loop.spec.ts 34 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816
  1. import { describe, expect, it } from 'vitest'
  2. import { Context } from 'cordis'
  3. import LlmService, { CallId, StreamChunk } from '@deepseek-ai/dsh-llm'
  4. import SessionStore, { SessionId, TurnEndReason } from '@deepseek-ai/dsh-session'
  5. import SystemPrompt from '@deepseek-ai/dsh-system-prompt'
  6. import ToolRegistry, { defineTool } from '@deepseek-ai/dsh-tools'
  7. import AgentRegistry, { AgentId } from '@deepseek-ai/dsh-agent'
  8. import AgentLoop, { ReactLoopAgent } from '@deepseek-ai/dsh-agent-loop'
  9. import { MockAdapter, maxTokensResponse, textResponse, toolCallResponse } from './mock-adapter.ts'
  10. async function harness(adapter: MockAdapter) {
  11. const ctx = new Context()
  12. await ctx.plugin(LlmService)
  13. await ctx.plugin(SessionStore)
  14. await ctx.plugin(SystemPrompt)
  15. await ctx.plugin(ToolRegistry)
  16. await ctx.plugin(AgentRegistry)
  17. await ctx.plugin(AgentLoop, { agents: [] })
  18. ctx.llm.registerAdapter(['mock'], adapter)
  19. return ctx
  20. }
  21. /**
  22. * Wait for the agent's NEXT transition to idle. Always event-based: callers
  23. * invoke this right after send(), when the loop hasn't woken yet (status is
  24. * still 'idle' synchronously), so polling the current status would lie.
  25. */
  26. function waitForIdle(ctx: Context, agent: ReactLoopAgent): Promise<void> {
  27. return new Promise((resolve) => {
  28. const dispose = ctx.on('agent/status', (subject, status) => {
  29. if (subject === agent && status === 'idle') {
  30. dispose()
  31. resolve()
  32. }
  33. })
  34. })
  35. }
  36. function send(agent: ReactLoopAgent, text: string) {
  37. agent.send([{ type: 'text', text }])
  38. }
  39. describe('agent loop', () => {
  40. it('runs a simple turn: queued message → model → idle, with ordered events', async () => {
  41. const adapter = new MockAdapter([textResponse('hello there')])
  42. const ctx = await harness(adapter)
  43. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  44. // All boundaries — turn and step — are durable session events on the
  45. // session/event feed (no agent/* mirror). Record them in fire order to
  46. // assert the full boundary nesting.
  47. const order: string[] = []
  48. ctx.on('session/event', (_session, event) => {
  49. if (event.type === 'turn/start' || event.type === 'step/start' || event.type === 'step/end' || event.type === 'turn/end') {
  50. order.push(event.type)
  51. }
  52. })
  53. send(agent, 'hi')
  54. await waitForIdle(ctx, agent)
  55. expect(order).toEqual(['turn/start', 'step/start', 'step/end', 'turn/end'])
  56. const types = agent.session.events.map(e => e.type)
  57. // turn/start opens the turn, THEN the queued user message is recorded inside
  58. // it (every event is turn-enclosed), then the assembled message (carrying the
  59. // step's usage).
  60. expect(types[0]).toBe('turn/start')
  61. expect(types[1]).toBe('user/message')
  62. expect(types).toContain('assistant/message')
  63. const assistantMessage = agent.session.events.find(e => e.type === 'assistant/message')
  64. expect(assistantMessage?.type === 'assistant/message' && assistantMessage.data.usage).toEqual({ inputTokens: 10, outputTokens: 'hello there'.length })
  65. expect(types.at(-1)).toBe('turn/end')
  66. // derived history: user + assistant
  67. const messages = agent.session.deriveMessages()
  68. expect(messages.map(m => m.role)).toEqual(['user', 'assistant'])
  69. expect(messages[1]!.content).toEqual([{ type: 'text', text: 'hello there' }])
  70. })
  71. it('round-trips tool calls: model requests tool → executes → result in next request', async () => {
  72. const adapter = new MockAdapter([
  73. toolCallResponse('c1', 'echo', { text: 'ping' }, 'calling echo'),
  74. textResponse('done'),
  75. ])
  76. const ctx = await harness(adapter)
  77. ctx.tools.register(defineTool({
  78. name: 'echo',
  79. description: 'echo back',
  80. parameters: { text: { type: 'string' } },
  81. async execute(args) {
  82. return [{ type: 'text', text: `echo: ${args.text}` }]
  83. },
  84. }))
  85. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  86. send(agent, 'use the tool')
  87. await waitForIdle(ctx, agent)
  88. // two model calls happened (tool-call step, then final step)
  89. expect(adapter.requests).toHaveLength(2)
  90. // the second request's derived history contains the tool result
  91. const secondMessages = adapter.requests[1]!.messages
  92. const toolResultMessage = secondMessages.find(m =>
  93. m.content.some(b => b.type === 'tool-result'))
  94. expect(toolResultMessage).toBeDefined()
  95. const block = toolResultMessage!.content.find(b => b.type === 'tool-result')!
