api-proxy-cold.spec.ts 33 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801
  1. /**
  2. * Cold-session and degenerate-composition paths of the host ApiProxy:
  3. * metadata-only listing, Agent-free history reads, subagent ownership
  4. * isolation, and prompt failure mapping.
  5. */
  6. import { mkdtempSync, writeFileSync } from 'node:fs'
  7. import { tmpdir } from 'node:os'
  8. import { join } from 'node:path'
  9. import { describe, expect, it, vi } from 'vitest'
  10. import { Context } from '@deepseek-ai/cordis'
  11. import SessionStore from '@deepseek-ai/dsh-session'
  12. import AgentRegistry from '@deepseek-ai/dsh-agent'
  13. import { TypertLookupFailure } from '@deepseek-ai/dsh-typert-protocol'
  14. import TypertRegistry from '@deepseek-ai/dsh-typert-registry'
  15. import { createUserMessage, MessageId } from '@deepseek-ai/dsh-llm'
  16. import type { Agent } from '@deepseek-ai/dsh-agent'
  17. import UserQuestionService from '@deepseek-ai/dsh-user-questions'
  18. import type { SessionEvent, SessionHeader, SessionId } from '@deepseek-ai/dsh-session'
  19. import {
  20. PersistenceCoordinator,
  21. SessionPersistenceRevision,
  22. type PersistenceBackend,
  23. type StoredPrefix,
  24. } from '@deepseek-ai/dsh-session-persistence'
  25. import type { RpcRequest } from '@deepseek-ai/dsh-host-apiproxy/api/rpc'
  26. import { RpcId } from '@deepseek-ai/dsh-host-apiproxy/api/rpc'
  27. import { createApiProxy } from '@deepseek-ai/dsh-host-apiproxy'
  28. const sid = (id: string): SessionId => id as SessionId
  29. let nextRpc = 1
  30. function request<P>(payload: P): RpcRequest<P> {
  31. return { rpcId: RpcId(`cold-${String(nextRpc++)}`), payload }
  32. }
  33. function header(id: string, createdAt: number, extra: Partial<SessionHeader> = {}): SessionHeader {
  34. return { version: 0, id: sid(id), createdAt, cwd: '/proj', ...extra }
  35. }
  36. describe('sessions.list cold merge', () => {
  37. it('verifies only small possibly-blank artifacts and treats every unavailable probe as visible', async () => {
  38. const ctx = new Context()
  39. await ctx.plugin(SessionStore)
  40. await ctx.plugin(UserQuestionService)
  41. const root = mkdtempSync(join(tmpdir(), 'dsh-cold-'))
  42. const smallPath = join(root, 'small.log')
  43. const largePath = join(root, 'large.log')
  44. writeFileSync(smallPath, 'x'.repeat(1024))
  45. writeFileSync(largePath, 'x'.repeat(1025))
  46. const metas = [
  47. header('small-blank', 100),
  48. header('small-conversation', 200),
  49. header('large-unknown', 300),
  50. header('cached-nonblank', 400),
  51. header('locationless', 500, { parentSession: sid('session-parent'), origin: 'subagent' }),
  52. header('vanished', 600),
  53. header('read-failure', 700),
  54. ]
  55. const readFrom = vi.fn(async (id: SessionId) => {
  56. if (id === sid('small-blank')) {
  57. return {
  58. meta: metas[0]!,
  59. events: [{ type: 'session/end-seed', seq: 0, time: 700, data: {} }] as SessionEvent[],
  60. }
  61. }
  62. if (id === sid('small-conversation')) {
  63. return {
  64. meta: metas[1]!,
  65. events: [
  66. { type: 'turn/start', seq: 0, time: 800, data: { turn: 1 } },
  67. {
  68. type: 'user/message', seq: 1, time: 1200,
  69. data: createUserMessage({ content: [{ type: 'text', text: 'worked' }], source: { kind: 'user' } }),
  70. surfaceOp: 'append',
  71. },
  72. ] as SessionEvent[],
  73. }
  74. }
  75. if (id === sid('read-failure')) throw new Error('simulated read failure')
  76. throw new Error(`unexpected cold read: ${id}`)
  77. })
  78. ctx.provide('sessionPersistence', {
  79. list: () => Promise.resolve(metas),
  80. locate: (meta: SessionHeader) => {
  81. if (meta.id === sid('large-unknown')) return { kind: 'jsonl', path: largePath }
  82. if (meta.id === sid('locationless')) return undefined
  83. if (meta.id === sid('vanished')) return { kind: 'jsonl', path: join(root, 'vanished.log') }
  84. return { kind: 'jsonl', path: smallPath }
  85. },
  86. readFrom,
  87. } as never)
  88. ctx.provide('sessionProjectionCache', {
  89. cachedSnapshot: async (meta: SessionHeader) => {
  90. if (meta.id === sid('small-blank')) {
  91. return { asOfSeq: 0, values: { sessionListMetadata: { blank: true, lastPromptAt: null } } }
  92. }
  93. if (meta.id === sid('small-conversation')) {
  94. return { asOfSeq: 0, values: { sessionListMetadata: { blank: true, lastPromptAt: 900 } } }
  95. }
  96. if (meta.id === sid('cached-nonblank')) {
  97. return { asOfSeq: 1, values: { sessionListMetadata: { blank: false, lastPromptAt: 1000 } } }
  98. }
  99. return undefined
  100. },
  101. } as never)
  102. const api = createApiProxy(ctx, { defaultModelSelection: () => ({ provider: 'p', model: 'm' }), cwd: '/tmp' })
  103. const response = await api.sessions.list(request({}))
  104. expect(response.result.ok).toBe(true)
  105. if (!response.result.ok) throw new Error('unreachable')
  106. const byId = Object.fromEntries(response.result.value.items.map(item => [item.sessionId, item]))
  107. expect(byId['small-blank']).toMatchObject({ blank: true, updatedAt: 100, running: false })
  108. // A stale true hint cannot hide the turn found in the bounded read.
