import type { GroupRule, GroupRuleMatch, GroupRuleMode } from '../models/groups.ts' /** * How a page rule is matched against a page, and which rule wins when several match. * * --------------------------------------------------------------------------------------------- * THE RULES OF PAGE PERMISSIONS * --------------------------------------------------------------------------------------------- * * A group grants page permissions through rules, never as a blanket. Every rule names a set of * permissions (`roles`), a way of addressing pages (`match`, with either `path` or `tags`), and what * it does with them (`mode`). A user's rules are all of their groups' rules pooled together — * belonging to a second group can therefore both widen and narrow what the first one said. * * The two tag kinds read `tags` and ignore `path`; every other kind is the other way round. They are * separate fields rather than one reused for both so that a rule keeps whichever it is not currently * matching on — changing a rule's kind and changing it back is not a way to lose what it said. * * **Nothing is granted by default.** A permission nobody wrote a rule for is denied: no rules at all * is the same as one DENY rule covering the whole site. This is why an empty group can read nothing. * * A rule only has a say over a page it is SCOPED to. `sites` and `locales` each name what the rule * is limited to, and an empty list means every one of them — so a rule left alone speaks for the * whole instance, and one naming a site says nothing at all about the others. * * When more than one rule names the permission being asked about and matches the page, exactly one * of them decides the answer — the most specific. Order in the array means nothing. * * 1. MATCH TYPE, as bands. From weakest to strongest: * * Path Starts With < Path Ends With < Path Matches Regex < * Has Any Tag < Has All Tags < Path Is Exactly * * The order runs from the vaguest way of naming pages to the most precise: a prefix is a whole * branch of the tree, a tag is something somebody put ON the page to say what it is, and an * exact path is one page and nothing else. * * Three BANDS, because path length below only settles a contest inside one of them: the three * path-shaped kinds, then the two tag kinds, then Path Is Exactly. A tag rule therefore beats a * prefix rule however deep that prefix is — `confidential` is denied under `docs` as surely as * anywhere else, and a guests group denying the whole site can still be opened on the pages * tagged `public` — while naming a page outright still beats saying what it is tagged. * * A tag rule has no path, so without the bands it would score zero at step 2 and lose to every * rule that named one, which is every rule a group starts with. * * 2. SPECIFICITY, highest first, within a band. A rule addressing `geography/countries` beats one * addressing `geography`, because it says something about a smaller part of the site. Measured * as the length of the path the rule addresses, so the deeper of two paths always wins, and a * rule for the whole site (empty path) is the least specific thing there is. Tag rules address * no path, so they all score zero and this settles nothing between them. * * 3. MATCH TYPE AGAIN, to separate two kinds sharing a band at the same specificity: Has All Tags * beats Has Any Tag, since every tag in a list is a stronger claim than any one of them, and a * regex beats a suffix beats a prefix. * * 4. MODE, when two rules are equally specific and of the same kind: * * ALLOW < DENY < FORCE ALLOW * * An ALLOW grants the permission. A DENY overrides any ALLOW. A FORCE ALLOW overrides any DENY, * which is what makes a hole in an otherwise closed branch possible. * * The four are applied in that order: mode only settles a tie between rules of the same kind at the * same specificity, so a DENY on `geography` does NOT override an ALLOW on `geography/countries` — * the deeper rule was more specific and had already won. And a path rule cannot make a hole in a tag * rule unless it names the page exactly; otherwise it takes another tag rule — a FORCE ALLOW on the * tag, or taking the tag off the page. * * --------------------------------------------------------------------------------------------- * * `manage:system` is not evaluated here: it bypasses this entirely, and does so before any rule is * read. See `models/groups.ts`. */ /** * A page as a rule sees it. * * `siteId` and `locale` scope it, `path` places it, and `tags` are what a tag rule matches on. * * `siteId` is REQUIRED, and deliberately so: a rule limited to particular sites has to be able to * tell whether this page is in one of them, and a reference that could leave it out would apply * every such rule to every site the moment a caller forgot. There is no page anywhere in this * codebase that does not belong to a site, so nothing is being asked for that is not in hand. */ export interface RulePageRef { siteId: string path: string locale?: string tags?: string[] } /** * Match kinds from weakest to strongest. The index IS the priority, so the order of this array