"""
Moltbot Hybrid Engine - Production v7.0.0
Multi-service: FastAPI + Ollama (optional) + OpenClaw/Clawdbot gateway (proxied at /gateway)
Runs on Hugging Face Spaces
Build: 2026-02-14 — Clawdbot installed in Space; gateway on 18789, proxied at /gateway

Endpoints:
  GET  /              - Health check
  GET  /health        - Detailed health status
  GET  /gateway       - OpenClaw/Clawdbot Control UI (reverse proxy to gateway :18789)
  GET  /gateway/{path} - OpenClaw proxy (path)
  POST /api/generate  - LLM text generation
  ...
"""
import os
import re
import json
import subprocess
import logging
from pathlib import Path
from fastapi import FastAPI, HTTPException, Header, UploadFile, File, Request
from fastapi.responses import StreamingResponse, Response
from pydantic import BaseModel
from typing import List, Optional, Dict, Any, Union

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger("moltbot-engine")

# Initialize App
app = FastAPI(
    title="Moltbot Hybrid Engine",
    description="AI agent for legal document processing - Dual LLM + file matching + Clawdbot gateway at /gateway",
    version="7.0.0"
)

# API Key for authentication
API_KEY = os.environ.get("MOLTBOT_API_KEY", "default_insecure_key")
if API_KEY == "default_insecure_key":
    logger.warning("MOLTBOT_API_KEY not set. Using insecure default.")

# HuggingFace token for Inference API
HF_TOKEN = os.environ.get("HF_TOKEN", "")

# Default HF model for inference API fallback
HF_MODEL = os.environ.get("HF_MODEL", "Qwen/Qwen2.5-7B-Instruct")


# ============================================================
# DATA MODELS
# ============================================================

class GenerateRequest(BaseModel):
    model: str = "qwen2.5:1.5b"
    prompt: str
    stream: bool = False

class FileSearchRequest(BaseModel):
    missing_filename: str
    available_files: List[str]
    context: Optional[str] = None

class FileSearchResponse(BaseModel):
    status: str
    missing_filename: str
    suggestions: List[Dict[str, Any]]
    confidence: float
    reasoning: str

class AnalysisRequest(BaseModel):
    report_data: Dict[str, Any]

class AnalysisResponse(BaseModel):
    status: str
    critical_issues: int
    suggestions: List[str]


# ============================================================
# HELPER FUNCTIONS
# ============================================================

def tokenize(text: str) -> set:
    """Tokenize a filename for fuzzy matching."""
    clean = re.sub(r'[_\-\.\(\)\[\]]', ' ', text.lower())
    tokens = set(clean.split())
    junk = {'pdf', 'mp4', 'jpg', 'jpeg', 'png', 'gif', 'doc', 'docx',
            'the', 'and', 'to', 'of', 'in', 'a', 'for', 'with', 'on'}
    return tokens - junk

def calculate_match_score(wanted: set, found: set) -> float:
    """Calculate token overlap score between two sets."""
    if not wanted:
        return 0.0
    common = wanted.intersection(found)
    return len(common) / len(wanted)

def find_best_matches(missing_filename: str, available_files: List[str], max_results: int = 5) -> List[Dict[str, Any]]:
    """Find best fuzzy matches for a missing filename."""
    wanted_tokens = tokenize(missing_filename)
    if not wanted_tokens:
        return []

    matches = []
    for filename in available_files:
        if filename == missing_filename:
            matches.append({"filename": filename, "score": 1.0, "match_type": "exact"})
            continue
        if filename.lower() == missing_filename.lower():
            matches.append({"filename": filename, "score": 0.99, "match_type": "case_insensitive"})
            continue
        found_tokens = tokenize(filename)
        score = calculate_match_score(wanted_tokens, found_tokens)
        if score >= 0.5:
            matches.append({"filename": filename, "score": round(score, 3), "match_type": "token_match"})

    matches.sort(key=lambda x: x["score"], reverse=True)
    return matches[:max_results]

def check_ollama_status() -> dict:
    """Check if Ollama is running and responsive."""
    try:
        result = subprocess.run(
            ["ollama", "list"],
            capture_output=True, text=True, timeout=10
        )
        if result.returncode == 0:
            models = [line.split()[0] for line in result.stdout.strip().split('\n')[1:] if line.strip()]
            return {"running": True, "models": models}
        return {"running": False, "error": result.stderr.strip()}
    except FileNotFoundError:
        return {"running": False, "error": "ollama binary not found"}
    except OSError as e:
        return {"running": False, "error": f"ollama exec error: {e}"}
    except subprocess.TimeoutExpired:
        return {"running": False, "error": "ollama list timed out"}
    except Exception as e:
        return {"running": False, "error": str(e)}


