@gemmein/sdk 0.4.5 → 0.4.7

package/REFERENCE.md

@@ -79,6 +79,8 @@ deleted, subscription row removed. Irreversible: put a real confirm in front.
 `name` must be **lowercase letters, numbers, and underscores** (`saved_games`,
 never `savedGames` — a bad name throws synchronously). Collections are created
 by the app owner in the dashboard, never by the SDK.
+Names JavaScript owns (`constructor`, `toString`, `__proto__`…) are refused as
+collection and field names — a plain object inherits them.
 
 `options.intent` — one sentence: what the collection is for and who should
 access it. It rides every call as a hint; against a **local `gemmein dev`
@@ -189,13 +191,20 @@ type ListOptions = {
 ```
 
 **Options notes.** `key` = create-if-absent (a second writer gets 409
-`conflict`; your own retry returns the record with `existing: true`). `for` =
+`conflict`; your own retry returns the record with `existing: true`). The
+retry doubles as the lookup: there is no separate get-by-key call — re-issue
+the same `create()` with the same key to fetch your own record (it counts
+as a write; a re-fetch idiom, not a read path). `for` =
 recipient id on `addressed`/`direct`. `published: false` = draft on a public
 rule. `ifVersion` = optimistic concurrency. Atomic counters go in *value*
 position: `update(id, { stock: { decrement: 1, floor: 0 } })`. `expand` throws
 on `private`/`public_read`/`admin_write` (no link shape) — join in memory there.
 `since` fixes the order (oldest change first) — pairing it with `sort` is a 400
-`invalid_since`. Deltas are at-least-once: apply by id.
+`invalid_since`. Deltas are at-least-once: apply by id. A field added to a
+sealed shape by a promote run may carry a **default**: records that never wrote
+it read that value, `expand` targets and `where` see it too, and a written value
+wins (writing `null` clears it back to the default) — `search` still matches
+stored text only.
 
 ### Live data — `watch()`
 
@@ -256,80 +265,53 @@ honestly, not hidden.)
 ## Reaffirm — prove your app's boundaries in CI
 
 Gemmein enforces the rules **server-side**, so your frontend is never the source
-of truth. That's exactly why you can move fast: a `reaffirm` script *exercises*
-your boundaries with live calls and fails the build on drift — so if a rule ever
-stopped matching what your UI assumes, you catch it on deploy, not in front of a
-user. You reaffirm **because** Gemmein enforces — never because these checks are
-the enforcement.
-
-Two tiers:
-
-**Tier A — anonymous + shape (zero setup, runs against any environment incl. live).**
-No session needed. Catches the loudest drift: a private collection an anon can
-read, a renamed/mis-cased collection (`collection("savedGames")` throws
-*synchronously*), the error-code contract, and the deliberate reminder that a
-`community`/`public_read` collection is readable by anyone.
-
-**Tier B — cross-user isolation (dev environments only).**
-Proving "user B genuinely can't read user A's private record" needs two real
-sessions. Mint them without an OTP round-trip:
+of truth. Because enforcement is server-side, a script of live calls can verify
+it: `reaffirm.mjs` ships inside this npm package — copy it next to your app,
+fill the CONFIG block at its top, run it on every deploy. You reaffirm
+**because** Gemmein enforces — never because these checks are the enforcement.
+
+Exit codes: `0` all proven · `1` boundary drift · `2` could not complete (a
+config or connectivity failure, named as such — never a boundary verdict).
+
+The CONFIG block: `PRIVATE_COLLECTION` (required) · `PUBLIC_COLLECTION`,
+`DIRECT_COLLECTION`, `COMMUNITY_COLLECTION`, `GATED_COLLECTION` (each `""` to
+skip — every skip prints its reason, so the CI log always says what was proven
+and what was not) · `PROBE_FIELD`/`TEXT_FIELD` (your shapes' own field names —
+dev shapes learn from writes, so the probes speak your app's shape) ·
+`TEST_USERS` (three dev test emails).
+
+**Tier A — anonymous, read/refusal only, safe against any environment
+including live:** an anonymous caller is refused reading and writing the
+private collection, and the public collection's exposure is stated with a
+record count. A format-invalid collection name throws at the `collection()`
+line; a wrong-but-well-formed name surfaces as `unknown_collection` and exits
+`2` with "fix the CONFIG block" — a typo is never reported as drift.
+
+**Tier B — dev environments only**, sessions minted without a sign-in code:
 
 ### `gemmeinServer(sk).testSession(email) → { token, expiresAt, user }`
-Mints a **member** session for a test email. **Dev only** — throws
-`test_session_forbidden_live` on an `sk_live` key, and the server refuses it on a
-live environment too (the invariant that keeps it off real user data). Pass the
-returned `token` to `gemmein(pk, { tokenStore })` to act as that user. Dev and
-live enforce the *same* rules, so isolation proven in dev holds in live.
-
-This exact harness ships as **`reaffirm.mjs` inside the npm package** — copy it
-out, edit the CONFIG block, run in CI.
-
-```js
-// reaffirm.mjs — run in CI: `node reaffirm.mjs` (exits non-zero on any drift).
