src/DAG2pr.py

import json
import os
import re
import time
import shutil
from pathlib import Path
from typing import Optional, List, Tuple, Any, Dict
from google import genai
from google.genai import types
from openai import OpenAI


# ========== Extract basic information for PR Generation ==========
def extract_basic_information(
    dag_path: str,
    extract_basic_information_prompt: str,
    model: str,
    auto_path: str,
    output_filename: str = "basic_information.txt",
    api_key: Optional[str] = None,
    base_url: Optional[str] = None,
    config: dict = None
) -> str:
    """
    读取 dag.json 的第一个 node（root），将其完整信息发给 LLM，提取并输出：
    Title / Author / Institution / Github
    然后把 LLM 输出写入 auto_path/basic_information.txt（不存在则创建）。

    返回：写入的 txt 的绝对路径（str）
    """
    dag_file = Path(dag_path).expanduser().resolve()
    auto_dir = Path(auto_path).expanduser().resolve()
    out_path = auto_dir / output_filename

    if not dag_file.exists() or not dag_file.is_file():
        raise FileNotFoundError(f"dag_path not found or not a file: {dag_file}")

    auto_dir.mkdir(parents=True, exist_ok=True)

    # 1) load dag.json
    with dag_file.open("r", encoding="utf-8") as f:
        dag = json.load(f)

    nodes = dag.get("nodes", [])
    if not isinstance(nodes, list) or len(nodes) == 0 or not isinstance(nodes[0], dict):
        raise ValueError("Invalid dag.json: missing or invalid 'nodes[0]' root node.")

    root_node = nodes[0]

    # 2) build LLM input (send the entire root node)
    root_payload = {
        "name": root_node.get("name", ""),
        "content": root_node.get("content", ""),
        "github": root_node.get("github", ""),
        "edge": root_node.get("edge", []),
        "level": root_node.get("level", 0),
        "visual_node": root_node.get("visual_node", []),
    }

    user_message = (
        "ROOT_NODE_JSON:\n"
        + json.dumps(root_payload, ensure_ascii=False, indent=2)
    )

    # 3) call LLM
    llm_text = ""
    api_keys_config = config.get("api_keys", {}) if config else {}
    
    # 判别平台
    is_gemini = "gemini" in model.lower()

    if is_gemini:
        raw_url = config.get('api_base_url', '').strip().rstrip("/")
        if raw_url.endswith("/v1"):
            base_url = raw_url[:-3].rstrip("/") # 去掉最后的 /v1
        else:
            base_url = raw_url
        # === Gemini Client Setup ===
        api_key = api_keys_config.get("gemini_api_key") or os.getenv("GOOGLE_API_KEY")
        
        client = genai.Client(api_key=api_key, http_options={'base_url': base_url} if base_url else None)
        
        # Gemini Call
        resp = client.models.generate_content(
            model=model,
            contents=user_message,
            config={
                "system_instruction": extract_basic_information_prompt,
                "temperature": 0,
            }
        )
        if resp.text:
            llm_text = resp.text
            
    else:
        # === OpenAI Client Setup ===
        api_key = api_keys_config.get("openai_api_key") or os.getenv("OPENAI_API_KEY")
        
        client = OpenAI(api_key=api_key)

        # OpenAI Call
        resp = client.chat.completions.create(
            model=model,
            messages=[
                {"role": "system", "content": extract_basic_information_prompt},
                {"role": "user", "content": user_message},
            ],
            temperature=0,
        )
        llm_text = resp.choices[0].message.content or ""

    if not llm_text:
        raise RuntimeError("LLM returned empty content for basic information extraction.")

    # 4) write to auto/basic_information.txt
    with out_path.open("w", encoding="utf-8") as f:
        f.write(llm_text.strip() + "\n")

    return str(out_path)

def initialize_pr_markdown(
    basic_information_path: str,
    auto_path: str,
    pr_template_path: str
) -> str:
    """
    basic_information_path: txt 文件路径（包含 Title/Author/Institution/Github）
    auto_path: 输出目录
    pr_template_path: PR markdown 模板路径
    输出：auto_path/markdown.md
    """

    info_txt_path = Path(basic_information_path)
    auto_dir = Path(auto_path)
    template_md = Path(pr_template_path)

    if not info_txt_path.exists():
        raise FileNotFoundError(f"basic_information_path not found: {basic_information_path}")
    if info_txt_path.suffix.lower() != ".txt":
        raise ValueError(f"basic_information_path must be a .txt file, got: {basic_information_path}")
    if not template_md.exists():
        raise FileNotFoundError(f"pr_template_path not found: {pr_template_path}")

    auto_dir.mkdir(parents=True, exist_ok=True)