def generate_with_ollama(model: str, prompt: str) -> Optional[str]:
    """Try to generate text with local Ollama. Returns None if unavailable."""
    # try:
    #     result = subprocess.run(
    #         ["ollama", "run", model, prompt],
    #         capture_output=True, text=True, timeout=120
    #     )
    #     if result.returncode == 0 and result.stdout.strip():
    #         return result.stdout.strip()
    #     logger.warning(f"[OLLAMA] Non-zero return or empty output: {result.stderr[:200]}")
    #     return None
    # except (FileNotFoundError, OSError) as e:
    #     logger.warning(f"[OLLAMA] Not available: {e}")
    #     return None
    # except subprocess.TimeoutExpired:
    #     logger.warning("[OLLAMA] Timeout after 120s")
    #     return None
    # except Exception as e:
    #     logger.warning(f"[OLLAMA] Error: {e}")
    #     return None
    try:
        import requests as req
        payload = {
            "model": model,
            "prompt": prompt,
            "stream": False,
            "options": {
                "num_ctx": 16384
            }
        }
        resp = req.post("http://localhost:11434/api/generate", json=payload, timeout=120)
        if resp.status_code == 200:
            data = resp.json()
            text = data.get("response", "").strip()
            if text:
                return text
        logger.warning(f"[OLLAMA] Status {resp.status_code}: {resp.text[:200]}")
        return None
    except req.exceptions.ConnectionError:
        logger.warning("[OLLAMA] Not available (connection refused)")
        return None
    except req.exceptions.Timeout:
        logger.warning("[OLLAMA] Timeout after 120s")
        return None
    except Exception as e:
        logger.warning(f"[OLLAMA] Error: {e}")
        return None

def generate_with_hf_api(prompt: str, model: str = None) -> Optional[str]:
    """Generate text using HuggingFace Inference API (free, no GPU needed).
    Uses chat_completion (conversational) — the only supported task for Qwen2.5-7B-Instruct.
    """
    try:
        from huggingface_hub import InferenceClient

        hf_model = model or HF_MODEL
        token = HF_TOKEN if HF_TOKEN else None

        client = InferenceClient(token=token)

        # Use chat_completion — Qwen2.5-7B-Instruct is conversational only
        # (text_generation fails: "Supported task: conversational")
        response = client.chat_completion(
            model=hf_model,
            messages=[
                {"role": "system", "content": "You are a helpful legal assistant."},
                {"role": "user", "content": prompt},
            ],
            max_tokens=1024,
            temperature=0.7,
        )
        if response and response.choices:
            text = response.choices[0].message.content
            if text:
                return text.strip()

        logger.warning("[HF_API] Empty response from chat_completion")
        return None

    except ImportError:
        logger.error("[HF_API] huggingface_hub not installed")
        return None
    except Exception as e:
        logger.warning(f"[HF_API] chat_completion error: {e}")
        return None


def analyze_report_data(data: dict) -> dict:
    """Analyze a verification report and return findings."""
    suggestions = []
    missing_files = data.get("missing_total", 0)
    structure_issues = data.get("page_structure_analysis", {}).get("bundles_with_structure_issues", 0)
    blank_pages = data.get("blank_placeholder_pages", 0)

    if missing_files > 0:
        suggestions.append(f"{missing_files} files missing - check paths in proven_paths_index.json")
    if structure_issues > 0:
        suggestions.append(f"{structure_issues} bundles have page structure problems - check embedding logic")
    if blank_pages > 0:
        suggestions.append(f"{blank_pages} blank placeholder pages found - files listed in TOC but not embedded")

    bundles = data.get("bundles", {})
    for bundle_name, bundle_data in bundles.items():
        if isinstance(bundle_data, dict):
            bundle_missing = bundle_data.get("missing_count", 0)
            if bundle_missing > 0:
                suggestions.append(f"Bundle {bundle_name}: {bundle_missing} files missing")

    critical_count = missing_files + structure_issues + blank_pages
    return {
        "status": "success",
        "critical_issues": critical_count,
        "suggestions": suggestions if suggestions else ["No critical issues found"],
        "summary": f"Analyzed report: {critical_count} critical issues across {len(bundles)} bundles"
    }


