@rebasepro/server-postgres 0.9.1-canary.fd3754b → 0.10.0

package/src/security/policy-drift.ts

@@ -25,6 +25,18 @@ export interface PolicyRef {
     roles: string[];
     /** SELECT / INSERT / UPDATE / DELETE / ALL. */
     command: string;
+    /** Whether a USING clause is present at all (not what it says). */
+    hasUsing: boolean;
+    /** Whether a WITH CHECK clause is present at all (not what it says). */
+    hasWithCheck: boolean;
+    /**
+     * The live clause text, when read from `pg_policies`. Present only for live
+     * policies (the expected side is parsed from DDL and does not carry it).
+     * Used solely for the insecure-tautology scan, not for divergence — Postgres
+     * rewrites this text, so it is not safe to diff against expected.
+     */
+    qual?: string | null;
+    withCheck?: string | null;
 }
 
 export interface PolicyDrift {
@@ -34,24 +46,76 @@ export interface PolicyDrift {
     orphaned: PolicyRef[];
     /** Same policy name, different roles or command. */
     diverged: { expected: PolicyRef; actual: PolicyRef; differences: string[] }[];
+    /**
+     * A live policy whose expression is the known-permissive tautology
+     * `auth.uid() IS NOT NULL` — true for anonymous visitors too, because the
+     * user path coerces a blank id to the `'anonymous'` sentinel. This is what
+     * `policy.authenticated()` used to compile to, so a database pushed before
+     * that fix carries it, and neither the name, roles, command nor clause
+     * *presence* differs from the corrected policy — the only thing that changed
+     * is the expression text, which this checker otherwise (correctly) ignores.
+     * So it is the one drift that hides from every other check here.
+     *
+     * @see reason  a sentence naming the clause and what to do.
+     */
+    insecure: { policy: PolicyRef; reason: string }[];
 }
 
 export interface Queryable {
     query<R>(text: string, values?: unknown[]): Promise<{ rows: R[] }>;
 }
 
-const CREATE_POLICY = /CREATE POLICY "([^"]+)" ON "([^"]+)"\."([^"]+)"\s+AS (\w+)\s+FOR (\w+)\s+TO ([^\n]+?)(?:\s+USING|\s+WITH CHECK|;)/gi;
+// The trailing group captures whichever clause follows the TO list, which is
+// what tells us the clause is present.
+const CREATE_POLICY = /CREATE POLICY "([^"]+)" ON "([^"]+)"\."([^"]+)"\s+AS (\w+)\s+FOR (\w+)\s+TO ([^\n]+?)(\s+USING\s*\(|\s+WITH CHECK\s*\(|;)/gi;
+
+const WITH_CHECK_NEXT = /^\s+WITH CHECK\s*\(/i;
+
+/**
+ * Index just past the `)` closing a clause whose `(` ends at `open`.
+ *
+ * Needed because a policy expression nests parens and can contain a quoted
+ * literal holding either character, so "find the next `)`" would stop early and
+ * miss the `WITH CHECK` that follows.
+ */
+function clauseEnd(ddl: string, open: number): number {
+    let depth = 1;
+    let inQuote = false;
+    for (let i = open; i < ddl.length; i++) {
+        const c = ddl[i];
+        if (inQuote) {
+            // '' is an escaped quote inside a string, not a close.
+            if (c === "'") {
+                if (ddl[i + 1] === "'") i++;
+                else inQuote = false;
+            }
+            continue;
+        }
+        if (c === "'") inQuote = true;
+        else if (c === "(") depth++;
+        else if (c === ")" && --depth === 0) return i + 1;
+    }
+    return ddl.length;
+}
 
