canned-prompts/workplans/CANP-WP-0006-hosted-registry-service.md

---
id: CANP-WP-0006
type: workplan
title: "Hosted registry and index service"
domain: agents
repo: canned-prompts
status: active
owner: codex
topic_slug: practice
created: "2026-09-06"
updated: "2026-09-06"
state_hub_workstream_id: "70c069e9-7abe-569a-aee2-ffcd15e8970d"
---

# Hosted registry and index service

Toward a production deployment on railiance01 as `rapp-canned-prompts`.

**Sequencing finding.** A `rapp-*` repo packages and operates a deployable its
`ownership_repo` already produces — `declarations/rapp.yaml` pins
`upstream_components.version` to a container digest. `canned-prompts` has no
deployable: a CLI module with no HTTP surface, no image, no database. A rapp
whose `upstream_components` points at nothing is a stub that tells the fleet's
tooling a deployment exists when none does. The service comes first.

**Decisions (operator, 2026-09-06):**

- *Hosted registry plus index* is the v1 surface. § 20 already permits an HTTP
  registry and § 20.3's index is the query surface, so this validates INTENT
  principle 10's "registry-ready" claim rather than inventing a new concept.
- *The service lives in `canned-prompts`*, so `rapp.yaml` gets
  `ownership_repo: canned-prompts` — the `sbom-nexus` shape, where product
  ownership stays out of the operations repo.
- *Platform service, tenant-keyed regardless.* Deployed as a platform service
  like `sbom-nexus`, but all owned data is keyed by tenant from the first
  migration. `business-app-service-contract_v0.1` § 1.3 makes that mandatory so
  a later consolidation is a data-copy rather than a rewrite; a collaborative
  prompting platform will need it, and retrofitting is the expensive path.
- *Service first, rapp when there is a digest to pin.*

**Stack**, matching `state-hub` and `sbom-nexus`: FastAPI, uvicorn, SQLAlchemy,
Alembic, Pydantic, PostgreSQL. `RailianceAppDeploymentGuide.md` additionally
requires unauthenticated `/healthz` and `/readyz`.

**Boundary.** The service hosts packages; it does not become an agent runtime.
`INTENT.md`'s deliberate boundary still holds — no model execution, no
orchestration. Rendering stays deterministic, and `derive` stays a declaration
the service does not satisfy.

## Service skeleton and health surface

```task
id: CANP-WP-0006-T01
status: done
priority: high
state_hub_task_id: "6e4178f0-a431-59de-8bd1-fb5b2de95302"
```

A `service/` package inside this repo: FastAPI application factory, settings
via pydantic-settings, and the health surface the deployment guide requires —
unauthenticated `/healthz` and `/readyz`, plus `/state/health` for consistency
with `state-hub` and `sbom-nexus`.

`/readyz` must actually check the database, not return 200 unconditionally. A
readiness probe that cannot fail is a liveness probe with a misleading name.

Keep `reference/` untouched. It is the format's conformance witness and must
stay dependency-light; the service is a separate consumer of the same package
semantics.

**Done.** `service/` with a FastAPI factory, pydantic-settings config, and the
three endpoints. `CANNED_PROMPTS_DATABASE_URL` has no default, so a
misconfigured service refuses to start rather than quietly using a local
database.

**Bug found by its own test.** `check_readiness` first caught every failure in
one `except` and reported "database unreachable". An unmigrated but perfectly
reachable database therefore reported a connection problem — sending an
operator to credentials and networking when the fix was `alembic upgrade`.
Connectivity and schema are now checked separately.

## Tenant-keyed schema and first migration

```task
id: CANP-WP-0006-T02
status: done
priority: high
state_hub_task_id: "bc0e4a49-63b1-5fa1-9ec9-5ed0f505baa1"
```

Alembic migration `0001` creating the store. Every table holding owned data
carries a tenant key from this migration onward — packages, versions, and index
entries — per the service contract § 1.3. No business logic may assume it is
the only tenant (§ 1.4).

Model the three things the format already distinguishes, and do not conflate
them:

- **package version** — immutable content addressed by `<id>@<version>`
  (§ 17), scoped to a registry because identity is registry-scoped (§ 3.2);
- **index entry** — how a version arrived in this store (§ 20.3): source,
  method, `included_at` which is never overwritten, `last_seen_at`;
- **manifest fields** worth indexing for discovery — name, summary, tags,
  license, type.

Store package files as content, not as rows per file.

**Done.** Migration `0001` creates `package_versions`, `package_files` and
`index_entries`; a test asserts every table in the metadata carries `tenant`, so
adding an unkeyed table fails the suite rather than being noticed at
consolidation time. Uniqueness is `(tenant, registry, package_id, version)`.

## Read API

```task
id: CANP-WP-0006-T03
status: done
priority: high
state_hub_task_id: "0ae11e65-68b0-543c-b8fa-f3188a4d239a"
```

`GET /packages` (search over id, name, summary, tags), `GET /packages/{id}`
(versions), `GET /packages/{id}/{version}` (manifest),
`GET /packages/{id}/{version}/archive` (the package files), and
`GET /index` (§ 20.3 entries).

Ambiguity is reported, never guessed (§ 3.2): a bare id matching packages from
more than one registry returns the candidates, not a choice.

