app.py

import base64
import os
import re
import tempfile

import gradio as gr
import requests
import inspect
import pandas as pd
from smolagents import (
    CodeAgent,
    DuckDuckGoSearchTool,
    VisitWebpageTool,
    OpenAIServerModel,
    tool,
)

# (Keep Constants as is)
# --- Constants ---
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"


# --- Agent prompt: enforces GAIA's exact-match answer format ----------------
# GAIA is graded by exact string match, so the final answer MUST follow these
# rules precisely or a correct answer still scores as wrong.
GAIA_INSTRUCTIONS = """You are a general AI assistant answering a question from the GAIA benchmark.
Reason step by step and use your tools to find the answer:
  - web_search / visit_webpage to research on the internet,
  - get_task_file to download a file attached to this question (if any),
  - Python code to parse files, do math, or process data.

The task_id for this question is: {task_id}
Only call get_task_file("{task_id}") if the question text explicitly mentions
an attached file (e.g. "the audio file", "the spreadsheet", "the image",
"the Python code", "this file"). Most GAIA questions have no attachment —
for those, do NOT call get_task_file. If you do call it and get back
"NO_FILE_ATTACHED", proceed without an attachment.

Available tools for multimodal / reference tasks:
  - wikipedia_search(query) returns the article INTRO only.
    - For tabular data (discography, filmography, awards, etc.) call
      wikipedia_search(query, section="Discography") to get JUST that
      section as Markdown with the table intact — much cleaner than
      slicing yourself.
    - As a last resort use wikipedia_search(query, full=True) for the
      whole article.
  - fetch_url(url) downloads a page with a real User-Agent and returns its
    content as Markdown (tables preserved). Prefer this over visit_webpage
    for Wikipedia and other sites that block default user agents.
  - youtube_transcript(url) when the question references a YouTube video.
    IMPORTANT: this fails on HF Spaces ~always (cloud IPs blocked). After
    ONE failed try, switch to youtube_info(url) for title/description,
    and use web_search to find an external summary of the video. Do not
    keep retrying the transcript.
  - youtube_info(url) scrapes the YouTube watch page (works from Spaces)
    for title, description, channel, duration, and the auto-caption URL.
    For "how long is the video" or "who is the speaker" this alone often
    answers the question. Visual-counting questions usually still need
    a web_search.
  - transcribe_audio(path) for MP3 / WAV / M4A attachments. After
    transcription, do NOT just split-and-sort every word — the audio
    usually contains a spoken sentence with filler words ("I need you to
    write down these ingredients..."). Re-read the transcript and extract
    ONLY the entities that match the question's category (ingredients,
    page numbers, names, etc.). Preserve multi-word items as one entity:
    "granulated sugar" is one ingredient, not two; "vanilla extract" is
    one, not two.
  - describe_image(path, question) for PNG / JPG attachments, charts,
    chess positions, etc. Ask a SPECIFIC question, e.g.
    "What text appears at the bottom?" or "List every fruit in the image."

GAIA is graded by EXACT MATCH. Use the OFFICIAL format spec verbatim:

  Your final answer should be a number OR as few words as possible OR a
  comma separated list of numbers and/or strings.
  - If you are asked for a number, don't use comma to write your number
    neither use units such as $ or percent sign unless specified otherwise.
  - If you are asked for a string, don't use articles, neither abbreviations
    (e.g. for cities), and write the digits in plain text unless specified
    otherwise. ("five" not "5" inside a string answer.)
  - If you are asked for a comma separated list, apply the above rules
    depending on whether the element is a number or a string.

Pass JUST the answer value to the final_answer tool — no "FINAL ANSWER:"
prefix, no explanation, no surrounding quotes, no trailing period.

Question:
{question}
"""


# --- Tools ------------------------------------------------------------------
@tool
def get_task_file(task_id: str) -> str:
    """Downloads the file attached to a GAIA task and saves it locally.