    # -------- 1) 读取并解析 txt --------
    txt = info_txt_path.read_text(encoding="utf-8", errors="ignore")

    def _extract_value(key: str) -> str:
        pattern = re.compile(
            rf"^\s*{re.escape(key)}\s*:\s*(.*?)\s*$",
            re.IGNORECASE | re.MULTILINE
        )
        m = pattern.search(txt)
        return m.group(1).strip() if m else ""

    title = _extract_value("Title")
    authors = _extract_value("Author") or _extract_value("Authors")
    institution = _extract_value("Institution")
    github = _extract_value("Github") or _extract_value("GitHub") or _extract_value("Direct Link")

    # -------- 2) 复制模板到 auto_path/markdown.md --------
    out_md_path = auto_dir / "markdown.md"
    shutil.copyfile(template_md, out_md_path)

    md = out_md_path.read_text(encoding="utf-8", errors="ignore")

    # 关键：统一换行，避免 \r\n 导致整行匹配失败
    md = md.replace("\r\n", "\n").replace("\r", "\n")

    # -------- 3) 按行替换（更鲁棒）--------
    def _fill_line_by_anchor(md_text: str, anchor_label: str, value: str) -> str:
        """
        anchor_label: 例如 "Authors"，用于匹配 **Authors**
        将包含 **anchor_label** 的那一行，改写为：
          <该行中 **anchor_label** 及其之前部分>: <value>
        例如：
          ✍️ **Authors**      -> ✍️ **Authors:** xxx
          🏛️**Institution**  -> 🏛️**Institution:** yyy
        """
        if not value:
            return md_text

        anchor = f"**{anchor_label}**"
        lines = md_text.split("\n")
        replaced = False

        for i, line in enumerate(lines):
            if anchor in line:
                # 保留锚点之前的所有内容 + 锚点本身
                left = line.split(anchor, 1)[0] + anchor
                lines[i] = f"{left}: {value}"
                replaced = True
                break

        return "\n".join(lines) if replaced else md_text

    md = _fill_line_by_anchor(md, "Authors", authors)
    md = _fill_line_by_anchor(md, "Direct Link", github)
    md = _fill_line_by_anchor(md, "Paper Title", title)
    md = _fill_line_by_anchor(md, "Institution", institution)

    out_md_path.write_text(md, encoding="utf-8")

    return str(out_md_path)


def generate_pr_from_dag(
    dag_path: str,
    pr_path: str,
    generate_pr_prompt: str,
    model: str = "gpt-4o-mini",
    timeout: int = 120,
    debug: bool = True,
    max_retries: int = 5,
    backoff_base: float = 1.5,
    max_content_chars: int = 12000,
    max_visuals: int = 20,
    config: dict = None
) -> None:
    """
    Generate a PR markdown from dag.json by iterating section nodes and querying an LLM.
    Supports both OpenAI and Google Gemini APIs using official SDKs.
    """

    def log(msg: str) -> None:
        if debug:
            print(msg)

    # -------------------------
    # 0) Env: API key & Client Setup
    # -------------------------
    if config is None:
        config = {}
    
    api_keys_conf = config.get("api_keys", {})
    
    # Simple heuristic to determine provider based on model name
    is_gemini = "gemini" in model.lower()
    
    client = None

    if is_gemini:
        raw_url = config.get('api_base_url', '').strip().rstrip("/")
        if raw_url.endswith("/v1"):
            base_url = raw_url[:-3].rstrip("/") # 去掉最后的 /v1
        else:
            base_url = raw_url
        # Gemini Configuration (New SDK)
        if genai is None:
            raise ImportError("Google GenAI SDK not installed. Please run `pip install google-genai`.")

        api_key = api_keys_conf.get("gemini_api_key") or os.getenv("GEMINI_API_KEY", "").strip()
        if not api_key:
            raise RuntimeError("API Key not found for Gemini.")
        
