git-codereview/app/lib/summary.js

/**
 * Review-history summariser.
 *
 * Deliberately separate from the OpenCodeReview pipeline: OCR decides what is
 * wrong with a diff; this module turns a finished review into a short,
 * uniformly formatted record for the history list. It shares the global LLM
 * settings but uses its own prompt, and it never influences blocking or merge
 * decisions.
 */

export const SUMMARY_MAX_CHARS = 500;
export const SUMMARY_PROMPT_VERSION = "review-history-summary/v1";
export const SUMMARY_SECTIONS = ["结论", "范围", "问题", "要点"];

const SYSTEM_PROMPT = [
  "你是一名代码审查记录整理员。你会收到一次自动代码审查的结构化结果。",
  "你的唯一任务是把这次审查整理成一段简洁、客观的中文记录，供团队在审查历史里快速浏览。",
  "",
  "严格要求：",
  "1. 只依据输入内容，不要推测、不要补充未出现的信息，不要提出新建议。",
  "2. 不要评价审查工具本身，不要输出“建议进一步检查”之类的话。",
  "3. 按下面的固定格式输出，保留小标题，每节一行到两行：",
  "结论：<通过 / 有阻断问题 / 有非阻断问题 / 未发现变更>",
  "范围：<审查了哪些文件、多少行；信息不足写“未知”>",
  "问题：<按严重级别归纳问题类型与数量；无问题写“无”>",
  "要点：<最值得关注的一到三条具体问题，写明文件与行为；无问题写“无”>",
  "4. 全文不超过 500 个字符，不要使用 Markdown 标题符号、代码块或列表符号。",
  "5. 直接输出上述四节内容，不要任何前言、后缀或解释。",
].join("\n");

function clip(text, max) {
  const value = String(text ?? "");
  return value.length <= max ? value : `${value.slice(0, max)}…`;
}

/** Build the user message describing one finished review. */
export function buildSummaryPrompt({ repo, record, findings, summary, requirement }) {
  const lines = [];
  lines.push(`仓库：${repo.owner}/${repo.name}`);
  lines.push(`分支：${record.ref_name}`);
  if (record.pr_number) lines.push(`Pull Request：#${record.pr_number}`);
  lines.push(`提交：${record.to_sha}`);
  if (record.from_sha) lines.push(`对比基准：${record.from_sha}`);
  if (requirement) lines.push(`本次需求/背景：${clip(requirement, 600)}`);
  if (summary) {
    const bits = [];
    if (summary.files_reviewed != null) bits.push(`审查文件 ${summary.files_reviewed} 个`);
    if (summary.comments != null) bits.push(`原始意见 ${summary.comments} 条`);
    if (summary.total_tokens) bits.push(`消耗 ${summary.total_tokens} tokens`);
    if (summary.elapsed) bits.push(`耗时 ${summary.elapsed}`);
    if (bits.length) lines.push(`运行数据：${bits.join("，")}`);
  }
  if (record.llm_status) lines.push(`审查状态：${record.llm_status}`);

  lines.push("", "审查发现：");
  if (!findings.length) {
    lines.push("（本次没有产生任何意见）");
  } else {
    for (const f of findings) {
      const where = f.start_line
        ? `${f.path}:${f.start_line}${f.end_line && f.end_line !== f.start_line ? `-${f.end_line}` : ""}`
        : f.path;
      const tags = [f.severity, f.category].filter(Boolean).join("/");
      lines.push(`- [${tags || "未分级"}] ${where} ${clip(f.content, 300)}`);
    }
  }
  return lines.join("\n");
}

