@hedeqiang/dsh-nl2sql 0.12.1 → 0.12.2

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hedeqiang/dsh-nl2sql",
-  "version": "0.12.1",
+  "version": "0.12.2",
   "description": "Ask MySQL, PostgreSQL, Elasticsearch and CSV/Excel in plain language inside DeepSeek Harness: charts, dashboards, alerts, and IM push and bots for Lark/Feishu, Telegram, WeCom and DingTalk.",
   "type": "module",
   "exports": {

package/service/alerts.js

@@ -6,6 +6,9 @@ import { evaluateAlert } from '../analysis.js'
 import { admitSql } from '../sql-guard.js'
 import { UserError, clampInt, newId, oneOf, requireText, upsert } from './shared.js'
 
+/** Fields written by checks (not by the edit form). */
+const ALERT_RUNTIME = ['state', 'lastCheckedAt', 'lastValue', 'lastResultId', 'lastFiredAt', 'lastError']
+
 export const methods = {
   // ---------------------------------------------------------------- alerts
 
@@ -53,6 +56,10 @@ export const methods = {
 
   async saveAlert(input) {
     const a = await this.normalizeAlert(input)
+    // A check may have finished between normalize and now: keep its runtime fields
+    // (last check / fire time, state) from the stored object, never the stale copy.
+    const current = (this.state.alerts ?? []).find(x => x.id === a.id)
+    if (current) for (const k of ALERT_RUNTIME) { if (current[k] === undefined) delete a[k]; else a[k] = current[k] }
     this.state.alerts = upsert(this.state.alerts ?? [], a)
     await this.save()
     this.scheduleAlerts()
@@ -73,20 +80,18 @@ export const methods = {
 
   /** Evaluate an unsaved alert definition (form "dry run"): nothing is stored or pushed. */
   async previewAlert(input) {
-    const tmpId = `preview-${newId('al')}`
-    const saved = this.state.alerts
-    try {
-      const draft = await this.normalizeAlert({ ...input, id: undefined }, tmpId)
-      this.state.alerts = [...(saved ?? []), draft]
-      return await this.checkAlertInner(tmpId, { dryRun: true })
-    } finally {
-      this.state.alerts = saved
-    }
+    // The draft is evaluated on its own: state.alerts is never touched, so concurrent
+    // saves are kept and the draft is never persisted.
+    const draft = await this.normalizeAlert({ ...input, id: undefined }, `preview-${newId('al')}`)
+    return this.checkAlertInner(draft.id, { dryRun: true, draft })
   },
 
-  async checkAlertInner(id, { dryRun = false, force = false } = {}) {
-    const a = (this.state.alerts ?? []).find(x => x.id === id)
-    if (!a) throw new UserError('nl2sql/invalid', '告警不存在')
+  async checkAlertInner(id, { dryRun = false, force = false, draft } = {}) {
+    const found = draft ?? (this.state.alerts ?? []).find(x => x.id === id)
+    if (!found) throw new UserError('nl2sql/invalid', '告警不存在')
+    // Evaluate a snapshot; runtime fields are applied below to the object stored
+    // *after* the query (saveAlert may have replaced it meanwhile).
+    let a = draft ? structuredClone(draft) : { ...found }
     const ds = this.requireDatasource(a.datasourceId)
     let result
     if (a.source === 'metric') {
@@ -100,6 +105,10 @@ export const methods = {
     const series = vi === undefined ? [] : result.rows.map(r => ({ t: r[0], v: r[vi] === null ? null : Number(r[vi]) }))
     if (a.condition.type !== 'missing' && result.columns.length === 1 && vi === 0) series.splice(0, series.length, ...result.rows.map((r, i) => ({ t: i, v: Number(r[0]) })))
     const ev = evaluateAlert(a.condition, series, { now: Date.now() })
+    if (!draft) {
+      const current = (this.state.alerts ?? []).find(x => x.id === id)
+      if (current) a = current
+    }
     a.lastCheckedAt = Date.now()
     a.lastValue = ev.value ?? null
     a.lastResultId = result.resultId
@@ -163,12 +172,20 @@ export const methods = {
     const last = a.lastCheckedAt ?? 0
     if (a.every === '15m') return last + 15 * 60000
     if (a.every === 'hour') return last + 60 * 60000
-    // daily at HH:MM local time: the most recent slot not yet covered by lastCheckedAt.
+    // daily at HH:MM local wall time: the most recent slot not yet covered by lastCheckedAt.
+    // DST (same rule as the Harness schedule service): a slot that does not exist
+    // that day (spring forward) is skipped; a repeated one (fall back) is the earlier instant.
     const [hh, mm] = String(a.at ?? '09:00').split(':').map(Number)
-    const slot = new Date(now)
-    slot.setHours(hh, mm, 0, 0)
-    if (slot.getTime() > now) slot.setDate(slot.getDate() - 1)
-    return last < slot.getTime() ? slot.getTime() : (() => { const n = new Date(slot); n.setDate(n.getDate() + 1); return n.getTime() })()
+    const today = new Date(now)
+    const slotOn = (offset) => {
+      const d = new Date(today.getFullYear(), today.getMonth(), today.getDate() + offset, hh, mm, 0, 0)
+      return d.getHours() === hh && d.getMinutes() === mm ? d.getTime() : undefined
+    }
+    let prev
+    for (let k = 0; k >= -3 && prev === undefined; k--) { const s = slotOn(k); if (s !== undefined && s <= now) prev = s }
+    if (prev !== undefined && last < prev) return prev
+    for (let k = 0; k <= 3; k++) { const s = slotOn(k); if (s !== undefined && s > now && s > last) return s }
+    return now + 86400000
   },
 
