Force a news search for trading-session questions

lib/llm/gate.py

@@ -34,6 +34,9 @@
 LIVE_WORDS = re.compile(
     r"\b(news|today|latest|yesterday|tonight|this week|right now|breaking|announce\w*|who won|"
     r"score|weather|happening|headlines?|current\w*|updates?|recent\w*|just (now|in))\b", re.I)
+# Asking about a trading session ("what happened at close"): answers must be fetched, not recalled.
+SESSION_WORDS = re.compile(
+    r"\b(?:(?:at|the|into|after) (?:the )?(?:close|open)|pre-?market|after[- ]hours|closing|opening)\b", re.I)
 
 
 @dataclass(frozen=True)
@@ -65,6 +68,10 @@ def is_portfolio_question(text: str, st: Settings) -> bool:
     return bool(PORTFOLIO_WEAK.search(text)) and not _has_ticker(text)
 
 
+def is_session_question(text: str) -> bool:
+    return bool(SESSION_WORDS.search(text))
+
+
 def route(text: str, st: Settings, registry: Registry, *, force_full: bool = False) -> Route:
     """Tools for this message. With TOKEN_SAVER off, or `force_full` (a search/data request such
     as a movers list, a preset button or a retry after NEEDS_TOOLS), everything is offered."""
@@ -73,7 +80,7 @@ def route(text: str, st: Settings, registry: Registry, *, force_full: bool = Fal
         return Route(up, st.search)
     portfolio = is_portfolio_question(text, st)
     market = portfolio or bool(MARKET_WORDS.search(text)) or _has_ticker(text)
-    live = bool(LIVE_WORDS.search(text))
+    live = is_session_question(text) or bool(LIVE_WORDS.search(text))
     servers = [s for s in up if (portfolio if s.cfg.get("gate") == "portfolio" else market)]
     search = st.search and (market or live)
     simple = not (servers or search)

lib/prompts.py

@@ -69,7 +69,7 @@ def down_note(down: list[MCPServer]) -> str:
 
 def chat_prompt(*, transcript: str, sender: str, private: bool, reply_quote: str | None,
                 msg_id: int, now: str, down: str = "", saver: bool = True, middle: str = "",
-                fresh: bool = False) -> str:
+                fresh: bool = False, session: bool = False) -> str:
     """The user turn: transcript, the replied-to message, and the volatile tail (kept last so the
     cached prefix before it is unchanged)."""
     head = (f"Private one-to-one chat with {sender}, not the group. Transcript:\n{transcript}"
@@ -89,6 +89,9 @@ def chat_prompt(*, transcript: str, sender: str, private: bool, reply_quote: str
                  "for anything else asked, such as prices or a comparison with market data); don't reuse "
                  "numbers or refusals from earlier replies. When comparing portfolios, or asked \"why\", "
                  "write 2-3 short paragraphs on the tailwinds and headwinds of each portfolio, not a summary.")
+    if session:  # a trading-session question: news and reasons in earlier replies are stale by now
+        tail += (" Search for current news before answering, and say why each stock moved. Don't reuse "
+                 "news or reasons from earlier messages in the transcript.")
     return head + tail + down
 
 

lib/telegram/handlers.py

@@ -18,7 +18,7 @@
 from ..features import dm_buttons, holding_news, movers, periods, post
 from ..history import compact, format_rows
 from ..llm.base import BUSY_REPLY, is_busy
-from ..llm.gate import Route, is_portfolio_question, route
+from ..llm.gate import Route, is_portfolio_question, is_session_question, route
 from ..msgtext import describe, sender_name
 from ..prompts import MOVERS_EXPLAIN, PRESET_PROMPT, chat_prompt, down_note
 from ..store import is_dm
@@ -179,15 +179,17 @@ async def _gate_text(self, msg: Message, trig: Trigger) -> str:
                     text += " " + compact(prev.text)[:FOLLOW_UP_CHARS]
         return text
 
