🛡️ Sentinel: [CRITICAL] Fix SQL injection risk in JSON extraction

.jules/sentinel.md

@@ -7,3 +7,8 @@
 **Vulnerability:** The daemon configuration file (`~/.agor/config.yaml`) and its parent directory (`~/.agor`) were created with default file permissions (e.g., `0o755`/`0o644`), which made them readable by other users on the system. This file stores extremely sensitive information such as API keys and master JWT secrets.
 **Learning:** Default Node.js filesystem operations (`fs.writeFile` and `fs.mkdir`) do not enforce strict permissions unless explicitly specified with a `mode` parameter. When handling sensitive files, relying on the system `umask` is insufficient.
 **Prevention:** Always specify `mode: 0o600` for sensitive files and `mode: 0o700` for their parent directories. Additionally, use `fs.chmod` to retroactively secure existing files and directories that might have been created with permissive defaults.
+
+## 2025-05-14 - [SQL Injection Risk via sql.raw in PostgreSQL JSON extraction]
+**Vulnerability:** A potential SQL injection vulnerability in `jsonExtract` inside `packages/core/src/db/database-wrapper.ts`. The code used `sql.raw` to construct JSON path access string dynamically.
+**Learning:** Using `sql.raw` with unparameterized variables makes the app susceptible to SQL injection. When generating dynamic JSON access paths for PostgreSQL `->` and `->>` operators, explicit cast to text (e.g. `sql\`->> (${key}::text)\``) is required so that Drizzle correctly handles the overloaded JSON operator without syntax errors, allowing us to safely parameterize paths.
+**Prevention:** Never use `sql.raw()` to interpolate dynamic variables. Always cast dynamic parameters to text (`::text`) when using Postgres JSON operators to maintain parameterization while avoiding operator overload ambiguity.

packages/core/src/db/database-wrapper.ts

@@ -105,15 +105,14 @@ export function jsonExtract(db: Database, column: SQL.Aliased | SQL | any, path:
   } else {
     // PostgreSQL: column->'path'->'to'->>'field'
     // Use -> for all but the last part (keeps as JSON), ->> for the last part (extracts as text)
-    // IMPORTANT: Use sql.raw() for JSON keys to avoid parameterization
     if (parts.length === 1) {
       // Single level: column->>'key'
-      return sql`${column}${sql.raw(`->>'${parts[0]}'`)}`;
+      return sql`${column}->>(${parts[0]}::text)`;
     } else {
       // Multiple levels: column->'key1'->'key2'->>'key3'
-      const objectParts = parts.slice(0, -1).map((p) => sql.raw(`->'${p}'`));
+      const objectParts = parts.slice(0, -1).map((p) => sql`->(${p}::text)`);
       const lastPart = parts[parts.length - 1];
-      return sql`${column}${sql.join(objectParts, sql``)}${sql.raw(`->>'${lastPart}'`)}`;
+      return sql`${column}${sql.join(objectParts, sql``)}->>(${lastPart}::text)`;
     }
   }
 }