index.ts 39 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103
  1. /**
  2. * Concrete session-query service with SQLite FTS5 over the live-preferred corpus.
  3. *
  4. * @module @deepseek-ai/dsh-session-query-sqlite
  5. */
  6. import { createHash, randomUUID } from 'node:crypto'
  7. import type { DatabaseSync } from 'node:sqlite'
  8. import { Context, Service, type Fiber } from '@deepseek-ai/cordis'
  9. import z from '@deepseek-ai/schemastery'
  10. import type { Session, SessionEvent, SessionHeader, SessionId } from '@deepseek-ai/dsh-session'
  11. import type SessionPersistence from '@deepseek-ai/dsh-session-persistence'
  12. import type {
  13. SessionPersistenceRevision,
  14. SessionPersistenceSnapshot,
  15. } from '@deepseek-ai/dsh-session-persistence'
  16. import SessionQueryEngine, {
  17. SESSION_QUERY_DEFAULT_PERSISTED_INSPECT_CONCURRENCY,
  18. SESSION_QUERY_READ_WINDOW_MAX,
  19. SessionQueryError,
  20. SessionSearchCursor,
  21. assertSessionHeadersCompatible,
  22. buildSessionEventSearchDocuments,
  23. } from '@deepseek-ai/dsh-session-query'
  24. import type {
  25. Config as SessionQueryConfig,
  26. SessionEventSearchDocument,
  27. SessionEventSearchHit,
  28. SessionEventSearchPage,
  29. SessionEventSearchRequest,
  30. SessionSearchExecContext,
  31. SessionSearchHit,
  32. SessionSearchCursor as SessionSearchCursorValue,
  33. SessionSearchPage,
  34. SessionSearchRequest,
  35. } from '@deepseek-ai/dsh-session-query'
  36. import {
  37. type JournalMode,
  38. openSearchDatabase,
  39. } from './schema.ts'
  40. import {
  41. type NormalizedEventRequest,
  42. type NormalizedSessionRequest,
  43. FTS_HIGHLIGHT_END,
  44. FTS_HIGHLIGHT_START,
  45. assertFts5OuterPredicateCount,
  46. assertPortableBindingCount,
  47. buildEventWhere,
  48. buildSessionWhere,
  49. makeSnippet,
  50. normalizeEventRequest,
  51. normalizeSessionRequest,
  52. quoteFtsData,
  53. requestFingerprint,
  54. sanitizeFtsText,
  55. SQLITE_MAX_PAGE_LIMIT,
  56. } from './query.ts'
  57. export {
  58. SESSION_QUERY_SQLITE_APPLICATION_ID,
  59. SESSION_QUERY_SQLITE_SCHEMA_VERSION,
  60. type JournalMode,
  61. } from './schema.ts'
  62. /** Boot-context slot for a launcher-owned absolute path to this process's derived query index. */
  63. export const SESSION_QUERY_SQLITE_PATH_KEY = 'launcherSessionQueryPath'
  64. declare module '@deepseek-ai/cordis' {
  65. interface Context {
  66. /** Launcher-owned absolute path to this process's disposable derived query index. */
  67. launcherSessionQueryPath?: string
  68. }
  69. }
  70. /** Default result page size. */
  71. export const SESSION_QUERY_SQLITE_DEFAULT_LIMIT = 20
  72. /** Maximum accepted result page size. */
  73. export const SESSION_QUERY_SQLITE_MAX_LIMIT = 100
  74. /** Default maximum snippet length in Unicode code points. */
  75. export const SESSION_QUERY_SQLITE_SNIPPET_CHARS = 240
  76. // One transient source change gets a retry; repeated churn fails rather than monopolizing the queue.
  77. const STABLE_OBSERVATION_ATTEMPTS = 2
  78. /** SQLite module/handle opening phase; `never` disables full-text search entirely. */
  79. export type OpenAt = 'startup' | 'first-search' | 'never'
  80. /** Combined session-query configuration backed by SQLite full-text search. */
  81. export interface Config extends SessionQueryConfig {
  82. /**
  83. * Dedicated derived-index path; `:memory:` is supported for ephemeral
  84. * indexes. Missing directories and database files are created owner-only on
  85. * POSIX filesystems; existing modes are preserved.
  86. */
  87. path: string
  88. /**
  89. * Open the SQLite module and handle at service activation or the first
  90. * search, or `never` to disable full-text search: the inherited exact
  91. * reads, filters, and traces stay available, while `searchSessions` and
  92. * `searchEvents` fail with `SESSION_QUERY_SEARCH_DISABLED` and SQLite is
  93. * never imported or opened. Defaults to `startup`.