  96. expect(block).toMatchObject({ toolCallId: 'c1', isError: false })
  97. expect((block).content).toEqual([{ type: 'text', text: 'echo: ping' }])
  98. // session log records call + result
  99. const types = agent.session.events.map(e => e.type)
  100. expect(types).toContain('tool/call')
  101. expect(types).toContain('tool/result')
  102. })
  103. it('passes assembled system prompt and tool schemas into the request', async () => {
  104. const adapter = new MockAdapter([textResponse('ok')])
  105. const ctx = await harness(adapter)
  106. ctx.systemPrompt.section({ name: 'persona', order: 0, text: 'You are a test agent.' })
  107. ctx.tools.register(defineTool({
  108. name: 'noop',
  109. description: 'does nothing',
  110. parameters: {},
  111. async execute() {
  112. return []
  113. },
  114. }))
  115. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock', systemPrompt: 'Agent-specific suffix.' })
  116. send(agent, 'hi')
  117. await waitForIdle(ctx, agent)
  118. const request = adapter.requests[0]
  119. expect(request!.system).toBe('You are a test agent.\n\nAgent-specific suffix.')
  120. expect(request!.tools?.map(t => t.name)).toEqual(['noop'])
  121. })
  122. it('records raw chunks for replay and emits agent/stream-chunk', async () => {
  123. const adapter = new MockAdapter([textResponse('abc')])
  124. const ctx = await harness(adapter)
  125. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  126. const streamed: StreamChunk[] = []
  127. ctx.on('agent/stream-chunk', (_agent, _turn, _step, chunk) => void streamed.push(chunk))
  128. send(agent, 'hi')
  129. await waitForIdle(ctx, agent)
  130. const chunkEvents = agent.session.events.filter(e => e.type === 'assistant/chunk')
  131. // textResponse('abc') = block-start + 3 deltas + block-end + usage + finish = 7
  132. expect(chunkEvents).toHaveLength(7)
  133. expect(streamed).toHaveLength(7)
  134. // replay: chunk events alone re-assemble to the recorded assistant message
  135. const deltaText = chunkEvents
  136. .flatMap(e => e.type === 'assistant/chunk' ? [e.data.chunk] : [])
  137. .filter((c: StreamChunk): c is Extract<StreamChunk, { type: 'text-delta' }> => c.type === 'text-delta')
  138. .map(c => c.text)
  139. .join('')
  140. expect(deltaText).toBe('abc')
  141. })
  142. it('injects steering between steps and continues the turn', async () => {
  143. const adapter = new MockAdapter([
  144. toolCallResponse('c1', 'slow', {}),
  145. textResponse('addressed the steering'),
  146. ])
  147. const ctx = await harness(adapter)
  148. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  149. ctx.tools.register(defineTool({
  150. name: 'slow',
  151. description: '',
  152. parameters: {},
  153. async execute() {
  154. // steer while the turn is running (during tool execution)
  155. agent.steer([{ type: 'text', text: 'change of plans' }])
  156. return [{ type: 'text', text: 'tool done' }]
  157. },
  158. }))
  159. send(agent, 'start')
  160. await waitForIdle(ctx, agent)
  161. const types = agent.session.events.map(e => e.type)
  162. expect(types).toContain('steering/message')
  163. // steering recorded before the second step's request derived its history
  164. const steeringSeq = agent.session.events.find(e => e.type === 'steering/message')!.seq
  165. const secondStepStart = agent.session.events.filter(e => e.type === 'step/start')[1]
  166. expect(secondStepStart).toBeDefined()
  167. expect(steeringSeq).toBeLessThan(secondStepStart!.seq)
  168. // the second model request saw the steering content
  169. const secondRequest = adapter.requests[1]
  170. const flat = JSON.stringify(secondRequest!.messages)
  171. expect(flat).toContain('change of plans')
  172. })
  173. it('steering while idle behaves like send (starts a turn)', async () => {
  174. const adapter = new MockAdapter([textResponse('ok')])
  175. const ctx = await harness(adapter)
  176. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  177. agent.steer([{ type: 'text', text: 'hello' }])
  178. await waitForIdle(ctx, agent)
  179. expect(agent.session.events.some(e => e.type === 'user/message')).toBe(true)
  180. })
  181. it('inject() while idle wraps context in a one-shot turn, visible to the next request', async () => {
  182. const adapter = new MockAdapter([textResponse('ok')])
  183. const ctx = await harness(adapter)
  184. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  185. agent.inject([{ type: 'text', text: 'file changed: a.ts' }], { source: { kind: 'plugin', plugin: 'watcher' } })
  186. // The idle inject records a self-contained turn (turn/start → context/message
  187. // → turn/end) so the event stays turn-enclosed, but does NOT run the model.
