sveltekit-admin 0.2.0 → 0.5.0

package/dist/server/query/listQuery.d.ts

@@ -0,0 +1,89 @@
+/**
+ * Query string parsing and Prisma `where` construction for list search
+ * and filters.
+ *
+ * Design reference: docs/design/list-search-filters.md.
+ *
+ * Golden rule (§4.3 of the design doc): the query string SELECTS from a
+ * finite whitelist of operators derived from the field's type. It never
+ * describes a Prisma clause. No operator ever flows from the URL into the
+ * `where` object as a key — the operator string from the URL is only ever
+ * used to look up a fixed table; the table's value (never the URL's raw
+ * string) becomes the Prisma operator key.
+ */
+import type { PrismaModel } from '../introspection/parser.js';
+export type FilterOp = 'equals' | 'contains' | 'startsWith' | 'gte' | 'lte' | 'isnull';
+export interface ActiveFilter {
+    field: string;
+    op: FilterOp;
+    /** Already coerced to the JS type Prisma expects for this field/op. */
+    value: unknown;
+    /** Original string from the query string, kept to re-render the UI. */
+    raw: string;
+}
+export interface IgnoredFilter {
+    /** Raw query param key, e.g. "f.passwordHash" or "f.nope__gte". */
+    param: string;
+    reason: 'unknown-field' | 'not-filterable' | 'bad-operator' | 'bad-value';
+}
+export interface ListQuery {
+    q: string | null;
+    searchFields: string[];
+    filters: ActiveFilter[];
+    ignored: IgnoredFilter[];
+}
+/** Field-name candidates used both for relation labels and for the default search heuristic (§2.1). */
+export declare const DEFAULT_LABEL_FIELDS: string[];
+/**
+ * Fields eligible for the free-text search box.
+ *
+ * Explicit `searchFields` config always wins. Otherwise: String fields,
+ * not sensitive, not relation/list/id, whose name is in `labelFields`
+ * (same list used for relation labels — one heuristic, not two that could
+ * drift apart). Empty result means "no search box rendered", never a
+ * fallback that scans every String column.
+ *
+ * `hidden` excludes a field in EVERY case, config included (§3.5): a
+ * field hidden from the list/form display must not remain a value-
+ * confirmation oracle via `?q=`/`contains` just because a developer
+ * explicitly listed it in `searchFields` — that's the exact §0.a leak
+ * closed for sensitive-by-name fields, `hidden` is a second independent
+ * source that must close the same way (§10).
+ */
+export declare function resolveSearchFields(model: PrismaModel, configured: string[] | undefined, labelFields?: string[], hidden?: Set<string>): string[];
+export interface DateRange {
+    gte: Date;
+    lt: Date;
+}
+/**
+ * DateTime shortcuts, à la Django. Upper bound is always EXCLUSIVE (`lt`),
+ * never `lte`: `lte 23:59:59.000` misses the last second's milliseconds,
+ * a classic bug invisible in tests unless caught explicitly (§5.5).
+ *
+ * `now` is injectable so tests are deterministic (not "will break at
+ * midnight UTC in CI").
+ */
+export declare function resolveDateShortcut(raw: string, now?: () => Date): DateRange | undefined;
+/**
+ * Parse `?q=` and `?f.*=` into a `ListQuery`. Pure function: no I/O, no
+ * Prisma. `filterableFields` is the set of field names the caller allows
+ * to be filtered (already validated against config/heuristics + the
+ * shared sensitive-name predicate) — this function does not decide
+ * *which* fields are filterable, only how to parse a value once a field
+ * is known to be eligible.
+ */
+export declare function parseListQuery(searchParams: URLSearchParams, model: PrismaModel, enums: Map<string, string[]>, searchFields: string[], filterableFields: Set<string>, now?: () => Date): ListQuery;
+/** A Prisma `where` clause built from a `ListQuery`. Opaque to callers — pass straight to Prisma. */
+export type PrismaWhere = Record<string, unknown>;
+/**
+ * Compose the final Prisma `where`: `AND: [scope, ...filters, {OR: search}]`.