        client = genai.Client(api_key=api_key, http_options={'base_url': base_url} if base_url else None)
        
    else:
        # OpenAI Configuration
        if OpenAI is None:
            raise ImportError("OpenAI SDK not installed. Please run `pip install openai`.")

        api_key = api_keys_conf.get("openai_api_key") or os.getenv("OPENAI_API_KEY", "").strip()
        
        if not api_key:
            raise RuntimeError("API Key not found for OpenAI.")

        client = OpenAI(
            api_key=api_key,
            timeout=timeout,
            max_retries=0 # We handle retries manually below
        )

    log(f"[INFO] Using provider = {'Gemini' if is_gemini else 'OpenAI'}")
    log(f"[INFO] Using model = {model}")

    # -------------------------
    # 1) Load DAG
    # -------------------------
    dag_obj = json.loads(Path(dag_path).read_text(encoding="utf-8"))
    if isinstance(dag_obj, list):
        nodes = dag_obj
    elif isinstance(dag_obj, dict) and isinstance(dag_obj.get("nodes"), list):
        nodes = dag_obj["nodes"]
    else:
        raise ValueError("Unsupported dag.json format (expect list or dict with 'nodes').")

    if not nodes:
        log("[WARN] dag.json has no nodes; exiting.")
        return

    root = nodes[0]
    root_edges = root.get("edge", [])
    if not isinstance(root_edges, list):
        log("[WARN] Root node edge is not a list; exiting.")
        return

    log(f"[INFO] Root node name = {root.get('name')}")
    log(f"[INFO] Root edges (section names) = {root_edges}")

    name2node: Dict[str, Dict[str, Any]] = {}
    for n in nodes:
        nm = n.get("name")
        if isinstance(nm, str) and nm:
            name2node[nm] = n

    section_nodes: List[Dict[str, Any]] = []
    for sec_name in root_edges:
        if sec_name in name2node:
            section_nodes.append(name2node[sec_name])
        else:
            log(f"[WARN] Section node '{sec_name}' not found in DAG; skipped.")

    log(f"[INFO] Resolved {len(section_nodes)} section nodes.")
    if not section_nodes:
        log("[WARN] No section nodes to process; exiting.")
        return

    # -------------------------
    # 2) Load PR markdown
    # -------------------------
    pr_file = Path(pr_path)
    pr_text = pr_file.read_text(encoding="utf-8")

    KNOWN_LABELS = ["Key Question", "Brilliant Idea", "Core Methods", "Core Results", "Significance/Impact"]

    if debug:
        log("[INFO] PR template header scan:")
        for lab in KNOWN_LABELS:
            found = bool(re.search(rf"(?mi)^.*\*\*{re.escape(lab)}\*\*.*$", pr_text))
            log(f"  - {lab}: found={found}")

    filled_key_question = False
    filled_brilliant_idea = False
    filled_significance = False
    core_methods_inlined = False
    core_results_inlined = False

    # -------------------------
    # 3) LLM call (OpenAI & Gemini SDKs)
    # -------------------------
    def chat_complete(prompt: str, payload_meta: str = "") -> str:
        system_msg = "You are a precise scientific writing assistant."
        
        last_err: Optional[Exception] = None
        
        for attempt in range(1, max_retries + 1):
            try:
                if is_gemini:
                    # Gemini (Google GenAI SDK)
                    response = client.models.generate_content(
                        model=model,
                        contents=prompt,
                        config=types.GenerateContentConfig(
                            system_instruction=system_msg,
                            temperature=0.4,
                        )
                    )
                    # 检查是否因为安全原因被拦截
                    if not response.text:
                        log(f"[WARN] Gemini response empty (possibly safety blocked) {payload_meta}")
                        return ""
                    return response.text.strip()
                
                else:
                    # OpenAI (Official SDK)
                    response = client.chat.completions.create(
                        model=model,
                        messages=[
                            {"role": "system", "content": system_msg},
                            {"role": "user", "content": prompt},
                        ],
                        temperature=0.4,
                    )
                    content = response.choices[0].message.content
                    return content.strip() if content else ""

            except Exception as e:
                # 捕获 SDK 抛出的各类异常 (RateLimit, APIError, ConnectionError 等)
                last_err = e
                sleep_s = backoff_base ** (attempt - 1)
                log(f"[LLM-RETRY] attempt {attempt}/{max_retries} error: {repr(e)}; sleep {sleep_s:.2f}s {payload_meta}")
                time.sleep(sleep_s)

        raise RuntimeError(f"LLM request failed after {max_retries} retries. Last error: {repr(last_err)}")

