key-cape/docs/tenant-claim-contract.md
tegwick f9812ab3b2 Carry approval-engine's store-isolation limit into the tenant contract
approval-engine chose the registration-bound shape for the consuming side and
asked that the limit making it admissible be carried into the contract rather
than left in a message.

Their check is store isolation -- does this caller belong to the store this
engine serves -- not membership. A registration-supplied tenant is admissible for
that question. It is not admissible for any doctrine turning on the approver's own
membership, which is a fact about the person that this claim cannot carry, and an
exact-match check does not become that claim merely by matching.

That is GH-DEC-2026-013 section 1 reached independently from the consuming side,
and it is the practical reason tenant_source exists: a consumer whose check means
store isolation can accept registration, one whose check means membership must
require directory. Recorded next to the provenance table so the two are read
together.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016uV8zoCKpA1WRAxsKRYbdH

Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 1182213@bnt-lap001
Assistant-Session: 966597b9-ae61-46a4-8b9e-1594ab3ec4ad
2026-09-10 13:44:02 +02:00

234 lines
13 KiB
Markdown

# Tenant claim contract
States what KeyCape owns and emits for the `tenant` claim. Raised by glas-harness
message `356f6977-d361-4e3b-83ab-b2c7f4759286` (GLAS-WP-0015) and resolved by
operator decision `5ed3fb35-eca9-413a-82b9-95171ba85bf6` — "Use tenant:platform
for the Glas approval dependency chain", recorded in
`glas-harness/docs/platform-tenant-decision.md`.
`tenant:platform` is the platform management, administration and services tenant
(the landlord zone). KeyCape emits it verbatim for the two approval clients and
still creates no mapping between tenant vocabularies: no alias to `platform` or
`tenant:coulomb`, and no implicit cross-tenant grant.
## What KeyCape emits
The `tenant` claim is bound at registration or directory-resolution time and is
never influenced by request parameters.
| Principal | Source of `tenant` | Value in the reviewed registrations |
| --- | --- | --- |
| Service (`client_credentials`) | The client's `tenant` field. Required by config validation — a `client_credentials` client without `serviceSubject` and `tenant` fails startup validation. | In `config/service-clients.example.yaml`: `tenant:platform` for `secrets-engine-approval` and `approval-engine-operator`; `tenant:coulomb` for `codex-railiance-platform` and `secrets-engine-openbao`. |
| Human (`authorization_code`) | The directory user's tenant, falling back to the platform default when unset. | Directory value, else `tenant:coulomb`. |
`tenant`, `tenant_hint`, `audience` and `resource` request parameters cannot
change the claim. `tenant_hint` on `/authorize` reaches registration/enrollment
handoffs only; it is not a claim input.
## The approval chain, after the decision
All three layers now use one exact spelling, `tenant:platform`:
| Layer | Value | Owner |
| --- | --- | --- |
| Approval store tenant | `tenant:platform` | approval-engine |
| Service-client JWT `tenant` claim | `tenant:platform` | KeyCape (this repo) |
| Lifecycle CheckRequest tenant | `tenant:platform` | flex-auth |
KeyCape performs **no** normalization, prefix-stripping or aliasing, so a
resource server compares `tenant` as an exact string. The previously observed
`platform` and `tenant:coulomb` spellings are **not** accepted aliases for
`tenant:platform`; a token carrying either must be denied by the resource server.
## Scope of the alignment
The decision changed the `tenant` field on exactly two registrations,
`secrets-engine-approval` and `approval-engine-operator`, plus the matching
entries in `docs/approval-engine-provisioning-request.yaml`. Deliberately
unchanged:
- `codex-railiance-platform` and `secrets-engine-openbao` keep `tenant:coulomb`.
- Human directory resolution keeps its `tenant:coulomb` default.
- No cross-tenant grant is implied: a `tenant:platform` token conveys no reach
into `tenant:coulomb` resources, and audiences, scopes, subjects, roles,
lifetimes and MFA requirements are untouched.
This resolves the *choice* of tenant only. Live provisioning of the two clients
remains gated on deployment-owned custody and the separate verification gates in
KEY-WP-0013-T02; KeyCape changed no live registration or policy subject.
## Evidence
`src/internal/server/oidc/tenant_test.go`:
- `TestServiceTenantIsBoundToRegistrationAndIgnoresRequestParameters` — a request
supplying `tenant=tenant:platform` and `tenant_hint=platform` still yields the
registered `tenant`, signature-verified against `/jwks`.
- `TestServiceTenantsAreDistinctPerRegistration` — two registrations with
different tenants yield their own values and never each other's; this is the
