POST /registrations passed its Documents list straight to Submit, which calls
DocumentStore.Link on every digital documentId in it — and linking a document
blocks its owner from ever deleting it (DeleteOwned returns 409 Linked). That
path had no ForeignIds ownership check, so any authenticated citizen could
post another citizen's document id and permanently block them from deleting
their own diploma scan. POST /applications/{id}/submit, the endpoint actually
in use, has had that guard since it was written.
Deleted rather than guarded: the endpoint is dead. No frontend caller, and
the whole registratie flow goes through /applications/{id}/submit.
RegistratieRequest went with it, and so did SubmissionRules.RejectRegistratie
— reachable only from here, and contradicted by the live path, which treats a
handmatig diploma as "does not auto-approve" rather than a 422 rejection. Its
own message said as much while being returned as a rejection. That last part
is a judgement call beyond the ticket's wording; reverting the two
SubmissionRules hunks restores it in isolation.
Coverage moved rather than vanished: the problem+json shape assertion is now
on /change-requests (the other endpoint on the same Submit helper), and the
linked-delete 409 test goes through the real submit path.
swagger.json, the generated client and the behaviour spec regenerated.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
ZgwHttpClient interpolated the full request uri and up to 500 characters of
the response body into its failure message. That message is persisted as
Aanvraag.ZgwError in SQLite and written to the log, and both halves can carry
a BSN: ZGW filters travel as query parameters (the citizen-scoped zaken list
filters on rol__betrokkeneIdentificatie__natuurlijkPersoon__inpBsn), and
OpenZaak echoes the offending request in its error bodies, so a rejected POST
/rollen comes back holding the owner BSN it was sent.
All three interpolation sites now use Redact(url) — the path without its
query — and the body snippet is replaced by the reason phrase. Status plus
path still routes a failure to the right endpoint; the lost detail already
has a deliberate home in ZGW_DEBUG_HTTP=1 (ZgwDiagnosticHandler), which is
opt-in, dev-only and not persisted.
The new test fails the one call in the fixture whose url carries a query
string and asserts the persisted ZgwError has neither the body snippet nor a
"?", while keeping the path and the 503. Verified red without the fix.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
DocumentStore wrote one audit row per upload and per user delete carrying the
acting citizen's raw BSN as AuditEntry.Actor, persisted to SQLite — on a
store whose own doc comment says it holds metadata only, never file content
"or other PII". Same shape as RB-02, in a second store.
Masked at the two citizen call sites rather than inside Audit, because the
third actor is the literal "admin" and MaskTail("admin", 3) is "**min";
masking centrally would mean guessing which actors are BSNs and which are
role names. Audit's doc comment now states that actors arrive redacted.
StoredDocument.Owner is untouched: it is the authorization key that
DeleteOwned, ForeignIds and RB-01's content check all compare against, so the
BSN stays where it is load-bearing and leaves the trail where it was only
decoration. No endpoint exposes AuditLog, so no response shape changes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Mappers.ToAdminSummaryDto set Owner to the raw BSN. Both consumers are
cross-owner lists read by someone who is not the subject — GET /admin/cases
and GET /werkvoorraad — while GET /beoordeling/{id}, the detail view of the
same data, already masked it. The detail screen showed ******782 and the list
one click earlier showed the whole thing.
Masked in the mapper rather than at each endpoint, so a third cross-owner
list cannot be added that forgets to.
MaskTail moves out of Program.cs into Domain/People/Pii.cs: it now has
callers in Contracts, Program.cs and (once RB-04 lands) Data, and a second
hand-rolled copy is how one of them drifts into leaking. Documented as
idempotent, which is what lets /beoordeling/{id} keep its own call —
IZaakSource has a second implementation whose Owner is mapped from the
OpenZaak zaak identificatie, so that endpoint should not depend on which
source answered.
No frontend change: all three consumers display the value, and the parse
boundaries only require a non-empty string.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Program.cs built the BIG-nummer reveal's audit resource ref as
"brief/" + ctx.Zorgverlener().Bsn. AuditAuthz persists that to the
AuthzAudit.Resource column in SQLite and /admin/audit renders it, so a BSN
reached durable storage and a UI on the one trail four documents describe as
data-minimised and PII-free — on the endpoint whose own comment promises the
audit carries no PII.
The ref is now "brief". Nothing is lost: BriefStore keys one brief per owner,
so the id named what the row's acting principal already implies.
The existing guard, The_audit_schema_carries_no_pii, asserts on column names,
so a BSN inside a column called Resource could never fail it. Added
No_audit_row_carries_a_subjects_bsn, which drives a denied reveal as a
non-default subject and scans every string field of every row for that BSN
and for DemoOwner — asserting on the two BSNs actually in play rather than a
\d{9} shape, since a hex correlation id can hold nine digits by chance.
Verified it goes red when only the Program.cs line is reverted.