  188. await new Promise(r => setTimeout(r, 20))
  189. expect(agent.status).toBe('idle')
  190. expect(adapter.requests).toHaveLength(0)
  191. const injectedTurn = agent.session.events.filter(e => e.type === 'turn/start')
  192. expect(injectedTurn).toHaveLength(1)
  193. const it0 = injectedTurn[0]!
  194. expect(it0.type === 'turn/start' && it0.data.trigger.kind).toBe('injection')
  195. expect(agent.session.events.at(-1)!.type).toBe('turn/end') // turn-enclosed
  196. send(agent, 'go')
  197. await waitForIdle(ctx, agent)
  198. const flat = JSON.stringify(adapter.requests[0]!.messages)
  199. expect(flat).toContain('file changed: a.ts')
  200. expect(flat).toContain('<context source=\\"plugin\\">')
  201. })
  202. it('inject() while running appends into the open turn (no extra synthetic turn)', async () => {
  203. const adapter = new MockAdapter([
  204. toolCallResponse('c1', 'noticer', {}, 'calling'),
  205. textResponse('done'),
  206. ])
  207. const ctx = await harness(adapter)
  208. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  209. // A tool that injects mid-execution: at this point the agent is running, so
  210. // inject must append the context/message into the ALREADY-open turn rather
  211. // than wrap it in its own one-shot turn.
  212. ctx.tools.register(defineTool({
  213. name: 'noticer',
  214. description: 'injects a notice',
  215. parameters: {},
  216. async execute() {
  217. agent.inject([{ type: 'text', text: 'mid-turn notice' }], { source: { kind: 'plugin', plugin: 'x' } })
  218. return [{ type: 'text', text: 'ok' }]
  219. },
  220. }))
  221. send(agent, 'go')
  222. await waitForIdle(ctx, agent)
  223. // Exactly ONE turn ran (no synthetic injection turn), and the mid-turn
  224. // context/message sits inside it.
  225. const turnStarts = agent.session.events.filter(e => e.type === 'turn/start')
  226. expect(turnStarts).toHaveLength(1)
  227. const ts0 = turnStarts[0]!
  228. expect(ts0.type === 'turn/start' && ts0.data.trigger.kind).toBe('message')
  229. expect(agent.session.events.some(e => e.type === 'context/message')).toBe(true)
  230. })
  231. it('agent/turn-continuation can force-continue (/loop pattern) and force-stop', async () => {
  232. // force-continue: model never calls tools, but a plugin forces 3 steps
  233. const adapter = new MockAdapter([
  234. textResponse('step 1'),
  235. textResponse('step 2'),
  236. textResponse('step 3'),
  237. ])
  238. const ctx = await harness(adapter)
  239. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  240. let steps = 0
  241. ctx.on('session/event', (_session, event) => { if (event.type === 'step/end') steps++ })
  242. ctx.on('agent/turn-continuation', async (_agent, _turn, _defaultDecision, next) => {
  243. if (steps < 3) return { action: 'continue' as const }
  244. return next()
  245. })
  246. send(agent, 'go')
  247. await waitForIdle(ctx, agent)
  248. expect(steps).toBe(3)
  249. expect(adapter.requests).toHaveLength(3)
  250. })
  251. it('agent/turn-continuation can veto continuation despite tool calls (budget-guard pattern)', async () => {
  252. const adapter = new MockAdapter([toolCallResponse('c1', 'echo', { text: 'x' })])
  253. const ctx = await harness(adapter)
  254. ctx.tools.register(defineTool({
  255. name: 'echo',
  256. description: '',
  257. parameters: { text: { type: 'string' } },
  258. async execute(args) {
  259. return [{ type: 'text', text: String(args.text) }]
  260. },
  261. }))
  262. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  263. ctx.on('agent/turn-continuation', async () => ({ action: 'stop' }) as const)
  264. send(agent, 'go')
  265. await waitForIdle(ctx, agent)
  266. // only one model call despite the tool call requesting a follow-up
  267. expect(adapter.requests).toHaveLength(1)
  268. // tool still executed before the decision
  269. expect(agent.session.events.some(e => e.type === 'tool/result')).toBe(true)
  270. })
  271. it('agent/request waterfall can rewrite the request (model-switch pattern)', async () => {
  272. const adapter = new MockAdapter([textResponse('ok')])
  273. const ctx = await harness(adapter)
  274. ctx.llm.registerAdapter(['other-model'], adapter)
  275. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  276. ctx.on('agent/request', async (_agent, _turn, _step, options, next) => {
  277. options.model = 'other-model'
  278. return next()
  279. })
  280. send(agent, 'hi')
  281. await waitForIdle(ctx, agent)
  282. expect(adapter.requests[0]!.model).toBe('other-model')
  283. })
  284. it('agent/pre-step fires once per step before the step is opened', async () => {
  285. // Two steps (a tool call, then a final text turn) → two model calls → two
  286. // pre-step fires, each carrying the assembled full system prompt, BEFORE
  287. // the step is opened and its request is derived (the request the adapter
  288. // sees reflects any surface state at fire time).
  289. const adapter = new MockAdapter([
  290. toolCallResponse('c1', 'echo', {}, 'calling echo'),
  291. textResponse('done'),
  292. ])
  293. const ctx = await harness(adapter)
  294. ctx.tools.register(defineTool({
  295. name: 'echo', description: 'echo', parameters: {},
  296. async execute() { return [{ type: 'text', text: 'echoed' }] },
  297. }))
  298. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  299. const fires: { turn: number; step: number; fullSystemPrompt: string }[] = []
  300. ctx.on('agent/pre-step', (subject, turn, step, fullSystemPrompt) => {
  301. if (subject === agent) fires.push({ turn, step, fullSystemPrompt })
  302. })
  303. send(agent, 'go')
  304. await waitForIdle(ctx, agent)
  305. // One fire per step, in order, each with the assembled system prompt.
  306. expect(fires).toEqual([
  307. { turn: 1, step: 1, fullSystemPrompt: '' },
  308. { turn: 1, step: 2, fullSystemPrompt: '' },
  309. ])
  310. })
  311. it('agent/pre-step fires BEFORE the step it precedes opens (events land outside the step)', async () => {
  312. // A listener appending a surface node in pre-step lands it BEFORE step/start
  313. // in the log — proving the seam fires outside the step. The node is still in
  314. // the derived request for that step (derive happens after step/start).
  315. const adapter = new MockAdapter([textResponse('ok')])
  316. const ctx = await harness(adapter)
  317. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  318. let injected = false
  319. ctx.on('agent/pre-step', (subject) => {
  320. if (subject === agent && !injected) {
  321. injected = true
  322. subject.session.append('context/message', {
  323. content: [{ type: 'text', text: 'INJECTED-IN-PRE-STEP' }],
  324. source: { kind: 'plugin', plugin: 'test' },
  325. }, { surfaceOp: 'append' })
  326. }
  327. })
  328. send(agent, 'go')
  329. await waitForIdle(ctx, agent)
  330. // The adapter's request includes the node injected during pre-step (derive
  331. // reflects it).
  332. const text = JSON.stringify(adapter.requests[0]!.messages)
  333. expect(text).toContain('INJECTED-IN-PRE-STEP')
  334. // And the injected event sits BEFORE the first step/start in the log —
  335. // the seam fired outside the step.
  336. const events = agent.session.events
  337. const injectedSeq = events.find(e => e.type === 'context/message')!.seq
  338. const firstStepStartSeq = events.find(e => e.type === 'step/start')!.seq
  339. expect(injectedSeq).toBeLessThan(firstStepStartSeq)
  340. })
  341. it('a throwing agent/pre-step listener ends the turn (error), not the loop', async () => {
  342. // The seam fires before step/start, so a throw escapes to runTurn's outer
  343. // catch: the not-yet-open step closes as a no-op, the failure surfaces via
  344. // agent/error, and the turn ends `error` (recorded on the durable turn/end).
  345. // The loop survives and a follow-up prompt still runs.
  346. const adapter = new MockAdapter([textResponse('second turn ok')])
  347. const ctx = await harness(adapter)
  348. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  349. let throwOnce = true
  350. ctx.on('agent/pre-step', () => {
  351. if (throwOnce) { throwOnce = false; throw new Error('boom in pre-step') }
  352. })
  353. const errors: Error[] = []
  354. ctx.on('agent/error', (_a, _t, _s, error) => void errors.push(error))
  355. send(agent, 'first')
  356. await waitForIdle(ctx, agent)
  357. // The first turn failed at step 1 (no model call happened), surfaced via
  358. // agent/error, with the durable failure on turn/end.reason.
  359. expect(errors).toHaveLength(1)
  360. expect(errors[0]!.message).toContain('boom in pre-step')
  361. expect(adapter.requests.length).toBe(0)
  362. const firstTurnEnd = agent.session.events.find(e => e.type === 'turn/end')
  363. expect(firstTurnEnd?.type === 'turn/end' && firstTurnEnd.data.reason).toMatchObject({ kind: 'error', step: 1 })
  364. // The step opened-and-closed count stays balanced even though it never ran.
  365. const types = agent.session.events.map(e => e.type)
  366. expect(types.filter(t => t === 'step/start').length).toBe(types.filter(t => t === 'step/end').length)
  367. // The loop survived: a second prompt runs a normal completed turn.
  368. send(agent, 'second')
  369. await waitForIdle(ctx, agent)
  370. expect(adapter.requests.length).toBe(1)
  371. const lastTurnEnd = agent.session.events.findLast(e => e.type === 'turn/end')
  372. expect(lastTurnEnd?.type === 'turn/end' && lastTurnEnd.data.reason).toEqual({ kind: 'completed' })
  373. })
  374. it('cancel() mid-stream ends the turn with reason aborted', async () => {
  375. const adapter = new MockAdapter(['hang'])
  376. const ctx = await harness(adapter)
  377. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  378. const reasons: TurnEndReason[] = []
  379. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  380. send(agent, 'go')
  381. // wait until the stream is hanging, then cancel
  382. await new Promise(r => setTimeout(r, 30))
  383. expect(agent.status).toBe('running')
  384. agent.cancel('user interrupt')
  385. await waitForIdle(ctx, agent)
  386. expect(reasons).toEqual([{ kind: 'aborted', reason: 'user interrupt' }])
  387. })
  388. it('surfaces max-tokens as the turn-end reason when the last step is cut off', async () => {
  389. // A single step that ends with a max-tokens finish (no tool calls): the
  390. // turn stops by default and ends max-tokens, not completed.
  391. const adapter = new MockAdapter([maxTokensResponse('truncat')])
  392. const ctx = await harness(adapter)
  393. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  394. const reasons: TurnEndReason[] = []
  395. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  396. send(agent, 'go')
  397. await waitForIdle(ctx, agent)
  398. expect(adapter.requests).toHaveLength(1)
  399. expect(reasons).toEqual([{ kind: 'max-tokens' }])
  400. // and the reason is recorded in the log's turn/end event
  401. const turnEnd = agent.session.events.findLast(e => e.type === 'turn/end')
  402. expect(turnEnd!.data.reason).toEqual({ kind: 'max-tokens' })
  403. })
  404. it('a max-tokens step earlier in a turn still surfaces as max-tokens after a later completed step', async () => {
  405. // Step 1 is cut off (max-tokens, no tool calls → would stop by default), so
  406. // continuation must be FORCED to reach step 2 which finishes normally
  407. // (stop). The rule "any max-tokens step surfaces as max-tokens" means the
  408. // turn ends max-tokens even though the LAST step completed cleanly.
  409. const adapter = new MockAdapter([
  410. maxTokensResponse('first half'),
  411. textResponse('second half'),
  412. ])
  413. const ctx = await harness(adapter)
  414. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  415. let steps = 0
  416. ctx.on('session/event', (_session, event) => { if (event.type === 'step/end') steps++ })
  417. // Force exactly one continuation (step 1 → step 2), then defer to default
  418. // (step 2 is a plain stop with no tool calls → stops).
  419. ctx.on('agent/turn-continuation', async (_agent, _turn, _defaultDecision, next) => {
  420. if (steps < 2) return { action: 'continue' as const }
  421. return next()
  422. })
  423. const reasons: TurnEndReason[] = []
  424. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  425. send(agent, 'go')
  426. await waitForIdle(ctx, agent)
  427. expect(steps).toBe(2)
  428. expect(adapter.requests).toHaveLength(2)
  429. expect(adapter.requests[1]!.messages).toEqual([
  430. { role: 'user', content: [{ type: 'text', text: 'go' }] },
  431. { role: 'assistant', content: [{ type: 'text', text: 'first half' }] },
  432. ])
  433. expect(reasons).toEqual([{ kind: 'max-tokens' }])
  434. })
  435. it('a completed step after no max-tokens keeps the turn completed (max-tokens does not leak across turns)', async () => {
  436. // Two consecutive turns: turn 1 is cut off (max-tokens), turn 2 is a clean
  437. // stop. The per-turn reason must be independent — turn 2 ends completed.
  438. const adapter = new MockAdapter([maxTokensResponse('cut'), textResponse('clean')])
  439. const ctx = await harness(adapter)
  440. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  441. const reasons: TurnEndReason[] = []
  442. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  443. send(agent, 'first')
  444. await waitForIdle(ctx, agent)
  445. send(agent, 'second')
  446. await waitForIdle(ctx, agent)
  447. expect(reasons).toEqual([{ kind: 'max-tokens' }, { kind: 'completed' }])
  448. })
  449. it('does not dispatch tool calls from a max-tokens-truncated step', async () => {
  450. const callId = CallId('c1')
  451. const adapter = new MockAdapter([[
  452. { type: 'block-start', index: 0, blockType: 'tool-call' },
  453. { type: 'tool-call-delta', index: 0, id: callId, name: 'echo', argumentsDelta: '{"text":"x"}' },
  454. { type: 'block-end', index: 0, block: { type: 'tool-call', id: callId, name: 'echo', arguments: '{"text":"x"}' } },
  455. { type: 'usage', usage: { inputTokens: 10, outputTokens: 5 } },
  456. { type: 'finish', reason: { kind: 'max-tokens' } },
  457. ]])
  458. const ctx = await harness(adapter)
  459. let executions = 0
  460. ctx.tools.register(defineTool({
  461. name: 'echo',
  462. description: '',
  463. parameters: { text: { type: 'string' } },
  464. async execute() {
  465. executions += 1
  466. return [{ type: 'text', text: 'should not run' }]
  467. },
  468. }))
  469. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  470. const reasons: TurnEndReason[] = []
  471. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  472. send(agent, 'go')
  473. await waitForIdle(ctx, agent)
  474. expect(executions).toBe(0)
  475. expect(agent.session.events.some(e => e.type === 'tool/call')).toBe(false)
  476. expect(agent.session.deriveMessages()).toEqual([{ role: 'user', content: [{ type: 'text', text: 'go' }] }])
  477. expect(reasons).toEqual([{ kind: 'max-tokens' }])
  478. // No-data-loss: a max-tokens step whose only content was a dropped tool call
  479. // has EMPTY assistant content, but its usage must still be represented. It
  480. // rides on an (empty-content) assistant/message — there is no standalone
  481. // usage event — and that empty message is skipped by deriveMessages(), so
  482. // the derived history above is NOT corrupted by a spurious assistant turn.
  483. const assistantMessage = agent.session.events.find(e => e.type === 'assistant/message')
  484. expect(assistantMessage?.type === 'assistant/message' && assistantMessage.data).toEqual({
  485. turn: 1, step: 1, content: [], usage: { inputTokens: 10, outputTokens: 5 },
  486. })
  487. })
  488. it('appends no assistant/message for a max-tokens step with empty content and no usage', async () => {
  489. // A max-tokens step truncated to a dropped tool call AND with no usage chunk
  490. // has nothing to record: empty content and no accounting → no assistant/message
  491. // (the empty-content host exists only to carry usage). The turn still ends
  492. // max-tokens.
  493. const callId = CallId('c1')
  494. const adapter = new MockAdapter([[
  495. { type: 'block-start', index: 0, blockType: 'tool-call' },
  496. { type: 'tool-call-delta', index: 0, id: callId, name: 'echo', argumentsDelta: '{"text":"x"}' },
  497. { type: 'block-end', index: 0, block: { type: 'tool-call', id: callId, name: 'echo', arguments: '{"text":"x"}' } },
  498. { type: 'finish', reason: { kind: 'max-tokens' } },
  499. ]])
  500. const ctx = await harness(adapter)
  501. ctx.tools.register(defineTool({
  502. name: 'echo',
  503. description: '',
  504. parameters: { text: { type: 'string' } },
  505. async execute() { return [{ type: 'text', text: 'should not run' }] },
  506. }))
  507. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  508. const reasons: TurnEndReason[] = []
  509. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  510. send(agent, 'go')
  511. await waitForIdle(ctx, agent)
  512. expect(reasons).toEqual([{ kind: 'max-tokens' }])
  513. expect(agent.session.events.some(e => e.type === 'assistant/message')).toBe(false)
  514. expect(agent.session.deriveMessages()).toEqual([{ role: 'user', content: [{ type: 'text', text: 'go' }] }])
  515. })
  516. it('appends no assistant/message for a normal stop finish with empty content and no usage', async () => {
  517. // A clean `stop` finish that streamed nothing assembled (no blocks) and
  518. // carried no usage chunk has nothing to record: the content-or-usage guard
  519. // on the normal step path suppresses a pure trace-only empty assistant/message.
