dsh-plugin-session-cleaner 0.0.0-stage → 0.2.0

package/src/cleaner.js

@@ -0,0 +1,676 @@
+/**
+ * session-cleaner core: pure, testable logic for completely removing DSH sessions.
+ *
+ * Design commitments:
+ *  - Never guess a path. Every deletion target is validated THREE times:
+ *      1. the session id matches the shipped id grammar,
+ *      2. the on-disk directory/file name equals that id exactly,
+ *      3. the normalized absolute path is contained in an allowed root.
+ *    A target failing any check is skipped and reported, never deleted.
+ *  - Removal order matters: stop the session BEFORE touching its files, and
+ *    detach registry references BEFORE deleting files, so the UI never shows a
+ *    row whose log has already vanished.
+ */
+
+import { existsSync } from 'node:fs'
+import { open, readdir, readFile, rename, rm, stat, writeFile } from 'node:fs/promises'
+import { homedir } from 'node:os'
+import path from 'node:path'
+import { zstdDecompressSync } from 'node:zlib'
+
+/**
+ * Shipped session directory grammar. Three forms coexist on disk:
+ *   - `session-<uuid>` — user sessions writing `session.v4.jsonl.zstd`,
+ *   - `sess_<uuid>`    — legacy user sessions writing `session.v3.jsonl.zstd`,
+ *   - `<uuid>`         — subagent / delegated child sessions.
+ * All three are valid deletion targets; anything else is not a session.
+ */
+const SESSION_ID = /^(?:(?:session-|sess_))?[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}$/
+
+/** A session log directly inside a session directory. */
+const LOG_NAME = /^session\.v\d+\.jsonl(?:\.zstd)?$/
+
+export function resolveDshHome(options = {}) {
+  const candidates = [
+    options.dshHome,
+    process.env.DSH_HOME,
+    path.join(homedir(), '.dsh'),
+  ]
+  for (const candidate of candidates) {
+    if (typeof candidate === 'string' && candidate.length > 0) return path.resolve(candidate)
+  }
+  throw new Error('session-cleaner: cannot resolve the DSH home directory')
+}
+
+export function isValidSessionId(id) {
+  return typeof id === 'string' && SESSION_ID.test(id)
+}
+
+export function normalizeForCompare(target) {
+  const resolved = path.resolve(target)
+  return process.platform === 'win32' ? resolved.toLowerCase() : resolved
+}
+
+/** True when `target` is `root` itself or lives underneath it. */
+export function isContained(root, target) {
+  const base = normalizeForCompare(root)
+  const candidate = normalizeForCompare(target)
+  if (candidate === base) return true
+  const withSep = base.endsWith(path.sep) ? base : base + path.sep
+  return candidate.startsWith(withSep)
+}
+
+/** A capture with a bounded argument vector — never a shell string. */
+
+/**
+ * Read the first line of a session log, transparently handling zstd compression.
+ *
+ * Decompression uses Node's built-in zstd support (Node >= 23.8) rather than a
+ * `zstd` executable: header linkage backs descendant discovery, and a missing
+ * external tool would silently degrade deletion into a no-op cascade.
+ *
+ * Returns `undefined` when the header cannot be read; linkage is an
+ * optimization, so an unreadable header must never fail a scan.
+ */
+export async function readSessionHeader(logFile) {
+  try {
+    let text
+    if (logFile.endsWith('.zstd')) {
+      const handle = await open(logFile, 'r')
+      let compressed
+      try {
+        const info = await handle.stat()
+        compressed = Buffer.alloc(Number(info.size))
+        let offset = 0
+        while (offset < compressed.length) {
+          const { bytesRead } = await handle.read(compressed, offset, compressed.length - offset, offset)
+          if (bytesRead <= 0) break
+          offset += bytesRead
+        }
+        compressed = compressed.subarray(0, offset)
+      } finally {
+        await handle.close()
+      }
+      text = zstdDecompressSync(compressed).toString('utf8')
+    } else {
+      const handle = await open(logFile, 'r')
+      try {
+        const buffer = Buffer.alloc(64 * 1024)
+        const { bytesRead } = await handle.read(buffer, 0, buffer.length, 0)
+        text = buffer.subarray(0, bytesRead).toString('utf8')
+      } finally {
+        await handle.close()
+      }
+    }
+    const firstLine = text.split('\n', 1)[0]?.trim()
+    if (!firstLine) return undefined
+    const parsed = JSON.parse(firstLine)
+    const meta = parsed?.meta ?? parsed?.header ?? parsed
+    if (!meta || typeof meta !== 'object') return undefined
+    return {
+      id: typeof meta.id === 'string' ? meta.id : undefined,
+      parent: typeof meta.parentSession === 'string' ? meta.parentSession : undefined,
+      cwd: typeof meta.cwd === 'string' ? meta.cwd : undefined,
+      title: typeof meta.title === 'string' ? meta.title : undefined,
+    }
+  } catch {
+    return undefined
+  }
+}
+
+/**
+ * Read the projection cache that rides beside a session log.