    ONLY call this when the question text explicitly mentions an attached
    file (spreadsheet, image, audio clip, Python file, etc.). If the
    question is pure text (no file mentioned), do not call this — most
    GAIA tasks have no attachment, and the server returns 404 for those.

    Returns the local path to the downloaded file, or one of:
      - "NO_FILE_ATTACHED": the task has no attachment. Move on with
        purely the question text; do not retry this tool.
      - "ERROR: ...": transient/network error worth surfacing.

    Args:
        task_id: The task_id of the current question.
    """
    url = f"{DEFAULT_API_URL}/files/{task_id}"
    try:
        resp = requests.get(url, timeout=30)
        if resp.status_code == 404:
            return "NO_FILE_ATTACHED"
        resp.raise_for_status()
    except requests.exceptions.HTTPError as e:
        if e.response is not None and e.response.status_code == 404:
            return "NO_FILE_ATTACHED"
        return f"ERROR: could not download file for task {task_id}: {e}"
    except Exception as e:  # noqa: BLE001 - surfaced to the agent as text
        return f"ERROR: could not download file for task {task_id}: {e}"

    # Try to preserve the original filename/extension from the response headers.
    filename = task_id
    disposition = resp.headers.get("content-disposition", "")
    match = re.search(r'filename="?([^"]+)"?', disposition)
    if match:
        filename = match.group(1)

    path = os.path.join(tempfile.gettempdir(), filename)
    with open(path, "wb") as f:
        f.write(resp.content)
    return path


def _html_to_markdown(html: str, *, is_wikipedia: bool) -> str:
    """Convert HTML to clean Markdown, preserving tables and headings.

    For Wikipedia pages we scope to the article body and drop navboxes,
    infoboxes, references, and edit links — those are pure noise for an
    agent reasoning about article facts.
    """
    try:
        from bs4 import BeautifulSoup
        from markdownify import markdownify as md
    except ImportError:
        return html  # fall back to raw HTML if deps are missing

    soup = BeautifulSoup(html, "html.parser")
    root = soup.select_one("#mw-content-text") or soup.body or soup if is_wikipedia else soup
    for selector in (
        "script", "style", "noscript",
        # Wikipedia chrome that pollutes article text:
        "table.navbox", "table.vertical-navbox", "table.infobox",
        "table.sidebar", "div.reflist", "div.refbegin", "ol.references",
        "sup.reference", "div.thumb", ".mw-editsection",
        ".mw-jump-link", "#toc", ".hatnote", ".navigation-not-searchable",
    ):
        for el in root.select(selector):
            el.decompose()
    # markdownify renders wikitables as proper Markdown tables, which keeps
    # the year ↔ title mapping the agent needs.
    return md(str(root), heading_style="ATX", strip=["a", "img"])


_WIKI_HEADERS = {
    "User-Agent": "GaiaAgent/1.0 (https://huggingface.co/spaces/agents-course)",
    "Accept": "application/json",
}


def _wiki_resolve_title(query: str) -> str | None:
    """Resolve a free-text query to the canonical Wikipedia title, or None."""
    try:
        r = requests.get(
            "https://en.wikipedia.org/w/api.php",
            params={
                "action": "query", "list": "search", "srsearch": query,
                "srlimit": 1, "format": "json",
            },
            headers=_WIKI_HEADERS, timeout=20,
        )
        r.raise_for_status()
        hits = r.json().get("query", {}).get("search", [])
        return hits[0]["title"] if hits else None
    except Exception:  # noqa: BLE001
        return None


@tool
def wikipedia_search(query: str, full: bool = False, section: str = "",
                     max_chars: int = 40000) -> str:
    """Look up a topic on Wikipedia.

    Prefer this over web_search for biographical, geographic, or other
    encyclopedic facts — Wikipedia answers most GAIA Level 1 lookups directly.

