KONT-WP-0013 renumbered, archived RAILIANCE-WP migrated, unqualified ADHOC
accepted per CUST-WP-0066, OPS-WP-* newly visible and undecided. Forge-side
collisions fell 45 to 16.
Central still reports 35 duplicates because archived workplans are never
registered — the registrar scans only top-level workplans, closed archives being
frozen under ADR-007. The renames re-derived projection UUIDs, so the old rows
are now unclaimed by any file, which is exactly the condition ADR-012 decision
7's reset retires. The duplicates clear when that reset runs, not before.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Discovered before implementation: discard-and-rebuild cannot exist. Progress
events, tasks, decisions and review contracts reference workplans with ON DELETE
RESTRICT, and 1067 of 1075 workplans carry such a reference. Deleting a workplan
would delete the hub-native records attached to it, which ADR-010 decision 4
forbids and the schema refuses.
The database is enforcing this ADR's own boundary one layer down. A progress
event is a fact that happened; destroying it to tidy a derived projection loses
hub-native truth to fix a derived-state problem.
Reset now reconciles: create what is missing, update what differs, retire what no
longer derives — and retirement must be visible rather than silent.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The fleet-wide re-sync took unbacked workplans from 265 to 29. ADR-010's orphan
rules applied to the remainder: all 29 are closed, none is an ADR-001 violation,
and all are retained as provenance under the "closed hub-first records" rule.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Fleet-wide re-sync across 120 repositories: 1075 bindings sent, 1015 written,
workplans recording no backing file fell from 265 to 29. ADR-010 predicted this
as the "broken links" class and it was 89% of the apparent orphans.
The 29 remaining are all closed: 23 finished, 6 archived, none open.
Corrects T09's scope: the prefix migration left 4 orphaned rows, not 26. The
other 22 are still backed by the archived RAILIANCE-WP files that migration
deliberately did not rename. The overstatement came from counting rows by slug
prefix rather than by whether a file backs them.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
An attempt to purge stale task rows established that the hub has no hard-delete
for work records at all: DELETE /tasks/{id} is cancel_task, DELETE
/workstreams/{id} is 410 Gone, and no session.delete exists for either.
Re-derivation is not one removal option among several, it is the only one
permitted — a deliberate property, since a read model that cannot drop a record
except by re-deriving cannot be quietly edited into disagreeing with its source.
Scope is now 26 workplan rows plus 305 task rows across 44 workplans, of which
only 9 show open. All wait on ADR-012's per-repository reset.
No records were mutated: eight cancel calls hit STATE-WP-0080 tasks that were
already cancel.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Migrating active workplans off the retired family prefix left 26 hub rows with
no backing file. Accepted deliberately: the alternative was one repository
carrying two prefixes indefinitely.
Blocked on ADR-012 decision 7's per-repository reset — C-07 is not fixable and
hand-deleting hub rows is the direct central edit ADR-010 decision 6 forbids.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
ADR-012 accepted 2026-08-25.
ADR-010 — partially superseded, notes added inline:
decision 1: "a reading of the repositories" never said which copy, and the
answer was neither the forge nor a working copy but whichever checkout last
synced. The hub holds no repository files and never reads one.
decision 5: "central derives, does not accept pushes" was policy while the
workstation pushed everything.
decision 6: "preliminary" was named but never built.
ADR-003 — partially superseded:
decision 2: fingerprints composed from filesystem mtime are invalidated in
part. mtime is a property of one workstation, differs across clones, and
says nothing about content. Evidence: git_fingerprint for the-custodian
held the initial commit while last_state_synced_at was minutes old.
decision 5: the rebuild principle was correct but never exercised, and its
"without data loss" claim holds only while nothing exists solely in the hub
— which was false for 111 records on 2026-08-25.
ADR-001, ADR-005, ADR-007 and ADR-011 reviewed and unaffected; ADR-007 is
reinforced, since derived identifiers let an overlay and a forge-derived
projection compute the same identity.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Refines decision 7. Per-repository rebuild is not a convenience over a
fleet-wide operation: the repository is the unit, and the fleet-wide form is
defined as iteration over it.
