fix: natural conversation flow — greet before querying DB

- Add premature_query_penalty (-0.15) when agent issues a DB query as
its first action before greeting the customer; wired into both
compute_step_reward and compute_hierarchy_reward, and into
_handle_query_action which builds its own Reward directly
- Update system prompts (SYSTEM_PROMPT, SUPPORT_AGENT_PROMPT in
inference.py; SUPPORT_AGENT_PROMPT in frontend ai-action route) to
mandate greet → gather info → query flow; remove old "query the DB
first" instruction that drove robotic behaviour
- Fix inference.py missing X-API-Key header on all httpx calls (401s
on reset/step); add _ENV_HEADERS from ENV_API_KEY env var
- Fix frontend Dockerfile healthcheck: wget resolves localhost to ::1
on BusyBox but Next.js only binds IPv4; switch to 127.0.0.1 and
target /demo (which returns 200, not the 307 that / redirects from)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

env/environment.py

@@ -260,15 +260,21 @@ class CustomerSupportEnv:
         # Minimal reward — actual signal computed in reward_engine via DB signals.
         # Using _clamp_db_total so spam queries can't push the reward above the
         # normal cap, and so wasted-query penalties are bounded.
-        from env.reward_engine import compute_db_signals, _clamp_db_total
+        from env.reward_engine import compute_db_signals, _clamp_db_total, compute_premature_query_penalty
         db_signals = compute_db_signals(action, self._ticket, self._history, self._retrieved_data)
         raw_db = _clamp_db_total(db_signals)
+        premature_penalty = compute_premature_query_penalty(action, self._history)
         import numpy as np
         reward = Reward(
-            value=float(np.clip(0.5 + raw_db, 0.0, 1.0)),
+            value=float(np.clip(0.5 + raw_db + premature_penalty, 0.0, 1.0)),
             resolution_score=0.0, tone_score=0.5,
             efficiency_score=0.0, accuracy_score=0.0,
-            breakdown={"db_signals": db_signals, "is_terminal": False, "action_type": at.value},
+            breakdown={
+                "db_signals": db_signals,
+                "premature_query_penalty": premature_penalty,
+                "is_terminal": False,
+                "action_type": at.value,
+            },
         )
 
         self._action_log.append({

env/reward_engine.py

@@ -215,6 +215,27 @@ def compute_contradiction_penalty(action: Action, history: List[Message]) -> flo
     return -0.15 if (prev_claimed_done and now_asking_info) else 0.0
 
 
+def compute_premature_query_penalty(action: Action, history: List[Message]) -> float:
+    """
+    -0.15 if the agent fires a DB query before ever greeting the customer.
+
+    A professional support agent always acknowledges the customer first.
+    Querying the DB as the opening move is robotic, hurts empathy scores, and
+    signals the model is pattern-matching identifiers rather than conversing.
+
+    Only triggers when there are zero prior agent messages in the history, so
+    legitimate second-or-later queries (after a greeting) are never penalized.
+    """
+    if action.action_type not in (
+        ActionType.QUERY_USER_PROFILE, ActionType.QUERY_ORDER_DETAILS
+    ):
+        return 0.0
+    prior_agent = [m for m in history if m.role == "agent" and m.content.strip()]
+    if not prior_agent:
+        return -0.15
+    return 0.0
+
+
 def compute_keyword_stuffing_penalty(message: str) -> float:
     """
     Detect and penalize keyword stuffing.
@@ -449,6 +470,7 @@ def compute_step_reward(
     loop_penalty = compute_loop_penalty(history)
     contradiction_penalty = compute_contradiction_penalty(action, history)
     stuffing_penalty = compute_keyword_stuffing_penalty(tone_msg)
+    premature_query_penalty = compute_premature_query_penalty(action, history)
 
     if is_terminal:
         resolution_score = compute_resolution_score(action, ticket, history)
@@ -481,6 +503,7 @@ def compute_step_reward(
         + escalation_penalty
         + stuffing_penalty
         + info_gathering_bonus
+        + premature_query_penalty
         + db_total
     )
 
@@ -499,6 +522,7 @@ def compute_step_reward(
             "escalation_penalty": escalation_penalty,
             "keyword_stuffing_penalty": stuffing_penalty,
             "info_gathering_bonus": info_gathering_bonus,
+            "premature_query_penalty": premature_query_penalty,
             **{f"db_{k}": v for k, v in db_signals.items()},
             "is_terminal": is_terminal,
         },
@@ -539,6 +563,7 @@ def compute_hierarchy_reward(
     loop_penalty = compute_loop_penalty(history)
     contradiction_penalty = compute_contradiction_penalty(action, history)
     stuffing_penalty = compute_keyword_stuffing_penalty(tone_msg)
+    premature_query_penalty = compute_premature_query_penalty(action, history)
     efficiency_score = compute_efficiency_score(steps_used, max_steps)
     accuracy_score = compute_accuracy_score(history, ticket)
 
