---
id: TEN-WP-0002
type: workplan
title: "Service skeleton, domain model, and the three boundary-contract APIs"
domain: infotech
repo: tenant-engine
status: finished
owner: codex
topic_slug: netkingdom
created: "2026-07-23"
updated: "2026-07-23"
state_hub_workstream_id: "f2fcafb4-edef-4e8e-9401-d781b4d0c585"
---

# Service skeleton, domain model, and the three boundary-contract APIs

First real implementation workplan. Builds the tenant-engine skeleton
against `net-kingdom/canon/standards/tenant-engine-boundary-contract_v0.1.md`:
tenant/grouping/capability-role/plan-grant domain model, an audited grant
trail, and the three API surfaces the contract defines — cache-read
(`key-cape`), live-lookup (`flex-auth`), and write (grant/revoke/plan
mutation). Guardrail/quota policy stays a reserved, unimplemented namespace
per ADR-0014 — not in scope here.

**Depends on:** `net-kingdom` canon — `iam-profile_v0.3.md`,
`tenant-engine-boundary-contract_v0.1.md`, `ADR-0013`, `ADR-0014` (all
ratified). **Non-goals:** payment processing, pricing-model definitions
(`adaptive-pricing`'s job), real `flex-auth` integration (the write/live
APIs get a policy-hook interface, not a working `flex-auth` client — that's
a follow-up once `flex-auth` has a reachable endpoint), guardrail/quota
enforcement.

## Task: Service skeleton

```task
id: TEN-WP-0002-T01
status: done
priority: high
state_hub_task_id: "d22b0ce8-3f25-4c8d-b54f-ba44c65159c1"
```

Python 3.12 + FastAPI, matching `qonto-assistant`'s layout convention:
`pyproject.toml`, `Makefile` (`install-dev`, `test`, `lint`, `run`), package
layout separating `domain/` (pure), `store/` (persistence), `api/` (FastAPI
routers), from day one — not refactored in later.

Done when: `make test` runs an empty/smoke suite; `make run` starts a bare
FastAPI app with a `/health` endpoint.

**Done 2026-07-23:** `pyproject.toml` + `Makefile` land, mirroring
`qonto-assistant`'s exactly (`install-dev`/`test`/`lint`/`run`).
`src/tenant_engine/{domain,store,app,main}.py` scaffolded per the planned
layering (`domain.py` pure, `store.py` the persistence seam, `app.py` the
FastAPI factory). Verified live: `python -m tenant_engine.main` starts on
`127.0.0.1:8090`, `GET /health` returns
`{"status": "ok", "service": "tenant-engine", "version": "0.1.0"}`.

## Task: Domain model — tenant, grouping, capability role, plan grant

```task
id: TEN-WP-0002-T02
status: done
priority: high
state_hub_task_id: "cfceea30-f383-4781-a7e4-226653aafca9"
```

Pure domain types, no framework dependency:

- `Tenant`: id, identifier (`tenant:<grouping>:<name>` or reserved
  `tenant:platform`/`tenant:coulomb`), grouping (ADR-0013 enum, nullable for
  the two reserved identifiers).
- `CapabilityRole`: enum `PLTF`/`IAM`/`VEN`/`CUS`.
- `RoleGrant`: tenant_id, role, grant_reason (`plan_assignment` /
  `manual_grant` / `platform_default`), plan_id (nullable), granted_by,
  granted_at, revoked_at (nullable), correlation_id — the audited record
  shape from the boundary contract's Tenant Role & Plan Grant Contract.
- `PlanAssignment`: tenant_id, plan_id (references an `adaptive-pricing`
  plan id — stored as an opaque string, never resolved or duplicated
  locally), assigned_at.

Validation rules encoded as domain invariants, not just API-layer checks:

- grouping must be one of ADR-0013's twelve values, or the tenant identifier
  must be exactly `tenant:platform`/`tenant:coulomb` (grouping-less);
- a `trial`-grouped tenant may hold any role with `grant_reason:
  platform_default` and no `plan_id` (ADR-0014) — this must be
  representable, not blocked by a plan-required constraint;
- non-`trial` roles other than `platform_default` require a `plan_id` when
  `grant_reason: plan_assignment`;
- revoking a grant sets `revoked_at`, never deletes the record (audit trail).

Done when: unit tests cover valid/invalid grouping values, the trial/
no-plan-required exception, and grant/revoke as append-only operations.