+ * NEVER a spread — a spread of `{...scope, ...filterWhere}` lets a filter
+ * on the same field as the developer's scoping silently overwrite it
+ * (docs/design §0.c, the exact IDOR the previous `?filter=` had). Two
+ * clauses on the same field inside `AND` intersect; they never merge.
+ *
+ * Returns `undefined` when nothing is active, so the query shape sent to
+ * Prisma is byte-for-byte identical to today's unfiltered call — no
+ * regression on existing snapshots/assertions.
+ */
+export declare function buildWhere(query: ListQuery, scope: Record<string, unknown> | undefined, caseInsensitiveSearch: boolean, model: PrismaModel): PrismaWhere | undefined;

package/dist/server/query/listQuery.js

@@ -0,0 +1,428 @@
+/**
+ * Query string parsing and Prisma `where` construction for list search
+ * and filters.
+ *
+ * Design reference: docs/design/list-search-filters.md.
+ *
+ * Golden rule (§4.3 of the design doc): the query string SELECTS from a
+ * finite whitelist of operators derived from the field's type. It never
+ * describes a Prisma clause. No operator ever flows from the URL into the
+ * `where` object as a key — the operator string from the URL is only ever
+ * used to look up a fixed table; the table's value (never the URL's raw
+ * string) becomes the Prisma operator key.
+ */
+import { isSensitiveFieldName } from '../introspection/parser.js';
+/** Max length accepted for the free-text search term. Longer input is truncated. */
+const MAX_SEARCH_LENGTH = 200;
+/** Field-name candidates used both for relation labels and for the default search heuristic (§2.1). */
+export const DEFAULT_LABEL_FIELDS = ['name', 'title', 'label', 'email', 'username', 'slug'];
+/** Types eligible for the free-text search heuristic (String only, see §2.1). */
+function isSearchableByHeuristic(field) {
+    return (field.type === 'String' &&
+        !field.isList &&
+        !field.relation &&
+        !isSensitiveFieldName(field.name) &&
+        !field.isId);
+}
+/**
+ * Fields eligible for the free-text search box.
+ *
+ * Explicit `searchFields` config always wins. Otherwise: String fields,
+ * not sensitive, not relation/list/id, whose name is in `labelFields`
+ * (same list used for relation labels — one heuristic, not two that could
+ * drift apart). Empty result means "no search box rendered", never a
+ * fallback that scans every String column.
+ *
+ * `hidden` excludes a field in EVERY case, config included (§3.5): a
+ * field hidden from the list/form display must not remain a value-
+ * confirmation oracle via `?q=`/`contains` just because a developer
+ * explicitly listed it in `searchFields` — that's the exact §0.a leak
+ * closed for sensitive-by-name fields, `hidden` is a second independent
+ * source that must close the same way (§10).
+ */
+export function resolveSearchFields(model, configured, labelFields = DEFAULT_LABEL_FIELDS, hidden = new Set()) {
+    if (configured) {
+        return configured.filter((name) => {
+            const field = model.fields.find((f) => f.name === name);
+            return field && isFilterableFieldType(field) && !isSensitiveFieldName(name) && !hidden.has(name);
+        });
+    }
+    return model.fields
+        .filter((f) => isSearchableByHeuristic(f) && labelFields.includes(f.name) && !hidden.has(f.name))
+        .map((f) => f.name);
+}
+/** Whether a field's Prisma type can ever appear in a where clause we build (excludes Json/Bytes/relations/lists). */
+function isFilterableFieldType(field) {
+    if (field.relation || field.isList)
+        return false;
+    return !['Json', 'Bytes'].includes(field.type);
+}
+/**
+ * Whitelist of operators per Prisma scalar type. The URL provides a
+ * *string* op name; this table is the only place that turns it into a
+ * real Prisma operator. Anything not listed here for the field's type is
+ * rejected (§4.3).
