SCOPE.md declared itself reviewed against b989de4, 63 commits behind, with its
limits kept current by appending to bullets. That works until the header asserts
a review date that no longer describes what was reviewed -- the drift this file
exists to prevent, in the document whose only job is to be accurate.
Every claim was rechecked against source rather than carried forward: routes from
the mux registrations, subcommands from the dispatch, flags from the flag sets,
claim behaviour from token.go and userinfo.go. Three were false. Caller commands
omitted verify-client, the container was said to package only keycape (untrue
since KEY-WP-0026), and the issuer API row omitted /readyz. Capabilities that had
reached SCOPE only piecemeal are now in the capability table: the
authorization-code bindings, upstream token verification, readiness and drain, the
human tenant rules, authentication-time assurance, the shared verifier.
The substantive addition is a distinction SCOPE did not draw: source behaviour
and deployed behaviour are different claims with different evidence -- the test
suite and harness runs for one, committed receipts in docs/evidence/ for the
other. What is proven live is narrower and now stated exactly, including that the
issuer pin's pre-pin state did not match, so that risk was real rather than
hypothetical, and that the deployed image predates this revision so several
described behaviours are running nowhere yet.
That distinction is the fix for a repeated error: four times in two days a claim
about current state was written from a workplan paragraph or a partial view
instead of the authoritative artifact. A SCOPE that does not separate "the code
does this" from "this is running" invites the same mistake from the next reader.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NV9oijZukGyGbRQGGKnK4P
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 713576@bnt-lap001
Assistant-Session: 384c511d-9bce-4cb8-a676-2aef6c0c8df6
G10 waits on custody and platform owners and cannot close from here. What was
doable: verify the handoffs actually went out, replace a remembered live state
with an observed one, and find out whether main is safe to deploy. The last
question found a defect in this repository's own recent work.
Handoffs verified independently rather than trusted: all seven messages are in
the hub with receipt ids. This gap was reopened once for claimed-but-unsent
delivery, so the claim deserved the same scrutiny.
Live state read from the cluster read-only: image main-153258b, only the Qonto
secret materialized so the approval clients remain unprovisioned, four registered
clients, no tenantEngine block. That also corrects an earlier claim of mine --
the deployed config sets userOU explicitly, so the KEY-WP-0023 default fix was
never a production issue.
The precondition: KEY-WP-0019 discovers the expected issuer from
authelia.tokenBaseURL, and the deployed Authelia derives its advertised issuer
from the request Host, advertising the in-cluster address to KeyCape and the
browser-facing one to browsers. Verification fails closed, so a mismatch breaks
every human login and looks like a broken login rather than a misconfiguration.
Which value the token carries needs a real login against production to settle and
was not determined here.
Two mitigations: docs/operations.md documents pinning authelia.issuer and
jwksUrl, with the curl that reveals what the provider advertises for a given
Host; and the authentication failure event now carries a specific reason, so
id_token_issuer_mismatch is distinguishable from a signature failure or an
unreachable key set. The browser still learns nothing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NV9oijZukGyGbRQGGKnK4P
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 713576@bnt-lap001
Assistant-Session: 384c511d-9bce-4cb8-a676-2aef6c0c8df6
Closes gap G09: five loosely related defects.
lldap-export took the service account password on argv, where ps exposes it to
any local user and shell history and process accounting capture it. It now
prefers KEYCAPE_LLDAP_BIND_PW or --bind-pw-file; --bind-pw still works but warns,
deprecated rather than removed because existing runbooks use it and breaking them
silently would be worse than one more cycle of exposure. Conflicting sources are
rejected instead of silently ranked, since an operator otherwise cannot tell
which bind was attempted. Both migration scripts pass the password by
environment now.
The canonical export, generated LDIF and Keycloak realm were written 0644. None
carries credential material, but the snapshot is every username, display name,
email and group membership in the estate, and it tends to land in /tmp. All three
are 0600.
The image packaged keycape alone, so the validator and migration binaries needed
a Go toolchain on the host -- which defeats shipping an image for the cutover
work they exist to support. All five ship; the issuer stays the entrypoint.
