manual-compaction.spec.ts 36 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885
  1. import { describe, expect, it, vi } from 'vitest'
  2. import { Context } from '@deepseek-ai/cordis'
  3. import AgentLoop from '@deepseek-ai/dsh-agent-loop'
  4. import { mountAgentLoopTestDependencies } from '@deepseek-ai/dsh-agent-loop-testkit'
  5. import InvariantRegistry from '@deepseek-ai/dsh-invariants'
  6. import { CommandId } from '@deepseek-ai/dsh-commands/brand'
  7. import * as SessionInvariant from '@deepseek-ai/dsh-session/invariant'
  8. import * as AgentInvariant from '@deepseek-ai/dsh-agent/invariant'
  9. import * as AgentLoopInvariant from '@deepseek-ai/dsh-agent-loop/invariant'
  10. import * as CompactionInvariant from '@deepseek-ai/dsh-compaction/invariant'
  11. import * as CompactionBasicInvariant from '@deepseek-ai/dsh-compaction-basic/invariant'
  12. import { BasicCompactionEngine } from '@deepseek-ai/dsh-compaction-basic'
  13. import { CompactionId, isCompactCheckpointSource, ManualCompactionError } from '@deepseek-ai/dsh-compaction'
  14. import type { CompactionResult } from '@deepseek-ai/dsh-compaction'
  15. import {
  16. createAssistantMessage,
  17. createUserMessage,
  18. LlmAdapter,
  19. } from '@deepseek-ai/dsh-llm'
  20. import type {
  21. ContentBlock,
  22. LlmResolvedModelInfo,
  23. Message,
  24. StreamChunk,
  25. TokenUsage,
  26. } from '@deepseek-ai/dsh-llm'
  27. import SessionStore, { Session, SessionId, type SessionEvent } from '@deepseek-ai/dsh-session'
  28. import LlmRuntime from '@deepseek-ai/dsh-llm'
  29. import TokenMeter from '@deepseek-ai/dsh-token-meter'
  30. import type { Agent } from '@deepseek-ai/dsh-agent'
  31. import type {
  32. SummarizationInput,
  33. SummaryResult,
  34. } from '@deepseek-ai/dsh-compaction-basic/src/summarizer.ts'
  35. const MODEL = 'mock'
  36. const SIGNAL = new AbortController().signal
  37. const PROMPT = 'older conversation history '.repeat(60)
  38. /** A summarizer under test control: it can block, fail, or mutate mid-call. */
  39. class GatedCompactionEngine extends BasicCompactionEngine {
  40. summary: ContentBlock[] = [{ type: 'text', text: 'checkpoint' }]
  41. rawOutput: ContentBlock[] | undefined
  42. usage: TokenUsage | undefined
  43. error: unknown
  44. gate: Promise<undefined> | undefined
  45. duringSummary: (() => void) | undefined
  46. calls: SummarizationInput[] = []
  47. override async summarize(
  48. input: SummarizationInput,
  49. _agent: Agent,
  50. _signal?: AbortSignal,
  51. ): Promise<SummaryResult> {
  52. this.calls.push(input)
  53. this.duringSummary?.()
  54. if (this.gate !== undefined) await this.gate
  55. if (this.error !== undefined) throw this.error
  56. return {
  57. summary: this.summary,
  58. ...this.rawOutput === undefined ? {} : { rawOutput: this.rawOutput },
  59. provider: 'summary-provider',
  60. model: 'summary-model',
  61. ...this.usage === undefined ? {} : { usage: this.usage },
  62. }
  63. }
  64. }
  65. /** One text answer per request, with a context window large enough to avoid pressure. */
  66. class TextAdapter extends LlmAdapter {
  67. readonly requests: Message[][] = []
  68. override resolveModel(provider: string, model: string): Promise<LlmResolvedModelInfo> {
  69. return Promise.resolve({
  70. provider,
  71. id: model,
  72. name: model,
  73. context: { contextWindow: 100_000 },
  74. })
  75. }
  76. override async * stream(options: { messages: readonly Message[] }): AsyncIterable<StreamChunk> {
  77. this.requests.push([...options.messages])
  78. yield { type: 'block-start', index: 0, blockType: 'text' }
  79. yield { type: 'block-end', index: 0, block: { type: 'text', text: 'answer' } }
  80. yield { type: 'finish', reason: { kind: 'stop' } }
  81. }
  82. }
  83. interface LoopHarness {
  84. readonly ctx: Context
  85. readonly agent: Agent
  86. readonly compact: GatedCompactionEngine
  87. readonly adapter: TextAdapter
  88. readonly log: string[]
  89. }
  90. /** Real loop, session store, and invariant companions around manual compaction. */
  91. async function loopHarness(): Promise<LoopHarness> {
  92. const ctx = new Context()
  93. await mountAgentLoopTestDependencies(ctx)
  94. await ctx.plugin(InvariantRegistry)
  95. await ctx.plugin(SessionInvariant)
  96. await ctx.plugin(AgentInvariant)
  97. await ctx.plugin(AgentLoopInvariant)
  98. await ctx.plugin(CompactionInvariant)
  99. await ctx.plugin(CompactionBasicInvariant)
  100. await ctx.plugin(AgentLoop, { agents: [] })
  101. await ctx.plugin(TokenMeter)
  102. const adapter = new TextAdapter()
  103. ctx.llm.registerAdapter([MODEL], adapter)
  104. const compact = new GatedCompactionEngine(ctx, { auto: false })