# ============================================================
# API ENDPOINTS
# ============================================================

@app.get("/")
def health_check():
    """Basic health check."""
    ollama = check_ollama_status()
    return {
        "status": "running",
        "service": "Moltbot Hybrid Engine",
        "version": "7.0.0",
        "ollama": ollama,
        "clawdbot": "OpenClaw gateway proxied at /gateway (if running)",
        "hf_inference_api": {
            "available": True,
            "model": HF_MODEL,
            "token_set": bool(HF_TOKEN)
        }
    }

@app.get("/health")
def detailed_health():
    """Detailed health check with LLM status."""
    ollama = check_ollama_status()
    return {
        "status": "healthy",
        "service": "moltbot-hybrid-engine",
        "version": "7.0.0",
        "llm_backends": {
            "ollama": {
                "running": ollama.get("running", False),
                "models": ollama.get("models", []),
                "error": ollama.get("error"),
            },
            "hf_inference_api": {
                "available": True,
                "model": HF_MODEL,
                "token_set": bool(HF_TOKEN),
                "note": "Always available as fallback, no GPU needed"
            }
        },
        "endpoints": ["/", "/health", "/api/generate", "/api/search",
                      "/api/analyze", "/api/extract_date", "/tools/analyze_report",
                      "/v1/chat/completions", "/v1/models", "/gateway (Clawdbot UI)"]
    }

@app.get("/security")
def security_info():
    """Report security posture."""
    return {
        "file_access": False,
        "network_access": "API only",
        "isolation": "Hugging Face container",
        "cannot_do": ["Read local files", "Write local files", "Delete files",
                      "Access host filesystem", "Execute arbitrary commands"]
    }


# ============================================================
# CLAWDBOT /gateway/verify — Direct verification endpoint
# Called by CoV hook (cov_integration_hook.py step 6)
# NOT proxied to OpenClaw; runs LLM verification directly
# ============================================================

class GatewayVerifyRequest(BaseModel):
    edit_text: str
    local_cov_results: Optional[Dict[str, Any]] = None
    exhibit_schedule: Optional[str] = None  # Authoritative exhibit schedule from build_exhibit_briefcase.py

class GatewayVerifyResponse(BaseModel):
    verdict: str  # "pass", "flag", "block"
    reasoning: str
    issues: List[str]

@app.post("/gateway/verify", response_model=GatewayVerifyResponse)
async def gateway_verify(request: GatewayVerifyRequest, x_api_key: str = Header(None)):
    """
    ClawdBot verification endpoint. Receives edit text + local CoV results,
    runs LLM verification with legal exhibit instruction + evidence context,
    and returns verdict (pass/flag/block), reasoning, and issues list.
    Called by courtBundleGenerator3/adapters/cov_integration_hook.py (step 6).
    """
    # Auth: accept API key or allow if default key
    if API_KEY != "default_insecure_key":
        if not x_api_key or x_api_key != API_KEY:
            raise HTTPException(status_code=401, detail="Invalid or missing API Key")

    edit_text = request.edit_text or ""
    local_results = request.local_cov_results or {}

    if not edit_text.strip():
        return GatewayVerifyResponse(verdict="block", reasoning="Empty edit text", issues=["No text provided"])

    # Build verification prompt with legal exhibit instruction + evidence context
    legal_instruction = _get_legal_exhibit_instruction()

    # Include exhibit schedule if provided (authoritative mapping from build_exhibit_briefcase.py)
    exhibit_schedule_section = ""
    if request.exhibit_schedule:
        exhibit_schedule_section = f"""

---

AUTHORITATIVE EXHIBIT SCHEDULE (from actual bundle output — use this to resolve all references):
{request.exhibit_schedule[:6000]}
"""

    # Summarise local CoV results so the LLM has context
    local_summary_parts = []
    scores = local_results.get("scores", {})
    if scores:
        local_summary_parts.append(f"Local CoV scores: {json.dumps(scores)}")
    critical = local_results.get("critical_issues", [])
    if critical:
        local_summary_parts.append(f"Critical issues from local checks: {critical}")
    warnings = local_results.get("warnings", [])
    if warnings:
        local_summary_parts.append(f"Warnings: {warnings[:5]}")
    local_summary = "\n".join(local_summary_parts) if local_summary_parts else "No local CoV results provided."

    verify_prompt = f"""You are ClawdBot, a legal document verification agent.