+ *
+ * The cache is the reliable TITLE source — the log header's title is usually
+ * empty, and the real one only ever arrives as a later `session/title` event —
+ * and it covers subagent sessions too. It also carries `identity.cwd` and
+ * `identity.createdAt`, which the header may lack. Returns `undefined` when the
+ * file is missing or malformed: like the header, this is display metadata and
+ * must never fail a scan.
+ */
+export async function readProjectionCache(cacheFile) {
+  try {
+    const parsed = JSON.parse(await readFile(cacheFile, 'utf8'))
+    const record = parsed?.record
+    if (record === null || typeof record !== 'object') return undefined
+    const title = record.rows?.title?.val
+    const createdAt = record.identity?.createdAt
+    const cwd = record.identity?.cwd
+    return {
+      title: typeof title === 'string' && title.length > 0 ? title : undefined,
+      createdAt: typeof createdAt === 'number' ? createdAt : undefined,
+      cwd: typeof cwd === 'string' && cwd.length > 0 ? cwd : undefined,
+    }
+  } catch {
+    return undefined
+  }
+}
+
+/** A line longer than this is a streaming carrier, not preview material. */
+const MAX_TRANSCRIPT_LINE_CHARS = 1024 * 1024
+
+/** Cheap pre-filter: only lines naming an event worth keeping get parsed. */
+const TRANSCRIPT_INTEREST = /user\/message|assistant\/message|session\/title|"session"/
+
+/** The zstd frame magic every frame starts with. */
+const ZSTD_MAGIC = Buffer.from([0x28, 0xb5, 0x2f, 0xfd])
+
+/**
+ * Decode a multi-frame zstd log.
+ *
+ * DSH appends one zstd frame per writing episode, and BOTH of Node's zstd
+ * APIs decode only the first frame of a file (the streaming decompressor
+ * ends at the first frame boundary too), so the file is split at the frame
+ * magic and every slice decompressed on its own. A magic sequence occurring
+ * inside compressed data — astronomically rare — would slice one real frame
+ * in half; both halves then fail to decode and only that one frame is lost,
+ * which a read-only preview can tolerate (and report).
+ */
+function decodeZstdFrames(buffer) {
+  const offsets = []
+  for (let pos = 0; (pos = buffer.indexOf(ZSTD_MAGIC, pos)) !== -1; pos += 1) offsets.push(pos)
+
+  const chunks = []
+  let failedFrames = 0
+  for (let index = 0; index < offsets.length; index += 1) {
+    const start = offsets[index]
+    const end = index + 1 < offsets.length ? offsets[index + 1] : buffer.length
+    try {
+      chunks.push(zstdDecompressSync(buffer.subarray(start, end)))
+    } catch {
+      failedFrames += 1
+    }
+  }
+  return {
+    text: chunks.length === 0 ? '' : Buffer.concat(chunks).toString('utf8'),
+    decodedFrames: chunks.length,
+    failedFrames,
+  }
+}
+
+/** Join the text blocks of a message content array; reasoning/tool blocks drop. */
+function joinTextBlocks(content) {
+  if (typeof content === 'string') return content
+  if (!Array.isArray(content)) return ''
+  const parts = []
+  for (const block of content) {
+    if (block?.type === 'text' && typeof block.text === 'string') parts.push(block.text)
+  }
+  return parts.join('\n\n')
+}
+
+function clampCount(value, fallback) {
+  const count = Math.trunc(Number(value))
+  return Number.isFinite(count) && count > 0 ? count : fallback
+}
+
+/**
+ * Read a session log as a display-oriented conversation transcript.
+ *
+ * Only the events a preview needs are kept: the conversation messages and the
+ * title event (the header's title is usually empty; the real one arrives as
+ * `session/title`). Messages are projected to `{ role, text, time, model? }`:
+ * text blocks are joined, turns without text (pure reasoning, tool calls) and
+ * DSH-injected context events are dropped, and both the message count and the
+ * per-message length are capped so a huge log cannot flood the panel. A
+ * missing log rejects and undecodable frames are reported as `failedFrames`
+ * instead of failing the read.