**Done**, with the route shape worth recording. Package ids contain `/`, so the
obvious `/packages/{id}/{version}` is ambiguous under a greedy path parameter.
Rather than invent an HTTP-specific identifier, the routes speak the format's
own `<registry>:<id>@<version>` syntax and parse it — `:` and `@` are both legal
in a path segment, each route keeps a distinct prefix, and the API therefore
tests § 3.2's reference notation rather than working around it.

Ambiguity returns **409 with the candidates**, not 300: the request is
answerable once the caller says which registry they meant. Omitting a version
applies § 17.1's selectors, so a prerelease is never chosen implicitly.

Validation is delegated to `reference/`, installed into the service
environment rather than reimplemented. One validator means the service and the
CLI cannot disagree about what a valid package is — a service accepting
something the CLI rejects would be exactly the divergence this project exists to
prevent. Importing it is not changing it; `reference/` stays the
dependency-light conformance witness.

**Test-vs-production difference found and handled.** SQLite autoincrements
`INTEGER PRIMARY KEY` only, never `BIGINT`, so the SQLite-backed tests could not
insert a row. `BigInteger().with_variant(Integer, "sqlite")` keeps `BIGINT` on
PostgreSQL, where this actually runs, while letting the tests exercise the same
models and the same migration.

## Publish API

```task
id: CANP-WP-0006-T04
status: done
priority: high
state_hub_task_id: "929932b6-931b-5409-a648-b6814f07f0b4"
```

`POST /packages` accepting a package, validating it with the same rules as the
CLI, and refusing a re-publish of an existing `<id>@<version>` with different
content (§ 17). Strict packaging applies: reserved paths and manifest-referenced
files only (§ 2).

Namespace policy (§ 20.1) is enforced here rather than advised, because unlike a
filesystem registry a service *can* authenticate a publisher. Note what identity
mechanism it uses; if there is none yet, say so and refuse writes rather than
accepting anonymous publishes into a closed namespace.

**Done.** `POST /packages` takes the same `files` shape `GET /archives`
returns, so an archive round-trips into a publish without translation and a
mirror is a GET followed by a POST — verified by a test, not just asserted.

**The identity mechanism, stated plainly.** A single shared bearer token
proving the caller is *the operator of this service*. Not per-publisher
identity: every token holder is indistinguishable. With no token configured the
service is read-only, which is the right default rather than an inconvenience —
§ 20.1 asks a registry to refuse publication into a closed namespace it does not
consider the publisher to own, and an unauthenticated service considers nobody
to own anything.

Namespace claims live in `namespace_claims` (migration `0002`) and are enforced,
which a filesystem registry cannot do at all — but only as precisely as the
identity allows. A closed namespace is protected from anonymous callers; it
cannot be attributed among several publishers. Until per-publisher identity
exists, a claim's `owner` is documentation rather than an access decision, and
`auth.py` says so rather than letting the code imply more than it delivers.

**Handed forward:** per-publisher identity is the main thing between this and a
registry several people can publish to.

**Migration hygiene found while adding `0002`.** Autogenerate proposed an
`ALTER COLUMN TYPE` on `package_files.package_version_id`, because the foreign
key's type was left to inference and compared as a variant against a reflected
plain type. SQLite cannot alter a column type, so `0002` failed halfway,
leaving the table created and the revision unstamped — the partially-applied
state that is worst to debug later. Fixed at the cause: the column is typed
explicitly, and `0001` was corrected rather than patched over, which is
legitimate only because it has never run outside this repo's tests. `alembic
check` now reports no drift.

## HTTP registry in the reference CLI

```task
id: CANP-WP-0006-T05
status: done
priority: medium
state_hub_task_id: "8dacbb68-aaa2-5f91-942d-a0667f1f1fc7"
```

Teach `--registry` to accept an `https://` URL, so `publish` and `install` work
against the service. § 20 already allows an HTTP registry; this is the first
implementation of one, and the first real test of whether package semantics
survive a transport change unmodified — which is what INTENT principle 10
claims.

If they do not survive it, that is a finding about the format, not a bug to
paper over in the client.

**Done, and it produced exactly one finding.**

*What survived unaltered*, verified against the running service: identity and
its ambiguity rules (a bare id in two registries came back 409 over HTTP just as
it does locally), immutability of a published `<id>@<version>` (identical
content accepted, changed content refused, a version bump accepted), strict
packaging, validation, and the index. A package published and installed over
HTTP was **byte-identical** to its source — `diff -r` clean — and its
`canonical-fidelity` eval still passed after the round trip.

*What did not*: **a URL is not a registry.** A filesystem registry *is* one
registry and § 20.1 names it from its directory; an HTTP service *hosts
several* behind one base URL. So the address cannot name the registry, and it
has to be named separately — `--as` when publishing, a qualified reference when
installing. Recorded as § 20.4 rather than papered over in the client, because
the gap is in the specification's list of registry kinds, not in the CLI.

Implemented with `urllib` rather than a library, so `reference/` keeps PyYAML as
its only dependency. Registry responses are treated as untrusted input (§ 19):
`decode_files` refuses path traversal, with a test.

## Image and smoke contract

```task
id: CANP-WP-0006-T06
status: todo
priority: medium
state_hub_task_id: "3c008ebe-bf07-5cee-8678-7b1ed27283ef"
```

Dockerfile and a published image, plus the checks a `rapp` smoke contract will
assert: health ok, migration at head, private service only, image digest match.

Completing this produces the digest that `rapp-canned-prompts` needs to pin, at
which point that repo can be created against something real.