fix(rating): read a free-text rating from the decision's own rating line (#1383)

- the first line opening with a rating label is the call, not a rating the text quotes later ("Consensus rating: Buy")
- "rating" must start a word, so "Operating margin: Sell-side" is not a label

tests/test_rating_integrity.py

@@ -218,3 +218,29 @@ def test_a_state_without_the_typed_rating_reads_it_from_the_decision():
     assert run_rating({"final_rating": "Hold", "final_trade_decision": "**Rating**: Buy"}) == "Hold"
     assert run_rating({"final_trade_decision": "**Rating**: Sell\n\nExit."}) == "Sell"
     assert run_rating({}) == RATING_REVIEW
+
+
+@pytest.mark.unit
+@pytest.mark.parametrize("quoted", [
+    "Street consensus rating: Buy (28 of 35 analysts).",
+    "Moody's affirmed the credit rating: Buy-side demand for the bonds stayed firm.",
+    "Operating margin: Sell-side estimates sit below guidance.",
+])
+def test_a_rating_the_text_quotes_does_not_replace_the_decision(quoted):
+    """A free-text decision opens with its own rating line; a rating it quotes
+    as evidence, or a word merely ending in 'rating', is not the call."""
+    text = f"**Rating**: Hold\n\n**Investment Thesis**: {quoted} We wait for margins."
+    assert extract_rating(text) == "Hold"
+
+
+@pytest.mark.unit
+@pytest.mark.parametrize("quoted", [
+    "- Rating: Buy (Goldman Sachs, 12m target 180)",
+    "| Rating: Buy | Morgan Stanley |",
+    "> Rating: Buy, per the sell-side note",
+    "Street consensus rating: Buy",
+    "Consensus rating: Buy (28 of 35 analysts)",
+])
+def test_a_quoted_rating_in_a_list_table_or_quote_is_not_the_decision(quoted):
+    text = f"Our rating: Hold\n\nWhat others say:\n{quoted}\n\nWe wait for margins."
+    assert extract_rating(text) == "Hold"

tradingagents/agents/rating.py

@@ -32,10 +32,20 @@ RATING_REVIEW = "REVIEW"
 _RATING_SET = {r.lower() for r in RATINGS_5_TIER}
 
 # Matches "Rating: X" / "rating - X" / "Rating — **X**" — tolerates markdown
-# bold wrappers and any dash or colon a model writes as the separator.
+# bold wrappers and any dash or colon a model writes as the separator. "rating"
-_RATING_LABEL_RE = re.compile(r"rating\b[^:\-\u2010-\u2015]*[:\-\u2010-\u2015][\s*]*(\w+)",
+# must start a word, so "Operating margin: Sell-side" is not a label.
+_RATING_LABEL_RE = re.compile(r"(?<![a-z])rating\b[^:\-\u2010-\u2015]*[:\-\u2010-\u2015][\s*]*(\w+)",
                               re.IGNORECASE)
 
+# The same label opening its own line ("**Rating**: X", "## Final Rating - X",
+# "Our rating: X"): the shape the Portfolio Manager is asked to write its
+# decision in. Only emphasis and heading marks may precede it, so a list item,
+# table row or blockquote quoting someone else's rating is not one.
+_RATING_LINE_RE = re.compile(
+    r"[\s*_#]*(?:\w+\s+)?rating[^\w:\-\u2010-\u2015]*[:\-\u2010-\u2015][\s*]*(\w+)",
+    re.IGNORECASE,
+)
+
 # A line presenting the scale rather than a decision ("Rating Scale: Buy, ...").
 _RATING_SCALE_RE = re.compile(r"rating\s*(scale|options|legend)", re.IGNORECASE)
 
@@ -50,25 +60,31 @@ def extract_rating(text: str) -> str | None:
 
     Two-pass strategy on the NFKC-normalized text (so fullwidth punctuation like
     ``Rating：Overweight`` is matched the same as ASCII):
-    1. An explicit "Rating: X" label (tolerant of markdown bold).
+    1. An explicit "Rating: X" label (tolerant of markdown bold): the first one
-    2. The first standalone 5-tier rating word found anywhere.
+       opening its own line, else the last one anywhere.
+    2. A single 5-tier rating word, when the text names only one.
     """
     if not text:
         return None
     norm = unicodedata.normalize("NFKC", text)
 
-    # The labelled rating, taking the last one written: a decision states its
+    # A decision is asked to open with its rating on its own line, so the first
-    # rating after discussing the alternatives. Lines presenting the scale
+    # such line is the call; later ones may quote someone else's ("Consensus
-    # itself are a legend the model echoed, not a call.
+    # rating: Buy"). Without one, the last label anywhere wins: prose states its
-    labelled = None
+    # rating after discussing the alternatives. Lines presenting the scale itself
+    # are a legend the model echoed, not a call.
+    on_own_line = anywhere = None
     for line in norm.splitlines():
         if _RATING_SCALE_RE.search(line):
             continue
+        m = _RATING_LINE_RE.match(line)
+        if on_own_line is None and m and m.group(1).lower() in _RATING_SET:
+            on_own_line = m.group(1).capitalize()
         m = _RATING_LABEL_RE.search(line)
         if m and m.group(1).lower() in _RATING_SET:
-            labelled = m.group(1).capitalize()
+            anywhere = m.group(1).capitalize()
-    if labelled:
+    if on_own_line or anywhere:
-        return labelled
+        return on_own_line or anywhere
 
     # No label. A single rating word in the text is the call; several are an
     # argument, and picking one of them reports a direction nobody decided --