Verified by building the image and running each binary inside it.
The publish workflow named 92.205.130.254:32166 while the cluster runs
forgejo.coulomb.social/coulomb/key-cape. It now defaults to the recorded name and
stays overridable by a repository variable. This repository cannot verify that
the runner resolves that hostname or that the registry credentials are valid for
it; if the next publish fails, set the REGISTRY variable back to the address.
docker-compose.dev.yml mounts a private key and Authelia material that are
correctly absent from the checkout. scripts/bootstrap-dev.sh generates them
locally under a restrictive umask rather than chmodding afterwards, so the key is
never briefly world-readable. Everything it writes is git-ignored.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NV9oijZukGyGbRQGGKnK4P
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 713576@bnt-lap001
Assistant-Session: 384c511d-9bce-4cb8-a676-2aef6c0c8df6
Closes gap G08. /healthz returned a constant without probing anything, the server
called ListenAndServe with no signal handling, and the operational limits of
in-memory state, startup-loaded keys and local-only logout lived in code comments
rather than anywhere an operator would look.
/readyz probes LLDAP, Authelia and privacyIDEA; /healthz stays liveness and
probes nothing. Keeping them distinct matters: wiring liveness to dependency
health means an orchestrator restarts KeyCape when a dependency blinks, and a
restart also discards every in-flight login, so the reaction is worse than the
condition it reacts to.
LLDAP is probed with a bind rather than a dial, since a rotated or revoked
service password leaves the port open and every lookup failing -- exactly what
readiness should catch and exactly what a dial would miss. The response names the
failing check but never the reason: the endpoint is unauthenticated and upstream
error text carries hostnames and sometimes credentials-in-URLs. Results are
cached for 2s so an unauthenticated endpoint cannot be used to drive unbounded
upstream traffic, and probes run concurrently under a 3s bound so a hung
dependency makes the endpoint answer rather than hang with it.
SIGTERM and SIGINT now drain in-flight requests for 15s, under the 30s read/write
timeouts so a stuck request cannot outlive the window before SIGKILL.
docs/operations.md states the single-replica topology and why, and three limits
easy to get wrong: the constant key-1 key ID makes same-kid rotation a trap for
consumers caching JWKS, removing a client does not revoke its issued tokens, and
/logout is local only. No throughput figures are given, since nothing here
benchmarks KeyCape. Shared storage and refresh tokens stay excluded, as G08
allows.
Verified in the running executable: 503 naming all three checks failed while
/healthz returned 200, the LLDAP check flipping to ok once started, and 40/40
requests succeeding across a SIGTERM.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NV9oijZukGyGbRQGGKnK4P
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 713576@bnt-lap001
Assistant-Session: 384c511d-9bce-4cb8-a676-2aef6c0c8df6
Closes gap G07. The tenant-engine client and TokenHandler.TenantEngine were
implemented and tested, but main.go supplied no client and exposed no
configuration, so the stock executable left the field nil and the capability
existed at library level only.
A tenantEngine block with baseURL and an optional timeout now wires it. An empty
baseURL leaves the stock server's behaviour exactly as it was, so enabling the
claim is a deliberate act. Validation treats a configured source as one that must
work: http/https with a host, and a timeout in (0, 10s] since it sits on the
synchronous token-issuance path. A timeout set without a baseURL is rejected
rather than ignored -- it means someone expected the claim to be on.
Verified in the built executable, which is what G07 asks for, rather than at the
wiring: against a stub source a real token carries tenant_roles; with no block
the claim is absent; with the source configured but down, issuance succeeds
without it, confirming the documented fail-open path end to end.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NV9oijZukGyGbRQGGKnK4P
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 713576@bnt-lap001
Assistant-Session: 384c511d-9bce-4cb8-a676-2aef6c0c8df6
Finishes the unproven half of gap G04. Running the harness against real LLDAP,
OpenLDAP and Keycloak found four defects that the full unit suite passed over.
