app.py

from __future__ import annotations

import asyncio
import os
import re
from typing import AsyncIterator, Dict, List

import gradio as gr
from gradio_wordleboard import WordleBoard
from openai import AsyncOpenAI

from envs.textarena_env import TextArenaAction, TextArenaEnv
from envs.textarena_env.models import TextArenaMessage


API_BASE_URL = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1")
API_KEY = os.getenv("API_KEY") or os.getenv("HF_TOKEN")
MODEL = os.getenv("MODEL", "openai/gpt-oss-120b:novita")
MAX_TURNS = int(os.getenv("MAX_TURNS", "6"))
DOCKER_IMAGE = os.getenv("TEXTARENA_IMAGE", "textarena-env:latest")


def _format_history(messages: List[TextArenaMessage]) -> str:
    lines: List[str] = []
    for message in messages:
        tag = message.category or "MESSAGE"
        lines.append(f"[{tag}] {message.content}")
    return "\n".join(lines)


def _make_user_prompt(prompt_text: str, messages: List[TextArenaMessage]) -> str:
    history = _format_history(messages)
    return (
        f"Current prompt:\n{prompt_text}\n\n"
        f"Conversation so far:\n{history}\n\n"
        "Reply with your next guess enclosed in square brackets."
    )


async def _generate_guesses(client: AsyncOpenAI, prompt: str, history: List[TextArenaMessage]) -> str:
    response = await client.chat.completions.create(
        model=MODEL,
        messages=[
            {
                "role": "system",
                "content": (
                    "You are an expert Wordle solver."
                    " Always respond with a single guess inside square brackets, e.g. [crane]."
                    " Use lowercase letters, exactly one five-letter word per reply."
                    " Reason about prior feedback before choosing the next guess."
                    " Words must be 5 letters long and real English words."
                    " Do not include any other text in your response."
                    " Do not repeat the same guess twice."
                ),
            },
            {"role": "user", "content": _make_user_prompt(prompt, history)},
        ],
        max_tokens=64,
        temperature=0.7,
    )

    content = response.choices[0].message.content
    response_text = content.strip() if content else ""
    print(f"Response text: {response_text}")
    return response_text


async def _play_wordle(env: TextArenaEnv, client: AsyncOpenAI) -> AsyncIterator[Dict[str, str]]:
    state = await asyncio.to_thread(env.reset)
    observation = state.observation

    for turn in range(1, MAX_TURNS + 1):
        if state.done:
            break

        model_output = await _generate_guesses(client, observation.prompt, observation.messages)
        guess = _extract_guess(model_output)

        state = await asyncio.to_thread(env.step, TextArenaAction(message=guess))
        observation = state.observation

        feedback = _collect_feedback(observation.messages)
        yield {"guess": guess, "feedback": feedback}

    yield {
        "guess": "",
        "feedback": _collect_feedback(observation.messages),
    }


def _extract_guess(text: str) -> str:
    if not text:
        return "[crane]"

    match = re.search(r"\[([A-Za-z]{5})\]", text)
    if match:
        guess = match.group(1).lower()
        return f"[{guess}]"

    cleaned = re.sub(r"[^a-zA-Z]", "", text).lower()
    if len(cleaned) >= 5:
        return f"[{cleaned[:5]}]"

    return "[crane]"


def _collect_feedback(messages: List[TextArenaMessage]) -> str:
    parts: List[str] = []
    for message in messages:
        tag = message.category or "MESSAGE"
        if tag.upper() in {"FEEDBACK", "SYSTEM", "MESSAGE"}:
            parts.append(message.content.strip())
    return "\n".join(parts).strip()


async def inference_handler(api_key: str) -> AsyncIterator[str]:
    if not api_key:
        raise RuntimeError("HF_TOKEN or API_KEY environment variable must be set.")

    client = AsyncOpenAI(base_url=API_BASE_URL, api_key=api_key)
    env = TextArenaEnv(base_url="https://burtenshaw-textarena.hf.space")

    try:
        async for result in _play_wordle(env, client):
            yield result["feedback"]
    finally:
        env.close()


wordle_component = WordleBoard()


async def run_inference() -> AsyncIterator[Dict]:
    feedback_history: List[str] = []

    async for feedback in inference_handler(API_KEY):
        stripped = feedback.strip()
        if not stripped:
            continue

        feedback_history.append(stripped)
        combined_feedback = "\n\n".join(feedback_history)
        state = wordle_component.parse_feedback(combined_feedback)
        yield wordle_component.to_public_dict(state)

    if not feedback_history:
        yield wordle_component.to_public_dict(wordle_component.create_game_state())


with gr.Blocks() as demo:
    gr.Markdown("# Wordle TextArena Inference Demo")

    board = WordleBoard(value=wordle_component.to_public_dict(wordle_component.create_game_state()))
    run_button = gr.Button("Run Inference", variant="primary")

    run_button.click(
        fn=run_inference,
        inputs=None,
        outputs=board,
        show_progress=True,
        api_name="run",
    )

demo.queue()


if __name__ == "__main__":
    if not API_KEY:
        raise SystemExit("HF_TOKEN (or API_KEY) must be set to query the model.")

    demo.launch()