docs: design for the inter-app communication system

Typed request/response between workspace apps over HTTP, behind a Transport
seam so gRPC stays additive rather than a rewrite. Contracts live in the
workspace's shared package and are imported by both sides, so types flow
through a normal import with no code generator.

Consumes the exportSubjectContext/importSubjectContext seam the permissions
system reserved, with one improvement on what that seam implied: the token
carries sub and tenant only, never roles. Every app shares the
PermissionStore, so the callee resolves roles itself - a stale or forged
roles claim becomes impossible by construction and there is no path to
injecting privileges through a claim. The token authenticates; it never
authorizes.

Records two properties that are easy to get wrong and expensive to discover:
/__wrnexus/rpc/* must be unreachable from the public internet, blocked at the
gateway AND verified at the app, or every permission check in the workspace
is bypassable; and only procedures explicitly marked idempotent may be
retried, because retrying a slow createInvoice is how a customer gets billed
twice.

Deliberately does not wrap pubsub or queue - they work, and an abstraction
over working code leaks and needs keeping in sync. They gain identity
propagation instead.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
2026-08-05 09:27:43 +05:30
co-authored by Claude Opus 5
parent 2ab7b1e762
commit 87d33a2ce1
@@ -0,0 +1,223 @@
# Inter-app communication design (`@wrnexus/rpc`)
Date: 2026-08-05
Status: approved, not yet implemented
Depends on: the permissions system (`@wrnexus/authz`), merged 2026-08-05
## Problem
Workspace apps are separate OS processes behind a Host-routed gateway — full isolation, own
database registry, own memory. They can currently reach each other in exactly two ways:
- `@wrnexus/helpers``appOrigin(name)` / `appUrl(name, path)`, string builders over the
`WRNEXUS_WORKSPACE_ORIGINS` environment map.
- `@wrnexus/pubsub` — fire-and-forget events, cross-process via the Redis, NATS or Kafka
drivers.
There is no request/response layer. An app that needs an answer from another hand-writes
`fetch(appUrl("billing", "/api/invoice"))`: no shared contract, no type safety, no timeout,
no retry policy, and no way for the callee to learn who the end user is. The permissions
system reserved a seam for that last part — `exportSubjectContext` / `importSubjectContext`
and left it unbuilt.
## Scope
Of the four cross-app call shapes, three already have implementations:
| Shape | Status |
| ---------------------- | ----------------------------------------------------------------- |
| Request/response | Missing. This is the gap. |
| Fire-and-forget events | `@wrnexus/pubsub` |
| Scheduled / deferred | `@wrnexus/queue``defineJob`, durable store, workflows, cron |
| Streaming | `@wrnexus/realtime` covers browser↔server; app↔app is not covered |
This design builds request/response and app-to-app streaming. It does **not** wrap `pubsub`
or `queue` in a new facade: they work, and an abstraction over working code leaks and needs
keeping in sync. They instead gain identity propagation, and all four shapes are documented
in one place so a developer knows what to reach for.
## Approach
Service contracts live in the workspace's shared package and are imported by both sides. The
callee `implement`s a contract; the caller gets a typed proxy client. Types flow through a
normal TypeScript import — no generator, no generated file to go stale, no build step. This
matches the idiom the framework already uses: `defineAuthz`, `defineJob`, `defineRbac`,
`defineWorkflow`.
Rejected alternatives:
- **Callee exports types, caller imports them.** No new abstraction, but the callee still
hand-writes and hand-mounts every route and nothing keeps the route path and the type in
agreement.
- **Generate a client from the callee's `app/api` routes.** Codegen over hand-written routes
is brittle, and the permissions work has already demonstrated that a generated file with no
consumer becomes dead weight.
## Module layout
```
@wrnexus/rpc
contract.ts defineService(), the procedure builder
server.ts implement(); routes discovered from app/services/*.ts
client.ts serviceClient(contract, { as: ctx }) -> typed proxy
transport.ts Transport interface, httpTransport(), inProcessTransport()
identity.ts exportSubjectContext / importSubjectContext
errors.ts ServiceError and retryability classification
```
## Contracts
```ts
// packages/shared/src/services/billing.ts — imported by both apps
import { defineService, procedure } from "@wrnexus/rpc";
import { v } from "@wrnexus/validation";
export const billing = defineService({
name: "billing",
procedures: {
createInvoice: procedure
.input(v.object({ userId: v.string(), amountCents: v.number().integer() }))
.output<{ invoiceId: string }>()
.permission("invoice:create"),
},
});
```
Input validation reuses `@wrnexus/validation`'s `v`, which already returns
`{ ok, value, errors }`. No new validator is introduced.
`.permission(id)` is **enforced**, not merely declarative: `implement` checks it against the
callee's own catalog before invoking the handler, and a denial never reaches the handler. A
handler may additionally perform finer-grained, resource-level checks — `.permission()` is
the coarse gate, not the whole authorization story. Declaring it on the contract also means
