feat(comprehension): OWASP_CX_COMPLETE asks for every part of multi-part and cross-standard questions

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

Author: northdpole

application/tests/comprehension_test.py

@@ -844,3 +844,21 @@ def execute(self, *a, **k):
         db = mock.Mock(session=Session())
         out = store.changed_chunks(db, "OWASP/X", chunks)
         self.assertEqual([c[:2] for c in out], [("a.md", 1), ("b.md", 0)])
+
+
+class CompleteRuleTest(unittest.TestCase):
+    def test_complete_flag_adds_rule(self):
+        from application.utils.comprehension.synthesize import COMPLETE_RULE, synthesize
+
+        seen = []
+
+        def llm(system, user):
+            seen.append(system)
+            return "A [1]"
+
+        with mock.patch.dict(os.environ, {"OWASP_CX_COMPLETE": "1"}):
+            synthesize("q", [("t", ["u"])], llm)
+        with mock.patch.dict(os.environ, {"OWASP_CX_COMPLETE": ""}):
+            synthesize("q", [("t", ["u"])], llm)
+        self.assertIn(COMPLETE_RULE, seen[0])
+        self.assertNotIn(COMPLETE_RULE, seen[1])

application/utils/comprehension/flags.py

@@ -22,6 +22,7 @@
     "LONGCTX",  # questions naming 1-3 projects: read the whole project text, not snippets
     "FAST",  # final-answer model runs with thinking off (latency)
     "PROJROUTE",  # no project named: pick one by similarity to project overviews
+    "COMPLETE",  # answer every part and every named standard; include all stated items
     "EDITION",  # answer for the edition/year/version the question names; never mix editions
     "DECOMPOSE",  # multi-part questions: answer each part separately, then combine
 )

application/utils/comprehension/synthesize.py

@@ -35,6 +35,14 @@
 )
 
 
+COMPLETE_RULE = (
+    " Cover every part of the question: if it has several parts or names several standards or "
+    "projects, answer each one in turn (one short line per part), and when asked to list or compare, "
+    "include every item the sources state. For a 'which is NOT ...' question, check each option "
+    "against the sources before choosing."
+)
+
+
 def synthesize(
     question: str,
     hits: Sequence[Tuple[str, List[str]]],
@@ -47,6 +55,8 @@ def synthesize(
 
     if flags.on("EDITION"):
         system += EDITION_RULE
+    if flags.on("COMPLETE"):
+        system += COMPLETE_RULE
     numbered = "\n\n".join(
         f"[{i + 1}] {t.split(chr(10) + chr(10) + 'Citations:')[0]}"
         for i, (t, _) in enumerate(hits)