{legal_instruction}
{exhibit_schedule_section}

---

LOCAL COV RESULTS (from timeline, accuracy, prosecutor, self-healing, Qwen checks):
{local_summary}

---

EDIT TEXT TO VERIFY:
{edit_text[:8000]}

---

TASK: Review the edit text above. Check:
1. Every evidence reference uses the format: Exhibit [Letter][Number] (DB-[N]) — [Filename]. Flag any bare DB-[N] or missing exhibit refs.
   Use the EXHIBIT SCHEDULE above to verify that the Letter+Number matches the correct DB number and filename.
   Flag any Exhibit [L][N] placeholders — these MUST be replaced with real letters from the schedule.
2. No sentences are cut off mid-word or mid-sentence (truncation).
3. Legal claims are supported by cited evidence. Flag unsupported assertions.
4. Dates, names, and case references are consistent.
5. No placeholder text like [●], [L][N], TBC, or TODO remains.

Respond with EXACTLY this JSON (no markdown, no extra text):
{{"verdict": "pass" or "flag" or "block", "reasoning": "one paragraph explanation", "issues": ["issue 1", "issue 2"]}}

If all checks pass, verdict is "pass" with empty issues list.
If minor issues found, verdict is "flag".
If critical issues (truncation, missing exhibit refs, unsupported claims, [L][N] placeholders), verdict is "block"."""

    # Run through LLM (same dual backend as /api/generate)
    response_text = None

    # Try Ollama first
    response_text = generate_with_ollama("qwen2.5:1.5b", verify_prompt)

    # Fallback to HF Inference API
    if not response_text:
        response_text = generate_with_hf_api(verify_prompt)

    if not response_text:
        return GatewayVerifyResponse(
            verdict="flag",
            reasoning="LLM backends unavailable; cannot verify. Passing as flag for human review.",
            issues=["LLM unavailable — manual review required"]
        )

    # Parse LLM response into structured verdict
    try:
        # Try to extract JSON from response (LLM may wrap it in markdown)
        json_match = re.search(r'\{[^{}]*"verdict"[^{}]*\}', response_text, re.DOTALL)
        if json_match:
            parsed = json.loads(json_match.group())
            verdict = parsed.get("verdict", "flag")
            if verdict not in ("pass", "flag", "block"):
                verdict = "flag"
            reasoning = parsed.get("reasoning", response_text[:300])
            issues = parsed.get("issues", [])
            if not isinstance(issues, list):
                issues = [str(issues)]
            return GatewayVerifyResponse(verdict=verdict, reasoning=reasoning, issues=issues)
    except (json.JSONDecodeError, AttributeError):
        pass

    # Fallback: couldn't parse JSON — return the raw response as a flag
    return GatewayVerifyResponse(
        verdict="flag",
        reasoning=response_text[:500],
        issues=["LLM response could not be parsed as structured JSON — manual review recommended"]
    )


# OpenClaw/Clawdbot gateway reverse proxy (gateway runs on 18789; Space exposes single port 7860)
OPENCLAW_GATEWAY_URL = "http://127.0.0.1:18789"

@app.api_route("/gateway", methods=["GET", "POST", "PUT", "DELETE", "PATCH", "OPTIONS"])
@app.api_route("/gateway/{path:path}", methods=["GET", "POST", "PUT", "DELETE", "PATCH", "OPTIONS"])
async def proxy_openclaw_gateway(request: Request, path: str = ""):
    """Proxy requests to OpenClaw/Clawdbot gateway so Control UI and WebChat are reachable at /gateway."""
    try:
        import httpx
    except ImportError:
        raise HTTPException(status_code=503, detail="httpx not installed; cannot proxy to Clawdbot gateway")
    target_path = request.url.path
    if target_path.startswith("/gateway"):
        target_path = target_path[8:] or "/"  # strip /gateway -> / or /foo
    target = f"{OPENCLAW_GATEWAY_URL}{target_path}"
    if request.url.query:
        target += "?" + request.url.query
    headers = {k: v for k, v in request.headers.raw if k.lower() not in (b"host", b"connection")}
    try:
        body = await request.body()
    except Exception:
        body = b""
    async with httpx.AsyncClient(timeout=30.0) as client:
        try:
            r = await client.request(
                request.method,
                target,
                headers=headers,
                content=body,
            )
        except httpx.ConnectError:
            return Response(
                content="Clawdbot gateway not reachable (is it running on 18789?). Start the Space and try again.",
                status_code=503,
                media_type="text/plain",
            )
        except Exception as e:
            logger.warning(f"[GATEWAY PROXY] {e}")
            return Response(content=str(e), status_code=502, media_type="text/plain")
    out_headers = {}
    for k, v in r.headers.items():
        if k.lower() not in ("transfer-encoding", "connection"):
            out_headers[k] = v
    return Response(
        content=r.content,
        status_code=r.status_code,
        headers=out_headers,
        media_type=r.headers.get("content-type", "application/octet-stream"),
    )


