Blame for classify_perception.py

# classify_perception.py
import json
import re
from typing import List, Dict, Any, Optional
from llm_client import client
from util_agent import load_short_term_history
from config import OPENAI_CHAT_MODEL

# Benchmarked 2026-07-15 (debug_perception_comparison.py): nano missed 2 of 7 scenarios
# (recognizing standalone brevity, picking the right tactic for a nostalgic register);
# mini matched Haiku 4.5 and Sonnet 5 on every scored case at a fraction of their cost.
PERCEPTION_MODEL = OPENAI_CHAT_MODEL

# Tune these
RECENT_USER_TURNS = 5            # use last 5 user turns for better stability
INCLUDE_ASSISTANT_TURNS = 2      # include last 2 assistant turns (optional)
DEFAULT_CATEGORY = "neutral"
EMOTION_CATALOG_SUPABASE = "supabase://Databases/Agent_Emotional_Evaluation_Reactions_List.json"

# In-memory cache — catalog is static at runtime, no need to re-fetch every turn
_CATALOG_CACHE: Optional[Dict[str, Any]] = None


# -------------------------------
# Helpers: history packing
# -------------------------------
def _extract_recent_turns(messages: List[Dict[str, str]]) -> List[Dict[str, str]]:
    """Return a compact history window: last N user turns + last M assistant turns near them."""
    if not messages:
        return []

    user_msgs = [m for m in messages if m.get("role") == "user"]
    keep_users = user_msgs[-RECENT_USER_TURNS:]

    assistant_msgs = [m for m in messages if m.get("role") == "assistant"]
    keep_assist = assistant_msgs[-INCLUDE_ASSISTANT_TURNS:] if INCLUDE_ASSISTANT_TURNS > 0 else []

    keep_set = {id(m) for m in keep_users + keep_assist}
    compact = [m for m in messages if id(m) in keep_set]

    if not compact:
        compact = messages[-(RECENT_USER_TURNS * 2):]

    return compact


# -------------------------------
# Lightweight keyword gates
# -------------------------------
def apply_keyword_gates(latest_user_msg: str, result: Dict[str, Any]) -> Dict[str, Any]:
    text = (latest_user_msg or "").lower()

    if any(tok in text for tok in ["lol", "haha", "😂", "😅", "just kidding", "i'm joking", "sto scherzando"]):
        result["tone"] = "playful"
        if result.get("receptivity") not in ("overwhelmed",) and result.get("emotion_intensity", 0.0) < 0.65:
            result["tactic"] = result.get("tactic") or "tease"

    if any(k in text for k in ["do you remember", "ti ricordi", "te l'ho già detto", "non ricordi"]):
        result["intent"] = "memory_check"

    # Accountability: user references something the assistant said/proposed
    if any(k in text for k in [
        "you said", "you suggested", "you mentioned", "you offered", "you asked",
        "you proposed", "you told me", "you were the one", "it was you",
        "in your last", "your previous", "your last message",
    ]):
        result["intent"] = "memory_check"
        result["tactic"] = "suggest"
        result.setdefault("constraints", {})
        result["constraints"]["no_questions"] = True

    if any(k in text for k in ["desperate", "can't sleep", "panic", "overwhelmed", "i can't do this", "disperato", "non dormo"]):
        result["emotion_label"] = result.get("emotion_label") or "stressed"
        result["emotion_intensity"] = max(float(result.get("emotion_intensity", 0.0)), 0.75)
        result.setdefault("constraints", {})
        result["constraints"]["no_teasing"] = True
        result["constraints"]["no_questions"] = True
        result["tactic"] = "suggest"

    # mirror_brevity is now decided by the LLM via the prompt rule below.
    # No hardcoded keyword gate here — language-agnostic semantic judgment
    # belongs to the LLM, not to a word list.