-import { gemmein, gemmeinServer } from "@gemmein/sdk";
-
-const API = process.env.GEMMEIN_API_URL;                 // your dev API url
-const g   = gemmein(process.env.PUBLIC_KEY, { apiUrl: API });
-const srv = gemmeinServer(process.env.SECRET_KEY, { apiUrl: API }); // sk_dev only
-let fail = 0;
-const refuse = async (label, code, fn) => {              // must throw `code`
-  try { await fn(); console.error("✗", label, "— expected", code, "got success"); fail++; }
-  catch (e) { e.code === code ? console.log("✓", label)
-            : (console.error("✗", label, "— got", e.code), fail++); }
-};
-const asUser = (token) => gemmein(process.env.PUBLIC_KEY, {
-  apiUrl: API,
-  tokenStore: { get: async () => token, set: async () => {}, clear: async () => {} },
-});
-
-// ── TIER A: functional + anonymous (no login) ──
-g.collection("private_notes");                   // misnamed → throws HERE, loudly, in CI
-await refuse("anon can't read private",    "denied", () => g.collection("private_notes").list());
-await refuse("anon can't write community", "denied", () => g.collection("board").create({ text: "x" }));
-
-// ── TIER B: cross-user isolation (dev only) ──
-const alice = await srv.testSession("alice@test.dev");
-const bob   = await srv.testSession("bob@test.dev");
-const A = asUser(alice.token), B = asUser(bob.token);
-
-const note = await A.collection("private_notes").create({ text: "alice-secret" });
-await refuse("B can't read A's private note", "not_found", () =>
-  B.collection("private_notes").get(note.id));
-const bobSees = await B.collection("private_notes").list();
-if (bobSees.records.length !== 0) { console.error("✗ B sees A's private records"); fail++; }
-else console.log("✓ B's private list is isolated");
-
-// shape the UI reads (userId, NOT id):
-const who = await A.auth.currentUser();
-if (!who.userId) { console.error("✗ currentUser().userId missing"); fail++; }
-else console.log("✓ currentUser().userId present");
-
-process.exit(fail ? 1 : 0);
-```
-
-Add a probe every time you add a feature. Point the harness at your **dev**
-environment (Tier B needs it); the Tier-A block alone can additionally smoke-test
-live, since it never mints a session.
+**Dev only** — throws `test_session_forbidden_live` on an `sk_live` key, and
+the server refuses it on a live environment too. Pass the `token` to
+`gemmein(pk, { tokenStore })` to act as that user. Dev and live enforce the
+*same* rules, so what is proven in dev holds in live.
+
+What Tier B proves, per probe: B can't read A's private record and B's list
+excludes it · `currentUser().userId` and `.data` shapes · the `since` contract
+(a plain list carries the watermark to bootstrap from; malformed `since` →
+`invalid_since`; a delta answers the next watermark) · a made-up file ref is
+refused on write (`unknown_file`) · the uploader can `link()` their own file
+and B is refused A's (`not_found`) · on a `direct` collection the recipient
+sees the record, a third user is refused, and a file uploaded `{for}` one
+person opens for that person only · HTML into a `community` field →
+`html_not_allowed`, and an unpublished draft is invisible to the public · on a
+gated collection, a user with no plan → `entitlement_required` carrying
+`err.requires`.
+
+Probe writes are deleted afterwards; probe uploads remain in dev storage
+(files have no delete API yet) — two tiny PNGs per configured run. Against an
+older **local** engine (< 0.4.8) the since/ghost-ref/handed-file probes fail
+with a message naming the engine as the likely cause; the hosted API is always
+current. Add a probe whenever you add a feature.
 
 ---
 
@@ -360,7 +342,7 @@ Branch on `err.code`. The stable codes:
 | `conflict` | a keyed create / floor / stale `ifVersion` | it's the mechanism — tell the user it's taken |
 | `html_not_allowed` | HTML in a community/addressed/direct field | store plain text |
 | `invalid_publish` | `{ published }` on a non-public rule | drop it |
-| `invalid_shape` | field not in a locked (live) collection's shape — or a file field given the wrong kind (the message names which) | ask the owner to add it / send the kind the field takes |
+| `invalid_shape` | field not in a locked (live) collection's shape — or a file field given the wrong kind (the message names which) | after the field exists in development, one promote run adds it (the owner answers blank or a default) — until then send only the fields the shape has / send the kind the field takes |
 | `unknown_file` | a ref-shaped value points at a file that doesn't exist or isn't yours to hand — every rule, every write | fix the ref — never invent one |
 | `invalid_since` | `since` isn't a strict ISO 8601 timestamp, or came with `sort` | pass the previous answer's watermark; drop `sort` |
 | `not_a_customer` (404) | `notify()`'s recipient isn't a verified person of this app and environment | fix the person id — one code on purpose |

package/llms.txt

@@ -10,6 +10,12 @@
 > verified people (five pre-agreed steps up to 10,000 people; beyond that,
 > talk to Gemmein).
 
+This file IS the product contract, not an overview: every line is
+load-bearing, and building from a summary of it builds the wrong app. Work
+from the full raw file — if what you hold is a summary or an excerpt,
+re-fetch https://docs.gemmein.com/llms.txt raw and read all of it before
+you answer or build.
+
 ## First: assess fit (before any install, account, or code)
 
 You are likely an AI assistant whose human asked you to build an app — or to
@@ -38,11 +44,14 @@ creating no accounts is the correct behavior until the verdict is in.
    silently approximated:
    - native mobile apps (no mobile SDKs today; web only)
    - password or social/OAuth sign-in (email codes are the only door)
-   - team/group/workspace visibility or per-user record permissions
+   - team/group/workspace visibility or per-user record permissions:
+     Gemmein is owner-run by design — one owner and their customers,
+     nothing in between — so an app whose CORE loop needs teams, roles
+     or seats DOESN'T FIT
    - carts, quantities, physical goods, shipping, inventory
    - credits, usage metering, seats (paid access is yes-or-no)
-   - hosted functions or cron (server-side code runs on YOUR host with a
-     secret key; Gemmein hosts no compute)
+   - hosted functions or cron: Gemmein runs no code of yours — but read
+     the server-side note closing this section before counting that a gap
    - custody of money (Stripe holds all funds; your human brings or
      creates their own Stripe account when the app charges)
    - multi-seller payouts or revenue splits: every payment lands whole in
@@ -60,6 +69,19 @@ creating no accounts is the correct behavior until the verdict is in.