# Legal document exhibit reference instruction — injected into every generate/chat so edit sources always get it
_prompts_dir = Path(__file__).resolve().parent / "prompts"
_LEGAL_EXHIBIT_PROMPT_PATH = _prompts_dir / "legal_exhibit_instruction.txt"
_FULL_EVIDENCE_REF_PATH = _prompts_dir / "full_evidence_reference.txt"
_LEGAL_EXHIBIT_INSTRUCTION_CACHED: Optional[str] = None

def _get_legal_exhibit_instruction() -> str:
    """Load legal exhibit instruction once; append full evidence list when present. Injected into all LLM requests."""
    global _LEGAL_EXHIBIT_INSTRUCTION_CACHED
    if _LEGAL_EXHIBIT_INSTRUCTION_CACHED is not None:
        return _LEGAL_EXHIBIT_INSTRUCTION_CACHED
    if _LEGAL_EXHIBIT_PROMPT_PATH.exists():
        _LEGAL_EXHIBIT_INSTRUCTION_CACHED = _LEGAL_EXHIBIT_PROMPT_PATH.read_text(encoding="utf-8", errors="replace")
    else:
        _LEGAL_EXHIBIT_INSTRUCTION_CACHED = "When referencing evidence use Exhibit [Letter][Number] (DB-[N]) — [Filename]. Do not use bare DB-[●]."
    if _FULL_EVIDENCE_REF_PATH.exists():
        try:
            full = _FULL_EVIDENCE_REF_PATH.read_text(encoding="utf-8", errors="replace").strip()
            if full:
                _LEGAL_EXHIBIT_INSTRUCTION_CACHED = _LEGAL_EXHIBIT_INSTRUCTION_CACHED + "\n\nFull evidence list (use for every cite):\n" + full
        except Exception:
            pass
    return _LEGAL_EXHIBIT_INSTRUCTION_CACHED

@app.get("/prompts/legal-exhibit-instruction")
def get_legal_exhibit_instruction():
    """Return the legal exhibit referencing instruction. Also injected automatically into /api/generate and /v1/chat/completions."""
    return {"instruction": _get_legal_exhibit_instruction()}


# --- LLM Generation (Dual Backend: Ollama → HF Inference API) ---

@app.post("/api/generate")
async def generate(request: GenerateRequest, x_api_key: str = Header(None)):
    """Generate text using LLM. Tries Ollama first, falls back to HF Inference API.
    Legal exhibit instruction is prepended so all edits/amendments use Exhibit [Letter][Number] (DB-[N]) — [Filename].
    """
    if not x_api_key or x_api_key != API_KEY:
        raise HTTPException(status_code=401, detail="Invalid or missing API Key")

    # Inject legal exhibit instruction so edit sources always get the rule
    prompt_with_legal = _get_legal_exhibit_instruction() + "\n\n---\n\n" + request.prompt
    logger.info(f"[GENERATE] model={request.model}, prompt_len={len(prompt_with_legal)}")

    backend_used = None
    response_text = None

    # Backend 1: Try Ollama (local)
    response_text = generate_with_ollama(request.model, prompt_with_legal)
    if response_text:
        backend_used = "ollama"
        logger.info(f"[GENERATE] Ollama success, response_len={len(response_text)}")

    # Backend 2: Fallback to HF Inference API
    if not response_text:
        logger.info("[GENERATE] Ollama unavailable, trying HF Inference API...")
        response_text = generate_with_hf_api(prompt_with_legal)
        if response_text:
            backend_used = "hf_inference_api"
            logger.info(f"[GENERATE] HF API success, response_len={len(response_text)}")