the caller can see, at the type level, what the call requires.
## Server side
```ts
// app/services/billing.ts
export default implement(billing, {
async createInvoice({ userId, amountCents }, ctx) {
/* ... */
},
});
```
`buildRouter` discovers `app/services/*.ts` exactly as it discovers `app/authz` and
`app/schemas`, skipping `*.gen.ts`. Procedures mount at
`/__wrnexus/rpc/<service>/<procedure>`.
## Client side
```ts
const client = serviceClient(billing, { as: ctx });
const { invoiceId } = await client.createInvoice({ userId, amountCents });
```
`{ as: ctx }` carries identity: it calls `exportSubjectContext(ctx)` to mint a short-lived
signed token.
## Identity
The token carries **`sub` and `tenant` only — never roles.** Every app shares the
`PermissionStore`, so the callee resolves roles itself. A stale or forged roles claim is
therefore impossible by construction, the token stays small, and there is no path to
injecting privileges through a claim.
Built on `@wrnexus/jwt`: `alg` pinned to HS256, signature verified before the payload is
parsed, `exp`/`nbf`/`aud`/`iss` enforced.
| Claim | Purpose |
| -------- | ------------------------------------------------------------------------------ |
| `iss` | calling app name |
| `aud` | target app name, so B cannot replay A's token to C |
| `sub` | subject id — a non-empty string, matching the permissions system's requirement |
| `tenant` | tenant id, absent for global |
| `exp` | 60 seconds |
**The token authenticates; it never authorizes.** The callee runs its own permission checks
against its own catalog. An app cannot grant access the callee's catalog does not allow.
### Trust boundary
With a workspace-wide signing secret, any app can mint a token claiming to be any user. All
workspace apps are therefore one trust boundary: compromising the lowest-privilege app
compromises user identity across all of them. This is the standard trade-off and is
acceptable for apps in one repository under one operator, but it must be documented rather
than discovered. Per-app signing keys are the hardening path if apps ever run at differing
trust levels.
The RPC secret MUST NOT be the session secret. Reusing it would make a leaked RPC token a
session-forgery primitive.
### The RPC path is a new attack surface
`/__wrnexus/rpc/*` must be unreachable from the public internet — blocked at the gateway
**and** verified at the app, because either alone is a single point of failure. If that
endpoint is reachable with a forged token, every permission check in the workspace is
bypassable. This is the single most important property in the design.
## Transport
```ts
interface Transport {
call(
target: { app: string; service: string; procedure: string },
payload: unknown,
options: { signal: AbortSignal; identity?: string },
): Promise<ServiceResult>;
}
```
`httpTransport()` is the default and the only production transport in v1. A gRPC transport,
if ever needed, implements this interface and nothing else changes.
`inProcessTransport()` implements the same interface by calling the implementation directly
with no network. It exists for tests.
## Errors and retries
Errors classify into three kinds, because the classification decides whether a retry is safe:
| Kind | Retryable |
| --------------------------------------------------------------------- | --------- |
| Transport failure — connection refused, timeout, 5xx | Yes |
| Application error — the procedure threw or returned a failure | No |
| Rejected — validation failed, authorization denied, unknown procedure | No |
Only procedures explicitly marked `.idempotent()` are retried at all. Retrying a
non-idempotent `createInvoice` because a response was slow is how a customer gets billed
twice: the default is no retry, and opting in is a deliberate act recorded in the contract.
`ServiceError` crossing an app boundary is **opaque by default** — code and message only, no
stack, no internal detail. An error that explains the callee's internals is an information
leak, the same defect class as `authorizeDecision`'s original 403 body.
## Testing
- **In-process transport.** Contract, identity round-trip and error classification are all
testable without standing up two servers.
- **Contract conformance suite.** An implementation must satisfy its contract at runtime, not
only at compile time.
- **Security regressions**, each proven by reverting the guard and confirming the test fails:
the RPC path rejects an external request; a forged signature is rejected; a token with the
wrong `aud` is rejected; an expired token is rejected; a missing token is rejected.
## Build order
1. Contract, server discovery, typed client, HTTP transport, identity — the usable core.
2. Failure handling: timeouts, idempotent retry, circuit breaking.
3. App-to-app streaming.
4. Identity propagation for `pubsub` and `queue`, and the unified "how apps talk" guide.
Phase 1 is what makes the system usable; everything after is additive.
## Out of scope
- gRPC and Protocol Buffers. The `Transport` seam exists so this stays additive. gRPC's value
is cross-language interop and streaming; workspace apps are all TypeScript in one
repository sharing types through workspace packages, so a typed HTTP transport delivers
end-to-end type safety without a code generator.
- Wrapping `pubsub` or `queue` behind a unified facade.
- Service discovery beyond the existing `WRNEXUS_WORKSPACE_ORIGINS` map.
- Cross-workspace federation.