Files
TREK/shared
Maurice e8b071aecd fix(tours): restoring a recorded track and saving a copy wait for the tour's queued edits
Both go straight to the server. While an edit of the same tour still waited in
the queue, the restore went out first and its answer replaced the cached edit,
and the edit replayed afterwards planned the track over again. A copy copied
the server's older line instead of the one on screen. Both now wait the way the
actual track writes do: offline, behind a queued write of the tour, or for a
tour the server has not got yet they refuse with OnlineOnlyError, and the
planner says a connection is needed and the waiting changes go first, in all
27 languages.
2026-10-24 02:18:08 +00:00
..
2026-06-16 22:22:45 +02:00
2026-06-16 22:22:45 +02:00
2026-08-27 13:22:39 +02:00

@trek/shared

Single source of truth for TREK's API contracts, expressed as Zod schemas and consumed by both the server (request validation + inferred DTO types) and the client (typed requests/responses).

This package is part of the incremental NestJS + React 19 migration (see the "Brownfield Rewrite" board). It is intentionally dormant until modules start importing it — adding it changes nothing for users.

Rules

  • One folder per domain: src/<domain>/<domain>.schema.ts (+ .spec.ts).
  • Domain-agnostic building blocks live in src/common/.
  • A route is only considered migrated once its contract lives here.
  • Schemas are the source of truth; server DTOs and client types are inferred from them (z.infer<typeof schema>), never hand-duplicated.

Consumption (dev)

Both apps resolve @trek/shared to this package's TypeScript source:

  • Server (tsx): via paths in server/tsconfig.json.
  • Client (vite): via resolve.alias in client/vite.config.ts (+ paths for the type-checker).

Production packaging (Docker / workspace wiring) is introduced in card F2, when the server first depends on this package at runtime. Until then prod builds are untouched.

Not yet here

The canonical error envelope is finalised in card F5 (it must match TREK's current Express error responses byte-for-byte), so it is deliberately not invented in F1.