  109. expect(byId['small-conversation']).toMatchObject({ blank: false, updatedAt: 1200 })
  110. expect(byId['large-unknown']).toMatchObject({ blank: false, updatedAt: 300 })
  111. // false is monotonic, so this row skips stat/read and keeps cached recency.
  112. expect(byId['cached-nonblank']).toMatchObject({ blank: false, updatedAt: 1000 })
  113. expect(byId['locationless']).toMatchObject({
  114. blank: false,
  115. updatedAt: 500,
  116. parentSessionId: 'session-parent',
  117. origin: 'subagent',
  118. })
  119. expect(byId['vanished']).toMatchObject({ blank: false, updatedAt: 600 })
  120. expect(byId['read-failure']).toMatchObject({ blank: false, updatedAt: 700 })
  121. expect(readFrom).toHaveBeenCalledTimes(3)
  122. expect(readFrom.mock.calls.map(([id]) => id)).toEqual(expect.arrayContaining([
  123. sid('small-blank'),
  124. sid('small-conversation'),
  125. sid('read-failure'),
  126. ]))
  127. })
  128. it('can disable bounded blank probes without hiding cold Sessions', async () => {
  129. const ctx = new Context()
  130. await ctx.plugin(SessionStore)
  131. await ctx.plugin(UserQuestionService)
  132. const meta = header('probe-disabled', 100)
  133. const readFrom = vi.fn()
  134. ctx.provide('sessionPersistence', {
  135. list: () => Promise.resolve([meta]),
  136. locate: () => ({ kind: 'jsonl', path: '/not-read' }),
  137. readFrom,
  138. } as never)
  139. const api = createApiProxy(ctx, {
  140. defaultModelSelection: () => ({ provider: 'p', model: 'm' }),
  141. cwd: '/tmp',
  142. coldBlankProbeMaxBytes: 0,
  143. })
  144. const response = await api.sessions.list(request({}))
  145. if (!response.result.ok) throw new Error('unreachable')
  146. expect(response.result.value.items).toEqual([
  147. expect.objectContaining({ sessionId: meta.id, blank: false, updatedAt: meta.createdAt }),
  148. ])
  149. expect(readFrom).not.toHaveBeenCalled()
  150. })
  151. it('replaces a probed cold row with the live Session that attached during the read', async () => {
  152. const ctx = new Context()
  153. await ctx.plugin(SessionStore)
  154. await ctx.plugin(UserQuestionService)
  155. await ctx.plugin(AgentRegistry)
  156. const meta = header('attached-during-probe', 100)
  157. const root = mkdtempSync(join(tmpdir(), 'dsh-cold-race-'))
  158. const path = join(root, 'small.log')
  159. writeFileSync(path, 'x')
  160. const started = Promise.withResolvers<undefined>()
  161. const release = Promise.withResolvers<undefined>()
  162. ctx.provide('sessionPersistence', {
  163. list: () => Promise.resolve([meta]),
  164. locate: () => ({ kind: 'jsonl', path }),
  165. readFrom: async () => {
  166. started.resolve(undefined)
  167. await release.promise
  168. return {
  169. meta,
  170. events: [{ type: 'session/end-seed', seq: 0, time: 110, data: {} }] as SessionEvent[],
  171. }
  172. },
  173. } as never)
  174. const api = createApiProxy(ctx, { defaultModelSelection: () => ({ provider: 'p', model: 'm' }), cwd: '/tmp' })
  175. const listing = api.sessions.list(request({}))
  176. await started.promise
  177. const session = ctx.sessions.create(meta.id, {
  178. seed: [
  179. { type: 'turn/start', seq: 0, time: 200, data: { turn: 1 } },
  180. {
  181. type: 'user/message', seq: 1, time: 300,
  182. data: createUserMessage({ content: [{ type: 'text', text: 'live' }], source: { kind: 'user' } }),
  183. surfaceOp: 'append',
  184. },
  185. ],
  186. meta: {
  187. ...meta.cwd === undefined ? {} : { cwd: meta.cwd },
  188. createdAt: meta.createdAt,
  189. },
  190. })
  191. ctx.agents.register({ id: session.id, session, status: 'running', ctx } as Agent)
  192. release.resolve(undefined)
  193. const response = await listing
  194. if (!response.result.ok) throw new Error('list failed')
  195. expect(response.result.value.items).toEqual([
  196. expect.objectContaining({
  197. sessionId: meta.id,
  198. blank: false,
  199. running: true,
  200. updatedAt: 300,
  201. }),
  202. ])
  203. })
  204. })
  205. describe('attached updatedAt tracks human prompts', () => {
  206. it('ignores pickup and non-prompt work after the latest human message', async () => {
  207. const ctx = new Context()
  208. await ctx.plugin(SessionStore)
  209. await ctx.plugin(UserQuestionService)
  210. await ctx.plugin(AgentRegistry)
  211. const api = createApiProxy(ctx, { defaultModelSelection: () => ({ provider: 'p', model: 'm' }), cwd: '/tmp' })
  212. // Old work, resumed just now: the log tail would report the pickup.