It follows from the source — each forge repository has its own history and head,
so its projection is derivable and provable in isolation. It bounds blast radius,
which is what makes "routine" achievable: a rebuild that only runs fleet-wide is
one nobody runs casually and therefore nobody runs, which is how ADR-010's
never-executed reconstruction claim came to be believed.
It is also the only form that composes with the refusals. A repository holding
projection-only records can be held back while healthy ones rebuild; an
all-or-nothing reset is blocked permanently by a single bad repository. The
refusal is therefore evaluated per repository, and a fleet run skips and reports
rather than aborting.
The fleet-wide form must share the per-repository implementation, so the rare
dangerous path is exercised by the common safe one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Adds decision 7: the projection must be discardable and rebuildable from what
Forgejo holds, per repository and fleet-wide.
This is what makes the other decisions checkable. A read model that cannot be
rebuilt from its source is a database with a projection's reputation, and the
difference only surfaces when someone needs to rebuild it. ADR-010 already
asserts a cache may be reconstructed at any time; that claim has never been
executed.
Three properties keep it real: routine rather than emergency, scoped per
repository, and idempotent enough to verify against the forge — which derived
identifiers make possible.
Two refusals are explicit. Reset cannot restore the preliminary overlay, since
the forge does not hold it. And reset must refuse when records exist only in
the projection: on 2026-08-25 that was 111 work records, which a rebuild would
have erased while reporting success.
Also flags an unresolved boundary: hub-native records are not forge-derived and
must survive a rebuild of forge-derived state.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
ADR-001 and ADR-010 both leave unanswered which repository files the hub reads.
Verified 2026-08-25: neither the forge nor a working copy — the projection
derives from whichever checkout most recently ran the sync, which is written
down nowhere.
Proposes the forge as projection source, unpushed local work as a labelled
preliminary overlay rather than a second store, and commit provenance on every
record. Implements ADR-010 decision 6's unbuilt notion of "preliminary" and
sharpens decision 1 rather than reversing it.
Status: proposed — needs review before anything is built on it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Central 670 -> 1018 workplans; cache-only gap 426 -> 111.
T06 was scoped as an identifier-migration decision and that framing was wrong.
Two defects did most of the blocking: 50 repositories carried retired
gitea-remote URLs on central so fix-consistency could not match them, and the
registrar was blind to records that had an identifier but were absent from the
hub. Both were masked by a generic error whose real message sat in a child
process's stderr.
Also records the repo-wide re-key that severed eight working identifiers on
kaizen-agentic, and the rule now enforced in code to prevent it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
These workplans exist only in the retired local hub. Their random pre-ADR-007
identifiers are refused by C-06 as stale references, so they cannot be
registered. Deriving from the canonical record id takes no identity from
anything: central does not hold them and the old ids die with the cache.
Records central already holds were deliberately left untouched.
Refs CUST-WP-0068-T06
The file carried an identifier minted by the retired local hub while central
held a different record for the same file — ADR-010's same-filename,
different-UUID duplicate-registration class.
Central is authoritative (ADR-010 decision 1) and decision 6 says to mitigate
by changing the repository file rather than editing central, so the file adopts
central's id. The cache-side id dies with the cache.
Refs CUST-WP-0068
The registrar now binds workplan files in the same pass as minting, verified
end to end. The unbound remainder is explained: those identifiers are the
cache-only records this workplan exists to recover, so the count is a progress
indicator rather than a fault.
Remaining under T07: re-sync fleet-wide, re-measure the 278, then disposition
whatever is genuinely orphaned.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
backing_filename is written only by the bindings sync, which fix-consistency
runs before the registrar mints the UUID. Newly registered workplans therefore
record no backing file. Reproduced on CUST-WP-0068 itself.
278 of 800 workplans on central record no backing file, which is what ADR-010
predicted as the "broken links" class. The 23 cache-only orphans were counted
with the same field, so that number is an investigation starting point rather
than a work list.
T07 now repairs the ordering and re-measures before dispositioning anything.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Promotes the remainder of CUST-WP-0067-T10. 297 workplans and their tasks
still exist only in the retired local hub, which is why its database cannot be
discarded.