-    async def _chat_text(self, msg: Message, trig: Trigger, middle: str = "", fresh: bool = False) -> str:
+    async def _chat_text(self, msg: Message, trig: Trigger, middle: str = "", fresh: bool = False,
+                         session: bool = False) -> str:
         ctx, st = self.ctx, self.ctx.st
         rows = await asyncio.to_thread(ctx.store.history, msg.chat_id, st.history_limit)
         transcript = format_rows(rows, ctx.self_name, st.tz, line_max=st.history_line_max,
                                  compact_text=st.token_saver, own_line_max=st.history_own_line_max)
         return chat_prompt(
             transcript=transcript, sender=sender_name(msg), private=trig.private,
             reply_quote=self._quote(trig), msg_id=msg.message_id, now=ctx.now(),
-            down=down_note(ctx.registry.down()), saver=st.token_saver, middle=middle, fresh=fresh)
+            down=down_note(ctx.registry.down()), saver=st.token_saver, middle=middle, fresh=fresh,
+            session=session)
 
     # -- answering ---------------------------------------------------------------------
     async def _answer(self, msg: Message, trig: Trigger) -> None:
@@ -257,9 +259,11 @@ async def _generate(self, msg: Message, trig: Trigger, draft: Draft | None) -> s
                 if chosen.simple:  # the model must not be told it has no tools
                     chosen = Route([], False)
             fresh = is_portfolio_question(context, st) and any(s.cfg.get("gate") == "portfolio" for s in chosen.servers)
-            parts = [{"type": "text", "text": await self._chat_text(msg, trig, fresh=fresh)}]
+            session = is_session_question(context) and not chosen.simple
+            parts = [{"type": "text", "text": await self._chat_text(msg, trig, fresh=fresh, session=session)}]
             parts += await self._images(msg, trig)
             answer = await ask(ctx, parts, chosen,
+                               must_search=session and chosen.search,  # a session recap must not be recalled
                                cache_key=f"chat-{chat_id}", on_text=on_text, chat_id=chat_id, extra_tools=extra,
                                admin=admin)
         return format_answer(answer.text, movers_bots=st.movers_bots, summary=trig.movers)

tests/test_handlers.py

@@ -90,6 +90,14 @@ async def test_partial_route_that_needs_more_is_asked_again_with_everything(env,
     assert "NEEDS_TOOLS" in backend.seen[0]["system"]  # the partial route told the model how to ask
 
 
+async def test_session_question_forces_a_search(env, store):
+    ctx, backend, h = make_ctx(env, store, [step("fresh recap", searches=1)], servers=[FakeMcp("yahoo")])
+    m = user_msg(ctx.bot, "@stockbot what happened at close?")
+    await run(h, m)
+    assert backend.seen[0]["choice"] == "required" and backend.seen[0]["search"] is True
+    assert "Search for current news" in backend.seen[0]["prompt"]
+
+
 async def test_market_question_offers_tools_up_front(env, store):
     mcp = FakeMcp("yahoo")
     ctx, backend, h = make_ctx(env, store, [step("ok")], servers=[mcp])

tests/test_runner.py

@@ -242,6 +242,22 @@ def test_gate_live_question_gets_search_only(settings):
     assert r.search and not r.servers
 
 
+@pytest.mark.parametrize("text", ["what happened at close?", "how did it go at the open", "any pre-market moves"])
+def test_gate_market_session_words_are_live(text, settings):
+    y, s = servers()
+    r = route(text, settings, registry(y, s))
+    assert r.search and not r.simple
+
+
+def test_session_question_prompt_demands_current_news():
+    from lib.llm.gate import is_session_question
+    from lib.prompts import chat_prompt
+    assert is_session_question("what happened at close?") and not is_session_question("hello")
+    kw = dict(transcript="t", sender="A", private=False, reply_quote=None, msg_id=1, now="now")
+    assert "Search for current news" in chat_prompt(**kw, session=True)
+    assert "Search for current news" not in chat_prompt(**kw)
+
+
 def test_gate_full_when_forced_or_token_saver_off(settings, env):
     y, s = servers()
     assert route("hello", settings, registry(y, s), force_full=True).servers == [y, s]