@gotcos/glasses-server 6.94.2 → 6.94.4

package/server/lib/session-find-collect.ts

@@ -0,0 +1,715 @@
+/**
+ * The sessions meaning search can ask about (Control 0.5.363, the "find" route): which sessions exist, what each is
+ * called, what its first typed prompt says, and when it last did something.
+ *
+ * WHY THIS IS NOT THE LIST OR THE SESSIONS-TAB COLLECTOR (validated 2026-11-01, all measured on the owner's Mac).
+ *   - The list keeps the first LINE of the first prompt, cut to 80 characters. 431 of 467 first prompts are longer, and
+ *     meaning search on titles alone found the right session 15 times in 20 against 19 with the first 150 characters.
+ *   - The list is capped at 7 days of file writes and 80 rows. mtime also lies: 234 of 854 Claude transcripts were touched
+ *     7 or more days after their last record (a relaunch appends bookkeeping). This collector orders by the transcript's own
+ *     last record, with its own memo, because the list's memo holds 480 entries and thrashes at this size.
+ *   - Machine sessions fill the newest 254 slots: raw, they reach 8 days; without them, about 33. Dropped here: no typed
+ *     first prompt, a title seen 5 or more times, a Claude Desktop scheduled task, a keep-warm title, an archived Desktop
+ *     session, and anything that ran on a key.
+ *   - Codex compresses finished rollouts (1,080 of 1,099 on that Mac are `.jsonl.zst`). `listCodexJsonlFiles` only takes
+ *     `.jsonl`, and every reader of it (detail, stream, completion, tails) is zstd-blind, so it is left alone. This module
+ *     has its own head-only lister and reader. A compressed row carries `compressed: true` and no file path ever leaves.
+ *
+ * ZSTD IS OPTIONAL. `createZstdDecompress` exists from Node 22.15 and 23.8; `engines` is `>=20.11`. A NAMED import of it
+ * fails when the module loads on an older Node and the server would not boot. The default import plus a `typeof` check at
+ * the call keeps boot safe; without the function a compressed file is skipped and COUNTED (`skippedNoDecoder`), never
+ * guessed at and never silent.
+ *
+ * Nothing here reads more than the first 16 MB of a file, calls a model, or logs a title or a prompt.
+ */
+import { createReadStream } from 'node:fs'
+import { execFile } from 'node:child_process'
+import { open } from 'node:fs/promises'
+import { promisify } from 'node:util'
+import { join } from 'node:path'
+import { StringDecoder } from 'node:string_decoder'
+import type { Transform } from 'node:stream'
+import zlib from 'node:zlib'
+import {
+  AGENT_SESSION_MAX_FILE_BYTES,
+  CLAUDE_UUID_JSONL,
+  agentSessionRoots,
+  codexUserText,
+  createdFromCodexFilename,
+  dirents,
+  fileStat,
+  firstLineTitle,
+  isKeepWarmSessionTitle,
+  isSkippedCursorFolder,
+  isWrapperPrompt,
+  isoFromMtime,
+  lastCustomTitle,
+  loadClaudeDesktopAliases,
+  loadCodexThreadNames,
+  parseJsonLine,
+  payloadText,
+  preferCursorCopy,
+  workspaceLabel,
+  type AgentProvider,
+  type AgentSessionRoots,
+} from './agent-session-store.js'
+import { ACTIVITY_READ_CONCURRENCY, ACTIVITY_TAIL_BYTES, ACTIVITY_TAIL_MAX_BYTES, activityFromTail, mapWithConcurrency } from './agent-session-activity.js'
+
+/**
+ * The most that is read from the top of a file (decompressed) to find its first typed prompt, stopping at the first one. A typed
+ * prompt is rarely deep: 256 KB found 135 of 172 Codex first prompts, 1 MB found 151, 16 MB all of them. Only a file that has none
+ * near the top reads further, so the cost lands on long machine runs, once per file (the memo keeps the answer).
+ */
+export const FIND_HEAD_BYTES = 16 * 1024 * 1024
+const FIND_CHUNK_BYTES = 64 * 1024
+/** A title seen this many times in the pool is a machine run (a routine, a proxy, a probe), not a conversation. */
+export const FIND_REPEAT_TITLE_LIMIT = 5
+/** The typed first prompt kept per session, unwrapped and on one line. Redaction reads all of it; the model gets 150. */
+export const FIND_PROMPT_KEEP_CHARS = 2_000
+/**
+ * A prompt is unwrapped from at most this much of its start, so an unclosed tag cannot make the scan quadratic (the wrapper
+ * pattern rescans to the end from every opener). Likewise the list's firstLineTitle, whose `<[^>]+>` is quadratic on a body of
+ * `a<b a<b ...` and blocked a server for 68 s on 800 KB: it only ever sees the first TITLE_SCAN_CHARS.