   async withJobLock(key, fn) {
@@ -192,12 +209,15 @@ export const methods = {
             this.checkAlert(a.id),
             new Promise((_, rej) => { const t = setTimeout(() => rej(new Error(`检查超过 ${Math.round(timeoutMs / 1000)} 秒`)), timeoutMs); t.unref?.() }),
           ])
-          a.lastError = undefined
+          // The stored object may have been replaced by an edit meanwhile.
+          const cur = (this.state.alerts ?? []).find(x => x.id === a.id)
+          if (cur) cur.lastError = undefined
           ran.push({ id: a.id, ok: true })
         } catch (error) {
           if (error.code === 'nl2sql/busy') return
-          a.lastError = error.message
-          a.lastCheckedAt = Date.now()
+          const cur = (this.state.alerts ?? []).find(x => x.id === a.id) ?? a
+          cur.lastError = error.message
+          cur.lastCheckedAt = Date.now()
           ran.push({ id: a.id, ok: false, error: error.message })
         }
       }))

package/service/bots.js

@@ -35,6 +35,29 @@ const toast = (type, content) => ({ toast: { type, content } })
 const label = (platform) => PLATFORM_LABEL[platform] ?? platform ?? 'IM'
 const SOURCE = (platform) => `${platform ?? 'im'}bot`
 
