Upload folder using huggingface_hub

app.py

@@ -230,7 +230,17 @@ async def plan_trip(user_text: str) -> str:
             if res.type == "final_answer":
                 agent_text = res.text or ""
                 break
-            # fallback_to_deterministic → drop through to the deterministic path
+            # fallback_to_deterministic → EXPLICIT + user-visible degrade. Never
+            # silently swap in the deterministic path. Most commonly this is a
+            # Modal cold-start timeout (see the agent_loop reason) — tell the user
+            # honestly so the wait / fast-mode result is understood, not hidden.
+            if res.type == "fallback_to_deterministic":
+                yield _ev(
+                    type="commentary",
+                    text="🌡️ Nemotron is warming up on Modal (cold start) — "
+                         "building your packages in fast mode now, then I'll "
+                         "still compare them live.",
+                )
             break
 
     # ── Deterministic fallback (K3): parse intent + build directly. Used when

index.html

@@ -10,7 +10,7 @@
 <script src="https://unpkg.com/leaflet@1.9.4/dist/leaflet.js"></script>
 <link rel="preconnect" href="https://fonts.googleapis.com">
 <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
-<link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700;800;900&display=swap" rel="stylesheet">
+<link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700;800;900&family=Space+Grotesk:wght@500;600;700&display=swap" rel="stylesheet">
 
 <style>
   :root{
@@ -22,6 +22,10 @@
     --bg:#f7f8fa; --surface:#ffffff;
     --shadow-sm:0 1px 2px rgba(17,24,39,.04),0 2px 6px -2px rgba(17,24,39,.08);
     --shadow-md:0 1px 2px rgba(17,24,39,.04),0 8px 22px -12px rgba(17,24,39,.16);
+    /* FIFA World Cup 2026 inspired palette (vibrant poster accents on dark) */
+    --wc-magenta:#db2777; --wc-violet:#7c3aed; --wc-blue:#2563eb;
+    --wc-teal:#0d9488; --wc-gold:#fbbf24; --wc-pink:#f472b6;
+    --display:'Space Grotesk',Inter,-apple-system,sans-serif;
   }
   *{box-sizing:border-box;}
   html,body{margin:0;height:100%;}
@@ -62,10 +66,24 @@
 
   /* ── Chat column ─────────────────────────────────────── */
   #chat{flex:1;overflow-y:auto;padding:20px 18px 10px;display:flex;flex-direction:column;gap:15px;}
-  .hero{background:var(--surface);border:1px solid var(--line2);border-radius:16px;padding:17px 18px;
-    box-shadow:var(--shadow-sm);}
-  .hero h2{margin:0 0 7px;font-size:15.5px;font-weight:800;letter-spacing:-.01em;}
-  .hero p{margin:0;font-size:13px;color:var(--mut);line-height:1.55;}
+  /* FIFA 2026 themed hero — real stadium imagery on a vibrant gradient fallback
+     (looks premium even before the photo finishes loading or if it's blocked). */
+  .hero{position:relative;overflow:hidden;border:0;border-radius:18px;min-height:220px;
+    background:
+      url('https://images.unsplash.com/photo-1522778526097-ce0a22ceb253?w=1600&q=80') center/cover no-repeat,
+      linear-gradient(135deg,#1e1b4b 0%,#6d28d9 45%,#db2777 100%);
+    box-shadow:var(--shadow-md);display:flex;align-items:flex-end;}
+  .hero-overlay{position:absolute;inset:0;
+    background:linear-gradient(180deg,rgba(15,23,42,.18) 0%,rgba(15,23,42,.55) 52%,rgba(15,23,42,.93) 100%);}
+  .hero-content{position:relative;z-index:2;padding:20px 22px 21px;color:#fff;width:100%;}
+  .hero-badge{display:inline-block;font-size:10.5px;font-weight:800;letter-spacing:.09em;text-transform:uppercase;
+    color:#fff;background:linear-gradient(135deg,var(--wc-magenta),var(--wc-violet));
+    padding:5px 11px;border-radius:999px;margin-bottom:11px;box-shadow:0 4px 12px -3px rgba(219,39,119,.5);}
+  .hero h2{margin:0 0 8px;font-family:var(--display);font-size:25px;font-weight:700;line-height:1.08;
+    letter-spacing:-.025em;color:#fff;}
+  .hero h2 .hero-accent{background:linear-gradient(135deg,var(--wc-gold),var(--wc-pink));
+    -webkit-background-clip:text;background-clip:text;-webkit-text-fill-color:transparent;color:transparent;}
+  .hero p{margin:0;font-size:13px;color:rgba(255,255,255,.92);line-height:1.55;max-width:48ch;}
 