    # -------------------------
    # 4) Parse LLM output
    # -------------------------
    def parse_llm_output(text: str) -> Dict[str, Any]:
        out: Dict[str, Any] = {
            "type": "unknown",
            "key_question": None,
            "brilliant_idea": None,
            "core_methods": None,
            "core_results": None,
            "significance": None,
            "image": None,
        }

        img = re.search(r'!\[\]\(([^)]+)\)', text)
        if img:
            out["image"] = f"![]({img.group(1).strip()})"

        def grab(label: str) -> Optional[str]:
            m = re.search(
                rf"(?is)(?:^{re.escape(label)}\s*[:：]\s*)(.*?)(?=^\w[\w /-]*\s*[:：]|\Z)",
                text,
                flags=re.MULTILINE,
            )
            return m.group(1).strip() if m else None

        out["key_question"] = grab("Key Question")
        out["brilliant_idea"] = grab("Brilliant Idea")
        out["core_methods"] = grab("Core Methods")
        out["core_results"] = grab("Core Results")
        out["significance"] = grab("Significance/Impact")

        if out["key_question"] or out["brilliant_idea"]:
            out["type"] = "intro"
        elif out["core_methods"]:
            out["type"] = "methods"
        elif out["core_results"]:
            out["type"] = "results"
        elif out["significance"]:
            out["type"] = "impact"

        return out

    # -------------------------
    # 5) PR template helpers (emoji-safe + APPEND inside block)
    # -------------------------
    def _find_header_line(md: str, label: str) -> Optional[re.Match]:
        return re.search(rf"(?mi)^(?P<line>.*\*\*{re.escape(label)}\*\*.*)$", md)

    def has_header(md: str, label: str) -> bool:
        found = _find_header_line(md, label) is not None
        log(f"[CHECK] Header '{label}' found = {found}")
        return found

    def set_inline(md: str, label: str, text: str) -> str:
        m = _find_header_line(md, label)
        if not m:
            log(f"[SKIP] set_inline: header '{label}' not found")
            return md

        line = m.group("line")
        token_pat = re.compile(rf"\*\*{re.escape(label)}\*\*", re.I)
        tm = token_pat.search(line)
        if not tm:
            log(f"[SKIP] set_inline: token '**{label}**' not found in line")
            return md

        prefix = line[:tm.end()]  # includes **Label**
        new_line = f"{prefix}: {text}".rstrip()

        start, end = m.start("line"), m.end("line")
        return md[:start] + new_line + md[end:]

    def _next_header_pos(md: str, start_pos: int) -> int:
        tail = md[start_pos:]
        next_pos = None
        for lab in KNOWN_LABELS:
            mm = re.search(rf"(?mi)^\s*.*\*\*{re.escape(lab)}\*\*.*$", tail)
            if mm:
                cand = start_pos + mm.start()
                if next_pos is None or cand < next_pos:
                    next_pos = cand
        return next_pos if next_pos is not None else len(md)

    def append_to_label_block(md: str, label: str, insertion_lines: List[str]) -> str:
        m = _find_header_line(md, label)
        if not m:
            log(f"[SKIP] append_to_label_block: header '{label}' not found")
            return md

        insertion = "\n".join([ln for ln in insertion_lines if ln and ln.strip()]).rstrip()
        if not insertion:
            log(f"[SKIP] append_to_label_block: empty insertion for '{label}'")
            return md

        hdr_end = m.end("line")
        block_end = _next_header_pos(md, hdr_end)

        before = md[:hdr_end]
        middle = md[hdr_end:block_end]
        after = md[block_end:]

        if not before.endswith("\n"):
            before += "\n"
        if middle and not middle.startswith("\n"):
            middle = "\n" + middle

        middle_stripped_right = middle.rstrip("\n")
        if middle_stripped_right.strip() == "":
            new_middle = "\n" + insertion + "\n"
        else:
            new_middle = middle_stripped_right + "\n\n" + insertion + "\n"

        return before + new_middle + after

    # -------------------------
    # 6) Main loop over sections
    # -------------------------
    for idx, sec in enumerate(section_nodes):
        log("\n" + "=" * 90)
        log(f"[SECTION {idx}] name = {sec.get('name')}")

