Upload folder using huggingface_hub

app.py

@@ -45,7 +45,7 @@ from matchday.agent_trace import (  # noqa: E402
     ranking_from_result,
     result_source_labels,
 )
-from matchday.intent import parse_intent  # noqa: E402
+from matchday.intent import parse_intent, _find_match  # noqa: E402
 from matchday.models import TripRequest  # noqa: E402
 from matchday.wc2026 import resolve_match  # noqa: E402
 from matchday.prompts import EXPLANATION_HINT  # noqa: E402
@@ -198,6 +198,38 @@ def _clean_match_name(name: str) -> str:
     return f"{_strip(a)} vs {_strip(b)}"
 
 
+_DEFAULT_GREETING = (
+    "I'd love to plan your FIFA 2026 World Cup trip! Tell me where you're "
+    "flying from (e.g. 'Montreal' or 'YUL'), which match you'd like to see, "
+    "and the dates."
+)
+
+
+def _precheck_chitchat(user_text: str):
+    """Deterministic reply for pure chit-chat / empty prompts, BEFORE the agent.
+
+    A greeting or content-free message ("hi", "hello", "thanks", "test") has no
+    origin, date, or match to plan around. Replying with ``parse_intent``'s
+    clarifying question deterministically turns what would be a multi-second to
+    multi-minute Modal cold-start wait (for a Nemotron call that would only
+    clarify anyway) into an instant answer. Same pre-agent seam as the fixture
+    validator — no agent_loop / Modal change.
+
+    Conservative: fires ONLY when BOTH origin and date are absent AND no 'X vs Y'
+    match is named, so any real (even partial) trip request still reaches the
+    fixture validator / agent. Returns ``(reply, missing_slots)`` for chit-chat,
+    or ``None`` to proceed normally. Never raises.
+    """
+    try:
+        parsed = parse_intent(user_text)
+    except Exception:  # noqa: BLE001 — must never break the turn
+        return None
+    # `missing` only ever holds origin and/or date; len >= 2 => both absent.
+    if len(parsed.missing) >= 2 and not _find_match(user_text):
+        return parsed.question or _DEFAULT_GREETING, list(parsed.missing)
+    return None
+
+
 def _finalize_trace(trace: AgentTrace, trip, result, built_by: str) -> None:
     """Populate the final intent/grounding/evidence/ranking/outcome on the trace.
 
@@ -315,6 +347,27 @@ async def plan_trip(user_text: str) -> str:
     yield _ev(type="progress", step="read", status="done", text="Read your request")
     yield _ev(type="progress", step="extract", status="running", text="Understanding your trip")
 
+    # ── Pre-agent chit-chat guard: a greeting / empty prompt ("hi", "thanks")
+    # has no origin, date, or match. Reply deterministically and instantly
+    # instead of waking Nemotron for a call that would only clarify — turns a
+    # Modal cold-start wait into an immediate answer. Conservative: only fires
+    # when nothing trip-related was said, so real (even partial) requests still
+    # reach the fixture validator / agent.
+    _chat = _precheck_chitchat(user_text)
+    if _chat is not None:
+        _chat_text, _chat_missing = _chat
+        trace.set_intent(None, missing=_chat_missing)
+        trace.set_outcome(
+            mode="deterministic", status="clarify",
+            notes=["Pre-agent chit-chat check: no trip details (origin / date / match) yet."],
+            model=_TRACE_MODEL, rounds=0,
+        )
+        yield _ev(type="trace", data=trace.to_dict())
+        yield _ev(type="progress", step="extract", status="done", text="Heard you")
+        yield _ev(type="progress", step="ready", status="fallback", text="Tell me your trip")
+        yield _ev(type="clarify", text=_chat_text)
+        return
+
     # ── Generic pre-agent fixture validator (grounding honesty). Ground the named
     # match deterministically BEFORE the agent picks a tool: a non-real 2026
     # fixture is refused with the closest real alternatives and we stop, so the

matchday/modal_spike.py

@@ -99,7 +99,7 @@ app = modal.App("matchday-spike")
     volumes={HF_CACHE_PATH: HF_CACHE_VOL},
     secrets=[hf_secret],
     startup_timeout=120 * MINUTES,  # time to download + load 60GB model
-    scaledown_window=300,
+    scaledown_window=1800,  # keep the replica warm 30 min after each request so an active demo stays hot (NOT min_containers — no 24/7 billing)
 )
 class Server:
     @modal.enter()