  94. */
  95. openAt?: OpenAt
  96. /** SQLite journal mode. Defaults to `wal`. */
  97. journalMode?: JournalMode
  98. /** Page size when a request omits `limit`. At most `Number.MAX_SAFE_INTEGER - 1`; defaults to 20. */
  99. defaultLimit?: number
  100. /** Largest accepted page size. At most `Number.MAX_SAFE_INTEGER - 1`; defaults to 100. */
  101. maxLimit?: number
  102. /** Maximum snippet length in Unicode code points. Defaults to 240. */
  103. snippetChars?: number
  104. /** Maximum concurrent persisted-log inspections in one inherited batch read. Defaults to 4. */
  105. persistedInspectConcurrency?: number
  106. }
  107. interface ResolvedConfig {
  108. path: string
  109. openAt: OpenAt
  110. journalMode: JournalMode
  111. defaultLimit: number
  112. maxLimit: number
  113. snippetChars: number
  114. readWindowMax: number
  115. persistedInspectConcurrency: number
  116. }
  117. interface ObservedSession {
  118. header: SessionHeader
  119. documents: SessionEventSearchDocument[]
  120. fingerprint: string
  121. }
  122. interface ObservedPersistedSession {
  123. header: SessionHeader
  124. revision: SessionPersistenceRevision
  125. loaded?: ObservedSession
  126. }
  127. interface PersistenceBinding {
  128. readonly identity: symbol
  129. readonly service?: SessionPersistence
  130. }
  131. interface Observation {
  132. persistenceBinding: PersistenceBinding
  133. persisted: Map<SessionId, ObservedPersistedSession>
  134. live: Map<SessionId, ObservedSession>
  135. }
  136. interface IndexedPersistedRow {
  137. id: string
  138. revision: string
  139. generation: number
  140. }
  141. interface IndexedLiveRow {
  142. id: string
  143. fingerprint: string
  144. persisted: number
  145. generation: number
  146. }
  147. interface SessionHeaderRow {
  148. session_id: string
  149. version: number
  150. created_at: number
  151. cwd: string | null
  152. parent_session: string | null
  153. seed_length: number | null
  154. delegation_depth: number | null
  155. agent_preset: string | null
  156. }
  157. interface SearchRow extends SessionHeaderRow {
  158. live: number
  159. persisted: number
  160. seq: number
  161. type: string
  162. time: number
  163. surface: string
  164. marked_text: string
  165. match_count: number
  166. document_length: number
  167. }
  168. interface CursorPayload {
  169. version: 1
  170. instance: string
  171. scope: 'sessions' | 'events'
  172. fingerprint: string
  173. generation: string
  174. offset: number
  175. }
  176. /** Concrete SQLite owner of the combined `ctx.sessionQuery` service. */
  177. export class SqliteSessionQueryEngine extends SessionQueryEngine {
  178. static override inject = ['sessions']
  179. static Config: z<Config> = z.object({
  180. path: z.string().required(),
  181. openAt: z.union(['startup', 'first-search', 'never'] as const).default('startup'),
  182. journalMode: z.union(['wal', 'delete', 'truncate', 'persist'] as const).default('wal'),
  183. defaultLimit: z.number().step(1).min(1).max(SQLITE_MAX_PAGE_LIMIT).default(SESSION_QUERY_SQLITE_DEFAULT_LIMIT),
  184. maxLimit: z.number().step(1).min(1).max(SQLITE_MAX_PAGE_LIMIT).default(SESSION_QUERY_SQLITE_MAX_LIMIT),
  185. snippetChars: z.number().step(1).min(1).default(SESSION_QUERY_SQLITE_SNIPPET_CHARS),
  186. readWindowMax: z.number().step(1).min(0).default(SESSION_QUERY_READ_WINDOW_MAX),
  187. persistedInspectConcurrency: z.number()
  188. .step(1)
  189. .min(1)
  190. .max(Number.MAX_SAFE_INTEGER)
  191. .default(SESSION_QUERY_DEFAULT_PERSISTED_INSPECT_CONCURRENCY),
  192. })
  193. /** Validated and defaulted backend configuration. */
  194. readonly config: ResolvedConfig
  195. private readonly _instance = randomUUID()
  196. private _ready: Promise<void> | undefined
  197. private _db: DatabaseSync | undefined
  198. private _persistenceBinding: PersistenceBinding = { identity: Symbol() }
  199. private _lastPersistenceIdentity: symbol | undefined
  200. private _persistenceEpoch = 0
  201. private _globalGeneration = 0
  202. private _localGeneration = 0
  203. private _tail: Promise<void> = Promise.resolve()
  204. private _closed = false
  205. private _closePromise: Promise<void> | undefined
  206. private readonly _optionalPersistenceFiber: Fiber
  207. constructor(ctx: Context, config: Config) {
  208. // The assignment expression resolves before the base constructor can
  209. // register `ctx.sessionQuery`; keep that same validated value afterward.
  210. super(ctx, config = resolveConfig(config))
  211. this.config = config as ResolvedConfig
  212. this._optionalPersistenceFiber = ctx.inject(['sessionPersistence'], (childCtx: Context) => {
  213. const service = childCtx.sessionPersistence
  214. const binding = { identity: Symbol(), service }
  215. this._persistenceBinding = binding
  216. childCtx.effect(() => () => {
  217. /* v8 ignore next -- a stale optional-service disposer cannot clear a replacement */
  218. if (this._persistenceBinding !== binding) return
  219. this._persistenceBinding = { identity: Symbol() }
  220. }, 'sessionQuerySqlite.persistenceBinding')
  221. })
  222. ctx.effect(() => {
  223. return () => this._optionalPersistenceFiber.dispose()
  224. }, 'sessionQuerySqlite.optionalPersistence')
  225. ctx.effect(() => async () => this.close(), 'sessionQuerySqlite.close')
  226. }