AuditEntry.Actor on document audit rows holds a raw BSN too; that is a
different store and stays with RB-04.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
GET /uploads/{documentId}/content took only (string documentId) — no
HttpContext, so no authorization was possible. It streams diploma and
identity scans, protected by GUID unguessability alone, while DELETE on the
same resource has always been owner-scoped. GET /uploads/status had the same
shape and confirmed whether any client-chosen localId exists, plus its
documentId.
Both now take HttpContext. Content is readable by the owning
ZorgverlenerCaller or a caller passing Authz.CanBeoordelen — matched on the
caller kind rather than branched on a boolean, because ctx.Zorgverlener()
throws for a MedewerkerCaller and the behandelportal's beoordeling screen is
a legitimate reader. Status is scoped to ctx.Zorgverlener().Bsn via a new
owner parameter on DocumentStore.ByLocalIds (one call site).
404, not 403, on both: a foreign document id must not be distinguishable
from one that never existed, and a foreign localId reads back as "unknown".
Residual, recorded in the implementation note: both callers reach the URL as
a plain browser navigation (<a href> / previewUrl), which carries no identity
header and no interceptor, so StubIdentityProvider resolves it to the seeded
citizen. That is BIO-002 and belongs to RB-09; the links keep working today
only because one citizen owns every document in the POC.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Completes the pipeline's analysis phase. Agent 07 (BIO2/ISO 27002:2022,
control set stated as an assumption since none was supplied) produced 20
findings — 12 "defect now", 8 "production gate" — and agent 08 consolidated
all 47 findings across 00/02/04/06/07 into 33 tickets, 5 ADR-fixes and a
release checklist.
Two findings are live defects rather than refactoring candidates, both
verified directly:
- RB-01/BIO-004: GET /uploads/{documentId}/content takes only (string
documentId) — no HttpContext, so no authorization is possible. It streams
diploma and identity scans, protected by GUID unguessability alone, while
DELETE on the same resource is owner-scoped.
- RB-02/BIO-008: Program.cs:674 concatenates the caller's BSN into the authz
audit Resource column, which is persisted to SQLite and rendered by the
admin audit page. Four doc comments claim that store holds no PII; the test
cited as enforcing it asserts on column names, so a BSN inside a column
called Resource is invisible to it.
07 also answered the handoff from 06: in a production behandelportal build no
X-Medewerker is sent, so StubIdentityProvider returns the seeded citizen. It
fails closed on backoffice capabilities but open on citizen-scoped ones,
including CanRevealBigNummer. Root cause is IIdentityProvider.Resolve
returning a non-nullable CallerIdentity — the interface cannot express "no
identity", so any provider must invent one.
08's gate was relaxed from all-seven to the four agents that ran; _status.md
records why 01/03/05 were skipped, and the backlog carries a "Coverage"
note naming what those skips leave unowned. It caught two errors in the
orchestrator's handoff: CQ-002 is not fixed (ApplicationsStore.cancel and
AdminCasesStore.delete still swallow errors -> RB-20), and CQ-004 shipped
with half its compliance criterion unmet (PUT /admin/flags/{key} writes no
audit row -> RB-07, which blocks signing ADR-C-009).
Both agents preserved a "verified clean — do not fix" list, so a later pass
does not re-spend effort on the controls that already hold.
Consolidation halted for human approval per its spec. No source file changed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Runs the multi-agent refactoring-backlog pipeline in docs/project/
refactor-backlog-setup/ up to and including three of the seven Phase 1
agents.
00-baseline.md establishes the metrics every later agent must cite, using
only tooling already in the repo (vitest lcov, coverlet cobertura, ESLint's
core `complexity` rule at threshold 0 for a full distribution, depcruise
--metrics). Duplication and C# complexity had no tooling, so
tools/baseline-scan.mjs adds a deterministic ~200-line text scan rather
than a new dependency; the approximations are labelled as such.
Headline: FE 75.1% line coverage but only over the 98 of 220 source files a
spec loads; BE 97.6% line / 79.6% branch; 0 layering violations; 7.1%
duplication; 25 of 2085 TS functions over CC 10.
Then 02-testability, 04-cqrs-light and 06-adr-conformance (27 findings).
01/03/05 were skipped deliberately — the baseline shows little for them to
find; 07 (BIO2) and 08 (consolidation) are still open.
Each agent corrected a baseline observation of mine, and in every case the
error was in something derived rather than measured:
- BL-007 counted ~13 adapter "mutations" from the `runSubmit` helper name;
5 of those call sites are reads. It also missed 3 real mutations that
reach the raw ApiClient and never return a Result.
- BL-002 diagnosed the 100%-duplicated auth folders as ADR-0002's
divergence prediction failing. It never had a chance to fail: §3's
`Principal` union was never built.
- BL-004 named libs/shared/domain and libs/beheer/contracts as coverage
gaps; both are pure type declarations where 0% is unimprovable.
All three corrections are recorded inline in 00-baseline.md §10, so agent
08 does not inherit the bad numbers.
.prettierignore excludes the agent prompt directories — reflowing their
markdown would edit the prompt text itself.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
ehoandClaude Opus 5