   .msg{display:flex;gap:10px;max-width:94%;animation:fade .3s ease;}
   .msg.user{align-self:flex-end;flex-direction:row-reverse;}
@@ -117,10 +135,11 @@
   .result{overflow-y:auto;background:var(--bg);}
   #result{padding:18px 20px 28px;}
   .empty{display:flex;flex-direction:column;align-items:center;justify-content:center;
-    height:100%;text-align:center;color:var(--mut2);padding:40px 30px;}
-  .empty .big{font-size:50px;margin-bottom:14px;filter:drop-shadow(0 6px 10px rgba(0,0,0,.08));}
-  .empty h3{margin:0 0 8px;font-size:17px;color:var(--ink2);font-weight:700;}
-  .empty p{margin:0;font-size:13.5px;max-width:340px;line-height:1.55;}
+    height:100%;text-align:center;padding:50px 30px;
+    background:radial-gradient(125% 90% at 50% 0%,#1e293b 0%,#0f172a 72%);color:#cbd5e1;}
+  .empty .big{font-size:66px;margin-bottom:16px;filter:drop-shadow(0 8px 26px rgba(56,189,248,.35));}
+  .empty h3{margin:0 0 9px;font-family:var(--display);font-size:19px;color:#fff;font-weight:700;letter-spacing:-.015em;}
+  .empty p{margin:0;font-size:13.5px;max-width:360px;line-height:1.6;color:#94a3b8;}
 
   /* ── "Building your trip" skeleton state (premium wait UX) ── */
   .building{padding:2px 0;}
@@ -199,8 +218,12 @@
     <section class="col chat">
       <div id="chat">
         <div class="hero">
-          <h2>👋 Plan your World Cup trip</h2>
-          <p>Tell me where you're flying from, the match you want, dates and budget. I'll build 3 ranked packages — cheapest flight, safest arrival, closest hotel to BC Place — with live prices and honest provenance.</p>
+          <div class="hero-overlay"></div>
+          <div class="hero-content">
+            <span class="hero-badge">⚽ FIFA World Cup 2026 · Vancouver</span>
+            <h2>Plan your <span class="hero-accent">World Cup</span> trip</h2>
+            <p>Tell me where you're flying from, the match, your dates &amp; budget. I'll build 3 ranked packages — cheapest flight, safest arrival, closest hotel to BC Place — with live prices &amp; honest provenance.</p>
+          </div>
         </div>
       </div>
 
@@ -225,8 +248,8 @@
       <div id="result">
         <div class="empty">
           <div class="big">🏟️</div>
-          <h3>Your trip packages will appear here</h3>
-          <p>Send your request and Nemotron picks the tools while Python scores real flights, hotels, weather and nearby spots — ranked on an interactive map.</p>
+          <h3>Your Vancouver 2026 packages appear here</h3>
+          <p>Send your request and Nemotron picks the tools while Python scores real flights, hotels, weather and nearby spots — ranked on an interactive map of BC Place.</p>
         </div>
       </div>
     </section>

matchday/agent.py

@@ -22,7 +22,12 @@ from typing import Any
 import modal
 
 from matchday.agent_loop import BuildTripPackagesArgs, ClarifyArgs, WebSearchArgs
-from matchday.errors import classify_error
+from matchday.errors import (
+    ClassifiedError,
+    FailoverReason,
+    classify_error,
+    user_message_for,
+)
 from matchday.prompts import build_system_prompt
 