is * the order documented above. */ const MATCH_PRIORITY: GroupRuleMatch[] = ['START', 'END', 'REGEX', 'TAG', 'TAGALL', 'EXACT'] /** * Which band of the ordering each kind sits in, weakest first — step 1 above. * * Path length settles a contest only INSIDE a band, which is the whole point of having them: a tag * rule names no path, so measuring it against one would put it below every rule that did. */ const MATCH_BAND: Record = { START: 0, END: 0, REGEX: 0, TAG: 1, TAGALL: 1, EXACT: 2 } /** Modes from weakest to strongest, used to break a tie between rules of the same kind. */ const MODE_PRIORITY: GroupRuleMode[] = ['ALLOW', 'DENY', 'FORCEALLOW'] /** The kinds that address pages by tag rather than by path. */ const TAG_MATCHES: GroupRuleMatch[] = ['TAG', 'TAGALL'] /** * The tags a rule addresses, as they are compared. * * Lowercased here rather than on the way in, because the page's own tags are stored as they were * typed: a page tagged `Meeting` and a rule naming `meeting` are the same tag, and the wiki has no * canonical case to hold either of them to. */ function ruleTags(rule: GroupRule): string[] { return (rule.tags ?? []).map((tag) => tag.trim().toLowerCase()).filter(Boolean) } /** Compared without leading slashes on either side, since neither is stored with one. */ function normalizePath(value: string): string { return value.replace(/^\/+/, '') } /** * How much of the site a rule is talking about, as a number where higher is narrower. * * The length of the path it addresses. A tag rule addresses no path, so every one of them scores * zero — which says nothing about it, since this is only ever read against another rule of the same * band and both tag kinds are in the same one. */ function specificityOf(rule: GroupRule): number { if (TAG_MATCHES.includes(rule.match)) { return 0 } return normalizePath(rule.path).length } /** * A rule as a sortable key: the four steps documented above, strongest first in each position. * * Compared lexicographically by `outranks`, so adding a step is adding an entry here rather than * another branch in the comparison. */ function rankOf(rule: GroupRule): number[] { return [ MATCH_BAND[rule.match], specificityOf(rule), MATCH_PRIORITY.indexOf(rule.match), MODE_PRIORITY.indexOf(rule.mode) ] } /** * Whether `a` beats `b`, comparing the ranks position by position. * * Strictly greater, so two identical ranks leave the incumbent in place: the first rule of an * otherwise identical pair wins and the outcome does not depend on the order they arrived in. */ function outranks(a: number[], b: number[]): boolean { for (let i = 0; i < a.length; i++) { if (a[i] !== b[i]) { return a[i] > b[i] } } return false } /** Whether a rule addresses this page at all, ignoring what it then says about it. */ export function ruleMatchesPage(rule: GroupRule, page: RulePageRef): boolean { // -> A rule may be limited to particular sites; an empty list means every one of them if (rule.sites?.length > 0 && !rule.sites.includes(page.siteId)) { return false } // -> And to particular locales, the same way. Unlike the site, a reference may leave the locale // out — an asset addressed by path alone — and such a page is not one a locale rule excludes if (rule.locales?.length > 0 && page.locale && !rule.locales.includes(page.locale)) { return false } const pagePath = normalizePath(page.path) const rulePath = normalizePath(rule.path) const pageTags = (page.tags ?? []).map((tag) => tag.toLowerCase()) switch (rule.match) { case 'START': return pagePath.startsWith(rulePath) case 'EXACT': return pagePath === rulePath case 'END': return pagePath.endsWith(rulePath) case 'REGEX': try { return new RegExp(rulePath).test(pagePath) } catch { // -> A rule that cannot compile addresses nothing, rather than everything return false } case 'TAG': return ruleTags(rule).some((tag) => pageTags.includes(tag)) case 'TAGALL': { const tags = ruleTags(rule) return tags.length > 0 && tags.every((tag) => pageTags.includes(tag)) } default: return false } } /** * The rule that decides a permission for a page, out of everything the caller's groups say. * * @param rules Every rule from every group the caller belongs to, pooled * @param permission The single permission being asked about, e.g. `read:pages` * @returns The deciding rule, or null when nothing addresses this — which means denied */ export function resolvePageRule( rules: GroupRule[], permission: string, page: RulePageRef ): GroupRule | null { let winner: GroupRule | null = null let winnerRank: number[] | null = null for (const rule of rules) { if (!rule.roles?.includes(permission) || !ruleMatchesPage(rule, page)) { continue } const rank = rankOf(rule) if (!winnerRank || outranks(rank, winnerRank)) { winner = rule winnerRank = rank } } return winner } /** * Whether the caller's rules grant a permission on a page. * * @returns False when no rule addresses it, which is the default for everything. */ export function rulesAllow(rules: GroupRule[], permission: string, page: RulePageRef): boolean { const rule = resolvePageRule(rules, permission, page) return rule ? rule.mode !== 'DENY' : false }