  105. const agent = ctx.agentLoop.create(SessionId('manual-compact'), { provider: MODEL, model: MODEL })
  106. const log: string[] = []
  107. ctx.on('session/event', (_session, event) => {
  108. if (event.type === 'turn/start') log.push('turn/start')
  109. if (event.type === 'turn/end') log.push('turn/end')
  110. if (event.type === 'compaction/start') log.push(`compaction/start:${String(event.data.turn)}`)
  111. if (event.type === 'compaction/summary') log.push('compaction/summary')
  112. if (event.type === 'compaction/end') log.push(`compaction/end:${String(event.data.turn)}`)
  113. if (event.type === 'user/message') log.push('user/message')
  114. })
  115. ctx.on('session/flush', () => { log.push('flush') })
  116. return { ctx, agent, compact, adapter, log }
  117. }
  118. /** Drive one real turn so the closed history holds a compactable older span. */
  119. async function seedHistory(harness: LoopHarness): Promise<void> {
  120. harness.agent.followup(createUserMessage({
  121. content: [{ type: 'text', text: PROMPT }],
  122. source: { kind: 'user' },
  123. }))
  124. await harness.agent.whenIdle()
  125. harness.log.length = 0
  126. }
  127. /** Text of every derived model-visible message, in request order. */
  128. function derivedText(session: Session): string[] {
  129. return session.deriveMessages().map((message: Message) => message.content
  130. .map(block => block.type === 'text' ? block.text : '')
  131. .join(''))
  132. }
  133. /** Await one classified manual-compaction rejection. */
  134. async function rejection(operation: Promise<unknown> | (() => Promise<unknown>)): Promise<ManualCompactionError> {
  135. let caught: unknown
  136. try {
  137. const value = await (typeof operation === 'function' ? operation() : operation)
  138. throw new Error(`expected a rejection, resolved with ${String(value)}`)
  139. } catch (error: unknown) {
  140. caught = error
  141. }
  142. if (!(caught instanceof ManualCompactionError)) {
  143. throw new Error(`expected a ManualCompactionError, got ${String(caught)}`)
  144. }
  145. return caught
  146. }
  147. /** The Error a classified failure wraps. */
  148. function causeOf(error: ManualCompactionError): Error {
  149. const { cause } = error
  150. if (!(cause instanceof Error)) throw new Error(`expected an Error cause, got ${String(cause)}`)
  151. return cause
  152. }
  153. function deferred(): { promise: Promise<undefined>; resolve: () => void } {
  154. const { promise, resolve } = Promise.withResolvers<undefined>()
  155. return { promise, resolve: () => { resolve(undefined) } }
  156. }
  157. /** A closed-tail session with compactable exchanges and no live agent. */
  158. function closedConversation(turns = 2, lastTurnNumber = turns): Session {
  159. const session = Session.create(SessionId(`closed-${turns}-${lastTurnNumber}`))
  160. for (let index = 1; index <= turns; index += 1) {
  161. const turn = index === turns ? lastTurnNumber : index
  162. session.append('turn/start', { turn })
  163. session.append('user/message', createUserMessage({
  164. content: [{ type: 'text', text: `${PROMPT} ${turn}` }],
  165. source: { kind: 'user' },
  166. }), { surfaceOp: 'append' })
  167. session.append('step/start', { turn, step: 1 })
  168. if (index === 1) {
  169. session.append('request/header', {
  170. header: { config: { provider: MODEL, model: MODEL } },
  171. reason: 'initial',
  172. })
  173. }
  174. session.append('assistant/message', {
  175. turn,
  176. step: 1,
  177. message: createAssistantMessage({
  178. content: [{ type: 'text', text: `answer ${turn}` }],
  179. source: { provider: MODEL, model: MODEL },
  180. }),
  181. }, { surfaceOp: 'append' })
  182. session.append('step/end', { turn, step: 1 })
  183. session.append('turn/end', { turn, reason: { kind: 'completed' } })
  184. }
  185. return session
  186. }
  187. /** A fake idle agent whose maintenance claim is scripted per test. */
  188. function fakeAgent(
  189. session: Session,
  190. reserve: () => (() => void) | undefined,
  191. maintenanceSignal = new AbortController().signal,
  192. ): Agent {
  193. return {
  194. session,
  195. options: { provider: MODEL, model: MODEL },
  196. runMaintenance<T>(task: (signal: AbortSignal) => Promise<T>): Promise<T> {
  197. const release = reserve()
  198. if (release === undefined) throw new Error('agent already has active work')
  199. return task(maintenanceSignal).finally(release)
  200. },
  201. } as unknown as Agent
  202. }
  203. /** Service over a store-detached session for failure classification. */
  204. function detachedService(): { ctx: Context; compact: GatedCompactionEngine; flushes: () => number } {