Eight tasks covering the registrar's self-defeating precondition, per-repo
unblocking, six non-conforming identifiers, clay-borg's 50 unparseable
workplans, the 176 with non-empty projections, the identifier ruling that
overlaps RMGR-WP-0005, 23 orphans, and final convergence.
Carries forward a correction: ten of sixteen records flagged as
non-conforming are ADHOC-YYYY-MM-DD, which is the documented convention. The
classifier's pattern was wrong, not the files.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Central 670 -> 799. All 403 cache-only workplans with a backing path have that
file on disk, confirming ADR-010's finding that nothing was lost.
Only 57 carry deterministic identifiers, so C-06 refuses most as stale
references. --bootstrap-empty-projection recovers them with identifiers
preserved, verified on one repository before rollout.
Classifies the 297 remaining and lists per-repository blockers from actual
error messages. Corrects this task's own earlier reading of
bootstrap_source_invalid as a single cause: it covers at least two.
Recommends promotion to its own workplan — the mechanical share is done; the
rest needs content fixes, an identifier decision overlapping RMGR-WP-0005, and
owner judgement across ~40 repositories.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
All 44 cache-only repository records resolved: 43 onboarded and classified,
agent-harness dropped by owner decision. Central holds 121 repositories.
Corrects an earlier claim in this workplan that the cache would hold nothing
unique once the repository records were resolved. That generalised a
repository-scoped result: the cache still holds ~400 workplans and ~1,950 tasks
that exist nowhere else, which is ADR-010's original divergence.
The cache database still must not be discarded.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Superseded by glas-harness and the rein concept, so there is nothing to
restore. Not onboarded to central; its cache-only record is left to disappear
when the cache database is discarded, rather than preserved as a hub-only
record with no backing file.
Successors verified present and classified on central before dropping.
Ten records remain outstanding.
Refs CUST-WP-0067-T05
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Release 20 carries the canon vocabulary; classification writes succeed against
central for the first time from outside a workstation. All 33 onboarded
repositories re-run and projected: 33 updated, no warnings. Central holds 111
repositories, 100 classified.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
178 AGENTS.md files moved off the 127.0.0.1:18000 reverse tunnel to the
in-cluster address; both machines report zero stale rows. 100 pushed from the
workstation, 74 picked up by pull on the node, 4 handled individually.
binky-control committed but not pushed — it carries 6 pre-existing unpushed
commits that are not this change's to publish. activity-core was already 39
behind with an unrelated dirty file; only AGENTS.md was taken from origin.
state-hub-railiance01 is now unblocked for removal.
Still open: the block has no single source, so this recurs on the next
topology change; and 68 non-AGENTS.md references need individual judgement
rather than substitution.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The remote row pointed at 127.0.0.1:18000, a reverse tunnel back to the
workstation. On railiance01 the State Hub runs in the cluster on that same
machine, so the request left the box and came back to reach a local service.
Refs CUST-WP-0067-T07
Three dated corrections, following the blockquote convention ADR-007 uses:
- Decision 2's merge/refresh dichotomy is not exhaustive. Repository records
took a third shape — cache-only records whose source exists and is reachable
but was never introduced to central. Refreshing would have destroyed them.
- The Context recorded the shared port but not why it was silent: an unbound
ssh -L binds every loopback family, and ExitOnForwardFailure fires only when
every bind fails.
- A new Outcome section states that the ADR's own remediation never landed and
the condition was still live seven weeks later.
Refs CUST-WP-0067-T06
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
rmgr repo-onboard is the missing managed_repos write path. 33 of 44 cache-only
repositories onboarded to central (78 -> 111); the other 11 carry written
dispositions rather than guessed classifications.
T04 uncovered why onboarding only worked from a workstation: the hub resolved
its classification allowed-values file from hardcoded developer checkouts, so
every classification write 500'd in the container. T09 delivers that file.
The local cache database must not be discarded yet — 11 records live only there.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
state-hub-mcp serves SSE on ClusterIP 10.43.110.80:8001, verified from the node
rather than through a tunnel, and reads central's data end-to-end (79 repos).
Remote port map and dev-hub registration repointed off the reverse tunnel.
state-hub-mcp-railiance01 is now safe to remove. state-hub-railiance01 is not:
~120 AGENTS.md files still depend on it until T07.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