        content = sec.get("content", "")
        if not isinstance(content, str):
            content = ""

        visuals = sec.get("visual_node", [])
        if not isinstance(visuals, list):
            visuals = []

        visuals_norm: List[str] = []
        for v in visuals:
            if isinstance(v, str):
                visuals_norm.append(v)
            elif isinstance(v, dict):
                for k in ("path", "src", "url", "md", "markdown"):
                    if k in v and isinstance(v[k], str):
                        visuals_norm.append(v[k])
                        break

        if len(content) > max_content_chars:
            content = content[:max_content_chars] + "\n...(truncated)"

        if len(visuals_norm) > max_visuals:
            visuals_norm = visuals_norm[:max_visuals]

        node_obj = {
            "name": sec.get("name", ""),
            "content": content,
            "visual_node": visuals_norm,
        }

        if debug:
            log(f"[PAYLOAD] content_chars={len(node_obj['content'])} visuals={len(node_obj['visual_node'])}")
            preview = dict(node_obj)
            if len(preview["content"]) > 800:
                preview["content"] = preview["content"][:800] + "...(truncated)"
            log("[SEND TO LLM] Node payload preview:")
            print(json.dumps(preview, indent=2, ensure_ascii=False))

        prompt = generate_pr_prompt.replace("{NODE_JSON}", json.dumps(node_obj, ensure_ascii=False))
        llm_text = chat_complete(prompt, payload_meta=f"(section_idx={idx}, section_name={node_obj['name']})")

        log("\n[LLM RAW OUTPUT]")
        print(llm_text)

        parsed = parse_llm_output(llm_text)
        log("\n[PARSED OUTPUT]")
        print(json.dumps(parsed, indent=2, ensure_ascii=False))

        if parsed["type"] == "intro":
            log("[TYPE] Introduction-like")
            kq = parsed.get("key_question")
            bi = parsed.get("brilliant_idea")
            img = parsed.get("image")

            if kq and not filled_key_question:
                if has_header(pr_text, "Key Question"):
                    pr_text = set_inline(pr_text, "Key Question", kq)
                    filled_key_question = True
                    log("[WRITE] Key Question filled (first time).")
                else:
                    log("[MISS] PR has no Key Question header.")
            else:
                log("[SKIP] Key Question ignored (empty or already filled).")

            if bi and not filled_brilliant_idea:
                if has_header(pr_text, "Brilliant Idea"):
                    pr_text = set_inline(pr_text, "Brilliant Idea", bi)
                    if img:
                        pr_text = append_to_label_block(pr_text, "Brilliant Idea", [img])
                    filled_brilliant_idea = True
                    log("[WRITE] Brilliant Idea filled (first time).")
                else:
                    log("[MISS] PR has no Brilliant Idea header.")
            else:
                log("[SKIP] Brilliant Idea ignored (empty or already filled).")

        elif parsed["type"] == "methods":
            log("[TYPE] Methods-like")
            cm = parsed.get("core_methods")
            img = parsed.get("image")

            if not cm:
                log("[SKIP] Core Methods empty.")
                continue
            if not has_header(pr_text, "Core Methods"):
                log("[MISS] PR has no Core Methods header.")
                continue

            if not core_methods_inlined:
                pr_text = set_inline(pr_text, "Core Methods", cm)
                core_methods_inlined = True
                log("[WRITE] Core Methods inlined (first time).")
                if img:
                    pr_text = append_to_label_block(pr_text, "Core Methods", [img])
                    log("[WRITE] Core Methods image appended.")
            else:
                lines = [cm] + ([img] if img else [])
                pr_text = append_to_label_block(pr_text, "Core Methods", lines)
                log("[WRITE] Core Methods appended as (text, image) pair.")

        elif parsed["type"] == "results":
            log("[TYPE] Results-like")
            cr = parsed.get("core_results")
            img = parsed.get("image")

            if not cr:
                log("[SKIP] Core Results empty.")
                continue
            if not has_header(pr_text, "Core Results"):
                log("[MISS] PR has no Core Results header.")
                continue

            if not core_results_inlined:
                pr_text = set_inline(pr_text, "Core Results", cr)
                core_results_inlined = True
                log("[WRITE] Core Results inlined (first time).")
                if img:
                    pr_text = append_to_label_block(pr_text, "Core Results", [img])
                    log("[WRITE] Core Results image appended.")
            else:
                lines = [cr] + ([img] if img else [])
                pr_text = append_to_label_block(pr_text, "Core Results", lines)
                log("[WRITE] Core Results appended as (text, image) pair.")

