fix: upgrade gemini model string to 2.5-flash-lite + add tom diagnostic script

.cursorrules

@@ -11,9 +11,9 @@ NO Anthropic API anywhere. NO npm. NO build step.
 
 ## LLM Client Rules (Gemini)
 
-- Model: `gemini-2.5-flash` everywhere. Never use other model names in game/agent code.
+- Model: `gemini-2.5-flash-lite` everywhere. Never use other model names in game/agent code.
 - Import: `from google import genai` and `from google.genai import types` — never `anthropic` or legacy `google.generativeai`.
-- Client: `genai.Client(api_key=os.environ.get("GOOGLE_API_KEY", ""))`; chats via `client.chats.create(model="gemini-2.5-flash", ...)`.
+- Client: `genai.Client(api_key=os.environ.get("GOOGLE_API_KEY", ""))`; chats via `client.chats.create(model="gemini-2.5-flash-lite", ...)`.
 - ALL Gemini calls wrapped in try/except returning `SYNTHETIC_RESPONSE` on failure.
 - JSON extraction: always strip markdown fences before `json.loads()`.
 - Async Gemini calls: use `asyncio.get_event_loop().run_in_executor(None, lambda: ...)`.

agent/gemini_client.py

@@ -42,7 +42,7 @@ SCENARIO_ROLE_CONTEXT: dict[str, dict[str, str]] = {
     },
 }
 
-GEMINI_MODEL = "gemini-2.5-flash"
+GEMINI_MODEL = "gemini-2.5-flash-lite"
 # Aliases for imports (dashboard, MCP, training all use flash-lite)
 MODEL_ID_DEMO = GEMINI_MODEL
 MODEL_ID_DATA = GEMINI_MODEL

scripts/diagnose_tom.py

@@ -0,0 +1,130 @@
+"""
+One-episode ToM belief diagnostic. Run from repo root:
+  - Put GOOGLE_API_KEY in .env, or: $env:GOOGLE_API_KEY="..."  # PowerShell
+  python scripts/diagnose_tom.py
+"""
+from __future__ import annotations
+
+import asyncio
+import copy
+import json
+import os
+import sys
+from pathlib import Path
+
+# Repo root on sys.path when launched as: python scripts/diagnose_tom.py
+_ROOT = Path(__file__).resolve().parent.parent
+if str(_ROOT) not in sys.path:
+    sys.path.insert(0, str(_ROOT))
+
+# Load .env from project root (same as main.py) so GOOGLE_API_KEY is available
+# when set in .env but not exported in the shell.
+try:
+    from dotenv import load_dotenv
+
+    load_dotenv(_ROOT / ".env")
+except ImportError:
+    pass
+
+from parlay_env.grader import _tom_accuracy
+from parlay_env.models import BeliefState
+from parlay_env.reward import BETA
+
+from dashboard.api import (  # noqa: E402
+    MoveRequest,
+    _build_observation,
+    _build_session,
+    _sessions,
+    make_move,
+)
+
+UTTERANCE_ODD = (
+    "I think $155,000 reflects fair value here. We'd like to move forward."
+)
+UTTERANCE_EVEN = (
+    "We can come down to $148,000 but that's near our floor."
+)
+
+
+def _tom_reward(belief: BeliefState, hidden) -> float:
+    return BETA * _tom_accuracy(belief, hidden)
+
+
+async def _run() -> None:
+    key = (os.environ.get("GOOGLE_API_KEY") or "").strip()
+    if not key:
+        print("ERROR: GOOGLE_API_KEY not set", file=sys.stderr)
+        raise SystemExit(1)
+
+    sid, session = _build_session("saas_enterprise", "diplomat", "ToM-Diagnose")
+    _sessions[sid] = session
+    tom = session["tom_tracker"]
+    state0 = session["state"]
+
+    initial = tom.current_belief
+    snapshots: list[BeliefState] = [copy.deepcopy(initial)]
+
+    init_obs = _build_observation(state0)
+    print(json.dumps(init_obs, indent=2, default=str))
+
+    for n in range(1, 9):
+        if n % 2 == 1:
+            utterance, amount = UTTERANCE_ODD, 155_000.0
+        else:
+            utterance, amount = UTTERANCE_EVEN, 148_000.0
+        result = await make_move(
+            MoveRequest(
+                session_id=sid,
+                amount=amount,
+                message=utterance,
+                tactical_move=None,
+            )
+        )
+        session = _sessions[sid]
+        st = session["state"]
+        tom = session["tom_tracker"]
+        b = tom.current_belief
+        snapshots.append(copy.deepcopy(b))
+
+        tr = _tom_reward(b, st.hidden_state)
+        print(
+            f"     [Turn {n}] belief_budget={b.est_budget:.3f}  "
+            f"belief_urgency={b.est_urgency:.3f}  "
+            f"belief_walkaway={b.est_walk_away:.3f}  "
+            f"tom_reward={tr:.4f}"
+        )
+        if result.get("done"):
+            break
+
+    s0, s1 = snapshots[0], snapshots[-1]
+    total_move = 0.0
+    for i in range(1, len(snapshots)):
+        a, snap_b = snapshots[i - 1], snapshots[i]
+        total_move += abs(snap_b.est_budget - a.est_budget)
+        total_move += abs(snap_b.est_urgency - a.est_urgency)
+        total_move += abs(snap_b.est_walk_away - a.est_walk_away)
+
+    print()
+    print("   === ToM Diagnostic Summary ===")
+    print(
+        f"   Initial beliefs: budget={s0.est_budget:.3f}  "
+        f"urgency={s0.est_urgency:.3f}  walkaway={s0.est_walk_away:.3f}"
+    )
+    print(
+        f"   Final beliefs:   budget={s1.est_budget:.3f}  "
+        f"urgency={s1.est_urgency:.3f}  walkaway={s1.est_walk_away:.3f}"
+    )
+    print(f"   Total belief movement: {total_move:.4f}")
+    if total_move > 0.05:
+        msg = "BELIEFS ARE MOVING — ToM reward is live"
+    else:
+        msg = "WARNING: beliefs stuck — ToM reward contributing ~0 to training signal"
+    print(f"   RESULT: {msg}")
+
+
+def main() -> None:
+    asyncio.run(_run())
+
+
+if __name__ == "__main__":
+    main()