  213. const worked = 1_000_000
  214. const resumed = ctx.sessions.create(sid('resumed-untouched'), {
  215. seed: [
  216. { type: 'turn/start', seq: 0, time: worked, data: { turn: 1 } },
  217. {
  218. type: 'user/message', seq: 1, time: worked,
  219. data: createUserMessage({ content: [{ type: 'text', text: 'worked' }], source: { kind: 'user' } }),
  220. surfaceOp: 'append',
  221. },
  222. { type: 'turn/end', seq: 2, time: worked + 1, data: { turn: 1, reason: { kind: 'completed' } } },
  223. ],
  224. meta: { cwd: '/proj', createdAt: 500 },
  225. })
  226. ctx.agents.register({ id: resumed.id, session: resumed, status: 'idle', ctx } as Agent)
  227. const boundary = resumed.events.at(-1)
  228. expect(boundary?.type).toBe('session/end-seed')
  229. expect(boundary?.time).toBeGreaterThan(worked)
  230. const listed = await api.sessions.list(request({}))
  231. if (!listed.result.ok) throw new Error('list failed')
  232. const summary = listed.result.value.items.find(item => item.sessionId === 'resumed-untouched')
  233. expect(summary?.updatedAt).toBe(worked)
  234. // A lifecycle boundary is not a human update.
  235. resumed.append('turn/start', { turn: 2 })
  236. const afterBoundary = await api.sessions.list(request({}))
  237. if (!afterBoundary.result.ok) throw new Error('list failed')
  238. expect(afterBoundary.result.value.items.find(item => item.sessionId === 'resumed-untouched')?.updatedAt)
  239. .toBe(worked)
  240. const prompt = resumed.append('user/message', createUserMessage({
  241. content: [{ type: 'text', text: 'new prompt' }],
  242. source: { kind: 'user' },
  243. }), { surfaceOp: 'append' })
  244. const after = await api.sessions.list(request({}))
  245. if (!after.result.ok) throw new Error('list failed')
  246. const moved = after.result.value.items.find(item => item.sessionId === 'resumed-untouched')
  247. expect(moved?.updatedAt).toBe(prompt.time)
  248. })
  249. })
  250. describe('cold history recovery view', () => {
  251. it('shows in-memory interruption repair without activating the session', async () => {
  252. const ctx = new Context()
  253. await ctx.plugin(SessionStore)
  254. await ctx.plugin(UserQuestionService)
  255. const sessionId = sid('session-interrupted')
  256. const meta = header(sessionId, 1000)
  257. const stored: StoredPrefix<never> = {
  258. meta,
  259. events: [{ type: 'turn/start', seq: 0, time: 1, data: { turn: 1 } }],
  260. revision: SessionPersistenceRevision('history-recovery-test:1'),
  261. }
  262. const backend: PersistenceBackend<never> = {
  263. name: 'history-recovery-test',
  264. loadStored: id => Promise.resolve(id === sessionId ? structuredClone(stored) : undefined),
  265. readStoredRevision: id => Promise.resolve(
  266. id === sessionId ? SessionPersistenceRevision('history-recovery-test:1') : undefined,
  267. ),
  268. appendBatch: () => Promise.resolve(),
  269. commitRepair: () => Promise.resolve(),
  270. list: () => Promise.resolve([structuredClone(meta)]),
  271. }
  272. const coordinator = new PersistenceCoordinator(ctx, backend)
  273. ctx.provide('sessionPersistence', {
  274. list: (signal?: AbortSignal) => backend.list(signal),
  275. inspect: (id: SessionId, signal?: AbortSignal) => coordinator.inspect(id, signal),
  276. locate: () => undefined,
  277. } as never)
  278. const api = createApiProxy(ctx, { defaultModelSelection: () => ({ provider: 'p', model: 'm' }), cwd: '/tmp' })
  279. const history = await api.sessions.history(request({ sessionId, beforeSeq: 2, maxMessages: 10 }))
  280. if (!history.result.ok) throw new Error('history failed')
  281. expect(history.result.value.events.map(entry => entry.event)).toMatchInlineSnapshot(`
  282. [
  283. {
  284. "data": {
  285. "turn": 1,
  286. },
  287. "seq": 0,
  288. "time": 1,
  289. "type": "turn/start",
  290. },
  291. {
  292. "data": {
  293. "reason": {
  294. "kind": "interrupted",
  295. },
  296. "turn": 1,
  297. },
  298. "seq": 1,
  299. "time": 1,
  300. "type": "turn/end",
  301. },
  302. ]
  303. `)
  304. expect(ctx.sessions.get(sessionId)).toBeUndefined()
  305. await ctx.fiber.dispose()
  306. })
  307. })
  308. describe('Remote Agent and Session lookup policy', () => {
