AgentHub/server/lib/cursor-usage.js

/**
 * @file Cursor sessions, read from the agent transcripts Cursor writes under
 * `~/.cursor/projects/<project slug>/agent-transcripts/<id>/<id>.jsonl` (the
 * Cursor CLI, `cursor-agent`, and the Cursor editor's agent both write there).
 * This is how a Cursor session reaches `sessions`, as `tool = 'cursor'`.
 *
 * NO tokens and NO model: a transcript line is only `{role, message:{content}}`
 * (plus `{type:"turn_ended"}`) — no usage, no model id, no timestamps — so a
 * Cursor session gets no `token_usage` rows and no cost. Nothing is guessed.
 *
 * - the session id is the transcript's file name;
 * - the name is the first `<user_query>` line (60 chars);
 * - the start is the file's birth time (its mtime where the filesystem keeps
 *   none), the last activity its mtime: active within 2 minutes, else
 *   completed, ended at that write;
 * - the cwd is the project slug walked back onto the filesystem — the slug
 *   turns every `/` and every other non-alphanumeric character into `-`, so it
 *   is resolved only when exactly such a path still exists, else left empty.
 * A file is re-read when its mtime changes, and the mtime per file lives in
 * `cursor_transcripts`, so a session the user deleted stays deleted until
 * Cursor writes to it again. No transcript is ever stored as `transcript_path`.
 * @author Nguyễn Ngọc Trí Vĩ <vinnt@smartgift.vn>
 */

const fs = require("node:fs");
const os = require("node:os");
const path = require("node:path");
const { db, stmts } = require("../db");

const ACTIVE_MS = 120_000; // the window codex-usage.js calls a session active
const FUTURE_SLACK_MS = 1000;
const ID = /^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/;
const HEAD_BYTES = 256 * 1024;

/** `~/.cursor`, or `AGENTHUB_CURSOR_DIR`; read at call time. */
function getCursorDir() {
  return process.env.AGENTHUB_CURSOR_DIR || path.join(os.homedir(), ".cursor");
}

const norm = (s) => s.replace(/[^A-Za-z0-9]/g, "-");

// A walk enters at most this many directories: a member can build a tree (or
// a symlink loop) whose matching paths grow exponentially with the slug.
const MAX_VISITS = 500;

/**
 * The existing directory whose path slugs to `slug`, or null. Walks from `/`,
 * trying every real directory (never a symlink) whose own slug is a prefix of
 * what is left; gives up (null) after MAX_VISITS directories.
 */
function cwdFromSlug(slug, root = "/") {
  let visits = 0;
  const walk = (dir, rest, depth) => {
    if (!rest) return dir;
    if (depth > 40 || ++visits > MAX_VISITS) return null;
    let entries;
    try {
      entries = fs.readdirSync(dir, { withFileTypes: true });
    } catch {
      return null;
    }
    for (const e of entries) {
      if (!e.isDirectory()) continue;
      const n = norm(e.name);
      if (rest !== n && !rest.startsWith(`${n}-`)) continue;
      const hit = walk(path.join(dir, e.name), rest.slice(n.length + 1), depth + 1);
      if (hit || visits > MAX_VISITS) return hit;
    }
    return null;
  };
  const want = norm(slug);
  return want ? walk(root, want, 0) : null;
}

/** The first prompt's first line (60 chars) from the transcript's head, or null. */
function firstPrompt(file) {
  let text;
  try {
    const fd = require("./run-as").openSafe(file);
    try {
      const buf = Buffer.alloc(Math.min(fs.fstatSync(fd).size, HEAD_BYTES));
      fs.readSync(fd, buf, 0, buf.length, 0);
      text = buf.toString("utf8");
    } finally {
      fs.closeSync(fd);
    }
  } catch {
    return null;
  }
  for (const line of text.split("\n")) {
    let o;
    try {
      o = JSON.parse(line);
    } catch {
      continue;
    }
    if (o?.role !== "user" || !Array.isArray(o.message?.content)) continue;
    for (const part of o.message.content) {
      if (part?.type !== "text" || typeof part.text !== "string") continue;
      const q = /<user_query>\s*([\s\S]*?)\s*<\/user_query>/.exec(part.text);
      const prompt = (q ? q[1] : part.text.replace(/<timestamp>[\s\S]*?<\/timestamp>/, "")).trim();
      if (prompt) return prompt.split("\n")[0].slice(0, 60);
    }
  }
  return null;
}