+ */
+function allowedOpsFor(field) {
+    if (field.isEnum)
+        return ['equals'];
+    switch (field.type) {
+        case 'String':
+            return ['equals', 'contains', 'startsWith'];
+        case 'Int':
+        case 'BigInt':
+        case 'Float':
+        case 'Decimal':
+            return ['equals', 'gte', 'lte'];
+        case 'Boolean':
+            return ['equals'];
+        case 'DateTime':
+            return ['equals', 'gte', 'lte'];
+        default:
+            return [];
+    }
+}
+/**
+ * Coerce a raw query-string value to the JS type Prisma expects, or return
+ * `undefined` if it can't be coerced.
+ *
+ * DateTime `equals`/`gte`/`lte` and enum membership are handled entirely
+ * inside `parseOneFilter` and never reach this function; `allowedOpsFor`
+ * already returns `[]` for Json/Bytes/relations/lists, so `parseOneFilter`
+ * rejects those before reaching this function too. `isnull` DOES reach
+ * here (it's field-type-agnostic), which is why it's handled first.
+ */
+function coerceValue(field, op, raw) {
+    if (op === 'isnull') {
+        if (raw === '1' || raw === 'true')
+            return true;
+        if (raw === '0' || raw === 'false')
+            return false;
+        return undefined;
+    }
+    switch (field.type) {
+        case 'String':
+            return raw;
+        case 'Int':
+        case 'BigInt': {
+            // Strict integer pattern — NOT parseInt: parseInt("12abc") === 12 and
+            // parseInt("") is NaN silently. Both are real bugs the previous
+            // `?filter=` implementation had (see docs/design §0.b).
+            if (!/^-?\d+$/.test(raw))
+                return undefined;
+            if (field.type === 'BigInt') {
+                // `BigInt(str)` never throws once `raw` has already matched the
+                // strict integer regex above — no try/catch needed, and a "just
+                // in case" catch here would be unreachable code the coverage
+                // threshold would force us to fake-test.
+                return BigInt(raw);
+            }
+            const n = Number(raw);
+            return Number.isSafeInteger(n) ? n : undefined;
+        }
+        case 'Float':
+        case 'Decimal': {
+            if (!/^-?\d+(\.\d+)?$/.test(raw))
+                return undefined;
+            // Decimal is passed to Prisma as a string to avoid precision loss on
+            // a round-trip through JS `Number`; Float uses the numeric value.
+            return field.type === 'Decimal' ? raw : Number(raw);
+        }
+        case 'Boolean':
+        default:
+            // Boolean is the only remaining branch reachable in practice; the
+            // `default` exists only so TypeScript accepts a non-exhaustive
+            // switch over `string`, it carries no distinct behaviour.
+            if (raw === 'true' || raw === '1')
+                return true;
+            if (raw === 'false' || raw === '0')
+                return false;
+            return undefined;
+    }
+}
+/**
+ * DateTime shortcuts, à la Django. Upper bound is always EXCLUSIVE (`lt`),
+ * never `lte`: `lte 23:59:59.000` misses the last second's milliseconds,
+ * a classic bug invisible in tests unless caught explicitly (§5.5).
+ *
+ * `now` is injectable so tests are deterministic (not "will break at
+ * midnight UTC in CI").