 # The Modal app name (must match `modal.App("matchday-spike")` in modal_spike.py
@@ -121,9 +126,26 @@ class MatchDayAgent:
             # Re-raise so the AgentLoop can degrade gracefully (fallback to the
             # deterministic parser); the classified reason is logged here.
             classified = classify_error(exc)
+            # A timeout on the Nemotron generate call is the model still LOADING
+            # (Modal cold start ~90-150s), not a network fault. classify_error
+            # would tag it NETWORK_TIMEOUT ("check your connection") — wrong, and
+            # the reason callers use to decide on degradation. Re-tag as
+            # MODEL_COLD_START so the user gets an honest "warming up" message and
+            # the loop degrades to deterministic cleanly.
+            if isinstance(exc, TimeoutError):  # asyncio.TimeoutError ⊂ TimeoutError
+                classified = ClassifiedError(
+                    reason=FailoverReason.MODEL_COLD_START,
+                    message=f"Nemotron generate timed out (cold start?): {exc!r}",
+                    status_code=None,
+                    retryable=True,
+                    retry_after_seconds=None,
+                    should_degrade=True,
+                    user_message=user_message_for(FailoverReason.MODEL_COLD_START),
+                )
             logger.error(
-                "Nemotron generate failed: %s (reason=%s, retryable=%s, degrade=%s)",
-                exc, classified.reason.value, classified.retryable, classified.should_degrade,
+                "Nemotron generate failed: %s: %r (reason=%s, retryable=%s, degrade=%s)",
+                type(exc).__name__, exc, classified.reason.value,
+                classified.retryable, classified.should_degrade,
             )
             raise
 

matchday/agent_loop.py

@@ -24,6 +24,7 @@ from __future__ import annotations
 import asyncio
 import json
 import logging
+import os
 from dataclasses import dataclass, field
 from datetime import datetime, timezone
 from typing import Any, Literal
@@ -47,8 +48,23 @@ happy-path turns exit at 2-3 rounds (Nemotron calls build_trip_packages, then
 explains). 5 fits the hard path (4 tool rounds + 1 explain) with the grace call
 as a safety net; reference agents cap far higher (hermes 90, deepagents 9999)."""
 
-PER_TOOL_TIMEOUT_SECONDS: float = 30.0
-"""Timeout per individual tool execution."""
+PER_TOOL_TIMEOUT_SECONDS: float = float(os.environ.get("MATCHDAY_WEB_TIMEOUT", "30"))
+"""Deadline for a single fast tool call (e.g. one SerpApi ``web_search``)."""
+
+# Agent (Nemotron) inference call deadline. This MUST tolerate Modal COLD START:
+# loading the 60GB Nemotron weights from the volume cache takes ~90-150s after
+# the container scales up from zero. A 30s deadline here silently timed out the
+# FIRST agent round on every cold-started query -> the smart loop fell back to
+# the deterministic parser and Nemotron never actually decided / grounded /
+# clarified (the crux bug behind "the agent isn't smart"). Warm calls finish in
+# ~5-15s, well under this ceiling. Tune via MATCHDAY_AGENT_TIMEOUT.
+AGENT_CALL_TIMEOUT_SECONDS: float = float(os.environ.get("MATCHDAY_AGENT_TIMEOUT", "150"))
+"""Deadline for one Nemotron inference call (cold-start tolerant)."""
+
+# build_trip_packages fans out flights + hotels + weather + OSM (multiple SerpApi
+# round-trips) and can take 30-60s; the old shared 30s could cut it mid-build.
+TOOL_EXEC_TIMEOUT_SECONDS: float = float(os.environ.get("MATCHDAY_TOOL_TIMEOUT", "90"))
+"""Deadline for a multi-API tool such as ``build_trip_packages``."""
 