        elif parsed["type"] == "impact":
            log("[TYPE] Impact-like")
            si = parsed.get("significance")
            if not si:
                log("[SKIP] Significance/Impact empty.")
                continue
            if filled_significance:
                log("[SKIP] Significance/Impact ignored (already filled).")
                continue
            if has_header(pr_text, "Significance/Impact"):
                pr_text = set_inline(pr_text, "Significance/Impact", si)
                filled_significance = True
                log("[WRITE] Significance/Impact filled (first time).")
            else:
                log("[MISS] PR has no Significance/Impact header.")

        else:
            log("[WARN] Unknown section type; ignored.")

    # -------------------------
    # 7) Save
    # -------------------------
    pr_file.write_text(pr_text, encoding="utf-8")
    log("\n[SAVED] PR markdown updated in-place.")


# ================================增加标题和hashtag=================================
def add_title_and_hashtag(pr_path: str, add_title_and_hashtag_prompt: str, model: str = "gpt-4o-mini", config: dict = None) -> None:
    """
    1) Read markdown from pr_path.
    2) Send it to LLM using add_title_and_hashtag_prompt (expects {MD_TEXT} placeholder).
    3) Parse LLM output.
    4) Update file in-place.
    """
    
    # 确保 config 不为空
    if config is None:
        config = {}
    api_keys = config.get('api_keys', {})

    # -------------------------
    # 0) Read markdown
    # -------------------------
    pr_file = Path(pr_path)
    if not pr_file.exists():
        raise FileNotFoundError(f"pr_path not found: {pr_path}")

    md_text = pr_file.read_text(encoding="utf-8")

    # -------------------------
    # 1) Call LLM (Modified for Dual Support)
    # -------------------------
    prompt_content = add_title_and_hashtag_prompt.replace("{MD_TEXT}", md_text)
    system_instruction = "You are a precise scientific social media copywriter."
    llm_out = ""

    # >>> 分支判断逻辑 >>>
    if "gemini" in model.lower():
        raw_url = config.get('api_base_url', '').strip().rstrip("/")
        if raw_url.endswith("/v1"):
            base_url = raw_url[:-3].rstrip("/") # 去掉最后的 /v1
        else:
            base_url = raw_url
        # --- Google Gemini (New SDK) ---
        api_key = api_keys.get("gemini_api_key", "").strip()
        
        if not api_key:
            raise ValueError("Missing config['api_keys']['gemini_api_key']")

        try:
            from google import genai
            from google.genai import types
        except ImportError as e:
            raise ImportError("google-genai package is required. Install with: pip install google-genai") from e

        # 配置 client        
        client = genai.Client(api_key=api_key, http_options={'base_url': base_url} if base_url else None)

        try:
            response = client.models.generate_content(
                model=model,
                contents=prompt_content,
                config=types.GenerateContentConfig(
                    system_instruction=system_instruction,
                    temperature=0.6
                )
            )
            llm_out = response.text
        except Exception as e:
            raise RuntimeError(f"Gemini API call failed: {e}")

    else:
        # --- OpenAI (Existing) ---
        api_key = api_keys.get("openai_api_key", "").strip()

        if not api_key:
            # 兼容旧逻辑，如果没有传 config，尝试读环境变量（可选，视你需求而定）
            api_key = os.getenv("OPENAI_API_KEY", "").strip()
        
        if not api_key:
            raise ValueError("Missing config['api_keys']['openai_api_key']")

        try:
            from openai import OpenAI
        except ImportError as e:
            raise ImportError("openai package is required. Install with: pip install openai") from e

        client = OpenAI(api_key=api_key)

        try:
            resp = client.chat.completions.create(
                model=model,
                messages=[
                    {"role": "system", "content": system_instruction},
                    {"role": "user", "content": prompt_content},
                ],
                temperature=0.6,
            )
            llm_out = (resp.choices[0].message.content or "").strip()
        except Exception as e:
            raise RuntimeError(f"OpenAI API call failed: {e}")

    if not llm_out:
        raise ValueError("LLM returned empty content.")