  227. /** Open eagerly only when activation owns the configured readiness boundary. */
  228. protected async [Service.init](): Promise<void> {
  229. if (this.config.openAt === 'startup') await this._ensureReady(undefined)
  230. }
  231. override async searchSessions(
  232. request: SessionSearchRequest,
  233. exec?: SessionSearchExecContext,
  234. ): Promise<SessionSearchPage<SessionSearchHit>> {
  235. this._assertSearchEnabled()
  236. const normalized = normalizeSessionRequest(request, this.config)
  237. const signal = exec?.signal
  238. return this._serialized(signal, async () => {
  239. await this._ensureReady(signal)
  240. const persistenceBinding = await this._reconcile(signal)
  241. assertNotAborted(signal)
  242. const generation = String(this._globalGeneration)
  243. const fingerprint = requestFingerprint(normalized)
  244. const offset = normalized.cursor === undefined
  245. ? 0
  246. : decodeCursor(normalized.cursor, this._instance, 'sessions', fingerprint, generation)
  247. const rows = this._querySessions(normalized, offset, persistenceBinding)
  248. return page(rows, normalized.limit, row => this._sessionHit(row), cursorOffset => encodeCursor({
  249. version: 1,
  250. instance: this._instance,
  251. scope: 'sessions',
  252. fingerprint,
  253. generation,
  254. offset: cursorOffset,
  255. }), offset)
  256. })
  257. }
  258. override async searchEvents(
  259. request: SessionEventSearchRequest,
  260. exec?: SessionSearchExecContext,
  261. ): Promise<SessionEventSearchPage> {
  262. this._assertSearchEnabled()
  263. const normalized = normalizeEventRequest(request, this.config)
  264. const signal = exec?.signal
  265. return this._serialized(signal, async () => {
  266. await this._ensureReady(signal)
  267. const persistenceBinding = await this._reconcile(signal)
  268. assertNotAborted(signal)
  269. const target = this._targetObservation(normalized.sessionId, persistenceBinding)
  270. const fingerprint = requestFingerprint(normalized)
  271. const offset = normalized.cursor === undefined
  272. ? 0
  273. : decodeCursor(normalized.cursor, this._instance, 'events', fingerprint, target.generation)
  274. const rows = this._queryEvents(normalized, offset, persistenceBinding)
  275. return {
  276. session: target.header,
  277. ...page(rows, normalized.limit, row => this._eventHit(row), cursorOffset => encodeCursor({
  278. version: 1,
  279. instance: this._instance,
  280. scope: 'events',
  281. fingerprint,
  282. generation: target.generation,
  283. offset: cursorOffset,
  284. }), offset),
  285. }
  286. })
  287. }
  288. /** Close the database after every accepted operation reaches quiescence. */
  289. close(): Promise<void> {
  290. this._closePromise ??= this._close()
  291. return this._closePromise
  292. }
  293. /**
  294. * Refuse full-text calls under `openAt: 'never'` before any request
  295. * normalization or SQLite work, so a disabled deployment never imports
  296. * node:sqlite, opens the index, or observes sources.
  297. */
  298. private _assertSearchEnabled(): void {
  299. if (this.config.openAt !== 'never') return
  300. throw new SessionQueryError(
  301. 'session search is disabled: this deployment configures the session-query index with openAt "never"',
  302. 'SESSION_QUERY_SEARCH_DISABLED',
  303. )
  304. }
  305. private async _close(): Promise<void> {
  306. this._closed = true
  307. await this._tail
  308. if (this._ready !== undefined) {
  309. try {
  310. await this._ready
  311. } catch {
  312. // Opening already closed a partially-created handle; disposal only waits.
  313. }
  314. }
  315. this._db?.close()
  316. this._db = undefined
  317. }
  318. private async _open(): Promise<void> {
  319. this._db = await openSearchDatabase(this.config.path, this.config.journalMode)
  320. const state = this._db.prepare(
  321. 'SELECT global_generation FROM search_state WHERE singleton = 1',
  322. ).get() as { global_generation: number }
  323. this._globalGeneration = state.global_generation
  324. this._localGeneration = state.global_generation
  325. }
  326. private async _ensureReady(signal: AbortSignal | undefined): Promise<void> {
  327. this._ready ??= this._open()
  328. try {
  329. await waitWithAbort(this._ready, signal)
  330. } catch (error: unknown) {
  331. if (isAbort(error)) throw error
  332. throw new SessionQueryError(
  333. `session-search SQLite index failed to open: ${errorMessage(error)}`,
  334. 'SESSION_QUERY_INDEX_FAILED',
  335. { cause: error },
  336. )
  337. }
  338. }
  339. private async _serialized<T>(signal: AbortSignal | undefined, operation: () => Promise<T>): Promise<T> {
  340. if (this._isClosed()) throw indexClosed()
  341. let release!: () => void
  342. const gate = new Promise<void>((resolve) => { release = resolve })
  343. const prior = this._tail
  344. this._tail = prior.then(() => gate)
  345. try {
  346. await waitWithAbort(prior, signal)
  347. } catch (error: unknown) {
  348. release()
  349. throw error
  350. }
  351. if (this._isClosed()) {
  352. release()
  353. throw indexClosed()
  354. }
  355. try {
  356. assertNotAborted(signal)
  357. return await operation()
  358. } finally {
  359. release()
  360. }
  361. }
  362. private async _reconcile(signal: AbortSignal | undefined): Promise<PersistenceBinding> {
  363. assertNotAborted(signal)
  364. const db = this._requireDb()
  365. const persistedRows = db.prepare(