+ */
+export function resolveDateShortcut(raw, now = () => new Date()) {
+    const today = () => {
+        const d = now();
+        return new Date(Date.UTC(d.getUTCFullYear(), d.getUTCMonth(), d.getUTCDate()));
+    };
+    const addDays = (d, n) => new Date(d.getTime() + n * 86_400_000);
+    if (raw === 'today') {
+        const start = today();
+        return { gte: start, lt: addDays(start, 1) };
+    }
+    if (raw === '7d') {
+        const end = addDays(today(), 1);
+        return { gte: addDays(today(), -6), lt: end };
+    }
+    if (raw === 'month') {
+        const d = now();
+        const start = new Date(Date.UTC(d.getUTCFullYear(), d.getUTCMonth(), 1));
+        const end = new Date(Date.UTC(d.getUTCFullYear(), d.getUTCMonth() + 1, 1));
+        return { gte: start, lt: end };
+    }
+    if (raw === 'year') {
+        const d = now();
+        const start = new Date(Date.UTC(d.getUTCFullYear(), 0, 1));
+        const end = new Date(Date.UTC(d.getUTCFullYear() + 1, 0, 1));
+        return { gte: start, lt: end };
+    }
+    // A single ISO date (no time component): treat as a day-long interval,
+    // never as `equals` — a DateTime stores a time, so `equals` on a bare
+    // date never matches anything (§5.5).
+    const dayMatch = raw.match(/^(\d{4})-(\d{2})-(\d{2})$/);
+    if (dayMatch) {
+        const [, y, m, d] = dayMatch;
+        const year = Number(y);
+        const month = Number(m);
+        const day = Number(d);
+        // `Date.UTC` silently rolls over out-of-range components (month 13 ->
+        // January next year); reject explicitly instead of trusting it.
+        if (month < 1 || month > 12 || day < 1 || day > 31)
+            return undefined;
+        const start = new Date(Date.UTC(year, month - 1, day));
+        if (Number.isNaN(start.getTime()) ||
+            start.getUTCFullYear() !== year ||
+            start.getUTCMonth() !== month - 1 ||
+            start.getUTCDate() !== day) {
+            return undefined;
+        }
+        return { gte: start, lt: addDays(start, 1) };
+    }
+    return undefined;
+}
+/** Parse a `gte`/`lte` DateTime bound: full ISO datetime or a bare date. */
+function parseDateBound(raw) {
+    const d = new Date(raw);
+    return Number.isNaN(d.getTime()) ? undefined : d;
+}
+/**
+ * Split a `f.<field>` or `f.<field>__<op>` param name into its parts.
+ * Cuts on the FIRST `__` after the `f.` prefix — Prisma identifiers can't
+ * contain `__` at all in practice for this parser's purposes, so this is
+ * unambiguous.
+ *
+ * Always called with a key that already starts with `f.` — both call
+ * sites in `parseListQuery` guarantee it (one filters on it explicitly,
+ * the other builds the string literally for the legacy `?filter=` path) —
+ * so there's no `null`-returning guard here to fake-test.
+ */
+function splitFilterParam(key) {
+    const rest = key.slice(2);
+    const sep = rest.indexOf('__');
+    if (sep === -1)
+        return { field: rest, op: null };
+    return { field: rest.slice(0, sep), op: rest.slice(sep + 2) };
+}
+function parseOneFilter(param, raw, ctx) {
+    // Empty value (e.g. the "All" option of a <select> FK/range filter,
+    // submitted via a plain GET form) means "no filter" — the design's
+    // nominal, expected case (§5.4: "?f.x= -> ignored. Important: it's
+    // what a <select> emits for its 'All' option"). This is checked BEFORE
+    // any field lookup and produces NEITHER an ActiveFilter NOR an
+    // IgnoredFilter: it must never surface as a rendered "unknown field"
+    // warning (that was a real regression found in review — classifying
+    // this as `ignored` made the UI show a false error banner on the most
+    // common sidebar interaction there is).
+    if (raw === '') {
+        return { skip: true };
+    }
+    const { field: fieldName, op: opParam } = splitFilterParam(param);
+    const field = ctx.model.fields.find((f) => f.name === fieldName);
+    // Sensitive field: rejected here too, not just by the caller's
+    // `filterableFields` set — defense in depth. Same predicate as
+    // `getDisplayFields`/relation labels (isSensitiveFieldName), and the
+    // exact fix for the oracle described in docs/design §0.a: a sensitive
+    // field is treated EXACTLY like an unknown one, never a distinct
+    // "forbidden" message that would confirm its existence.