+ */
+const UNWRAP_SCAN_CHARS = 16_000
+const TITLE_SCAN_CHARS = 8_000
+/** A line longer than this is skipped through to its newline: it is an attachment or a dump, never a typed prompt. */
+const MAX_LINE_CHARS = 2 * 1024 * 1024
+/** Files opened at most per walk. The newest-first stop below ends nearly every walk long before this. */
+export const FIND_MAX_OPEN = 4_000
+/** Files read at once, and files per batch before the stop rule is checked. */
+const BATCH = 32
+/** A Desktop record read whole to find its scheduled-task and archived flags. The largest seen is 0.7 MB. */
+const DESKTOP_RECORD_MAX_BYTES = 8 * 1024 * 1024
+export const FIND_MEMO_MAX = 8_000
+/** What the memo may hold in all, by the characters it keeps (each entry is also flattened, see flat()). */
+export const FIND_MEMO_MAX_CHARS = 16 * 1024 * 1024
+
+const WRAP_TAGS = 'system-reminder|relevant-memories|cache-health|daily-bulletin|cos-alarms|device-handoff|now|local-command-stdout|local-command-stderr|command-name|command-message|command-args|task-notification|user-prompt-submit-hook|ide_opened_file|ide_selection|environment_context|attached_files|open_files|user_info|git_status|rules|agent_transcripts|image'
+const WRAP_RE = new RegExp(`<(${WRAP_TAGS})(?:\\s[^>]*)?>[\\s\\S]*?</\\1>`, 'gi')
+
+/**
+ * What the person typed, out of a user turn: the Cursor `<user_query>` body when there is one, the harness wrapper blocks
+ * out, whitespace on one line. The canary's rule (2026-11-01), which found the right session first in 19 of 20 searches.
+ */
+export function unwrapPrompt(text: string): string {
+  let body = text.slice(0, UNWRAP_SCAN_CHARS)
+  const open = body.indexOf('<user_query>'), close = body.indexOf('</user_query>')
+  if (open >= 0 && close > open) body = body.slice(open + '<user_query>'.length, close)
+  return body.replace(WRAP_RE, ' ').replace(/\s+/g, ' ').trim()
+}
+
+export interface FindCandidate {
+  provider: AgentProvider
+  session_id: string
+  /** `provider:id` in lower case: one identity across the pool, the cache and the merge. */
+  key: string
+  /** As the list titles it (Desktop alias, then custom title, then the first line of the first prompt). Never sent as is. */
+  title: string
+  /** The list's first line of the first prompt, at most 80 characters. */
+  listFirstLine: string
+  /** The first typed prompt, unwrapped, on one line, at most FIND_PROMPT_KEEP_CHARS. Never sent as is. */
+  typed: string
+  project: string
+  /** The newest record's time (or the file's write time where records carry none), never later than the write time. */
+  clockMs: number
+  mtimeMs: number
+  birthtimeMs: number
+  created: string
+  /** A `.jsonl.zst` Codex rollout: no tail can be read, so it cannot be opened from COS, only from the Codex app. */
+  compressed: boolean
+  /** Where the transcript is. Server side only; no row carries it. */
+  file: string
+}
+
+export interface FindPool {
+  candidates: FindCandidate[]
+  /** Compressed files left out because this Node cannot decompress zstd. */
+  skippedNoDecoder: number
+  /** Compressed files left out because they could not be read (a damaged or half-written file). */
+  skippedUnreadable: number
+  /** Files whose heads were read this walk (memo hits are free and not counted). */
+  opened: number
+  considered: number
+  ms: number
+}
+
+type ZstdFactory = () => Transform | null
+
+/** Resolved at the call, so a Node without it (and a test that removes it) both read as "absent". */
+export function zstdDecompressor(): Transform | null {
+  const make = (zlib as { createZstdDecompress?: () => Transform }).createZstdDecompress
+  return typeof make === 'function' ? make.call(zlib) : null
+}
+
+/** `rollout-<stamp>-<uuid>.jsonl` or `.jsonl.zst`: the uuid, or null. Not `idFromCodexFilename`, which stays `.jsonl` only. */
+export function rolloutFileId(name: string): string | null {
+  const base = rolloutBaseName(name)
+  if (!base) return null
+  const parts = base.split('-')
+  if (parts.length < 5) return null
+  const uuid = parts.slice(-5).join('-').toLowerCase()
+  return uuid.length >= 36 ? uuid : null
+}
+function rolloutBaseName(name: string): string | null {
+  if (name.endsWith('.jsonl.zst')) return name.slice(0, -'.jsonl.zst'.length)
+  if (name.endsWith('.jsonl')) return name.slice(0, -'.jsonl'.length)
+  return null
+}
+
+/** Every Codex rollout under one sessions root, plain and compressed, in directory order. Head-only: nothing is opened. */
+export async function listCodexRolloutFiles(sessionsRoot: string): Promise<Array<{ file: string; compressed: boolean }>> {
+  const files: Array<{ file: string; compressed: boolean }> = []
+  for (const year of await dirents(sessionsRoot)) {
+    const yearDir = join(sessionsRoot, year)
+    const yearSt = await fileStat(yearDir)
+    if (!yearSt || yearSt.isFile) continue
+    for (const month of await dirents(yearDir)) {
+      const monthDir = join(yearDir, month)
+      const monthSt = await fileStat(monthDir)
+      if (!monthSt || monthSt.isFile) continue
+      for (const day of await dirents(monthDir)) {
+        const dayDir = join(monthDir, day)
+        const daySt = await fileStat(dayDir)
+        if (!daySt || daySt.isFile) continue
+        for (const name of await dirents(dayDir)) {
+          if (name.endsWith('.jsonl.zst')) files.push({ file: join(dayDir, name), compressed: true })
+          else if (name.endsWith('.jsonl')) files.push({ file: join(dayDir, name), compressed: false })
+        }
+      }
+    }
+  }
+  return files
+}
+
+class NoDecoderError extends Error {}
+
+/**
+ * The first `maxBytes` of a file's (decompressed) content, one line at a time, until `onLine` says stop. A line cut by the
+ * limit is handed over as it stands and will usually not parse, which is the reader's cue to skip it.
