feat: adopt the owner-documented PDP access path, and prove it live
Adoption asked for by flex-auth (FLEX-WP-0021-T05) and glas-harness
(GLAS-WP-0015), plus the first real decision this engine has obtained from
the deployed pin -- which found a defect the fixtures could not.
ACCESS PATH. require_supported_pdp_address refuses in-cluster Service names
and any non-loopback host. This is no longer a unilateral call: the owner
path is documented as loopback kubectl port-forward over the authenticated
Kubernetes API, which is what authenticates the responder transitively
(FLEX-DEC-2026-010). A Service name from a workstation does not fail, it
resolves through the DNS search suffix to an unrelated public host, and
since decision records carry no signature, a responder knowing the
published package and version can return an allow that passes every check
we make. Fail-closed protects against a PDP that is absent, not one that
lies. The guard runs before the token is read, so a misdirected request
cannot leak it; a test pins that ordering.
LIVE PROOF. Minted a 10-minute TokenRequest token (audience flex-auth, SA
secrets-engine/secrets-engine, mode 0600 outside the worktree, shredded
after), forwarded to the named pod, and sent a real CheckRequest for
glas-claude-agent-dev-anthropic. Result: allow, catalog_lane_policy_matched,
served by v2 (sha256:bd11c5fe...) -- so the redeploy flex-auth flagged as
outstanding has landed and the pin no longer serves the tenant-blind v1.
Our tenant fix is confirmed against the real service: binding.tenant is
tenant:platform.
THE DEFECT IT FOUND. The evaluator enriches from its registry before
hashing -- subject gains attributes and tenant, resource gains tenant --
so binding.request_digest is over material we never sent and cannot
reproduce. validate_decision_envelope rejects every real allow.
Every replay test passes because _request_from() rebuilds the request out
of the binding, i.e. the enriched form: a self-consistent fake agreeing
with itself, which hid this through three rounds of digest work. Third
time a real artifact has beaten a fake in this integration.
NOT FIXED, DELIBERATELY. Rejecting a valid allow is wrong in the safe
direction. Which fields may be enriched is flex-auth's contract to publish;
inferring it means accepting a binding that differs from our proposal in a
way we decided was benign -- the fail-open shape GH-DEC-2026-008 rejected
for vocabularies and FLEX-DEC-2026-007 for digests. Raised with them.
Tenant question closed by operator decision 5ed3fb35: tenant:platform
exactly, and service_auth.TENANT stays tenant:coulomb because the two
identity layers are to remain distinct. Declining to author that mapping
was right -- the answer was neither reading offered.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E4tNMAYcSQmZWUE4wqP4ij
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 715726@bnt-lap001
Assistant-Session: 80a42b32-cba6-4b23-8be0-68819b1a6092
2026-09-07 00:22:12 +02:00
|
|
|
# Reaching the access-engine pin
|
|
|
|
|
|
|
|
|
|
The supported way for this engine to reach `flex-auth-secrets-engine`, and the
|
|
|
|
|
reason the address is enforced in code rather than left to configuration.
|
|
|
|
|
|
|
|
|
|
## Never the Service DNS name
|
|
|
|
|
|
|
|
|
|
```
|
|
|
|
|
SECRETS_ENGINE_PDP_URL=http://flex-auth-secrets-engine.flex-auth.svc.cluster.local:8080
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
is refused by `decision_check.require_supported_pdp_address`. From a workstation
|
|
|
|
|
that name does not fail — it **resolves, to the wrong host**:
|
|
|
|
|
|
|
|
|
|
```text
|
|
|
|
|
$ getent hosts flex-auth-secrets-engine.flex-auth.svc.cluster.local
|
|
|
|
|
80.158.43.29 ...svc.cluster.local.ad.binect.de
|
|
|
|
|
$ getent hosts this-service-does-not-exist.flex-auth.svc.cluster.local
|
|
|
|
|
80.158.43.29 (identical — proves suffix expansion, not a record)
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
`resolv.conf` carries `search ad.binect.de`, a wildcard zone, so every
|
|
|
|
|
`*.svc.cluster.local` name answers with one unrelated public address. Sending
|
|
|
|
|
there ships the CheckRequest body and our bearer token to a third party.
