fix(search): strip only JSON cards from MCP answers and keep tag-shaped prose

- Remove interactive cards by their JSON payload so closing markers inside strings cannot end a card and tag-shaped prose stays
- Keep the Slack stream's existing tag withholding unchanged
- Require valid element syntax in looksLikeHtml so an address like <a@acme.com> stays text
- Keep verified results when verification is rate-limited and name the rate limit

Author: waleedlatif1

apps/sim/connectors/utils.test.ts

@@ -72,6 +72,7 @@ import {
   isIndexableConnectorFile,
   isSkippableMicrosoftGraphFolderError,
   isSkippedDocument,
+  looksLikeHtml,
   MICROSOFT_GRAPH_MAX_CURSOR_ENCODED_BYTES,
   MICROSOFT_GRAPH_MAX_ITEM_ID_BYTES,
   MICROSOFT_GRAPH_MAX_PENDING_FOLDERS,
@@ -1754,3 +1755,14 @@ describe('BoundedLines', () => {
     })
   })
 })
+
+describe('looksLikeHtml', () => {
+  it('recognizes elements with valid tag syntax', () => {
+    for (const markup of ['<p>', '</p>', '<br>', '<br/>', '<a href="x">', '<td class="c">'])
+      expect(looksLikeHtml(`text ${markup} text`)).toBe(true)
+  })
+  it('does not mistake an address whose name starts like a tag for markup', () => {
+    for (const text of ['Invite <a@acme.com>', 'From <b.smith@acme.com>', '<i.e. later>'])
+      expect(looksLikeHtml(text)).toBe(false)
+  })
+})

apps/sim/connectors/utils.ts

@@ -339,11 +339,11 @@ function decodeCharacterReference(raw: string, code: number): string {
  * A false positive therefore does not merely pass text through untouched — it
  * deletes the bracketed span and flattens the document's line structure. Plain
  * text routinely contains angle brackets that are not markup: an email address
- * (`Reply from John <john@acme.com>`), a markdown autolink
+ * (`Reply from John <john@acme.com>`, or `<a@acme.com>`, whose name is a tag's), a markdown autolink
  * (`<https://acme.com>`), or a placeholder (`<redacted>`).
  */
 const HTML_TAG_PATTERN =
-  /<\/?(?:p|div|br|hr|ul|ol|li|h[1-6]|table|thead|tbody|tr|td|th|span|strong|em|b|i|u|a|code|pre|blockquote|img|figure)\b[^>]*>/i
+  /<\/?(?:p|div|br|hr|ul|ol|li|h[1-6]|table|thead|tbody|tr|td|th|span|strong|em|b|i|u|a|code|pre|blockquote|img|figure)(?=[\s/>])[^>]*>/i
 
 /**
  * Reports whether a value carries real HTML markup and is therefore worth routing

apps/sim/lib/knowledge/application/chat-citations.test.ts

@@ -0,0 +1,21 @@
+/** @vitest-environment node */
+import { describe, expect, it } from 'vitest'
+import { stripInteractiveCards } from '@/lib/knowledge/application/chat-citations'
+
+describe('stripInteractiveCards', () => {
+  it('removes a JSON card, including closing markers inside its strings', () => {
+    expect(
+      stripInteractiveCards('Answer <options>{"1":{"title":"Say </options> x"}}</options> end')
+    ).toBe('Answer  end')
+  })
+  it('removes thinking blocks', () => {
+    expect(stripInteractiveCards('<thinking>plan</thinking>Answer')).toBe('Answer')
+  })
+  it('keeps tag-shaped prose', () => {
+    const prose = 'Wrap choices in <options>your list</options> tags, or use <question> to ask.'
+    expect(stripInteractiveCards(prose)).toBe(prose)
+  })
+  it('drops a card left open before its JSON payload', () => {
+    expect(stripInteractiveCards('Connect here <credential>{"type":"link"')).toBe('Connect here ')
+  })
+})

apps/sim/lib/knowledge/application/chat-citations.ts

@@ -8,6 +8,28 @@ export interface SearchChatCitation {
   url: string
 }
 
+/** Cards the Chat UI renders from a JSON payload; an MCP text answer drops them. */
+const CARD_TAGS = 'options|question|usage_upgrade|credential|workspace_resource'
+/** A JSON string owns its escaped quotes and any tag-shaped text inside it. */
+const JSON_STRING = '"(?:\\\\.|[^"\\\\\\r\\n])*"'
+/**
+ * A card whose body is a JSON value, a thinking block, or a card opener left unclosed before a
+ * JSON payload. A tag-shaped pair or opener in ordinary prose is not a card and stays.
+ */
+const INTERACTIVE_CARD = new RegExp(
+  [
+    `<(${CARD_TAGS})>\\s*[[{](?:${JSON_STRING}|[^"<])*?[\\]}]\\s*</\\1>`,
+    '<thinking>[\\s\\S]*?</thinking>',
+    `<(?:${CARD_TAGS})>\\s*[[{][\\s\\S]*$`,
+  ].join('|'),
+  'g'
+)
+
+/** Removes interactive Chat cards so a text-only surface receives only the answer prose. */
+export function stripInteractiveCards(content: string): string {
+  return content.replace(INTERACTIVE_CARD, '')
+}
+
 /** Resolves Assistant source tags only against successful, bounded retrieval evidence. */
 export function resolveSearchChatCitations(content: string, toolCalls: ToolCallSummary[]) {
   const evidence = new Map<string, SearchChatCitation>()