+ */
+async function readHeadLines(file: string, compressed: boolean, maxBytes: number, zstd: ZstdFactory, onLine: (line: string) => boolean): Promise<void> {
+  const decoder = new StringDecoder('utf8')
+  let carry = ''
+  let scanFrom = 0
+  let skipping = false
+  let total = 0
+  let stopped = false
+  const feed = (chunk: Buffer): boolean => {
+    total += chunk.length
+    carry += decoder.write(chunk)
+    for (;;) {
+      // Resume the newline search where the last one stopped: one enormous line is not rescanned from 0 on every chunk.
+      const at = carry.indexOf('\n', scanFrom)
+      if (at < 0) {
+        scanFrom = carry.length
+        if (carry.length > MAX_LINE_CHARS) { skipping = true; carry = ''; scanFrom = 0 }
+        return false
+      }
+      const line = carry.slice(0, at)
+      carry = carry.slice(at + 1)
+      scanFrom = 0
+      if (skipping) { skipping = false; continue }
+      if (line && line.length <= MAX_LINE_CHARS && onLine(line)) return true
+    }
+  }
+  if (!compressed) {
+    const handle = await open(file, 'r')
+    try {
+      const buffer = Buffer.alloc(FIND_CHUNK_BYTES)
+      while (total < maxBytes && !stopped) {
+        const { bytesRead } = await handle.read(buffer, 0, Math.min(FIND_CHUNK_BYTES, maxBytes - total), total)
+        if (bytesRead <= 0) break
+        stopped = feed(Buffer.from(buffer.subarray(0, bytesRead)))
+      }
+    } finally { await handle.close() }
+  } else {
+    const inflate = zstd()
+    if (!inflate) throw new NoDecoderError('zstd unavailable')
+    const raw = createReadStream(file, { highWaterMark: FIND_CHUNK_BYTES })
+    raw.on('error', error => inflate.destroy(error))
+    raw.pipe(inflate)
+    try {
+      for await (const chunk of inflate) {
+        stopped = feed(chunk as Buffer)
+        if (stopped || total >= maxBytes) break
+      }
+    } finally { raw.destroy(); inflate.destroy() }
+  }
+  if (!stopped && !skipping && carry.trim()) onLine(carry + decoder.end())
+}
+
+interface HeadInfo {
+  /** A typed first prompt was found (or this is a subagent, which is left out whatever it says). */
+  typed: string
+  listFirstLine: string
+  subagent: boolean
+  metaId: string
+  cwd: string
+  created: string
+}
+
+/** A fresh flat copy. A slice of a big string keeps the whole parent alive in V8, and the memo outlives the parse (418 MB for 400 files). */
+const flat = (text: string): string => Buffer.from(text, 'utf8').toString('utf8')
+
+const emptyHead = (): HeadInfo => ({ typed: '', listFirstLine: '', subagent: false, metaId: '', cwd: '', created: '' })
+
+function parseClaudeHead(): { head: HeadInfo; onLine: (line: string) => boolean } {
+  const head = emptyHead()
+  return { head, onLine: line => {
+    if (!line.includes('"user"')) return false
+    const obj = parseJsonLine(line)
+    if (!obj || obj.type !== 'user' || obj.toolUseResult || obj.isSidechain === true) return false
+    const message = obj.message && typeof obj.message === 'object' ? obj.message as Record<string, unknown> : null
+    const body = message ? payloadText(message) : null
+    if (!body) return false
+    // Not from a wrapper body: the list titles a session from whatever the first record says, which can be harness text.
