subagent-inprocess.spec.ts 12 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248
  1. import { createUserMessage } from '@deepseek-ai/dsh-llm'
  2. import { describe, expect, it } from 'vitest'
  3. import { Context } from 'cordis'
  4. import { type Agent, type AgentOptions } from '@deepseek-ai/dsh-agent'
  5. import { SessionId } from '@deepseek-ai/dsh-session'
  6. import AgentLoop from '@deepseek-ai/dsh-agent-loop'
  7. import { mountAgentLoopTestDependencies } from '@deepseek-ai/dsh-agent-loop-testkit'
  8. import InvariantService from '@deepseek-ai/dsh-invariants'
  9. import * as SessionInvariant from '@deepseek-ai/dsh-session/invariant'
  10. import * as AgentInvariant from '@deepseek-ai/dsh-agent/invariant'
  11. import * as AgentLoopInvariant from '@deepseek-ai/dsh-agent-loop/invariant'
  12. import SubagentService from '@deepseek-ai/dsh-subagent'
  13. import { maxTokensResponse, MockAdapter, textResponse } from '../../../core/agent-loop/tests/mock-adapter.ts'
  14. import { startInProcessRun } from '../src/index.ts'
  15. type Script = ConstructorParameters<typeof MockAdapter>[0]
  16. async function mountInvariants(ctx: Context): Promise<void> {
  17. await ctx.plugin(InvariantService)
  18. await ctx.plugin(SessionInvariant)
  19. await ctx.plugin(AgentInvariant)
  20. await ctx.plugin(AgentLoopInvariant)
  21. }
  22. async function setup(script: Script, parentOptions: Partial<AgentOptions> = {}) {
  23. const ctx = new Context()
  24. await mountAgentLoopTestDependencies(ctx)
  25. await mountInvariants(ctx)
  26. await ctx.plugin(AgentLoop, { agents: [] })
  27. await ctx.plugin(SubagentService)
  28. const adapter = new MockAdapter(script)
  29. ctx.llm.registerAdapter(['mock'], adapter)
  30. const parent = ctx.agentLoop.create(SessionId('parent'), { provider: 'mock', model: 'mock', ...parentOptions })
  31. return { ctx, parent, adapter }
  32. }
  33. function request(parent: Agent, signal = new AbortController().signal) {
  34. return { prompt: [{ type: 'text' as const, text: 'child task' }], parent, signal }
  35. }
  36. function text(blocks: readonly { type: string; text?: string }[]): string {
  37. return blocks.filter(block => block.type === 'text').map(block => block.text).join('')
  38. }
  39. describe('startInProcessRun', () => {
  40. it('returns only after publication, drives a fresh child, and disposes it', async () => {
  41. const { ctx, parent } = await setup([textResponse('driver answer')])
  42. const run = await startInProcessRun(request(parent), {})
  43. expect(ctx.agents.get(run.id)).toBeDefined()
  44. const result = await run.result
  45. expect(result.stopReason).toBe('completed')
  46. expect(text(result.output)).toBe('driver answer')
  47. expect(ctx.agents.get(run.id)!.options.subagentDepth).toBe(1)
  48. await run.dispose()
  49. await run.dispose()
  50. expect(ctx.agents.get(run.id)).toBeUndefined()
  51. })
  52. it('reports the message-turn outcome when a later non-message turn completes during flush', async () => {
  53. const { ctx, parent } = await setup([maxTokensResponse('partial answer')])
  54. let injected = false
  55. ctx.on('session/flush', (session) => {
  56. if (injected || session.header.parentSession === undefined) return
  57. const lastEnd = session.events.findLast(event => event.type === 'turn/end')
  58. if (lastEnd?.type !== 'turn/end' || lastEnd.data.reason.kind !== 'max-tokens') return
  59. injected = true
  60. const turn = lastEnd.data.turn + 1
  61. session.append('turn/start', {
  62. turn,
  63. trigger: { kind: 'injection', source: { kind: 'plugin', plugin: 'late-metadata' } },
  64. })
  65. session.append('user/message', createUserMessage({
  66. content: [{ type: 'text', text: 'late metadata' }],
  67. source: { kind: 'plugin', plugin: 'late-metadata' },
  68. }), { surfaceOp: 'append' })
  69. session.append('turn/end', { turn, reason: { kind: 'completed' } })
  70. })
  71. const run = await startInProcessRun(request(parent), {})
  72. const result = await run.result
  73. const child = ctx.agents.get(run.id)!