  366. 'SELECT id, revision, generation FROM persisted_sessions',
  367. ).all() as unknown as IndexedPersistedRow[]
  368. const liveRows = db.prepare(
  369. 'SELECT id, fingerprint, persisted, generation FROM temp.live_sessions',
  370. ).all() as unknown as IndexedLiveRow[]
  371. const persistedById = new Map(persistedRows.map(row => [row.id as SessionId, row]))
  372. const liveById = new Map(liveRows.map(row => [row.id as SessionId, row]))
  373. const observation = await this._observeStable(persistedById, signal)
  374. assertNotAborted(signal)
  375. const persistentChanges = observation.persistenceBinding.service === undefined
  376. ? []
  377. : [...observation.persisted.values()].filter(entry => entry.loaded !== undefined)
  378. const persistentDeletes = observation.persistenceBinding.service === undefined
  379. ? []
  380. : persistedRows.filter(row => !observation.persisted.has(row.id as SessionId))
  381. const liveChanges = [...observation.live.values()].filter((entry) => {
  382. const indexed = liveById.get(entry.header.id)
  383. const persisted = observation.persisted.has(entry.header.id) ? 1 : 0
  384. return indexed?.fingerprint !== entry.fingerprint || indexed.persisted !== persisted
  385. })
  386. const liveDeletes = liveRows.filter(row => !observation.live.has(row.id as SessionId))
  387. const pointerChanged = this._lastPersistenceIdentity !== undefined
  388. && this._lastPersistenceIdentity !== observation.persistenceBinding.identity
  389. const hasWrites = persistentChanges.length > 0
  390. || persistentDeletes.length > 0
  391. || liveChanges.length > 0
  392. || liveDeletes.length > 0
  393. let nextMainGeneration = this._mainGeneration()
  394. let nextLocalGeneration = this._localGeneration
  395. if (persistentChanges.length > 0 || persistentDeletes.length > 0) nextMainGeneration += 1
  396. const liveReplacements = liveChanges.map((entry) => {
  397. nextLocalGeneration = Math.max(nextLocalGeneration, nextMainGeneration) + 1
  398. return {
  399. entry,
  400. generation: nextLocalGeneration,
  401. persisted: observation.persisted.has(entry.header.id),
  402. }
  403. })
  404. if (hasWrites) {
  405. let began = false
  406. try {
  407. db.exec('BEGIN IMMEDIATE')
  408. began = true
  409. for (const row of persistentDeletes) this._deleteSession('persisted', row.id as SessionId)
  410. for (const entry of persistentChanges) {
  411. /* v8 ignore next -- observation loads every entry whose revision differs */
  412. if (entry.loaded === undefined) throw new Error(`missing loaded revision for session "${entry.header.id}"`)
  413. this._replacePersistedSession(entry.loaded, entry.revision, nextMainGeneration)
  414. }
  415. if (persistentChanges.length > 0 || persistentDeletes.length > 0) {
  416. db.prepare('UPDATE search_state SET global_generation = ? WHERE singleton = 1').run(nextMainGeneration)
  417. }
  418. for (const row of liveDeletes) this._deleteSession('live', row.id as SessionId)
  419. for (const { entry, generation, persisted } of liveReplacements) {
  420. this._replaceLiveSession(entry, generation, persisted)
  421. }
  422. db.exec('COMMIT')
  423. } catch (error: unknown) {
  424. /* v8 ignore next -- a BEGIN failure has no transaction to roll back; the common wrapper still reports it. */
  425. if (began) {
  426. /* v8 ignore next 5 -- ROLLBACK failure requires a SQLite double fault; the original failure remains actionable. */
  427. try {
  428. db.exec('ROLLBACK')
  429. } catch {
  430. // The original SQLite failure remains the actionable cause.
  431. }
  432. }
  433. throw new SessionQueryError(
  434. `session-search reconciliation failed: ${errorMessage(error)}`,
  435. 'SESSION_QUERY_INDEX_FAILED',
  436. { cause: error },
  437. )
  438. }
  439. }
  440. if (hasWrites || pointerChanged) this._globalGeneration += 1
  441. if (pointerChanged) this._persistenceEpoch += 1
  442. this._localGeneration = nextLocalGeneration
  443. this._lastPersistenceIdentity = observation.persistenceBinding.identity
  444. return observation.persistenceBinding
  445. }
  446. private async _observeStable(
  447. indexed: ReadonlyMap<SessionId, IndexedPersistedRow>,
  448. signal: AbortSignal | undefined,
  449. ): Promise<Observation> {
  450. for (let attempt = 0; attempt < STABLE_OBSERVATION_ATTEMPTS; attempt += 1) {
  451. assertNotAborted(signal)
  452. const persistenceBinding = this._persistenceBinding
  453. const persistence = persistenceBinding.service
  454. const initiallyLive = new Set(this.ctx.sessions.list().map(session => session.id))
  455. let persisted = new Map<SessionId, ObservedPersistedSession>()
  456. if (persistence !== undefined) {
  457. try {
  458. const canReuseIndexed = this._lastPersistenceIdentity === undefined
  459. || this._lastPersistenceIdentity === persistenceBinding.identity
  460. const before = await persistence.listSnapshots(signal)
  461. assertNotAborted(signal)
  462. persisted = materializePersistenceSnapshots(before)
  463. for (const entry of persisted.values()) {
  464. if (canReuseIndexed && indexed.get(entry.header.id)?.revision === entry.revision) continue
  465. // Skip work already shadowed by a live owner. `inspect()` is
  466. // non-mutating, so an owner attaching after this check cannot cause
  467. // crash-repair side effects; the live-membership retry below makes
  468. // the returned observation live-preferred.