|
|
|
|
|
|
2026-10-16 10:48:40 +02:00
|
|
|
As of source `ff1a13d` (2026-10-16), this consumer requires the producer's
|
|
|
|
|
detached Ed25519 signature using independently admitted local keys and exact
|
|
|
|
|
signed evaluator/mode/policy pin. The complete CLI wheel-plus-helper bundle and
|
|
|
|
|
strict task handoff are implemented in source. Native signer custody, key/pin
|
|
|
|
|
admission, installed placement and actual consumer receipts remain unaccepted;
|
|
|
|
|
source enforcement does not establish those facts. See
|
|
|
|
|
[the scoped native preparation](proposals/native-cli-admission/README.md).
|
|
|
|
|
|
|
|
|
|
Historical 2026-10-12 finding: prior consumer source did not verify producer
|
|
|
|
|
signatures, so a forger could reproduce public request digests. The literal
|
|
|
|
|
loopback address restriction limited the destination and did not prove responder
|
|
|
|
|
identity. That source defect is repaired; FLEX-DEC-2026-010's historical warning that an absent
|
|
|
|
|
PDP and a lying responder require different controls remains relevant. The
|
|
|
|
|
independently admitted native trust and rollout obligations in SECRETS-WP-0014 T02/T03 remain open.
|
2026-10-12 08:27:37 +02:00
|
|
|
|
|
|
|
|
The Check client disables automatic redirects, environment/system proxies and
|
|
|
|
|
the process-global urllib opener. A 3xx response fails closed without contacting
|
|
|
|
|
the redirect destination. URLs must be HTTP(S) bare origins with a literal
|
|
|
|
|
loopback IP and optional nonzero port: no hostname (including `localhost`),
|
|
|
|
|
userinfo, path prefix, query, fragment or whitespace/control characters. Existing
|
|
|
|
|
`localhost` configuration must use `127.0.0.1` or `[::1]`. Invalid configuration
|
|
|
|
|
is rejected before reading the bearer file. Responses are limited to 512 KiB;
|
|
|
|
|
oversized, invalid UTF-8 and non-object JSON bodies fail closed. These rules also
|
|
|
|
|
apply when the approval flow invokes the PDP client.
|
|
|
|
|
|
|
|
|
|
Regression tests use disposable local HTTP servers and synthetic credentials;
|
|
|
|
|
they do not demonstrate identity or signature verification on the native pin.
|
feat: adopt the owner-documented PDP access path, and prove it live
Adoption asked for by flex-auth (FLEX-WP-0021-T05) and glas-harness
(GLAS-WP-0015), plus the first real decision this engine has obtained from
the deployed pin -- which found a defect the fixtures could not.
ACCESS PATH. require_supported_pdp_address refuses in-cluster Service names
and any non-loopback host. This is no longer a unilateral call: the owner
path is documented as loopback kubectl port-forward over the authenticated
Kubernetes API, which is what authenticates the responder transitively
(FLEX-DEC-2026-010). A Service name from a workstation does not fail, it
resolves through the DNS search suffix to an unrelated public host, and
since decision records carry no signature, a responder knowing the
published package and version can return an allow that passes every check
we make. Fail-closed protects against a PDP that is absent, not one that
lies. The guard runs before the token is read, so a misdirected request
cannot leak it; a test pins that ordering.
LIVE PROOF. Minted a 10-minute TokenRequest token (audience flex-auth, SA
secrets-engine/secrets-engine, mode 0600 outside the worktree, shredded
after), forwarded to the named pod, and sent a real CheckRequest for
glas-claude-agent-dev-anthropic. Result: allow, catalog_lane_policy_matched,
served by v2 (sha256:bd11c5fe...) -- so the redeploy flex-auth flagged as
outstanding has landed and the pin no longer serves the tenant-blind v1.
Our tenant fix is confirmed against the real service: binding.tenant is
tenant:platform.
THE DEFECT IT FOUND. The evaluator enriches from its registry before
hashing -- subject gains attributes and tenant, resource gains tenant --
so binding.request_digest is over material we never sent and cannot
reproduce. validate_decision_envelope rejects every real allow.
Every replay test passes because _request_from() rebuilds the request out
of the binding, i.e. the enriched form: a self-consistent fake agreeing
with itself, which hid this through three rounds of digest work. Third
time a real artifact has beaten a fake in this integration.