    # Both failed
    if not response_text:
        raise HTTPException(
            status_code=503,
            detail="Both LLM backends unavailable. Ollama not running + HF Inference API failed. Check HF_TOKEN."
        )

    return {
        "model": request.model,
        "response": response_text,
        "backend": backend_used,
        "done": True
    }


# --- File Search ---

@app.post("/api/search", response_model=FileSearchResponse)
async def search_file(request: FileSearchRequest, x_api_key: str = Header(None)):
    """Fuzzy file matching for missing evidence files."""
    if not x_api_key or x_api_key != API_KEY:
        raise HTTPException(status_code=401, detail="Invalid or missing API Key")

    if len(request.missing_filename) > 200:
        return FileSearchResponse(
            status="error", missing_filename=request.missing_filename[:50] + "...",
            suggestions=[], confidence=0.0,
            reasoning="Filename too long - likely concatenated filenames"
        )

    matches = find_best_matches(request.missing_filename, request.available_files)
    confidence = matches[0]["score"] if matches else 0.0

    if not matches:
        reasoning = f"No matches found in {len(request.available_files)} files"
    elif matches[0]["match_type"] == "exact":
        reasoning = f"Exact match: {matches[0]['filename']}"
    else:
        reasoning = f"Token match with {int(confidence * 100)}% similarity"

    return FileSearchResponse(
        status="success", missing_filename=request.missing_filename,
        suggestions=matches, confidence=confidence, reasoning=reasoning
    )


# --- Report Analysis via JSON body ---

@app.post("/api/analyze", response_model=AnalysisResponse)
async def analyze_report_json(request: AnalysisRequest, x_api_key: str = Header(None)):
    """Analyze a verification report (JSON body)."""
    if not x_api_key or x_api_key != API_KEY:
        raise HTTPException(status_code=401, detail="Invalid or missing API Key")

    result = analyze_report_data(request.report_data)
    return AnalysisResponse(
        status=result["status"],
        critical_issues=result["critical_issues"],
        suggestions=result["suggestions"]
    )


# --- Report Analysis via file upload ---

@app.post("/tools/analyze_report")
async def analyze_report_upload(
    report_file: UploadFile = File(...),
    x_api_key: str = Header(None)
):
    """Analyze a verification report uploaded as a file."""
    if not x_api_key or x_api_key != API_KEY:
        raise HTTPException(status_code=401, detail="Invalid or missing API Key")

    logger.info(f"[ANALYZE_REPORT] Received file: {report_file.filename}")

    try:
        content = await report_file.read()
        data = json.loads(content)
    except json.JSONDecodeError:
        raise HTTPException(status_code=400, detail="Invalid JSON in uploaded file")
    except Exception as e:
        raise HTTPException(status_code=400, detail=f"Error reading file: {str(e)}")

    result = analyze_report_data(data)
    logger.info(f"[ANALYZE_REPORT] Found {result['critical_issues']} critical issues")
    return result


# --- Date Extraction ---

@app.post("/api/extract_date")
async def extract_date(filename: str, x_api_key: str = Header(None)):
    """Extract date from a filename string."""
    if not x_api_key or x_api_key != API_KEY:
        raise HTTPException(status_code=401, detail="Invalid or missing API Key")

    patterns = [
        (r'(\d{4})-(\d{2})-(\d{2})', 'ISO'),
        (r'(\d{4})_(\d{2})_(\d{2})', 'underscore'),
        (r'(\d{1,2})-(\d{1,2})-(\d{2,4})', 'UK_dash'),
        (r'(\d{4})(\d{2})(\d{2})', 'compact'),
    ]
    for pattern, fmt in patterns:
        match = re.search(pattern, filename)
        if match:
            groups = match.groups()
            if fmt == 'UK_dash':
                day, month, year = groups
                if len(year) == 2:
                    year = f"20{year}"
                return {"status": "found", "date": f"{year}-{int(month):02d}-{int(day):02d}", "format": fmt}
            else:
                year, month, day = groups
                return {"status": "found", "date": f"{year}-{int(month):02d}-{int(day):02d}", "format": fmt}

    return {"status": "not_found", "date": None}


# ============================================================
# OPENAI-COMPATIBLE ENDPOINT (for Cursor IDE integration)
# Enhanced: streaming, full message history, multi-model
# ============================================================

class ChatMessage(BaseModel):
    role: str
    content: Union[str, List[Dict[str, Any]]]  # OpenAI: string or array of parts (vision/R1)

class ChatCompletionRequest(BaseModel):
    model: str = "moltbot-legal"
    messages: List[ChatMessage]
    temperature: Optional[float] = 0.7
    max_tokens: Optional[int] = 2048
    stream: Optional[bool] = False
    # Accept but ignore these (Cursor sends them)
    top_p: Optional[float] = None
    frequency_penalty: Optional[float] = None
    presence_penalty: Optional[float] = None
    stop: Optional[List[str]] = None
    n: Optional[int] = None