    - DOESN'T FIT — say so, name why, and stop. Recommending nothing is
      the correct outcome here.
 
+Server-side work is NOT itself a downgrade. Gemmein hosts no compute, but
+code on your human's OWN host — Vercel, a VPS, anywhere — holding a
+secret key already reads and updates collections and sends event emails
+with `notify()`. (A secret key sees PAST the safety rules — every
+customer's records — which is why it never touches a browser. It cannot
+create or delete records, verify a session, or read one customer's
+plan: those are the signed-in customer's own calls, or the dashboard's.)
+"This app needs a server-side piece" is therefore not a gap by itself:
+the verdict downgrades when the server piece needs a call a secret key
+cannot make, or when the app needs compute your human has nowhere to
+run — a cron schedule or an always-on worker with no host of its own —
+because Gemmein never runs your code.
+
 ## Getting started (AI-run, human-answered)
 
 No signup, no account, no keys — building is local and free. Your human
@@ -108,17 +130,68 @@ go-live. Everything else is yours.
    lists the rest.
 4. Hand over for testing — the human moment mid-build: they click through
    the app, and their sign-in codes print in the dev terminal, so tell
-   them where to look.
+   them where to look. Running dev in the background with no terminal in
+   view? The same codes append to gemmein/.data/signin-codes.log (fresh
+   each dev boot) — read the latest line instead of scraping stdout.
 5. Done building? `npx gemmein check` reads the project and says what's
    ready and what go-live still needs. Then `npx gemmein sync` — THIS is
    the moment your human signs up (free) at app.gemmein.com and pastes two
    dev keys — and `npx gemmein go-live` walks the rest: plans, Payment
    Links, the live flip. A card enters at go-live, never before.
+6. After go-live, the DATA MODEL reaches live only by PROMOTION from
+   development: new collections built in dev promote (`npx gemmein
+   go-live` again, or the dashboard's Go-live page), and so do new FIELDS
+   for a live collection whose shape sealed at go-live (see Shapes). Ship
+   in that order: promote first, then deploy the frontend that depends on
+   the new collection or field — a frontend shipped ahead of its
+   promotion reads blanks or refusals from live. Each added field is one
+   question for YOUR HUMAN, asked by the promote run: existing records
+   don't have it, so what should they show? Blank is a fine answer, or
+   they give a default those records will show. What is already
+   sealed never moves — a live field never changes type or name and never
+   leaves — and rules still never change through promotion. Plans,
+   Payment Links and domains stay live-editable in the dashboard, and a
+   field's default joins them there (changeable or cleared later, with no
+   promotion needed).
+
+Two runtimes, one dashboard — keep your human oriented on where things
+live, or the dashboard will look broken to them. The local runtime
+(`gemmein dev`, `pk_local_` keys) is a private rehearsal room: every
+record, person and simulated payment stays in gemmein/.data on this
+machine and never appears in any dashboard. The dashboard at
+app.gemmein.com is the back office of the CLOUD app only — there is no
+local dashboard and nothing to log into locally. `npx gemmein sync`
+carries STRUCTURE to the cloud app's development environment —
+collections, rules, plan locks — never records or people; after a sync
+the dashboard shows the shape of the app with no data in it, and that is
+correct, not a bug. Data reaches the dashboard only when the app runs on
+cloud keys: the dev pair from app.gemmein.com (`pk_test_` in the
+browser, `sk_dev_` on a server) drives the real API's development
+environment, and `pk_live_` drives production after go-live. The key
+prefix always names the rail — local, test, live. The dev rail is where
+a REAL rehearsal happens before any card enters: real emailed sign-in
+codes, real data in the dashboard — and it can face the public. Put the
+app on a real https domain and list that domain on the dashboard's
+Domains page: before go-live a listed domain works WITHOUT verification
+(http://localhost and http://127.0.0.1 also work unlisted until go-live
+drops them; any other raw IP never works — the Domains page accepts
+names only). Go-live is when the keys lock to proven domains. The
+development environment carries its own free monthly allowance (10,000
+API requests, 500 writes, 500 sign-in events, 1 GiB of files); crossing
+a counter never stops the app and never turns into a bill — the
+dashboard's usage page shows what crossed, and storage is the one thing
+that can fill (see Pricing). Tell your human before they first open the
+dashboard: local building is invisible there on purpose — nothing
+leaves this machine until sync, and sync moves the blueprint, not the
+contents.
 
 ## What it is
 
 - Audience: people who build web apps with AI tools (Cursor, Bolt, Lovable, Claude).
 - Scope: web apps (TypeScript/JavaScript SDK). Mobile SDKs are not offered today.
+- Shape: built for owner-run businesses — one owner and their customers,
+  nothing in between. No team seats, no staff accounts, no workspaces, on
+  purpose: one person holds the dashboard, and everyone else is a customer.