wrong-tenant denial basis for an exact-comparison resource server.
- `TestHumanTenantClaimUsesDirectoryValueThenPlatformDefault` — human tokens carry
the directory tenant, defaulting to `tenant:coulomb` rather than an empty claim.
- `TestApprovalClientIssuesExactPlatformTenantAndRejectsAliases` — the
approval-client shape issues `tenant:platform` exactly, with
`aud=approval-engine`, even when the caller asks for `platform`,
`tenant:coulomb` or `TENANT:PLATFORM`.
- `TestUnrelatedServiceClientKeepsCoulombTenant` — the OpenBao login client keeps
`tenant:coulomb` even when the request asks for `tenant:platform`, so the
alignment grants no cross-tenant reach.
`src/cmd/keycape/clients_test.go`:
- `TestServiceRegistrationTenantsAreExactPerDecision` — loads the real
registration fixture and pins the exact tenant of every reviewed client, so
drift or an alias reintroduced into `config/service-clients.example.yaml` fails
the build.
## How a human token's tenant is resolved (KEY-WP-0013-T05)
A human's tenant is normally a property of the person, read from the directory
record. That alone could not serve the approval chain: decision
`5ed3fb35-eca9-413a-82b9-95171ba85bf6` binds it to the landlord zone,
approval-engine compares the claim by exact string equality, and no adapter
populates `domain.User.Tenant` — so every human token fell back to
`tenant:coulomb` and an approver token would have been refused downstream. It
would have presented as a failed approval rather than as a registration defect.
A client registration may therefore declare a `tenant`, and `humanTenant` in
`src/internal/server/oidc/token.go` resolves it by these four rules:
| Client declares | Directory assigns | Result |
| --- | --- | --- |
| nothing | anything | the directory answer, or `tenant:coulomb` — unchanged |
| a zone | nothing | the declared zone |
| a zone | the same zone | that zone; client and directory agree |
| a zone | a **different** zone | **issuance is refused** |
The last row is the point, and it is **normative** — GH-DEC-2026-013 condition
(a). A registration can bind a zone for users the directory has not placed, and
can never relabel a user it *has* placed. That case fails closed rather than
picking a winner, because either answer would be a silent cross-tenant
assertion.
Note the direction that cuts, which is not the intuitive one: **a future change
preferring the directory is also void.** "The directory won" is still a winner,
and picking any winner converts a refusal into a silent cross-tenant assertion.
Row four is not a design choice available for later optimisation in either
direction.
This shape is a **declared bounded gap**, not the terminal state. Directory-
sourced is (GH-DEC-2026-013 §1): a registration is a statement about the *actor*,
a tenant is a property of the *principal*, and sourcing the second from the first
collapses two identities in the direction that widens. It was granted because the
one case distinguishing it from the correct resolution fails closed, and because
it converges on the terminal state by subtraction rather than needing to be
unwound — the same code stops supplying the zone and starts enforcing agreement
with it the day the directory carries tenants. On that day §2 is *spent*, not
merely unused. Registered as a §13 gap at `net-kingdom@f9e1611`; the directory
adapter has no named owner yet. The refusal is a `403` with
`error_type: tenant_binding`, distinct from an authentication failure, so an
operator can tell a misconfigured registration from a rejected login.
This also means the design survives the other resolution. If the directory later
carries tenants, the same code stops supplying the zone and starts enforcing
agreement with it — no second migration, and no window in which a stale
registration silently wins.
### Condition of the capability
This is not a caveat on the reasoning; it is a condition of the capability, and
it is stated here so a future change to registration policy has to confront it.
> **A client-declared tenant is safe only while client registrations are static
> and deployment-owned.** If KeyCape ever admits dynamic client registration,
> anyone able to register a client can name its users' tenant and relabel
> unplaced users into a zone. The deliberate binding becomes an escalation.
>
> **Lifting the exclusion voids this rule rather than reopening it**
> (GH-DEC-2026-013 condition (b)). This document previously said the rule "must
> be revisited", and gate-house strengthened that against us: *revisited* implies
> the answer might survive review, and it would not. If dynamic client
> registration is ever admitted, the registration-bound shape is void **that
> day** and the directory becomes the only source, whatever state the adapter is
> in. Removing this paragraph is not the decision.