 /** Parse the generated DDL rather than rebuilding the shape by hand. */
 export function parseExpectedPolicies(ddl: string): PolicyRef[] {
     const found: PolicyRef[] = [];
     for (const m of ddl.matchAll(CREATE_POLICY)) {
-        const [, name, schema, table, , command, rolesRaw] = m;
+        const [, name, schema, table, , command, rolesRaw, clause] = m;
         const roles = rolesRaw
             .split(",")
             .map((r) => r.trim().replace(/^"|"$/g, ""))
             .filter(Boolean);
-        found.push({ schema, table, name, roles, command: command.toUpperCase() });
+
+        // The generator emits USING before WITH CHECK, so what follows the TO
+        // list settles USING; WITH CHECK is then whatever follows that clause.
+        const hasUsing = /USING/i.test(clause);
+        const hasWithCheck = hasUsing
+            ? WITH_CHECK_NEXT.test(ddl.slice(clauseEnd(ddl, m.index + m[0].length)))
+            : /WITH CHECK/i.test(clause);
+
+        found.push({ schema, table, name, roles, command: command.toUpperCase(), hasUsing, hasWithCheck });
     }
     return found;
 }
@@ -59,8 +123,9 @@ export function parseExpectedPolicies(ddl: string): PolicyRef[] {
 async function readLivePolicies(client: Queryable, schemas: string[]): Promise<PolicyRef[]> {
     const { rows } = await client.query<{
         schemaname: string; tablename: string; policyname: string; roles: string[] | string; cmd: string;
+        qual: string | null; with_check: string | null;
     }>(
-        `SELECT schemaname, tablename, policyname, roles, cmd
+        `SELECT schemaname, tablename, policyname, roles, cmd, qual, with_check
          FROM pg_policies
          WHERE schemaname = ANY($1)`,
         [schemas]
@@ -74,10 +139,34 @@ async function readLivePolicies(client: Queryable, schemas: string[]): Promise<P
         roles: Array.isArray(r.roles)
             ? r.roles
             : String(r.roles ?? "").replace(/^\{|\}$/g, "").split(",").filter(Boolean),
-        command: (r.cmd ?? "ALL").toUpperCase()
+        command: (r.cmd ?? "ALL").toUpperCase(),
+        // Presence only. Postgres rewrites the text, but it does not invent or
+        // drop a clause: NULL here means the policy genuinely has none.
+        hasUsing: r.qual != null,
+        hasWithCheck: r.with_check != null,
+        qual: r.qual,
+        withCheck: r.with_check
     }));
 }
 
+/**
+ * The permissive tautology `auth.uid() IS NOT NULL`, without the
+ * `<> 'anonymous'` guard that makes it mean "signed in".
+ *
+ * Whitespace varies with Postgres's rewrite, so match on a collapsed form. The
+ * guard clause (`<> 'anonymous'`, in any spelling) is what distinguishes the
+ * corrected policy from the stale one, so its presence clears the text.
+ */
+function isPermissiveAuthTautology(clause: string | null | undefined): boolean {
+    if (!clause) return false;
+    const flat = clause.toLowerCase().replace(/\s+/g, " ");
+    if (!/auth\.uid\(\)\s*is not null/.test(flat)) return false;
+    // The fix appends `AND auth.uid() <> 'anonymous'`; Postgres may store the
+    // literal as `'anonymous'::text`. Either spelling means it is the corrected
+    // policy, not the tautology.
+    return !/<>\s*'anonymous'/.test(flat) && !/!=\s*'anonymous'/.test(flat);
+}
+
 const keyOf = (p: PolicyRef) => `${p.schema}.${p.table}.${p.name}`;
 const sameRoles = (a: string[], b: string[]) =>
     a.length === b.length && [...a].sort().join(",") === [...b].sort().join(",");
@@ -85,11 +174,18 @@ const sameRoles = (a: string[], b: string[]) =>
 /**
  * Diff expected against live.
  *
- * Compares names, roles and command only — all exact values. Policy
- * *expressions* are deliberately not compared: Postgres rewrites `qual`/
- * `with_check` when storing them (parenthesising, casting, schema-qualifying),
- * so text comparison reports drift that does not exist, and a check that cries
- * wolf gets ignored. Roles alone catch the failure this exists for.
+ * Compares names, roles, command, and whether each clause exists — all exact
+ * values. Policy expression *text* is deliberately not compared: Postgres
+ * rewrites `qual`/`with_check` when storing them (parenthesising, casting,
+ * schema-qualifying), so text comparison reports drift that does not exist, and
+ * a check that cries wolf gets ignored.
+ *
+ * Presence is not text, though. A NULL `qual` is not a rewrite of an
+ * expression, it is the absence of one, and absence has no false-positive risk:
+ * either the generator emitted a clause or it did not. That distinction is worth
+ * the extra comparison — a production database was found with a SELECT policy
+ * whose `qual` was NULL, matching on every field this checked and denying 100%
+ * of reads. The same blindness would hide a policy that fails open.
  */
 export async function checkPolicyDrift(
     client: Queryable,
@@ -99,13 +195,31 @@ export async function checkPolicyDrift(
     const schemas = [...new Set(expected.map((p) => p.schema))];
     // Nothing expected means nothing to reconcile against; scanning every
     // schema would report the whole database as orphaned.
-    if (schemas.length === 0) return { missing: [], orphaned: [], diverged: [] };
+    if (schemas.length === 0) return { missing: [], orphaned: [], diverged: [], insecure: [] };
 