    # -------------------------
    # 2) Parse LLM output (Remaining logic unchanged)
    # -------------------------
    # 注意：这里假设 _parse_title_and_tags 和 _line_ending 已经在外部定义或在此作用域内可用
    title, specific_tags, community_tag = _parse_title_and_tags(llm_out)

    # -------------------------
    # 3) Update first line: "# {Title}"
    # -------------------------
    lines = md_text.splitlines(True) 
    if not lines:
        raise ValueError("Markdown file is empty.")

    first_line = lines[0].rstrip("\r\n")
    # 辅助函数：如果原代码中 _line_ending 未定义，需自行补充。此处沿用原逻辑。
    # 假设 _line_ending(s) 返回 s 的换行符
    
    def _local_line_ending(s):
        if s.endswith("\r\n"): return "\r\n"
        if s.endswith("\n"): return "\n"
        return "\n" # default

    current_ending = _local_line_ending(lines[0])

    if re.fullmatch(r"#\s*", first_line):
        lines[0] = f"# 🔥{title}😯{current_ending}"
    else:
        if first_line.startswith("#"):
            lines[0] = re.sub(r"^#\s*.*$", f"# 🔥{title}😯", first_line) + current_ending
        else:
            lines.insert(0, f"# 🔥{title}😯\n")

    updated = "".join(lines)

    # -------------------------
    # 4) Replace "Specific:" line
    # -------------------------
    def _replace_specific_line(match: re.Match) -> str:
        prefix = match.group(1)
        tail = match.group(2) or ""
        has_semicolon = ";" in tail
        end = ";" if has_semicolon else ""
        return f"{prefix}{specific_tags[0]} {specific_tags[1]} {specific_tags[2]}{end}"

    updated, n1 = re.subn(
        r"(?mi)^(Specific:\s*)(.*)$",
        lambda m: _replace_specific_line(m),
        updated,
        count=1,
    )

    # -------------------------
    # 5) Replace "Community:" line
    # -------------------------
    updated, n2 = re.subn(
        r"(Community:\s*)(#[^\s;]+)",
        lambda m: f"{m.group(1)}{community_tag}",
        updated,
        count=1,
        flags=re.IGNORECASE,
    )

    if n1 == 0:
        raise ValueError("Could not find a line starting with 'Specific:' to replace.")
    if n2 == 0:
        raise ValueError("Could not find the 'Community: #Tag1' pattern to replace.")

    # -------------------------
    # 6) Write back
    # -------------------------
    pr_file.write_text(updated, encoding="utf-8")


def _parse_title_and_tags(llm_out: str) -> Tuple[str, List[str], str]:
    """
    Parse:
      Title: ...
      Specific Tag: #A #B #C
      Community Tag: #X
    """
    def pick_line(prefix: str) -> str:
        m = re.search(rf"^{re.escape(prefix)}\s*(.+)$", llm_out, flags=re.MULTILINE)
        if not m:
            raise ValueError(f"LLM output missing line: '{prefix} ...'")
        return m.group(1).strip()

    title = pick_line("Title:")
    spec = pick_line("Specific Tag:")
    comm = pick_line("Community Tag:")

    spec_tags = re.findall(r"#[A-Za-z0-9_]+", spec)
    comm_tags = re.findall(r"#[A-Za-z0-9_]+", comm)

    if len(spec_tags) != 3:
        raise ValueError(f"Expected exactly 3 specific tags, got {len(spec_tags)}. Raw: {spec}")
    if len(comm_tags) != 1:
        raise ValueError(f"Expected exactly 1 community tag, got {len(comm_tags)}. Raw: {comm}")

    title = title.strip().strip('"').strip("'")
    if not title:
        raise ValueError("Parsed title is empty.")

    return title, spec_tags, comm_tags[0]


def _line_ending(original_line: str) -> str:
    if original_line.endswith("\r\n"):
        return "\r\n"
    if original_line.endswith("\n"):
        return "\n"
    return "\n"




# ===========================增加机构的tag===============================
def add_institution_tag(pr_path: str) -> None:
    """
    读取 markdown 中 🏛️**Institution**: 后的所有机构名（鲁棒分割），然后：
    1) 将所有机构名拼成一行，写入 Strategic Mentions：# 后面，例如：
       Strategic Mentions：# NVIDIA, Tel Aviv University
    2) 在 Strategic Mentions 这一行的下一行追加一行：
       @NVIDIA, Tel Aviv University
    原地修改文件。
    """

    md_path = Path(pr_path)
    if not md_path.exists():
        raise FileNotFoundError(f"Markdown file not found: {pr_path}")

    text = md_path.read_text(encoding="utf-8")