+const BARE_COMMANDS = {
+  帮助: 'help', help: 'help', start: 'help',
+  数据源: 'sources', sources: 'sources', datasources: 'sources',
+  自动: 'auto', auto: 'auto',
+  新对话: 'reset', new: 'reset', reset: 'reset',
+}
+
+/**
+ * A bot command, or null for a question. Argument-less commands must be the whole
+ * message (optionally `/`-prefixed, trailing punctuation ignored), so questions like
+ * 「自动续费用户有多少」 reach the agent. `切换` / `切换数据源` / `switch` / `use` take an
+ * argument after whitespace; alone they ask for usage.
+ * → { verb: 'help'|'sources'|'auto'|'reset'|'switch', arg? }
+ */
+export function parseBotCommand(text) {
+  const s = String(text ?? '').trim().replace(/^\//, '').replace(/[\s?？。.!！~～]+$/u, '')
+  const bare = BARE_COMMANDS[s.toLowerCase()]
+  if (bare) return { verb: bare }
+  const m = /^(切换数据源|切换|switch|use)(?:\s+([\s\S]+))?$/i.exec(s)
+  if (m) return { verb: 'switch', arg: (m[2] ?? '').trim() }
+  return null
+}
+
 function hasAggregate(r, dialect) {
   const parsed = parseSelect(r.baseSql ?? r.sql, { dialect })
   return Boolean(parsed?.select?.some(i => i.isAggregate))
@@ -286,8 +309,8 @@ export const methods = {
     if (m.chatType === 'group' && m.chatId) this.noteSeenIdentity({ platform, kind: 'chat', id: String(m.chatId), name: chatName, chatName, channelId: m.channelId, refused: Boolean(resolved.refused) })
     if (resolved.refused) { await this.save(); await reply(resolved.refused); return }
     const actor = resolved.actor
-    // Commands: 帮助 / 数据源 / 切换 <名称> / 新对话
-    const cmd = text.match(/^\/?(帮助|help|数据源|切换数据源|切换|自动|auto|新对话|reset|start)\s*(.*)$/i)
+    // Commands: 帮助 / 数据源 / 切换 <名称> / 自动 / 新对话 (and English aliases).
+    const cmd = parseBotCommand(text)
     // Members with different scopes must not share a conversation: per-user sessions
     // whenever the sender's effective scope or rules differ from the chat's.
     const scopeSig = JSON.stringify([actor.datasourceIds, actor.role, actor.rowRules ?? []])
@@ -301,14 +324,15 @@ export const methods = {
     const allowedDs = this.state.datasources.filter(d => !Array.isArray(actor.datasourceIds) || actor.datasourceIds.includes(d.id))
     if (conv.datasourceId && !allowedDs.some(d => d.id === conv.datasourceId)) conv.datasourceId = allowedDs[0]?.id ?? null
     if (cmd) {
-      const verb = cmd[1].toLowerCase()
-      if (verb === '帮助' || verb === 'help' || verb === 'start') { await reply(`直接${m.chatType === 'group' ? ' @我 ' : ''}提问即可，例如「上周每天的订单量」。\n命令：数据源（查看可用数据源）、切换 <名称>、自动（按问题自动选择数据源）、新对话。\n当前数据源：${this.findDatasource(conv.datasourceId)?.name ?? '未选择'}${conv.datasourcePinned ? '（已指定）' : allowedDs.length > 1 ? '（按问题自动选择）' : ''}${actor.role === 'viewer' ? '\n你是只读成员：可以查询指标库中的指标、已验证的常见问题和看板。' : ''}`); return }
-      if (verb === '数据源') { await reply(`可用数据源：\n${allowedDs.map(d => `· ${d.name}${d.description ? ` — ${d.description}` : ''}`).join('\n') || '无'}\n当前：${this.findDatasource(conv.datasourceId)?.name ?? '未选择'}${conv.datasourcePinned ? '（已指定，发送「自动」恢复按问题自动选择）' : allowedDs.length > 1 ? '（按问题自动选择；发送「切换 <名称>」固定）' : ''}`); return }
-      if (verb === '自动' || verb === 'auto') { delete conv.datasourcePinned; convs[key] = conv; await this.save(); await reply(`已改为按问题自动选择数据源（可用：${allowedDs.map(d => d.name).join('、')}）。`); return }
-      if (verb === '新对话' || verb === 'reset') { conv.sessionId = null; convs[key] = conv; await this.save(); await reply('已开始新对话。'); return }
-      const want = cmd[2].trim().toLowerCase()
-      const target = allowedDs.find(d => d.name.toLowerCase() === want || d.id === cmd[2].trim()) ?? (want ? allowedDs.filter(d => d.name.toLowerCase().includes(want) || (d.database ?? '').toLowerCase() === want) : []).find((_, __, arr) => arr.length === 1)
-      if (!target) { await reply(`找不到可用的数据源「${cmd[2]}」。可用：${allowedDs.map(d => d.name).join('、')}`); return }
+      const { verb } = cmd
+      if (verb === 'help') { await reply(`直接${m.chatType === 'group' ? ' @我 ' : ''}提问即可，例如「上周每天的订单量」。\n命令（单独发送）：\n· 数据源 / sources — 查看可用数据源\n· 切换 <名称> / switch <name> — 固定使用某个数据源\n· 自动 / auto — 按问题自动选择数据源\n· 新对话 / new — 开始新对话\n· 帮助 / help — 显示本说明\n当前数据源：${this.findDatasource(conv.datasourceId)?.name ?? '未选择'}${conv.datasourcePinned ? '（已指定）' : allowedDs.length > 1 ? '（按问题自动选择）' : ''}${actor.role === 'viewer' ? '\n你是只读成员：可以查询指标库中的指标、已验证的常见问题和看板。' : ''}`); return }
+      if (verb === 'sources') { await reply(`可用数据源：\n${allowedDs.map(d => `· ${d.name}${d.description ? ` — ${d.description}` : ''}`).join('\n') || '无'}\n当前：${this.findDatasource(conv.datasourceId)?.name ?? '未选择'}${conv.datasourcePinned ? '（已指定，发送「自动」恢复按问题自动选择）' : allowedDs.length > 1 ? '（按问题自动选择；发送「切换 <名称>」固定）' : ''}`); return }
+      if (verb === 'auto') { delete conv.datasourcePinned; convs[key] = conv; await this.save(); await reply(`已改为按问题自动选择数据源（可用：${allowedDs.map(d => d.name).join('、')}）。`); return }
+      if (verb === 'reset') { conv.sessionId = null; convs[key] = conv; await this.save(); await reply('已开始新对话。'); return }
+      if (!cmd.arg) { await reply(`用法：切换 <数据源名称>（或 switch <name>）。可用：${allowedDs.map(d => d.name).join('、') || '无'}`); return }
+      const want = cmd.arg.toLowerCase()
+      const target = allowedDs.find(d => d.name.toLowerCase() === want || d.id === cmd.arg) ?? allowedDs.filter(d => d.name.toLowerCase().includes(want) || (d.database ?? '').toLowerCase() === want).find((_, __, arr) => arr.length === 1)
+      if (!target) { await reply(`找不到可用的数据源「${cmd.arg}」。可用：${allowedDs.map(d => d.name).join('、')}`); return }
       conv.datasourceId = target.id
       conv.datasourcePinned = true
       if (conv.sessionId) { this.state.sessions[conv.sessionId] = target.id }
@@ -417,7 +441,8 @@ export const methods = {
     for (const d of list) {
       let tables = ''
       try {
-        const schema = this.schemas.get(d.id) ?? await this.readSchemaFile(d.id)
+        // Policy-filtered, like nl2sql_schema: hidden tables never reach the bot prompt.
+        const schema = await this.cachedVisibleSchema(d)
         if (schema?.tables?.length) {
           const rank = { fact: 0, dimension: 1, view: 2, other: 3, log: 4, config: 5 }
           const top = [...schema.tables].sort((a, b) => (rank[a.role] ?? 3) - (rank[b.role] ?? 3) || (b.rows ?? 0) - (a.rows ?? 0)).slice(0, 12)
@@ -475,7 +500,8 @@ export const methods = {
     }
     const waited = await controller.wait({ sessionId: sid, timeoutMs, maxMessages: 1 }, new AbortController().signal)
     const reason = waited?.reason ?? 'idle'
-    if (reason !== 'idle') { try { controller.cancel({ sessionId: sid }) } catch {} }
+    // cancel() is async on DSH controllers: a rejection must not become unhandled.
+    if (reason !== 'idle') void Promise.resolve().then(() => controller.cancel({ sessionId: sid })).catch(() => {})
     const answer = await this.collectTurnByRequest(sid, requestId)
     return { sessionId: sid, reason, ...answer }
   },