  469. if (initiallyLive.has(entry.header.id) || this.ctx.sessions.get(entry.header.id) !== undefined) continue
  470. assertNotAborted(signal)
  471. const loaded = await persistence.inspect(entry.header.id, signal)
  472. assertNotAborted(signal)
  473. assertSessionHeadersCompatible(entry.header, loaded.meta)
  474. entry.loaded = observeSession(loaded.meta, loaded.events)
  475. }
  476. assertNotAborted(signal)
  477. const afterSnapshots = await persistence.listSnapshots(signal)
  478. assertNotAborted(signal)
  479. const after = materializePersistenceSnapshots(afterSnapshots)
  480. if (!samePersistenceSnapshots(persisted, after)) continue
  481. if (this._persistenceBinding !== persistenceBinding) continue
  482. } catch (error: unknown) {
  483. if (isAbort(error) || signal?.aborted) {
  484. throw new SessionQueryError('session-search aborted', 'SESSION_QUERY_ABORTED', {
  485. cause: error,
  486. })
  487. }
  488. if (this._persistenceBinding !== persistenceBinding) continue
  489. if (error instanceof SessionQueryError) throw error
  490. throw new SessionQueryError(
  491. `session-search persistence observation failed: ${errorMessage(error)}`,
  492. 'SESSION_QUERY_PERSISTENCE_FAILED',
  493. { cause: error },
  494. )
  495. }
  496. }
  497. const live = new Map<SessionId, ObservedSession>()
  498. for (const session of this.ctx.sessions.list()) {
  499. const observed = observeLive(session)
  500. const durable = persisted.get(session.id)
  501. if (durable !== undefined) assertSessionHeadersCompatible(observed.header, durable.header)
  502. live.set(session.id, observed)
  503. }
  504. if (!sameSessionIds(initiallyLive, live)) continue
  505. return { persistenceBinding, persisted, live }
  506. }
  507. throw new SessionQueryError(
  508. 'session-search persistence observation did not stabilize after one retry',
  509. 'SESSION_QUERY_PERSISTENCE_FAILED',
  510. )
  511. }
  512. private _mainGeneration(): number {
  513. const row = this._requireDb().prepare(
  514. 'SELECT global_generation FROM search_state WHERE singleton = 1',
  515. ).get() as { global_generation: number }
  516. return row.global_generation
  517. }
  518. private _deleteSession(source: 'persisted' | 'live', id: SessionId): void {
  519. const db = this._requireDb()
  520. if (source === 'persisted') {
  521. db.prepare('DELETE FROM persisted_docs WHERE session_id = ?').run(id)
  522. db.prepare('DELETE FROM persisted_sessions WHERE id = ?').run(id)
  523. } else {
  524. db.prepare('DELETE FROM temp.live_docs WHERE session_id = ?').run(id)
  525. db.prepare('DELETE FROM temp.live_sessions WHERE id = ?').run(id)
  526. }
  527. }
  528. private _replacePersistedSession(
  529. entry: ObservedSession,
  530. revision: SessionPersistenceRevision,
  531. generation: number,
  532. ): void {
  533. this._deleteSession('persisted', entry.header.id)
  534. const db = this._requireDb()
  535. db.prepare(`
  536. INSERT INTO persisted_sessions
  537. (id, version, created_at, cwd, parent_session, seed_length, delegation_depth, agent_preset, revision, generation)
  538. VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  539. `).run(
  540. ...headerBindings(entry.header),
  541. revision,
  542. generation,
  543. )
  544. const insert = db.prepare(`
  545. INSERT INTO persisted_docs (text, session_id, seq, type, time, surface, codepoint_length)
  546. VALUES (?, ?, ?, ?, ?, ?, ?)
  547. `)
  548. for (const document of entry.documents) {
  549. const text = sanitizeFtsText(document.text)
  550. insert.run(
  551. text,
  552. document.sessionId,
  553. document.seq,
  554. document.type,
  555. document.time,
  556. document.surface,
  557. Array.from(text).length,
  558. )
  559. }
  560. }
  561. private _replaceLiveSession(entry: ObservedSession, generation: number, persisted: boolean): void {
  562. this._deleteSession('live', entry.header.id)
  563. const db = this._requireDb()
  564. db.prepare(`
  565. INSERT INTO temp.live_sessions
  566. (id, version, created_at, cwd, parent_session, seed_length, delegation_depth, agent_preset, fingerprint, persisted, generation)
  567. VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
  568. `).run(
  569. ...headerBindings(entry.header),
  570. entry.fingerprint,
  571. persisted ? 1 : 0,
  572. generation,
  573. )
  574. const insert = db.prepare(`
  575. INSERT INTO temp.live_docs (text, session_id, seq, type, time, surface, codepoint_length)
  576. VALUES (?, ?, ?, ?, ?, ?, ?)
  577. `)
  578. for (const document of entry.documents) {
  579. const text = sanitizeFtsText(document.text)
  580. insert.run(
  581. text,
  582. document.sessionId,
  583. document.seq,
  584. document.type,
  585. document.time,
  586. document.surface,
  587. Array.from(text).length,
  588. )
  589. }
  590. }
  591. private _querySessions(
  592. request: NormalizedSessionRequest,
  593. offset: number,
  594. persistenceBinding: PersistenceBinding,
  595. ): SearchRow[] {
  596. const selected = selectedDocumentsSql()
  597. const sessionWhere = buildSessionWhere(request.sessionFilters)
  598. const eventWhere = buildEventWhere(request.eventFilters)
  599. assertFts5OuterPredicateCount(sessionWhere.predicateCount + eventWhere.predicateCount)
  600. const where = [sessionWhere.sql, eventWhere.sql].filter(Boolean).join(' AND ')
  601. const bindings = [
  602. ...selectedDocumentsParams(request.query, persistenceBinding.service !== undefined),
  603. ...sessionWhere.params,
  604. ...eventWhere.params,
  605. request.limit + 1,
  606. offset,
  607. ]
  608. assertPortableBindingCount(bindings.length)
  609. // The browser fixture mirrors these rank keys in
  610. // `packages/client/connection/src/client/fixture.ts`; update both together.
  611. return this._requireDb().prepare(`
  612. ${selected.sql},
  613. filtered AS (
  614. SELECT * FROM matched ${where.length === 0 ? '' : `WHERE ${where}`}
  615. ),
  616. ranked AS (
  617. SELECT *, ROW_NUMBER() OVER (
  618. PARTITION BY session_id
  619. ORDER BY match_count DESC, document_length ASC, time DESC, seq DESC
  620. ) AS event_rank
  621. FROM filtered
  622. )
  623. SELECT * FROM ranked
  624. WHERE event_rank = 1
  625. ORDER BY match_count DESC, document_length ASC, time DESC, session_id ASC, seq DESC
  626. LIMIT ? OFFSET ?
  627. `).all(...bindings) as unknown as SearchRow[]
  628. }
  629. private _queryEvents(
  630. request: NormalizedEventRequest,
  631. offset: number,
  632. persistenceBinding: PersistenceBinding,
  633. ): SearchRow[] {
  634. const selected = selectedDocumentsSql()
  635. const eventWhere = buildEventWhere(request.filters)
  636. assertFts5OuterPredicateCount(1 + eventWhere.predicateCount)
  637. const where = ['session_id = ?', eventWhere.sql].filter(Boolean).join(' AND ')
  638. const bindings = [
  639. ...selectedDocumentsParams(request.query, persistenceBinding.service !== undefined),
  640. request.sessionId,
  641. ...eventWhere.params,
  642. request.limit + 1,
  643. offset,
  644. ]
  645. assertPortableBindingCount(bindings.length)