# Model routing table — maps Cursor model names to backends
# Additive: new models can be added here without touching any other code
MODEL_ROUTING = {
    "moltbot-legal":   {"backend": "hf", "hf_model": "Qwen/Qwen2.5-7B-Instruct"},
    "qwen2.5":         {"backend": "hf", "hf_model": "Qwen/Qwen2.5-7B-Instruct"},
    "qwen2.5:1.5b":    {"backend": "ollama", "ollama_model": "qwen2.5:1.5b"},
    "deepseek-chat":   {"backend": "hf", "hf_model": "deepseek-ai/DeepSeek-V2.5"},
    "deepseek-coder":  {"backend": "hf", "hf_model": "deepseek-ai/DeepSeek-Coder-V2-Instruct"},
}


def _message_content_to_str(content: Union[str, List[Dict[str, Any]]]) -> str:
    """Normalize message content to string (OpenAI can send string or array of parts, e.g. vision/R1)."""
    if isinstance(content, str):
        return content
    if isinstance(content, list):
        return " ".join(
            p.get("text", "") for p in content
            if isinstance(p, dict) and isinstance(p.get("text"), str)
        )
    return str(content)


def _build_prompt_from_messages(messages: List[ChatMessage]) -> str:
    """Build a full prompt from the chat message history.
    Preserves conversation context, not just last message.
    """
    parts = []
    for msg in messages:
        text = _message_content_to_str(msg.content)
        if msg.role == "system":
            parts.append(f"System: {text}")
        elif msg.role == "user":
            parts.append(f"User: {text}")
        elif msg.role == "assistant":
            parts.append(f"Assistant: {text}")
    parts.append("Assistant:")
    return "\n\n".join(parts)


def _build_hf_chat_messages(messages: List[ChatMessage]) -> List[Dict[str, str]]:
    """Build HF chat_completion message list from OpenAI messages."""
    return [{"role": msg.role, "content": _message_content_to_str(msg.content)} for msg in messages]


def _generate_for_model(model_name: str, messages: List[ChatMessage], temperature: float = 0.7, max_tokens: int = 2048) -> Optional[str]:
    """Route generation to the right backend based on model name.
    Falls back through: model-specific backend → HF chat → HF text → Ollama.
    """
    route = MODEL_ROUTING.get(model_name, MODEL_ROUTING.get("moltbot-legal"))

    # Try HF chat_completion first (best quality for instruct models)
    try:
        from huggingface_hub import InferenceClient
        hf_model = route.get("hf_model", HF_MODEL)
        token = HF_TOKEN if HF_TOKEN else None
        client = InferenceClient(token=token)
        chat_msgs = _build_hf_chat_messages(messages)
        response = client.chat_completion(
            model=hf_model,
            messages=chat_msgs,
            max_tokens=max_tokens,
            temperature=temperature,
        )
        if response and response.choices:
            text = response.choices[0].message.content.strip()
            if text:
                logger.info(f"[CHAT] HF chat_completion success, model={hf_model}, len={len(text)}")
                return text
    except Exception as e:
        logger.warning(f"[CHAT] HF chat_completion failed: {e}")

    # Fallback: HF text_generation
    prompt = _build_prompt_from_messages(messages)
    text = generate_with_hf_api(prompt, model=route.get("hf_model"))
    if text:
        logger.info(f"[CHAT] HF text_generation fallback success, len={len(text)}")
        return text

    # Fallback: Ollama
    ollama_model = route.get("ollama_model", "qwen2.5:1.5b")
    text = generate_with_ollama(ollama_model, prompt)
    if text:
        logger.info(f"[CHAT] Ollama fallback success, len={len(text)}")
        return text

    return None


def _stream_sse_response(model: str, response_text: str, request_id: str):
    """Yield Server-Sent Events chunks that Cursor expects for streaming."""
    import time
    now = int(time.time())