package/service/channels.js

@@ -19,6 +19,15 @@ import { UserError, newId, pick, requireText, upsert } from './shared.js'
 const LARK_LEGACY = ['domain', 'appId', 'receiveIdType', 'receiveId']
 const BOT_DEFAULTS = { enabled: false, allowP2p: true, sessionScope: 'chat' }
 
+/** Truncate to `max` UTF-16 units with an ellipsis, without splitting a surrogate pair. */
+export function clip(s, max) {
+  const str = String(s ?? '')
+  if (str.length <= max) return str
+  let cut = max - 1
+  if (/[\uD800-\uDBFF]/.test(str[cut - 1] ?? '')) cut--
+  return `${str.slice(0, cut)}…`
+}
+
 /** ChannelError (or legacy ImError) → UserError with the same code and message. */
 export function channelUserError(error, code) {
   if (!(error instanceof ChannelError || (error?.code && /^(im|channel)\//.test(error.code)))) return error
@@ -172,8 +181,10 @@ export const methods = {
     c.platform = adapter.id
     c.description = String(input.description ?? c.description ?? '').slice(0, 300)
     const { config: configIn, secrets: changed } = splitPayload(input, { ...kind, platform: adapter.id })
-    const stored = existing && !typeChanged ? await this.channelSecrets(existing) : {}
     const keys = secretKeys(kind)
+    // A type change keeps stored secrets whose key the new kind also has (e.g. webhookUrl, token).
+    const stored = existing ? await this.channelSecrets(existing) : {}
+    if (typeChanged) for (const k of Object.keys(stored)) if (!keys.includes(k)) delete stored[k]
     const merged = {}
     for (const k of keys) {
       if (Object.hasOwn(changed, k)) { if (changed[k] !== null) merged[k] = changed[k] } else if (stored[k] !== undefined) merged[k] = stored[k]
@@ -224,9 +235,19 @@ export const methods = {
     await this.stopBot(id)
     this.state.channels = this.state.channels.filter(x => x.id !== id)
     if (c.isDefault && this.state.channels[0]) this.state.channels[0].isDefault = true
-    for (const r of this.state.reports) if (r.channelId === id) delete r.channelId
+    // Drop every reference to the channel: alerts, schema notifications, bot conversations, reports.
+    for (const a of this.state.alerts ?? []) if (a.channelId === id) a.channelId = null
+    if (this.state.settings?.schemaNotifyChannel === id) this.state.settings.schemaNotifyChannel = null
+    for (const [k, conv] of Object.entries(this.state.botConversations ?? {})) if (conv?.channelId === id) delete this.state.botConversations[k]
+    const reports = (this.state.reports ?? []).filter(r => r.channelId === id)
+    for (const r of reports) delete r.channelId
     try { await this.ctx.get('credentials')?.deleteRecord(this.channelSecretKey(id)) } catch {}
     await this.save()
+    // Scheduled report prompts name the channel: rewrite them through the normal report save.
+    for (const r of reports) {
+      if (!r.scheduleId || !this.ctx.get('schedule')) continue
+      try { await this.saveReport({ ...r, channelId: undefined }) } catch (error) { console.warn('nl2sql: cannot update report %s after channel delete: %s', r.id, error.message) }
+    }
     return { deleted: id }
   },
 