  309. it('deduplicates a cold resume across Agent and Session parameters', async () => {
  310. const ctx = new Context()
  311. await ctx.plugin(TypertRegistry)
  312. await ctx.plugin(SessionStore)
  313. await ctx.plugin(AgentRegistry)
  314. await ctx.plugin(UserQuestionService)
  315. const sessionId = sid('session-remote-cold')
  316. const meta = header(sessionId, 1000)
  317. const inspect = vi.fn(() => Promise.resolve({ meta, events: [] as SessionEvent[] }))
  318. ctx.provide('sessionPersistence', {
  319. list: () => Promise.resolve([meta]),
  320. inspect,
  321. locate: () => undefined,
  322. } as never)
  323. const resumedSession = { id: sessionId, header: meta, events: [] } as unknown as import('@deepseek-ai/dsh-session').Session
  324. const resumedAgent = { id: sessionId, session: resumedSession, status: 'idle', ctx } as Agent
  325. const release = Promise.withResolvers<undefined>()
  326. const resume = vi.spyOn(ctx.agents, 'resume').mockImplementation(async () => {
  327. await release.promise
  328. return { agent: resumedAgent, dispose: () => Promise.resolve() }
  329. })
  330. const defaultAgentLookup = ctx.typert.lookups.get('agent')
  331. const defaultSessionLookup = ctx.typert.lookups.get('session')
  332. createApiProxy(ctx, { defaultModelSelection: () => ({ provider: 'p', model: 'm' }), cwd: '/tmp' })
  333. await vi.waitFor(() => {
  334. expect(ctx.typert.lookups.get('agent')).not.toBe(defaultAgentLookup)
  335. expect(ctx.typert.lookups.get('session')).not.toBe(defaultSessionLookup)
  336. })
  337. const agentLookup = ctx.typert.lookups.get('agent')
  338. const sessionLookup = ctx.typert.lookups.get('session')
  339. if (agentLookup === undefined || sessionLookup === undefined) throw new Error('core lookup providers were not mounted')
  340. const resolvedAgent = Promise.resolve(agentLookup.resolve(sessionId))
  341. const resolvedSession = Promise.resolve(sessionLookup.resolve(sessionId))
  342. await vi.waitFor(() => { expect(resume).toHaveBeenCalledOnce() })
  343. release.resolve(undefined)
  344. await expect(resolvedAgent).resolves.toBe(resumedAgent)
  345. await expect(resolvedSession).resolves.toBe(resumedSession)
  346. expect(inspect).toHaveBeenCalledOnce()
  347. })
  348. it('preserves the subagent ownership fence for cold and live Remote lookups', async () => {
  349. const ctx = new Context()
  350. await ctx.plugin(TypertRegistry)
  351. await ctx.plugin(SessionStore)
  352. await ctx.plugin(AgentRegistry)
  353. await ctx.plugin(UserQuestionService)
  354. const coldId = sid('session-remote-cold-child')
  355. const coldMeta = header(coldId, 1000, {
  356. parentSession: sid('session-parent'),
  357. origin: 'subagent',
  358. })
  359. const inspect = vi.fn(() => Promise.resolve({ meta: coldMeta, events: [] as SessionEvent[] }))
  360. ctx.provide('sessionPersistence', {
  361. list: () => Promise.resolve([coldMeta]),
  362. inspect,
  363. locate: () => undefined,
  364. } as never)
  365. const liveSession = ctx.sessions.create(sid('session-remote-live-child'), {
  366. meta: { cwd: '/proj', parentSession: sid('session-parent'), origin: 'subagent' },
  367. })
  368. const liveAgent = { id: liveSession.id, session: liveSession, status: 'idle', ctx } as Agent
  369. ctx.agents.register(liveAgent)
  370. const resume = vi.spyOn(ctx.agents, 'resume')
  371. const defaultAgentLookup = ctx.typert.lookups.get('agent')
  372. const defaultSessionLookup = ctx.typert.lookups.get('session')
  373. createApiProxy(ctx, { defaultModelSelection: () => ({ provider: 'p', model: 'm' }), cwd: '/tmp' })
  374. await vi.waitFor(() => {
  375. expect(ctx.typert.lookups.get('agent')).not.toBe(defaultAgentLookup)
  376. expect(ctx.typert.lookups.get('session')).not.toBe(defaultSessionLookup)
  377. })
  378. const agentLookup = ctx.typert.lookups.get('agent')
  379. const sessionLookup = ctx.typert.lookups.get('session')
  380. if (agentLookup === undefined || sessionLookup === undefined) throw new Error('core lookup providers were not mounted')
  381. const ownershipFailure = {
  382. failure: {
  383. code: 'agent-busy',
  384. details: { reason: 'use subagent delivery for this child session' },
  385. },
  386. }
  387. const coldFailure = Promise.resolve(agentLookup.resolve(coldId))