     const live = await readLivePolicies(client, schemas);
     const liveByKey = new Map(live.map((p) => [keyOf(p), p]));
     const expectedByKey = new Map(expected.map((p) => [keyOf(p), p]));
 
-    const drift: PolicyDrift = { missing: [], orphaned: [], diverged: [] };
+    const drift: PolicyDrift = { missing: [], orphaned: [], diverged: [], insecure: [] };
+
+    // Scan every live policy for the permissive tautology. This is deliberately
+    // independent of the name-keyed diff below: a database pushed before the
+    // `authenticated()` fix matches its expected policy on name, roles, command
+    // and clause presence, so nothing else here would flag it.
+    for (const p of live) {
+        const clause = isPermissiveAuthTautology(p.qual)
+            ? "USING"
+            : isPermissiveAuthTautology(p.withCheck) ? "WITH CHECK" : null;
+        if (clause) {
+            drift.insecure.push({
+                policy: p,
+                reason: `${clause} is \`auth.uid() IS NOT NULL\`, which is true for anonymous ` +
+                    `visitors too — this grants access to signed-out requests. It predates the ` +
+                    `\`policy.authenticated()\` fix; re-run \`rebase db push\` to tighten it.`
+            });
+        }
+    }
 
     for (const [key, want] of expectedByKey) {
         const got = liveByKey.get(key);
@@ -120,6 +234,13 @@ export async function checkPolicyDrift(
         if (want.command !== got.command) {
             differences.push(`command: expected ${want.command}, database has ${got.command}`);
         }
+        for (const clause of ["USING", "WITH CHECK"] as const) {
+            const key = clause === "USING" ? "hasUsing" : "hasWithCheck";
+            if (want[key] === got[key]) continue;
+            differences.push(want[key]
+                ? `${clause}: expected an expression, database has none — this policy matches no rows`
+                : `${clause}: expected none, database has an expression`);
+        }
         if (differences.length > 0) drift.diverged.push({ expected: want, actual: got, differences });
     }
 
@@ -130,8 +251,64 @@ export async function checkPolicyDrift(
     return drift;
 }
 