@@ -311,10 +332,10 @@ export const methods = {
     const modes = chartModesOf(found.kind)
     const body = text ? String(text).slice(0, 20000) : ''
     const message = {
-      title: requireText(title || resolved[0]?.title || '数据问答', 'title', 100),
+      title: clip(String(title || resolved[0]?.title || '').trim() || '数据问答', 100),
       text: body,
       chartMode: modes.includes(chartMode) ? chartMode : modes.includes(c.chartMode) ? c.chartMode : modes[0],
-      results: resolved.map(r => pick(r, ['title', 'columns', 'rows', 'rowCount', 'truncated', 'chart', 'sql', 'filters', 'error', ...(found.kind.capabilities?.buttons ? ['actions'] : [])])),
+      results: resolved.map(r => ({ ...pick(r, ['title', 'columns', 'rows', 'rowCount', 'truncated', 'chart', 'sql', 'filters', 'error', ...(found.kind.capabilities?.buttons ? ['actions'] : [])]), ...(r.title ? { title: clip(String(r.title), 100) } : {}) })),
       includeSql: Boolean(includeSql),
       mentionAll: mentionAll ?? /<at\s+id\s*=\s*["']?all["']?\s*>/i.test(body),
       ...(template ? { template } : {}),
@@ -328,7 +349,7 @@ export const methods = {
     if (caps.files && wantAttach) {
       const idx = message.results.map((r, i) => (needsAttachment(r, shown) ? i : -1)).filter(i => i >= 0)
       if (idx.length) {
-        const sheets = idx.map(i => ({ name: (message.results[i].title || `结果${i + 1}`).slice(0, 31), columns: message.results[i].columns, rows: message.results[i].rows }))
+        const sheets = idx.map(i => ({ name: message.results[i].title || `结果${i + 1}`, columns: message.results[i].columns, rows: message.results[i].rows }))
         if (includeSql) sheets.push({ name: 'SQL', columns: [{ name: '结果', kind: 'text' }, { name: 'SQL', kind: 'text' }], rows: idx.map(i => [message.results[i].title ?? '', message.results[i].sql ?? '']) })
         const buffer = writeXlsx(sheets, { creator: 'DeepSeek Harness 数据问答' })
         const safe = String(message.title).replace(/[\\/:*?"<>|\s]+/g, '_').slice(0, 60) || '数据'

package/service/dashboards.js

@@ -296,8 +296,11 @@ export const methods = {
   // --------------------------------------------------------- auto overview
 
   /** Build overview widgets for one table from its column roles; runs each query and returns results. */
-  async tableOverview({ datasourceId, table, pin = false }) {
+  async tableOverview({ datasourceId, table, pin = false, actor, sessionId, source = 'ui' }) {
     const ds = this.requireDatasource(datasourceId)
+    // Agent / bot callers run as their actor: scope, row-level rules, viewer limits and audit attribution apply.
+    actor ??= sessionId ? this.actorFor(sessionId) : undefined
+    if (actor) this.assertAccess(actor, ds)
     const schema = await this.visibleSchema(ds)
     const t = schema.tables.find(x => x.name === table)
     if (!t) throw new UserError('nl2sql/unknown-table', `表不存在或无权访问：${table}`)
@@ -305,7 +308,7 @@ export const methods = {
     const results = []
     for (const plan of plans) {
       try {
-        const r = await this.runQuery({ sql: plan.sql, datasourceId: ds.id, title: plan.title, chart: plan.chart, source: 'ui', confirmScan: true })
+        const r = await this.runQuery({ sql: plan.sql, datasourceId: ds.id, sessionId, actor, title: plan.title, chart: plan.chart, source, confirmScan: true })
         results.push(r)
         if (pin) await this.pinWidget({ title: plan.title, sql: plan.sql, datasourceId: ds.id, chart: r.chart, width: plan.width })
       } catch (error) {

package/service/datasources.js

@@ -3,7 +3,7 @@
  * Mixed into the service prototype by index.js; `this` is the service.
  */
 import { DIALECTS, createDriver, describeError } from '../drivers.js'
-import { normalizePolicy, policyActive } from '../policy.js'
+import { filterSchema, normalizePolicy, policyActive } from '../policy.js'
 import { sanitizeTableName } from '../sqlite-driver.js'
 import { join } from 'node:path'
 import { CREDENTIAL_SCOPE, DEFAULT_SETTINGS, RETENTION_MAX, RETENTION_MIN, SUPPORTED_TYPES, UserError, clampInt, newId, oneOf, pick, requireText } from './shared.js'
@@ -308,14 +308,27 @@ export const methods = {
     return current
   },
 
-  async bindSession(sessionId, datasourceId) {
+  /**
+   * The already-analyzed schema (memory or file, no introspection) restricted by the
+   * access policy — the same view nl2sql_schema gives the model. Undefined when not analyzed.
+   */
+  async cachedVisibleSchema(ds) {
+    const schema = this.schemas.get(ds.id) ?? await this.readSchemaFile(ds.id)
+    return schema ? filterSchema(schema, this.policyFor(ds)) : undefined
+  },
+
+  /**
+   * Bind a session to a data source. The page RPC also makes it the page's current data source
+   * (default); agent tools (including bot sessions) pass setCurrent: false so they never move it.
+   */
+  async bindSession(sessionId, datasourceId, { setCurrent = true } = {}) {
     if (!sessionId) throw new UserError('nl2sql/invalid', '缺少会话')
     if (datasourceId === null) {
       delete this.state.sessions[sessionId]
     } else {
       const ds = this.requireDatasource(datasourceId)
       this.state.sessions[sessionId] = ds.id
-      this.state.current = ds.id
+      if (setCurrent) this.state.current = ds.id
     }
     await this.save()
     return this.sessionInfo(sessionId)

package/service/query.js

@@ -3,10 +3,11 @@
  * Mixed into the service prototype by index.js; `this` is the service.
  */
 import { aliasExpr, buildAttributionSql, buildBreakdownSql, buildDetailSql, buildMetricSql, computeAttribution, isAdditive, parseSelect, quoteName, rankDimensions } from '../analysis.js'
-import { applyFilters, normalizeFilters } from '../filters.js'
+import { FilterInjectionError, applyFilters, normalizeFilters } from '../filters.js'
 import { checkAccess, filterSchema, policyActive, referencedTables, resultColumnPolicy } from '../policy.js'
 import { admitSql, prepareSql, stripSql } from '../sql-guard.js'
 import { cacheKey } from '../store.js'
+import { createHash } from 'node:crypto'
 import { writeXlsx } from '../xlsx.js'
 import { RESULT_CACHE_LIMIT, SENSITIVE_NAME, UserError, columnKind, fillChartHint, fmtNum, maskValue, newId, normalizeChart, suggestChart, summarize } from './shared.js'
 
@@ -36,12 +37,19 @@ export const methods = {
     }
     let appliedFilters = []
     if (filters?.length) {
-      const f = applyFilters(admitted.sql, normalizeFilters(filters), { dialect, schema: filterSchema(fullSchema, policy) })
+      let f
+      try { f = applyFilters(admitted.sql, normalizeFilters(filters), { dialect, schema: filterSchema(fullSchema, policy) }) } catch (error) {
+        if (!(error instanceof FilterInjectionError)) throw error
+        this.audit({ ds, sessionId, source, sql: admitted.sql, title, status: 'blocked', error: 'filter injection refused', actor })
+        await this.save()
+        throw new UserError('nl2sql/invalid', '无法安全地为这条查询加上筛选条件，已拒绝。请改写查询（检查括号与注释）。')
+      }
       appliedFilters = f.applied
       if (f.applied.length) admitted = admitSql(f.sql, { dialect })
     }
     // Row-level security: every rule must be injected into the outer query; otherwise refuse.
     const rowRules = this.rowRulesFor(actor, ds)
+    const rowRulesSig = rowRules.length ? rowRulesSignature(rowRules) : undefined
     if (rowRules.length) {
       const schemaForRules = fullSchema ?? await this.analyzeSchema(ds.id)
       const refs = referencedTables(admitted.sql, schemaForRules.tables, { dialect })
@@ -55,7 +63,13 @@ export const methods = {
           await this.save()
           throw new UserError('nl2sql/row-policy', `当前成员受行级权限限制（${needed.map(r => `${r.column} ∈ {${r.values.join(', ')}}`).join('；')}），查询不能包含子查询、CTE 或 UNION，请改写为单层查询。`)
         }
-        const f = applyFilters(admitted.sql, needed.map(r => ({ kind: 'in', column: r.column, values: r.values })), { dialect, schema: schemaForRules })
+        let f
+        try { f = applyFilters(admitted.sql, needed.map(r => ({ kind: 'in', column: r.column, values: r.values })), { dialect, schema: schemaForRules }) } catch (error) {
+          if (!(error instanceof FilterInjectionError)) throw error
+          this.audit({ ds, sessionId, source, sql: admitted.sql, title, status: 'blocked', error: 'row-level policy: injection refused', actor })
+          await this.save()
+          throw new UserError('nl2sql/row-policy', `当前成员受行级权限限制（${needed.map(r => `${r.column} ∈ {${r.values.join(', ')}}`).join('；')}），这条查询无法安全地加上该过滤，已拒绝。`)
+        }
         const missing = needed.filter(r => !f.applied.some(a => a.column.toLowerCase() === r.column.toLowerCase()))
         // Every referenced table that has the column must be covered.
         const tablesWithCol = (col) => refs.filter(t => schemaForRules.tables.find(x => x.name === t)?.columns.some(c => c.name.toLowerCase() === col.toLowerCase()))
@@ -71,10 +85,10 @@ export const methods = {
     }
     // Query cache: dashboards, reruns and reports reuse a recent identical result.
     const ttlMs = (settings.cacheTtlSec ?? 0) * 1000
-    const key = ttlMs > 0 && useCache && this.store ? cacheKey([ds.id, admitted.sql, settings.maxRows, settings.maskSensitive, ds.policy ?? null, ds.updatedAt ?? 0, rowRules]) : null
+    const key = ttlMs > 0 && useCache && this.store ? cacheKey([ds.id, admitted.sql, settings.maxRows, settings.maskSensitive, ds.policy ?? null, ds.updatedAt ?? 0, rowRules, rowRulesSig ?? null]) : null
     if (key) {
       const hit = this.store.cached(key, ttlMs)
-      if (hit) {
+      if (hit && (hit.rowRulesSig ?? undefined) === rowRulesSig) {
         const result = { ...hit, resultId: newId('r'), title: title ? String(title).slice(0, 120) : hit.title, chart: normalizeChart(chart, hit.columns) ?? hit.chart, cached: true, cachedAt: hit.createdAt, createdAt: Date.now(), ...(appliedFilters.length ? { filters: appliedFilters, baseSql: sql } : {}) }
         this.remember(result)
         this.audit({ ds, sessionId, source, sql: admitted.sql, title, status: 'ok', rows: result.rowCount, elapsedMs: 0, resultId: result.resultId, cached: true, actor })
@@ -134,6 +148,7 @@ export const methods = {
       executedSql: prepared,
       ...(appliedFilters.length ? { filters: appliedFilters, baseSql: sql } : {}),
       ...(colPolicy.drop.size ? { hiddenColumns: colPolicy.drop.size } : {}),
+      ...(rowRulesSig ? { rowRulesSig } : {}),
       title: title ? String(title).slice(0, 120) : '',
       columns,
       rows: cells,
@@ -211,11 +226,17 @@ export const methods = {
     return this.lookupResult(resultId) ?? null
   },
 