  388. const liveFailure = Promise.resolve(sessionLookup.resolve(liveSession.id))
  389. await expect(coldFailure).rejects.toBeInstanceOf(TypertLookupFailure)
  390. await expect(coldFailure).rejects.toMatchObject(ownershipFailure)
  391. await expect(liveFailure).rejects.toBeInstanceOf(TypertLookupFailure)
  392. await expect(liveFailure).rejects.toMatchObject(ownershipFailure)
  393. expect(resume).not.toHaveBeenCalled()
  394. expect(inspect).toHaveBeenCalledOnce()
  395. })
  396. })
  397. describe('subagent ownership fence', () => {
  398. it('reads a cold child without an Agent and rejects generic resume or adoption', async () => {
  399. const ctx = new Context()
  400. await ctx.plugin(SessionStore)
  401. await ctx.plugin(AgentRegistry)
  402. await ctx.plugin(UserQuestionService)
  403. const sessionId = sid('session-child')
  404. const meta = header('session-child', 1000, {
  405. parentSession: sid('session-parent'),
  406. seedLength: 0,
  407. origin: 'subagent',
  408. })
  409. const events = [
  410. { type: 'turn/start', seq: 0, time: 1, data: { turn: 1, trigger: { kind: 'message', source: { kind: 'user' } } } },
  411. {
  412. type: 'user/message',
  413. seq: 1,
  414. time: 2,
  415. data: { content: [{ type: 'text', text: 'work' }], source: { kind: 'user' } },
  416. surfaceOp: 'append',
  417. },
  418. {
  419. type: 'subagent/descriptor',
  420. seq: 2,
  421. time: 3,
  422. data: { version: 2, mode: 'continuable', provider: 'spawn', label: 'child' },
  423. },
  424. { type: 'turn/end', seq: 3, time: 4, data: { turn: 1, reason: { kind: 'completed' } } },
  425. ] as SessionEvent[]
  426. const inspect = vi.fn(() => Promise.resolve({ meta, events }))
  427. ctx.provide('sessionPersistence', {
  428. list: () => Promise.resolve([meta]),
  429. inspect,
  430. locate: () => undefined,
  431. } as never)
  432. const resume = vi.spyOn(ctx.agents, 'resume')
  433. const api = createApiProxy(ctx, { defaultModelSelection: () => ({ provider: 'p', model: 'm' }), cwd: '/tmp' })
  434. const history = await api.sessions.history(request({ sessionId }))
  435. expect(history.result.ok).toBe(true)
  436. if (history.result.ok) {
  437. expect(history.result.value.events.map(entry => entry.event.type)).toEqual(events.map(event => event.type))
  438. }
  439. expect(ctx.agents.get(sessionId)).toBeUndefined()
  440. const prompt = await api.sessions.prompt(request({
  441. sessionId,
  442. mode: 'queue',
  443. content: [{ type: 'text', text: 'follow up' }],
  444. }))
  445. expect(prompt.result.ok).toBe(false)
  446. if (!prompt.result.ok) {
  447. expect(prompt.result.error).toMatchObject({
  448. code: 'agent-busy',
  449. details: { reason: 'use subagent delivery for this child session' },
  450. })
  451. }
  452. const create = await api.sessions.create(request({ sessionId, cwd: '/proj' }))
  453. expect(create.result.ok).toBe(false)
  454. if (!create.result.ok) expect(create.result.error.code).toBe('agent-busy')
  455. expect(resume).not.toHaveBeenCalled()
  456. expect(ctx.agents.get(sessionId)).toBeUndefined()
  457. expect(inspect).toHaveBeenCalledTimes(3)
  458. })
  459. it('no longer treats a descriptor-only cold child without origin as subagent-owned', async () => {
  460. const ctx = new Context()
  461. await ctx.plugin(SessionStore)
  462. await ctx.plugin(AgentRegistry)
  463. await ctx.plugin(UserQuestionService)
  464. const sessionId = sid('session-legacy-child')
  465. const meta = header('session-legacy-child', 1000, {
  466. parentSession: sid('session-parent'),
  467. seedLength: 0,
  468. })
  469. const events = [
  470. {
  471. type: 'subagent/descriptor',
  472. seq: 0,
  473. time: 1,
  474. data: { version: 2, mode: 'continuable', provider: 'spawn', label: 'child' },
  475. },
  476. ] as SessionEvent[]
  477. ctx.provide('sessionPersistence', {
  478. list: () => Promise.resolve([meta]),
  479. inspect: () => Promise.resolve({ meta, events }),
  480. locate: () => undefined,
  481. } as never)
  482. // Stores whose headers predate `origin` classify a child only through the
  483. // descriptor event; the pre-release decision stops recognizing them, so
  484. // the ownership fence lets generic resume reach the registry instead of
  485. // answering `agent-busy`.