NOT FIXED, DELIBERATELY. Rejecting a valid allow is wrong in the safe
direction. Which fields may be enriched is flex-auth's contract to publish;
inferring it means accepting a binding that differs from our proposal in a
way we decided was benign -- the fail-open shape GH-DEC-2026-008 rejected
for vocabularies and FLEX-DEC-2026-007 for digests. Raised with them.
Tenant question closed by operator decision 5ed3fb35: tenant:platform
exactly, and service_auth.TENANT stays tenant:coulomb because the two
identity layers are to remain distinct. Declining to author that mapping
was right -- the answer was neither reading offered.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E4tNMAYcSQmZWUE4wqP4ij
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 715726@bnt-lap001
Assistant-Session: 80a42b32-cba6-4b23-8be0-68819b1a6092
2026-09-07 00:22:12 +02:00
|
|
|
|
2026-10-16 10:48:40 +02:00
|
|
|
## Historical transport procedure and current admission gate
|
feat: adopt the owner-documented PDP access path, and prove it live
Adoption asked for by flex-auth (FLEX-WP-0021-T05) and glas-harness
(GLAS-WP-0015), plus the first real decision this engine has obtained from
the deployed pin -- which found a defect the fixtures could not.
ACCESS PATH. require_supported_pdp_address refuses in-cluster Service names
and any non-loopback host. This is no longer a unilateral call: the owner
path is documented as loopback kubectl port-forward over the authenticated
Kubernetes API, which is what authenticates the responder transitively
(FLEX-DEC-2026-010). A Service name from a workstation does not fail, it
resolves through the DNS search suffix to an unrelated public host, and
since decision records carry no signature, a responder knowing the
published package and version can return an allow that passes every check
we make. Fail-closed protects against a PDP that is absent, not one that
lies. The guard runs before the token is read, so a misdirected request
cannot leak it; a test pins that ordering.
LIVE PROOF. Minted a 10-minute TokenRequest token (audience flex-auth, SA
secrets-engine/secrets-engine, mode 0600 outside the worktree, shredded
after), forwarded to the named pod, and sent a real CheckRequest for
glas-claude-agent-dev-anthropic. Result: allow, catalog_lane_policy_matched,
served by v2 (sha256:bd11c5fe...) -- so the redeploy flex-auth flagged as
outstanding has landed and the pin no longer serves the tenant-blind v1.
Our tenant fix is confirmed against the real service: binding.tenant is
tenant:platform.
THE DEFECT IT FOUND. The evaluator enriches from its registry before
hashing -- subject gains attributes and tenant, resource gains tenant --
so binding.request_digest is over material we never sent and cannot
reproduce. validate_decision_envelope rejects every real allow.
Every replay test passes because _request_from() rebuilds the request out
of the binding, i.e. the enriched form: a self-consistent fake agreeing
with itself, which hid this through three rounds of digest work. Third
time a real artifact has beaten a fake in this integration.
NOT FIXED, DELIBERATELY. Rejecting a valid allow is wrong in the safe
direction. Which fields may be enriched is flex-auth's contract to publish;
inferring it means accepting a binding that differs from our proposal in a
way we decided was benign -- the fail-open shape GH-DEC-2026-008 rejected
for vocabularies and FLEX-DEC-2026-007 for digests. Raised with them.
Tenant question closed by operator decision 5ed3fb35: tenant:platform
exactly, and service_auth.TENANT stays tenant:coulomb because the two
identity layers are to remain distinct. Declining to author that mapping
was right -- the answer was neither reading offered.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E4tNMAYcSQmZWUE4wqP4ij
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 715726@bnt-lap001
Assistant-Session: 80a42b32-cba6-4b23-8be0-68819b1a6092
2026-09-07 00:22:12 +02:00
|
|
|
|
2026-10-16 10:48:40 +02:00
|
|
|
The following transport example records the earlier supported access shape; it
|
|
|
|
|
is incomplete for current native use without independently admitted signing trust,
|
|
|
|
|
the reviewed CLI bundle and exact owner configuration. It grants no token mint,
|
|
|
|
|
installation or live action. Use literal loopback `kubectl port-forward` to the admitted named pod after
|
2026-10-12 08:27:37 +02:00
|
|
|
verifying the cluster identity and pod identity. That path rides the Kubernetes
|
|
|
|
|
API server's TLS with the operator's cluster credentials; the client cannot infer
|
|
|
|
|
that binding from its local URL. Note this
|
feat: adopt the owner-documented PDP access path, and prove it live
Adoption asked for by flex-auth (FLEX-WP-0021-T05) and glas-harness
(GLAS-WP-0015), plus the first real decision this engine has obtained from
the deployed pin -- which found a defect the fixtures could not.