    Three modes:
      - default: returns the article's intro summary.
      - full=True: returns the entire article body as Markdown (tables kept).
      - section="Discography": returns ONLY that section (recommended when
        you already know which section holds the answer — discography,
        filmography, awards, etc.). Much smaller than full=True.

    Args:
        query: The topic to search for (e.g. "Mercedes Sosa").
        full: If True, return the full article body as Markdown.
        section: If set, return only the named section (case-sensitive H2/H3
            heading text, e.g. "Discography" or "Studio albums").
        max_chars: Truncate output to this length (default 40000).
    """
    title = _wiki_resolve_title(query)
    if not title:
        return f"No Wikipedia page found for '{query}'."

    safe_title = requests.utils.quote(title.replace(" ", "_"), safe="_()")
    page_url = f"https://en.wikipedia.org/wiki/{safe_title}"
    header = f"{title}\nURL: {page_url}\n\n"

    if not full and not section:
        # Use the REST summary endpoint — fast and gives a clean intro.
        try:
            r = requests.get(
                f"https://en.wikipedia.org/api/rest_v1/page/summary/{safe_title}",
                headers=_WIKI_HEADERS, timeout=20,
            )
            r.raise_for_status()
            return header + (r.json().get("extract") or "")
        except Exception as e:  # noqa: BLE001
            return f"ERROR: wikipedia summary failed: {e}"

    # Full HTML, converted to Markdown.
    try:
        r = requests.get(
            f"https://en.wikipedia.org/api/rest_v1/page/html/{safe_title}",
            headers={**_WIKI_HEADERS, "Accept": "text/html"}, timeout=30,
        )
        r.raise_for_status()
        html = r.text
    except Exception as e:  # noqa: BLE001
        return f"ERROR: wikipedia HTML fetch failed: {e}"

    markdown = _html_to_markdown(html, is_wikipedia=True)

    if section:
        # Locate "## Section" or "### Section" exactly (anchored, case-insensitive).
        pattern = re.compile(
            rf"^(#{{2,4}})\s+{re.escape(section)}\s*$",
            re.MULTILINE | re.IGNORECASE,
        )
        match = pattern.search(markdown)
        if not match:
            headings = [ln for ln in markdown.split("\n") if ln.startswith("## ")]
            return (f"Section '{section}' not found. Available H2 headings:\n"
                    + "\n".join(headings[:30]))
        start = match.start()
        level = len(match.group(1))
        # Slice until the next heading at the same or higher level.
        sibling = re.compile(rf"^#{{1,{level}}}(?!#)\s+", re.MULTILINE)
        end_match = sibling.search(markdown, pos=match.end())
        end = end_match.start() if end_match else len(markdown)
        return (header + markdown[start:end])[:max_chars]

    return (header + markdown)[:max_chars]


@tool
def fetch_url(url: str, max_chars: int = 20000) -> str:
    """Download a web page and return its content as Markdown.

    Use this when visit_webpage returns a 403 / 429 — many sites (Wikipedia
    included) block requests that lack a real browser User-Agent. This tool
    sends one, then converts the page to Markdown so that tables, headings,
    and lists keep their structure.

    Args:
        url: The URL to fetch.
        max_chars: Truncate the result to this many characters (default 20000).
    """
    try:
        resp = requests.get(
            url,
            headers={
                "User-Agent": (
                    "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) "
                    "AppleWebKit/537.36 (KHTML, like Gecko) "
                    "Chrome/124.0 Safari/537.36"
                ),
                "Accept": "text/html,application/xhtml+xml",
            },
            timeout=30,
        )
        resp.raise_for_status()
    except Exception as e:  # noqa: BLE001
        return f"ERROR: could not fetch {url}: {e}"

    is_wiki = "wikipedia.org" in url
    try:
        text = _html_to_markdown(resp.text, is_wikipedia=is_wiki)
    except Exception:  # noqa: BLE001
        text = resp.text

    return text[:max_chars]


def _extract_youtube_id(url: str) -> str | None:
    m = re.search(r"(?:v=|youtu\.be/|/shorts/|/embed/)([0-9A-Za-z_-]{11})", url)
    if m:
        return m.group(1)
    return url if re.fullmatch(r"[0-9A-Za-z_-]{11}", url) else None


@tool
def youtube_transcript(url: str) -> str:
    """Fetch the transcript (subtitles) of a YouTube video.