+    if (!field || !ctx.filterableFields.has(fieldName) || isSensitiveFieldName(fieldName)) {
+        return { ignored: { param, reason: 'unknown-field' } };
+    }
+    const allowed = allowedOpsFor(field);
+    if (allowed.length === 0) {
+        return { ignored: { param, reason: 'not-filterable' } };
+    }
+    const wantsIsnull = opParam === 'isnull';
+    const op = wantsIsnull ? 'isnull' : (opParam || 'equals');
+    if (op === 'isnull') {
+        if (field.isRequired)
+            return { ignored: { param, reason: 'not-filterable' } };
+    }
+    else if (!allowed.includes(op)) {
+        return { ignored: { param, reason: 'bad-operator' } };
+    }
+    if (field.type === 'DateTime' && op !== 'isnull') {
+        if (op === 'equals') {
+            const range = resolveDateShortcut(raw, ctx.now);
+            if (!range)
+                return { ignored: { param, reason: 'bad-value' } };
+            // A day/shortcut becomes two filters merged by the caller into one
+            // AND pair; represented here as a single filter carrying both
+            // bounds so buildWhere can emit `{gte, lt}` from one ActiveFilter.
+            return {
+                filter: { field: fieldName, op: 'gte', value: range, raw }
+            };
+        }
+        const bound = parseDateBound(raw);
+        if (!bound)
+            return { ignored: { param, reason: 'bad-value' } };
+        return { filter: { field: fieldName, op, value: bound, raw } };
+    }
+    if (field.isEnum) {
+        const members = ctx.enums.get(field.type) ?? [];
+        if (!members.includes(raw))
+            return { ignored: { param, reason: 'bad-value' } };
+        return { filter: { field: fieldName, op: 'equals', value: raw, raw } };
+    }
+    const value = coerceValue(field, op, raw);
+    if (value === undefined)
+        return { ignored: { param, reason: 'bad-value' } };
+    return { filter: { field: fieldName, op, value, raw } };
+}
+/**
+ * Parse `?q=` and `?f.*=` into a `ListQuery`. Pure function: no I/O, no
+ * Prisma. `filterableFields` is the set of field names the caller allows
+ * to be filtered (already validated against config/heuristics + the
+ * shared sensitive-name predicate) — this function does not decide
+ * *which* fields are filterable, only how to parse a value once a field
+ * is known to be eligible.
+ */
+export function parseListQuery(searchParams, model, enums, searchFields, filterableFields, now) {
+    const rawQ = searchParams.get('q');
+    const q = rawQ && rawQ.trim() ? rawQ.trim().slice(0, MAX_SEARCH_LENGTH) : null;
+    const ctx = { model, enums, searchFields, filterableFields, now };
+    const filters = [];
+    const ignored = [];
+    for (const [key] of searchParams) {
+        if (!key.startsWith('f.'))
+            continue;
+        const raw = searchParams.get(key);
+        const result = parseOneFilter(key, raw, ctx);
+        if ('filter' in result)
+            filters.push(result.filter);
+        else if ('ignored' in result)
+            ignored.push(result.ignored);
+        // else: { skip: true } — empty "All" value, silently dropped, not
+        // even recorded in `ignored` (see parseOneFilter's comment).
+    }
+    // Legacy `?filter=field:value` — routed through the exact same
+    // whitelist/coercion path as `f.*`, so it inherits the security fix
+    // (docs/design §4.4, §0.a). If both are present for the same field,
+    // `f.*` wins (parsed above, so it's already in `filters`); the legacy
+    // value is only added if that field has no `f.*` entry.