  646. return this._requireDb().prepare(`
  647. ${selected.sql}
  648. SELECT * FROM matched
  649. WHERE ${where}
  650. ORDER BY match_count DESC, document_length ASC, time DESC, seq DESC
  651. LIMIT ? OFFSET ?
  652. `).all(...bindings) as unknown as SearchRow[]
  653. }
  654. private _targetObservation(
  655. sessionId: SessionId,
  656. persistenceBinding: PersistenceBinding,
  657. ): { header: SessionHeader; generation: string } {
  658. const db = this._requireDb()
  659. const live = db.prepare(
  660. `SELECT
  661. id AS session_id, version, created_at, cwd, parent_session, seed_length, delegation_depth, agent_preset, generation
  662. FROM temp.live_sessions
  663. WHERE id = ?`,
  664. ).get(sessionId) as (SessionHeaderRow & { generation: number }) | undefined
  665. if (live !== undefined) {
  666. return { header: rowHeader(live), generation: `live:${live.generation}` }
  667. }
  668. if (persistenceBinding.service !== undefined) {
  669. const persisted = db.prepare(
  670. `SELECT
  671. id AS session_id, version, created_at, cwd, parent_session, seed_length, delegation_depth, agent_preset, generation
  672. FROM persisted_sessions
  673. WHERE id = ?`,
  674. ).get(sessionId) as (SessionHeaderRow & { generation: number }) | undefined
  675. if (persisted !== undefined) {
  676. return {
  677. header: rowHeader(persisted),
  678. generation: `persisted:${this._persistenceEpoch}:${persisted.generation}`,
  679. }
  680. }
  681. }
  682. throw new SessionQueryError(
  683. `session "${sessionId}" not found`,
  684. 'SESSION_QUERY_SESSION_NOT_FOUND',
  685. )
  686. }
  687. private _sessionHit(row: SearchRow): SessionSearchHit {
  688. return {
  689. header: rowHeader(row),
  690. live: row.live === 1,
  691. persisted: row.persisted === 1,
  692. bestMatch: this._eventHit(row),
  693. }
  694. }
  695. private _eventHit(row: SearchRow): SessionEventSearchHit {
  696. return {
  697. sessionId: row.session_id as SessionId,
  698. seq: row.seq,
  699. type: row.type as SessionEventSearchHit['type'],
  700. time: row.time,
  701. surface: row.surface as SessionEventSearchHit['surface'],
  702. snippet: makeSnippet(row.marked_text, this.config.snippetChars),
  703. }
  704. }
  705. private _requireDb(): DatabaseSync {
  706. /* v8 ignore next -- callers await `_ready`; this guards lifecycle misuse */
  707. if (this._db === undefined) throw indexClosed()
  708. return this._db
  709. }
  710. private _isClosed(): boolean {
  711. return this._closed
  712. }
  713. }
  714. /**
  715. * The header columns both session upserts bind, in the order their INSERT
  716. * lists them. The two statements differ only in what they append after these.
  717. * @param header - the session header being written.
  718. * @returns one bound value per header column.
  719. */
  720. function headerBindings(header: SessionHeader): (string | number | null)[] {
  721. return [
  722. header.id,
  723. header.version,
  724. header.createdAt,
  725. header.cwd ?? null,
  726. header.parentSession ?? null,
  727. header.seedLength ?? null,
  728. header.delegationDepth ?? null,
  729. header.agentPreset ?? null,
  730. ]