 - The platform holds identity, data, and subscription state. It never holds,
   moves, or processes money — Stripe handles all payments end-to-end; Gemmein
   only receives Stripe's webhooks and records who is on which plan. App
@@ -143,7 +216,7 @@ go-live. Everything else is yours.
   those local declarations. A declaration is `gemmein/collections/<name>.json`:
   `{ "rule": "shared", "means": "<why, in your human's words>" }`, plus an
   optional `"unlockedBy": ["pro"]` — plan or product NAMES from
-  gemmein/payments.json (never access keys) — which is the console's
+  gemmein/payments.json (never access keys) — which is the dashboard's
   "Unlocked by": the local engine refuses members without one of them
   (403 entitlement_required) the moment the file lands, and sync carries
   it to the cloud app. An unknown name is refused out loud in the dev
@@ -315,7 +388,7 @@ go-live. Everything else is yours.
   retry. A 403 from `link()` means the customer isn't allowed this
   file right now — signed out, not theirs, or an entitlement they no longer
   hold (`entitlement_required` — `err.requires` is the plan's key, the one
-  the console shows by name).
+  the dashboard shows by name).
   Honest bound: revoking access stops NEW links immediately; a link already
   issued works until it expires. Gemmein controls delivery, it can't take
   back a file someone already downloaded. A file can name ONE other reader:
@@ -339,11 +412,22 @@ go-live. Everything else is yours.
   If a stored ref's file was later deleted, re-sending it refuses the same
   way: clear the field (null) or upload a fresh file. Once sealed, class
   is law too: a ZIP into a profile photo field is refused with the teach;
-  before the seal, what you upload is what the field learns. A 400 invalid_shape means the field isn't in the locked
-  shape or the value doesn't fit its kind. A live shape is sealed and
-  cannot take new fields — stop, tell your human which field you needed,
-  and send only the fields the shape already has. Never rename fields to
-  dodge it.
+  before the seal, what you upload is what the field learns. A 400
+  invalid_shape means the field isn't in the locked shape or the value
+  doesn't fit its kind. A sealed shape CAN still gain a new field, but
+  only through a promote run: build it in development, tell your human
+  which field you needed, and until that promotion runs send only the
+  fields the shape already has. Never rename fields to dodge it. What is
+  sealed never moves — a field never changes type or name and never
+  leaves. A field promoted with a DEFAULT is served to the records that
+  never wrote it: get, list, expand targets and `where` filters all see
+  the default (an atomic increment counts from it too), and a written
+  value wins; writing `null` clears it back to the default. Only text,
+  number and yes/no fields can carry a default; list, link, json and
+  file fields join blank. Honest bound: full-text `search` scans stored
+  text only, so a record showing a default is not found by searching for
+  it. New collections built in dev reach live through the same PROMOTION,
+  sealing on the way.
 - Contention (bookings, slugs, stock, shared edits): when two users can race
   for the same thing, a permission model can't save you — preconditions do,
   and they're just arguments on calls you already make. A 409 `conflict` from
@@ -352,14 +436,19 @@ go-live. Everything else is yours.
   - Uniqueness: `create(data, { key: "slot:2026-07-15T15:00" })` — derive the
     key from the thing that must be unique; the second writer gets 409, your
     own retry gets your existing record back (`existing: true`), deleting
-    frees the key. Keys are unique across the WHOLE collection — every
-    writer, every recipient, under every safety rule (live records only) —
-    so a claim is race-proof even in `direct`/`addressed`. Never
+    frees the key. That retry IS the lookup — there is no separate
+    get-by-key call: to fetch your own record by key later, issue the
+    same `create()` with the same key and read `existing: true` off the
+    result. It counts as a write (rate and allowance), so use it to
+    re-fetch a claim — never on a render path. Keys are unique across
+    the WHOLE collection — every writer, every recipient, under every
+    safety rule (live records only) — so a claim is race-proof even in
+    `direct`/`addressed`. Never
     find-then-create — that races. Keys are 1-120 chars of letters,
     numbers, and `: _ . @ / -` only. A claimed key holds until its record
     is deleted: if a claim must be PAID to stick (book then pay), expiring
-    unpaid claims is your app's job — the owner deletes them from their
-    dashboard, or your own server does with a secret key; there is no cron.
+    unpaid claims is the owner's job — they delete them from their
+    dashboard (a secret key cannot delete records); there is no cron.
   - Limited stock (N units anyone can buy): claim units with keyed creates —
     try `create({...}, { key: "unit:item42:1" })`, on conflict try `:2` … `:N`;
     all taken = sold out. Race-proof under every safety rule.
@@ -404,7 +493,7 @@ go-live. Everything else is yours.
   checks that the lock holds and names collections left open while the app
   sells something) — and the
   engine refuses customers without it, under all seven rules. (Server
-  secret keys and the owner's console are exempt by design; link/expand
+  secret keys and the owner's dashboard are exempt by design; link/expand
   silently hide gated records rather than naming them.) Every plan and
   product carries its own key, `access:<slug of its name>` (plan "pro" →
   `access:pro`); the paid webhook grants that key, and a FULL refund or a
@@ -413,7 +502,8 @@ go-live. Everything else is yours.
   Owners also grant and revoke by hand (trials, comps, support). Effective
   access is the UNION of a customer's live grants. A signed-in customer
   without access gets `403 entitlement_required`: `err.requires` is the
-  plan's key (the console shows the plan by name; the key is `access:<slug>`).
+  plan's key (the dashboard shows the plan by name; the key is
+  `access:<slug>`).
   A collection unlocked by several plans (OR) still names ONE key — offer
   your plans by name through checkout, never a key list. Show your upgrade
   screen and send them to checkout; retry only after they hold one. Proof surfaces: `await g.purchases.mine()`
@@ -471,7 +561,9 @@ go-live. Everything else is yours.
   sign in first). Render `err.message`; it reads correctly in every case.