+    if (isWrapperPrompt(body)) return false
+    const typed = unwrapPrompt(body)
+    if (typed.length < 3) return false
+    head.listFirstLine = flat(firstLineTitle(body.slice(0, TITLE_SCAN_CHARS)))
+    head.typed = flat(typed.slice(0, FIND_PROMPT_KEEP_CHARS))
+    return true
+  } }
+}
+
+function parseCodexHead(): { head: HeadInfo; onLine: (line: string) => boolean } {
+  const head = emptyHead()
+  return { head, onLine: line => {
+    if (!line.includes('"session_meta"') && !line.includes('"response_item"')) return false
+    const obj = parseJsonLine(line)
+    if (!obj || !obj.payload || typeof obj.payload !== 'object') return false
+    const payload = obj.payload as Record<string, unknown>
+    if (obj.type === 'session_meta') {
+      head.metaId = flat(String(payload.id ?? payload.session_id ?? head.metaId))
+      head.cwd = flat(String(payload.cwd ?? head.cwd))
+      head.subagent = payload.thread_source === 'subagent'
+      head.created = flat(String(payload.timestamp ?? obj.timestamp ?? head.created).trim())
+      return head.subagent
+    }
+    if (obj.type !== 'response_item' || payload.type !== 'message' || payload.role !== 'user') return false
+    const body = codexUserText(payload)
+    if (!body) return false
+    if (isWrapperPrompt(body)) return false
+    const typed = unwrapPrompt(body)
+    if (typed.length < 3) return false
+    head.listFirstLine = flat(firstLineTitle(body.slice(0, TITLE_SCAN_CHARS)))
+    head.typed = flat(typed.slice(0, FIND_PROMPT_KEEP_CHARS))
+    return true
+  } }
+}
+
+function parseCursorHead(): { head: HeadInfo; onLine: (line: string) => boolean } {
+  const head = emptyHead()
+  return { head, onLine: line => {
+    if (!line.includes('"user"')) return false
+    const obj = parseJsonLine(line)
+    if (!obj || obj.role !== 'user') return false
+    const message = obj.message && typeof obj.message === 'object' ? obj.message as Record<string, unknown> : null
+    const body = message ? payloadText(message) : null
+    if (!body) return false
+    if (isWrapperPrompt(body) && !body.includes('<user_query>')) return false
+    const typed = unwrapPrompt(body)
+    if (typed.length < 3) return false
+    head.listFirstLine = flat(firstLineTitle(body.slice(0, TITLE_SCAN_CHARS)))
+    head.typed = flat(typed.slice(0, FIND_PROMPT_KEEP_CHARS))
+    return true
+  } }
+}
+
+interface FileRef {
+  provider: AgentProvider
+  file: string
+  /** Claude: the uuid. Codex: the file's uuid. Cursor: the session directory. */
+  id: string
+  /** Claude: the project folder. Cursor: the project folder. Codex: unused. */
+  folder: string
+  mtimeMs: number
+  birthtimeMs: number
+  size: number
+  compressed: boolean
+}
+
+interface FileInfo {
+  mtimeMs: number
+  size: number
+  head: HeadInfo
+  /** Claude only, read the first time it is needed (a Desktop alias title wins over it). `null` is "none found". */
+  customTitle?: string | null
+  clockMs?: number
+}
+
+interface DesktopFlags { mtimeMs: number; size: number; scheduled: boolean; archived: boolean }
+
+/** Both flags are keys of the Desktop record: archived near its top, scheduledTaskId anywhere in its first megabytes. */
+async function readDesktopFlags(file: string, size: number): Promise<{ scheduled: boolean; archived: boolean } | null> {
+  if (size > DESKTOP_RECORD_MAX_BYTES) return null
+  try {
+    const handle = await open(file, 'r')
+    try {
+      const buffer = Buffer.alloc(size)
+      const { bytesRead } = await handle.read(buffer, 0, size, 0)
+      const text = buffer.subarray(0, bytesRead).toString('utf8')
+      return {
+        scheduled: /"scheduledTaskId"\s*:\s*"[^"]+"/.test(text),
+        archived: /"isArchived"\s*:\s*true/.test(text.slice(0, 64 * 1024)),
+      }
+    } finally { await handle.close() }
+  } catch { return null }
+}
+
+export interface FindPoolOptions {
+  roots?: AgentSessionRoots
+  /** How many sessions the pool should hold at most: 254 for the first request of a search, 508 for both. */
+  limit: number
+  signal?: AbortSignal
+  /** Lower-case ids (Claude session ids, Codex thread ids) that ran on a key. They never leave the Mac. */
+  keyedIds?: ReadonlySet<string>
+  /** Test seam: where the zstd decoder comes from. Defaults to node:zlib at the call. */
+  zstd?: ZstdFactory
+  maxOpen?: number
+  /** Test seam: how far into a file to look for the first typed prompt. Defaults to FIND_HEAD_BYTES. */
+  headBytes?: number
+}
+
+const normTitle = (title: string) => title.toLowerCase().replace(/\s+/g, ' ').trim()
+
+const execFileAsync = promisify(execFile)
+
+/**
+ * Cursor's sidebar names and which of its sessions are archived, read-only, from composerHeaders. loadCursorComposerNames
+ * (the list's reader) keeps names only; an archived session still has one, and 21 of 828 headers here are archived.