+ */
+export async function readSessionTranscript(logFile, options = {}) {
+  const maxMessages = clampCount(options.maxMessages, 500)
+  const maxChars = clampCount(options.maxCharsPerMessage, 8000)
+
+  const raw = await readFile(logFile)
+  let text
+  let failedFrames = 0
+  if (logFile.endsWith('.zstd')) {
+    const decoded = decodeZstdFrames(raw)
+    text = decoded.text
+    failedFrames = decoded.failedFrames
+  } else {
+    text = raw.toString('utf8')
+  }
+
+  const collected = []
+  let title
+  let createdAt
+
+  for (const line of text.split('\n')) {
+    if (line.length > MAX_TRANSCRIPT_LINE_CHARS) continue
+    if (!TRANSCRIPT_INTEREST.test(line)) continue
+    let event
+    try { event = JSON.parse(line) } catch { continue }
+
+    if (event?.type === 'session/title') {
+      if (typeof event.data?.title === 'string' && event.data.title.length > 0) title = event.data.title
+    } else if (event?.type === 'session') {
+      const meta = event.meta ?? event.header ?? event
+      if (typeof meta?.createdAt === 'number') createdAt = meta.createdAt
+    } else if (event?.type === 'user/message' || event?.type === 'assistant/message') {
+      const fromUser = event.type === 'user/message'
+      const holder = fromUser ? event.data : (event.data?.message ?? event.data)
+      const textOfMessage = joinTextBlocks(holder?.content)
+      if (textOfMessage.trim() === '') continue
+      const entry = { role: fromUser ? 'user' : 'assistant', text: textOfMessage }
+      if (typeof event.time === 'number') entry.time = event.time
+      if (!fromUser && typeof holder?.source?.model === 'string') entry.model = holder.source.model
+      if (fromUser && holder?.source?.kind !== 'user') entry.injected = true
+      collected.push(entry)
+    }
+  }
+
+  // DSH's runtime-context snapshots and skill catalogs ride user/message
+  // events too, under a source.kind other than 'user'. When a real human
+  // turn exists they are noise; when none does, keep everything — a preview
+  // must never come back empty over a kind-value difference.
+  const hasHumanInput = collected.some((entry) => entry.injected !== true)
+  const kept = hasHumanInput ? collected.filter((entry) => entry.injected !== true) : collected
+
+  const messages = []
+  let truncated = false
+  for (const entry of kept) {
+    if (messages.length >= maxMessages) {
+      truncated = true
+      break
+    }
+    const message = {
+      role: entry.role,
+      text: entry.text.length > maxChars ? `${entry.text.slice(0, maxChars)}…` : entry.text,
+    }
+    if (entry.time !== undefined) message.time = entry.time
+    if (entry.model !== undefined) message.model = entry.model
+    messages.push(message)
+  }
+
+  return { title, createdAt, messages, totalMessages: kept.length, truncated, failedFrames }
+}
+
+/**
+ * Index every session directory under `<dshHome>/sessions/<workspaceKey>/<sessionId>/`.
+ * The directory listing is the authority: a session unknown to the listing is
+ * not on disk, and a directory whose name is not a valid session id is ignored.
+ */
+export async function scanSessions(roots) {
+  const records = new Map()
+  let workspaceKeys = []
+  try {
+    const entries = await readdir(roots.sessions, { withFileTypes: true })
+    workspaceKeys = entries.filter((entry) => entry.isDirectory()).map((entry) => entry.name)
+  } catch {
+    return { records, workspaceKeys, skipped: [] }
+  }
+
+  const skipped = []
+  for (const workspaceKey of workspaceKeys) {
+    const workspaceDir = path.join(roots.sessions, workspaceKey)
+    let sessionDirs
+    try {
+      sessionDirs = await readdir(workspaceDir, { withFileTypes: true })
+    } catch {
+      continue
+    }
+    for (const dir of sessionDirs) {
+      if (!dir.isDirectory()) continue
+      const sessionDir = path.join(workspaceDir, dir.name)
+      if (!isValidSessionId(dir.name)) {
+        skipped.push({ path: sessionDir, reason: 'directory name is not a valid session id' })
+        continue
+      }
+      if (!isContained(roots.sessions, sessionDir)) {
+        skipped.push({ path: sessionDir, reason: 'path escapes the sessions root' })
+        continue
+      }
+      let files = []
+      try {
+        files = (await readdir(sessionDir, { withFileTypes: true }))
+          .filter((entry) => entry.isFile())
+          .map((entry) => entry.name)
+      } catch {
+        continue
+      }
+      const logName = files.find((name) => LOG_NAME.test(name))
+      if (logName === undefined) {
+        skipped.push({ path: sessionDir, reason: 'no session log found in the session directory' })
+        continue
+      }
+      const logFile = path.join(sessionDir, logName)
+      let bytes = 0
+      try {
+        bytes = (await stat(logFile)).size
+      } catch {
+        bytes = 0
+      }
+      const header = await readSessionHeader(logFile)
+      const cache = await readProjectionCache(path.join(roots.projectionCache, `${dir.name}.json`))
+      records.set(dir.name, {
+        id: dir.name,
+        workspaceKey,
+        dir: sessionDir,
+        logFile,
+        logName,
+        bytes,
+        parent: header?.parent,
+        cwd: header?.cwd ?? cache?.cwd,
+        // The cache title beats the header's (the header's is usually empty);
+        // the header cwd is authoritative, the cache's only fills a gap.