  74. expect(injected).toBe(true)
  75. expect(child.session.events.findLast(event => event.type === 'turn/end'))
  76. .toMatchObject({ data: { reason: { kind: 'completed' } } })
  77. expect(result.stopReason).toBe('max-tokens')
  78. await run.dispose()
  79. })
  80. it('seeds a forked child but reads only the child-owned output', async () => {
  81. const { ctx, parent } = await setup([textResponse('parent answer'), textResponse('child answer')])
  82. parent.followup(createUserMessage({ content: [{ type: 'text', text: 'parent question' }], source: { kind: 'user' } }))
  83. await parent.whenIdle()
  84. const seed = parent.session.events.slice()
  85. const run = await startInProcessRun(request(parent), { seed })
  86. const result = await run.result
  87. expect(text(result.output)).toBe('child answer')
  88. const child = ctx.agents.get(run.id)!
  89. expect(child.session.header.seedLength).toBe(seed.length)
  90. expect(child.session.events.slice(0, seed.length)).toEqual(seed)
  91. await run.dispose()
  92. })
  93. it('persists the child depth in its session header', async () => {
  94. const { ctx, parent } = await setup([textResponse('child answer')])
  95. const run = await startInProcessRun(request(parent), {})
  96. await run.result
  97. // The recursion budget is durable session data, not only runtime options —
  98. // a depth that lived only in AgentOptions would reset to 0 on resume.
  99. expect(ctx.agents.get(run.id)!.session.header.delegationDepth).toBe(1)
  100. await run.dispose()
  101. })
  102. it('inherits the parent output-token cap and accepts an explicit child override', async () => {
  103. const { ctx, parent, adapter } = await setup(
  104. [textResponse('inherited'), textResponse('overridden')],
  105. { maxTokens: 111 },
  106. )
  107. const inherited = await startInProcessRun(request(parent), {})
  108. await inherited.result
  109. expect(adapter.requests[0]?.maxTokens).toBe(111)
  110. expect(ctx.agents.get(inherited.id)?.options.maxTokens).toBe(111)
  111. await inherited.dispose()
  112. const overridden = await startInProcessRun({
  113. ...request(parent),
  114. agentOptions: { maxTokens: 222 },
  115. }, {})
  116. await overridden.result
  117. expect(adapter.requests[1]?.maxTokens).toBe(222)
  118. expect(ctx.agents.get(overridden.id)?.options.maxTokens).toBe(222)
  119. await overridden.dispose()
  120. })
  121. it('counts a RESUMED child by its persisted header depth, not the absent runtime depth', async () => {
  122. // Resume rebuilds runtime options, so the durable header must keep this
  123. // depth-1 child from delegating as though it were top-level.
  124. const { ctx } = await setup([textResponse('unused')])
  125. const resumed = (await ctx.agents.create({
  126. sessionId: SessionId('resumed-child'),
  127. meta: { parentSession: SessionId('root'), delegationDepth: 1 },
  128. agentOptions: { provider: 'mock', model: 'mock' },
  129. signal: new AbortController().signal,
  130. })).agent
  131. await expect(startInProcessRun({ ...request(resumed), maxDepth: 1 }, {}))
  132. .rejects.toMatchObject({ name: 'SubagentDepthError', attemptedDepth: 2, maxDepth: 1 })
  133. })
  134. it('lets runtime options deepen but never lower the persisted depth', async () => {
  135. const { ctx } = await setup([textResponse('unused')])
  136. const parent = (await ctx.agents.create({
  137. sessionId: SessionId('deep-parent'),
  138. meta: { delegationDepth: 2 },
  139. agentOptions: { provider: 'mock', model: 'mock', subagentDepth: 1 },
  140. signal: new AbortController().signal,
  141. })).agent
  142. // Persisted 2 vs runtime 1: the child is depth 3, so maxDepth 2 rejects.