+/**
+ * Does this name look like one the generator produced for this table?
+ *
+ * Unnamed rules compile to `<table>_<op>_<sha1[0:7]>` (plus `_<idx>` when one
+ * rule spans several operations), and the hash covers the rule's semantics — so
+ * *editing* a rule renames its policy. The policy under the old name is left
+ * behind by `db push`, which only DROPs the names it is about to CREATE, and
+ * Postgres ORs PERMISSIVE policies together: a superseded `USING (true)` keeps
+ * granting everything no matter how tight its replacement is.
+ *
+ * Matching the shape is what makes dropping them safe. A hand-written policy
+ * would have to collide with a 7-hex digest to be mistaken for generated one;
+ * a policy named anything else is left alone and merely reported, because a
+ * custom name is indistinguishable from one someone wrote in SQL on purpose.
+ */
+export function isGeneratedPolicyName(name: string, table: string): boolean {
+    return new RegExp(`^${table.replace(/[.*+?^${}()|[\]\\]/g, "\\$&")}_(select|insert|update|delete|all)_[0-9a-f]{7}(_\\d+)?$`)
+        .test(name);
+}
+
+export interface OrphanCleanup {
+    /** Superseded generated policies that were dropped. */
+    dropped: PolicyRef[];
+    /** Orphans left in place because their names are not generator-shaped. */
+    kept: PolicyRef[];
+}
+
+/**
+ * Drop the policies an earlier push superseded but never removed.
+ *
+ * Only touches tables the collections describe — a table with no expected
+ * policy is not ours to reconcile, and scanning by schema alone would sweep up
+ * policies belonging to something else sharing the database.
+ */
+export async function dropOrphanedPolicies(
+    client: Queryable,
+    drift: PolicyDrift,
+    collections: CollectionConfig[]
+): Promise<OrphanCleanup> {
+    const expected = parseExpectedPolicies(generatePostgresPoliciesDdl(collections));
+    const managed = new Set(expected.map((p) => `${p.schema}.${p.table}`));
+
+    const cleanup: OrphanCleanup = { dropped: [], kept: [] };
+    for (const p of drift.orphaned) {
+        if (!managed.has(`${p.schema}.${p.table}`) || !isGeneratedPolicyName(p.name, p.table)) {
+            cleanup.kept.push(p);
+            continue;
+        }
+        // Identifiers are quoted, and the name came from pg_policies rather than
+        // from user input, so it is already a valid identifier.
+        await client.query(`DROP POLICY IF EXISTS "${p.name}" ON "${p.schema}"."${p.table}"`);
+        cleanup.dropped.push(p);
+    }
+    return cleanup;
+}
+
 export const hasDrift = (d: PolicyDrift): boolean =>
-    d.missing.length > 0 || d.orphaned.length > 0 || d.diverged.length > 0;
+    d.missing.length > 0 || d.orphaned.length > 0 || d.diverged.length > 0 || d.insecure.length > 0;
 
 /** Human-readable report; empty string when the database matches the config. */
 export function formatPolicyDrift(drift: PolicyDrift): string {
@@ -155,5 +332,12 @@ export function formatPolicyDrift(drift: PolicyDrift): string {
             for (const diff of d.differences) lines.push(`        ${diff}`);
         }
     }
+    if (drift.insecure.length > 0) {
+        lines.push("  Insecure — a live policy grants access it should not:");
+        for (const i of drift.insecure) {
+            lines.push(`    • ${i.policy.schema}.${i.policy.table} → "${i.policy.name}"`);
+            lines.push(`        ${i.reason}`);
+        }
+    }
     return lines.join("\n");
 }

package/src/security/rls-enforcement.ts

@@ -1,4 +1,6 @@
 import { sql as drizzleSql, SQL } from "drizzle-orm";
+import { ANONYMOUS_USER_ID, PolicyExpression, SecurityRule } from "@rebasepro/types";
+import { AnonymousGrantRisk, findAnonymousGrants, securityRuleToConditions } from "@rebasepro/common";
 import { logger } from "@rebasepro/server";
 
