test(agent-bff): document the deliberate non-string html cast

Author: Tonours

packages/agent-bff/src/action/action-execute-mapper.ts

@@ -92,8 +92,6 @@ export function mapActionExecuteResult(raw: unknown): ActionExecuteMapped {
         type: 'success',
         message: typeof body.success === 'string' ? body.success : null,
         invalidated: relationships.filter((name): name is string => typeof name === 'string'),
-        // Untrusted agent output sanitized at the BFF boundary (PRD-1095); a non-string value
-        // relays as null exactly as before.
         html: sanitizeActionHtml(body.html),
       },
     };

packages/agent-bff/src/action/action-routes-middleware.ts

@@ -138,8 +138,6 @@ async function handleExecute({
     }
 
     if (error instanceof ActionFormValidationError) {
-      // The agent's Error-result html is the same untrusted output as the success html, so it
-      // goes through the same sanitizer; a non-string value yields no details rather than a 500.
       const html = sanitizeActionHtml(error.html);
 
       throw actionError(error.message, html === null ? undefined : { html });

packages/agent-bff/src/action/sanitize-action-html.ts

@@ -1,13 +1,5 @@
 import sanitizeHtml from 'sanitize-html';
 
-/**
- * The agent's action-result html is untrusted output: an action can interpolate record data into
- * it, so the BFF sanitizes it at this trust boundary before relaying (PRD-1095). The library's
- * default allowlist keeps rich text (paragraphs, emphasis, links, tables, lists) and drops active
- * content: script/style/iframe/svg/img tags, on* attributes, javascript:/data: URLs. A non-string
- * value is an unchecked agent-JSON cast, not html: null is relayed instead of throwing inside the
- * sanitizer.
- */
 export default function sanitizeActionHtml(html: unknown): string | null {
   return typeof html === 'string' ? sanitizeHtml(html) : null;
 }

packages/agent-bff/test/action/action-execute-mapper.test.ts

@@ -19,8 +19,6 @@ describe('mapActionExecuteResult', () => {
     });
   });
 
-  // PRD-1095: the agent's html is untrusted output (it can interpolate record data), so the BFF
-  // sanitizes it with an allowlist before relaying. Safe markup survives; active markup does not.
   it('sanitizes the success html: safe markup kept, active markup stripped', () => {
     expect(
       mapActionExecuteResult({

packages/agent-bff/test/action/action-routes-middleware.test.ts

@@ -665,8 +665,6 @@ describe('action execute', () => {
     });
   });
 
-  // PRD-1095: the same untrusted agent html travels in a native action Error's details; the
-  // sanitizer applies here too, and a non-string html yields no details rather than a 500.
   it('sanitizes the html carried by a native action Error in details', async () => {
     const form = makeAction({
       execute: jest.fn(async () => {