  486. const resume = vi.spyOn(ctx.agents, 'resume')
  487. .mockRejectedValue(new Error('registry unavailable in this bench'))
  488. const api = createApiProxy(ctx, { defaultModelSelection: () => ({ provider: 'p', model: 'm' }), cwd: '/tmp' })
  489. const prompt = await api.sessions.prompt(request({
  490. sessionId,
  491. mode: 'queue',
  492. content: [{ type: 'text', text: 'follow up' }],
  493. }))
  494. expect(resume).toHaveBeenCalledTimes(1)
  495. expect(prompt.result.ok).toBe(false)
  496. if (!prompt.result.ok) expect(prompt.result.error.code).toBe('internal')
  497. })
  498. it('rejects origin-marked and runtime-owned live children from generic controls', async () => {
  499. const ctx = new Context()
  500. await ctx.plugin(SessionStore)
  501. await ctx.plugin(AgentRegistry)
  502. await ctx.plugin(UserQuestionService)
  503. const parentSession = ctx.sessions.create(sid('session-parent'), { meta: { cwd: '/proj' } })
  504. const parent = { id: parentSession.id, session: parentSession, status: 'idle', ctx } as Agent
  505. ctx.agents.register(parent)
  506. const originSession = ctx.sessions.create(sid('session-origin-child'), {
  507. meta: { cwd: '/proj', parentSession: parent.id, origin: 'subagent' },
  508. })
  509. const cancel = vi.fn()
  510. const updateInbox = vi.fn(() => 'applied' as const)
  511. const originChild = {
  512. id: originSession.id,
  513. session: originSession,
  514. status: 'idle',
  515. ctx,
  516. cancel,
  517. updateInbox,
  518. } as unknown as Agent
  519. ctx.agents.register(originChild)
  520. const startingSession = ctx.sessions.create(sid('session-starting-child'), {
  521. meta: { cwd: '/proj', parentSession: parent.id },
  522. })
  523. const startingChild = { id: startingSession.id, session: startingSession, status: 'idle', ctx } as Agent
  524. ctx.agents.enter(startingChild, parent)
  525. const api = createApiProxy(ctx, { defaultModelSelection: () => ({ provider: 'p', model: 'm' }), cwd: '/tmp' })
  526. const stopped = await api.sessions.cancel(request({ sessionId: originChild.id }))
  527. expect(stopped.result.ok).toBe(false)
  528. if (!stopped.result.ok) expect(stopped.result.error.code).toBe('agent-busy')
  529. expect(cancel).not.toHaveBeenCalled()
  530. const queued = await api.sessions.updateQueue(request({
  531. sessionId: originChild.id,
  532. itemId: MessageId('queued-item'),
  533. action: { kind: 'remove' },
  534. }))
  535. expect(queued.result.ok).toBe(false)
  536. if (!queued.result.ok) expect(queued.result.error.code).toBe('agent-busy')
  537. expect(updateInbox).not.toHaveBeenCalled()
  538. const models = await api.sessions.models(request({ sessionId: startingChild.id }))
  539. expect(models.result.ok).toBe(false)
  540. if (!models.result.ok) expect(models.result.error.code).toBe('agent-busy')
  541. const create = await api.sessions.create(request({ sessionId: originChild.id, cwd: '/proj' }))
  542. expect(create.result.ok).toBe(false)
  543. if (!create.result.ok) expect(create.result.error.code).toBe('agent-busy')
  544. const history = await api.sessions.history(request({ sessionId: originChild.id }))
  545. expect(history.result.ok).toBe(true)
  546. expect(ctx.agents.get(originChild.id)).toBe(originChild)
  547. })
  548. it('does not classify an ordinary fork from an inherited ancestor descriptor', async () => {
  549. const ctx = new Context()
  550. await ctx.plugin(SessionStore)
  551. await ctx.plugin(AgentRegistry)
  552. await ctx.plugin(UserQuestionService)
  553. const session = ctx.sessions.create(sid('session-ordinary-fork'), {
  554. seed: [{
  555. type: 'subagent/descriptor',
  556. seq: 0,
  557. time: 1,
  558. data: { version: 2, mode: 'continuable', provider: 'spawn', label: 'ancestor' },
  559. }],
  560. meta: { cwd: '/proj', parentSession: sid('session-source'), seedLength: 1 },
  561. })
  562. const followup = vi.fn()
  563. const agent = { id: session.id, session, status: 'idle', ctx, followup } as unknown as Agent