/** Every agent transcript, with its project slug, birth and mtime — this
 *  account's, then every member's (S3: their own home). */
function transcriptFiles() {
  const members = require("./run-as")
    .memberHomes()
    .map((h) => path.join(h, ".cursor"));
  return [getCursorDir(), ...members].flatMap((dir) => transcriptFilesIn(dir));
}

function transcriptFilesIn(cursorDir) {
  const projects = path.join(cursorDir, "projects");
  let slugs;
  try {
    slugs = fs.readdirSync(projects);
  } catch {
    return [];
  }
  const out = [];
  for (const slug of slugs) {
    const dir = path.join(projects, slug, "agent-transcripts");
    let ids;
    try {
      ids = fs.readdirSync(dir);
    } catch {
      continue;
    }
    for (const id of ids) {
      if (!ID.test(id)) continue;
      const file = path.join(dir, id, `${id}.jsonl`);
      try {
        const st = fs.statSync(file);
        if (!st.isFile()) continue;
        const mtime = Math.floor(st.mtimeMs);
        out.push({ path: file, id, slug, mtime, birth: Math.floor(st.birthtimeMs) || mtime });
      } catch {
        /* no transcript (yet) */
      }
    }
  }
  return out;
}

const getSeen = db.prepare("SELECT * FROM cursor_transcripts WHERE path = ?");
const upsertSeen = db.prepare(
  `INSERT INTO cursor_transcripts (path, session_id, mtime_ms, updated_at)
   VALUES (?, ?, ?, strftime('%Y-%m-%dT%H:%M:%fZ', 'now'))
   ON CONFLICT(path) DO UPDATE SET session_id = excluded.session_id,
     mtime_ms = excluded.mtime_ms, updated_at = excluded.updated_at`
);
const insertSession = db.prepare(
  `INSERT INTO sessions (id, name, status, cwd, started_at, ended_at, updated_at, metadata, tool)
   VALUES (?, ?, ?, ?, ?, ?, ?, ?, 'cursor')`
);
const updateSession = db.prepare(
  `UPDATE sessions SET name = ?, status = ?, ended_at = ?, updated_at = ?, cwd = COALESCE(cwd, ?)
   WHERE id = ? AND tool = 'cursor'`
);
const settle = db.prepare(
  "UPDATE sessions SET status = 'completed', ended_at = ? WHERE id = ? AND tool = 'cursor'"
);

// A write "in the future" (a clock that runs ahead) is not proof of activity.
const fresh = (mtime, now) => mtime - now < FUTURE_SLACK_MS && now - mtime < ACTIVE_MS;

/** Write one transcript's session row. @returns the broadcast type, or null. */
function writeSession(f, now) {
  const existing = stmts.getSession.get(f.id);
  if (existing && existing.tool !== "cursor") return null;
  // S3: a member's transcript names only their own folder and session.
  const cwd = existing?.cwd ? null : cwdFromSlug(f.slug);
  if (!require("./run-as").sameOwner(f.path, existing?.cwd || cwd)) return null;
  const lastWrite = new Date(f.mtime).toISOString();
  const active = fresh(f.mtime, now);
  const status = active ? "active" : "completed";
  const endedAt = active ? null : lastWrite;
  const name = firstPrompt(f.path) || existing?.name || `Cursor ${f.id.slice(0, 8)}`;
  if (existing) {
    updateSession.run(
      name,
      status,
      endedAt,
      lastWrite,
      cwd,
      f.id
    );
    return "session_updated";
  }
  insertSession.run(
    f.id,
    name,
    status,
    cwd,
    new Date(Math.min(f.birth, f.mtime)).toISOString(),
    endedAt,
    lastWrite,
    JSON.stringify({ source: "cursor-transcript" })
  );
  return "session_created";
}

let running = null;

/**
 * One pass over every transcript: a changed one is re-read, an unchanged one
 * only has an `active` session settled once it went quiet.
 * @returns {Promise<{files: number, sessions: number}>}
 */
function ingestCursorUsage({ broadcast = () => {}, now = Date.now() } = {}) {
  if (running) return running;
  running = (async () => {
    let read = 0;
    let written = 0;
    for (const f of transcriptFiles()) {
      try {
        const seen = getSeen.get(f.path);
        if (seen && seen.mtime_ms === f.mtime) {
          if (fresh(f.mtime, now)) continue;
          const s = stmts.getSession.get(f.id);
          if (s && s.tool === "cursor" && s.status === "active") {
            settle.run(new Date(f.mtime).toISOString(), s.id);
            broadcast("session_updated", stmts.getSession.get(s.id));
            written += 1;
          }
          continue;
        }
        read += 1;
        const type = writeSession(f, now);
        upsertSeen.run(f.path, f.id, f.mtime);
        if (type) {
          written += 1;
          broadcast(type, stmts.getSession.get(f.id));
        }
      } catch (err) {
        console.warn(`[cursor-usage] skipped ${f.path}: ${err.message}`);
      }
      await new Promise((r) => setImmediate(r));
    }
    return { files: read, sessions: written };
  })().finally(() => {
    running = null;
  });
  return running;
}

/** First pass at once (in the background), then every DASHBOARD_CURSOR_POLL_MS (default 15 s; 0 disables). */
function startCursorUsagePoll(broadcast) {
  const POLL_MS = process.env.DASHBOARD_CURSOR_POLL_MS
    ? Number(process.env.DASHBOARD_CURSOR_POLL_MS)
    : 15_000;
  if (!Number.isFinite(POLL_MS) || POLL_MS <= 0) return;
  const tick = () => {
    ingestCursorUsage({ broadcast }).catch(() => {});
  };
  tick();
  const timer = setInterval(tick, POLL_MS);
  if (timer.unref) timer.unref();
}

module.exports = {
  getCursorDir,
  cwdFromSlug,
  firstPrompt,
  ingestCursorUsage,
  startCursorUsagePoll,
};