    NOTE: On HuggingFace Spaces this often fails with an IP-block error
    because YouTube blocks cloud-provider IPs. If you see that error, do
    NOT keep retrying — call youtube_info(url) for title/description, and
    use web_search() to find someone else's summary of the video content.

    Args:
        url: A YouTube URL or bare 11-character video id.
    """
    try:
        from youtube_transcript_api import YouTubeTranscriptApi
    except ImportError:
        return ("ERROR: the 'youtube-transcript-api' package is not installed "
                "in this Space.")

    video_id = _extract_youtube_id(url)
    if not video_id:
        return "ERROR: could not extract YouTube video id from URL."
    try:
        # v1.0+ exposes a `.fetch()` instance method; older versions had the
        # static `.get_transcript()`. Support both.
        if hasattr(YouTubeTranscriptApi, "get_transcript"):
            chunks = YouTubeTranscriptApi.get_transcript(video_id)
            return " ".join(c["text"] for c in chunks)
        fetched = YouTubeTranscriptApi().fetch(video_id)
        return " ".join(snippet.text for snippet in fetched)
    except Exception as e:  # noqa: BLE001
        msg = str(e)
        if "IpBlocked" in msg or "RequestBlocked" in msg or "blocking requests" in msg:
            return ("ERROR: YouTube IP-blocked this Space's transcript request. "
                    "DO NOT RETRY transcript. Instead call youtube_info(url) "
                    "for the title/description, then web_search for an external "
                    "summary or transcript.")
        return f"ERROR: could not fetch transcript: {e}"


@tool
def youtube_info(url: str) -> str:
    """Scrape a YouTube watch page for title, description, channel, and duration.

    Works from cloud IPs even when the transcript API is blocked — the watch
    page itself is publicly fetchable. Use as a fallback when
    youtube_transcript fails. The description sometimes contains the answer
    outright; otherwise pair this with web_search.

    Args:
        url: A YouTube URL or bare 11-character video id.
    """
    video_id = _extract_youtube_id(url)
    if not video_id:
        return "ERROR: could not extract YouTube video id from URL."
    page_url = f"https://www.youtube.com/watch?v={video_id}"
    try:
        resp = requests.get(
            page_url,
            headers={
                "User-Agent": (
                    "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) "
                    "AppleWebKit/537.36 (KHTML, like Gecko) "
                    "Chrome/124.0 Safari/537.36"
                ),
            },
            timeout=20,
        )
        resp.raise_for_status()
    except Exception as e:  # noqa: BLE001
        return f"ERROR: could not fetch YouTube page: {e}"

    html = resp.text
    # YouTube embeds structured data in og:* meta tags and a JSON blob
    # called ytInitialPlayerResponse. Pulling from both is most reliable.
    def og(prop: str) -> str:
        m = re.search(
            rf'<meta\s+property="og:{prop}"\s+content="([^"]+)"', html
        )
        return m.group(1) if m else ""

    title = og("title")
    description = og("description")
    duration_match = re.search(r'"lengthSeconds":"(\d+)"', html)
    duration_s = int(duration_match.group(1)) if duration_match else None
    channel_match = re.search(r'"ownerChannelName":"([^"]+)"', html)
    channel = channel_match.group(1) if channel_match else ""