  520. const adapter = new MockAdapter([[{ type: 'finish', reason: { kind: 'stop' } }]])
  521. const ctx = await harness(adapter)
  522. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  523. const reasons: TurnEndReason[] = []
  524. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  525. send(agent, 'go')
  526. await waitForIdle(ctx, agent)
  527. expect(reasons).toEqual([{ kind: 'completed' }])
  528. expect(agent.session.events.some(e => e.type === 'assistant/message')).toBe(false)
  529. expect(agent.session.deriveMessages()).toEqual([{ role: 'user', content: [{ type: 'text', text: 'go' }] }])
  530. })
  531. it('keeps safe max-tokens assistant content while dropping truncated tool calls', async () => {
  532. const callId = CallId('c1')
  533. const adapter = new MockAdapter([[
  534. { type: 'block-start', index: 0, blockType: 'text' },
  535. { type: 'text-delta', index: 0, text: 'partial text' },
  536. { type: 'block-end', index: 0, block: { type: 'text', text: 'partial text' } },
  537. { type: 'block-start', index: 1, blockType: 'tool-call' },
  538. { type: 'tool-call-delta', index: 1, id: callId, name: 'echo', argumentsDelta: '{"text"' },
  539. { type: 'finish', reason: { kind: 'max-tokens' } },
  540. ]])
  541. const ctx = await harness(adapter)
  542. let stepResults = 0
  543. ctx.on('agent/step-result', async (_agent, _turn, _step, message, next) => {
  544. stepResults += 1
  545. expect(message.content).toEqual([{ type: 'text', text: 'partial text' }])
  546. return next()
  547. })
  548. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  549. send(agent, 'go')
  550. await waitForIdle(ctx, agent)
  551. expect(stepResults).toBe(1)
  552. expect(agent.session.events.some(e => e.type === 'tool/call')).toBe(false)
  553. expect(agent.session.deriveMessages()).toEqual([
  554. { role: 'user', content: [{ type: 'text', text: 'go' }] },
  555. { role: 'assistant', content: [{ type: 'text', text: 'partial text' }] },
  556. ])
  557. })
  558. it('stops the turn when a step/end session-event listener failure has recorded an error', async () => {
  559. const adapter = new MockAdapter([
  560. toolCallResponse('c1', 'echo', { text: 'x' }),
  561. textResponse('should not run'),
  562. ])
  563. const ctx = await harness(adapter)
  564. ctx.tools.register(defineTool({
  565. name: 'echo',
  566. description: '',
  567. parameters: { text: { type: 'string' } },
  568. async execute(args) {
  569. return [{ type: 'text', text: String(args.text) }]
  570. },
  571. }))
  572. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  573. let threw = false
  574. // A throwing step/end session-event listener is the surviving boundary-listener
  575. // failure path (step boundaries have no agent/* mirror): closeStep contains it
  576. // and surfaces it as a turn error rather than stranding the turn open.
  577. ctx.on('session/event', (_session, event) => {
  578. if (event.type === 'step/end' && !threw) { threw = true; throw new Error('bad step/end listener') }
  579. })
  580. send(agent, 'go')
  581. await waitForIdle(ctx, agent)
  582. expect(adapter.requests).toHaveLength(1)
  583. const turnEnd = agent.session.events.findLast(e => e.type === 'turn/end')
  584. expect(turnEnd?.type === 'turn/end' && turnEnd.data.reason.kind).toBe('error')
  585. })
  586. it('chains queued messages into consecutive turns', async () => {
  587. const adapter = new MockAdapter([textResponse('first'), textResponse('second')])
  588. const ctx = await harness(adapter)
  589. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  590. const turns: number[] = []
  591. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/start') turns.push(event.data.turn) })
  592. // queue two messages while idle — first starts turn 1 immediately;
  593. // queue the second during turn 1 via a stream-chunk hook
  594. let queued = false
  595. ctx.on('agent/stream-chunk', () => {
  596. if (!queued) {
  597. queued = true
  598. send(agent, 'second message')
  599. }
  600. })
  601. send(agent, 'first message')
  602. await waitForIdle(ctx, agent)
  603. expect(turns).toEqual([1, 2])
  604. expect(adapter.requests).toHaveLength(2)
  605. })
  606. it('awaits session/flush at turn end (persistence checkpoint)', async () => {
  607. const adapter = new MockAdapter([textResponse('ok')])
  608. const ctx = await harness(adapter)
  609. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  610. let flushed = 0
  611. let flushedBeforeIdle = false
  612. ctx.on('session/flush', async (session) => {
  613. await new Promise(r => setTimeout(r, 10))
  614. flushed++
  615. flushedBeforeIdle = agent.status !== 'idle'
  616. void session
  617. })
  618. send(agent, 'hi')
  619. await waitForIdle(ctx, agent)
  620. expect(flushed).toBe(1)
  621. expect(flushedBeforeIdle).toBe(true)
  622. })
  623. it('errors from the model surface as agent/error and end the turn', async () => {
  624. const adapter = new MockAdapter([]) // script exhausted → throws
  625. const ctx = await harness(adapter)
  626. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  627. const errors: Error[] = []
  628. const reasons: TurnEndReason[] = []
  629. ctx.on('agent/error', (_agent, _turn, _step, error) => void errors.push(error))
  630. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  631. send(agent, 'hi')
  632. await waitForIdle(ctx, agent)
  633. expect(errors).toHaveLength(1)
  634. expect(errors[0]!.message).toContain('script exhausted')
  635. expect(reasons[0]).toMatchObject({ kind: 'error' })
  636. // The durable failure lives entirely on turn/end.reason (with the failing
  637. // step), not a standalone error event.
