chore(agent): uncommitted changes from task

backend/apps/operations/services/insight_service.py

@@ -519,6 +519,79 @@ def _normalize_string_newlines(fragment: str) -> str:
     return "".join(out)
 
 
+def _strip_json_noise(fragment: str) -> str:
+    """Remove JS-style comments and trailing commas OUTSIDE string literals.
+
+    String-state-aware (same scanner idea as `_normalize_string_newlines`):
+    `"… // …"` and `"…, }…"` inside JSON string values survive — the old blind
+    regexes corrupted valid JSON into the silent one-liner fallback.
+    """
+    out: list[str] = []
+    in_str = False
+    i = 0
+    n = len(fragment)
+    while i < n:
+        ch = fragment[i]
+        if in_str:
+            if ch == "\\" and i + 1 < n:
+                out.append(ch)
+                out.append(fragment[i + 1])
+                i += 2
+                continue
+            out.append(ch)
+            if ch == '"':
+                in_str = False
+            i += 1
+            continue
+        if ch == '"':
+            in_str = True
+            out.append(ch)
+            i += 1
+            continue
+        if (
+            ch == "/"
+            and i + 1 < n
+            and fragment[i + 1] == "/"
+            and (i == 0 or fragment[i - 1] != ":")  # keep "http://"
+        ):
+            while i < n and fragment[i] != "\n":
+                i += 1  # drop comment, keep the newline
+            continue
+        if ch == "/" and i + 1 < n and fragment[i + 1] == "*":
+            end = fragment.find("*/", i + 2)
+            if end != -1:
+                i = end + 2
+                continue
+            # unclosed block comment — keep as-is (legacy regex was a no-op)
+        if ch == ",":
+            j = i + 1
+            while j < n:
+                cj = fragment[j]
+                if cj in " \t\r\n":
+                    j += 1
+                elif (
+                    cj == "/"
+                    and j + 1 < n
+                    and fragment[j + 1] == "/"
+                    and (j == 0 or fragment[j - 1] != ":")
+                ):
+                    while j < n and fragment[j] != "\n":
+                        j += 1
+                elif cj == "/" and j + 1 < n and fragment[j + 1] == "*":
+                    end = fragment.find("*/", j + 2)
+                    if end == -1:
+                        break
+                    j = end + 2
+                else:
+                    break
+            if j < n and fragment[j] in "}]":
+                i += 1  # drop only the comma; keep following whitespace
+                continue
+        out.append(ch)
+        i += 1
+    return "".join(out)
+
+
 def parse_llm_json(text: str) -> dict[str, str] | None:
     if not text:
         return None
@@ -527,10 +600,8 @@ def parse_llm_json(text: str) -> dict[str, str] | None:
     cleaned = re.sub(r"^```(?:json)?\s*", "", text.strip(), flags=re.I)
     cleaned = re.sub(r"\s*```\s*$", "", cleaned)
     # Tolerate small-model JSON noise (FE parser parity): // and /* */ comments,
-    # trailing commas. Lookbehind keeps "http://" URLs intact.
-    cleaned = re.sub(r"(?<!:)//[^\n]*", "", cleaned)
-    cleaned = re.sub(r"/\*.*?\*/", "", cleaned, flags=re.S)
-    cleaned = re.sub(r",(\s*[}\]])", r"\1", cleaned)
+    # trailing commas — string-aware, see _strip_json_noise.
+    cleaned = _strip_json_noise(cleaned)
     first = cleaned.find("{")
     last = cleaned.rfind("}")
     if first < 0 or last <= first:
@@ -543,7 +614,13 @@ def parse_llm_json(text: str) -> dict[str, str] | None:
             return None
         parts = re.split(r"(?<=[.!?])\s+", prose, maxsplit=1)
         kesimpulan = parts[0].strip()
-        insight = parts[1].strip() if len(parts) > 1 else kesimpulan
+        # No sentence split → don't store the same text twice (FE shows both
+        # fields); reuse the JSON-path message for a missing recommendation.
+        insight = (
+            parts[1].strip()
+            if len(parts) > 1
+            else "Model tidak mengembalikan rekomendasi eksplisit."
+        )
         return {"kesimpulan": kesimpulan, "insight": insight}
     fragment = cleaned[first : last + 1]
     try:

backend/apps/operations/tests_insight_narrate.py

@@ -122,6 +122,38 @@ class ParseLlmJsonTests(SimpleTestCase):
         self.assertIsNotNone(out)
         self.assertIn("FCR 1.70", out["kesimpulan"])
 
+    def test_line_comment_inside_string_value_survives(self):
+        """`//` inside a JSON string must not be stripped as a comment (review MEDIUM)."""
+        raw = (
+            '{"kesimpulan": "Pakan // revisi besok dikirim ke kandang 1.", '
+            '"insight": "Manajemen pakan perlu penyesuaian rasio harian."}'
+        )
+        out = parse_llm_json(raw)
+        self.assertIsNotNone(out)
+        self.assertIn("// revisi besok", out["kesimpulan"])
+
+    def test_trailing_comma_lookalike_inside_string_value_survives(self):
+        """`, }` inside a JSON string must not be eaten by trailing-comma cleanup (review LOW)."""
+        raw = (
+            '{"kesimpulan": "Kandang 2, } blok selatan aman dari insiden.", '
+            '"insight": "Tidak ada anomali lingkungan pada periode berjalan ini."}'
+        )
+        out = parse_llm_json(raw)
+        self.assertIsNotNone(out)
+        self.assertIn(", }", out["kesimpulan"])
+
+    def test_prose_without_sentence_split_not_duplicated(self):
+        """No `.!?` split: prose stored once as kesimpulan, insight must differ (review LOW)."""
+        raw = (
+            "FCR membaik ke 0,18 dan lebih baik dari standar CP 707 berkat "
+            "manajemen pakan yang konsisten ditinjau setiap hari tanpa kendala "
+            "berarti di kandang blok selatan periode berjalan ini"
+        )
+        out = parse_llm_json(raw)
+        self.assertIsNotNone(out)
+        self.assertIn("FCR membaik", out["kesimpulan"])
+        self.assertNotEqual(out["insight"], out["kesimpulan"])
+
 
 class TopicWhitelistTests(SimpleTestCase):
     def test_unknown_topic_rejected_before_any_work(self):