@@ -666,6 +691,7 @@ def compute_hierarchy_reward(
             + escalation_penalty
             + ignored_feedback_penalty
             + unnecessary_manager_penalty
+            + premature_query_penalty
             + db_total
         )
     else:
@@ -681,6 +707,7 @@ def compute_hierarchy_reward(
             + stuffing_penalty
             + ignored_feedback_penalty
             + unnecessary_manager_penalty
+            + premature_query_penalty
             + db_total
         )
 
@@ -754,6 +781,7 @@ def compute_hierarchy_reward(
             "keyword_stuffing_penalty": stuffing_penalty,
             "ignored_feedback_penalty": ignored_feedback_penalty,
             "unnecessary_manager_penalty": unnecessary_manager_penalty,
+            "premature_query_penalty": premature_query_penalty,
             **{f"db_{k}": v for k, v in db_signals.items()},
             "is_terminal": is_terminal,
             "role": role,

frontend/Dockerfile

@@ -45,6 +45,6 @@ ENV PORT=3000
 ENV HOSTNAME=0.0.0.0
 
 HEALTHCHECK --interval=30s --timeout=10s --start-period=20s --retries=3 \
-    CMD wget -qO- http://localhost:3000/ > /dev/null 2>&1 || exit 1
+    CMD wget -qO- http://127.0.0.1:3000/demo > /dev/null 2>&1 || exit 1
 
 CMD ["node", "server.js"]

frontend/src/app/api/ai-action/route.ts

@@ -19,11 +19,20 @@ ACTION TYPES — output exactly one per step:
 - "query_user_profile"  → look up customer DB (internal)    → requires: "email"
 - "query_order_details" → look up order DB (internal)       → requires: "order_id"
 
+CONVERSATION FLOW — MANDATORY ORDER:
+1. FIRST action MUST be "respond": greet the customer warmly and acknowledge their specific issue.
+2. THEN use "request_info" if you still need details before you can help.
+3. ONLY THEN query the DB — but only when the customer has confirmed the email or order ID
+   in this conversation AND you need that data to answer accurately.
+4. Respond with grounded facts, then "close" or "escalate".
+
+CRITICAL DB QUERY RULES:
+- NEVER query the DB as your very first action, even if the ticket already shows an email or order ID.
+  Always greet and acknowledge first. Querying without greeting feels robotic and is penalized.
+- After querying, cite ONLY values from KNOWN DATA or the customer's own messages — never invent facts.
+- If supervisor gave feedback, INCORPORATE it into your next action.
+
 SCORING: Empathy(30%) + Accuracy(25%) + Resolution(25%) + Efficiency(20%)
-Be warm, gather info from "Unresolved issues", use specific resolution language.
-If supervisor gave feedback, INCORPORATE it into your next action.
-If the ticket references an email or order-id you haven't looked up yet, query
-the DB first — never invent facts not present in KNOWN DATA or the conversation.
 
 OUTPUT FORMAT — return ONLY this JSON, no code fences, no explanation:
 {"action_type": "respond", "message": "..."} or {"action_type": "escalate", "reason": "..."}`;

inference.py

@@ -27,6 +27,8 @@ INFERENCE_MODEL   = (
     or "unsloth/Qwen2.5-1.5B-Instruct"
 )
 ENV_URL    = os.getenv("ENV_URL", "http://localhost:7860")
+ENV_API_KEY = os.getenv("ENV_API_KEY", os.getenv("API_KEY", "meta_hack_2026"))
+_ENV_HEADERS = {"X-API-Key": ENV_API_KEY}
 BENCHMARK  = "customer-support-env"
 TASKS           = ["easy", "medium", "hard"]
 HIERARCHY_TASKS = ["hierarchy_easy", "hierarchy_medium", "hierarchy_hard"]
@@ -158,6 +160,20 @@ ACTION TYPES — output exactly one per step:
 - "query_user_profile"    -> look up customer account (internal)    -> requires: "email"
 - "query_order_details"   -> look up order details (internal)       -> requires: "order_id"
 