ACCESS PATH. require_supported_pdp_address refuses in-cluster Service names
and any non-loopback host. This is no longer a unilateral call: the owner
path is documented as loopback kubectl port-forward over the authenticated
Kubernetes API, which is what authenticates the responder transitively
(FLEX-DEC-2026-010). A Service name from a workstation does not fail, it
resolves through the DNS search suffix to an unrelated public host, and
since decision records carry no signature, a responder knowing the
published package and version can return an allow that passes every check
we make. Fail-closed protects against a PDP that is absent, not one that
lies. The guard runs before the token is read, so a misdirected request
cannot leak it; a test pins that ordering.
LIVE PROOF. Minted a 10-minute TokenRequest token (audience flex-auth, SA
secrets-engine/secrets-engine, mode 0600 outside the worktree, shredded
after), forwarded to the named pod, and sent a real CheckRequest for
glas-claude-agent-dev-anthropic. Result: allow, catalog_lane_policy_matched,
served by v2 (sha256:bd11c5fe...) -- so the redeploy flex-auth flagged as
outstanding has landed and the pin no longer serves the tenant-blind v1.
Our tenant fix is confirmed against the real service: binding.tenant is
tenant:platform.
THE DEFECT IT FOUND. The evaluator enriches from its registry before
hashing -- subject gains attributes and tenant, resource gains tenant --
so binding.request_digest is over material we never sent and cannot
reproduce. validate_decision_envelope rejects every real allow.
Every replay test passes because _request_from() rebuilds the request out
of the binding, i.e. the enriched form: a self-consistent fake agreeing
with itself, which hid this through three rounds of digest work. Third
time a real artifact has beaten a fake in this integration.
NOT FIXED, DELIBERATELY. Rejecting a valid allow is wrong in the safe
direction. Which fields may be enriched is flex-auth's contract to publish;
inferring it means accepting a binding that differs from our proposal in a
way we decided was benign -- the fail-open shape GH-DEC-2026-008 rejected
for vocabularies and FLEX-DEC-2026-007 for digests. Raised with them.
Tenant question closed by operator decision 5ed3fb35: tenant:platform
exactly, and service_auth.TENANT stays tenant:coulomb because the two
identity layers are to remain distinct. Declining to author that mapping
was right -- the answer was neither reading offered.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E4tNMAYcSQmZWUE4wqP4ij
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 715726@bnt-lap001
Assistant-Session: 80a42b32-cba6-4b23-8be0-68819b1a6092
2026-09-07 00:22:12 +02:00
|
|
|
is the exact reverse of the caller direction, where port-forward bypasses the
|
|
|
|
|
NetworkPolicy (`FLEX-WP-0023`); the two properties point opposite ways and
|
|
|
|
|
summarising either as "the network protects it" gets one backwards.
|
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
# 1. Target the pod by name, not the Service.
|
|
|
|
|
POD=$(kubectl -n flex-auth get pods \
|
|
|
|
|
-l app.kubernetes.io/name=flex-auth-secrets-engine \
|
|
|
|
|
-o jsonpath='{.items[0].metadata.name}')
|
|
|
|
|
kubectl -n flex-auth port-forward "pod/$POD" 18080:8080 --address 127.0.0.1 &
|
|
|
|
|
|
|
|
|
|
# 2. Fresh bounded caller token per session. Audience MUST be flex-auth;
|
|
|
|
|
# caller auth is `enforce` as of Helm revision 3.
|
|
|
|
|
umask 077
|
|
|
|
|
kubectl -n secrets-engine create token secrets-engine \
|
|
|
|
|
--audience flex-auth --duration 10m > "$PDP_TOKEN"
|
|
|
|
|
chmod 600 "$PDP_TOKEN" # outside any Git worktree
|
|
|
|
|
|
|
|
|
|
# 3. Point the engine at the forward.
|
|
|
|
|
export SECRETS_ENGINE_PDP_URL=http://127.0.0.1:18080
|
|
|
|
|
export SECRETS_ENGINE_PDP_TOKEN_FILE="$PDP_TOKEN"
|
|
|
|
|
export SECRETS_ENGINE_AUTHORIZATION_POLICY_PACKAGE=secrets-engine.catalog-lane.lifecycle
|
|
|
|
|
export SECRETS_ENGINE_AUTHORIZATION_POLICY_VERSION=v2
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
Shred the token file and stop the forward when the session ends. The token is
|
|
|
|
|
ten minutes; it is a caller credential, not an authorization.