    # Try to pull the auto-generated caption track URL from the page, if any.
    cap_match = re.search(r'"captionTracks":(\[[^\]]+\])', html)
    captions_url = ""
    if cap_match:
        u = re.search(r'"baseUrl":"([^"]+)"', cap_match.group(1))
        if u:
            captions_url = u.group(1).encode().decode("unicode_escape")

    parts = [
        f"Title: {title}",
        f"Channel: {channel}" if channel else "",
        f"Duration: {duration_s}s" if duration_s else "",
        f"Description: {description}",
    ]
    if captions_url:
        parts.append(
            f"Caption track URL (try fetch_url on this): {captions_url}"
        )
    return "\n".join(p for p in parts if p)


def _hf_client():
    """Build a HuggingFace InferenceClient. Token is read from HF_TOKEN
    (auto-injected in HF Spaces) or HUGGINGFACEHUB_API_TOKEN."""
    try:
        from huggingface_hub import InferenceClient
    except ImportError as e:
        raise RuntimeError(
            "huggingface_hub is not installed; cannot use audio/vision tools."
        ) from e

    token = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACEHUB_API_TOKEN")
    if not token:
        raise RuntimeError(
            "HF_TOKEN is not set — required for transcribe_audio / describe_image."
        )
    return InferenceClient(token=token)


# Whisper is currently served on HF Inference Providers via fal-ai (recommended),
# hf-inference, and replicate. Letting the client auto-route is safest.
_ASR_PROVIDERS = ("auto", "fal-ai", "hf-inference")

# As of 2026 the HF docs' top-recommended VLM is GLM-4.5V; Llama-4 Scout is
# widely deployed via Groq and a good fast fallback.
_VLM_FALLBACKS = (
    "zai-org/GLM-4.5V",
    "meta-llama/Llama-4-Scout-17B-16E-Instruct",
    "Qwen/Qwen3.6-27B",
)


@tool
def transcribe_audio(file_path: str) -> str:
    """Transcribe an audio file (mp3, wav, m4a, flac) to text using Whisper.

    Use after get_task_file for any audio attachment. Returns the full text;
    you can then parse it with Python (regex, splits) to extract the answer.

    Args:
        file_path: Local path to the audio file.
    """
    model = os.getenv("HF_ASR_MODEL", "openai/whisper-large-v3")
    client = _hf_client()
    errors = []
    for provider in _ASR_PROVIDERS:
        try:
            # InferenceClient.automatic_speech_recognition accepts a provider
            # kwarg in recent huggingface_hub; passing "auto" lets HF route.
            kwargs = {"model": model}
            if provider != "auto":
                kwargs["provider"] = provider  # type: ignore[arg-type]
            result = client.automatic_speech_recognition(file_path, **kwargs)
            return getattr(result, "text", None) or str(result)
        except TypeError:
            # Older huggingface_hub: no `provider` kwarg. Try without it once.
            try:
                result = client.automatic_speech_recognition(file_path, model=model)
                return getattr(result, "text", None) or str(result)
            except Exception as e:  # noqa: BLE001
                errors.append(f"{provider}: {e}")
        except Exception as e:  # noqa: BLE001
            errors.append(f"{provider}: {e}")
    return ("ERROR: transcription failed on all providers. "
            "Tried: " + " | ".join(errors))


@tool
def describe_image(file_path: str, question: str) -> str:
    """Ask a vision-language model a question about an image.

    Use for any image attachment (photo, chart, chess board, diagram). Ask a
    SPECIFIC question — vague prompts give vague answers. Examples:
        describe_image(p, "List every item visible on the table, comma-separated.")
        describe_image(p, "What move should White play? Use algebraic notation.")