  205. const ctx = new Context()
  206. void new LlmRuntime(ctx)
  207. void new SessionStore(ctx)
  208. void new TokenMeter(ctx)
  209. ctx.llm.registerAdapter([MODEL], new TextAdapter())
  210. let flushes = 0
  211. vi.spyOn(ctx.sessions, 'flush').mockImplementation(() => {
  212. flushes += 1
  213. return Promise.resolve(false)
  214. })
  215. return { ctx, compact: new GatedCompactionEngine(ctx, { auto: false }), flushes: () => flushes }
  216. }
  217. function compactEvents(session: Session): Array<Session['events'][number]> {
  218. return session.events.filter(event => event.type.startsWith('compaction/'))
  219. }
  220. describe('compactNow through the real loop', () => {
  221. it('holds a prompt accepted during summarization until the standalone bracket is flushed', async () => {
  222. const harness = await loopHarness()
  223. const { agent, compact, adapter, log } = harness
  224. await seedHistory(harness)
  225. const gate = deferred()
  226. compact.gate = gate.promise
  227. const running = compact.compactNow(agent, SIGNAL)
  228. await Promise.resolve()
  229. expect(log).toEqual(['compaction/start:null'])
  230. agent.followup(createUserMessage({
  231. content: [{ type: 'text', text: 'after compaction' }],
  232. source: { kind: 'user' },
  233. }))
  234. await new Promise<void>((resolve) => { setTimeout(resolve, 5) })
  235. expect(agent.status).toBe('idle')
  236. expect(adapter.requests).toHaveLength(1)
  237. expect(log).toEqual(['compaction/start:null'])
  238. gate.resolve()
  239. const result = await running
  240. expect(result).not.toBeNull()
  241. await agent.whenIdle()
  242. const start = log.indexOf('compaction/start:null')
  243. const summary = log.indexOf('compaction/summary')
  244. const end = log.indexOf('compaction/end:null')
  245. const flush = log.indexOf('flush')
  246. const nextTurn = log.indexOf('turn/start')
  247. expect(start).toBeLessThan(summary)
  248. expect(summary).toBeLessThan(end)
  249. expect(end).toBeLessThan(flush)
  250. expect(flush).toBeLessThan(nextTurn)
  251. expect(adapter.requests).toHaveLength(2)
  252. const second = (adapter.requests[1] ?? []).map(message => message.content
  253. .map(block => block.type === 'text' ? block.text : '')
  254. .join(''))
  255. expect(second[0]).toContain('checkpoint')
  256. expect(second.at(-1)).toBe('after compaction')
  257. expect(second.some(text => text.includes(PROMPT))).toBe(false)
  258. })
  259. it('keeps context injected during summarization pending for the next step', async () => {
  260. const harness = await loopHarness()
  261. const { agent, compact } = harness
  262. await seedHistory(harness)
  263. compact.duringSummary = () => {
  264. agent.inject(createUserMessage({
  265. content: [{ type: 'text', text: 'INJECTED CONTEXT' }],
  266. source: { kind: 'plugin', plugin: 'test' },
  267. }))
  268. }
  269. const result = await compact.compactNow(agent, SIGNAL)
  270. expect(result).not.toBeNull()
  271. const start = agent.session.events.findLast(event => event.type === 'compaction/start')
  272. const injected = agent.inbox.nextStep.find(message =>
  273. message.source.kind === 'plugin' && message.source.plugin === 'test')
  274. const end = agent.session.events.findLast(event => event.type === 'compaction/end')
  275. expect(start).toBeDefined()
  276. expect(injected).toBeDefined()
  277. expect(end).toBeDefined()
  278. expect(agent.session.events.some(event => event.type === 'user/message'
  279. && event.data.id === injected?.id)).toBe(false)
  280. agent.followup(createUserMessage({
  281. content: [{ type: 'text', text: 'after compaction' }],
  282. source: { kind: 'user' },
  283. }))
  284. await agent.whenIdle()
  285. const messages = derivedText(agent.session)
  286. expect(messages[0]).toContain('checkpoint')
  287. expect(messages.filter(text => text.includes('INJECTED CONTEXT'))).toHaveLength(1)
  288. })
  289. it('keeps the marker order when listeners attempt a re-entrant injection', async () => {
  290. const harness = await loopHarness()
  291. const { ctx, agent, compact } = harness
  292. await seedHistory(harness)
  293. const attempts: string[] = []
  294. ctx.on('session/event', (_session, event) => {
  295. if (event.type !== 'compaction/start' && event.type !== 'compaction/summary') return
  296. attempts.push(event.type)
  297. agent.inject(createUserMessage({
  298. content: [{ type: 'text', text: `from ${event.type}` }],
  299. source: { kind: 'plugin', plugin: 'listener' },
  300. }))
  301. })
  302. const result = await compact.compactNow(agent, SIGNAL)
  303. expect(attempts).toEqual(['compaction/start', 'compaction/summary'])
  304. expect(result).not.toBeNull()