|
|
|
|
|
|
|
|
|
|
`ServiceAccount secrets-engine/secrets-engine` has `automountServiceAccountToken:
|
|
|
|
|
false` and no role bindings — it exists to be a caller identity and nothing else.
|
|
|
|
|
|
|
|
|
|
## Pin v2, never v1
|
|
|
|
|
|
|
|
|
|
v1 shipped and was deployed with no `input.tenant` rule at all; a `rotate` under
|
|
|
|
|
`tenant:coulomb` returned **allow**. v2 supersedes rather than amends it so the
|
2026-10-16 10:48:40 +02:00
|
|
|
change is visible in the version string. Current consumer configuration and signed
|
|
|
|
|
trust must explicitly pin the admitted v2 artifact; the validator does not impose
|
|
|
|
|
a universal hard-coded package/version default. A v1 configuration is not admitted
|
|
|
|
|
by this proposal. See `docs/tenant-alignment.md`.
|
feat: adopt the owner-documented PDP access path, and prove it live
Adoption asked for by flex-auth (FLEX-WP-0021-T05) and glas-harness
(GLAS-WP-0015), plus the first real decision this engine has obtained from
the deployed pin -- which found a defect the fixtures could not.
ACCESS PATH. require_supported_pdp_address refuses in-cluster Service names
and any non-loopback host. This is no longer a unilateral call: the owner
path is documented as loopback kubectl port-forward over the authenticated
Kubernetes API, which is what authenticates the responder transitively
(FLEX-DEC-2026-010). A Service name from a workstation does not fail, it
resolves through the DNS search suffix to an unrelated public host, and
since decision records carry no signature, a responder knowing the
published package and version can return an allow that passes every check
we make. Fail-closed protects against a PDP that is absent, not one that
lies. The guard runs before the token is read, so a misdirected request
cannot leak it; a test pins that ordering.
LIVE PROOF. Minted a 10-minute TokenRequest token (audience flex-auth, SA
secrets-engine/secrets-engine, mode 0600 outside the worktree, shredded
after), forwarded to the named pod, and sent a real CheckRequest for
glas-claude-agent-dev-anthropic. Result: allow, catalog_lane_policy_matched,
served by v2 (sha256:bd11c5fe...) -- so the redeploy flex-auth flagged as
outstanding has landed and the pin no longer serves the tenant-blind v1.
Our tenant fix is confirmed against the real service: binding.tenant is
tenant:platform.
THE DEFECT IT FOUND. The evaluator enriches from its registry before
hashing -- subject gains attributes and tenant, resource gains tenant --
so binding.request_digest is over material we never sent and cannot
reproduce. validate_decision_envelope rejects every real allow.
Every replay test passes because _request_from() rebuilds the request out
of the binding, i.e. the enriched form: a self-consistent fake agreeing
with itself, which hid this through three rounds of digest work. Third
time a real artifact has beaten a fake in this integration.
NOT FIXED, DELIBERATELY. Rejecting a valid allow is wrong in the safe
direction. Which fields may be enriched is flex-auth's contract to publish;
inferring it means accepting a binding that differs from our proposal in a
way we decided was benign -- the fail-open shape GH-DEC-2026-008 rejected
for vocabularies and FLEX-DEC-2026-007 for digests. Raised with them.
Tenant question closed by operator decision 5ed3fb35: tenant:platform
exactly, and service_auth.TENANT stays tenant:coulomb because the two
identity layers are to remain distinct. Declining to author that mapping
was right -- the answer was neither reading offered.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E4tNMAYcSQmZWUE4wqP4ij
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 715726@bnt-lap001
Assistant-Session: 80a42b32-cba6-4b23-8be0-68819b1a6092
2026-09-07 00:22:12 +02:00
|
|
|
|
fix: compare structured binding fields, not a digest we cannot reproduce
The live proof in 03c0569 showed validate_decision_envelope rejecting every
real allow. The evaluator normalizes before hashing -- the request tenant is
copied onto subject and resource, and a registry hit copies type, tenant and
selected attributes onto the refs -- so binding.request_digest covers
material we never sent. Byte-equality against our unenriched request was
unsatisfiable, not merely mismatched.