**Done 2026-07-23:** `domain.py` implements `Tenant`, `CapabilityRole`,
`RoleGrant`, `PlanAssignment`, and `create_role_grant()` exactly as
specified. One refinement made while implementing, not pre-specified in the
task text: `platform_default` is valid for `trial`-grouped tenants **or**
the reserved, ungrouped `tenant:platform`/`tenant:coulomb` tenants (their
baseline roles were never purchased either) — the task only mentioned the
`trial` case. Also enforced `platform_default` grants must not carry a
`plan_id` (implied by "no `plan_id`" in the task text, made an explicit
invariant). `tests/test_domain.py` covers all listed cases plus the
reserved-tenant extension and double-revoke rejection. `pytest` →
29 passed (whole-repo total, includes T03's tests below).

## Task: Storage layer

```task
id: TEN-WP-0002-T03
status: done
priority: high
state_hub_task_id: "ade9374a-ed26-45da-a424-a0d21effe520"
```

In-memory store behind a `TenantStore` protocol/interface (mirrors
`qonto_client.QontoClientProtocol`'s pattern in `qonto-assistant`), so a
real backend can be swapped in later without touching `domain/` or `api/`.
Every mutation emits a domain event (tenant created, role granted, role
revoked, plan assigned) per the boundary contract's Audit Correlation
Contract — the event bus itself can be a simple in-process list for now.

Done when: unit tests cover create/read/grant/revoke/assign-plan through the
store interface, plus event emission for each mutation.

**Done 2026-07-23:** `store.py`'s `TenantStore` Protocol +
`InMemoryTenantStore` implement create/get/grant/revoke/active_roles/
assign_plan, each emitting a `DomainEvent`. `tests/test_store.py` covers
the full lifecycle, non-exclusive role coexistence (`CUS` + `VEN`
simultaneously, proving ADR-0014's non-exclusivity isn't just a comment),
revoke-keeps-record-but-clears-active-set, and asserts the exact event
sequence for a full mutation chain. `python -m compileall src tests` clean.

## Task: Cache-read API (for key-cape)

```task
id: TEN-WP-0002-T04
status: done
priority: high
state_hub_task_id: "425cab98-ab7f-47fb-a8b6-ec0f52d62cd5"
```

`GET /tenants/{tenant_id}/roles` — returns current, non-revoked capability
roles for a tenant. This is the endpoint `key-cape` calls at token-issuance
time to source the cached `tenant_roles` claim (IAM Profile v0.3). No
authorization gate of its own beyond service-to-service auth (out of scope
here — see Non-Goals); this endpoint's whole purpose is to be cheap and
fast, per the boundary contract's performance model.

Done when: integration test hits the endpoint against the in-memory store
and returns the expected role set for a seeded tenant.

**Done 2026-07-23:** Implemented in `app.py`. Returns
`{"tenant_id": ..., "roles": [...]}`, sorted role-value strings; 404 for an
unknown tenant. Extended beyond the task's minimum: also fails closed (503)
on store unavailability (see T05) — there was no good reason for the two
read endpoints to behave differently on that axis, and keeping them
identical avoids a second, subtly-different error-handling path to drift
later. `tests/test_api_reads.py` covers the happy path and the 404 case;
verified live over real HTTP against the running service.

## Task: Live-lookup API (for flex-auth) — fail closed

```task
id: TEN-WP-0002-T05
status: done
priority: high
state_hub_task_id: "2a04ce4c-c967-4540-8873-70ac419378e6"
```

`GET /tenants/{tenant_id}/roles/live` — same data as the cache-read
endpoint, but explicitly documented and tested as the path `flex-auth` must
call before authorizing `aal2`-class actions. The distinction from T04 is
operational intent (freshness guarantee, called synchronously on a
privileged-decision path), not payload shape.

Per the boundary contract's performance model: **fail closed, never open**.
If the store is unavailable, this endpoint must return an error response
that a policy caller would treat as "deny", not a default-allow or an empty
role list indistinguishable from "no roles granted".

Done when: a test simulates store unavailability and asserts the endpoint
signals failure distinctly from "zero roles", not silently as 200 + `[]`.

**Done 2026-07-23:** New `store.StoreUnavailableError`; both read endpoints
convert it to `503 {"detail": "tenant_roles_unavailable"}`. A `_BrokenStore`
test double (`tests/test_api_reads.py`) always raises it from
`active_roles()`, simulating an outage; the test asserts `503` and
explicitly asserts the response body is *not*
`{"tenant_id": ..., "roles": []}` — the exact ambiguity the task exists to
prevent. `TenantNotFoundError` stays a distinct `404`, so "tenant doesn't
exist," "store is down," and "tenant exists with zero roles" are three
different, distinguishable responses, never collapsed into one shape.