   Branch only on the specifically-named codes (unknown_collection,
   unknown_product, invalid_shape, html_not_allowed, invalid_publish,
-  conflict, …) plus the forbidden-means-stop rule.
+  conflict, …) plus the forbidden-means-stop rule. If something is stuck
+  in a way this file doesn't explain, `npx gemmein feedback` reaches the
+  builders — it works from anywhere, with no project and no account.
 - Draft state reads back as a TOP-LEVEL boolean `record.published` (next to
   id/updatedAt), not under `.data`. Non-authors only ever receive published
   records, so you see `false` only on your own drafts (or on everything, as
@@ -485,10 +577,27 @@ go-live. Everything else is yours.
   `await g.account.delete()` — for every app it APPLIES to (GDPR right to
   erasure; Apple 5.1.1(v) for apps with account creation). It's the full
   server-side cascade (sessions, records, files, subscription row) and
-  irreversible — put a real confirm in front of it. Suspensions, bans, and
-  owner-side erasure are dashboard actions, not SDK calls.
+  irreversible — put a real confirm in front of it. Suspensions, forced
+  sign-out everywhere, and owner-side erasure are dashboard actions, not
+  SDK calls.
+- The owner's dashboard already handles these — do not build them into the
+  app: creating collections and choosing their safety rule; naming plans
+  and products and pasting their Stripe Payment Links; the "Unlocked by"
+  lock on a collection; suspending, signing out and erasing customers;
+  moderation and status flips on other users' records; scoping a secret
+  key to collections and actions; verifying the sending and receiving
+  domain; usage and logs; the Inbox, where every notify() send is a
+  conversation and customer replies land; holding the billing band. None
+  of them has an SDK equivalent, so an owner screen you build for one is
+  a page that already exists, without the enforcement.
 - Secret keys (`sk_`) must never appear in browser code; app keys (`pk_`)
-  are public and domain-locked.
+  are public and domain-locked. A secret key can be SCOPED in the
+  dashboard's Keys room to particular collections, and to read-only or
+  read-and-update; a call outside that scope answers 403 `scope_denied`.
+  Read that error as deliberate — the human narrowed the key on purpose.
+  Name the collection and access your code needs and ask them to mint a
+  replacement key that includes it — scopes are fixed when a key is
+  created; never reach for a broader key than the job needs.
 
 ## Return values & shapes (get these exactly right)
 
@@ -500,6 +609,11 @@ rule. The specifics:
   synchronously from `g.collection(name)`; if you call that at module
   load, it can blank your whole app with no browser-console error. Name them
   right.
+  One more refusal, for collection AND field names alike: anything
+  JavaScript itself owns — `constructor`, `toString`, `__proto__`,
+  `hasOwnProperty` and their kin — is refused the same way a reserved
+  server-managed field is. A plain object inherits those names, so a
+  seal must never carry one.
 - The signed-in user: `await g.auth.currentUser()` →
   `{ authenticated: true, userId, email }` or `{ authenticated: false }`. The
   id field is **`userId`, not `id`** — `user.id` is `undefined`, and feeding
@@ -522,29 +636,34 @@ rule. The specifics:
   `{ url, ... }`. Just `await` them on the click — don't also redirect to the
   returned `url` (you'll double-navigate), and don't build the URL yourself.
 
-## Reaffirm your app (don't trust your frontend — prove it)
+## Reaffirm your app (the server enforces — prove it in CI)
 
 Gemmein enforces the rules on the server, so your UI is never the source of
-truth. That's what lets you move fast: ship a small `reaffirm` script that
-*exercises* your boundaries with live calls and run it in CI. When a rule ever
-stops matching what your screens assume, this catches it on deploy — not in
-front of a user.
-
-- **Anywhere, no login:** anonymous reads/writes that must be refused (an anon
-  reading a `private` collection → `denied`; an anon writing `community` →
-  `denied`), plus the shapes your UI reads (`currentUser().userId`,
-  `record.data.*`) and that your collection names don't throw.
-- **Isolation (dev environment):** to prove one user can't see another's
-  private records, sign in two test users without an OTP round-trip —
-  `gemmeinServer(sk_dev).testSession(email)` returns a session `token` you pass
-  to `gemmein(pk, { tokenStore })`. It works in a **dev environment only**
-  (`sk_live` throws `test_session_forbidden_live`), and dev and live enforce the
-  same rules, so isolation proven in dev holds in live.
-
-Add a probe whenever you add a feature. You reaffirm **because** Gemmein
-enforces — never because these checks are the enforcement. A ready-to-edit
-`reaffirm.mjs` ships inside the `@gemmein/sdk` npm package (next to this file and
-REFERENCE.md) — copy it out, name your collections, run it in CI.
+truth. Because enforcement is server-side, a script of live calls can verify
+it: a ready-to-edit `reaffirm.mjs` ships inside the `@gemmein/sdk` npm package
+(next to this file and REFERENCE.md). Copy it next to the app, fill the CONFIG
+block at its top (the private collection is required; direct/community/gated
+collections each unlock more probes, `""` skips with the reason printed; set
+PROBE_FIELD/TEXT_FIELD to the app's own field names), and run it on every
+deploy. Exit 0 = all proven, 1 = boundary drift, 2 = could not complete
+(config/connectivity — named as such, never reported as drift).
+
+- **Anywhere, no login (Tier A):** anonymous reads and writes of the private
+  collection must be refused (`denied`); the public collection's exposure is
+  stated out loud. Read-and-refusal only — safe against live.
