ops-warden/tenancy.yaml
tegwick 7ce58ae638
All checks were successful
CI Smoke / host-smoke (push) Successful in 0s
CI Smoke / container-smoke (push) Successful in 1s
feat: adopt security zones and explicit workload refs
Assistant: codex
Assistant-Model: gpt-5.6-sol
Assistant-Session: 01a0291a-1e87-7151-9934-fcbfe3f65eb1
2026-08-22 15:36:37 +02:00

172 lines
7.7 KiB
YAML

# ops-warden tenancy posture declaration
# Framework: net-kingdom/canon/standards/tenancy-posture_v0.1.md draft-9
# Conformance rule (§6): accuracy, not altitude. Nothing here is aspirational.
# Validate: python3 ~/net-kingdom/tools/tenancy-posture/validate.py tenancy.yaml
schema_version: "0.1"
framework: netkingdom-tenancy-posture
service: ops-warden
role: ssh-certificate-authority
workload_identity:
name: ops-warden
kind: operational-control-plane
responsible_repo: ops-warden
identity_bindings:
- scheme: kubernetes-service-account
authority: railiance01
subject: system:serviceaccount:ops-warden:ops-warden
principal_type: service
environment: prod
tenancy:
current:
I: 1
A: 1
E: 0
P: "n/a"
R: "n/a"
V: 0
implemented:
A: 3
target:
I: 1
A: 3
E: 0
P: "n/a"
R: "n/a"
V: 1
reviewed: "2026-08-19"
review_due: "2027-02-18"
service_class: interactive
permanent: [P, R]
gap:
I: >-
ops-warden has a tenant notion — `policy.tenant` in warden.yaml, and the
tenant/platform path split introduced by WP-0028 — but it is a static
configuration constant, not a claim verified on an inbound call. `warden`
is a CLI invoked by an operator or agent; its caller is an OS user and no
token is presented to it. Subject id comes from `WARDEN_POLICY_SUBJECT`
or falls back to the actor name, which is §4.1's "taken from the request"
case exactly. I1.
Not declared permanent. `warden desk` (WP-0029) already performs an OIDC
login against key-cape, so a verified inbound identity is reachable
rather than structurally excluded. It is simply not built, and I2 would
require it on the signing path, not just the desk.
A: >-
There is a single choke point on the signing path — inventory membership,
actor type, and the TTL ceiling — but it binds an *actor* to principals,
not a request to the tenants it may act for. Tenant context is a
constant, so "bound once, centrally" (§4.2 A2) would be true only in a
trivial sense that overclaims. A1.
`implemented: A3` is the honest separate fact: delegation to flex-auth as
PDP is built (`src/warden/policy.py`, `check_sign_policy`) and was
verified live on 2026-08-19 against the enforcing `flex-auth-ops-warden`
pin — `decision:f3f7c88f9585582a`, with an anonymous `/v1/check`
returning 401. It is not `current` because `policy.enabled` is false, and
it is false by decision rather than by blocker: `ADR-0006` scopes
enforcement to security zones, which `zone-engine` is defining
(`ZONE-WP-0001`). Evidence:
`history/2026-08-19-flex-auth-caller-identity-evidence.md`.
A4 is not a target here. The AuthZEN interface question belongs to
flex-auth as the decision point; ops-warden would follow it, not lead it.
E: >-
E0 is accurate and is not a defect to remediate. ops-warden holds no
tenant-partitioned data: `registry/routing/catalog.yaml` is a pointer
layer carrying no secret values (`ADR-0001`, CI-enforced) and no `tenant`
field on any entry; local state (`signatures.log`, `audit.jsonl`,
`access-audit.log`) is operator-scoped and keyed by actor, not tenant.