  305. expect(derivedText(agent.session)[0]).toContain('checkpoint')
  306. expect(agent.session.events.filter(event => event.type === 'user/message'
  307. && event.data.source.kind === 'plugin' && event.data.source.plugin === 'listener')).toHaveLength(0)
  308. const types = compactEvents(agent.session).map(event => event.type)
  309. expect(types).toEqual(['compaction/start', 'compaction/summary', 'compaction/end'])
  310. })
  311. it('reports busy without summarizing when a prompt already owns the next turn', async () => {
  312. const harness = await loopHarness()
  313. const { agent, compact, adapter } = harness
  314. await seedHistory(harness)
  315. agent.followup(createUserMessage({
  316. content: [{ type: 'text', text: 'first in line' }],
  317. source: { kind: 'user' },
  318. }))
  319. expect((await rejection(() => compact.compactNow(agent, SIGNAL))).code).toBe('busy')
  320. expect(compact.calls).toHaveLength(0)
  321. await agent.whenIdle()
  322. expect(adapter.requests).toHaveLength(2)
  323. expect(agent.session.events.some(event => event.type === 'compaction/start')).toBe(false)
  324. })
  325. it('releases turn admission after a summarizer failure and records the failed attempt', async () => {
  326. const harness = await loopHarness()
  327. const { agent, compact, adapter } = harness
  328. await seedHistory(harness)
  329. compact.error = new Error('summarizer unavailable')
  330. const before = [...agent.session.surface.nodes]
  331. expect((await rejection(compact.compactNow(agent, SIGNAL))).code).toBe('summary')
  332. expect(agent.session.surface.nodes).toEqual(before)
  333. const markers = compactEvents(agent.session)
  334. expect(markers.map(event => event.type)).toEqual(['compaction/start', 'compaction/end'])
  335. expect(markers[1]?.type === 'compaction/end' && markers[1].data.error)
  336. .toContain('summarizer unavailable')
  337. agent.followup(createUserMessage({
  338. content: [{ type: 'text', text: 'runs after the failure' }],
  339. source: { kind: 'user' },
  340. }))
  341. await agent.whenIdle()
  342. expect(adapter.requests).toHaveLength(2)
  343. })
  344. })
  345. describe('compactNow transaction and failure classification', () => {
  346. it('returns null without writing a bracket for history that cannot be compacted', async () => {
  347. const { compact } = detachedService()
  348. const session = Session.create(SessionId('empty'))
  349. let released = 0
  350. const agent = fakeAgent(session, () => () => { released += 1 })
  351. expect(await compact.compactNow(agent, SIGNAL)).toBeNull()
  352. expect(released).toBe(1)
  353. expect(compact.calls).toHaveLength(0)
  354. expect(compactEvents(session)).toEqual([])
  355. })
  356. it('commits a standalone bracket without consuming a turn number and checkpoints durability', async () => {
  357. const { compact, flushes } = detachedService()
  358. const session = closedConversation(2, 7)
  359. const agent = fakeAgent(session, () => () => undefined)
  360. const commandId = CommandId('manual-compact-command')
  361. const result = await compact.compactNow(agent, SIGNAL, commandId)
  362. expect(result).not.toBeNull()
  363. expect(result?.sourceCommandId).toBe(commandId)
  364. expect(flushes()).toBe(1)
  365. expect(session.events.filter(event => event.type === 'turn/start').at(-1)?.data.turn).toBe(7)
  366. const start = session.events.findLast(event => event.type === 'compaction/start')
  367. const summaryEvent = session.events.findLast(event => event.type === 'compaction/summary')
  368. const checkpoint = session.events.findLast(
  369. (event): event is SessionEvent<'user/message'> => event.type === 'user/message'
  370. && isCompactCheckpointSource(event.data.source),
  371. )
  372. const end = session.events.findLast(event => event.type === 'compaction/end')
  373. const correlated = { compactionId: result?.compactionId, sourceCommandId: commandId }
  374. expect(start?.data).toEqual({ ...correlated, turn: null })
  375. expect(summaryEvent?.data.sourceCommandId).toBe(commandId)
  376. expect(checkpoint?.data.source).toMatchObject(correlated)
  377. expect(end?.data).toEqual({ ...correlated, turn: null })
  378. })
  379. it('reports a live unmatched bracket as busy without summarizing', async () => {
  380. const { compact } = detachedService()
  381. const session = closedConversation(2)
  382. session.append('compaction/start', {
  383. compactionId: CompactionId('live-manual-compaction'),
  384. turn: null,
  385. })
  386. const agent = fakeAgent(session, () => () => undefined)
  387. const error = await rejection(() => compact.compactNow(agent, SIGNAL))
  388. expect(error.code).toBe('busy')
  389. expect(error.message).toContain('compaction lock is already active')