    Args:
        file_path: Local path to the image file.
        question: A precise question about the image.
    """
    # Build the candidate model list: HF_VLM_MODEL override first, then fallbacks.
    primary = os.getenv("HF_VLM_MODEL")
    candidates = ([primary] if primary else []) + [
        m for m in _VLM_FALLBACKS if m != primary
    ]

    ext = os.path.splitext(file_path)[1].lower().lstrip(".")
    mime = {"jpg": "image/jpeg", "jpeg": "image/jpeg", "png": "image/png",
            "gif": "image/gif", "webp": "image/webp"}.get(ext, "image/jpeg")
    try:
        with open(file_path, "rb") as f:
            b64 = base64.b64encode(f.read()).decode()
    except Exception as e:  # noqa: BLE001
        return f"ERROR: could not read image {file_path}: {e}"

    messages = [{
        "role": "user",
        "content": [
            {"type": "text", "text": question},
            {"type": "image_url",
             "image_url": {"url": f"data:{mime};base64,{b64}"}},
        ],
    }]

    client = _hf_client()
    errors = []
    for model in candidates:
        try:
            resp = client.chat_completion(model=model, messages=messages, max_tokens=512)
            return resp.choices[0].message.content
        except Exception as e:  # noqa: BLE001
            errors.append(f"{model}: {e}")
    return ("ERROR: vision call failed on all candidate models. "
            "Tried: " + " | ".join(errors))


# --- Agent ------------------------------------------------------------------
# ----- THIS IS WHERE YOU CAN BUILD WHAT YOU WANT ------
class GaiaAgent:
    """A tool-using agent (smolagents CodeAgent + DeepSeek V4 Pro) for GAIA.

    Required environment variables:
        DEEPSEEK_API_KEY   - your DeepSeek key (https://platform.deepseek.com)
        HF_TOKEN           - HuggingFace token (auto-set in HF Spaces), used for
                             audio transcription and image VLM calls
    Optional:
        DEEPSEEK_MODEL     - default: "deepseek-v4-flash" (lighter/cheaper V4
                             variant; thinking mode is on by default)
        DEEPSEEK_API_BASE  - default: "https://api.deepseek.com/v1"
        HF_ASR_MODEL       - override ASR model (default: openai/whisper-large-v3,
                             auto-routed via HF Inference Providers)
        HF_VLM_MODEL       - override VLM model. If unset, tries GLM-4.5V then
                             falls back to Llama-4-Scout-17B and Qwen3.6-27B.
    """

    def __init__(self) -> None:
        api_key = os.getenv("DEEPSEEK_API_KEY")
        if not api_key:
            raise RuntimeError(
                "DEEPSEEK_API_KEY is not set. Add it as a Space secret "
                "(Settings > Variables and secrets) or export it locally."
            )

        model = OpenAIServerModel(
            model_id=os.getenv("DEEPSEEK_MODEL", "deepseek-v4-flash"),
            api_base=os.getenv("DEEPSEEK_API_BASE", "https://api.deepseek.com/v1"),
            api_key=api_key,
            temperature=0.0,
        )

        self.agent = CodeAgent(
            tools=[
                DuckDuckGoSearchTool(),
                VisitWebpageTool(),
                fetch_url,
                get_task_file,
                wikipedia_search,
                youtube_transcript,
                youtube_info,
                transcribe_audio,
                describe_image,
            ],
            model=model,
            # Let the sandboxed Python interpreter use these for file parsing.
            additional_authorized_imports=[
                "pandas",
                "numpy",
                "openpyxl",
                "csv",
                "json",
                "math",
                "statistics",
                "datetime",
                "io",
                "re",
                "os",
                "requests",
                "bs4",
            ],
            max_steps=25,
            verbosity_level=1,
        )
        print("GaiaAgent initialized (smolagents CodeAgent + deepseek-v4-flash).")

    def __call__(self, question: str, task_id: str | None = None) -> str:
        print(f"Agent received question (first 80 chars): {question[:80]}...")
        prompt = GAIA_INSTRUCTIONS.format(task_id=task_id or "", question=question)
        result = self.agent.run(prompt)
        answer = self._clean(result)
        print(f"Agent returning answer: {answer}")
        return answer

    @staticmethod
    def _clean(result) -> str:
        """Normalize the agent output to the bare answer string.