The tenant boundaries ops-warden *routes* to — `tenants/binky/...` versus
`platform/workloads/...` — are enforced by OpenBao policy, which is the
owner's control, not ops-warden's. Claiming E1 on the strength of someone
else's enforcement is the overclaim §6 prohibits. Under `ADR-0002`
ops-warden is a transparent conduit and takes no custody, so it has no
tenant data to key. Raising E would mean acquiring data it is out of
scope to hold.
P: >-
n/a and permanent. No primary datastore: state is files under
`~/.local/state/warden` on the invoking operator's machine, and
§3.3 scopes P to the primary datastore. Custody of secrets is explicitly
out of scope (`SCOPE.md`, `ADR-0002`), so ops-warden will not acquire a
tenant-bearing substrate.
R: >-
n/a and permanent, on the same ground as P: no tenant data at rest.
Named honestly rather than hidden behind the n/a: ops-warden *does* keep
local operator records indefinitely — `audit.jsonl` is append-only by
design (WP-0022) and `signatures.log` has no retention position. They are
metadata-only and guarded against secret material, and they are not
tenant data, so they do not move this axis. But "no declared retention
for local audit" is a real gap on a different axis than this file grades,
and is recorded here so it is not lost.
V: >-
V0. `warden sign` depends synchronously on OpenBao at
`https://bao.coulomb.social` (railiance01) and, once enabled, on the
flex-auth pin reached through the ops-bridge tunnel
`flex-auth-ops-warden-railiance01`. §4.6.1 takes the minimum across that
path, and none of it has an exercised recovery objective.
A `backend: local` CA exists in the code and is covered by tests, but it
has never been exercised as a production degraded mode, and §13 does not
accept "the code path exists" as V evidence. Target V1 means documenting
and actually rehearsing recovery of the signing path — not adding
redundancy, which the substrate cannot currently support: `reef-railiance`
is single-node with a shared control plane, and under Decision 4.6.1 that
caps V for everything bound to it (see NK-WP-0027).
zones:
standard: security-zones_v0.1
membership: z1-operational
responsible_party: team:platform-security
justification: >-
The attended platform signing service has a bounded operational scope and
internal metadata exposure. It has M1 evidence, but no SLO history, on-call
rotation, or incident exercise that would support z2-protected.
context:
maturity: M1
criticality: medium
data_classification: internal
evidence:
- ref: docs/evidence/security-zone-admission-2026-08-22.md
supports: [M1, platform-only-scope, basic-slo, data-handling-note]
- ref: history/2026-08-19-flex-auth-caller-identity-evidence.md
supports: [production-policy-path, authenticated-caller]
reviewed: "2026-08-22"
review_due: "2026-11-22"
evidence:
A: >-
Built and verified live 2026-08-19 against the enforcing
`flex-auth-ops-warden` pin: `decision:f3f7c88f9585582a`, anonymous
`/v1/check` 401. Record:
`history/2026-08-19-flex-auth-caller-identity-evidence.md`. Re-establish on
demand with `scripts/check_policy_caller_identity.py`. Why it is implemented
and not current: `docs/adr/ADR-0006-enforcement-is-zone-scoped.md`.
E: >-
`docs/adr/ADR-0002-conduit-not-broker.md` and `ADR-0001` (catalog is a
pointer layer, CI-enforced) — why no tenant-keyed data exists to enforce on.
P: >-
No datastore. State is operator-local files under `~/.local/state/warden`.
V: >-
Production signing path verified 2026-06-17
(`history/2026-06-17-openbao-production-verify.md`). No recovery exercise
exists for that path, which is why this is V0 and not V1.
notes:
- >-
ops-warden issues short-lived SSH certificates and routes every other
credential need to its owner.
- >-
Five of six axes are low because it deliberately holds nothing. The low
grades on E, P and R are the intended consequence of ADR-0002 and ADR-0005,
not deferred work, and raising them would mean acquiring data ops-warden is
out of scope to hold.
- >-
The two axes with real movement are A — built, deferred by ADR-0006 pending
the zone model — and V, which needs a rehearsed recovery and is capped by
the substrate until NK-WP-0027 lands.