+ */
+async function loadComposerHeaders(db: string): Promise<{ names: Map<string, string>; archived: Set<string> }> {
+  const names = new Map<string, string>(), archived = new Set<string>()
+  try {
+    const { stdout } = await execFileAsync('/usr/bin/sqlite3', ['-readonly', '-json', db,
+      "SELECT composerId AS id, json_extract(value, '$.name') AS name, json_extract(value, '$.isArchived') AS archived FROM composerHeaders"],
+    { timeout: 4000, maxBuffer: 16 * 1024 * 1024 })
+    const rows = JSON.parse(stdout || '[]') as Array<{ id?: string; name?: string | null; archived?: number | boolean | null }>
+    for (const row of rows) {
+      const id = String(row.id ?? '').trim()
+      if (!id) continue
+      const name = String(row.name ?? '').trim()
+      if (name) names.set(id, flat(name.slice(0, 120)))
+      if (row.archived === 1 || row.archived === true) archived.add(id)
+    }
+  } catch { /* unreadable: no names and nothing marked archived, as the list's reader does */ }
+  return { names, archived }
+}
+
+/** The newest conversation record's time from the transcript's own tail, widening x4 only while the tail has none. Not the list's memo. */
+async function lastRecordMs(provider: AgentProvider, file: string, size: number): Promise<number | null> {
+  try {
+    const handle = await open(file, 'r')
+    try {
+      for (let window = ACTIVITY_TAIL_BYTES; ; window *= 4) {
+        const length = Math.min(size, window)
+        const start = size - length
+        const buffer = Buffer.alloc(length)
+        const { bytesRead } = await handle.read(buffer, 0, length, start)
+        const at = activityFromTail(provider, buffer.subarray(0, bytesRead).toString('utf8'), start > 0).lastActivityAt
+        if (at) { const ms = Date.parse(at); return Number.isFinite(ms) ? ms : null }
+        if (start === 0 || window >= ACTIVITY_TAIL_MAX_BYTES) return null
+      }
+    } finally { await handle.close() }
+  } catch { return null }
+}
+
+export class SessionFindCollector {
+  private memo = new Map<string, FileInfo>()
+  private flags = new Map<string, DesktopFlags>()
+  private cursorNames: { stamp: string; names: Map<string, string>; archived: Set<string> } | null = null
+  private flights = new Map<string, { promise: Promise<FindPool>; waiters: number; controller: AbortController }>()
+
+  constructor(private readonly memoMaxChars: number = FIND_MEMO_MAX_CHARS) {}
+
+  /** Memo entries held. A test seam: the memo must never outgrow the files it mirrors. */
+  memoSize(): number { return this.memo.size }
+
+  /**
+   * The pool. Concurrent calls for the same roots and limit share one walk; the walk stops when its LAST caller has gone
+   * (a request that closed), and what it had read stays in the memo, so the next try starts further along.
+   */
+  collect(options: FindPoolOptions): Promise<FindPool> {
+    const roots = options.roots ?? agentSessionRoots()
+    const key = `${options.limit}\0${JSON.stringify(roots)}\0${options.keyedIds ? [...options.keyedIds].sort().join(',') : ''}`
+    let flight = this.flights.get(key)
+    // A flight whose last caller left is stopping; a new caller starts a fresh walk (the memo keeps what the old one read).
+    if (!flight || flight.controller.signal.aborted) {
+      const controller = new AbortController()
+      const created = { promise: undefined as unknown as Promise<FindPool>, waiters: 0, controller }
+      created.promise = this.walk({ ...options, roots }, controller.signal).finally(() => { if (this.flights.get(key) === created) this.flights.delete(key) })
+      // A walk nobody waits for any more rejects; that is not an unhandled error.