+CONVERSATION FLOW — MANDATORY ORDER:
+1. FIRST action MUST be "respond": greet the customer warmly and acknowledge their specific issue.
+2. THEN use "request_info" if you still need information to resolve the issue.
+3. ONLY THEN use "query_user_profile" or "query_order_details" — and only when the customer has
+   explicitly confirmed or provided the email / order ID during the conversation AND you need that
+   data to answer their question.
+4. Finally "respond" with grounded facts, then "close" or "escalate".
+
+CRITICAL DB QUERY RULES:
+- NEVER query the DB as your first action, even if the ticket already shows an email or order ID.
+  Always greet and acknowledge first — querying without greeting is robotic and is penalized.
+- Only query AFTER the customer has confirmed the identifier in the current conversation thread.
+- Never invent data not present in KNOWN DATA or the customer's own messages.
+
 OUTPUT FORMAT — return ONLY this JSON, no code fences, no preamble:
 {"action_type": "...", "message": "..."}   <- for respond / request_info / close
 {"action_type": "escalate", "reason": "..."} <- for escalate
@@ -190,9 +206,22 @@ ACTION TYPES — output exactly one per step:
 - "query_user_profile"    -> look up customer account (internal)    -> requires: "email"
 - "query_order_details"   -> look up order details (internal)       -> requires: "order_id"
 
+CONVERSATION FLOW — MANDATORY ORDER:
+1. FIRST action MUST be "respond": greet the customer warmly by name if possible, acknowledge their issue.
+2. THEN use "request_info" if you still need details to help them.
+3. ONLY THEN query the DB — but ONLY when the customer has explicitly confirmed the email or order ID
+   during this conversation AND you need that data to answer accurately.
+4. Respond with the grounded facts, then "close" or "escalate" as appropriate.
+
+CRITICAL DB QUERY RULES:
+- NEVER issue a DB query as your very first action — always acknowledge the customer first.
+  The ticket description may mention an email or order ID, but greet first. Skipping the greeting
+  is robotic, hurts empathy scores, and incurs a penalty.
+- After querying, cite ONLY values from KNOWN DATA or from the customer's own messages.
+  Never invent facts (amounts, dates, status) not present in retrieved data.
+- If supervisor gave feedback, INCORPORATE it into your next action.
+
 SCORING: Empathy(30%) + Accuracy(25%) + Resolution(25%) + Efficiency(20%)
-Be warm, gather info from "Unresolved issues", use specific resolution language.
-If supervisor gave feedback, INCORPORATE it into your next action.
 
 OUTPUT FORMAT — return ONLY this JSON:
 {{"action_type": "...", "message": "..."}} or {{"action_type": "escalate", "reason": "..."}}"""
@@ -337,7 +366,7 @@ _FALLBACKS = {
 
 def run_task(task_name: str) -> None:
     try:
-        r = httpx.post(f"{ENV_URL}/reset", params={"task": task_name}, timeout=30)
+        r = httpx.post(f"{ENV_URL}/reset", params={"task": task_name}, headers=_ENV_HEADERS, timeout=30)
         r.raise_for_status()
     except Exception as exc:
         print(f"[ERROR] Reset failed task={task_name}: {exc}", file=sys.stderr)
@@ -359,7 +388,7 @@ def run_task(task_name: str) -> None:
             action = _FALLBACKS["support_agent"]
 
         try:
-            sr = httpx.post(f"{ENV_URL}/step", params={"session_id": session_id}, json=action, timeout=30)
+            sr = httpx.post(f"{ENV_URL}/step", params={"session_id": session_id}, json=action, headers=_ENV_HEADERS, timeout=30)
             sr.raise_for_status()
             result = sr.json()
         except Exception as exc:
@@ -379,7 +408,7 @@ def run_task(task_name: str) -> None:
 
 def run_hierarchy_task(task_name: str) -> None:
     try:
-        r = httpx.post(f"{ENV_URL}/reset", params={"task": task_name}, timeout=30)
+        r = httpx.post(f"{ENV_URL}/reset", params={"task": task_name}, headers=_ENV_HEADERS, timeout=30)
         r.raise_for_status()
     except Exception as exc:
         print(f"[ERROR] Reset failed task={task_name}: {exc}", file=sys.stderr)
@@ -403,7 +432,7 @@ def run_hierarchy_task(task_name: str) -> None:
             action = _FALLBACKS.get(role, _FALLBACKS["support_agent"])
 
         try:
-            sr = httpx.post(f"{ENV_URL}/step", params={"session_id": session_id}, json=action, timeout=60)
+            sr = httpx.post(f"{ENV_URL}/step", params={"session_id": session_id}, json=action, headers=_ENV_HEADERS, timeout=60)
             sr.raise_for_status()
             result = sr.json()
         except Exception as exc: