@ljwei-stak/dsh-model-route-bot 0.66.4 → 0.67.0

package/modules/bot-runtime/product-overlay/base/src/host/privacy.js

@@ -0,0 +1,368 @@
+// Private Bots. A private Bot reads the team like any Bot, but no other Bot reads it, and
+// what it sends out waits for the user. DSH does not fence reads, so this closes the
+// tools DS Bot owns and checks the paths that the file tools and the shell are given;
+// a shell can still reach any file the account can (see docs/plans).
+import { join, resolve } from 'node:path'
+import { within } from './folders.js'
+import { unquote } from './text.js'
+
+const FILE_TOOLS = { read: 'file_path', write: 'file_path', edit: 'file_path', read_image: 'file_path', str_replace_editor: 'path' }
+const WRITE_TOOLS = ['write', 'edit', 'str_replace_editor']
+const SEARCH_TOOLS = ['glob', 'grep']
+const SHELL_TOOLS = ['bash', 'pwsh', 'terminal_open', 'terminal_send']
+const SESSION_TOOLS = ['session_search', 'session_event_search', 'session_trace', 'session_event_trace', 'session_event_read']
+
+// Windows and macOS file systems ignore case.
+const fold = text => (process.platform === 'linux' ? text : text.toLowerCase())
+const overlap = (a, b) => within(a, b) || within(b, a)
+const strings = value => (typeof value === 'string' ? [value] : Array.isArray(value) ? value.flatMap(strings) : value && typeof value === 'object' ? Object.values(value).flatMap(strings) : [])
+// The same path as a shell writes it: forward slashes, and on Windows both drive forms.
+const asShell = (path) => {
+  const slashed = path.replace(/\\/g, '/')
+  const drive = /^([a-z]):\//i.exec(slashed)
+  return drive ? [slashed, `/${drive[1]}${slashed.slice(2)}`] : [slashed]
+}
+
+export function install(rt) {
+  const { ctx, state } = rt
+
+  const isPrivate = id => rt.botOf(id)?.private === true
+  const privateBots = () => Object.values(state.bots).filter(bot => bot.private === true)
+  // Deleting a private Bot does not make the files and memory it leaves behind public.
+  const protectedBots = () => [...new Map([...(state.privateArchives ?? []), ...privateBots()].map(bot => [bot.id, bot])).values()]
+  const sessionsOf = bot => [...new Set([bot.id, ...(bot.sessionIds ?? [rt.chatOf(bot.id), ...(bot.segments ?? []).map(segment => segment.sessionId), ...rt.groupSessionsOf(bot.id)])])]
+  const dataRoots = () => [rt.dshHome, rt.home, process.env.DS_BOT_USER_DSH_HOME?.trim()].filter(Boolean)
+
+  // A path as it is and as the file system resolves it, folded for comparing. realBeyond
+  // also resolves a link above a file that does not exist yet.
+  const spellings = async (path) => [...new Set([fold(resolve(path)), fold(await rt.realBeyond(resolve(path)))])]
+
+  // The Bot a call speaks for. A subagent speaks for the Bot at the root of its
+  // parent chain, however many subagents sit between them.
+  const callerOf = async (exec) => {
+    const own = rt.selfOf(exec.agent?.id) ?? rt.selfOf(exec.agent?.session?.id)
+    if (own !== undefined) return own
+    let parent = exec.agent?.session?.header?.parentSession
+    const seen = new Set()
+    while (parent !== undefined && !seen.has(parent)) {
+      seen.add(parent)
+      const owner = rt.selfOf(parent)
+      if (owner !== undefined) return owner
+      const agent = await rt.resolveAgent(parent).catch(() => undefined)
+      parent = agent?.session?.header?.parentSession
+    }
+    return undefined
+  }
+
+  // Where a private Bot keeps what must stay closed: its folder and its memory. Its
+  // Sessions and its id are in file names under DSH's and the team's data.
+  const regionsOf = async bot => (await Promise.all([
+    bot.workDir,
+    join(rt.home, 'memory', 'bots', bot.id),
+    join(rt.home, 'browser-profiles', bot.id),
+    join(rt.home, 'browser-downloads', bot.id),
+    join(rt.home, 'side', `${bot.id}.json`),
+    join(rt.home, 'btw', bot.id),
+    bot.archivePath,
+  ].filter(Boolean).map(spellings))).flat()
+  const sharedFiles = async () => (await Promise.all([
+    rt.statePath,
+    join(rt.home, 'chat-index.sqlite'),
+    join(rt.home, 'chat-index.sqlite-wal'),
+    join(rt.home, 'chat-index.sqlite-shm'),
+    join(rt.home, 'usage.sqlite'),
+    join(rt.home, 'usage.sqlite-wal'),
+    join(rt.home, 'usage.sqlite-shm'),
+    join(rt.home, 'inflight.json'),
+    join(rt.home, 'learning', 'state.json'),
+    join(rt.home, 'ocr.json'),
+    join(rt.home, 'secrets.json'),
+    join(rt.home, 'memory', 'journal.jsonl'),
+    join(rt.home, 'memory', 'journal.torn'),
+    join(rt.home, 'memory', 'pending.json'),
+    join(rt.home, 'logs', 'diagnostics.jsonl'),
+  ].filter(Boolean).map(spellings))).flat()
+
+  const teamFolder = () => rt.teamFolder() ?? resolve(rt.defaultTeamFolder())
+  const teamMemory = join(rt.home, 'memory', 'team')
+  // The memory of deleted Bots and groups. What a private Bot left there is a region of
+  // its own (archivePath) and is checked first; the rest was never closed, and the
+  // memory of a Bot whose protection the user lifted is in it.
+  const memoryArchive = join(rt.home, 'memory', 'archive')
+
+  // The private Bot that `path` belongs to, other than `self`; undefined when it is open.
+  // `search` is for tools that walk a folder: a folder that holds a closed one is closed
+  // too, and so are the data folders, where the walk cannot be told which files to skip.
+  async function ownerOfPath(path, { self, search = false, base } = {}) {
+    const bots = protectedBots()
+    if (bots.length === 0) return undefined
+    const keys = await spellings(base === undefined ? path : resolve(base, path))
+    for (const bot of bots) {
+      if (bot.id === self) continue
+      const needles = sessionsOf(bot).map(fold)
+      const regions = await regionsOf(bot)
+      for (const key of keys) {
+        if (regions.some(region => within(key, region) || (search && within(region, key)))) return bot
+        if (needles.some(needle => key.includes(needle))) return bot
+      }
+    }
+    const other = bots.find(bot => bot.id !== self)
+    for (const file of await sharedFiles()) if (other !== undefined && keys.some(key => key.startsWith(file))) return other
+    if (other !== undefined) {
+      const ownBot = rt.botOf(self)
+      const ownIds = ownBot === undefined ? [] : sessionsOf(ownBot).map(fold)
+      const ownRegions = ownBot === undefined ? [] : await regionsOf(ownBot)
+      // A walk of the archive stays closed while a private Bot's memory is on its way
+      // there: until the path is known, the walk cannot be told which folder to skip.
+      const moving = search && bots.some(bot => bot.archiving === true)
+      const open = [rt.workDirOf({ botId: self }), teamFolder(), teamMemory, moving ? undefined : memoryArchive, ...ownRegions].filter(Boolean)
+      const folders = (await Promise.all(open.map(spellings))).flat()
+      const roots = (await Promise.all(dataRoots().map(spellings))).flat()
+      for (const key of keys) {
+        if ((!search && ownIds.some(id => key.includes(id))) || folders.some(folder => within(key, folder))) continue
+        // Direct reads are closed too, not only walks: an index, pending record or log
+        // can hold the same private text without naming the Session in its file name.
+        if (roots.some(root => within(key, root) || (search && within(root, key)))) return other
+      }
+    }
+    return undefined
+  }
+
+  // Text a shell runs: a closed path anywhere in it closes the command. A path that the
+  // command builds, or reaches by walking up from a folder, is not seen.