  143. await expect(startInProcessRun({ ...request(parent), maxDepth: 2 }, {}))
  144. .rejects.toMatchObject({ name: 'SubagentDepthError', attemptedDepth: 3, maxDepth: 2 })
  145. })
  146. it('rejects invalid and exceeded depth before publication', async () => {
  147. const { parent } = await setup([])
  148. await expect(startInProcessRun({ ...request(parent), maxDepth: -1 }, {}))
  149. .rejects.toThrow('non-negative safe integer')
  150. await expect(startInProcessRun({ ...request(parent), maxDepth: 0 }, {}))
  151. .rejects.toMatchObject({ name: 'SubagentDepthError' })
  152. for (const value of [Number.NaN, 1.5, -1, -0, Number.MAX_SAFE_INTEGER + 1]) {
  153. const malformed = { options: { subagentDepth: value }, session: { header: {} } } as unknown as Agent
  154. await expect(startInProcessRun(request(malformed), {}))
  155. .rejects.toThrow('agent subagentDepth must be a non-negative safe integer')
  156. }
  157. const maxParent = { options: { subagentDepth: Number.MAX_SAFE_INTEGER }, session: { header: {} } } as unknown as Agent
  158. await expect(startInProcessRun(request(maxParent), {})).rejects.toBeInstanceOf(RangeError)
  159. })
  160. it('rejects an already-aborted request without publishing a child', async () => {
  161. const { ctx, parent } = await setup([])
  162. const beforeAgents = ctx.agents.list().length
  163. const beforeSessions = ctx.sessions.list().length
  164. const controller = new AbortController()
  165. controller.abort('too late')
  166. await expect(startInProcessRun(request(parent, controller.signal), {}))
  167. .rejects.toThrow('aborted before child publication')
  168. expect(ctx.agents.list()).toHaveLength(beforeAgents)
  169. expect(ctx.sessions.list()).toHaveLength(beforeSessions)
  170. })
  171. it('uses the request signal after publication and dispose as cancellation paths', async () => {
  172. const { parent, adapter } = await setup(['hang', 'hang'])
  173. const controller = new AbortController()
  174. const signalled = await startInProcessRun(request(parent, controller.signal), {})
  175. await new Promise(resolve => setTimeout(resolve, 30))
  176. controller.abort('stop child')
  177. await expect(signalled.result).resolves.toMatchObject({ stopReason: 'aborted' })
  178. expect(adapter.requests[0]?.signal?.reason).toEqual({ kind: 'parent' })
  179. const child = parent.ctx.agents.get(signalled.id)
  180. const turnEnd = child?.session.events.findLast(event => event.type === 'turn/end')
  181. expect(turnEnd?.type === 'turn/end' && turnEnd.data.reason).toEqual({ kind: 'aborted' })
  182. await signalled.dispose()
  183. const disposed = await startInProcessRun(request(parent), {})
  184. await new Promise(resolve => setTimeout(resolve, 30))
  185. await disposed.dispose()
  186. await expect(disposed.result).resolves.toMatchObject({ stopReason: 'aborted' })
  187. })
  188. it('cleans a failed unpublished setup before rejecting', async () => {
  189. const { ctx, parent } = await setup([])
  190. const beforeAgents = ctx.agents.list().length
  191. const beforeSessions = ctx.sessions.list().length
  192. await expect(startInProcessRun({
  193. ...request(parent),
  194. toolFilter: { deny: ['unknown-tool'] },
  195. }, {})).rejects.toThrow('unknown global tool')
  196. expect(ctx.agents.list()).toHaveLength(beforeAgents)
  197. expect(ctx.sessions.list()).toHaveLength(beforeSessions)
  198. })
  199. it('closes the abort handoff after the factory detaches its creation listener', async () => {
  200. const { ctx, parent } = await setup([])
  201. const controller = new AbortController()
  202. const beforeAgents = ctx.agents.list().length
  203. const beforeSessions = ctx.sessions.list().length
  204. const parentWithAbortAtHandoff = {
  205. options: parent.options,
  206. session: parent.session,
  207. ctx: {
  208. // The driver's synchronous inheritance capture probes both policy
  209. // services opportunistically; this stub composes neither.
  210. get: () => undefined,
  211. agents: {
  212. create: async (options: Parameters<typeof ctx.agents.create>[0]) => {
  213. const handle = await ctx.agents.create(options)
  214. // `create()` has detached its creation-only listener, but the
  215. // provider continuation has not installed its live-run listener.
  216. controller.abort('handoff race')
  217. return handle
  218. },
  219. },
  220. },
  221. } as unknown as Agent
  222. await expect(startInProcessRun(request(parentWithAbortAtHandoff, controller.signal), {}))
  223. .rejects.toThrow('aborted before child publication')
  224. expect(ctx.agents.list()).toHaveLength(beforeAgents)
  225. expect(ctx.sessions.list()).toHaveLength(beforeSessions)
  226. })
  227. })