+      created.promise.catch(() => undefined)
+      this.flights.set(key, created)
+      flight = created
+    }
+    const shared = flight
+    shared.waiters += 1
+    return new Promise<FindPool>((resolve, reject) => {
+      let left = false
+      const leave = () => {
+        if (left) return
+        left = true
+        options.signal?.removeEventListener('abort', onAbort)
+        shared.waiters -= 1
+        if (shared.waiters <= 0) shared.controller.abort()
+      }
+      const onAbort = () => { leave(); reject(abortError()) }
+      if (options.signal?.aborted) { onAbort(); return }
+      options.signal?.addEventListener('abort', onAbort, { once: true })
+      shared.promise.then(pool => { leave(); resolve(pool) }, error => { leave(); reject(error) })
+    })
+  }
+
+  private async walk(options: FindPoolOptions & { roots: AgentSessionRoots }, signal: AbortSignal): Promise<FindPool> {
+    const started = performance.now()
+    const { roots } = options
+    const zstd = options.zstd ?? zstdDecompressor
+    const keyed = options.keyedIds ?? new Set<string>()
+    const maxOpen = options.maxOpen ?? FIND_MAX_OPEN
+    const throwIfStopped = () => { if (signal.aborted) throw abortError() }
+
+    const aliases = await loadClaudeDesktopAliases(roots.claudeCodeSessions)
+    throwIfStopped()
+    const desktopByCli = new Map<string, { title: string; file: string; mtimeMs: number }>()
+    for (const [id, head] of aliases) {
+      if (!head) continue
+      const cli = head.cliSessionId && CLAUDE_UUID_JSONL.test(head.cliSessionId + '.jsonl') ? head.cliSessionId : id
+      const previous = desktopByCli.get(cli)
+      if (!previous || head.mtimeMs >= previous.mtimeMs) desktopByCli.set(cli, { title: head.title, file: head.file, mtimeMs: head.mtimeMs })
+    }
+    const codexNames = await loadCodexThreadNames(roots.codexSessions)
+    const composers = await this.composerNames(roots.cursorComposerDb)
+    throwIfStopped()
+
+    const refs = await this.enumerate(roots)
+    throwIfStopped()
+    this.prune(refs)
+    refs.sort((a, b) => b.mtimeMs - a.mtimeMs)
+
+    const assembled = new Map<string, FindCandidate>()
+    let skippedNoDecoder = 0, skippedUnreadable = 0, opened = 0
+    const noDecoderFiles = new Set<string>(), unreadableFiles = new Set<string>()
+    let pool: FindCandidate[] = []
+    let next = 0
+    while (next < refs.length) {
+      throwIfStopped()
+      // Everything not yet opened was written no later than refs[next], and a transcript's last record is never later than its
+      // file's write time. So once the pool is full and its oldest entry is newer than that, no unopened file can displace it.
+      if (pool.length >= options.limit && pool[options.limit - 1]!.clockMs >= refs[next]!.mtimeMs) break
+      if (opened >= maxOpen) break
+      const batch = refs.slice(next, next + BATCH)
+      next += batch.length
+      const results = await mapWithConcurrency(batch, ACTIVITY_READ_CONCURRENCY, async ref => {
+        if (signal.aborted) return null
+        return this.inspect(ref, { desktopByCli, codexNames, cursorNames: composers.names, cursorArchived: composers.archived, keyed, zstd, headBytes: options.headBytes ?? FIND_HEAD_BYTES })
+      })
+      for (let i = 0; i < batch.length; i++) {
+        const outcome = results[i]
+        const ref = batch[i]!
+        if (!outcome) continue
+        if (outcome.read) opened += 1
+        if (outcome.skip === 'nodecoder') noDecoderFiles.add(ref.file)
+        if (outcome.skip === 'unreadable') unreadableFiles.add(ref.file)
+        if (outcome.candidate) assembled.set(ref.file, outcome.candidate)
+      }
+      pool = this.settle([...assembled.values()])
+    }
+    throwIfStopped()
+    this.trim()
+    skippedNoDecoder = noDecoderFiles.size
+    skippedUnreadable = unreadableFiles.size
+    return { candidates: pool.slice(0, options.limit), skippedNoDecoder, skippedUnreadable, opened, considered: refs.length, ms: Math.round(performance.now() - started) }
+  }
+
+  /** The pool-wide rules (a title seen 5 or more times is a machine), one identity per session, newest record first. */
+  private settle(all: FindCandidate[]): FindCandidate[] {
+    const seen = new Map<string, number>()
+    for (const candidate of all) {
+      const t = normTitle(candidate.title)
