RPF-WP-0035-T06. Adds CCR-2026-0019 (secrets-engine) and CCR-2026-0020 (approval-engine-operator) with their exact-path read policies, reusing the existing version-1 custody from the verifier activation. No reseed, rotation, shared reader or verifier Secret reuse; both requests are in_flight and nothing is applied. The two shapes were decided by read-only survey rather than assumed. secrets-engine consumes its client secret through an operator-run CLI reading a protected file, and its namespace holds no workload, so reader 1 is an attended operator-workstation OIDC lane rather than an ESO lane; its one missing input is the operator group claim, which NetKingdom and KeyCape own. approval-engine is not deployed and no owner source names who presents the operator client, so reader 2 records the undetermined actor instead of guessing one for the widest approval scope in the pair. Both declare openbao.auth missing rather than carrying a placeholder binding. T06 moves to wait on those two owner inputs. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WLUjpv3ssxNRAEPPgLFnEB Assistant: claude-code Assistant-Model: opus Assistant-Process: 1275505@bnt-lap001 Assistant-Session: 97265baa-f08f-4032-b290-a1e2965a69c5
4.8 KiB
Approval client-side readers
RPF-WP-0035-T06, the residual handed over by the completed verifier custody
(T05). Requests:
CCR-2026-0019
and CCR-2026-0020.
Both are in_flight, neither is approved, and nothing is applied.
Why this exists as separate requests
CCR-2026-0017/0018 admitted the verifier copies: KeyCape holds the two client
secrets so it can authenticate credentials presented to it. Presenting them is a
different power with a different holder, so it needs its own admission. The two
Warden fetch selectors that the verifier review left unresolved are exactly this
gap. Both readers use the existing version-1 custody — no reseed, no
rotation, no shared reader, no reuse of the sso verifier Secrets.
Survey, 2026-09-09
Read-only, against owner source and the live cluster.
| Question | Finding |
|---|---|
| Is there an in-cluster secrets-engine consumer? | No. Namespace secrets-engine holds only ServiceAccount secrets-engine and no workload. flex-auth/flex-auth-secrets-engine is the flex-auth PDP for that consumer, not the service. |
| How does secrets-engine consume the client secret? | Operator-run CLI reading SECRETS_ENGINE_APPROVAL_CLIENT_SECRET_FILE, exchanged for a fresh token before each request, held in memory only, no refresh token or fallback identity. Implemented 2026-09-09, SECRETS-WP-0009-T03, secrets-engine revision 9eb07fd. |
| Is approval-engine deployed? | No. No namespace, workload or Service. The owner manifest targets namespace approval-engine with a digest-pinned image and is unapplied. |
Who presents approval-engine-operator? |
Not identified by any owner source. approval-engine verifies these tokens and never presents them. |
Reader 1 — secrets-engine-approval (CCR-2026-0019)
Shape is determined; one input is missing.
| Item | Value |
|---|---|
| Path / field | platform/workloads/secrets-engine/approval-client / CLIENT_SECRET |
| Policy | workload-kv-read-secrets-engine-approval-client (exact data + metadata read, written and committed) |
| Auth | oidc on mount netkingdom, role secrets-engine-approval-client-read, standard three redirect URIs, scopes openid profile email groups, user_claim: sub, groups_claim: groups, ttl 15m |
| Front door | warden access secrets-engine-approval-client --out FILE |
| Delivery | Protected owner-only file outside any Git work tree, consumed as SECRETS_ENGINE_APPROVAL_CLIENT_SECRET_FILE, removed at end of session |
Missing: the exact group claim for the authorized operator. NetKingdom and
KeyCape own it; this repo and secrets-engine cannot supply it. The role cannot
be created without it, so openbao.auth is declared missing rather than
recorded with a placeholder that would later be mistaken for a confirmed
binding.
This is an operator-workstation lane, not an External Secrets lane — because the consumer is a CLI, not a pod. A file-delivered credential outlives the process that read it, so removal depends on operator hygiene. That is weaker than the pod-lifetime bound of the verifier delivery and is recorded as a consequence of the consumer's shape rather than engineered away.
Reader 2 — approval-engine-operator (CCR-2026-0020)
Held open deliberately. The exact-path read policy
workload-kv-read-approval-engine-operator-client is written and committed so
the eventual reader inherits a bounded grant rather than one drafted under time
pressure. Writing the policy is not admitting a reader.
Missing: the presenting actor and its placement. This client carries the
widest scope set in the pair — approval:create, read, approve, revoke,
supersede, observe, emit, and deliberately not consume. approval-engine
records principal_type on every entry rather than restricting who may supply
approver evidence, and its owner records that whether a non-human principal
should is approval doctrine belonging to gate-house. Naming a reader here
would hand the widest approval scope to a guessed identity, so the request names
what is undetermined instead.
The delivery surface follows the actor: an attended operator implies a protected
workstation file; an in-cluster component implies an ExternalSecret into the
eventual approval-engine namespace; an agent surface implies neither.
Not in scope
- No reseed or rotation of either custody path. Rotation is one version-guarded operation shared with the verifier copies and needs both owners.
- No
approval:consumeon the operator client, now or by later scope edit. - No reuse of the
ssoverifier Secrets or the KeyCape verifier roles. - No audit-sender custody; that stays with the existing owner records.