+  async function ownerInText(text, self) {
+    const flat = fold(text.replace(/\\/g, '/'))
+    const bots = protectedBots()
+    for (const bot of bots) {
+      if (bot.id === self) continue
+      const spelled = (await regionsOf(bot)).flatMap(asShell)
+      if (spelled.some(region => flat.includes(region)) || sessionsOf(bot).some(id => flat.includes(fold(id)))) return bot
+    }
+    const other = bots.find(bot => bot.id !== self)
+    for (const file of await sharedFiles()) if (other !== undefined && asShell(file).some(spelled => flat.includes(spelled))) return other
+    return undefined
+  }
+
+  const closed = (bot, what) => new Error(`${bot.name} is a private Bot, and other Bots cannot reach its ${what}. Ask the user if ${bot.name} should share something.`)
+
+  // What a private Bot sends out is shown to the user first, as the user sees it.
+  async function approve(exec, { en, zh }) {
+    const approval = ctx.get?.('approval')
+    if (approval === undefined || exec.agent?.session === undefined) throw new Error('A private Bot needs the user\'s approval for this, and there is no way to ask for it here, so it was not done.')
+    const outcome = await approval.request({
+      agent: exec.agent,
+      toolName: exec.name,
+      ...(exec.callId !== undefined ? { callId: exec.callId } : {}),
+      reason: en,
+      displayReason: { en, zh },
+      ...(exec.signal ? { signal: exec.signal } : {}),
+    })
+    if (outcome === 'allowed-once') return
+    throw new Error(outcome === 'rejected' ? 'The user did not allow it, so it was not done.'
+      : outcome === 'cancelled' ? 'The approval was cancelled, so it was not done.'
+        : 'The approval could not be answered, so it was not done.')
+  }
+
+  // The Bot a tool argument names, as the tools and the memory scopes read it.
+  const botNamed = ref => (typeof ref === 'string' && ref !== '' ? rt.findBot(ref) ?? rt.findBot(unquote(ref)) : undefined)
+
+  // Tools between Bots. Other Bots may not reach a private Bot; what a private Bot sends
+  // to them is the user's to allow. A private Bot is no member of a group, which the
+  // roster checks as the members are written.
+  async function guardTeam(exec, self) {
+    const args = exec.arguments ?? {}
+    const target = (ref) => {
+      const bot = botNamed(ref)
+      return bot !== undefined && isPrivate(bot.id) && bot.id !== self ? bot : undefined
+    }
+    switch (exec.name) {
+      case 'message_bot': {
+        const bot = target(args.to)
+        if (bot) throw new Error(`Not sent: ${bot.name} is a private Bot, and other Bots cannot message it. Ask the user to pass it on.`)
+        const to = botNamed(args.to)
+        if (isPrivate(self) && to !== undefined) await approve(exec, { en: `Let ${rt.botOf(self).name} send this to ${to.name}?\n\n${String(args.message ?? '')}`, zh: `允许 ${rt.botOf(self).name} 把下面的内容发给 ${to.name} 吗？\n\n${String(args.message ?? '')}` })
+        return
+      }
+      // Every Bot reads the shared capabilities, so saving one shares what is in it.
+      case 'save_capability': {
+        if (!isPrivate(self)) return
+        const shown = JSON.stringify(args, null, 2)
+        await approve(exec, { en: `Let ${rt.botOf(self).name} share this with the team?\n\n${shown}`, zh: `允许 ${rt.botOf(self).name} 把下面的内容共享给团队吗？\n\n${shown}` })
+        return
+      }
+      case 'stop_bot': {
+        const bot = target(args.bot)
+        if (bot) throw closed(bot, 'turns')
+        return
+      }
+      case 'attach_capability':
+      case 'detach_capability': {
+        const bot = target(args.bot)
+        if (bot) throw closed(bot, 'capabilities')
+        return
+      }
+      case 'update_bot': {
+        const bot = target(args.bot)
+        if (bot) throw closed(bot, 'settings')
+        return
+      }
+      case 'recall':
+      case 'remember':
+      case 'forget': {
+        const bot = target(args.scope)
+        if (bot) throw closed(bot, 'memory')
+        return
+      }
+      default:
+    }
+  }
+
+  // A Session started under a private Bot is closed too, even when its own id is not in
+  // the Bot's chat history. Walk the same parent chain as the caller check.