  390. expect(compact.calls).toHaveLength(0)
  391. })
  392. it('ignores an unmatched bracket inherited before a later end-seed marker', async () => {
  393. const { compact } = detachedService()
  394. const original = closedConversation(2)
  395. original.append('compaction/start', {
  396. compactionId: CompactionId('stale-manual-compaction'),
  397. turn: null,
  398. })
  399. const reloaded = Session.create(SessionId('stale-orphan'), [...original.events])
  400. const boundary = reloaded.events.findLast(event => event.type === 'session/end-seed')
  401. const orphan = reloaded.events.find(event => event.type === 'compaction/start')
  402. const agent = fakeAgent(reloaded, () => () => undefined)
  403. expect(boundary?.seq).toBeGreaterThan(orphan?.seq ?? Number.MAX_SAFE_INTEGER)
  404. await expect(compact.compactNow(agent, SIGNAL)).resolves.not.toBeNull()
  405. expect(compact.calls).toHaveLength(1)
  406. })
  407. it('scans a stale orphan independently of later repaired turn state', async () => {
  408. const { compact } = detachedService()
  409. const original = closedConversation(2)
  410. original.append('compaction/start', {
  411. compactionId: CompactionId('reloaded-manual-compaction'),
  412. turn: null,
  413. })
  414. original.append('turn/start', { turn: 3 })
  415. original.append('turn/end', { turn: 3, reason: { kind: 'interrupted' } })
  416. const reloaded = Session.create(SessionId('reloaded-orphan'), [...original.events])
  417. const agent = fakeAgent(reloaded, () => () => undefined)
  418. await expect(compact.compactNow(agent, SIGNAL)).resolves.not.toBeNull()
  419. expect(compact.calls).toHaveLength(1)
  420. })
  421. it('refuses an open turn in the log', async () => {
  422. const { compact } = detachedService()
  423. const session = closedConversation(2)
  424. session.append('turn/start', { turn: 3 })
  425. const agent = fakeAgent(session, () => () => undefined)
  426. const error = await rejection(compact.compactNow(agent, SIGNAL))
  427. expect(error.code).toBe('busy')
  428. expect(error.message).toContain('already has an open turn')
  429. })
  430. it('reports busy and skips summarization when admission is unavailable', async () => {
  431. const { compact } = detachedService()
  432. const agent = fakeAgent(closedConversation(2), () => undefined)
  433. expect((await rejection(() => compact.compactNow(agent, SIGNAL))).code).toBe('busy')
  434. expect(compact.calls).toHaveLength(0)
  435. })
  436. it('rejects a selected span replaced during summarization and records an error close', async () => {
  437. const { compact, flushes } = detachedService()
  438. const session = closedConversation(2)
  439. let released = 0
  440. const agent = fakeAgent(session, () => () => { released += 1 })
  441. compact.duringSummary = () => {
  442. const [head] = session.surface.nodes
  443. session.append('user/message', createUserMessage({
  444. content: [{ type: 'text', text: 'competing replacement' }],
  445. source: { kind: 'plugin', plugin: 'rival' },
  446. }), {
  447. surfaceOp: { op: 'replace', start: head!, end: head! },
  448. sourceEventSeqs: [head!],
  449. })
  450. }
  451. expect((await rejection(compact.compactNow(agent, SIGNAL))).code).toBe('changed')
  452. expect(released).toBe(1)
  453. expect(flushes()).toBe(1)
  454. expect(compactEvents(session).map(event => event.type)).toEqual(['compaction/start', 'compaction/end'])
  455. })
  456. it('rejects a selected span whose middle node was replaced during summarization', async () => {
  457. const { compact } = detachedService()
  458. const session = closedConversation(3)
  459. const agent = fakeAgent(session, () => () => undefined)
  460. compact.duringSummary = () => {
  461. const middle = session.surface.nodes[1]
  462. session.append('user/message', createUserMessage({
  463. content: [{ type: 'text', text: 'rewritten middle node' }],
  464. source: { kind: 'plugin', plugin: 'rival' },
  465. }), {
  466. surfaceOp: { op: 'replace', start: middle!, end: middle! },
  467. sourceEventSeqs: [middle!],
  468. })
  469. }
  470. const error = await rejection(compact.compactNow(agent, SIGNAL))
  471. expect(error.code).toBe('changed')
  472. expect(causeOf(error).message).toContain('span changed during summarization')
  473. })
  474. it('revalidates the selected span after the summarizer continuation settles', async () => {
  475. const { compact, flushes } = detachedService()
  476. const session = closedConversation(2)
  477. const gate = deferred()
  478. compact.gate = gate.promise
  479. let released = 0
  480. const agent = fakeAgent(session, () => () => { released += 1 })
  481. const head = session.surface.nodes[0]!