apps/sim/lib/knowledge/application/chat.test.ts

@@ -304,7 +304,8 @@ describe('organization Search Assistant chat', () => {
   })
 
   it('drops interactive Chat tags from the MCP answer but keeps them in the transcript', async () => {
-    const tags = '<options>{"1":"Open it"}</options><question>Which kit?</question>'
+    const tags =
+      '<options>{"1":"Open it"}</options><question>{"prompt":"Which </question> kit?"}</question>'
     mocks.lifecycle.mockResolvedValue(createResult({ content: `Violet suitcase.${tags}` }))
     const result = await execute()
     expect(result.content).toBe('Violet suitcase.')

apps/sim/lib/knowledge/application/chat.ts

@@ -19,10 +19,10 @@ import { requireOrganizationSearchAvailable } from '@/lib/knowledge/access/avail
 import {
   resolveSearchChatCitations,
   type SearchChatCitation,
+  stripInteractiveCards,
 } from '@/lib/knowledge/application/chat-citations'
 import { organizationSearchChatOperation } from '@/lib/knowledge/application/chat-operations'
 import { loadCopilotSearchIntegrations } from '@/lib/mothership/application/load-search-integrations'
-import { stripInteractiveTags } from '@/lib/mothership/chat/interactive-tags'
 import { persistCopilotChatTurn } from '@/lib/mothership/chat/messages-store'
 import { buildCopilotRequestPayload } from '@/lib/mothership/chat/payload'
 import {
@@ -204,7 +204,7 @@ export const organizationSearchChat: OperationUseCase<
       )
       /** MCP clients receive text; interactive Chat cards stay only in the saved transcript. */
       const answer = resolveSearchChatCitations(
-        stripInteractiveTags(assistantMessage.content, { complete: true }),
+        stripInteractiveCards(assistantMessage.content),
         result.toolCalls
       )
       if (!answer.content.trim())

apps/sim/lib/mothership/chat/interactive-tags.test.ts

@@ -274,6 +274,25 @@ describe('authorized live retrieval', () => {
       message: expect.stringContaining('could not be verified'),
     })
   })
+  it('keeps verified results and names a rate limit hit during verification', async () => {
+    mocks.search.mockResolvedValue({
+      documents: [document, { ...document, id: 'other', url: 'https://docs.google.com/other' }],
+    })
+    mocks.service.mockResolvedValue({
+      policy: defaultLiveSearchPolicy(),
+      partial: false,
+      verify: async ({ id }: { id: string }) => {
+        if (id === 'other') throw new NativeSearchError('rate_limited', 'Later', 30)
+        return true
+      },
+    })
+    const result = await searchLiveKnowledge.execute({ principal, input })
+    expect(result.results).toHaveLength(1)
+    expect(result.live?.accounts[0]).toMatchObject({
+      status: 'partial',
+      message: expect.stringContaining('rate-limited verification'),
+    })
+  })
   it('fails the account when a grant is revoked during verification', async () => {
     mocks.service.mockResolvedValue({
       policy: defaultLiveSearchPolicy(),

apps/sim/lib/mothership/chat/interactive-tags.ts

@@ -224,13 +224,16 @@ async function verifyCandidates(session: LiveAccountSession, candidates: LiveCan
         return (await session.verify(document)) ? ('permitted' as const) : ('denied' as const)
       } catch (error) {
         if (error instanceof NativeSearchError && error.status === 'reconnect') throw error
-        return 'unverified' as const
+        return error instanceof NativeSearchError && error.status === 'rate_limited'
+          ? ('rate_limited' as const)
+          : ('unverified' as const)
       }
     }
   )
   return {
     permitted: candidates.filter((_, index) => outcomes[index] === 'permitted'),
-    unverified: outcomes.includes('unverified'),
+    unverified: outcomes.some((outcome) => outcome === 'unverified' || outcome === 'rate_limited'),
+    rateLimited: outcomes.includes('rate_limited'),
   }
 }
 
@@ -359,7 +362,7 @@ export const searchLiveKnowledge = defineAuthorizedKnowledgeUseCase({
          * Local filters run first so provider verification is spent only on eligible results.
          * Undated results are verified too, so their exclusion is reported only when readable.
          */
-        const { permitted, unverified } = await measureSearchStage('live.verify', () =>
+        const { permitted, unverified, rateLimited } = await measureSearchStage('live.verify', () =>
           verifyCandidates(
             session,
             candidates.filter(
@@ -403,7 +406,9 @@ export const searchLiveKnowledge = defineAuthorizedKnowledgeUseCase({
                 ? 'Service account verification covered a bounded subset of the configured users. Narrow the source user list for complete coverage; external Drive users can only search files also visible to the source administrator.'
                 : undefined,
               unverified
-                ? 'Some results could not be verified against the source settings and were omitted.'
+                ? rateLimited
+                  ? 'The provider rate-limited verification, so some results were omitted. Try again later.'
+                  : 'Some results could not be verified against the source settings and were omitted.'
                 : undefined,
               undatedExcluded
                 ? 'Some results lacked date metadata and were excluded; date coverage is incomplete.'