+- **Isolation (Tier B, dev environment only):** sessions are minted without a
+  sign-in code via `gemmeinServer(sk_dev).testSession(email)` (`sk_live`
+  throws `test_session_forbidden_live`). It then proves: cross-user private
+  isolation; the `since` contract (bootstrap from a plain list's watermark;
+  junk → `invalid_since`); a made-up file ref is refused (`unknown_file`);
+  sealed file delivery (own file links, another user's is `not_found`);
+  `direct` recipient scoping and handed files (`upload(blob, {for})` opens for
+  the named person only); `community` plain text (`html_not_allowed`) and
+  invisible drafts; entitlements (`entitlement_required` + `err.requires`).
+
+Dev and live enforce the same rules, so what is proven in dev holds in live.
+Add a probe whenever you add a feature. You reaffirm BECAUSE Gemmein
+enforces — never because these checks are the enforcement.
 
 ## Pricing (current, v4 — one banded plan)
 

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@gemmein/sdk",
-  "version": "0.4.5",
-  "description": "Gemmein SDK \u2014 passwordless auth, safe storage, and Stripe-driven record flips for AI-built apps. Small enough that one prompt teaches the whole API.",
+  "version": "0.4.7",
+  "description": "Gemmein SDK — passwordless auth, safe storage, and Stripe-driven record flips for AI-built apps. Small enough that one prompt teaches the whole API.",
   "license": "MIT",
   "type": "module",
   "main": "./dist/index.cjs",

package/reaffirm.mjs

@@ -5,19 +5,39 @@
 //
 //   PUBLIC_KEY=pk_test_... SECRET_KEY=sk_dev_... node reaffirm.mjs
 //
-// Exits non-zero on any drift. Gemmein enforces the rules server-side — you
-// reaffirm BECAUSE the server enforces, never because these checks are the
-// enforcement. Tier A (anonymous) runs against any environment, live included.
-// Tier B (cross-user isolation) mints sessions via testSession, which works
-// ONLY in a development environment (sk_live is refused, by design) — dev and
-// live enforce the same rules, so isolation proven in dev holds in live.
+// Exits 0 all-proven · 1 boundary drift · 2 could-not-complete (config or
+// connectivity — NOT a boundary verdict). Gemmein enforces the rules
+// server-side — you reaffirm BECAUSE the server enforces, never because
+// these checks are the enforcement. Tier A (anonymous) runs against any
+// environment, live included. Tier B (cross-user isolation) mints sessions
+// via testSession, which works ONLY in a development environment (sk_live
+// is refused, by design) — dev and live enforce the same rules, so what is
+// proven in dev holds in live.
+//
+// Every probe past the first is optional: name a collection to prove its
+// surface, leave it "" and the probe is skipped WITH ITS REASON PRINTED —
+// your CI log always says what was proven and what was not.
+//
+// Probes write only to the collections you name, using YOUR field names
+// (dev shapes learn from writes — the probes must speak your app's shape).
+// Probe uploads stay in dev storage (files have no delete API yet): two
+// tiny PNGs per configured run, a known cost.
+//
+// The since/ghost-ref/handed-file probes need engine ≥ 0.4.8 when pointed
+// at a LOCAL runtime (GEMMEIN_API_URL) — an older engine fails them even
+// though your hosted app is fine. `npx -y gemmein@latest dev` updates.
 
 import { gemmein, gemmeinServer } from "@gemmein/sdk";
 
 // ── CONFIG — edit for your app ──────────────────────────────────────────────
-const PRIVATE_COLLECTION = "notes";   // a collection with the `private` rule
-const PUBLIC_COLLECTION  = "board";   // a `community` or `public_read` collection (or "" to skip)
-const TEST_USERS = ["reaffirm-a@test.dev", "reaffirm-b@test.dev"];
+const PRIVATE_COLLECTION   = "notes";  // a collection with the `private` rule (required)
+const PUBLIC_COLLECTION    = "";       // a `community`/`public_read` collection, e.g. "board" ("" to skip)
+const DIRECT_COLLECTION    = "";       // a `direct` collection — proves recipient scoping + handed files
+const COMMUNITY_COLLECTION = "";       // a `community` collection — proves plain-text + drafts
+const GATED_COLLECTION     = "";       // a collection unlocked by a paid plan — proves entitlements
+const PROBE_FIELD = "probe";           // a field YOUR private collection's shape allows
+const TEXT_FIELD  = "text";            // the text field YOUR direct/community shapes use
+const TEST_USERS = ["reaffirm-a@test.dev", "reaffirm-b@test.dev", "reaffirm-c@test.dev"];
 // ────────────────────────────────────────────────────────────────────────────
 
 const API = process.env.GEMMEIN_API_URL;               // omit for production api
@@ -30,46 +50,142 @@ const g = gemmein(PK, opts);
 let fail = 0;
 
 const refuse = async (label, code, fn) => {            // the call MUST throw `code`
-  try { await fn(); console.error("✗", label, "— expected", code, "but it succeeded"); fail++; }
-  catch (e) { e.code === code ? console.log("✓", label)
-            : (console.error("✗", label, "— expected", code, "got", e.code ?? e.message), fail++); }
+  try { await fn(); console.error("✗", label, "— expected", code, "but it succeeded"); fail++; return null; }
+  catch (e) {
+    if (e.code === code) { console.log("✓", label); return e; }
+    console.error("✗", label, "— expected", code, "got", e.code ?? "(no code)", "—", e.message); fail++; return e;
+  }
 };
-const check = (label, cond) => cond ? console.log("✓", label) : (console.error("✗", label), fail++);
+const check = (label, cond, got) => cond ? console.log("✓", label)
+  : (console.error("✗", label, got !== undefined ? `— got ${JSON.stringify(got)}` : ""), fail++);
+const skip = (what, why) => console.log(`· ${what} skipped — ${why}`);
 const asUser = (token) => gemmein(PK, { ...opts, tokenStore: {
   get: async () => token, set: async () => {}, clear: async () => {} } });
+// A tiny real PNG head: enough for the server's content check, no meaning.
