server/app.py

from __future__ import annotations

import json
from typing import Any, Dict, Optional

from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import JSONResponse
from pydantic import BaseModel

from models import TrustAction, TrustObservation, TrustState, ContentSignals
from your_environment import TrustSafetyEnvironment

# ── Force manual FastAPI (openenv_core create_app causes 422 on /step) ────────
print("[app] Using manual FastAPI ✅")

_env = TrustSafetyEnvironment(seed=42)

app = FastAPI(
    title="Trust & Safety RL Environment",
    description="Risk-aware content moderation environment for agent training.",
    version="1.0.0",
)

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_methods=["*"],
    allow_headers=["*"],
)


# ── Serializers ───────────────────────────────────────────────────────────────

def _obs_to_dict(obs: TrustObservation) -> Dict[str, Any]:
    return {
        "ticket_id":           obs.ticket_id,
        "post_text":           obs.post_text,
        "image_description":   obs.image_description,
        "comments_found":      obs.comments_found,
        "user_history_found":  obs.user_history_found,
        "entity_status_found": obs.entity_status_found,
        "policy_found":        obs.policy_found,
        "extracted_signals":   obs.extracted_signals,
        "validation_result":   obs.validation_result,
        "step_number":         obs.step_number,
        "info":                obs.info,
        "done":                obs.done,
        "reward":              obs.reward,
    }


def _state_to_dict(s: TrustState) -> Dict[str, Any]:
    return {
        "episode_id":        s.episode_id,
        "step_count":        s.step_count,
        "current_task_id":   s.current_task_id,
        "difficulty":        s.difficulty,
        "ambiguity_level":   s.ambiguity_level,
        "risk_level":        s.risk_level,
        "tools_used":        s.tools_used,
        "signals_extracted": s.signals_extracted,
        "is_done":           s.is_done,
    }


# ── Request bodies ─────────────────────────────────────────────────────────────

class ResetRequest(BaseModel):
    seed:       Any = None
    episode_id: Any = None

    model_config = {"extra": "ignore"}


class ActionRequest(BaseModel):
    action_type:    str                      = ""
    tool_name:      Optional[str]            = None
    signals:        Optional[Dict[str, Any]] = None   # raw dict — validated below
    final_decision: Optional[str]            = None

    model_config = {"extra": "ignore"}   # ← ignore unknown keys from LLM


# ── Helpers ────────────────────────────────────────────────────────────────────

def _parse_signals(raw: Dict[str, Any]) -> ContentSignals:
    """Defensively normalise LLM signal output before Pydantic validation."""
    # Clamp floats
    raw["toxicity_level"] = float(raw.get("toxicity_level", 0.5))
    raw["confidence"]     = float(raw.get("confidence",     0.5))

    # content_flags must be a list of strings
    flags = raw.get("content_flags", [])
    if not isinstance(flags, list):
        flags = [flags] if isinstance(flags, str) else []
    raw["content_flags"] = [str(f) for f in flags]

    # boolean coercion
    raw["is_protected_class"]       = bool(raw.get("is_protected_class",       False))
    raw["is_direct_attack"]         = bool(raw.get("is_direct_attack",         False))
    raw["abusive_language_present"] = bool(raw.get("abusive_language_present", False))

    # string fields — fallback to sensible defaults
    raw.setdefault("target",       "none")
    raw.setdefault("intent",       "ambiguous")
    raw.setdefault("context_type", "statement")

    return ContentSignals(**raw)


# ── Routes ─────────────────────────────────────────────────────────────────────

@app.get("/health")
async def health():
    return {"status": "ok", "environment": "trust-safety-env", "version": "1.0.0"}


@app.get("/")
async def root():
    return {"status": "ok", "docs": "/docs"}


@app.post("/reset")
async def reset(body: ResetRequest = ResetRequest()):
    obs = _env.reset(seed=body.seed, episode_id=body.episode_id)
    return JSONResponse(_obs_to_dict(obs))


@app.post("/step")
async def step(body: ActionRequest):
    # Parse + validate signals defensively
    signals: Optional[ContentSignals] = None
    if body.signals:
        try:
            signals = _parse_signals(dict(body.signals))   # copy so we don't mutate
        except Exception as e:
            raise HTTPException(status_code=400, detail=f"Invalid signals payload: {e}")

    action = TrustAction(
        action_type    = body.action_type,
        tool_name      = body.tool_name,
        signals        = signals,
        final_decision = body.final_decision,
    )

    try:
        obs = _env.step(action)
    except (RuntimeError, ValueError) as e:
        raise HTTPException(status_code=400, detail=str(e))

    return JSONResponse(_obs_to_dict(obs))


@app.get("/state")
async def state():
    return JSONResponse(_state_to_dict(_env.state))