-  /** A scoped actor may only use results of data sources in scope (and with its own row rules applied). */
+  /**
+   * A scoped actor may only use results of data sources in scope, and — when it
+   * has row rules — only results filtered by exactly its own rules (signature
+   * match; results without a signature, e.g. from older versions, are refused).
+   */
   assertResultAccess(r, actor) {
     if (!actor || actor.kind === 'owner') return
-    this.assertAccess(actor, this.requireDatasource(r.datasourceId))
-    if (this.rowRulesFor(actor, this.requireDatasource(r.datasourceId)).length && !r.filters?.some(f => f.rowPolicy)) throw new UserError('nl2sql/row-policy', '该结果未按你的行级权限过滤，不能使用，请重新查询')
+    const ds = this.requireDatasource(r.datasourceId)
+    this.assertAccess(actor, ds)
+    const rules = this.rowRulesFor(actor, ds)
+    if (rules.length && (!r.filters?.some(f => f.rowPolicy) || r.rowRulesSig !== rowRulesSignature(rules))) throw new UserError('nl2sql/row-policy', '该结果未按你的行级权限过滤，不能使用，请重新查询')
   },
 
   cacheStats() {
@@ -409,3 +430,10 @@ export const methods = {
     return { fileName: file, mediaType: 'application/vnd.openxmlformats-officedocument.spreadsheetml.sheet', base64: buf.toString('base64'), bytes: buf.length, sheets: sheets.length }
   },
 }
+
+/** Stable signature of a set of row rules (order of rules and of values does not matter). */
+export function rowRulesSignature(rules) {
+  const norm = rules.map(r => ({ column: String(r.column).toLowerCase(), values: [...new Set((r.values ?? []).map(String))].sort() }))
+    .sort((a, b) => (a.column + '\0' + a.values.join('\0')).localeCompare(b.column + '\0' + b.values.join('\0')))
+  return createHash('sha256').update(JSON.stringify(norm)).digest('hex').slice(0, 32)
+}

package/service/uploads.js

@@ -202,8 +202,11 @@ export const methods = {
       fileName = fileName || staged.fileName
       raw = staged.raw
     }
+    fileName = String(fileName ?? '').trim().slice(0, 255)
+    if (!raw && !fileName) throw new UserError('nl2sql/invalid', '缺少文件名')
     let ds = datasourceId ? this.requireDatasource(datasourceId) : undefined
-    if (!ds) ds = await this.saveDatasource({ type: 'sqlite', name: newName || String(fileName ?? '本地文件').replace(/\.[^.]+$/, '').slice(0, 60), description: `上传的文件：${fileName}` })
+    // The description is left empty for the user to fill in (it is shown to the model and in bot catalogs).
+    if (!ds) ds = await this.saveDatasource({ type: 'sqlite', name: newName || (fileName || 'uploaded-file').replace(/\.[^.]+$/, '').slice(0, 60) })
     if (ds.type !== 'sqlite') throw new UserError('nl2sql/invalid', '只能导入到「本地文件」类型的数据源')
     const ds2 = this.requireDatasource(ds.id)
     if (!raw) {

package/sql-guard.js

@@ -34,7 +34,7 @@ const FORBIDDEN = [
  * Per-dialect lexical rules. Getting these exactly right matters: if the guard
  * thinks a quote is escaped when the server does not (or vice versa), text the
  * server executes would be hidden inside a "string".
- *   mysql:          `-- ` needs whitespace after it, `#` comments, backslash escapes in '…' and "…", `…` identifiers
+ *   mysql:          `--` starts a comment only before a char <= 0x20 or end of input, `#` comments, backslash escapes in '…' and "…", `…` identifiers
  *   postgres:       any `--` comments, no `#`, backslash escapes only inside E'…', $tag$…$tag$ dollar quoting, "…" identifiers
  *   sqlite:         any `--` comments, no `#`, no backslash escapes, "…" / `…` / […] identifiers
  *   elasticsearch:  any `--` comments, no `#`, no backslash escapes, "…" identifiers
@@ -52,6 +52,35 @@ export function dialectRules(dialect = 'mysql') {
   }
 }
 
