secrets-engine's first live request rejected our allow: binding.
request_digest is computed over material they never sent, because we
enrich subject and resource from the registry before hashing. Answering
that meant reading the enrichment path, which had a worse defect in it.
Enrichment was additive-if-absent — addAttribute wrote a registry value
only where the request had no value for that key. So where a caller
supplied a key, the caller's value won and the registry's never applied.
Every registry ceiling and allowlist was advisory. Verified against the
shipped ops-warden package, each one added key on an otherwise-denied
request:
max_ttl_hours: 99 registry says 8 -> allowed a 12h certificate
allowed_principals registry allowlist -> disallowed_principal bypassed
allowed_subjects registry allowlist -> unknown_subject bypassed
The third is the one to read twice: a subject the registry does not know
authorized itself by naming itself in the allowlist it was being checked
against.
Not remotely reachable today — the PEP builds the CheckRequest,
ops-warden sends no resource.attributes, and enforce admits one identity.
It is a defence-in-depth failure: any path that lets attacker-influenced
data into a CheckRequest field became a full policy bypass rather than a
bounded input problem. Callers sending resource.attributes is not
hypothetical; secrets-engine does it on every request.
Registry facts now win, and diagnostics.registry_overrode names every
displaced key, because a registry that silently discards a contradicting
claim hides that a caller asserted authority it did not have.
subject.type is carved out, and the reason is a finding of its own.
Making the registry win there denied every secrets-engine allow: the
registry's type is CARING vocabulary (Human, Agent, Automation, Service)
and the request's is the protected system's actor vocabulary (service,
adm, agt, atm). Two fields sharing a name; substituting one for the other
is translation rather than identity, which GH-DEC-2026-008 ruled against.
Note what surfaced it — the registry's type had been dead data since the
field existed, because the caller's value always won.
Also publishes binding.submitted_request_digest, over the request exactly
as sent. request_digest was published as the consumer replay test and
cannot be one. Nothing is lost hashing the pre-enrichment form:
enrichment is a function of the request and the snapshot, and
registry_snapshot_digest already pins the snapshot.
Existing pins do not move. All three replay fixtures' request_digest and
approval_binding_digest values are byte-identical — those requests
contradict no registry fact. A field to add, not a value to correct.
Regression tests verified failing against the old behaviour before being
kept. FLEX-DEC-2026-012; FLEX-WP-0025 carries the residual, that a policy
still cannot tell a fact from an assertion.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aQMM1dPXaPiXVn6DwwtLd
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 715613@bnt-lap001
Assistant-Session: fabd95c1-4c9e-4080-8849-8707ae025f80
secrets-engine.catalog-lane.lifecycle v1 contained no reference to
input.tenant — not in well_formed, not in the denial ladder, not in a
test. A rotate on lane:glas-primary under tenant:coulomb returned allow
against the deployed package (decision:066e629bbf0c0924).
Found while answering glas-harness's tenant-alignment request, which had
asked for wrong-tenant denial evidence. There was none to return.
Three covers failed the same way: every one of the 29 fixtures carried
tenant:platform, so the suite could not report on the field; T02's own
gate named "wrong-tenant deny" and was recorded done unmet; and the
engine hashes tenant into request_digest but never compares it. Four
other published packages carry the branch — this one was the outlier.
v2 adds wrong_tenant above wrong_system, three Rego tests and three
fixtures (28/28, 32/32). The absent-tenant test caught a second defect
in the first draft: a bare input.tenant != comparison is undefined on a
missing key, so the branch dropped and the ladder reported the wrong
rung. request_tenant := object.get(input, "tenant", "") fixes it.
v2 supersedes rather than amends v1 because the defect failed open: a
consumer pinned to _VERSION=v1 would keep receiving allows with no
signal the rule beneath the version string had changed. The earlier
dual-control correction stayed at v1 because it denied everything.
Replay envelopes regenerated at v2; both request_digest values are
byte-identical, so secrets-engine's digest join needs no re-pinning.
The sweep this prompted found tenant-engine unscoped on tenant as well —
deployed, and verified allowing tenant:coulomb. Not the same fix: its
request tenant names the target rather than the caller, so a constant
would break it. Recorded and carried by FLEX-WP-0022 rather than patched
unilaterally.
FLEX-WP-0021 closes at T05; the pin still serves v1 until a redeploy.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aQMM1dPXaPiXVn6DwwtLd
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 715613@bnt-lap001
Assistant-Session: fabd95c1-4c9e-4080-8849-8707ae025f80
secrets-engine confirmed T03, and re-verifying against the regenerated
destroy fixture found something neither repository can fix alone: a
pdp_digest recorded at issue time can never equal the request_digest of a
request that carries the claim in its context, because the claim is part
of the context that is hashed. Embedding the claim changes the very
digest the claim would need to name.