+    const legacy = searchParams.get('filter');
+    if (legacy && legacy.includes(':')) {
+        const sep = legacy.indexOf(':');
+        const legacyField = legacy.slice(0, sep);
+        const legacyValue = legacy.slice(sep + 1);
+        const alreadyHasField = filters.some((f) => f.field === legacyField);
+        if (!alreadyHasField) {
+            const result = parseOneFilter(`f.${legacyField}`, legacyValue, ctx);
+            if ('filter' in result)
+                filters.push(result.filter);
+            else if ('ignored' in result)
+                ignored.push({ param: 'filter', reason: result.ignored.reason });
+            // else: { skip: true } — an empty legacy value (`?filter=field:`)
+            // is likewise a no-op, never surfaced as an "ignored" warning.
+        }
+    }
+    return { q, searchFields, filters, ignored };
+}
+function clauseOf(filter) {
+    if (filter.op === 'gte' && filter.value && typeof filter.value === 'object' && 'gte' in filter.value) {
+        // Date shortcut carrying both bounds (see parseOneFilter's DateTime branch).
+        const range = filter.value;
+        return { [filter.field]: { gte: range.gte, lt: range.lt } };
+    }
+    if (filter.op === 'isnull') {
+        return { [filter.field]: filter.value ? { equals: null } : { not: null } };
+    }
+    if (filter.op === 'equals') {
+        return { [filter.field]: filter.value };
+    }
+    return { [filter.field]: { [filter.op]: filter.value } };
+}
+/**
+ * Compose the final Prisma `where`: `AND: [scope, ...filters, {OR: search}]`.
+ * NEVER a spread — a spread of `{...scope, ...filterWhere}` lets a filter
+ * on the same field as the developer's scoping silently overwrite it
+ * (docs/design §0.c, the exact IDOR the previous `?filter=` had). Two
+ * clauses on the same field inside `AND` intersect; they never merge.
+ *
+ * Returns `undefined` when nothing is active, so the query shape sent to
+ * Prisma is byte-for-byte identical to today's unfiltered call — no
+ * regression on existing snapshots/assertions.
+ */
+export function buildWhere(query, scope, caseInsensitiveSearch, model) {
+    const and = [];
+    if (scope)
+        and.push(scope);
+    for (const f of query.filters)
+        and.push(clauseOf(f));
+    if (query.q && query.searchFields.length > 0) {
+        const or = [];
+        for (const fieldName of query.searchFields) {
+            const field = model.fields.find((f) => f.name === fieldName);
+            const clause = searchClauseFor(field, query.q, caseInsensitiveSearch);
+            if (clause)
+                or.push({ [fieldName]: clause });
+        }
+        // Never emit `{OR: []}` — in Prisma that matches nothing, which would
+        // silently turn "no searchable field" (or "every clause omitted", §2.4)
+        // into "empty result". A no-op search must add nothing to the where.
+        if (or.length > 0)
+            and.push({ OR: or });
+    }
+    if (and.length === 0)
+        return undefined;
+    if (and.length === 1)
+        return and[0];
+    return { AND: and };
+}
+/**
+ * The per-field-type clause for a `searchFields` entry (§2.4):
+ * - String @id -> `equals` (a `contains` on a cuid/uuid can't use the
+ *   index and never makes semantic sense; §2.1 talks ONLY about the id
+ *   here — an earlier version of this function over-generalized to
+ *   `@id || @unique`, which silently broke fragment search on the most
+ *   common real-world case: `email`/`slug` fields are `@unique` in
+ *   nearly every Prisma schema and are exactly what §2.3's "a title, an
+ *   email" example means by free-text search. `@unique` alone is NOT a
+ *   reason to switch to `equals` — only `@id` is).
+ * - other String (including @unique) -> `contains` (+ `mode:
+ *   'insensitive'` when the provider supports it).
+ * - Int/BigInt/Float/Decimal -> `equals` if `q` coerces to that type,
+ *   otherwise the clause is OMITTED — never `contains` on a numeric
+ *   column, which Prisma rejects with a hard error (`Unknown argument
+ *   contains`), turning any legitimate `?q=` into a 500 (§10's known
+ *   trap, discovered via review — the original implementation searched
+ *   this exactly wrong).