+        title: cache?.title ?? header?.title,
+        createdAt: cache?.createdAt,
+      })
+    }
+  }
+  return { records, workspaceKeys, skipped }
+}
+
+/**
+ * Expand a set of session ids into `{ id, depth }` pairs including every
+ * descendant, following the durable `parentSession` link.
+ */
+export function expandWithDescendants(records, rootIds) {
+  const children = new Map()
+  for (const record of records.values()) {
+    if (typeof record.parent !== 'string' || record.parent.length === 0) continue
+    const list = children.get(record.parent)
+    if (list === undefined) children.set(record.parent, [record.id])
+    else list.push(record.id)
+  }
+
+  const seen = new Set()
+  const out = []
+  const queue = []
+  for (const id of rootIds) {
+    if (typeof id !== 'string' || seen.has(id)) continue
+    seen.add(id)
+    queue.push({ id, depth: 0 })
+  }
+  while (queue.length > 0) {
+    const current = queue.shift()
+    out.push(current)
+    for (const child of children.get(current.id) ?? []) {
+      if (seen.has(child)) continue
+      seen.add(child)
+      queue.push({ id: child, depth: current.depth + 1 })
+    }
+  }
+  return out
+}
+
+/** Read the durable workspace registry; a missing or malformed file yields an empty shape. */
+export async function readWorkspaceRegistry(roots) {
+  try {
+    const raw = await readFile(roots.workspaceRegistry, 'utf8')
+    const parsed = JSON.parse(raw)
+    if (parsed === null || typeof parsed !== 'object') return undefined
+    return parsed
+  } catch {
+    return undefined
+  }
+}
+
+/**
+ * Remove session ids from every place the workspace registry can hold one:
+ * each workspace's `sessionIds`, plus the global `archivedSessionIds` and
+ * `pinnedSessionIds` lists.
+ *
+ * Returns the SAME object when nothing matched, so callers can cheaply detect
+ * a no-op without a deep clone.
+ */
+export function detachFromRegistry(registry, ids) {
+  const remove = new Set(ids)
+  if (registry === undefined || registry === null || typeof registry !== 'object') {
+    return { registry, changes: [] }
+  }
+
+  // Collect first so an unchanged registry is never cloned.
