ForexBot2/llm_client.py

"""
ForexBot2 — DeepSeek LLM client
================================
Thin requests-based client for DeepSeek's OpenAI-compatible chat API,
used by `news_engine` for display/warning-only enrichment (headline
sentiment, event surprise direction, market summary). Signals are NEVER
gated on LLM output — every caller has a non-LLM fallback and treats a
`None` return as "use the fallback".

Design constraints:
- One attempt per call, no retries: news runs inside scan cycles; the
  news engine's 600 s context cache is the natural retry cadence.
- A UTC-day request cap bounds spend even if a bug causes call storms.
- Cost: DeepSeek caches identical prompt PREFIXES server-side (cache-hit
  input is ~50x cheaper), so callers pass a STATIC system prompt (no
  timestamps or interpolation) and put all variable content in the user
  payload.

This is a module-level singleton (`get_llm_client`) because the daily
cap counter must survive the per-scan reconstruction of `NewsEngine`.
"""

import datetime as _dt
import json
import logging
import threading
import time
from typing import Any, Dict, Optional, Tuple

import requests

from app_config import deepseek_api_key, get_app_config

logger = logging.getLogger(__name__)

DEEPSEEK_URL = "https://api.deepseek.com/chat/completions"
DEFAULT_MODEL = "deepseek-v4-flash"
# 8s timed out on a measured cold-prefix-cache headline batch (2026-08-03);
# warm-cache calls come back well under that. 12s buys cold-start headroom
# while still bounding worst-case scan latency (≤3 calls per cache window).
DEFAULT_TIMEOUT = 12.0
DEFAULT_DAILY_CAP = 500


class DeepSeekClient:
    """JSON-in/JSON-out chat wrapper. All failures return None."""

    def __init__(
        self,
        api_key: str,
        model: str = DEFAULT_MODEL,
        timeout: float = DEFAULT_TIMEOUT,
        daily_cap: int = DEFAULT_DAILY_CAP,
        enabled: bool = True,
    ):
        self._api_key = api_key or ""
        self._model = model
        self._timeout = timeout
        self._daily_cap = daily_cap
        self._enabled = enabled
        self._session = requests.Session()
        self._session.headers.update({
            "Content-Type": "application/json",
            "Authorization": f"Bearer {self._api_key}",
        })
        self._lock = threading.Lock()
        self._count_day: Optional[_dt.date] = None
        self._count = 0
        self._warned: set = set()
        # Health stamps for status displays: last successful call and
        # last failure (ts, kind). Read-only for consumers.
        self.last_ok: float = 0.0
        self.last_error: Optional[Tuple[float, str]] = None

    @property
    def available(self) -> bool:
        return self._enabled and bool(self._api_key)

    def _warn_once(self, kind: str, msg: str, *args: Any) -> None:
        if kind in self._warned:
            logger.debug(msg, *args)
        else:
            self._warned.add(kind)
            logger.warning(msg, *args)

    def _under_cap(self) -> bool:
        """Count a request against the UTC-day cap; False = don't call."""
        today = _dt.datetime.now(_dt.timezone.utc).date()
        with self._lock:
            if self._count_day != today:
                self._count_day = today
                self._count = 0
                # Clear ALL warn-once kinds, not just "cap": the network/
                # http_*/parse flags used to be set once and never reset,
                # so in a daemon that runs for weeks the second and every
                # later outage logged at DEBUG only.
                self._warned.clear()
            if self._count >= self._daily_cap:
                self._warn_once(
                    "cap", "DeepSeek daily request cap (%d) exhausted — "
                    "falling back to keyword analysis until UTC midnight",
                    self._daily_cap)
                return False
            self._count += 1
            return True

    def chat_json(self, system_prompt: str,
                  user_payload: Dict[str, Any]) -> Optional[Dict[str, Any]]:
        """One chat call; returns the model's JSON object or None."""
        if not self.available or not self._under_cap():
            return None
        body = {
            "model": self._model,
            "messages": [
                {"role": "system", "content": system_prompt},
                {"role": "user",
                 "content": json.dumps(user_payload, sort_keys=True)},
            ],
            "response_format": {"type": "json_object"},
            "temperature": 0,
            # V4-Flash is a hybrid thinking model; without this flag it
            # burns hundreds of reasoning tokens per call — measured
            # 2026-08-03: a 25-headline batch blew straight through the
            # timeout AND could exhaust max_tokens before the answer,
            # returning empty content. The doctrine + few-shot examples
            # in the system prompts carry the quality instead.
            "thinking": {"type": "disabled"},
            # Must fit the largest reply: a full sentiment batch is
            # ~10 tokens per headline; 1000 truncated mid-JSON on a real
            # RSS cycle (2026-08-03).
            "max_tokens": 2000,
        }
        try:
            resp = self._session.post(
                DEEPSEEK_URL, json=body, timeout=self._timeout)
        except requests.RequestException as e:
            self.last_error = (time.time(), "network")
            self._warn_once("network", "DeepSeek call failed: %s", e)
            return None
        if resp.status_code != 200:
            self.last_error = (time.time(), f"http_{resp.status_code}")
            self._warn_once(
                f"http_{resp.status_code}",
                "DeepSeek HTTP %s: %s", resp.status_code, resp.text[:200])
            return None
        try:
            parsed = self._parse_response(resp.json())
        except ValueError:
            parsed = None
        if parsed is None:
            self.last_error = (time.time(), "parse")
            self._warn_once("parse", "DeepSeek response was not the "
                            "expected JSON object — ignoring")
        else:
            self.last_ok = time.time()
            # A recovered call re-arms the failure warnings so the NEXT
            # outage is visible at WARNING again. "cap" keeps its own
            # UTC-day reset in _under_cap. (Taken under the lock; safe
            # here because chat_json holds no lock — _under_cap has
            # already returned.)
            if self._warned - {"cap"}:
                with self._lock:
                    self._warned = {k for k in self._warned if k == "cap"}
        return parsed

    @staticmethod
    def _parse_response(body: Any) -> Optional[Dict[str, Any]]:
        """Extract the assistant's JSON object from a chat completion."""
        try:
            content = body["choices"][0]["message"]["content"]
        except (KeyError, IndexError, TypeError):
            return None
        try:
            obj = json.loads(content)
        except (TypeError, ValueError):
            return None
        return obj if isinstance(obj, dict) else None


# ── Process-wide singleton ────────────────────────────────────────────

_client: Optional[DeepSeekClient] = None
_client_lock = threading.Lock()


def get_llm_client() -> DeepSeekClient:
    global _client
    with _client_lock:
        if _client is None:
            settings = get_app_config().deepseek_settings
            _client = DeepSeekClient(
                api_key=deepseek_api_key(),
                model=str(settings.get("model", DEFAULT_MODEL)),
                timeout=float(settings.get("timeout_seconds",
                                           DEFAULT_TIMEOUT)),
                daily_cap=int(settings.get("daily_request_cap",
                                           DEFAULT_DAILY_CAP)),
                enabled=settings.get("enabled", True) is not False,
            )
        return _client


def reset_llm_client() -> None:
    """Drop the singleton (tests / config reloads)."""
    global _client
    with _client_lock:
        _client = None