/**
 * Call an OpenAI- or Anthropic-compatible endpoint with the built-in fetch, so
 * the runtime image needs no extra HTTP client.
 */
async function runLlm(llm, messages, timeoutMs) {
  const base = String(llm.url).replace(/\/+$/, "");
  const isAnthropic = String(llm.protocol || "").toLowerCase().includes("anthropic");
  const endpoint = isAnthropic
    ? (base.endsWith("/v1") ? `${base}/messages` : `${base}/v1/messages`)
    : (base.endsWith("/v1") ? `${base}/chat/completions` : `${base}/v1/chat/completions`);

  const headers = { "Content-Type": "application/json" };
  if (isAnthropic) {
    headers[llm.authHeader || "x-api-key"] = llm.token;
    headers["anthropic-version"] = "2023-06-01";
  } else {
    headers.Authorization = `Bearer ${llm.token}`;
  }
  for (const pair of String(llm.extraHeaders || "").split(",")) {
    const [k, ...rest] = pair.split("=");
    if (k && rest.length) headers[k.trim()] = rest.join("=").trim();
  }

  const payload = isAnthropic
    ? {
        model: llm.model,
        max_tokens: 1024,
        system: messages[0].content,
        messages: [{ role: "user", content: messages[1].content }],
      }
    : { model: llm.model, messages, max_tokens: 1024, stream: false };

  let res;
  try {
    res = await fetch(endpoint, {
      method: "POST",
      headers,
      body: JSON.stringify(payload),
      signal: AbortSignal.timeout(timeoutMs),
    });
  } catch (err) {
    const reason = err.name === "TimeoutError" ? `请求超时（${timeoutMs} ms）` : err.message;
    return { ok: false, error: `调用 LLM 失败：${reason}` };
  }

  const text = await res.text();
  let parsed;
  try {
    parsed = JSON.parse(text);
  } catch {
    return { ok: false, error: `无法解析 LLM 响应（HTTP ${res.status}）：${clip(text, 200)}` };
  }
  if (!res.ok) {
    const msg = typeof parsed.error === "string"
      ? parsed.error
      : (parsed.error?.message || JSON.stringify(parsed.error || parsed));
    return { ok: false, error: `LLM 返回 HTTP ${res.status}：${clip(msg, 200)}` };
  }
  const content = isAnthropic
    ? (parsed.content || []).map((p) => p.text || "").join("")
    : (parsed.choices?.[0]?.message?.content ?? "");
  if (!content) return { ok: false, error: `LLM 返回空内容：${clip(text, 200)}` };
  return { ok: true, text: String(content).trim(), servedModel: parsed.model || null };
}

/** Normalise the model output into the fixed four-section shape. */
export function normaliseSummary(text) {
  const cleaned = String(text || "")
    .replace(/```[a-z]*\n?/gi, "")
    .replace(/^\s*#{1,6}\s*/gm, "")
    .trim();
  const found = SUMMARY_SECTIONS.filter((s) => new RegExp(`^\\s*${s}[:：]`, "m").test(cleaned));
  const collapsed = cleaned.replace(/\n{3,}/g, "\n\n");
  const clipped = collapsed.length > SUMMARY_MAX_CHARS
    ? `${collapsed.slice(0, SUMMARY_MAX_CHARS - 1)}…`
    : collapsed;
  return { text: clipped, complete: found.length === SUMMARY_SECTIONS.length, sectionsFound: found };
}

/**
 * Summarise one review record.
 * @returns {Promise<{ok: boolean, text?: string, error?: string, model?: string}>}
 */
export async function summariseReview({ llm, repo, record, findings, summary, requirement, timeoutMs = 120000 }) {
  if (!llm?.url || !llm?.token || !llm?.model) {
    return { ok: false, error: "未配置全局 LLM，无法生成审查摘要" };
  }
  const messages = [
    { role: "system", content: SYSTEM_PROMPT },
    { role: "user", content: buildSummaryPrompt({ repo, record, findings, summary, requirement }) },
  ];
  const res = await runLlm(llm, messages, timeoutMs);
  if (!res.ok) return res;
  const normalised = normaliseSummary(res.text);
  return { ok: true, text: normalised.text, model: res.servedModel || llm.model, complete: normalised.complete };
}

export { SYSTEM_PROMPT };