+ * - anything else (enum, Boolean, DateTime, relation, Json/Bytes):
+ *   omitted. `resolveSearchFields`'s auto heuristic never proposes these,
+ *   but explicit `searchFields` config isn't type-checked against §2.4 at
+ *   boot (only against `isFilterableFieldType`), so this is reached in
+ *   practice for a misconfigured field — degrading to "omitted" here
+ *   keeps the guarantee that a legitimate URL never 500s, without adding
+ *   a boot-time validation pass this design doc doesn't ask for.
+ */
+function searchClauseFor(field, q, caseInsensitiveSearch) {
+    if (!field)
+        return undefined;
+    if (field.type === 'String') {
+        if (field.isId)
+            return { equals: q };
+        return caseInsensitiveSearch ? { contains: q, mode: 'insensitive' } : { contains: q };
+    }
+    if (['Int', 'BigInt', 'Float', 'Decimal'].includes(field.type)) {
+        const coerced = coerceValue(field, 'equals', q);
+        return coerced === undefined ? undefined : { equals: coerced };
+    }
+    return undefined;
+}

package/dist/server/query/urls.d.ts

@@ -0,0 +1,33 @@
+/**
+ * URL helpers shared by every piece of UI that builds a list-view URL:
+ * the search box, the filter sidebar links, and pagination.
+ *
+ * Every list URL MUST go through `buildListUrl` — including pagination,
+ * which today builds `?page=N` by hand. A hand-rolled URL bypasses the
+ * "always drop `page` on filter change" invariant (docs/design §3.3, §7.1)
+ * and risks parameter-order drift, which makes snapshots flaky.
+ */
+/**
+ * Build a list-view URL from the current one, applying a patch of query
+ * params. `null` in the patch removes that key. `page` is ALWAYS dropped
+ * unless the patch explicitly sets it (pagination is the one caller that
+ * does) — every other caller (search box, filter link) resets to page 1
+ * implicitly by omitting `page` from its patch.
+ *
+ * Keys are sorted before serialization so the resulting URL is
+ * deterministic — required for stable snapshot tests, and incidentally
+ * nicer to read/bookmark.
+ */
+export declare function buildListUrl(currentUrl: URL, patch: Record<string, string | null>): string;
+/**
+ * Hidden `<input>` params to re-emit inside a `<form method="GET">` so
+ * submitting it doesn't wipe out every other active param — a bare GET
+ * form REPLACES the whole query string with just its own fields (docs/design
+ * §3.3). `exclude` lists the param(s) the form itself controls (e.g. `q`
+ * for the search box, `f.published` for a single-field filter form);
+ * `page` is always excluded too, since any new search/filter resets it.
+ */
+export declare function hiddenParams(currentUrl: URL, exclude: string[]): {
+    name: string;
+    value: string;
+}[];

package/dist/server/query/urls.js

@@ -0,0 +1,58 @@
+/**
+ * URL helpers shared by every piece of UI that builds a list-view URL:
+ * the search box, the filter sidebar links, and pagination.
+ *
+ * Every list URL MUST go through `buildListUrl` — including pagination,
+ * which today builds `?page=N` by hand. A hand-rolled URL bypasses the
+ * "always drop `page` on filter change" invariant (docs/design §3.3, §7.1)
+ * and risks parameter-order drift, which makes snapshots flaky.
+ */
+/**
+ * Build a list-view URL from the current one, applying a patch of query
+ * params. `null` in the patch removes that key. `page` is ALWAYS dropped
+ * unless the patch explicitly sets it (pagination is the one caller that
+ * does) — every other caller (search box, filter link) resets to page 1
+ * implicitly by omitting `page` from its patch.
+ *
+ * Keys are sorted before serialization so the resulting URL is
+ * deterministic — required for stable snapshot tests, and incidentally
+ * nicer to read/bookmark.