 MAX_RESULT_TOKENS: int = 2_000
 """Upper bound on the tool-result string fed back to the model. Tool results
@@ -705,6 +721,7 @@ class AgentLoop:
         round_num = 0
         tool_history: list[tuple[str, dict[str, Any]]] = []
         correction_used = False
+        cold_start_retried = False  # one retry tolerated for Modal cold start
 
         while round_num < self._max_rounds:
             round_num += 1
@@ -713,10 +730,41 @@ class AgentLoop:
             try:
                 response = await self._call_agent(messages)
             except Exception as exc:
-                logger.error("Agent call failed on round %d: %s", round_num, exc)
+                # Never log an empty str(exc) — asyncio.TimeoutError stringifies
+                # to "" (the blank "Agent call failed on round 1: " in the logs).
+                # Capture the type + repr so cold-start vs parse vs OOM is
+                # diagnosable. A TimeoutError here is a Modal COLD START (the 60GB
+                # model is still loading), NOT a network fault — flag it so the
+                # reason callers use to degrade is honest.
+                is_cold = isinstance(exc, TimeoutError)  # asyncio.TimeoutError ⊂ TimeoutError (3.11+)
+                logger.error(
+                    "Agent call failed on round %d: %s: %r%s",
+                    round_num, type(exc).__name__, exc,
+                    " [MODEL_COLD_START — Nemotron still loading]" if is_cold else "",
+                )
+                # Modal cold start has HIGH variance (90-250s observed — a single
+                # deadline can't cover it). But the first (timed-out) call already
+                # TRIGGERED the server-side model load; by the time we loop back,
+                # the container is usually warm. Retry round 1 ONCE before
+                # degrading — this is what turns a guaranteed cold-start failure
+                # into a successful first query.
+                if is_cold and round_num == 1 and not cold_start_retried:
+                    cold_start_retried = True
+                    logger.warning(
+                        "Round 1 cold-start timeout (>=%.0fs) — retrying once "
+                        "(container should be warm now).",
+                        AGENT_CALL_TIMEOUT_SECONDS,
+                    )
+                    continue
                 return AgentLoopResult(
                     type="fallback_to_deterministic",
-                    reason=f"Agent call failed on round {round_num}: {exc}",
+                    reason=(
+                        f"Nemotron cold-start timeout on round {round_num} "
+                        f"(>={AGENT_CALL_TIMEOUT_SECONDS:.0f}s)"
+                        if is_cold else
+                        f"Agent call failed on round {round_num}: "
+                        f"{type(exc).__name__}: {exc!r}"
+                    ),
                 )
 
             # ---- Step 2: Check for tool calls ----
@@ -833,7 +881,7 @@ class AgentLoop:
         try:
             return await asyncio.wait_for(
                 self._agent.run(messages),
-                timeout=PER_TOOL_TIMEOUT_SECONDS,
+                timeout=AGENT_CALL_TIMEOUT_SECONDS,
             )
         except asyncio.TimeoutError:
             raise  # re-raised so the caller handles it
@@ -928,14 +976,14 @@ class AgentLoop:
         try:
             tool_result = await asyncio.wait_for(
                 impl(validated_args),
-                timeout=PER_TOOL_TIMEOUT_SECONDS,
+                timeout=TOOL_EXEC_TIMEOUT_SECONDS,
             )
         except asyncio.TimeoutError:
             return AgentLoopResult(
                 type="fallback_to_deterministic",
                 reason=(
                     f"Tool {tool_name!r} timed out after "
-                    f"{PER_TOOL_TIMEOUT_SECONDS}s."
+                    f"{TOOL_EXEC_TIMEOUT_SECONDS}s."
                 ),
             )
         except Exception as exc:

matchday/render.py

@@ -11,6 +11,7 @@ from __future__ import annotations
 
 from html import escape
 from typing import Any
+from urllib.parse import quote as _urlquote
 