        GAIA grades by exact string match, so we strip the noise the model
        commonly leaves around the answer (label prefix, surrounding quotes,
        trailing punctuation) without altering the answer itself.
        """
        text = str(result).strip()
        text = re.sub(r"(?i)^\s*final answer\s*:?\s*", "", text).strip()
        # Drop matching wrapping quotes — "Paris" -> Paris, but leave inner
        # quotes alone.
        if len(text) >= 2 and text[0] == text[-1] and text[0] in ("'", '"', "`"):
            text = text[1:-1].strip()
        # Drop a single trailing period / comma / semicolon.
        text = re.sub(r"[.,;]\s*$", "", text).strip()
        return text

def run_and_submit_all( profile: gr.OAuthProfile | None):
    """
    Fetches all questions, runs the BasicAgent on them, submits all answers,
    and displays the results.
    """
    # --- Determine HF Space Runtime URL and Repo URL ---
    space_id = os.getenv("SPACE_ID") # Get the SPACE_ID for sending link to the code

    if profile:
        username= f"{profile.username}"
        print(f"User logged in: {username}")
    else:
        print("User not logged in.")
        return "Please Login to Hugging Face with the button.", None

    api_url = DEFAULT_API_URL
    questions_url = f"{api_url}/questions"
    submit_url = f"{api_url}/submit"

    # 1. Instantiate Agent ( modify this part to create your agent)
    try:
        agent = GaiaAgent()
    except Exception as e:
        print(f"Error instantiating agent: {e}")
        return f"Error initializing agent: {e}", None
    # In the case of an app running as a hugging Face space, this link points toward your codebase ( usefull for others so please keep it public)
    agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main"
    print(agent_code)

    # 2. Fetch Questions
    print(f"Fetching questions from: {questions_url}")
    try:
        response = requests.get(questions_url, timeout=15)
        response.raise_for_status()
        questions_data = response.json()
        if not questions_data:
             print("Fetched questions list is empty.")
             return "Fetched questions list is empty or invalid format.", None
        print(f"Fetched {len(questions_data)} questions.")
    except requests.exceptions.RequestException as e:
        print(f"Error fetching questions: {e}")
        return f"Error fetching questions: {e}", None
    except requests.exceptions.JSONDecodeError as e:
         print(f"Error decoding JSON response from questions endpoint: {e}")
         print(f"Response text: {response.text[:500]}")
         return f"Error decoding server response for questions: {e}", None
    except Exception as e:
        print(f"An unexpected error occurred fetching questions: {e}")
        return f"An unexpected error occurred fetching questions: {e}", None

    # 3. Run your Agent
    results_log = []
    answers_payload = []
    print(f"Running agent on {len(questions_data)} questions...")
    for item in questions_data:
        task_id = item.get("task_id")
        question_text = item.get("question")
        if not task_id or question_text is None:
            print(f"Skipping item with missing task_id or question: {item}")
            continue
        try:
            submitted_answer = agent(question_text, task_id)
            answers_payload.append({"task_id": task_id, "submitted_answer": submitted_answer})
            results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": submitted_answer})
        except Exception as e:
             print(f"Error running agent on task {task_id}: {e}")
             results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": f"AGENT ERROR: {e}"})

    if not answers_payload:
        print("Agent did not produce any answers to submit.")
        return "Agent did not produce any answers to submit.", pd.DataFrame(results_log)

    # 4. Prepare Submission 
    submission_data = {"username": username.strip(), "agent_code": agent_code, "answers": answers_payload}
    status_update = f"Agent finished. Submitting {len(answers_payload)} answers for user '{username}'..."
    print(status_update)