  482. const generation = session.surface.replaceGeneration
  483. const running = compact.compactNow(agent, SIGNAL)
  484. await Promise.resolve()
  485. expect(compact.calls).toHaveLength(1)
  486. gate.resolve()
  487. queueMicrotask(() => {
  488. queueMicrotask(() => {
  489. session.append('user/message', createUserMessage({
  490. content: [{ type: 'text', text: 'late competing replacement' }],
  491. source: { kind: 'plugin', plugin: 'rival' },
  492. }), {
  493. surfaceOp: { op: 'replace', start: head, end: head },
  494. sourceEventSeqs: [head],
  495. })
  496. })
  497. })
  498. const error = await rejection(running)
  499. expect(error.code).toBe('changed')
  500. expect(causeOf(error).message).toContain('selected span')
  501. expect(released).toBe(1)
  502. expect(flushes()).toBe(1)
  503. expect(session.surface.replaceGeneration).toBe(generation + 1)
  504. expect(session.surface.nodes).not.toContain(head)
  505. expect(compactEvents(session).map(event => event.type)).toEqual(['compaction/start', 'compaction/end'])
  506. expect(session.events.some(event => event.type === 'user/message'
  507. && isCompactCheckpointSource(event.data.source))).toBe(false)
  508. })
  509. it('classifies a failing compaction/end as commit failure and leaves one orphan', async () => {
  510. const { compact, flushes } = detachedService()
  511. const session = closedConversation(2)
  512. const agent = fakeAgent(session, () => () => undefined)
  513. const append = session.append.bind(session)
  514. vi.spyOn(session, 'append').mockImplementation(((type: string, ...rest: never[]) => {
  515. if (type === 'compaction/end') throw new Error('boundary rejected')
  516. return (append as (...args: never[]) => unknown)(type as never, ...rest)
  517. }) as never)
  518. const error = await rejection(compact.compactNow(agent, SIGNAL))
  519. expect(error.code).toBe('commit')
  520. expect(causeOf(error).message).toBe('boundary rejected')
  521. vi.restoreAllMocks()
  522. expect(flushes()).toBe(0)
  523. expect(session.events.findLast(event => event.type.startsWith('compaction/'))?.type)
  524. .toBe('compaction/summary')
  525. expect(compactEvents(session).filter(event => event.type === 'compaction/start')).toHaveLength(1)
  526. const calls = compact.calls.length
  527. expect((await rejection(compact.compactNow(agent, SIGNAL))).code).toBe('busy')
  528. expect(compact.calls).toHaveLength(calls)
  529. })
  530. it('keeps a failed error-close as the commit failure and does not flush', async () => {
  531. const { compact, flushes } = detachedService()
  532. const session = closedConversation(2)
  533. let released = 0
  534. const agent = fakeAgent(session, () => () => { released += 1 })
  535. compact.error = new Error('summary rejected')
  536. const append = session.append.bind(session)
  537. vi.spyOn(session, 'append').mockImplementation(((type: string, ...rest: never[]) => {
  538. if (type === 'compaction/end') throw new Error('error boundary rejected')
  539. return (append as (...args: never[]) => unknown)(type as never, ...rest)
  540. }) as never)
  541. const error = await rejection(compact.compactNow(agent, SIGNAL))
  542. vi.restoreAllMocks()
  543. expect(error.code).toBe('commit')
  544. expect(causeOf(error).message).toBe('error boundary rejected')
  545. expect(released).toBe(1)
  546. expect(flushes()).toBe(0)
  547. expect(compactEvents(session).map(event => event.type)).toEqual(['compaction/start'])
  548. })
  549. it('rejects a selected span whose pricing changed during summarization', async () => {
  550. const { ctx, compact } = detachedService()
  551. const session = closedConversation(2)
  552. const agent = fakeAgent(session, () => () => undefined)
  553. const meter = ctx.tokenMeter
  554. const original = meter.measure.bind(meter)
  555. compact.duringSummary = () => {
  556. vi.spyOn(meter, 'measure').mockImplementationOnce((target) => {
  557. const measurement = original(target)
  558. return {
  559. ...measurement,
  560. nodes: measurement.nodes.map((node, index) =>
  561. index === 0 ? { ...node, tokens: node.tokens + 1 } : node),
  562. }
  563. })
  564. }
  565. expect((await rejection(compact.compactNow(agent, SIGNAL))).code).toBe('changed')
  566. vi.restoreAllMocks()
  567. })
  568. it('classifies a commit-body failure and still releases admission', async () => {
  569. const { compact } = detachedService()
  570. const session = closedConversation(2)
  571. let released = 0
  572. const agent = fakeAgent(session, () => () => { released += 1 })
  573. const append = session.append.bind(session)
  574. vi.spyOn(session, 'append').mockImplementation(((type: string, ...rest: never[]) => {
  575. if (type === 'compaction/summary') throw new Error('summary record rejected')
  576. return (append as (...args: never[]) => unknown)(type as never, ...rest)
  577. }) as never)
  578. const error = await rejection(compact.compactNow(agent, SIGNAL))
  579. vi.restoreAllMocks()
  580. expect(error.code).toBe('commit')
  581. expect(released).toBe(1)
  582. const end = session.events.findLast(event => event.type === 'compaction/end')
  583. expect(end?.type === 'compaction/end' && end.data.error).toContain('summary record rejected')