 /**
@@ -57,7 +59,7 @@ export interface ConnectionPosture {
 }
 
 export interface AuthContext {
-    userId: string;
+    uid: string;
     /** Raw roles as carried on the user (strings or `{ id }` objects). */
     roles: unknown[];
 }
@@ -198,21 +200,29 @@ export async function ensureAppRole(run: RawSqlRunner, schemas: string[]): Promi
  * SECURITY: this function is only ever called on the **user** path (the server
  * context uses the base/owner driver and never calls it). The default policies
  * treat `auth.uid() IS NULL` as the trusted server context, and `auth.uid()`
- * is `NULLIF(current_setting('app.user_id'), '')` — so an EMPTY user id would
+ * is `NULLIF(current_setting('app.uid'), '')` — so an EMPTY user id would
  * be read as NULL and silently escalate a user request to server privileges.
- * Coerce empty/blank ids to a sentinel here, at the single chokepoint, rather
- * than trusting every caller (e.g. realtime subscription auth) to do it.
+ * Coerce empty/blank ids to `ANONYMOUS_USER_ID` here, at the single chokepoint,
+ * rather than trusting every caller (e.g. realtime subscription auth) to do it.
+ * That sentinel is exported from `@rebasepro/types` because it leaks into rule
+ * semantics: it is why `auth.uid() IS NOT NULL` is true for anonymous requests.
  */
 export async function applyAuthContext(tx: SqlTx, auth: AuthContext, userRole?: string): Promise<void> {
-    const userId = typeof auth.userId === "string" && auth.userId.trim() !== "" ? auth.userId : "anonymous";
+    const uid = typeof auth.uid === "string" && auth.uid.trim() !== "" ? auth.uid : ANONYMOUS_USER_ID;
     const normalizedRoles = auth.roles.map((r: unknown) =>
         typeof r === "string" ? r : (r as Record<string, unknown>)?.id ?? String(r)
     );
+    // `app.user_id` is the pre-rename spelling, still written because policies
+    // are data: a database provisioned before the rename holds rules compiled
+    // to `current_setting('app.user_id')`, and those predicates would evaluate
+    // to NULL — failing open or locking out — if we stopped setting it. Drop
+    // the alias only once no live database carries a legacy policy.
     await tx.execute(drizzleSql`
         SELECT
-            set_config('app.user_id', ${userId}, true),
+            set_config('app.uid', ${uid}, true),
+            set_config('app.user_id', ${uid}, true),
             set_config('app.user_roles', ${normalizedRoles.join(",")}, true),
-            set_config('app.jwt', ${JSON.stringify({ sub: userId, roles: auth.roles })}, true)
+            set_config('app.jwt', ${JSON.stringify({ sub: uid, roles: auth.roles })}, true)
     `);
     if (userRole) {
         await tx.execute(drizzleSql.raw(`SET LOCAL ROLE ${quoteIdent(userRole)}`));
@@ -226,6 +236,63 @@ const FOREIGN_CONVENTION_ROLES: Record<string, string> = {
     service_role: "Supabase"
 };
 
+/**
+ * Warn about rules that read as "signed-in users only" but admit anonymous
+ * callers — `auth.uid() IS NOT NULL`, or a comparison against another
+ * platform's magic user id such as `'anon'`.
+ *
+ * The sibling of {@link validatePolicyPgRoles}, for the more dangerous spelling
+ * of the same habit. A foreign `pgRoles` value makes a policy unreachable and
+ * the table reads empty — loud, and that guard throws. These do the opposite:
+ * the rule compiles to a grant, and nothing looks wrong until the data is
+ * already public.
+ *
+ * Warns rather than throws. Unlike an unreachable `pgRoles`, these rules are
+ * serving traffic today: refusing to boot would take an app offline to report a
+ * problem it already has, and on the read path it would take it offline
+ * *because* its data was exposed. Rewriting the author's SQL is not an option
+ * either — this is the escape hatch whose whole promise is that it means what it
+ * says. So: say so, loudly, and leave the rule alone.
+ */
+export function warnOnAnonymousGrants(
+    collections: { slug?: string; securityRules?: readonly SecurityRule[] }[]
+): void {
+    // Grouped by the mistake, not by the rule: one habit typically repeats
+    // across every collection an author wrote, and a per-rule list would repeat
+    // the same paragraph dozens of times and get skimmed.
+    const byRisk = new Map<string, { risk: AnonymousGrantRisk; sites: string[] }>();
+
+    for (const collection of collections) {
+        for (const rule of collection.securityRules ?? []) {
+            const { usingExpr, withCheckExpr } = securityRuleToConditions(rule);
+            const risks = [usingExpr, withCheckExpr]
+                .filter((e): e is PolicyExpression => e !== null)
+                .flatMap(findAnonymousGrants);
+
+            for (const risk of risks) {
+                const key = `${risk.pattern}:${risk.detail}`;
+                const site = `${collection.slug ?? "(unnamed)"} → "${rule.name ?? "(unnamed rule)"}"`;
+                const entry = byRisk.get(key) ?? { risk, sites: [] };
+                if (!entry.sites.includes(site)) entry.sites.push(site);
+                byRisk.set(key, entry);
+            }
+        }
+    }
+
+    if (byRisk.size === 0) return;
+
+    const problems = [...byRisk.values()].map(({ risk, sites }) =>
+        `  • ${risk.explanation}\n    ${sites.length} rule(s): ${sites.join(", ")}`
+    );
+
+    logger.warn(
+        `Security rules that read as a lockdown but grant access to anonymous requests. Every caller from a ` +
+        `client carries a user id ('${ANONYMOUS_USER_ID}' when nobody is signed in), so these clauses are ` +
+        `true for everyone:\n\n` +
+        problems.join("\n\n") + "\n"
+    );
+}
+
 /**
  * Reject `pgRoles` that this server can never satisfy.
  *