    # 5. Submit
    print(f"Submitting {len(answers_payload)} answers to: {submit_url}")
    try:
        response = requests.post(submit_url, json=submission_data, timeout=60)
        response.raise_for_status()
        result_data = response.json()
        final_status = (
            f"Submission Successful!\n"
            f"User: {result_data.get('username')}\n"
            f"Overall Score: {result_data.get('score', 'N/A')}% "
            f"({result_data.get('correct_count', '?')}/{result_data.get('total_attempted', '?')} correct)\n"
            f"Message: {result_data.get('message', 'No message received.')}"
        )
        print("Submission successful.")
        results_df = pd.DataFrame(results_log)
        return final_status, results_df
    except requests.exceptions.HTTPError as e:
        error_detail = f"Server responded with status {e.response.status_code}."
        try:
            error_json = e.response.json()
            error_detail += f" Detail: {error_json.get('detail', e.response.text)}"
        except requests.exceptions.JSONDecodeError:
            error_detail += f" Response: {e.response.text[:500]}"
        status_message = f"Submission Failed: {error_detail}"
        print(status_message)
        results_df = pd.DataFrame(results_log)
        return status_message, results_df
    except requests.exceptions.Timeout:
        status_message = "Submission Failed: The request timed out."
        print(status_message)
        results_df = pd.DataFrame(results_log)
        return status_message, results_df
    except requests.exceptions.RequestException as e:
        status_message = f"Submission Failed: Network error - {e}"
        print(status_message)
        results_df = pd.DataFrame(results_log)
        return status_message, results_df
    except Exception as e:
        status_message = f"An unexpected error occurred during submission: {e}"
        print(status_message)
        results_df = pd.DataFrame(results_log)
        return status_message, results_df


# --- Build Gradio Interface using Blocks ---
with gr.Blocks() as demo:
    gr.Markdown("# Basic Agent Evaluation Runner")
    gr.Markdown(
        """
        **Instructions:**

        1.  Please clone this space, then modify the code to define your agent's logic, the tools, the necessary packages, etc ...
        2.  Log in to your Hugging Face account using the button below. This uses your HF username for submission.
        3.  Click 'Run Evaluation & Submit All Answers' to fetch questions, run your agent, submit answers, and see the score.

        ---
        **Disclaimers:**
        Once clicking on the "submit button, it can take quite some time ( this is the time for the agent to go through all the questions).
        This space provides a basic setup and is intentionally sub-optimal to encourage you to develop your own, more robust solution. For instance for the delay process of the submit button, a solution could be to cache the answers and submit in a seperate action or even to answer the questions in async.
        """
    )

    gr.LoginButton()

    run_button = gr.Button("Run Evaluation & Submit All Answers")

    status_output = gr.Textbox(label="Run Status / Submission Result", lines=5, interactive=False)
    # Removed max_rows=10 from DataFrame constructor
    results_table = gr.DataFrame(label="Questions and Agent Answers", wrap=True)

    run_button.click(
        fn=run_and_submit_all,
        outputs=[status_output, results_table]
    )

if __name__ == "__main__":
    print("\n" + "-"*30 + " App Starting " + "-"*30)
    # Check for SPACE_HOST and SPACE_ID at startup for information
    space_host_startup = os.getenv("SPACE_HOST")
    space_id_startup = os.getenv("SPACE_ID") # Get SPACE_ID at startup

    if space_host_startup:
        print(f"✅ SPACE_HOST found: {space_host_startup}")
        print(f"   Runtime URL should be: https://{space_host_startup}.hf.space")
    else:
        print("ℹ️  SPACE_HOST environment variable not found (running locally?).")

    if space_id_startup: # Print repo URLs if SPACE_ID is found
        print(f"✅ SPACE_ID found: {space_id_startup}")
        print(f"   Repo URL: https://huggingface.co/spaces/{space_id_startup}")
        print(f"   Repo Tree URL: https://huggingface.co/spaces/{space_id_startup}/tree/main")
    else:
        print("ℹ️  SPACE_ID environment variable not found (running locally?). Repo URL cannot be determined.")

    print("-"*(60 + len(" App Starting ")) + "\n")

    print("Launching Gradio Interface for Basic Agent Evaluation...")
    demo.launch(debug=True, share=False)