Commit graph

2 commits

Author SHA1 Message Date
03c0569820 feat: adopt the owner-documented PDP access path, and prove it live
All checks were successful
CI Smoke / host-smoke (push) Successful in 0s
CI Smoke / container-smoke (push) Successful in 2s
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
f62d3fe789 feat: complete and prove the authorization chain end to end
Implements step 2 (access-engine POST /v1/check) and wires the whole
GH-DEC-2026-003 sequence together, then proves it against a live throwaway
OpenBao rather than only unit-level fakes.

- decision_check.check_decision performs the PDP call; an unreachable or
  non-200 PDP raises, since silence is never permission.
- approval_consume.authorize_action coordinates steps 1 and 2 and returns an
  AuthorizedAction. Both steps build the same CheckRequest via a shared
  _expected_request, since two descriptions of the action cannot produce
  corresponding digests.
- apply_unreachable_engine_stance takes authorized=. The published map
  defines fail_closed as no side effect WITHOUT a durable decision record,
  so holding a validated one means the residue does not apply. Not a bypass:
  both steps must have succeeded and CAS consume still precedes OpenBao.
  Unconfigured still returns None and fails closed.

The end-to-end test caught one more instance of the cross-vocabulary bug: a
leftover comparison of the claim's binding.action against ours. The claim
says secrets.kv.destroy where we say destroy, so it would have failed against
every real claim. Removed; the tie is pdp_digest.

Integration coverage asserts PIP-then-PDP ordering, that consume is the last
step before the backend, and that an unreachable PDP, denied decision,
invalid claim, missing pdp_digest, consume conflict and action mismatch each
stop before OpenBao. 284 tests pass; production still fails closed.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M65ovP3eiiPHubibvWs9mD

Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 393550@bnt-lap001
Assistant-Session: 4bb359f9-1f12-4410-9e76-079cf23c82e4
2026-09-06 14:56:02 +02:00