  638. const turnEnd = agent.session.events.find(e => e.type === 'turn/end')
  639. expect(turnEnd?.type === 'turn/end' && turnEnd.data.reason).toMatchObject({ kind: 'error', step: 1 })
  640. })
  641. it('disposing the loop fiber mid-turn stops the loop (HMR safety)', async () => {
  642. const adapter = new MockAdapter(['hang'])
  643. const ctx = await harness(adapter)
  644. let agent!: ReactLoopAgent
  645. const fiber = await ctx.plugin(Object.assign((inner: Context) => {
  646. agent = inner.agentLoop.create(AgentId('scoped'), { model: 'mock' })
  647. }, { inject: ['agentLoop'] }))
  648. expect(ctx.agents.get(AgentId('scoped'))).toBe(agent)
  649. send(agent, 'go')
  650. await new Promise(r => setTimeout(r, 30))
  651. expect(agent.status).toBe('running')
  652. await fiber.dispose()
  653. await agent.done
  654. expect(agent.status).toBe('disposed')
  655. expect(ctx.agents.get(AgentId('scoped'))).toBeUndefined()
  656. expect(() => { send(agent, 'too late') }).toThrow('disposed')
  657. })
  658. it('creates agents from config on startup', async () => {
  659. const adapter = new MockAdapter([textResponse('from config')])
  660. const ctx = new Context()
  661. await ctx.plugin(LlmService)
  662. await ctx.plugin(SessionStore)
  663. await ctx.plugin(SystemPrompt)
  664. await ctx.plugin(ToolRegistry)
  665. await ctx.plugin(AgentRegistry)
  666. await ctx.plugin(AgentLoop, {
  667. agents: [{ id: AgentId('config-agent'), model: 'mock', systemPrompt: 'Config prompt' }],
  668. })
  669. ctx.llm.registerAdapter(['mock'], adapter)
  670. const agent = ctx.agents.get(AgentId('config-agent'))! as ReactLoopAgent
  671. expect(agent).toBeDefined()
  672. expect(agent.id).toBe('config-agent')
  673. expect(agent.options.model).toBe('mock')
  674. // the agent is alive: send triggers a turn
  675. send(agent, 'hi')
  676. await waitForIdle(ctx, agent)
  677. expect(adapter.requests).toHaveLength(1)
  678. })
  679. it('replays a session log into an identical derived history', async () => {
  680. const adapter = new MockAdapter([
  681. toolCallResponse('c1', 'echo', { text: 'x' }),
  682. textResponse('done'),
  683. ])
  684. const ctx = await harness(adapter)
  685. ctx.tools.register(defineTool({
  686. name: 'echo',
  687. description: '',
  688. parameters: { text: { type: 'string' } },
  689. async execute(args) {
  690. return [{ type: 'text', text: String(args.text) }]
  691. },
  692. }))
  693. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  694. send(agent, 'run')
  695. await waitForIdle(ctx, agent)
  696. const replayed = ctx.sessions.create(SessionId('replayed'), { seed: [...agent.session.events] })
  697. expect(replayed.deriveMessages()).toEqual(agent.session.deriveMessages())
  698. // event-by-event identity of types
  699. expect(replayed.events.map(e => e.type)).toEqual(
  700. agent.session.events.map(e => e.type))
  701. })
  702. })