+ */
+export function buildListUrl(currentUrl, patch) {
+    const params = new URLSearchParams(currentUrl.search);
+    if (!('page' in patch)) {
+        params.delete('page');
+    }
+    for (const [key, value] of Object.entries(patch)) {
+        if (value === null)
+            params.delete(key);
+        else
+            params.set(key, value);
+    }
+    const sorted = new URLSearchParams();
+    for (const key of [...params.keys()].sort()) {
+        for (const value of params.getAll(key)) {
+            sorted.append(key, value);
+        }
+    }
+    const qs = sorted.toString();
+    return qs ? `${currentUrl.pathname}?${qs}` : currentUrl.pathname;
+}
+/**
+ * Hidden `<input>` params to re-emit inside a `<form method="GET">` so
+ * submitting it doesn't wipe out every other active param — a bare GET
+ * form REPLACES the whole query string with just its own fields (docs/design
+ * §3.3). `exclude` lists the param(s) the form itself controls (e.g. `q`
+ * for the search box, `f.published` for a single-field filter form);
+ * `page` is always excluded too, since any new search/filter resets it.
+ */
+export function hiddenParams(currentUrl, exclude) {
+    const excluded = new Set([...exclude, 'page']);
+    const out = [];
+    for (const [key, value] of currentUrl.searchParams) {
+        if (excluded.has(key))
+            continue;
+        out.push({ name: key, value });
+    }
+    return out;
+}

package/dist/server/router.d.ts

@@ -0,0 +1,6 @@
+export interface ParsedRoute {
+    view: 'dashboard' | 'list' | 'create' | 'edit' | 'notFound' | 'search' | 'logout';
+    model?: string;
+    id?: string;
+}
+export declare function parseRoute(pathname: string, basePath: string): ParsedRoute;

package/dist/server/router.js

@@ -0,0 +1,27 @@
+export function parseRoute(pathname, basePath) {
+    // Le `replace` n'est PAS redondant avec le `filter(Boolean)` plus bas : il est ce
+    // qui fait que `/admin/` et `/admin///` donnent un `path` vide, donc le dashboard.
+    // Sans lui, `path` vaudrait '/' — truthy — et le chemin tomberait sur `notFound`.
+    const path = pathname.slice(basePath.length).replace(/^\/+|\/+$/g, '');
+    if (!path) {
+        return { view: 'dashboard' };
+    }
+    const segments = path.split('/').filter(Boolean);
+    if (segments.length === 1) {
+        if (segments[0] === '_search') {
+            return { view: 'search' };
+        }
+        if (segments[0] === '_logout') {
+            return { view: 'logout' };
+        }
+        return { view: 'list', model: segments[0] };
+    }
+    if (segments.length === 2) {
+        if (segments[1] === 'new') {
+            return { view: 'create', model: segments[0] };
+        }
+        return { view: 'edit', model: segments[0], id: segments[1] };
+    }
+    // 3 segments ou plus : aucune vue ne correspond.
+    return { view: 'notFound' };
+}

package/dist/server/views/Dashboard.svelte

@@ -0,0 +1,29 @@
+<script lang="ts">
+  import StatCard from './StatCard.svelte';
+  import ModelCard from './ModelCard.svelte';
+
+  let {
+    models,
+    stats,
+    basePath
+  }: {
+    models: Array<{ name: string; label: string; count: number }>;
+    stats: { total: number; models: number };
+    basePath: string;
+  } = $props();
+</script>
+
+<h1>Dashboard</h1>
+<p class="ska-subtitle">Welcome to your admin panel</p>
+
+<div class="ska-stats">
+  <StatCard icon="models" value={stats.models} label="Models" />
+  <StatCard icon="records" value={stats.total} label="Total Records" />
+</div>
+
+<h2>Models</h2>
+<div class="ska-models">
+  {#each models as m (m.name)}
+    <ModelCard href="{basePath}/{m.name.toLowerCase()}" label={m.label} count={m.count} />
+  {/each}
+</div>