  564. ctx.agents.register(agent)
  565. const api = createApiProxy(ctx, { defaultModelSelection: () => ({ provider: 'p', model: 'm' }), cwd: '/tmp' })
  566. const response = await api.sessions.prompt(request({
  567. sessionId: agent.id,
  568. mode: 'queue',
  569. content: [{ type: 'text', text: 'ordinary work' }],
  570. }))
  571. expect(response.result.ok).toBe(true)
  572. expect(followup).toHaveBeenCalledOnce()
  573. })
  574. it('canonicalizes a supplied browser zone on the exact prompt and rejects invalid names', async () => {
  575. const ctx = new Context()
  576. await ctx.plugin(SessionStore)
  577. await ctx.plugin(AgentRegistry)
  578. await ctx.plugin(UserQuestionService)
  579. const session = ctx.sessions.create(sid('session-browser-zone'), { meta: { cwd: '/proj' } })
  580. const followup = vi.fn()
  581. const agent = { id: session.id, session, status: 'idle', ctx, followup } as unknown as Agent
  582. ctx.agents.register(agent)
  583. const api = createApiProxy(ctx, {
  584. defaultModelSelection: () => ({ provider: 'p', model: 'm' }),
  585. cwd: '/tmp',
  586. })
  587. const alias = 'US/Pacific'
  588. const canonical = new Intl.DateTimeFormat('en-US', { timeZone: alias })
  589. .resolvedOptions().timeZone
  590. const zonedRequest = request({
  591. sessionId: agent.id,
  592. mode: 'queue' as const,
  593. content: [{ type: 'text' as const, text: 'zoned work' }],
  594. clientTimeZone: alias,
  595. })
  596. await expect(api.sessions.prompt(zonedRequest)).resolves.toMatchObject({
  597. result: { ok: true },
  598. })
  599. expect(followup).toHaveBeenNthCalledWith(1, expect.objectContaining({
  600. source: { kind: 'user', rpcId: zonedRequest.rpcId, clientTimeZone: canonical },
  601. }))
  602. const utcRequest = request({
  603. sessionId: agent.id,
  604. mode: 'queue' as const,
  605. content: [{ type: 'text' as const, text: 'UTC work' }],
  606. clientTimeZone: 'UTC',
  607. })
  608. await expect(api.sessions.prompt(utcRequest)).resolves.toMatchObject({
  609. result: { ok: true },
  610. })
  611. expect(followup).toHaveBeenNthCalledWith(2, expect.objectContaining({
  612. source: { kind: 'user', rpcId: utcRequest.rpcId, clientTimeZone: 'UTC' },
  613. }))
  614. const unzonedRequest = request({
  615. sessionId: agent.id,
  616. mode: 'queue' as const,
  617. content: [{ type: 'text' as const, text: 'headless work' }],
  618. })
  619. await expect(api.sessions.prompt(unzonedRequest)).resolves.toMatchObject({
  620. result: { ok: true },
  621. })
  622. expect(followup).toHaveBeenNthCalledWith(3, expect.objectContaining({
  623. source: { kind: 'user', rpcId: unzonedRequest.rpcId },
  624. }))
  625. for (const clientTimeZone of ['', ' UTC', 'CST', 'Not/A_Real_Zone']) {
  626. const invalid = await api.sessions.prompt(request({
  627. sessionId: agent.id,
  628. mode: 'queue' as const,
  629. content: [{ type: 'text' as const, text: 'invalid zone' }],
  630. clientTimeZone,
  631. }))
  632. expect(invalid.result).toEqual({
  633. ok: false,
  634. error: {
  635. code: 'invalid-time-zone',
  636. message: 'clientTimeZone must be UTC or a valid IANA Area/Location name',
  637. details: { value: clientTimeZone },
  638. },
  639. })
  640. }
  641. expect(followup).toHaveBeenCalledTimes(3)
  642. })
  643. })
  644. describe('degenerate composition (no persistence, no factory)', () => {
  645. it('list skips the cold merge and history reports missing persistence as internal', async () => {
  646. const ctx = new Context()
  647. await ctx.plugin(SessionStore)
  648. await ctx.plugin(AgentRegistry)
  649. await ctx.plugin(UserQuestionService)
  650. const api = createApiProxy(ctx, { defaultModelSelection: () => ({ provider: 'p', model: 'm' }), cwd: '/tmp' })
  651. const listed = await api.sessions.list(request({}))
  652. expect(listed.result.ok).toBe(true)
  653. if (listed.result.ok) expect(listed.result.value.items).toEqual([])
  654. // No persistence means cold history cannot inspect a transcript.