Not fixture staleness. It holds for every dual-control request whose
claim travels in context -- the shape GH-DEC-2026-008 had just ruled
mandatory. Left unresolved that ruling was unimplementable for exactly
the case it was written for, and destroy would have been permanently
un-allowable in production, failing closed forever on a check that could
never pass.
flex-auth owns the canonical request digest, so the fix is ours.
binding.approval_binding_digest is the same material with
context.approval removed, emitted only when a claim was carried. An
approval issued against a claim-free Check records that Check's
request_digest; the claim-bearing request reproduces it here.
DELIBERATELY ADDITIVE, and the reason matters. The tempting fix is to
drop context.approval from request_digest entirely. That is wrong:
request_digest is the replay identity, and two requests differing only in
which approval was presented must not share one, because their decisions
differ -- one allows, the other denies dual_control_required. Collapsing
them would let an allow obtained with a valid claim be replayed against a
request carrying none. So request_digest still covers the claim and still
moves; approval_binding_digest deliberately does not, and is documented
as not a replay identity. The tests assert the two functions DISAGREE on
a claim-bearing request, which is approval-engine's formulation of how to
defend a distinction that looks like duplication.
The fixture now demonstrates the property rather than asserting it: its
claim's pdp_digest equals the envelope's approval_binding_digest with
pdp_path true, and changing the claim's contents moved request_digest
while leaving approval_binding_digest untouched. Two files a consumer can
diff.
Also picked up approval-engine's new required binding.pdp_path via the
cross-repo schema test added yesterday -- which is the test doing exactly
what it was built for, one day later.
T03 is done. secrets-engine's own digest-material defect, which our two
real envelopes caught, is recorded in the workplan.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JTbVXpEiXA7mNJVpDnEPcB
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 412054@bnt-lap001
Assistant-Session: 3968fae1-8d59-4209-9bd6-c22594b8ab19
approval-engine suggested a conformance check after the examples-contradict-
prose class hit a third repository -- theirs. Fifteen lines, they said, and
it would have caught our caring fixture and our provenance omission. Worth
stealing, so we stole it.
internal/schemaguard validates published examples against published schemas.
It implements only the JSON Schema subset these schemas use, and the property
that makes it trustworthy is that an unrecognised keyword FAILS rather than
skips: a validator that silently approves what it does not understand invites
reliance it cannot support. It found three things on first run.
ONE, AND THE LARGEST: check_request.schema.json pointed subject.type at
CARING's subject_type enum (Human, Service, ...), and no consumer sends that
vocabulary. user-engine sends human, tenant-engine and secrets-engine send
service, and ops-warden sends adm/agt/atm -- an actor-type vocabulary CARING
does not model at all. Our published schema declared three live integrations
non-conformant. A rule that outlaws shipped correct behaviour is the rule
that is wrong, so the $ref is replaced with an opaque non-empty string and a
description saying why. CARING's enum remains correct where it belongs: the
registry's subject_manifest.yaml, where Service is right.
TWO: policy_package_note, which this session added to the caring example's
decision provenance, is undeclared under additionalProperties:false. Our own
annotation broke the conformance it was annotating. Moved to the envelope's
outer provenance.
THREE: the secrets-engine fixtures carried partial approval-claims, missing
binding, freshness and validity. A partial claim in a fixture is how a
consumer learns the wrong shape -- the same mechanism that produced the
destroy defect. They are now complete and valid against approval-engine's
schema, including the now-required binding.pdp_digest, and a test validates
them against that schema when the sibling repo is present.
The destroy replay fixture is regenerated accordingly and the replay README's
pinned digests updated, since a stale digest table is the same defect wearing
a different hat.
Closed the binding-mapping open item. approval-engine declined to publish a
vocabulary mapping and their reasoning is better than the request: a PIP
asserting secrets.kv.destroy MEANS destroy would author semantics over two
vocabularies it owns neither of, and a wrong mapping silently accepts a claim
approved for a different action. pdp_digest is the mapping precisely because
it does not translate. It is now always present and nullable, so the destroy
gate is pdp_digest non-null and equal, enforced at the PEP.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JTbVXpEiXA7mNJVpDnEPcB
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 412054@bnt-lap001
Assistant-Session: 3968fae1-8d59-4209-9bd6-c22594b8ab19
Two real DecisionEnvelopes in examples/secrets-engine/replay/ for
secrets-engine to verify its digest join unchanged: a plain allow
(rotate, empty context) and the dual-control allow (destroy with a valid
approval-claim).
Both are included deliberately. input_claim_digests.context appears only
when the request carries a non-empty context, so a consumer asserting the
field is always present would pass on destroy and fail on rotate. One
fixture would have hidden that.
request_digest, policy_package_digest, registry_snapshot_digest and the
context claim digest are verified identical across two runs and are the
fields to pin. id, decision_time and the lifetime bounds move with the
clock; the README says so rather than leaving a consumer to discover it
by flake. lifetime.ttl is 15m from the package allow_ttl.
Emitted from flex-auth/local in standalone mode, not from a cluster pin.
T03 is progress, not done -- it closes when secrets-engine confirms their
validator accepts the records unchanged.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JTbVXpEiXA7mNJVpDnEPcB
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 412054@bnt-lap001
Assistant-Session: 3968fae1-8d59-4209-9bd6-c22594b8ab19
