improvement(chat): require key-value colon evidence before dropping an unparsable tag body

Author: j15z

apps/sim/app/workspace/[workspaceId]/home/components/message-content/components/special-tags/special-tags.test.ts

@@ -330,8 +330,8 @@ describe('parseSpecialTags with <question>', () => {
     // fails JSON.parse, so the old was-it-ever-JSON test called them prose and
     // rendered the whole payload verbatim — the markdown layer then swallowed
     // the tag markers and the reader saw a wall of raw JSON. They open `{"` or
-    // `[{`, which marks them as attempted payloads: droppable, like any other
-    // broken emission.
+    // `[{` and carry key-value colons, which marks them as attempted payloads:
+    // droppable, like any other broken emission.
     const cases = [
       'Prose before. <question>{"type": "single_select", "prompt": "How should I proceed?", "options": [{"id": "a", "label": "Confirm the id"}}]}</question>',
       'Prose before. <question>{"type":"multi_select","prompt":"What should I build now?",options": [{"id":"lib","label":"Pattern library"}]}</question>',
@@ -377,15 +377,17 @@ describe('parseSpecialTags with <question>', () => {
     expect(renderedText(parseSpecialTags(raw, false).segments)).toBe(raw)
   })
 
-  it('treats brace-wrapped quoted prose as an attempted payload, by design', () => {
-    // The deliberate edge of the opener heuristic, pinned so it stays a
-    // decision: `{"..."}` reads as a payload the model started and botched (a
-    // key with no value), not prose — prose the reader was meant to see arrives
-    // unwrapped or in bare quotes, and both of those stay rendered (see the
-    // cases above). The array twin parses as JSON, so it was already dropped as
-    // `wrong-shape` before the opener heuristic existed.
+  it('renders brace-wrapped quoted prose — an opener alone is not an attempt', () => {
+    // The attempted-payload call takes BOTH kinds of evidence: the `{"` opener
+    // and a key-value colon outside string literals. `{"the Q4 report"}` has
+    // the opener but no colon — prose in costume, so it renders; a colon
+    // inside the quotes changes nothing. The array twin parses as JSON, so it
+    // was dropped as `wrong-shape` before this heuristic existed and still is —
+    // that verdict comes from a real parse, not from the opener.
     const braceWrapped = 'see <options>{"the Q4 report"}</options> end'
-    expect(renderedText(parseSpecialTags(braceWrapped, false).segments)).toBe('see  end')
+    expect(renderedText(parseSpecialTags(braceWrapped, false).segments)).toBe(braceWrapped)
+    const quotedColon = 'see <options>{"ratio: 4:5"}</options> end'
+    expect(renderedText(parseSpecialTags(quotedColon, false).segments)).toBe(quotedColon)
     const arrayWrapped = 'see <options>["some list item"]</options> end'
     expect(renderedText(parseSpecialTags(arrayWrapped, false).segments)).toBe('see  end')
   })

apps/sim/app/workspace/[workspaceId]/home/components/message-content/components/special-tags/special-tags.tsx

@@ -609,8 +609,8 @@ const LONGEST_TAG_MARKER = Math.max(...SPECIAL_TAG_NAMES.map((name) => `</${name
  */
 function blankJsonStringLiterals(body: string): string {
   // With no quote there is no string literal, so the loop below would copy the
-  // body to itself character by character. Both callers reach here on bodies
-  // that are usually plain prose, and this runs per opener per streamed chunk.
+  // body to itself character by character. Callers reach here on bodies that
+  // are usually plain prose, and this runs per opener per streamed chunk.
   if (!body.includes('"')) return body
 
   let out = ''
@@ -984,20 +984,26 @@ type BodyClass =
  * JSON payload — the line between `not-parsable` (droppable) and
  * `not-a-payload` (must render).
  *
- * The test is the opener pair, whitespace-tolerant: every payload these tags
- * carry is an object of quoted keys or an array of objects/strings, so an
- * attempt opens `{"`, `[{`, or `["`. Prose falls outside it by construction —
+ * Two pieces of evidence, both required, both structural: the opener pair and
+ * a key-value colon. Every payload these tags carry is built from objects of
+ * quoted keys, so an attempt opens `{"`, `[{`, or `["` AND carries a `:`
+ * outside its string literals. Prose fails one or the other by construction —
  * `{the Q4 report}` opens `{t`, `{type: "file"}` opens `{t`, `{'type':'file'}`
- * opens `{'`, a bare scalar opens with its own first character — so the
- * wrapped-prose cases stay rendered while a payload one typo away from valid
- * (`{"type":"multi_select",options": …`) is recognized as the broken emission
- * it is. Named for the question it answers, not the check it performs: the
- * class names assert meaning, and this predicate is what earns the assertion.
+ * opens `{'`, a bare scalar opens with its own first character, and a
+ * brace-wrapped quoted phrase (`{"the Q4 report"}`) has no colon outside its
+ * quotes — so every wrapped-prose shape stays rendered while a payload one
+ * typo away from valid (`{"type":"multi_select",options": …`) is recognized as
+ * the broken emission it is. Deleting text is the harm here, so the predicate
+ * fails toward rendering. Named for the question it answers, not the checks it performs:
+ * the class names assert meaning, and this predicate is what earns the
+ * assertion. Blanks on the rare path only, like {@link isParseableJson} — the
+ * common cases never reach it.
  */
 function wasAttemptedPayload(body: string): boolean {
   const opener = /^\s*([{[])\s*(["{])/.exec(body)
   if (!opener) return false
-  return opener[1] === '{' ? opener[2] === '"' : true
+  if (opener[1] === '{' && opener[2] !== '"') return false
+  return blankJsonStringLiterals(body).includes(':')
 }
 
 /**