  584. expect(end?.type === 'compaction/end' && end.data.turn).toBeNull()
  585. })
  586. it('keeps a commit failure when the durability checkpoint also fails', async () => {
  587. const { ctx, compact } = detachedService()
  588. const session = closedConversation(2)
  589. const agent = fakeAgent(session, () => () => undefined)
  590. const append = session.append.bind(session)
  591. vi.spyOn(session, 'append').mockImplementation(((type: string, ...rest: never[]) => {
  592. if (type === 'compaction/summary') throw new Error('summary record rejected')
  593. return (append as (...args: never[]) => unknown)(type as never, ...rest)
  594. }) as never)
  595. vi.spyOn(ctx.sessions, 'flush').mockRejectedValueOnce(new Error('disk full'))
  596. const error = await rejection(compact.compactNow(agent, SIGNAL))
  597. expect(error.code).toBe('commit')
  598. expect(causeOf(error).message).toBe('summary record rejected')
  599. vi.restoreAllMocks()
  600. })
  601. it('compacts a session with no durable turn boundary without creating one', async () => {
  602. const { compact } = detachedService()
  603. const session = Session.create(SessionId('turnless'))
  604. for (const text of [PROMPT, 'recent tail']) {
  605. session.append('user/message', createUserMessage({
  606. content: [{ type: 'text', text }],
  607. source: { kind: 'user' },
  608. }), { surfaceOp: 'append' })
  609. }
  610. const agent = fakeAgent(session, () => () => undefined)
  611. const result = await compact.compactNow(agent, SIGNAL)
  612. expect(result).not.toBeNull()
  613. expect(session.events.some(event => event.type === 'turn/start')).toBe(false)
  614. expect(session.events.find(event => event.type === 'compaction/start')?.data)
  615. .toEqual({ compactionId: result?.compactionId, turn: null })
  616. })
  617. it('classifies a durability failure after the standalone bracket committed', async () => {
  618. const { ctx, compact } = detachedService()
  619. const session = closedConversation(2)
  620. const agent = fakeAgent(session, () => () => undefined)
  621. vi.spyOn(ctx.sessions, 'flush').mockRejectedValueOnce(new Error('disk full'))
  622. expect((await rejection(compact.compactNow(agent, SIGNAL))).code).toBe('persistence')
  623. vi.restoreAllMocks()
  624. expect(session.events.some(event => event.type === 'compaction/summary')).toBe(true)
  625. const start = session.events.findLast(event => event.type === 'compaction/start')
  626. const end = session.events.findLast(event => event.type === 'compaction/end')
  627. expect(end?.data).toEqual({ compactionId: start?.data.compactionId, turn: null })
  628. })
  629. it('lets a pre-aborted signal win before reservation, measurement, or summarization', async () => {
  630. const cases = [
  631. { name: 'busy', session: closedConversation(2), release: undefined },
  632. { name: 'empty', session: Session.create(SessionId('pre-aborted-empty')), release: () => undefined },
  633. { name: 'compactable', session: closedConversation(2, 9), release: () => undefined },
  634. ] as const
  635. for (const testCase of cases) {
  636. const { ctx, compact } = detachedService()
  637. const reserve = vi.fn(() => testCase.release)
  638. const measure = vi.spyOn(ctx.tokenMeter, 'measure')
  639. const agent = fakeAgent(testCase.session, reserve)
  640. const before = [...testCase.session.events]
  641. const reason = Object.freeze({ kind: 'cancelled', case: testCase.name })
  642. const controller = new AbortController()
  643. controller.abort(reason)
  644. let thrown: unknown
  645. try {
  646. void compact.compactNow(agent, controller.signal)
  647. } catch (error: unknown) {
  648. thrown = error
  649. }
  650. expect(thrown).toBe(reason)
  651. expect(reserve).not.toHaveBeenCalled()
  652. expect(measure).not.toHaveBeenCalled()
  653. expect(compact.calls).toHaveLength(0)
  654. expect(testCase.session.events).toEqual(before)
  655. vi.restoreAllMocks()
  656. }
  657. })
  658. it('preserves the exact cancellation reason when the summarizer also rejects', async () => {
  659. const { compact, flushes } = detachedService()
  660. const controller = new AbortController()
  661. const reason = new Error('cancelled by the caller')
  662. let released = 0
  663. const session = closedConversation(2)
  664. const agent = fakeAgent(session, () => () => { released += 1 })
  665. compact.duringSummary = () => { controller.abort(reason) }
  666. compact.error = new Error('summarizer aborted')
  667. await expect(compact.compactNow(agent, controller.signal)).rejects.toBe(reason)
  668. expect(released).toBe(1)
  669. expect(flushes()).toBe(1)
  670. const events = compactEvents(session)
  671. expect(events.map(event => event.type)).toEqual(['compaction/start', 'compaction/end'])
  672. expect(events[1]?.type === 'compaction/end' && events[1].data.error)
  673. .toContain('summarizer aborted')
  674. })
  675. it('classifies agent cancellation during maintenance as an expected cancellation', async () => {
  676. const { compact } = detachedService()
  677. const controller = new AbortController()
  678. const reason = new Error('agent cancelled maintenance')
  679. const session = closedConversation(2)