+const PNG = new Blob([Uint8Array.from([0x89,0x50,0x4e,0x47,0x0d,0x0a,0x1a,0x0a,
+  0,0,0,0x0d,0x49,0x48,0x44,0x52, ...new Array(64).fill(0)])], { type: "image/png" });
+// A well-formed reference no upload ever returned — the write must refuse it.
+const GHOST_REF = "file:99999999-9999-4999-8999-999999999999";
 
-// ── TIER A — functional + anonymous. No login; safe against live. ───────────
-g.collection(PRIVATE_COLLECTION);              // a misnamed collection throws HERE, loudly
-await refuse("anon can't read the private collection", "denied",
-  () => g.collection(PRIVATE_COLLECTION).list());
-await refuse("anon can't write the private collection", "denied",
-  () => g.collection(PRIVATE_COLLECTION).create({ probe: "x" }));
-if (PUBLIC_COLLECTION) {
-  const open = await g.collection(PUBLIC_COLLECTION).list();
-  console.log(`ℹ "${PUBLIC_COLLECTION}" is public by rule — ${open.records.length} records visible to ANYONE. Never put secrets in it.`);
-}
+try {
+  // ── TIER A — functional + anonymous. No login; read/refusal only. ─────────
+  // A format-invalid name (uppercase, spaces) throws right here. A
+  // wrong-but-well-formed name surfaces below as `unknown_collection` —
+  // that is a CONFIG fix, not a boundary failure.
+  g.collection(PRIVATE_COLLECTION);
+  await refuse("anon can't read the private collection", "denied",
+    () => g.collection(PRIVATE_COLLECTION).list());
+  await refuse("anon can't write the private collection", "denied",
+    () => g.collection(PRIVATE_COLLECTION).create({ [PROBE_FIELD]: "x" }));
+  if (PUBLIC_COLLECTION) {
+    const open = await g.collection(PUBLIC_COLLECTION).list({ limit: 100 });
+    const shown = open.hasMore ? `${open.records.length}+` : `${open.records.length}`;
+    console.log(`ℹ "${PUBLIC_COLLECTION}" is public by rule — ${shown} records visible to ANYONE. Never put secrets in it.`);
+  } else skip("public-collection reminder", "set PUBLIC_COLLECTION");
+
+  // ── TIER B — cross-user isolation. Dev environments only. ─────────────────
+  if (SK && !SK.startsWith("sk_live")) {
+    const srv = gemmeinServer(SK, opts);
+    const [a, b] = await Promise.all(TEST_USERS.slice(0, 2).map((e) => srv.testSession(e)));
+    const A = asUser(a.token), B = asUser(b.token);
+
+    // The original core: private means private.
+    const note = await A.collection(PRIVATE_COLLECTION).create({ [PROBE_FIELD]: "a-secret" });
+    await refuse("B can't read A's private record", "not_found",
+      () => B.collection(PRIVATE_COLLECTION).get(note.id));
+    const bSees = await B.collection(PRIVATE_COLLECTION).list();
+    check("B's private list contains none of A's records",
+      !bSees.records.some((r) => r.id === note.id));
+
+    const who = await A.auth.currentUser();            // the shape your UI reads
+    check("currentUser() exposes userId (not id)", !!who.userId, who);
+    check("record fields live under .data", note.data?.[PROBE_FIELD] === "a-secret", note.data);
+
+    // Live data: bootstrap `since` from a plain list's watermark (that is
+    // the real pattern — never an ancient timestamp, which pages through
+    // history) and prove junk is refused.
+    const seed = await A.collection(PRIVATE_COLLECTION).list({ limit: 1 });
+    check("a plain list carries the watermark to start from", typeof seed.watermark === "string", seed.watermark);
+    await refuse("a malformed `since` is refused", "invalid_since",
+      () => A.collection(PRIVATE_COLLECTION).list({ since: "not-a-timestamp" }));
+    const delta = await A.collection(PRIVATE_COLLECTION).list({ since: seed.watermark });
+    check("a delta read returns the next watermark", typeof delta.watermark === "string", delta.watermark);
+
+    // The file law: a made-up reference never lands in a record. (If this
+    // SUCCEEDS against a local runtime, your engine predates 0.4.8 —
+    // update it; the hosted API always enforces this.)
+    try {
+      const polluted = await A.collection(PRIVATE_COLLECTION).create({ [PROBE_FIELD]: GHOST_REF });
+      console.error("✗ a made-up file ref is refused on write — it SUCCEEDED (old local engine? run `npx -y gemmein@latest dev`)"); fail++;
+      await A.collection(PRIVATE_COLLECTION).delete(polluted.id);       // never leave the ghost behind
+    } catch (e) {
+      e.code === "unknown_file" ? console.log("✓ a made-up file ref is refused on write")
+        : (console.error("✗ a made-up file ref is refused on write — expected unknown_file got", e.code ?? "(no code)", "—", e.message), fail++);
+    }
+
+    // Sealed delivery: your file is not their file.