+  const changes = []
+  const global = registry.global
+  if (global !== null && typeof global === 'object') {
+    for (const key of ['archivedSessionIds', 'pinnedSessionIds', 'sessionIds']) {
+      const list = global[key]
+      if (!Array.isArray(list)) continue
+      const removed = list.filter((id) => remove.has(id)).length
+      if (removed > 0) changes.push({ where: `global.${key}`, removed })
+    }
+  }
+  const workspaces = registry.tables?.workspaces
+  if (workspaces !== null && typeof workspaces === 'object') {
+    for (const [workspaceId, workspace] of Object.entries(workspaces)) {
+      const list = workspace?.sessionIds
+      if (!Array.isArray(list)) continue
+      const removed = list.filter((id) => remove.has(id)).length
+      if (removed > 0) changes.push({ where: `workspaces[${workspaceId}].sessionIds`, removed })
+    }
+  }
+  if (changes.length === 0) return { registry, changes }
+
+  const next = structuredClone(registry)
+  const nextGlobal = next.global
+  if (nextGlobal !== null && typeof nextGlobal === 'object') {
+    for (const key of ['archivedSessionIds', 'pinnedSessionIds', 'sessionIds']) {
+      const list = nextGlobal[key]
+      if (Array.isArray(list)) nextGlobal[key] = list.filter((id) => !remove.has(id))
+    }
+  }
+  const nextWorkspaces = next.tables?.workspaces
+  if (nextWorkspaces !== null && typeof nextWorkspaces === 'object') {
+    for (const workspace of Object.values(nextWorkspaces)) {
+      const list = workspace?.sessionIds
+      if (Array.isArray(list)) workspace.sessionIds = list.filter((id) => !remove.has(id))
+    }
+  }
+  return { registry: next, changes }
+}
+
+/** Write JSON atomically: a sibling temp file, then a rename over the target. */
+export async function writeJsonAtomic(file, value) {
+  const temp = `${file}.session-cleaner.${process.pid}.tmp`
+  await writeFile(temp, `${JSON.stringify(value, null, 2)}\n`, 'utf8')
+  await rename(temp, file)
+}
+
+/** Candidate paths for the optional SQLite full-text search index. */
+export async function findSearchIndexes(roots) {
+  const found = []
+  const storages = roots.storages
+  let entries = []
+  try {
+    entries = await readdir(storages, { withFileTypes: true })
+  } catch {
+    return found
+  }
+  for (const entry of entries) {
+    if (!entry.isFile()) continue
+    if (!/\.(?:db|sqlite|sqlite3)$/i.test(entry.name)) continue
+    found.push(path.join(storages, entry.name))
+  }
+  return found
+}
+
+/**
+ * Build a deletion plan. Pure with respect to the filesystem apart from reads.
+ *
+ * Returns `{ root, targets, descendants, registry, registryChanges, searchIndexes, skipped, totals }`.
+ */
+export async function planDeletion(roots, sessionIds, options = {}) {
+  const requested = [...new Set(sessionIds.filter(isValidSessionId))]
+  const rejectedIds = sessionIds.filter((id) => !isValidSessionId(id))
+
+  const { records, skipped } = await scanSessions(roots)
+  const expanded = expandWithDescendants(records, requested)
+
+  const targets = []
+  const missing = []
+  for (const entry of expanded) {
+    const record = records.get(entry.id)
+    if (record === undefined) {
+      // A requested root that is absent from the listing cannot be deleted; a
+      // descendant that is absent simply has no files.
+      if (entry.depth === 0) missing.push(entry.id)
+      continue
+    }
+    targets.push({
+      id: record.id,
+      depth: entry.depth,
+      isRoot: entry.depth === 0,
+      workspaceKey: record.workspaceKey,
+      dir: record.dir,
+      logFile: record.logFile,
+      bytes: record.bytes,
+      cwd: record.cwd,
+      title: record.title,
+    })
+  }
+
+  const registry = await readWorkspaceRegistry(roots)
+  const projected = detachFromRegistry(registry, targets.map((target) => target.id))
+  const searchIndexes = await findSearchIndexes(roots)
+
+  const totals = targets.reduce((sum, target) => sum + target.bytes, 0)
+  return {
+    root: options.rootSessionId,
+    dshHome: roots.dshHome,
+    targets,
+    requested,
+    rejectedIds,
+    missing,
+    registryChanges: projected.changes,
+    searchIndexes,
+    skipped,
+    totals: { sessions: targets.length, bytes: totals },
+  }
+}
+
+/** Validate one concrete path immediately before deleting it. */
+export function assertDeletable(target, allowedRoots) {
+  const normalized = path.resolve(target)
+  if (!allowedRoots.some((root) => isContained(root, normalized))) {
+    return { ok: false, reason: `outside the allowed roots: ${normalized}` }
+  }
+  return { ok: true }
+}
+
+/**
+ * Execute a plan. Order is deliberate and must not be reordered:
+ *   1. stop every running session,
+ *   2. detach registry references,
+ *   3. delete session directories and projection caches,
+ *   4. purge derived search indexes that can no longer be reconciled.
+ *
+ * `stop` is injected so this module stays testable and free of runtime coupling.
+ */
+export async function executeDeletion(roots, plan, options = {}) {
+  const { stop, now = () => Date.now() } = options
+  const allowedRoots = [roots.sessions, roots.storages]
+  const result = {
+    startedAt: now(),
+    stopped: [],
+    registryChanges: [],
+    deleted: [],
+    failed: [],
+    searchIndexesRemoved: [],
+    searchIndexesKept: [],
+  }
+
+  // 1. Stop each session before any file is touched. A still-running agent
+  //    would keep appending to a log we just deleted, resurrecting a half
+  //    session. A target that refuses to stop aborts the whole run before
+  //    anything is deleted, so a partial outcome is never reported as success.