  680. const agent = fakeAgent(session, () => () => undefined, controller.signal)
  681. compact.duringSummary = () => { controller.abort(reason) }
  682. compact.error = new Error('summarizer observed cancellation')
  683. const error = await rejection(compact.compactNow(agent, SIGNAL))
  684. expect(error.code).toBe('cancelled')
  685. expect(error.cause).toBe(reason)
  686. })
  687. it('aborts before committing when cancellation lands after summarization', async () => {
  688. const { compact } = detachedService()
  689. const controller = new AbortController()
  690. const reason = new Error('cancelled by the caller')
  691. const session = closedConversation(2)
  692. const agent = fakeAgent(session, () => () => undefined)
  693. compact.duringSummary = () => { controller.abort(reason) }
  694. await expect(compact.compactNow(agent, controller.signal)).rejects.toBe(reason)
  695. expect(compactEvents(session).map(event => event.type)).toEqual(['compaction/start', 'compaction/end'])
  696. expect(session.events.some(event => event.type === 'compaction/summary')).toBe(false)
  697. })
  698. it('waits for the durability checkpoint before cancellation wins and admission releases', async () => {
  699. const { ctx, compact } = detachedService()
  700. const controller = new AbortController()
  701. const reason = new Error('cancelled during flush')
  702. const flushGate = Promise.withResolvers<boolean>()
  703. const flush = vi.spyOn(ctx.sessions, 'flush').mockReturnValueOnce(flushGate.promise)
  704. const session = closedConversation(2)
  705. let released = 0
  706. const agent = fakeAgent(session, () => () => { released += 1 })
  707. const running = compact.compactNow(agent, controller.signal)
  708. let settled = false
  709. void running.then(
  710. () => { settled = true },
  711. () => { settled = true },
  712. )
  713. await vi.waitFor(() => {
  714. expect(flush).toHaveBeenCalledWith(session)
  715. })
  716. controller.abort(reason)
  717. await Promise.resolve()
  718. expect(settled).toBe(false)
  719. expect(released).toBe(0)
  720. flushGate.resolve(false)
  721. await expect(running).rejects.toBe(reason)
  722. expect(released).toBe(1)
  723. })
  724. it('preserves raw output and usage in the manual summary event', async () => {
  725. const { compact } = detachedService()
  726. const session = closedConversation(2)
  727. const agent = fakeAgent(session, () => () => undefined)
  728. compact.rawOutput = [
  729. { type: 'text', text: 'checkpoint' },
  730. { type: 'reasoning', text: 'hidden reasoning' },
  731. ]
  732. compact.usage = { inputTokens: 40, outputTokens: 5 }
  733. await compact.compactNow(agent, SIGNAL)
  734. const summary = session.events.find(event => event.type === 'compaction/summary')
  735. expect(summary?.type === 'compaction/summary' && summary.data.rawOutput).toEqual(compact.rawOutput)
  736. expect(summary?.type === 'compaction/summary' && summary.data.usage).toEqual(compact.usage)
  737. })
  738. it('makes duration derivable from the opening and closing marker times', async () => {
  739. const { compact } = detachedService()
  740. const session = closedConversation(2)
  741. const agent = fakeAgent(session, () => () => undefined)
  742. compact.gate = new Promise<undefined>((resolve) => {
  743. setTimeout(() => { resolve(undefined) }, 5)
  744. })
  745. await compact.compactNow(agent, SIGNAL)
  746. const start = session.events.findLast(event => event.type === 'compaction/start')
  747. const end = session.events.findLast(event => event.type === 'compaction/end')
  748. expect(start).toBeDefined()
  749. expect(end).toBeDefined()
  750. expect(end!.time - start!.time).toBeGreaterThan(0)
  751. })
  752. it('excludes concurrent automatic and manual compaction of one session', async () => {
  753. const { compact } = detachedService()
  754. const session = closedConversation(3)
  755. const agent = fakeAgent(session, () => () => undefined)
  756. const gate = deferred()
  757. compact.gate = gate.promise
  758. const manual = compact.compactNow(agent, SIGNAL)
  759. await Promise.resolve()
  760. const nodes = session.surface.nodes
  761. await expect(compact.compactRegion(
  762. nodes[0]!,
  763. nodes[1]!,
  764. agent,
  765. )).rejects.toThrow('compaction lock is already active')
  766. gate.resolve()
  767. compact.gate = undefined
  768. const result: CompactionResult | null = await manual
  769. expect(result).not.toBeNull()
  770. })
  771. it('excludes a manual request while an explicit region compaction runs', async () => {
  772. const { compact } = detachedService()
  773. const session = closedConversation(3)
  774. session.append('turn/start', { turn: 4 })
  775. const agent = fakeAgent(session, () => () => undefined)
  776. const gate = deferred()
  777. compact.gate = gate.promise
  778. const nodes = session.surface.nodes
  779. const region = compact.compactRegion(nodes[0]!, nodes[1]!, agent)
  780. await Promise.resolve()
  781. expect((await rejection(compact.compactNow(agent, SIGNAL))).code).toBe('busy')
  782. gate.resolve()
  783. compact.gate = undefined
  784. await expect(region).resolves.toMatchObject({ shadowedSeqs: nodes.slice(0, 2) })
  785. })
  786. })