+      seen.set(t, (seen.get(t) ?? 0) + 1)
+    }
+    const byKey = new Map<string, FindCandidate>()
+    for (const candidate of all) {
+      if ((seen.get(normTitle(candidate.title)) ?? 0) >= FIND_REPEAT_TITLE_LIMIT) continue
+      const existing = byKey.get(candidate.key)
+      if (!existing || candidate.clockMs > existing.clockMs) byKey.set(candidate.key, candidate)
+    }
+    return [...byKey.values()].sort((a, b) => b.clockMs - a.clockMs || a.key.localeCompare(b.key))
+  }
+
+  private async composerNames(db: string): Promise<{ names: Map<string, string>; archived: Set<string> }> {
+    const st = db ? await fileStat(db) : null
+    if (!st?.isFile) return { names: new Map(), archived: new Set() }
+    const stamp = `${db}:${st.mtimeMs}:${st.size}`
+    if (this.cursorNames?.stamp === stamp) return this.cursorNames
+    const loaded = { stamp, ...(await loadComposerHeaders(db)) }
+    this.cursorNames = loaded
+    return loaded
+  }
+
+  /** Stat every transcript. Claude and Cursor from their own stores; Codex from the SERVER home only (never a keyed home). */
+  private async enumerate(roots: AgentSessionRoots): Promise<FileRef[]> {
+    const refs: FileRef[] = []
+    for (const folder of await dirents(roots.claudeProjects)) {
+      const dir = join(roots.claudeProjects, folder)
+      for (const name of await dirents(dir)) {
+        if (!CLAUDE_UUID_JSONL.test(name)) continue
+        const file = join(dir, name)
+        const st = await fileStat(file)
+        if (!st?.isFile) continue
+        refs.push({ provider: 'claude', file, id: name.slice(0, -'.jsonl'.length), folder, mtimeMs: st.mtimeMs, birthtimeMs: st.birthtimeMs, size: st.size, compressed: false })
+      }
+    }
+    for (const { file, compressed } of await listCodexRolloutFiles(roots.codexSessions)) {
+      const st = await fileStat(file)
+      if (!st?.isFile) continue
+      refs.push({ provider: 'codex', file, id: file.split('/').pop() || file, folder: '', mtimeMs: st.mtimeMs, birthtimeMs: st.birthtimeMs, size: st.size, compressed })
+    }
+    const cursor = new Map<string, FileRef>()
+    for (const folder of await dirents(roots.cursorProjects)) {
+      if (isSkippedCursorFolder(folder)) continue
+      const transcripts = join(roots.cursorProjects, folder, 'agent-transcripts')
+      for (const sessionDir of await dirents(transcripts)) {
+        if (sessionDir === 'subagents') continue
+        const file = join(transcripts, sessionDir, `${sessionDir}.jsonl`)
+        const st = await fileStat(file)
+        if (!st?.isFile || st.size > AGENT_SESSION_MAX_FILE_BYTES) continue
+        const next: FileRef = { provider: 'cursor', file, id: sessionDir, folder, mtimeMs: st.mtimeMs, birthtimeMs: st.birthtimeMs, size: st.size, compressed: false }
+        const existing = cursor.get(sessionDir)
+        if (!existing) { cursor.set(sessionDir, next); continue }
+        const pick = preferCursorCopy(
+          { file: next.file, sessionDir, project: workspaceLabel(next.folder), mtimeMs: next.mtimeMs, birthtimeMs: next.birthtimeMs },
+          { file: existing.file, sessionDir, project: workspaceLabel(existing.folder), mtimeMs: existing.mtimeMs, birthtimeMs: existing.birthtimeMs })
+        cursor.set(sessionDir, pick.file === next.file ? next : existing)
+      }
+    }
+    refs.push(...cursor.values())
+    return refs
+  }
+
+  /** Forget memo entries for files that are gone, and keep the memo inside its bound. */
+  private prune(refs: readonly FileRef[]): void {
+    const live = new Set(refs.map(ref => ref.file))
+    for (const file of this.memo.keys()) if (!live.has(file)) this.memo.delete(file)
+    this.trim()
+  }
+
+  /** Evict least recently used entries until the memo is inside both bounds: entries and characters held. */
+  private trim(): void {
+    const chars = (info: FileInfo) => info.head.typed.length + info.head.listFirstLine.length + info.head.metaId.length + info.head.cwd.length + (info.customTitle?.length ?? 0) + 64
+    let held = 0
+    for (const info of this.memo.values()) held += chars(info)
+    while (this.memo.size > FIND_MEMO_MAX || held > this.memoMaxChars) {
+      const oldestKey = this.memo.keys().next().value
+      if (oldestKey === undefined) break
+      held -= chars(this.memo.get(oldestKey)!)