  655. const response = await api.sessions.history(request({ sessionId: sid('session-ghost') }))
  656. expect(response.result.ok).toBe(false)
  657. if (!response.result.ok) {
  658. expect(response.result.error.code).toBe('internal')
  659. expect(response.result.error.message).toMatch(/history unavailable for session "session-ghost"/)
  660. }
  661. })
  662. it('maps a persistence catalog miss to session-not-found without inspection', async () => {
  663. const ctx = new Context()
  664. await ctx.plugin(SessionStore)
  665. await ctx.plugin(AgentRegistry)
  666. await ctx.plugin(UserQuestionService)
  667. const inspect = vi.fn()
  668. ctx.provide('sessionPersistence', {
  669. list: () => Promise.resolve([]),
  670. inspect,
  671. } as never)
  672. const api = createApiProxy(ctx, { defaultModelSelection: () => ({ provider: 'p', model: 'm' }), cwd: '/tmp' })
  673. const response = await api.sessions.history(request({ sessionId: sid('session-missing') }))
  674. expect(response.result.ok).toBe(false)
  675. if (!response.result.ok) expect(response.result.error.code).toBe('session-not-found')
  676. expect(inspect).not.toHaveBeenCalled()
  677. })
  678. })
  679. describe('sessions.prompt synchronous rejection', () => {
  680. it('maps a synchronous send throw (disposed/invalid input) to agent-busy with the reason attached', async () => {
  681. const ctx = new Context()
  682. await ctx.plugin(SessionStore)
  683. await ctx.plugin(AgentRegistry)
  684. await ctx.plugin(UserQuestionService)
  685. const session = ctx.sessions.create(sid('session-throwing'))
  686. // A live structural stub whose delivery verbs throw synchronously, the
  687. // shape a disposed loop presents at this gateway boundary.
  688. ctx.agents.register({
  689. id: session.id,
  690. session,
  691. status: 'idle',
  692. ctx,
  693. followup: () => { throw new Error('agent "session-throwing" lifecycle disposed') },
  694. steer: () => { throw new Error('agent "session-throwing" lifecycle disposed') },
  695. } as unknown as Agent)
  696. const api = createApiProxy(ctx, { defaultModelSelection: () => ({ provider: 'p', model: 'm' }), cwd: '/tmp' })
  697. for (const mode of ['queue', 'steer'] as const) {
  698. const response = await api.sessions.prompt(request({
  699. sessionId: session.id, mode, content: [{ type: 'text' as const, text: 'x' }],
  700. }))
  701. expect(response.result.ok).toBe(false)
  702. if (!response.result.ok) {
  703. expect(response.result.error.code).toBe('agent-busy')
  704. expect(response.result.error.message).toBe('prompt rejected')
  705. expect(response.result.error.details).toEqual({
  706. reason: 'Error: agent "session-throwing" lifecycle disposed',
  707. })
  708. }
  709. }
  710. })
  711. it('classifies a raced cold-resume ID collision as agent-busy', async () => {
  712. const ctx = new Context()
  713. await ctx.plugin(SessionStore)
  714. await ctx.plugin(AgentRegistry)
  715. await ctx.plugin(UserQuestionService)
  716. const sessionId = sid('race-resume')
  717. const meta: SessionHeader = header('race-resume', 1000)
  718. ctx.provide('sessionPersistence', {
  719. list: () => Promise.resolve([meta]),
  720. inspect: () => Promise.resolve({ meta, events: [] as SessionEvent[] }),
  721. locate: () => undefined,
  722. } as never)
  723. // The raced winner: a live parent-owned subagent publishes the identity
  724. // while the generic cold resume is in flight, so the resume collides.
  725. const parentSession = ctx.sessions.create(sid('race-parent'), { meta: { cwd: '/proj' } })
  726. const parent = { id: parentSession.id, session: parentSession, status: 'idle', ctx } as Agent
  727. ctx.agents.register(parent)
  728. const childSession = ctx.sessions.create(sessionId, {
  729. meta: { cwd: '/proj', parentSession: parent.id, origin: 'subagent' },
  730. })
  731. const child = { id: sessionId, session: childSession, status: 'idle', ctx } as unknown as Agent
  732. vi.spyOn(ctx.agents, 'resume').mockImplementationOnce(async () => {
  733. // The parent's `enter()` wins the identity between the pre-resume
  734. // re-check and publication; the generic resume then collides.
  735. ctx.agents.register(child)
  736. throw new Error('session id already published')
  737. })
  738. const api = createApiProxy(ctx, { defaultModelSelection: () => ({ provider: 'p', model: 'm' }), cwd: '/tmp' })
  739. const models = await api.sessions.models(request({ sessionId }))
  740. expect(models.result.ok).toBe(false)
  741. if (!models.result.ok) {
  742. expect(models.result.error).toMatchObject({
  743. code: 'agent-busy',
  744. details: { reason: 'use subagent delivery for this child session' },
  745. })
  746. }
  747. })
  748. })