THE RULE WAS ALREADY PUBLISHED. flex-auth's canonical-request-digest.md
section "Normalization" states the enrichment and tells consumers what to do
instead: compare structured binding fields to the proposed action, treat
request_digest as the evaluator's statement of what it hashed, and recompute
independently over the tuple the binding carries. I raised this with them as
an unpublished gap and asked them to pick between three shapes; it was in
their contract already and the answer was the first of the three. Nothing
was blocked on them, and this follows the published rule rather than one I
inferred.
- _require_binding_corresponds: everything we proposed must survive
unchanged -- tenant, action, context, subject.id/type,
resource.id/type/system, and every attribute we sent.
- Enrichment may add only type, tenant, attributes. Any other added field is
refused, and an enriched tenant must be the request tenant, so a
cross-tenant binding cannot arrive wearing our request's clothes.
- request_digest is still verified, now against binding_tuple(binding) for
self-consistency rather than against material we never sent.
- The envelope's top-level subject/resource get the same rule; they are
enriched too.
Proved against the artifact: the real decision:0f9c98f14545c42d now
validates, and the unrefreshed envelope is refused on lifetime -- reaching
the lifetime check at all is the evidence the binding checks pass on a real
decision. Negatives cover a restated resource.attributes.stage, a foreign
subject.tenant, and an unexpected enrichment field.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E4tNMAYcSQmZWUE4wqP4ij
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 715726@bnt-lap001
Assistant-Session: 80a42b32-cba6-4b23-8be0-68819b1a6092
2026-09-07 09:04:57 +02:00
|
|
|
## The digest join, and why it is not a digest comparison
|
|
|
|
|
|
|
|
|
|
Proved live on 2026-09-07, fixed the same day.
|
|
|
|
|
|
|
|
|
|
The evaluator normalizes before hashing: the request tenant is copied onto
|
|
|
|
|
`subject` and `resource`, and a registry hit copies `type`, `tenant` and
|
|
|
|
|
selected `attributes` onto the refs. So `binding.request_digest` is over
|
|
|
|
|
material we did not send and cannot reproduce, and the validator's original
|
|
|
|
|
byte-equality check against our unenriched request rejected every real allow.
|
|
|
|
|
|
|
|
|
|
flex-auth's `canonical-request-digest.md` already published the consumer rule,
|
|
|
|
|
and the fix follows it rather than inventing one:
|
|
|
|
|
|
|
|
|
|
> A consumer that re-hashes the **original** unenriched request will not match a
|
|
|
|
|
> decision that turned on registry attributes. Compare structured `binding`
|
|
|
|
|
> fields to the proposed action, and treat `request_digest` as the evaluator's
|
|
|
|
|
> statement of what it hashed. To recompute independently, hash the same
|
|
|
|
|
> normalized tuple the binding carries.
|
|
|
|
|
|
2026-10-16 10:48:40 +02:00
|
|
|
The bullets in earlier revisions described the September structured-enrichment
|
|
|
|
|
check. Current source has since moved to the producer's submitted-request replay
|
|
|
|
|
contract: `validate_decision_envelope` validates the exact submitted digest,
|
|
|
|
|
opaque enriched digest shape, approval-free correspondence, policy pin and
|
|
|
|
|
lifetime on an authenticated envelope. It does not independently reconstruct the
|
|
|
|
|
full enriched tuple or native five-field claim in the generic CLI path.
|
|
|
|
|
|
|
|
|
|
The separate strict `task_execution_handoff` additionally recomputes the evaluated
|
|
|
|
|
and approval-free digests, verifies exact descriptor/context/native claim joins
|
|
|
|
|
(including the served native digest), and binds signed enforced caller provenance.
|
|
|
|
|
That source seam is not an `execute-ccr` CLI command and does not admit the proposed
|
|
|
|
|
service caller mapping or enable either task policy's runtime false gate.
|
fix: compare structured binding fields, not a digest we cannot reproduce
The live proof in 03c0569 showed validate_decision_envelope rejecting every
real allow. The evaluator normalizes before hashing -- the request tenant is
copied onto subject and resource, and a registry hit copies type, tenant and
selected attributes onto the refs -- so binding.request_digest covers
material we never sent. Byte-equality against our unenriched request was
unsatisfiable, not merely mismatched.
