Files
wild-pc/docs/relationships.md
Paul Payne e438f45b54 feat: deployment-scoped requires + UI editors, gateway nav, services filter
Model:
- Drop `requires` from ProgramSpec; requires is now deployment-only.
  Collapse Requirement to {ref, bind} with kind defaulting to "deployment"
  (drop the unused `version`). System-package preconditions stay on the
  program as `system_dependencies`, synthesized into {kind: system} for the
  functional check. relations.py/deploy.py read only the deployment's requires.
- create.py seeds a stack's substrate dependency onto the deployment, not the
  program. Docs + tests updated.

UI:
- Add a `requires` editor (useRequires) on service, job, and static detail
  pages — edit service→service deps (ref + optional bind env var).
- Restore the gateway route table "Open" column: a kind-aware in-app link to
  each route's deployment.
- Reformat the reach control: name baked into each option
  (localhost:PORT (local) / <name>.<domain> (internal) / <name>.<pubdomain> (public)).
- Services page: sort statics in with systemd services by name, add a
  search + kind-filter bar mirroring the programs page.
2026-07-06 04:00:24 -07:00

102 lines
5.0 KiB
Markdown

# Relationships: requires, repos, and derived predicates
How castle models the relationships between **programs**, **deployments**, and
**repos** — and answers questions like *"is this functional?"*, *"is it fresh?"*,
*"is it deployed?"* — with the smallest possible amount of stored state.
## The governing principle
> **Predicates are always derived. Encode only what is not derivable.**
A *predicate* is a question we ask about a program or deployment: `functional?`,
`fresh?`, `deployed?`. None of these are ever stored — each is a **function** over
data castle already has (git, config, the registry). When a predicate can't be
answered from derived data, find the one missing datum and ask: is it about the
**thing** (a node property) or about a **relationship** (an edge property)? Encode
*only* that datum. Everything else stays computed.
This is the same instinct as `kind` (derived from `manager`) — we don't store what
we can compute, and a relationship that proves real and stable in the derived graph
is a candidate to *promote* into a first-class concept. Diagnostic → evidence →
abstraction, in that order.
## Entities
- **program** — the software catalog entry.
- **deployment** — a program realized on this node (`kind` derived from `manager`).
- **repo** — a git working copy. **Derived** from `git rev-parse --show-toplevel`
on each program's source; several programs sharing one toplevel is a *monorepo*.
Never stored.
## Encoded preconditions: `requires` + `system_dependencies`
Everything we were calling "substrate", "wiring", or "dependency" reduces to
preconditions ("A must have B to be functional"), encoded on the layer each belongs
to. The relationship model unifies them into one requirement set with a typed target;
the **kind fixes the meaning and the check**:
| kind | source (where encoded) | means | checked by |
|------|------------------------|-------|-----------|
| `deployment` | the **deployment**'s `requires` | another deployment must **exist** | registry / config |
| `system` | the **program**'s `system_dependencies` | the host package/binary must be **installed** | `which` / `dpkg` |
A **deployment** declares the deployments it depends on. `kind` defaults to
`deployment`, so an entry is just a `ref` (+ optional `bind`):
```yaml
# deployments/<kind>/astro.yaml
requires:
- ref: astro-guru
- ref: supabase
- { ref: litellm, bind: LITELLM_URL } # bind: project the target's URL into env
```
A **program**'s host-package preconditions stay on the program as
`system_dependencies` (a plain list of package names); the model synthesizes the
`{kind: system}` requirements from it for the `functional?` check. This split keeps
each precondition on its natural layer — a deployment-ref is node-level wiring
(belongs on the deployment), a host package is intrinsic to the software (belongs on
the program). There is no `requires` on the program, and no `kind: system` written
into a deployment's `requires`.
Only encode a `requires` edge that is **not derivable** and that **castle itself
must traverse** for an operation (status, bring-up order, group ops). Do **not**
duplicate what another layer already owns — systemd `Requires=`/`After=` for unit
ordering, uv/pnpm for build graphs. This is *castle's* slice, uncoupled from any
one package ecosystem.
### Env is derived *from* `requires`, never scraped *into* it
Reading dependencies out of env strings is unstable (formats vary; a static
frontend's API URL is baked into its bundle and invisible). The stable direction is
the reverse: from an encoded `{ref, bind}` deployment requirement castle **generates**
the wiring env — it knows the target's address (`<ref>.<domain>` / its port) and
projects it into the consumer's env, optionally under the var named by `bind`. Same
move as `${public_url}`, one step further. Dependency → env, never env → dependency.
## Derived predicates
Computed on demand from encoded `requires` + git + registry; nothing stored:
- **`functional?`** — every `requires` is satisfied (system installed, deployment
exists). The unmet ones *are* the node's status (`doctor`/`status`).
- **`fresh?`** — the program's repo is at latest and clean (git status).
- **`deployed?`** — the deployment is active in the registry.
New predicates are just new functions over the same preconditions — there is
nothing to "unify" in storage.
## What's encoded vs derived (the whole surface)
| datum | source | stored? |
|-------|--------|---------|
| repo / monorepo | git toplevel | derived |
| `fresh?` / `deployed?` / `functional?` | git / registry / requires | derived |
| env wiring for a dependency | the `requires` edge + target address | derived |
| fan-in ("widely depended-on") | count of requires | derived |
| a **non-derivable** requirement (frontend→backend, host package) | — | **encoded** (`requires`) |
| the env var to bind a dep's URL to, when non-conventional | — | **encoded** (`requires[].bind`) |
Every irreducible found so far is an **edge** (a relationship); no new **node**
property has been needed yet — a sign the encoded surface stays tiny.