package/src/services/BranchService.ts

@@ -11,10 +11,45 @@ import { sql } from "drizzle-orm";
 import { BranchInfo } from "@rebasepro/types";
 import { DrizzleClient } from "../interfaces";
 import { DatabasePoolManager } from "../databasePoolManager";
+import { extractPgError, extractCauseMessage } from "../utils/pg-error-utils";
 
 /** Internal prefix applied to branch database names to avoid collisions. */
 const BRANCH_DB_PREFIX = "rb_";
 
+/** `duplicate_database` — the target database name is already taken. */
+const PG_DUPLICATE_DATABASE = "42P04";
+
+/** `object_in_use` — the database still has connections attached. */
+const PG_OBJECT_IN_USE = "55006";
+
+/**
+ * Describe a failed branch DDL statement in terms a user can act on.
+ *
+ * Drizzle reports failures as `Failed query: <sql> params:` and hides the real
+ * PostgreSQL error in the `cause` chain, so matching on `err.message` never sees
+ * the actual problem. Match on the PG error code instead — it survives wrapping
+ * and, unlike the message text, is not locale-dependent.
+ */
+function describeBranchDdlError(err: unknown, fallbackContext: string): Error {
+    const pgError = extractPgError(err);
+
+    if (pgError?.code === PG_DUPLICATE_DATABASE) {
+        return new Error(`Database "${fallbackContext}" already exists on the server. Choose a different branch name.`);
+    }
+    if (pgError?.code === PG_OBJECT_IN_USE) {
+        return new Error(
+            `Cannot complete the operation: the database "${fallbackContext}" has active connections. ` +
+            "Close other clients or connections and try again."
+        );
+    }
+
+    // Unknown failure: surface the real PG message rather than the Drizzle
+    // wrapper, which would otherwise show the raw SQL and no reason at all.
+    const detail = pgError?.message ?? extractCauseMessage(err);
+    if (detail) return new Error(detail);
+    return err instanceof Error ? err : new Error(String(err));
+}
+
 /** Fully-qualified metadata table in the rebase schema. */
 const BRANCHES_TABLE = "rebase.branches";
 
@@ -109,18 +144,15 @@ export class BranchService {
                 sql.raw(`CREATE DATABASE "${safeDbName}" TEMPLATE "${safeSourceDb}"`)
             );
         } catch (err) {
-            const msg = err instanceof Error ? err.message : String(err);
-            if (msg.includes("already exists")) {
-                throw new Error(`Database "${dbName}" already exists on the server. Choose a different branch name.`);
-            }
-            // If template fails due to active connections, provide a helpful error
-            if (msg.includes("being accessed by other users")) {
+            const pgError = extractPgError(err);
+            if (pgError?.code === PG_OBJECT_IN_USE) {
+                // The template — not the new database — is the one still in use.
                 throw new Error(
                     `Cannot create branch: the source database "${sourceDb}" has active connections. ` +
                     "Close other clients or connections and try again."
                 );
             }
-            throw err;
+            throw describeBranchDdlError(err, dbName);
         }
 
         // Record metadata in the default database
@@ -166,14 +198,14 @@ export class BranchService {
         try {
             await this.db.execute(sql.raw(`DROP DATABASE "${safeDbName}"`));
         } catch (err) {
-            const msg = err instanceof Error ? err.message : String(err);
-            if (msg.includes("being accessed by other users")) {
+            const pgError = extractPgError(err);
+            if (pgError?.code === PG_OBJECT_IN_USE) {
                 throw new Error(
                     `Cannot delete branch "${sanitizedName}": the database has active connections. ` +
                     "Close other clients and try again."
                 );
             }
-            throw err;
+            throw describeBranchDdlError(err, dbName);
         }
 
         // Remove metadata