+  for (const target of plan.targets) {
+    const outcome = await stop(target.id)
+    result.stopped.push({ id: target.id, ...outcome })
+    if (outcome?.stillRunning === true && options.requireStopped !== false) {
+      result.failed.push({ id: target.id, stage: 'stop', reason: 'session is still running' })
+      return { ...result, aborted: true }
+    }
+  }
+
+  const blocked = new Set(result.failed.map((entry) => entry.id))
+
+  // 2. Detach registry references first, so no row survives its own log.
+  let registry = await readWorkspaceRegistry(roots)
+  if (registry !== undefined) {
+    const removable = plan.targets.filter((target) => !blocked.has(target.id)).map((target) => target.id)
+    const projected = detachFromRegistry(registry, removable)
+    if (projected.changes.length > 0) {
+      try {
+        await writeJsonAtomic(roots.workspaceRegistry, projected.registry)
+        registry = projected.registry
+        result.registryChanges = projected.changes
+      } catch (error) {
+        result.failed.push({ id: '*', stage: 'registry', reason: String(error?.message ?? error) })
+      }
+    }
+  }
+
+  // 3. Delete each session directory and its projection cache.
+  for (const target of plan.targets) {
+    if (blocked.has(target.id)) continue
+    const check = assertDeletable(target.dir, allowedRoots)
+    if (!check.ok) {
+      result.failed.push({ id: target.id, stage: 'validate', reason: check.reason })
+      continue
+    }
+    try {
+      await rm(target.dir, { recursive: true, force: true })
+      result.deleted.push({ id: target.id, kind: 'session-dir', path: target.dir, bytes: target.bytes })
+    } catch (error) {
+      result.failed.push({ id: target.id, stage: 'session-dir', reason: String(error?.message ?? error) })
+      continue
+    }
+    const cacheFile = path.join(roots.projectionCache, `${target.id}.json`)
+    const cacheCheck = assertDeletable(cacheFile, allowedRoots)
+    if (!cacheCheck.ok) {
+      result.failed.push({ id: target.id, stage: 'cache-validate', reason: cacheCheck.reason })
+      continue
+    }
+    if (existsSync(cacheFile)) {
+      try {
+        await rm(cacheFile, { force: true })
+        result.deleted.push({ id: target.id, kind: 'projection-cache', path: cacheFile })
+      } catch (error) {
+        result.failed.push({ id: target.id, stage: 'projection-cache', reason: String(error?.message ?? error) })
+      }
+    }
+  }
+
+  // 4. Derived search indexes are disposable: removing unreconcilable ones is
+  //    safe because they are rebuilt from the surviving logs on next search.
+  for (const indexFile of plan.searchIndexes) {
+    const check = assertDeletable(indexFile, allowedRoots)
+    if (!check.ok) {
+      result.searchIndexesKept.push({ path: indexFile, reason: check.reason })
+      continue
+    }
+    try {
+      const info = await stat(indexFile)
+      await rm(indexFile, { force: true })
+      result.searchIndexesRemoved.push({ path: indexFile, bytes: info.size })
+      for (const suffix of ['-wal', '-shm']) {
+        const sidecar = `${indexFile}${suffix}`
+        if (existsSync(sidecar)) await rm(sidecar, { force: true })
+      }
+    } catch (error) {
+      result.searchIndexesKept.push({ path: indexFile, reason: String(error?.message ?? error) })
+    }
+  }
+
+  result.finishedAt = now()
+  result.summary = {
+    sessionsDeleted: result.deleted.filter((entry) => entry.kind === 'session-dir').length,
+    cachesDeleted: result.deleted.filter((entry) => entry.kind === 'projection-cache').length,
+    bytesDeleted: result.deleted.reduce((sum, entry) => sum + (entry.bytes ?? 0), 0),
+    failures: result.failed.length,
+  }
+  return result
+}
+
+/** Default roots derived from a DSH home directory. */
+export function rootsFor(dshHome) {
+  const home = path.resolve(dshHome)
+  return {
+    dshHome: home,
+    sessions: path.join(home, 'sessions'),
+    storages: path.join(home, 'storages'),
+    projectionCache: path.join(home, 'storages', 'session_projcache', 'sessions'),
+    workspaceRegistry: path.join(home, 'storages', 'workspace.json'),
+  }
+}