+  async function ownerOfSession(id, self) {
+    const bots = protectedBots()
+    const owns = (sessionId, bot) => bot.id !== self && sessionsOf(bot).includes(sessionId)
+    const direct = bots.find(bot => owns(id, bot))
+    if (direct) return direct
+    const seen = new Set([id])
+    let current = id
+    for (;;) {
+      const agent = await rt.resolveAgent(current).catch(() => undefined)
+      const parent = agent?.session?.header?.parentSession
+      if (parent === undefined || seen.has(parent)) return undefined
+      const owner = bots.find(bot => owns(parent, bot))
+      if (owner) return owner
+      seen.add(parent)
+      current = parent
+    }
+  }
+
+  // The file tools, the shell and the session search: where the path or text names a
+  // closed place, the call fails before it runs.
+  async function guardFiles(exec, self) {
+    const args = exec.arguments ?? {}
+    const folder = () => rt.sessionFolder(exec.agent?.session ?? exec.agent)
+    const refuse = (bot) => { throw closed(bot, 'files') }
+    const name = exec.name
+    if (name in FILE_TOOLS) {
+      const path = args[FILE_TOOLS[name]]
+      if (typeof path !== 'string' || path === '') return
+      const bot = await ownerOfPath(path, { self, base: folder() })
+      if (bot) refuse(bot)
+      // A private Bot's file changes stay in its own folder unless the user allows them.
+      if (isPrivate(self) && WRITE_TOOLS.includes(name)) {
+        const own = await spellings(rt.botOf(self).workDir)
+        const base = folder()
+        const target = await spellings(base === undefined ? path : resolve(base, path))
+        if (!target.every(key => own.some(root => within(key, root)))) await approve(exec, { en: `Let ${rt.botOf(self).name} change ${path}? It is outside its own folder.`, zh: `允许 ${rt.botOf(self).name} 修改 ${path} 吗？它在自己的文件夹之外。` })
+      }
+      return
+    }
+    if (SEARCH_TOOLS.includes(name)) {
+      const where = typeof args.path === 'string' && args.path !== '' ? args.path : '.'
+      const bot = await ownerOfPath(where, { self, search: true, base: folder() })
+        ?? (typeof args.pattern === 'string' && /^([a-z]:|[\\/]|~)/i.test(args.pattern) ? await ownerOfPath(args.pattern.replace(/[*?[{].*$/, ''), { self, search: true, base: folder() }) : undefined)
+      if (bot) refuse(bot)
+      return
+    }
+    if (SHELL_TOOLS.includes(name)) {
+      const bot = await ownerInText(strings(args).join('\n'), self)
+      if (bot) refuse(bot)
+      return
+    }
+    if ((name === 'subagent' || name === 'subagent_fork') && typeof args.cwd === 'string') {
+      const bot = await ownerOfPath(args.cwd, { self, search: true, base: folder() })
+      if (bot) refuse(bot)
+      return
+    }
+    if (name === 'working_directory' && typeof args.cd === 'string') {
+      const bot = await ownerOfPath(args.cd, { self, search: true, base: folder() })
+      if (bot) refuse(bot)
+      return
+    }
+    if (SESSION_TOOLS.includes(name)) {
+      const asked = strings([args.session_id, args.session_ids, args.parent_session_ids])
+      for (const id of asked) {
+        const bot = await ownerOfSession(id, self)
+        if (bot) refuse(bot)
+      }
+      // DSH scopes Session search to the caller's cwd. A subagent can have another cwd,
+      // so an unfiltered search is closed too; explicit safe ids still narrow it.
+      if (name === 'session_search' && args.session_ids === undefined && args.parent_session_ids === undefined) {
+        const closedBot = protectedBots().find(found => found.id !== self)
+        if (closedBot) throw closed(closedBot, 'Sessions')
+      }
+    }
+  }
+
+  // Before the tools' own hooks, so that a closed call never reaches a card they show.