  731. }
  732. function selectedDocumentsSql(): { sql: string } {
  733. return {
  734. sql: `WITH candidates AS (
  735. SELECT
  736. pd.session_id AS session_id,
  737. ps.version AS version,
  738. ps.created_at AS created_at,
  739. ps.cwd AS cwd,
  740. ps.parent_session AS parent_session,
  741. ps.seed_length AS seed_length,
  742. ps.delegation_depth AS delegation_depth,
  743. ps.agent_preset AS agent_preset,
  744. 0 AS live,
  745. 1 AS persisted,
  746. CAST(pd.seq AS INTEGER) AS seq,
  747. pd.type AS type,
  748. CAST(pd.time AS INTEGER) AS time,
  749. pd.surface AS surface,
  750. highlight(persisted_docs, 0, ?, ?) AS marked_text,
  751. CAST(pd.codepoint_length AS INTEGER) AS document_length
  752. FROM persisted_docs AS pd
  753. JOIN persisted_sessions AS ps ON ps.id = pd.session_id
  754. WHERE persisted_docs MATCH ?
  755. AND ? = 1
  756. AND NOT EXISTS (SELECT 1 FROM temp.live_sessions AS ls WHERE ls.id = pd.session_id)
  757. UNION ALL
  758. SELECT
  759. ld.session_id AS session_id,
  760. ls.version AS version,
  761. ls.created_at AS created_at,
  762. ls.cwd AS cwd,
  763. ls.parent_session AS parent_session,
  764. ls.seed_length AS seed_length,
  765. ls.delegation_depth AS delegation_depth,
  766. ls.agent_preset AS agent_preset,
  767. 1 AS live,
  768. CASE WHEN ? = 1 THEN ls.persisted ELSE 0 END AS persisted,
  769. CAST(ld.seq AS INTEGER) AS seq,
  770. ld.type AS type,
  771. CAST(ld.time AS INTEGER) AS time,
  772. ld.surface AS surface,
  773. highlight(live_docs, 0, ?, ?) AS marked_text,
  774. CAST(ld.codepoint_length AS INTEGER) AS document_length
  775. FROM temp.live_docs AS ld
  776. JOIN temp.live_sessions AS ls ON ls.id = ld.session_id
  777. WHERE live_docs MATCH ?
  778. ), matched AS (
  779. SELECT *,
  780. (
  781. length(CAST(marked_text AS BLOB))
  782. - length(CAST(replace(marked_text, ?, '') AS BLOB))
  783. ) / ? AS match_count
  784. FROM candidates
  785. )`,
  786. }
  787. }
  788. function selectedDocumentsParams(query: string, persistenceVisible: boolean): Array<string | number> {
  789. const expression = quoteFtsData(query)
  790. const visible = persistenceVisible ? 1 : 0
  791. return [
  792. FTS_HIGHLIGHT_START,
  793. FTS_HIGHLIGHT_END,
  794. expression,
  795. visible,
  796. visible,
  797. FTS_HIGHLIGHT_START,
  798. FTS_HIGHLIGHT_END,
  799. expression,
  800. FTS_HIGHLIGHT_START,
  801. Buffer.byteLength(FTS_HIGHLIGHT_START, 'utf8'),
  802. ]
  803. }
  804. function observeLive(session: Session): ObservedSession {
  805. return observeSession(session.header, session.events)
  806. }
  807. function observeSession(header: SessionHeader, events: readonly SessionEvent[]): ObservedSession {
  808. const detachedHeader = structuredClone(header)
  809. const detachedEvents = events.map(event => structuredClone(event))
  810. return {
  811. header: detachedHeader,
  812. documents: buildSessionEventSearchDocuments(detachedHeader.id, detachedEvents),
  813. fingerprint: createHash('sha256')
  814. .update(JSON.stringify({ header: detachedHeader, events: detachedEvents }))
  815. .digest('base64url'),
  816. }
  817. }
  818. function materializePersistenceSnapshots(
  819. snapshots: readonly SessionPersistenceSnapshot[],
  820. ): Map<SessionId, ObservedPersistedSession> {
  821. if (!isRuntimeArray(snapshots)) throw new Error('persistence snapshots must be an array')
  822. const result = new Map<SessionId, ObservedPersistedSession>()
  823. for (const snapshot of snapshots) {
  824. if (typeof snapshot.revision !== 'string') {
  825. throw new Error('persistence snapshot revision must be a string')
  826. }
  827. const header = structuredClone(snapshot.header)
  828. if (result.has(header.id)) {
  829. throw new Error(`persistence listed duplicate session "${header.id}"`)
  830. }
  831. result.set(header.id, { header, revision: snapshot.revision })
  832. }
  833. return result
  834. }
  835. function samePersistenceSnapshots(
  836. before: ReadonlyMap<SessionId, ObservedPersistedSession>,
  837. after: ReadonlyMap<SessionId, ObservedPersistedSession>,
  838. ): boolean {
  839. if (before.size !== after.size) return false
  840. for (const [id, first] of before) {
  841. const second = after.get(id)
  842. if (
  843. second === undefined
  844. || first.revision !== second.revision
  845. || !sameHeader(first.header, second.header)
  846. ) return false
  847. }
  848. return true
  849. }
  850. function sameSessionIds(
  851. before: ReadonlySet<SessionId>,
  852. after: ReadonlyMap<SessionId, ObservedSession>,
  853. ): boolean {
  854. if (before.size !== after.size) return false
  855. for (const id of before) {
  856. if (!after.has(id)) return false
  857. }
  858. return true
  859. }
  860. function sameHeader(a: SessionHeader, b: SessionHeader): boolean {
  861. return a.version === b.version
  862. && a.id === b.id
  863. && a.createdAt === b.createdAt
  864. && a.cwd === b.cwd
  865. && a.parentSession === b.parentSession
  866. && a.seedLength === b.seedLength
  867. && (a.delegationDepth ?? 0) === (b.delegationDepth ?? 0)
  868. && a.agentPreset === b.agentPreset
  869. }
  870. function rowHeader(row: SessionHeaderRow): SessionHeader {
  871. return {
  872. version: row.version,
  873. id: row.session_id as SessionId,
  874. createdAt: row.created_at,
  875. ...row.cwd === null ? {} : { cwd: row.cwd },
  876. ...row.parent_session === null ? {} : { parentSession: row.parent_session as SessionId },
  877. ...row.seed_length === null ? {} : { seedLength: row.seed_length },
  878. ...row.delegation_depth === null ? {} : { delegationDepth: row.delegation_depth },
  879. ...row.agent_preset === null ? {} : { agentPreset: row.agent_preset },
  880. }
  881. }
  882. function page<Row, Item>(
  883. rows: readonly Row[],
  884. limit: number,
  885. convert: (row: Row) => Item,
  886. nextCursor: (offset: number) => SessionSearchCursorValue,
  887. offset: number,
  888. ): SessionSearchPage<Item> {
  889. const hasMore = rows.length > limit
  890. return {
  891. items: rows.slice(0, limit).map(convert),