+      this.memo.delete(oldestKey)
+    }
+  }
+
+  private async inspect(ref: FileRef, ctx: {
+    desktopByCli: Map<string, { title: string; file: string; mtimeMs: number }>
+    codexNames: Map<string, string>
+    cursorNames: Map<string, string>
+    cursorArchived: ReadonlySet<string>
+    keyed: ReadonlySet<string>
+    zstd: ZstdFactory
+    headBytes: number
+  }): Promise<{ candidate: FindCandidate | null; read: boolean; skip?: 'nodecoder' | 'unreadable' } | null> {
+    let info = this.memo.get(ref.file)
+    let read = false
+    if (!info || info.mtimeMs !== ref.mtimeMs || info.size !== ref.size) {
+      const parser = ref.provider === 'claude' ? parseClaudeHead() : ref.provider === 'codex' ? parseCodexHead() : parseCursorHead()
+      try {
+        await readHeadLines(ref.file, ref.compressed, ctx.headBytes, ctx.zstd, parser.onLine)
+      } catch (error) {
+        return { candidate: null, read: false, skip: error instanceof NoDecoderError ? 'nodecoder' : 'unreadable' }
+      }
+      read = true
+      info = { mtimeMs: ref.mtimeMs, size: ref.size, head: parser.head }
+      this.memo.delete(ref.file)
+      this.memo.set(ref.file, info)
+    } else {
+      this.memo.delete(ref.file)
+      this.memo.set(ref.file, info)
+    }
+    const { head } = info
+    if (head.subagent || !head.typed) return { candidate: null, read }
+
+    let sessionId: string
+    let title: string
+    let project: string
+    let created = isoFromMtime(ref.birthtimeMs)
+    if (ref.provider === 'claude') {
+      sessionId = ref.id
+      if (ctx.keyed.has(sessionId.toLowerCase())) return { candidate: null, read }
+      const desktop = ctx.desktopByCli.get(sessionId)
+      if (desktop) {
+        // The Desktop record says whether it was a scheduled run or put away. Unreadable counts as yes: it fails toward not sending.
+        const flags = await this.desktopFlags(desktop.file)
+        if (!flags || flags.scheduled || flags.archived) return { candidate: null, read }
+      }
+      if (!desktop?.title && info.customTitle === undefined) info.customTitle = await lastCustomTitle(ref.file).then(t => t === null ? null : flat(t), () => null)
+      title = desktop?.title || info.customTitle || head.listFirstLine || 'Claude session'
+      project = workspaceLabel(ref.folder)
+    } else if (ref.provider === 'codex') {
+      const name = ref.file.split('/').pop() || ref.file
+      const fileId = rolloutFileId(name)
+      const metaId = head.metaId.trim().toLowerCase()
+      // The filename wins when it disagrees: a cloned thread carries its parent's id in session_meta (listCodexSessions).
+      sessionId = fileId && metaId && fileId !== metaId ? fileId : (head.metaId || rolloutBaseName(name) || name)
+      if (ctx.keyed.has(sessionId.toLowerCase())) return { candidate: null, read }
+      title = ctx.codexNames.get(sessionId) || head.listFirstLine || 'Codex session'
+      project = workspaceLabel(head.cwd)
+      created = head.created || createdFromCodexFilename(name) || created
+    } else {
+      sessionId = ref.id
+      if (ctx.keyed.has(sessionId.toLowerCase()) || ctx.cursorArchived.has(sessionId)) return { candidate: null, read }
+      title = ctx.cursorNames.get(sessionId) || head.listFirstLine || 'Cursor session'
+      project = workspaceLabel(ref.folder)
+    }
+    if (isKeepWarmSessionTitle(title) || isKeepWarmSessionTitle(head.listFirstLine) || isKeepWarmSessionTitle(head.typed)) return { candidate: null, read }
+
+    if (info.clockMs === undefined) {
+      let clock = ref.mtimeMs
+      // A compressed rollout cannot be tailed; its write time was preserved from the last write (median 0.21 days after the start, which is the stamp's own offset).
+      if (ref.provider !== 'cursor' && !ref.compressed) {
+        const at = await lastRecordMs(ref.provider, ref.file, ref.size)
+        if (at !== null) clock = Math.min(at, ref.mtimeMs)
+      }
+      info.clockMs = clock
+    }
+    return { read, candidate: {
+      provider: ref.provider, session_id: sessionId, key: `${ref.provider}:${sessionId.toLowerCase()}`, title: title.slice(0, 120),
+      listFirstLine: head.listFirstLine, typed: head.typed, project, clockMs: info.clockMs, mtimeMs: ref.mtimeMs, birthtimeMs: ref.birthtimeMs,
+      created, compressed: ref.compressed, file: ref.file,
+    } }
+  }
+
+  private async desktopFlags(file: string): Promise<{ scheduled: boolean; archived: boolean } | null> {
+    const st = await fileStat(file)
+    if (!st?.isFile) return null
+    const known = this.flags.get(file)
+    if (known && known.mtimeMs === st.mtimeMs && known.size === st.size) return known
+    const read = await readDesktopFlags(file, st.size)
+    if (!read) { this.flags.delete(file); return null }
+    const flags: DesktopFlags = { mtimeMs: st.mtimeMs, size: st.size, ...read }
+    this.flags.set(file, flags)
+    return flags
+  }
+}
+
+function abortError(): Error {
+  const error = new Error('cancelled')
+  error.name = 'AbortError'
+  return error
+}
+
+export function isAbortError(error: unknown): boolean { return error instanceof Error && error.name === 'AbortError' }
+
+let shared: SessionFindCollector | null = null
+/** One collector per process: it owns the memo that makes the second search cheap. */
+export function sharedSessionFindCollector(): SessionFindCollector { return shared ??= new SessionFindCollector() }