honey-biscuit-workshop/tests/llm/client.py

"""
LLM client supporting both Anthropic and OpenAI APIs.

Prefers Anthropic format (per project convention).
Falls back to OpenAI format if Anthropic is unavailable.
"""

import os
import logging
from .cache import LLMCache

logger = logging.getLogger(__name__)


class LLMClient:
    """LLM client with caching support.  Prefers Anthropic API."""

    def __init__(self, cache: LLMCache | None = None):
        self.cache = cache or LLMCache()
        self._anthropic_client = None
        self._openai_client = None
        self._init_clients()

    def _init_clients(self):
        """Initialize API clients based on available environment variables."""
        # Anthropic (preferred)
        anthropic_key = os.environ.get("ANTHROPIC_API_KEY")
        anthropic_base = os.environ.get("ANTHROPIC_BASE_URL")
        if anthropic_key:
            try:
                import anthropic

                kwargs: dict = {"api_key": anthropic_key}
                if anthropic_base:
                    kwargs["base_url"] = anthropic_base
                self._anthropic_client = anthropic.Anthropic(**kwargs)
                logger.info("Anthropic client initialized")
            except ImportError:
                logger.warning("anthropic package not installed")

        # OpenAI (fallback)
        openai_key = os.environ.get("OPENAI_API_KEY")
        openai_base = os.environ.get("OPENAI_BASE_URL")
        if openai_key:
            try:
                import openai

                kwargs = {"api_key": openai_key}
                if openai_base:
                    kwargs["base_url"] = openai_base
                self._openai_client = openai.OpenAI(**kwargs)
                logger.info("OpenAI client initialized")
            except ImportError:
                logger.warning("openai package not installed")

    @property
    def available(self) -> bool:
        """Check if any LLM client is available."""
        return (
            self._anthropic_client is not None or self._openai_client is not None
        )

    def complete(
        self, prompt: str, system_prompt: str = "", max_tokens: int = 2000
    ) -> str:
        """Send a completion request, with caching.

        Args:
            prompt: User message.
            system_prompt: System message (sets role / persona).
            max_tokens: Max response tokens.

        Returns:
            LLM response text.

        Raises:
            RuntimeError: If no client is available.
        """
        model = os.environ.get("MODEL", "deepseek-v4-flash")

        # Check cache
        cached = self.cache.get(prompt, model, system_prompt)
        if cached:
            logger.info("Cache hit for LLM request")
            return cached["content"]

        # Try Anthropic first
        if self._anthropic_client:
            content = self._call_anthropic(
                prompt, system_prompt, model, max_tokens
            )
        elif self._openai_client:
            content = self._call_openai(prompt, system_prompt, model, max_tokens)
        else:
            raise RuntimeError(
                "No LLM client available. "
                "Set ANTHROPIC_API_KEY or OPENAI_API_KEY."
            )

        # Cache the response
        self.cache.set(prompt, model, system_prompt, {"content": content})
        return content

    # ------------------------------------------------------------------
    # Private helpers
    # ------------------------------------------------------------------

    def _call_anthropic(
        self, prompt: str, system_prompt: str, model: str, max_tokens: int
    ) -> str:
        """Call Anthropic API."""
        kwargs: dict = {
            "model": model,
            "max_tokens": max_tokens,
            "messages": [{"role": "user", "content": prompt}],
        }
        if system_prompt:
            kwargs["system"] = system_prompt

        # Only add reasoning_effort if the model supports it
        reasoning_effort = os.environ.get("REASONING_EFFORT")
        if reasoning_effort:
            try:
                kwargs["reasoning_effort"] = reasoning_effort
                response = self._anthropic_client.messages.create(**kwargs)
            except Exception as e:
                if (
                    "reasoning_effort" in str(e).lower()
                    or "unexpected" in str(e).lower()
                ):
                    del kwargs["reasoning_effort"]
                    response = self._anthropic_client.messages.create(**kwargs)
                else:
                    raise
        else:
            response = self._anthropic_client.messages.create(**kwargs)

        # Extract text from response content blocks
        if hasattr(response, "content") and response.content:
            block = response.content[0]
            if hasattr(block, "text"):
                return block.text
            if hasattr(block, "type") and block.type == "thinking":
                for b in response.content:
                    if hasattr(b, "text"):
                        return b.text
            return str(block)
        return str(response)

    def _call_openai(
        self, prompt: str, system_prompt: str, model: str, max_tokens: int
    ) -> str:
        """Call OpenAI-compatible API."""
        messages: list[dict] = []
        if system_prompt:
            messages.append({"role": "system", "content": system_prompt})
        messages.append({"role": "user", "content": prompt})

        kwargs: dict = {
            "model": model,
            "messages": messages,
            "max_tokens": max_tokens,
        }

        # Only add reasoning_effort if supported
        reasoning_effort = os.environ.get("REASONING_EFFORT")
        if reasoning_effort:
            try:
                kwargs["reasoning_effort"] = reasoning_effort
                response = self._openai_client.chat.completions.create(
                    **kwargs
                )
            except Exception as e:
                if (
                    "reasoning_effort" in str(e).lower()
                    or "unexpected" in str(e).lower()
                ):
                    del kwargs["reasoning_effort"]
                    response = self._openai_client.chat.completions.create(
                        **kwargs
                    )
                else:
                    raise
        else:
            response = self._openai_client.chat.completions.create(**kwargs)

        return response.choices[0].message.content