    # -------------------------------------------------
    # 1) 提取 Institution 行内容
    # -------------------------------------------------
    institution_pattern = re.compile(r"🏛️\s*\*\*Institution\*\*\s*:\s*(.+)")
    m = institution_pattern.search(text)
    if not m:
        raise ValueError("Institution section not found in markdown.")

    institution_raw = m.group(1).strip()

    # -------------------------------------------------
    # 2) 鲁棒切分：提取所有机构名
    #    支持中英文分号/逗号/顿号/竖线/斜杠/换行/制表符等
    # -------------------------------------------------
    split_pattern = re.compile(r"\s*(?:;|；|,|，|、|\||｜|/|\n|\t)\s*")
    parts = [p.strip() for p in split_pattern.split(institution_raw) if p.strip()]
    if not parts:
        raise ValueError("Institution content is empty after parsing.")

    # 去除常见尾部标点噪声，并保持原顺序去重
    seen = set()
    institutions = []
    for p in parts:
        p = p.strip(" .。")
        if p and p not in seen:
            institutions.append(p)
            seen.add(p)

    if not institutions:
        raise ValueError("No valid institution names parsed.")

    institutions_str = ", ".join(institutions)

    # -------------------------------------------------
    # 3) 替换 Strategic Mentions：# 行，并在其下一行插入 @...
    #    仅处理第一次出现
    # -------------------------------------------------
    strategic_line_pattern = re.compile(r"^(Strategic Mentions：\s*#).*$", re.MULTILINE)

    def repl(mm: re.Match) -> str:
        prefix = mm.group(1)
        return f"{prefix} {institutions_str}\n@{institutions_str}"

    new_text, n = strategic_line_pattern.subn(repl, text, count=1)
    if n == 0:
        raise ValueError("Strategic Mentions section not found in markdown.")

    md_path.write_text(new_text, encoding="utf-8")


# =======================删除重复图片引用======================
def dedup_consecutive_markdown_images(md_path: str, inplace: bool = True) -> Tuple[str, int]:
    """
    去除 Markdown 中“连续出现”的相同图片引用（按 src 判重），只保留一个。
    连续的定义：两张图片之间只要是空白字符（空格/Tab/换行）也视为连续；
              若中间出现任何非空白内容（文字、代码、列表符号等），则不算连续。

    支持：
      ![](path)
      ![alt](path)
      ![alt](path "title")

    返回：
      (new_text, removed_count)
    """

    p = Path(md_path)
    text = p.read_text(encoding="utf-8")

    img_pat = re.compile(
        r'!\[(?P<alt>[^\]]*)\]\((?P<src>[^)\s]+)(?:\s+"[^"]*")?\)',
        flags=re.MULTILINE
    )

    parts = []
    last = 0
    for m in img_pat.finditer(text):
        if m.start() > last:
            parts.append(("text", text[last:m.start()]))
        parts.append(("img", m.group(0), m.group("src")))
        last = m.end()
    if last < len(text):
        parts.append(("text", text[last:]))

    removed = 0
    new_parts = []

    prev_img_src = None
    # 只有遇到“非空白文本”才会真正打断连续性
    saw_non_whitespace_since_prev_img = True  # 初始视为打断状态

    for part in parts:
        if part[0] == "img":
            _, raw_img, src = part

            # 连续重复判定：上一张图存在 + 中间没有非空白内容 + src 相同
            if prev_img_src is not None and (not saw_non_whitespace_since_prev_img) and prev_img_src == src:
                removed += 1
                # 删除该图片（不追加）
                continue

            # 保留该图片
            new_parts.append(raw_img)
            prev_img_src = src
            saw_non_whitespace_since_prev_img = False

        else:
            _, raw_text = part
            new_parts.append(raw_text)

            # 只有出现非空白，才视为打断连续图片序列
            if raw_text.strip() != "":
                saw_non_whitespace_since_prev_img = True

    new_text = "".join(new_parts)

    if inplace and removed > 0:
        p.write_text(new_text, encoding="utf-8")

    return new_text, removed