  892. ...hasMore ? { nextCursor: nextCursor(offset + limit) } : {},
  893. }
  894. }
  895. function encodeCursor(payload: CursorPayload): SessionSearchCursorValue {
  896. return SessionSearchCursor(Buffer.from(JSON.stringify(payload), 'utf8').toString('base64url'))
  897. }
  898. function decodeCursor(
  899. cursor: SessionSearchCursorValue,
  900. instance: string,
  901. scope: CursorPayload['scope'],
  902. fingerprint: string,
  903. generation: string,
  904. ): number {
  905. let decoded: Partial<CursorPayload>
  906. try {
  907. decoded = JSON.parse(Buffer.from(cursor, 'base64url').toString('utf8')) as Partial<CursorPayload>
  908. } catch (error: unknown) {
  909. throw invalidCursor(error)
  910. }
  911. if (
  912. decoded.version !== 1
  913. || decoded.instance !== instance
  914. || decoded.scope !== scope
  915. || decoded.fingerprint !== fingerprint
  916. || !Number.isSafeInteger(decoded.offset)
  917. || decoded.offset === undefined
  918. || decoded.offset < 0
  919. ) {
  920. throw invalidCursor(new Error('cursor does not belong to this normalized request'))
  921. }
  922. if (decoded.generation !== generation) {
  923. throw new SessionQueryError(
  924. 'session-search cursor is stale because its relevant corpus changed',
  925. 'SESSION_QUERY_STALE_CURSOR',
  926. )
  927. }
  928. return decoded.offset
  929. }
  930. function invalidCursor(cause: unknown): SessionQueryError {
  931. return new SessionQueryError(
  932. 'session-search cursor is invalid',
  933. 'SESSION_QUERY_INVALID_CURSOR',
  934. { cause },
  935. )
  936. }
  937. function resolveConfig(config: Config): ResolvedConfig {
  938. const resolved: ResolvedConfig = {
  939. path: config.path,
  940. openAt: config.openAt ?? 'startup',
  941. journalMode: config.journalMode ?? 'wal',
  942. defaultLimit: config.defaultLimit ?? SESSION_QUERY_SQLITE_DEFAULT_LIMIT,
  943. maxLimit: config.maxLimit ?? SESSION_QUERY_SQLITE_MAX_LIMIT,
  944. snippetChars: config.snippetChars ?? SESSION_QUERY_SQLITE_SNIPPET_CHARS,
  945. readWindowMax: config.readWindowMax ?? SESSION_QUERY_READ_WINDOW_MAX,
  946. persistedInspectConcurrency: config.persistedInspectConcurrency
  947. ?? SESSION_QUERY_DEFAULT_PERSISTED_INSPECT_CONCURRENCY,
  948. }
  949. if (typeof resolved.path !== 'string' || resolved.path.trim().length === 0) {
  950. throw invalidConfig('path must not be blank')
  951. }
  952. const openPhases: readonly string[] = ['startup', 'first-search', 'never']
  953. if (!openPhases.includes(resolved.openAt)) throw invalidConfig('openAt is not supported')
  954. assertPageLimit('defaultLimit', resolved.defaultLimit)
  955. assertPageLimit('maxLimit', resolved.maxLimit)
  956. assertPositiveInteger('snippetChars', resolved.snippetChars)
  957. if (!Number.isInteger(resolved.readWindowMax) || resolved.readWindowMax < 0) {
  958. throw invalidConfig('readWindowMax must be a non-negative integer')
  959. }
  960. if (
  961. !Number.isSafeInteger(resolved.persistedInspectConcurrency)
  962. || resolved.persistedInspectConcurrency < 1
  963. ) {
  964. throw invalidConfig('persistedInspectConcurrency must be a positive safe integer')
  965. }
  966. if (resolved.defaultLimit > resolved.maxLimit) {
  967. throw invalidConfig('defaultLimit must be less than or equal to maxLimit')
  968. }
  969. const journalModes: readonly string[] = ['wal', 'delete', 'truncate', 'persist']
  970. if (!journalModes.includes(resolved.journalMode)) throw invalidConfig('journalMode is not supported')
  971. return resolved
  972. }
  973. function assertPositiveInteger(name: string, value: number): void {
  974. if (!Number.isInteger(value) || value < 1) throw invalidConfig(`${name} must be a positive integer`)
  975. }
  976. function assertPageLimit(name: string, value: number): void {
  977. if (!Number.isSafeInteger(value) || value < 1 || value > SQLITE_MAX_PAGE_LIMIT) {
  978. throw invalidConfig(`${name} must be an integer between 1 and ${SQLITE_MAX_PAGE_LIMIT}`)
  979. }
  980. }
  981. function invalidConfig(detail: string): SessionQueryError {
  982. return new SessionQueryError(
  983. `session-search SQLite config: ${detail}`,
  984. 'SESSION_QUERY_INVALID_CONFIG',
  985. )
  986. }
  987. function indexClosed(): SessionQueryError {
  988. return new SessionQueryError('session-search SQLite index is closed', 'SESSION_QUERY_INDEX_FAILED')
  989. }
  990. function assertNotAborted(signal: AbortSignal | undefined): void {
  991. if (signal?.aborted) {
  992. throw new SessionQueryError('session-search aborted', 'SESSION_QUERY_ABORTED')
  993. }
  994. }
  995. function waitWithAbort<T>(promise: Promise<T>, signal: AbortSignal | undefined): Promise<T> {
  996. if (signal === undefined) return promise
  997. if (signal.aborted) return Promise.reject(new SessionQueryError('session-search aborted', 'SESSION_QUERY_ABORTED'))
  998. return new Promise<T>((resolve, reject) => {
  999. const onAbort = () => {
  1000. reject(new SessionQueryError('session-search aborted', 'SESSION_QUERY_ABORTED'))
  1001. }
  1002. signal.addEventListener('abort', onAbort, { once: true })
  1003. promise.then(
  1004. (value) => {
  1005. signal.removeEventListener('abort', onAbort)
  1006. resolve(value)
  1007. },
  1008. (error: unknown) => {
  1009. signal.removeEventListener('abort', onAbort)
  1010. reject(asError(error))
  1011. },
  1012. )
  1013. })
  1014. }
  1015. function isAbort(error: unknown): boolean {
  1016. return error instanceof SessionQueryError && error.code === 'SESSION_QUERY_ABORTED'
  1017. }
  1018. function asError(error: unknown): Error {
  1019. return error instanceof Error
  1020. ? error
  1021. : new Error('session-search dependency rejected with a non-Error value', { cause: error })
  1022. }
  1023. function errorMessage(error: unknown): string {
  1024. return error instanceof Error ? error.message : 'unknown error'
  1025. }
  1026. function isRuntimeArray(value: unknown): boolean {
  1027. return Array.isArray(value)
  1028. }
  1029. export default SqliteSessionQueryEngine