    # Split response into word-level chunks for realistic streaming
    words = response_text.split(" ")
    for i, word in enumerate(words):
        chunk_text = word + (" " if i < len(words) - 1 else "")
        chunk = {
            "id": request_id,
            "object": "chat.completion.chunk",
            "created": now,
            "model": model,
            "choices": [{
                "index": 0,
                "delta": {"content": chunk_text},
                "finish_reason": None
            }]
        }
        yield f"data: {json.dumps(chunk)}\n\n"

    # Final chunk with finish_reason
    final = {
        "id": request_id,
        "object": "chat.completion.chunk",
        "created": now,
        "model": model,
        "choices": [{
            "index": 0,
            "delta": {},
            "finish_reason": "stop"
        }]
    }
    yield f"data: {json.dumps(final)}\n\n"
    yield "data: [DONE]\n\n"


@app.post("/v1/chat/completions")
async def chat_completions(
    request: ChatCompletionRequest,
    authorization: str = Header(None),
    x_api_key: str = Header(None),
):
    """OpenAI-compatible chat completions endpoint.
    Supports both streaming and non-streaming modes.
    Accepts 'Authorization: Bearer <key>' or 'x-api-key: <key>'.
    Routes to appropriate backend based on model name.
    """
    import time

    # Auth: accept both header styles.
    # Accept MOLTBOT_API_KEY, HF_TOKEN, or any non-empty key when using default key.
    # This is a personal Space running a free model — HF provides its own rate limiting.
    api_key = None
    if authorization and authorization.startswith("Bearer "):
        api_key = authorization[7:].strip()
    if not api_key:
        api_key = x_api_key

    # If MOLTBOT_API_KEY was explicitly set (not default), enforce it strictly
    if API_KEY != "default_insecure_key":
        valid = api_key and (api_key == API_KEY or (HF_TOKEN and api_key == HF_TOKEN))
    else:
        # Default key in use: accept any non-empty key (so Cursor can send DeepSeek/HF/any key)
        valid = bool(api_key)

    if not valid:
        raise HTTPException(status_code=401, detail="Invalid or missing API Key")

    logger.info(f"[CHAT] model={request.model}, messages={len(request.messages)}, stream={request.stream}")

    # Inject legal exhibit instruction so every edit/amendment/insert uses Exhibit [Letter][Number] (DB-[N]) — [Filename]
    legal_system = ChatMessage(role="system", content=_get_legal_exhibit_instruction())
    messages_with_legal = [legal_system] + list(request.messages)

    # Generate response via model routing
    response_text = _generate_for_model(
        request.model, messages_with_legal,
        temperature=request.temperature or 0.7,
        max_tokens=request.max_tokens or 2048,
    )

    if not response_text:
        raise HTTPException(status_code=503, detail="All LLM backends unavailable")

    now = int(time.time())
    request_id = f"chatcmpl-moltbot-{now}"

    # STREAMING MODE — Cursor sends stream=true by default
    if request.stream:
        return StreamingResponse(
            _stream_sse_response(request.model, response_text, request_id),
            media_type="text/event-stream",
            headers={
                "Cache-Control": "no-cache",
                "Connection": "keep-alive",
                "X-Accel-Buffering": "no",
            },
        )

    # NON-STREAMING MODE — standard OpenAI response
    return {
        "id": request_id,
        "object": "chat.completion",
        "created": now,
        "model": request.model,
        "choices": [{
            "index": 0,
            "message": {"role": "assistant", "content": response_text},
            "finish_reason": "stop"
        }],
        "usage": {
            "prompt_tokens": len(" ".join(_message_content_to_str(m.content) for m in request.messages).split()),
            "completion_tokens": len(response_text.split()),
            "total_tokens": len(" ".join(_message_content_to_str(m.content) for m in request.messages).split()) + len(response_text.split()),
        }
    }


@app.get("/v1/models")
async def list_models():
    """OpenAI-compatible model listing. Cursor discovers models via this endpoint."""
    return {
        "object": "list",
        "data": [
            {"id": "moltbot-legal", "object": "model", "created": 1700000000, "owned_by": "moltbot"},
            {"id": "qwen2.5", "object": "model", "created": 1700000000, "owned_by": "alibaba-cloud"},
            {"id": "deepseek-chat", "object": "model", "created": 1700000000, "owned_by": "deepseek"},
            {"id": "deepseek-coder", "object": "model", "created": 1700000000, "owned_by": "deepseek"},
        ]
    }


if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=7860)