THE RULE WAS ALREADY PUBLISHED. flex-auth's canonical-request-digest.md
section "Normalization" states the enrichment and tells consumers what to do
instead: compare structured binding fields to the proposed action, treat
request_digest as the evaluator's statement of what it hashed, and recompute
independently over the tuple the binding carries. I raised this with them as
an unpublished gap and asked them to pick between three shapes; it was in
their contract already and the answer was the first of the three. Nothing
was blocked on them, and this follows the published rule rather than one I
inferred.
- _require_binding_corresponds: everything we proposed must survive
unchanged -- tenant, action, context, subject.id/type,
resource.id/type/system, and every attribute we sent.
- Enrichment may add only type, tenant, attributes. Any other added field is
refused, and an enriched tenant must be the request tenant, so a
cross-tenant binding cannot arrive wearing our request's clothes.
- request_digest is still verified, now against binding_tuple(binding) for
self-consistency rather than against material we never sent.
- The envelope's top-level subject/resource get the same rule; they are
enriched too.
Proved against the artifact: the real decision:0f9c98f14545c42d now
validates, and the unrefreshed envelope is refused on lifetime -- reaching
the lifetime check at all is the evidence the binding checks pass on a real
decision. Negatives cover a restated resource.attributes.stage, a foreign
subject.tenant, and an unexpected enrichment field.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E4tNMAYcSQmZWUE4wqP4ij
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 715726@bnt-lap001
Assistant-Session: 80a42b32-cba6-4b23-8be0-68819b1a6092
2026-09-07 09:04:57 +02:00
|
|
|
|
|
|
|
|
The lesson worth keeping: the replay fixtures could not catch this, because
|
|
|
|
|
`_request_from()` rebuilds the request from the binding, so the tests hashed the
|
|
|
|
|
evaluator's output and compared it to the evaluator's output. Only a real
|
|
|
|
|
request through this access path exposed it.
|
2026-10-16 07:51:44 +02:00
|
|
|
|
|
|
|
|
## Mandatory responder authentication (WP14 source preparation)
|
|
|
|
|
|
|
|
|
|
Every serving `authorize_action` path requires `SECRETS_ENGINE_PDP_TRUST_FILE`
|
|
|
|
|
and `SECRETS_ENGINE_PDP_VERIFIER`. These fail closed before approval/PDP credential
|
|
|
|
|
access when a PDP path is configured. Trust must be an absolute regular JSON file,
|
|
|
|
|
not group/world writable, admitted independently of the responder or its keys
|
|
|
|
|
endpoint. Neither a loopback URL nor installing source admits production keys.
|
|
|
|
|
See `verifier/README.md` for build, source pin/update and installation requirements.
|
|
|
|
|
|
|
|
|
|
Trust schema `secrets-engine.pdp-trust.v1` requires exactly `schema`, `evaluator`,
|
|
|
|
|
`mode`, `policy_package`, `policy_version`, `policy_package_digest`, and `keys`
|
|
|
|
|
(a map of signer kid to base64url-unpadded Ed25519 public key). The producer has
|
|
|
|
|
no separate issuer or instance identifier: issuer binding is the admitted signer
|
|
|
|
|
plus signed provenance evaluator/mode; the exact policy coordinate and artifact
|
|
|
|
|
digest identify the admitted pin. Deployment key custody must scope the signer to
|
|
|
|
|
that responder; shared signing keys cannot prove a distinct pod identity.
|
|
|
|
|
|
|
|
|
|
Rotation is an explicit reviewed admission update: admit old/new keys for a bounded
|
|
|
|
|
overlap, switch producer signer, verify the actual consumer, then remove the old
|
|
|
|
|
kid and prove refusal. Keys are reloaded each call; no endpoint discovery, cached
|
|
|
|
|
key fallback, unsigned exception, or automatic admission exists. These source
|
|
|
|
|
checks do not constitute Platform key admission or native rollout acceptance.
|