+/**
+ * MySQL starts a `--` comment only when the next character is a control
+ * character or space (code <= 0x20) or the input ends. Unicode spaces such as
+ * U+00A0 do NOT start a comment (JS `\s` would wrongly say they do).
+ */
+export function dashDashCommentFollower(ch) {
+  return ch === undefined || ch.charCodeAt(0) <= 0x20
+}
+
+/**
+ * Characters refused outside string literals / quoted identifiers / comments:
+ * C0 controls other than \t \n \r, DEL, and any non-ASCII whitespace or
+ * format character. Servers disagree with each other (and with JS) about these,
+ * so they are a lexer-confusion hazard and never needed in real SQL.
+ */
+const SUSPICIOUS_CODE_CHAR = /[\x00-\x08\x0B\x0C\x0E-\x1F\x7F\x80-\x9F\u00A0\u1680\u180E\u2000-\u200F\u2028-\u202F\u205F-\u2064\u3000\uFEFF]/
+
+/** First suspicious character in the code (non-literal) parts of the SQL, or undefined. */
+export function suspiciousCodeChar(sql, { dialect = 'mysql' } = {}) {
+  const { segments } = lexSql(sql, { dialect })
+  for (const s of segments) {
+    // Comments are checked too: a control character right after `--` is exactly the ambiguity being refused.
+    if (s.kind === 'string' || s.kind === 'ident') continue
+    const m = SUSPICIOUS_CODE_CHAR.exec(s.text)
+    if (m) return m[0]
+  }
+  return undefined
+}
+
 /**
  * Tokenize SQL into code / string / identifier / comment segments using the
  * dialect's lexical rules. Unterminated strings, identifiers and comments are
@@ -72,7 +101,7 @@ export function lexSql(sql, { dialect = 'mysql' } = {}) {
     const c = sql[i]
     const n = sql[i + 1]
     // -- comments
-    if (c === '-' && n === '-' && (!r.dashDashNeedsSpace || sql[i + 2] === undefined || /\s/.test(sql[i + 2]))) {
+    if (c === '-' && n === '-' && (!r.dashDashNeedsSpace || dashDashCommentFollower(sql[i + 2]))) {
       const start = i
       while (i < sql.length && sql[i] !== '\n') i++
       push('comment', start, i)
@@ -167,6 +196,8 @@ export function admitSql(input, { dialect = 'mysql' } = {}) {
   const { code, executableComment, unterminated } = stripSql(sql, { dialect })
   if (executableComment) throw new Error('MySQL executable comments (/*! ... */) are not allowed')
   if (unterminated) throw new Error('unterminated string, identifier or comment')
+  const bad = suspiciousCodeChar(sql, { dialect })
+  if (bad) throw new Error(`control or non-ASCII whitespace character U+${bad.charCodeAt(0).toString(16).toUpperCase().padStart(4, '0')} is only allowed inside string literals or quoted identifiers`)
   if (code.includes(';')) throw new Error('only one statement may be executed at a time')
   const head = (code.trim().match(/^\(*\s*([A-Za-z]+)/)?.[1] ?? '').toUpperCase()
   const allowed = ALLOWED_HEADS[dialect] ?? ALLOWED_HEADS.mysql

package/xlsx.js

@@ -208,6 +208,16 @@ const STYLES_XML = XML_HEAD +
 
 const SHEET_NAME_BAD_RE = /[\[\]:*?/\\]/g;
 
+/** First `max` UTF-16 units of `s` (Excel's sheet-name limit), never splitting a surrogate pair. */
+function utf16Prefix(s, max) {
+  let out = '';
+  for (const ch of s) {
+    if (out.length + ch.length > max) break;
+    out += ch;
+  }
+  return out;
+}
+
 /** Make Excel-safe, unique (case-insensitive), non-empty sheet names. */
 function sanitizeSheetNames(names) {
   const used = new Set(['history']); // "History" is reserved by Excel
@@ -215,11 +225,11 @@ function sanitizeSheetNames(names) {
     let base = stripIllegal(String(raw ?? '')).replace(SHEET_NAME_BAD_RE, '_').replace(/[\r\n\t]/g, ' ').trim();
     base = base.replace(/^'+|'+$/g, '').trim();
     if (!base) base = `Sheet${i + 1}`;
-    base = Array.from(base).slice(0, 31).join('');
+    base = utf16Prefix(base, 31);
     let name = base;
     for (let n = 2; used.has(name.toLowerCase()); n++) {
       const suffix = ` (${n})`;
-      name = Array.from(base).slice(0, 31 - suffix.length).join('').trimEnd() + suffix;
+      name = utf16Prefix(base, 31 - suffix.length).trimEnd() + suffix;
     }
     used.add(name.toLowerCase());
     return name;