Every LLDAP user search pointed at a branch that does not exist. Config.userOU()
defaulted to ou=users while LLDAP stores users under ou=people, and nothing set
UserOU. LookupUser, ListUsers and ValidatePassword all derive from it, so all
three silently found nothing against a stock LLDAP -- human login included, not
only the export. Exposed by an export returning zero users while still emitting a
membership referencing uid=admin, a snapshot the repo's own validator rejects.
raw_attributes_well_formed, added one workplan earlier, rejected the LLDAP
adapter's own _validation_warning annotation, so the exporter's output failed its
own validation. Tooling annotations are exempt now, and a test proves the
exemption does not weaken the rule.
Migrated group memberships were dangling: resolveMemberDN passed a source DN
through unchanged while entries were written to the target branch, so groups
named entries the migrated directory does not contain. And empty groups could
not load at all, since groupOfNames makes member a MUST -- they reference a
placeholder entry the LDIF creates, an organizationalRole rather than a person,
emitted only when some group needs it.
The scenario-c compose file could not start: bitnami/openldap:2.6 does not
exist, though it passed docker compose config. Pinned to the image the scenario
was proved against.
Proof: LLDAP -> export -> validate -> LDIF -> ldapadd into OpenLDAP 1.5.0, every
entry added and every member resolving; a realm from the same export imports into
Keycloak and serves discovery. KeyCape passes 5/5 conformance checks, a migrated
Keycloak 4/5.
Relying-party behaviour and MFA against a migrated realm remain unexercised, and
credential/MFA migration is not supplied at all, so no harness can establish it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NV9oijZukGyGbRQGGKnK4P
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 713576@bnt-lap001
Assistant-Session: 384c511d-9bce-4cb8-a676-2aef6c0c8df6
Closes the runnable half of gap G04. The Scenario B and C scripts could not
execute: absent compose files, binaries sought at src/bin where the Makefile
builds to bin/, --base-dn passed to a generator whose flag is --basedn, and a
hardcoded workstation Go path invoked from outside the module.
Repairing the shell alone would have proved nothing. Both scripts set
KEYCAPE_TEST_ISSUER while the profile suite built its own httptest server and
never read it, so they passed identically whether or not a provider was running.
A harness that cannot fail for the reason it exists is worse than a missing one.
src/tests/conformance targets the issuer named by KEYCAPE_TEST_ISSUER over HTTP:
discovery, the profile authorization surface, published keys parsed under the
runtime's own rules, excluded grants, and -- with credentials -- a real token
exchange verified against those keys. It skips when the variable is unset, so
make test is unchanged.
Run against Keycloak 26.0 rather than asserted to work. Discovery, authorization
surface and key checks passed, and a client_credentials exchange produced a token
that verified against Keycloak's published JWKS through internal/jose.
It also failed, correctly: stock Keycloak advertises the excluded implicit and
password grants, and in Keycloak those are server capabilities rather than
per-client toggles, so no emitted realm removes them. A migrated Keycloak has a
wider grant surface than KeyCape, which substantiates with evidence what SCOPE
previously asserted without it. Scenario B legitimately reports failure today.
Directory migration, credential and MFA preservation and relying-party behaviour
remain unexercised, and Scenario C has never been run end to end, so G04 does not
fully close.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NV9oijZukGyGbRQGGKnK4P
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 713576@bnt-lap001
Assistant-Session: 384c511d-9bce-4cb8-a676-2aef6c0c8df6
Closes gap G06. checkNoUnknownAttributes was named for a rule it did not
implement: it rejected blank keys, which is not an allow-list, so a snapshot
carrying malformed or shadowing attribute names passed a check whose name said
otherwise. A rule that passes for the wrong reason is worse than an absent one,
because the report is read as evidence.
raw_attributes_well_formed requires each key to be a valid LDAP attribute
descriptor, forbids shadowing an attribute the canonical model owns (uid, cn,
mail) in any casing, and rejects keys differing only by case, which LDAP treats
as one attribute. spec/ldap-schema.yaml carries the same wording.