Recorded at the request of informed-decision, who will hold the approver
registration and asked for it in the contract rather than in a message.
A document can be missed, so the condition is also enforced.
`TestRegistrationBoundTenantRequiresStaticRegistration`
(`src/internal/server/oidc/tenant_precondition_test.go`, KEY-WP-0030) asserts
both halves together: that a client-declared tenant still issues, and that
discovery advertises no registration endpoint. Whichever is removed first, the
failure points at the other — which is the part that matters, since
`tests/profile` has long asserted the endpoint's absence on its own and that
assertion reads as discovery metadata rather than as a warning about relabelling
users. Dynamic client registration is a deliberate exclusion in
[SCOPE](../SCOPE.md); this ties the tenant capability to that exclusion so the
two cannot drift apart silently.
Covered by `src/internal/server/oidc/human_tenant_test.go`, including the
relabel refusal.
### The claim carries its provenance
GH-DEC-2026-013 §5 found what the rules above leave unsaid: `tenant` is a bare
string, so a consumer cannot tell a zone **the directory asserted about the
person** from one **a registration supplied about the client they came through**.
approval-engine admits an approver by exact-matching that string while its
contract reads as though it relies on the first. The check is sound; the property
a reader infers from it is absent. gate-house named the property and left the
mechanism to us.
Every token therefore carries `tenant_source` alongside `tenant`:
| `tenant_source` | Meaning |
| --- | --- |
| `directory` | The identity layer asserted this zone about this person. Includes the agreement case — if a registration declared the same zone the directory did, the directory still asserted it, so the stronger provenance is the true one. |
| `registration` | Supplied by the client registration for a person the directory has placed nowhere. A fact about the client, not about the person. Always the value on service tokens: there is no directory principal behind one. |
| `default` | Nobody asserted a zone. This is the profile's non-empty default. |
**The third value is not padding.** The ruling names two sources, but the code
has three states, and labelling an unasserted default as `directory` would
reproduce the same defect one level down — a consumer reading an assertion the
identity layer never made. That is the unknown-versus-absent distinction
GH-DEC-2026-011 §3 requires, applied to our own fallback rather than only to the
case we were asked about.
**For consumers.** Do not treat the three as equivalent for any decision that
turns on a fact about the *person*. `registration` and `default` are not weaker
evidence of the same thing; they are evidence of a different thing. What follows
from that is the consumer's call — gate-house explicitly did **not** rule on
whether approval-engine's exact-match admission is the right gate. Carrying
provenance makes that question answerable; it does not answer it.
**What approval-engine's check actually asks, in their words.** Asked which
resolution they wanted, approval-engine chose the registration-bound shape and
stated the limit that makes it admissible for them, which they asked to be
carried here so it cannot be over-read later:
> Their tenant comparison is **store isolation**. It answers *does this caller
> belong to the store this engine serves*, not *is this human a member of the
> platform zone*. For that question a registration-supplied tenant is admissible:
> admission means arriving through a channel the platform registered, and the
> approver client is static and deployment-owned.
>
> **A registration-supplied tenant is not admissible for any doctrine that turns
> on the approver's own membership.** "An approver must be a member of the
> platform zone" is a fact about the person, and this claim cannot carry it. No
> such doctrine exists today. If one is issued, it needs a directory-sourced
> claim — and an exact-match check does not become that claim merely by matching.
That is the same conclusion as GH-DEC-2026-013 §1 reached independently from the
consuming side, and it is the practical reason `tenant_source` exists: a consumer
whose check means *store isolation* can accept `registration`, and a consumer
whose check means *membership* must require `directory`. The claim does not know
which question is being asked; the consumer does.
`tenant_source` is advertised in `claims_supported`, and
`TestIssuedTokensCarryTenantProvenance` asserts it reaches issued tokens on both
grants, including the unasserted-default case.
These are local issuance proofs. They are not live-rollout evidence; see
`docs/approval-engine-auth-contract.md` and KEY-WP-0013-T02 for the deployment
boundary. No token or secret values appear in this document or in test output.