    return result


# -------------------------------
# Prompt builder
# -------------------------------
def build_perception_prompt(
    recent_msgs: List[Dict[str, str]],
    catalog: Optional[Dict[str, Any]] = None,
    dialogue_context: Optional[str] = None,
) -> str:
    if catalog and catalog.get("categories"):
        categories_description = "\n".join(
            [f"- {c['name']} (priority {c.get('priority', 99)}): {c.get('description','')}"
             for c in catalog["categories"]]
        )
    else:
        categories_description = "- neutral (priority 9): default, no strong signals."

    schema = {
        "category": "neutral",
        "priority": 9,
        "confidence": 0.7,

        "emotion_label": "neutral",
        "emotion_intensity": 0.2,

        "intent": "explain",
        "tone": "serious",
        "receptivity": "neutral",
        "tactic": "reflect",

        "constraints": {
            "no_questions": False,
            "no_teasing": False,
            "mirror_brevity": False
        },

        "evidence": "short quote (<=12 words) from the user"
    }

    prompt = f"""
You are an evaluation agent for a Socratic companion.

Goal:
Return ACTIONABLE control signals for how the assistant should respond next.

You must:
1) Use the recent chat snippet (history-aware).
2) Prioritize the LAST user message, but interpret it in context.
3) Choose ONE 'category' from the list below (use highest priority if multiple apply).
4) Also output human affect + intent + tone + receptivity + tactic.

Categories (priority order applies: top = highest priority):
{categories_description}

Definitions:
- emotion_label: neutral|stressed|anxious|sad|angry|excited|frustrated|proud|nostalgic|vulnerable
- intent: vent (wants empathy), guidance (wants steps), explain (wants understanding),
          debate (wants challenge), banter (playful), memory_check, closure,
          celebrate (user sharing excitement about something, wants joy mirrored back).
- tone: serious|playful|sarcastic|defensive|warm
- receptivity: open|neutral|defensive|overwhelmed
- tactic: tease|question|reflect|suggest|celebrate|listen

Rules:
- If emotional intensity >= 0.75 OR receptivity = overwhelmed:
    set constraints.no_teasing = true
    set constraints.no_questions = true (unless absolutely necessary for clarification)
    prefer tactic = suggest or reflect
- If tone is playful AND receptivity is open AND intensity < 0.65:
    teasing may be allowed; tactic can be tease OR question (choose one)
- If the user is frustrated with the assistant:
    be calm, concise; prefer reflect then suggest; avoid end-question by default.
- If the user expresses excitement, joy, or enthusiasm about any event (sport, culture, personal):
    set intent = celebrate, tactic = celebrate, constraints.no_teasing = true, constraints.no_questions = true.
    Socrates mirrors the joy — no sarcasm, no philosophy, no redirecting questions.
- If the user expresses sadness, grief, or emotional pain:
    set tactic = listen, constraints.no_teasing = true, constraints.no_questions = true.
    Prefer empathetic category.
- If the user reflects on a past memory or nostalgia:
    set tactic = question (one gentle curiosity question only), constraints.no_teasing = true.
    Prefer nostalgic category.
- If the user shares something deeply personal or vulnerable:
    set tactic = listen, constraints.no_teasing = true, constraints.no_questions = true.
    Prefer vulnerable category.
- constraints.mirror_brevity — set to true ONLY when the user's message is short AND it is a
  genuinely standalone brief exchange: a greeting, a casual one-liner, a closing remark, a
  reaction with no follow-up intent. Do NOT set it when the short message is:
    (a) answering a direct question the assistant just asked (any language);
    (b) asking to continue a narration or explanation ("go on", "yes", "tell me more", or
        equivalent in any language);
    (c) part of an ongoing structured dialogue where short replies are normal step responses.
  The purpose of mirror_brevity is to keep the assistant brief when the user is being casual
  or winding down — not to truncate responses when the user wants more content.
- If the LAST ASSISTANT MESSAGE posed a binary or multiple-choice question (e.g. "should you do X or Y?",
  "does it lead you to A or B?") AND the user is now asking about one of those options (e.g. "do you think
  I should do X?"): set intent = guidance, tactic = suggest, constraints.no_questions = true.
  The user wants a direct answer, not more exploration.
- Output ONLY valid JSON matching this schema exactly (keys must exist):
{json.dumps(schema, indent=2)}