Two corrections to the gap's description rather than implementations of it. An
allow-list is not derivable: ldapAttributes is defined as what the canonical
model does not cover, so the schema cannot enumerate what may appear there. And
the reference constraint already existed -- checkValidGroupMemberships only
checks emptiness, but the semantic rule checkReferencedUsersExist resolves every
member against the user set.
Neutering the rule fails four of the five new tests.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NV9oijZukGyGbRQGGKnK4P
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 713576@bnt-lap001
Assistant-Session: 384c511d-9bce-4cb8-a676-2aef6c0c8df6
keycape-to-keycloak called Transform, which passes no clients, so it wrote a
realm with an empty clients array and nothing said the service-identity contract
had not been migrated. Where clients were supplied, mapClient hardcoded
standardFlowEnabled — silently giving every client_credentials registration the
browser flow — and dropped audience, service subject, tenant, roles, lifetime,
MFA policy, secret reference and handoff URLs. Realm roles and client scopes were
emitted as empty containers.
The defect was not the missing mapping but that a dropped field and an
inapplicable one looked identical in the output.
Add -clients, reading registrations through a new config.Registrations() that
converts without resolving secrets, so migration tooling cannot load material it
has no business holding. Derive flows from the declared grants. Carry the profile
claims as protocol mappers, since Keycloak has no native concept for them, and
lifetime, handoff URLs and the secret reference as attributes — the reference,
never a value. Derive realm roles and client scopes from what is present.
Report what cannot be carried, in UnpreservedReport, kept deliberately separate
from ValidationReport: consistency with the snapshot and completeness of the
migration are different questions and one list cannot answer both. It names the
unmigrated secret, the unenforceable MFA policy, passwords and factor enrolment,
and subject continuity. An incomplete transform emits partial telemetry.
Closes the semantic-preservation half of G03; proof against a live provider is
G04 and stays open.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012WAsfsfQmDu4vcBhiMcmQp
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 867844@bnt-lap001
Assistant-Session: 3d45905e-0016-4b49-b828-231406881f7b
KEY-WP-0019-T05 rested on "existing tests pass unchanged", which shows the move
onto internal/jose preserved behaviour but not that the behaviour is checked.
Disabling the RSA comparison in jose.Verify fails both callers' suites, so the
shared verifier is load-bearing on each path.
Correct the comment on the caller-side cases added with the move. It claimed the
existing tamper case fails on the signature segment's shape before any key is
used; it does not — appending eight characters leaves a decodable segment, so
that case does reach and does check the signature. The two new cases are still
worth their place for what a byte-level tamper cannot reach: a structurally valid
token signed by an unpublished key under a published kid tests that key selection
is bound to the key set, and an undersized modulus in the published set tests
that ParseJWKS strictness denies rather than falling through to the claims.
Record both in the workplan and in G01's status.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012WAsfsfQmDu4vcBhiMcmQp
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 867844@bnt-lap001
Assistant-Session: 3d45905e-0016-4b49-b828-231406881f7b
Closes the remaining half of gap G01. The adapter decoded upstream ID-token
claims without verifying anything, justified in a comment by a server-to-server
TLS boundary that nothing enforced.
Operator decision: verify the token rather than police the transport. The hop is
to be HTTPS as defence in depth, but KeyCape does not monitor, check or gate on
that -- a transport check helps only when it is configured correctly, which is
the assumption it was meant to remove. Verification holds regardless of how the
token arrived, so no HTTPS validation or opt-in flag is added.
HandleCallback now verifies the RS256 signature against Authelia's published
keys, the issuer Authelia advertises, KeyCape's own client ID in the audience,
and a sane validity window, before any claim is trusted. It fails closed: an
unreachable or unparseable key set denies the login. The advertised jwks_uri
path is rebased onto the server-side token base URL so split-horizon deployments
resolve, with config overrides where that inference is wrong, and an unknown key
id triggers one refresh so provider rotation needs no restart.
The reusable half lives in internal/jose rather than being copied from
authclient's verifier, since duplicated verification is how two copies drift and
one misses a fix. Migrating authclient onto it is tracked as KEY-WP-0019-T05,
kept separate so it does not destabilise a tested path in this change.
Thirteen rejection cases plus algorithm and rotation coverage; with the
unverified parse restored all fifteen fail, so they test the fix rather than
merely passing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NV9oijZukGyGbRQGGKnK4P
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 713576@bnt-lap001
Assistant-Session: 384c511d-9bce-4cb8-a676-2aef6c0c8df6
The exporter discovered groups by walking each user's memberships, so a group
nobody belongs to never reached the snapshot, and a failed lookup was skipped by
a `continue` under a comment claiming it was recorded in the incompatibility
report. The run then emitted `result: "success"`.
Add an optional `domain.GroupLister` capability and implement `ListGroups` on
the LLDAP adapter as a direct group-subtree search, kept off `UserRepository`
because the OIDC layer never enumerates the directory. Record `groupEnumeration`
on every result and a `Complete()` predicate over it; abort rather than write a
smaller snapshot when the enumeration fails; report a failed per-user lookup on
the fallback path; emit `partial` telemetry and name the mode from the CLI.
Reading the adapter to write this surfaced a defect the assessment had not
listed: `LookupGroups` never populated `Group.Members`, and the exporter built
every membership from that field, so against a real directory the `memberships`
block was always empty while the fixture-backed tests passed. Memberships on the
fallback path now come from the user/group pair actually observed.
Sort users, groups and memberships so an unchanged directory exports
identically. Closes G05 of the scope/intent assessment.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012WAsfsfQmDu4vcBhiMcmQp
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 867844@bnt-lap001
Assistant-Session: 3d45905e-0016-4b49-b828-231406881f7b
Closes gap G02 of the scope assessment for the client-registration and discovery
surface. spec/canonical-model.yaml and domain/model.go both claimed to be the
source of truth and disagreed: the spec restricted grants to authorization_code,
required redirect URIs of every client, and omitted the audience, service
subject, tenant, role, MFA and handoff fields the runtime reads.
The durable part is the link, not the edit. A two-way conformance test compares
the spec against the Go model by reflection and fails when a runtime field has
no spec entry or a spec entry is not read by the runtime, the latter unless
marked runtime: false. It found drift beyond the assessment's list on its first
run -- User.tenant was undeclared -- which is the argument for the check over a
one-time reconciliation.
Discovery now advertises the core profile claims that appear on every token and
derives scopes_supported from the registered clients rather than a fixed list.
The Go model is stated as the runtime authority and the YAML as the reviewed
contract, in both files. This covers client registration and discovery, not
schema enforcement in general, which remains G06.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NV9oijZukGyGbRQGGKnK4P
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 713576@bnt-lap001
Assistant-Session: 384c511d-9bce-4cb8-a676-2aef6c0c8df6
Closes the local protocol surface of gap G01 from the scope assessment
(KEY-WP-0016). The browser grant validated PKCE, client id and scopes but left
four bindings unenforced, and UserInfo verified less than the caller CLI does.
Authorization-code path: bind the exchange to the redirect URI the code was
issued for, refuse clients whose registration does not permit the grant, and
authenticate confidential clients with a digest-based constant-time comparison
over the same credential sources as the service grant. An empty grantTypes stays
an implicit authorization-code client, matching config validation.
Code consumption: SessionStore.Consume reads and deletes under one lock. The
previous Get/Delete pair spanned JWT signing, and the added test reproduces the
race against that version -- 9 of 16 concurrent exchanges succeeded, and a
failed exchange left the code replayable.
UserInfo: check the JOSE header algorithm before trusting the signature, require
the configured issuer, and require an access token rather than accepting an ID
token of the right shape. Purpose is decided on the scope claim so the issued
token contract, which consumers pin exactly, does not change.
SCOPE.md and the assessment record which bindings are now enforced and that the
Authelia upstream-trust assumption remains open, so G01 is not fully closed and
no profile-conformance claim is made.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NV9oijZukGyGbRQGGKnK4P
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 713576@bnt-lap001
Assistant-Session: 384c511d-9bce-4cb8-a676-2aef6c0c8df6