+  ctx.on('tools/execute', async (exec, next) => {
+    if (protectedBots().length === 0) return next()
+    const self = await callerOf(exec)
+    await guardTeam(exec, self)
+    await guardFiles(exec, self)
+    return next()
+  })
+
+  // A folder that a private Bot's folder would overlap, or that overlaps one.
+  async function privateClash(path, { except } = {}) {
+    const keys = await spellings(path)
+    for (const bot of protectedBots()) {
+      if (bot.id === except || bot.workDir === undefined) continue
+      const dirs = await spellings(bot.workDir)
+      if (keys.some(key => dirs.some(dir => overlap(key, dir)))) return bot
+    }
+    return undefined
+  }
+
+  // Checks a change of `private` or of the folder of a Bot before it is written. A
+  // private Bot's folder is no other Bot's, no group's and not the team workspace —
+  // a member's own pick in a group counts as that member's folder — because DSH
+  // cannot be told which files of a shared folder to hide.
+  async function checkChange(bot, { flag, moving, workDir }) {
+    const makes = flag === undefined ? bot.private === true : flag
+    const dir = moving ? workDir : bot.workDir
+    if (!makes) {
+      const clash = moving && dir !== undefined ? await privateClash(dir, { except: bot.id }) : undefined
+      if (clash) throw new Error(`${dir} is in or around the folder of ${clash.name}, a private Bot, so ${bot.name} cannot work there`)
+      return
+    }
+    if (rt.isMain(bot.id)) throw new Error(`${bot.name} is a Main Bot, so it cannot be private`)
+    if (dir === undefined) throw new Error(`${bot.name} needs a folder of its own to be private`)
+    if (Object.values(state.rooms).some(room => room.members.includes(bot.id))) throw new Error(`${bot.name} is in a group chat, where every member reads what is said. Remove it from its groups first`)
+    const clash = await privateClash(dir, { except: bot.id })
+    if (clash) throw new Error(`${dir} is in or around the folder of ${clash.name}, a private Bot, so ${bot.name} cannot work there`)
+    const keys = await spellings(dir)
+    const others = [
+      ['the team workspace', teamFolder()],
+      ...Object.values(state.bots).filter(other => other.id !== bot.id && other.workDir).map(other => [`the folder of ${other.name}`, other.workDir]),
+      ...Object.values(state.rooms).filter(room => room.workDir).map(room => [`the folder of "${room.name}"`, room.workDir]),
+      ...Object.values(state.rooms).flatMap(room => Object.entries(room.memberDirs ?? {})
+        .filter(([member, dir]) => room.members.includes(member) && dir !== null)
+        .map(([member, dir]) => [`the folder of ${rt.botOf(member)?.name ?? 'a member'} in "${room.name}"`, dir])),
+    ]
+    for (const [what, path] of others) {
+      const spelled = await spellings(path)
+      if (keys.some(key => spelled.some(other => overlap(key, other)))) throw new Error(`${dir} is in or around ${what}, so ${bot.name} cannot be private there. Give it a folder of its own`)
+    }
+  }
+
+  const checkNoPrivate = (ids) => {
+    const found = ids.map(id => rt.botOf(id)).find(bot => bot?.private === true)
+    if (found) throw new Error(`${found.name} is a private Bot and cannot be in a group chat: what is said in a group is read by every member`)
+  }
+
+  // Deleting a private Bot keeps its places closed (see roster.js deleteBot); only the
+  // user opens them again, one Bot at a time. The files stay where they are.
+  async function releaseArchive(id) {
+    const kept = state.privateArchives ?? []
+    if (!kept.some(entry => entry.id === id)) throw new Error('Nothing kept closed for this Bot')
+    state.privateArchives = kept.filter(entry => entry.id !== id)
+    await rt.save()
+  }
+
+  // What the settings list shows. Session ids stay on the host.
+  const archivesView = () => (state.privateArchives ?? []).map(({ id, name, workDir }) => ({ id, name, ...(workDir === undefined ? {} : { workDir }) }))
+
+  Object.assign(rt, { isPrivate, privateClash, checkPrivacyChange: checkChange, checkNoPrivate, releaseArchive, archivesView })
+}