## Task: Write API — grant, revoke, assign-plan

```task
id: TEN-WP-0002-T06
status: done
priority: medium
state_hub_task_id: "3a2d1ee7-6e66-4080-9345-32ba457acf5f"
```

`POST /tenants` (create), `POST /tenants/{id}/roles/grant`,
`POST /tenants/{id}/roles/revoke`, `POST /tenants/{id}/plan` — all mutating
endpoints from the boundary contract's Write API. Per the contract,
`tenant-engine` does not self-authorize these writes; `flex-auth` is meant
to gate them. Since a reachable `flex-auth` integration is explicitly a
non-goal here, implement a `WriteAuthorizer` protocol/interface point (one
class, default-deny stub) so every mutation already flows through a single
seam — swapping in a real `flex-auth` client later touches one file, not
every endpoint.

Done when: unit tests confirm every write endpoint calls the
`WriteAuthorizer` seam and is denied by the default-deny stub; a
test-only authorizer override proves the seam actually gates the mutation
when swapped.

**Done 2026-07-23:** `authz.py`'s `WriteAuthorizer` Protocol +
`DefaultDenyWriteAuthorizer`; every write endpoint in `app.py` calls
`authorizer.authorize(...)` before touching the store, and a
`WriteAuthorizationDeniedError` exception handler maps denial to
`403 {"error_code": "write_denied", ...}`. `tests/test_api_writes.py`
asserts all four write endpoints are `403` under the default authorizer,
then swaps in an `_AllowAllAuthorizer` test double and exercises the full
create → grant → read → revoke → read → assign-plan lifecycle over real
HTTP requests (`TestClient`), confirming the seam actually gates rather
than just existing decoratively. Domain/store errors surfacing after
authorization passes get separate status codes: `400` invalid
identifier/grant, `404` unknown tenant, `409` duplicate tenant. Verified
live: unauthenticated `POST /tenants` over real HTTP returns `403` with the
expected body, matching the test suite. One known simplification, not
resolved here: `actor` is a request-body field rather than extracted from a
real auth context — there is no real auth context yet, since wiring one in
is exactly what a `flex-auth`-backed `WriteAuthorizer` will do; noted for
whoever picks up that follow-up.

## Task: Closure review

```task
id: TEN-WP-0002-T07
status: done
priority: low
state_hub_task_id: "eecf4bc8-d20b-4b18-986a-6518f2f74d7b"
```

Confirm T01–T06 done; run `make test`/`make lint`; verify the three API
surfaces match the boundary contract's Ownership Model and Source-of-Truth
Matrix with no drift. Note follow-ups: real `flex-auth` `WriteAuthorizer`
integration, `key-cape` wiring to actually call the cache-read endpoint at
issuance, guardrail/quota policy design (ADR-0014's reserved item), and
Binky Hedgehog GmbH as the first real tenant record once `key-cape`'s
`KEY-WP-0004` reaches that point. Run `statehub fix-consistency`.

**Closed 2026-07-23.** T01–T06 all done. `PYTHONPATH=src pytest` → `43
passed`; `python -m compileall src tests` clean; `ruff` unavailable in this
workstation's shared venv (same environment gap noted in `qonto-assistant`'s
own workplans) — `compileall` substituted, `make lint` untested against a
real `ruff` install. Verified live over real HTTP, not just `TestClient`:
`/health` 200, unauthenticated `POST /tenants` 403 with the expected error
body, `GET` on an unknown tenant 404.

**Ownership check against the boundary contract:** all three API surfaces
present (cache-read, live-lookup, write); every write routes through the
`WriteAuthorizer` seam, never self-authorized; every mutation emits a
`DomainEvent`; `plan_id` is stored and returned as an opaque string, never
resolved against `adaptive-pricing` locally; nothing here stores user data,
issues tokens, or makes an authorization decision. No drift found.

**Follow-ups, not started:**
- Real `flex-auth`-backed `WriteAuthorizer` (replaces the default-deny
  stub) — also where a real `actor` identity would come from instead of a
  request-body field.
- `key-cape` wiring to actually call the cache-read endpoint at token
  issuance.
- Guardrail/quota policy design (ADR-0014's reserved item — spend limits,
  entity/action counts).
- Binky Hedgehog GmbH as the first real tenant record, once `key-cape`'s
  `KEY-WP-0004` reaches that point.
- Persistence beyond in-memory (`TenantStore` is already a swappable seam
  for this).