 from matchday.models import ScoredPackage
 from matchday.trip_tool import TripPackageResult
@@ -100,12 +101,17 @@ _CSS = """
 .md-facts .md-fbody{min-width:0;}
 .md-facts .md-fbody b{color:#111827;font-weight:700;display:block;}
 .md-facts .md-fbody .md-fdet{display:block;color:#6b7280;font-size:12.5px;margin-top:2px;}
-/* booking CTA (Layla checklist) — honest search link, never a fabricated URL */
+/* per-option booking CTAs (Layla-style): a real booking/directions button for
+   flight, hotel, and transit — not one generic web search. */
 .md-card-cta{padding:2px 22px 18px;}
+.md-cta-row{display:flex;flex-wrap:wrap;gap:8px;}
 .md-cta{display:inline-flex;align-items:center;gap:6px;text-decoration:none;font-size:13px;font-weight:700;
-  color:#fff;background:var(--accent,#7c3aed);padding:9px 16px;border-radius:11px;
-  box-shadow:0 6px 14px -6px rgba(124,58,237,.45);transition:.15s;}
+  color:#fff;background:var(--accent,#7c3aed);padding:9px 15px;border-radius:11px;
+  box-shadow:0 6px 14px -6px rgba(17,24,39,.35);transition:.15s;}
 .md-cta:hover{filter:brightness(1.07);transform:translateY(-1px);}
+.md-cta-f{background:#1d4ed8;}          /* flight — sky blue */
+.md-cta-h{background:#047857;}          /* hotel — emerald */
+.md-cta-t{background:#374151;}          /* transit — slate */
 
 /* ── Map ── */
 .md-map-wrap{position:relative;border:1px solid #eef0f3;border-radius:18px;overflow:hidden;
@@ -231,11 +237,34 @@ def render_card(pkg: ScoredPackage, is_top: bool = False) -> str:
 
     topflag = '<span class="md-topflag">★ Best match</span>' if is_top else ""
 
-    # Honest booking CTA: a real Google search for this exact route + date
-    # (never a fabricated booking URL — provenance stays clean).
-    cta_url = (
-        f"https://www.google.com/search?q=flights+{f.origin}+to+{f.destination}+"
-        f"{f.arrival_time:%B}+{f.arrival_time.day}+2026"
+    # Per-option booking CTAs (Layla-style). Hotels carry a REAL SerpApi deep
+    # link (Hotel.booking_url from Google Hotels). Flights expose no resolved
+    # booking URL (booking_token needs a deferred SerpApi call), so we deep-link
+    # to Skyscanner for the exact route+date — a real booking site, not a generic
+    # web search. Transit links to Google Maps directions to BC Place.
+    fdate = f.arrival_time.strftime("%Y-%m-%d")
+    flight_url = (
+        f"https://www.skyscanner.com/transport/flights/{f.origin.lower()}/"
+        f"{f.destination.lower()}/{fdate}/"
+    )
+    if h and getattr(h, "booking_url", None):
+        hotel_url, hotel_label = h.booking_url, "🏨 Book hotel"
+    elif h:
+        hotel_url = (
+            "https://www.google.com/travel/hotels?q="
+            + _urlquote(f"{h.name} Vancouver BC Place")
+        )
+        hotel_label = "🏨 Find hotel"
+    else:
+        hotel_url = hotel_label = None
+    transit_url = (
+        "https://www.google.com/maps/dir/?api=1&destination=BC%20Place%20"
+        "Stadium%2C%20Vancouver&travelmode=transit"
+    )
+    hotel_btn = (
+        f'<a class="md-cta md-cta-h" href="{_e(hotel_url)}" target="_blank" '
+        f'rel="noopener noreferrer">{hotel_label} →</a>'
+        if hotel_url else ""
     )
 
     return f"""
@@ -255,7 +284,11 @@ def render_card(pkg: ScoredPackage, is_top: bool = False) -> str:
         <li><span class="md-fico">🚶</span><span class="md-fbody"><b>{pkg.hotel_to_stadium_min} min walk</b>
           <span class="md-fdet">hotel → BC Place · {len(pkg.amenities)} nearby spots</span></span></li>
       </ul>
-      <div class="md-card-cta"><a class="md-cta" href="{cta_url}" target="_blank" rel="noopener noreferrer">✈️ Book this flight →</a></div>
+      <div class="md-card-cta"><div class="md-cta-row">
+        <a class="md-cta md-cta-f" href="{flight_url}" target="_blank" rel="noopener noreferrer">✈️ Book flight →</a>
+        {hotel_btn}
+        <a class="md-cta md-cta-t" href="{transit_url}" target="_blank" rel="noopener noreferrer">🚖 Transit →</a>
+      </div></div>
     </article>
     """