Recent chat snippet:
{json.dumps(recent_msgs, ensure_ascii=False, indent=2)}
""".strip()

    if dialogue_context:
        prompt += f"\n\nDIALOGUE CONTEXT:\n{dialogue_context}"

    return prompt


# -------------------------------
# LLM call
# -------------------------------
def call_perception_llm(
    recent_msgs: List[Dict[str, str]],
    catalog: Optional[Dict[str, Any]] = None,
    dialogue_context: Optional[str] = None,
    model: Optional[str] = None,
) -> Dict[str, Any]:
    prompt = build_perception_prompt(recent_msgs, catalog, dialogue_context=dialogue_context)

    resp = client.chat.completions.create(
        model=model or PERCEPTION_MODEL,
        messages=[
            {"role": "system", "content": "Return ONLY valid JSON. No extra text."},
            {"role": "user", "content": prompt},
        ],
        temperature=0,
        max_tokens=350,
        response_format={"type": "json_object"},
    )

    raw = (resp.choices[0].message.content or "").strip()

    m = re.search(r"\{.*\}", raw, re.S)
    if not m:
        return {
            "category": DEFAULT_CATEGORY,
            "priority": 9,
            "confidence": 0.0,
            "emotion_label": "neutral",
            "emotion_intensity": 0.2,
            "intent": "explain",
            "tone": "serious",
            "receptivity": "neutral",
            "tactic": "reflect",
            "constraints": {"no_questions": False, "no_teasing": False, "mirror_brevity": False},
            "evidence": "parse_failed"
        }

    out = json.loads(m.group(0))

    try:
        out["emotion_intensity"] = float(out.get("emotion_intensity", 0.2))
    except Exception:
        out["emotion_intensity"] = 0.2
    out["emotion_intensity"] = max(0.0, min(1.0, out["emotion_intensity"]))

    try:
        out["confidence"] = float(out.get("confidence", 0.7))
    except Exception:
        out["confidence"] = 0.7
    out["confidence"] = max(0.0, min(1.0, out["confidence"]))

    if not isinstance(out.get("constraints"), dict):
        out["constraints"] = {"no_questions": False, "no_teasing": False, "mirror_brevity": False}
    for k in ["no_questions", "no_teasing", "mirror_brevity"]:
        out["constraints"][k] = bool(out["constraints"].get(k, False))

    return out


# -------------------------------
# Public API
# -------------------------------
def analyze_perception_from_history(
    latest_user_msg: str,
    user_id: str,
    catalog: Optional[Dict[str, Any]] = None,
    catalog_path: Optional[str] = None,
    dialogue_context: Optional[str] = None,
    model: Optional[str] = None,
) -> Dict[str, Any]:
    """
    Loads short-term history, appends latest user msg (not persisted yet),
    runs perception analysis, then applies keyword gates.

    Catalog resolution order: dict > local file > Supabase default.

    dialogue_context: optional string passed to the LLM to inform it of the
    current dialogue state (e.g. "user is mid-step in a structured dialogue").
    When set, the LLM adjusts its disengagement/intent thresholds accordingly.
    """
    history = load_short_term_history(user_id)

    if latest_user_msg:
        history.append({"role": "user", "content": latest_user_msg})

    compact = _extract_recent_turns(history)

    if catalog is None and catalog_path:
        with open(catalog_path, "r", encoding="utf-8") as f:
            catalog = json.load(f)

    if catalog is None:
        global _CATALOG_CACHE
        if _CATALOG_CACHE is None:
            try:
                from db_user import load_json
                _CATALOG_CACHE = load_json(EMOTION_CATALOG_SUPABASE)
            except Exception:
                _CATALOG_CACHE = None
        catalog = _CATALOG_CACHE

    result = call_perception_llm(compact, catalog=catalog, dialogue_context=dialogue_context, model=model)
    result = apply_keyword_gates(latest_user_msg, result)

    return result