+    try {
+      const up = await A.collection(PRIVATE_COLLECTION).upload(PNG, { name: "probe.png" });
+      const mine = await A.files.link(up.ref);
+      check("the uploader can link their own file", typeof mine.url === "string");
+      await refuse("B can't link A's file", "not_found", () => B.files.link(up.ref));
+    } catch (e) {
+      if (e.code === "unsupported_file_type") skip("file-delivery probes", `"${PRIVATE_COLLECTION}" doesn't accept PNG uploads — point the probes at a collection that takes images`);
+      else throw e;
+    }
+
+    await A.collection(PRIVATE_COLLECTION).delete(note.id);             // leave dev tidy
+
+    // direct: the recipient reads it; nobody else does. And a file handed
+    // to one person opens for that person only.
+    if (DIRECT_COLLECTION) {
+      const [cSess] = await Promise.all([srv.testSession(TEST_USERS[2])]);
+      const C = asUser(cSess.token);
+      const msg = await A.collection(DIRECT_COLLECTION).create({ [TEXT_FIELD]: "for b" }, { for: b.user.id });
+      const bBox = await B.collection(DIRECT_COLLECTION).list({ limit: 100 });
+      check("the recipient sees the direct record", bBox.records.some((r) => r.id === msg.id));
+      await refuse("a third user can't read it", "not_found",
+        () => C.collection(DIRECT_COLLECTION).get(msg.id));
+      try {
+        const handed = await A.collection(DIRECT_COLLECTION).upload(PNG, { name: "handed.png", for: b.user.id });
+        const bGets = await B.files.link(handed.ref);
+        check("the person a file was handed to can open it", typeof bGets.url === "string");
+        await refuse("anyone else is refused the handed file", "not_found",
+          () => C.files.link(handed.ref));
+      } catch (e) {
+        if (e.code === "unsupported_file_type") skip("handed-file probes", `"${DIRECT_COLLECTION}" doesn't accept PNG uploads`);
+        else throw e;
+      }
+      await A.collection(DIRECT_COLLECTION).delete(msg.id);
+    } else skip("direct-rule probes (recipient scoping, handed files)", "set DIRECT_COLLECTION");
+
+    // community: other people's screens get text, never markup — and
+    // drafts stay invisible until published.
+    if (COMMUNITY_COLLECTION) {
+      await refuse("HTML into a community field is refused", "html_not_allowed",
+        () => A.collection(COMMUNITY_COLLECTION).create({ [TEXT_FIELD]: "<b>hi</b>" }));
+      const draft = await A.collection(COMMUNITY_COLLECTION).create({ [TEXT_FIELD]: "draft probe" }, { published: false });
+      const anon = await g.collection(COMMUNITY_COLLECTION).list({ limit: 100 });
+      check("an unpublished draft is invisible to the public", !anon.records.some((r) => r.id === draft.id));
+      await A.collection(COMMUNITY_COLLECTION).delete(draft.id);
+    } else skip("community probes (plain text, drafts)", "set COMMUNITY_COLLECTION");
 
-// ── TIER B — cross-user isolation. Dev environments only. ───────────────────
-if (SK && !SK.startsWith("sk_live")) {
-  const srv = gemmeinServer(SK, opts);
-  const [a, b] = await Promise.all(TEST_USERS.map((e) => srv.testSession(e)));
-  const A = asUser(a.token), B = asUser(b.token);
-
-  const note = await A.collection(PRIVATE_COLLECTION).create({ probe: "a-secret" });
-  await refuse("B can't read A's private record", "not_found",
-    () => B.collection(PRIVATE_COLLECTION).get(note.id));
-  const bSees = await B.collection(PRIVATE_COLLECTION).list();
-  check("B's private list contains none of A's records",
-    !bSees.records.some((r) => r.id === note.id));
-
-  const who = await A.auth.currentUser();              // the shape your UI reads
-  check("currentUser() exposes userId (not id)", !!who.userId);
-  check("record fields live under .data", note.data?.probe === "a-secret");
-
-  await A.collection(PRIVATE_COLLECTION).delete(note.id);   // leave dev tidy
-} else {
-  console.log(SK ? "· Tier B skipped — sk_live can never mint test sessions (by design)"
-                 : "· Tier B skipped — set SECRET_KEY (sk_dev) to prove cross-user isolation");
+    // entitlements: no plan, no access — and the error names the plan.
+    if (GATED_COLLECTION) {
+      const e = await refuse("no plan → entitlement_required", "entitlement_required",
+        () => B.collection(GATED_COLLECTION).list());
+      if (e?.code === "entitlement_required")
+        check("the refusal names the plan key (err.requires)", typeof e.requires === "string" && e.requires.length > 0, e.requires);
+    } else skip("entitlement probe (paid access)", "set GATED_COLLECTION");
+  } else {
+    console.log(SK ? "· Tier B skipped — sk_live can never mint test sessions (by design)"
+                   : "· Tier B skipped — set SECRET_KEY (sk_dev) to prove cross-user isolation");
+  }
+} catch (e) {
+  console.error(`\nreaffirm could not complete — this is a config or connectivity failure, NOT a boundary verdict:`);
+  console.error(`  ${e.code ?? "(no code)"} — ${e.message}`);
+  if (e.code === "unknown_collection") console.error("  → a CONFIG name doesn't exist in this app. Fix the CONFIG block at the top of this file.");
+  process.exit(2);
 }
 
 console.log(fail ? `\n${fail} boundary check(s) FAILED` : "\nall boundaries reaffirmed");