawn as a new version naming why, ends
a waiting operator's ask and clears a pending review in the same version,
and refuses an already accepted, superseded or withdrawn record. The number
stays spent either way.
The old Withdraw (the background job for reservations never proposed) is
renamed WithdrawStale to make room for the new, caller-driven WithdrawRecord.
mesh/merge-gate head bf3b3a5e362b merged into trunk f076cadecedc — pass: builds mesh-issues, mesh-knowledge → novox; no bus step; every machine composes wi…
mesh/repo-check head bf3b3a5e362b merged into trunk f076cadecedc — pass: its merge-check.sh passed, each part in its toolchain: merge-check.sh (go), merge-…
jschoubben
mesh-tools
jochen