loop.spec.ts 35 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838
  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('threads a tool-attached meta (execute object return) onto the tool/result event', async () => {
  104. const adapter = new MockAdapter([
  105. toolCallResponse('c1', 'writer', { path: 'a.txt' }, 'writing'),
  106. textResponse('done'),
  107. ])
  108. const ctx = await harness(adapter)
  109. // A tool that returns the { content, meta } object form: the loop must
  110. // persist `meta` on the tool/result event so a UI reproduces the card on replay.
  111. ctx.tools.register(defineTool({
  112. name: 'writer',
  113. description: 'writes a file',
  114. parameters: { path: { type: 'string' } },
  115. async execute() {
  116. return { content: [{ type: 'text', text: 'ok' }], meta: { diffs: [{ path: 'a.txt', oldText: null, newText: 'x' }] } }
  117. },
  118. }))
  119. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  120. send(agent, 'use the tool')
  121. await waitForIdle(ctx, agent)
  122. const toolResult = agent.session.events.find(e => e.type === 'tool/result')
  123. expect(toolResult?.type === 'tool/result' && toolResult.data.meta)
  124. .toEqual({ diffs: [{ path: 'a.txt', oldText: null, newText: 'x' }] })
  125. })
  126. it('passes assembled system prompt and tool schemas into the request', async () => {
  127. const adapter = new MockAdapter([textResponse('ok')])
  128. const ctx = await harness(adapter)
  129. ctx.systemPrompt.section({ name: 'persona', order: 0, text: 'You are a test agent.' })
  130. ctx.tools.register(defineTool({
  131. name: 'noop',
  132. description: 'does nothing',
  133. parameters: {},
  134. async execute() {
  135. return []
  136. },
  137. }))
  138. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock', systemPrompt: 'Agent-specific suffix.' })
  139. send(agent, 'hi')
  140. await waitForIdle(ctx, agent)
  141. const request = adapter.requests[0]
  142. expect(request!.system).toBe('You are a test agent.\n\nAgent-specific suffix.')
  143. expect(request!.tools?.map(t => t.name)).toEqual(['noop'])
  144. })
  145. it('records raw chunks for replay as assistant/chunk session events', async () => {
  146. const adapter = new MockAdapter([textResponse('abc')])
  147. const ctx = await harness(adapter)
  148. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  149. send(agent, 'hi')
  150. await waitForIdle(ctx, agent)
  151. const chunkEvents = agent.session.events.filter(e => e.type === 'assistant/chunk')
  152. // textResponse('abc') = block-start + 3 deltas + block-end + usage + finish = 7
  153. expect(chunkEvents).toHaveLength(7)
  154. // replay: chunk events alone re-assemble to the recorded assistant message
  155. const deltaText = chunkEvents
  156. .flatMap(e => e.type === 'assistant/chunk' ? [e.data.chunk] : [])
  157. .filter((c: StreamChunk): c is Extract<StreamChunk, { type: 'text-delta' }> => c.type === 'text-delta')
  158. .map(c => c.text)
  159. .join('')
  160. expect(deltaText).toBe('abc')
  161. })
  162. it('injects steering between steps and continues the turn', async () => {
  163. const adapter = new MockAdapter([
  164. toolCallResponse('c1', 'slow', {}),
  165. textResponse('addressed the steering'),
  166. ])
  167. const ctx = await harness(adapter)
  168. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  169. ctx.tools.register(defineTool({
  170. name: 'slow',
  171. description: '',
  172. parameters: {},
  173. async execute() {
  174. // steer while the turn is running (during tool execution)
  175. agent.steer([{ type: 'text', text: 'change of plans' }])
  176. return [{ type: 'text', text: 'tool done' }]
  177. },
  178. }))
  179. send(agent, 'start')
  180. await waitForIdle(ctx, agent)
  181. const types = agent.session.events.map(e => e.type)
  182. expect(types).toContain('steering/message')
  183. // steering recorded before the second step's request derived its history
  184. const steeringSeq = agent.session.events.find(e => e.type === 'steering/message')!.seq
  185. const secondStepStart = agent.session.events.filter(e => e.type === 'step/start')[1]
  186. expect(secondStepStart).toBeDefined()
  187. expect(steeringSeq).toBeLessThan(secondStepStart!.seq)
  188. // the second model request saw the steering content
  189. const secondRequest = adapter.requests[1]
  190. const flat = JSON.stringify(secondRequest!.messages)
  191. expect(flat).toContain('change of plans')
  192. })
  193. it('steering while idle behaves like send (starts a turn)', async () => {
  194. const adapter = new MockAdapter([textResponse('ok')])
  195. const ctx = await harness(adapter)
  196. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  197. agent.steer([{ type: 'text', text: 'hello' }])
  198. await waitForIdle(ctx, agent)
  199. expect(agent.session.events.some(e => e.type === 'user/message')).toBe(true)
  200. })
  201. it('inject() while idle wraps context in a one-shot turn, visible to the next request', async () => {
  202. const adapter = new MockAdapter([textResponse('ok')])
  203. const ctx = await harness(adapter)
  204. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  205. agent.inject([{ type: 'text', text: 'file changed: a.ts' }], { source: { kind: 'plugin', plugin: 'watcher' } })
  206. // The idle inject records a self-contained turn (turn/start → context/message
  207. // → turn/end) so the event stays turn-enclosed, but does NOT run the model.
  208. await new Promise(r => setTimeout(r, 20))
  209. expect(agent.status).toBe('idle')
  210. expect(adapter.requests).toHaveLength(0)
  211. const injectedTurn = agent.session.events.filter(e => e.type === 'turn/start')
  212. expect(injectedTurn).toHaveLength(1)
  213. const it0 = injectedTurn[0]!
  214. expect(it0.type === 'turn/start' && it0.data.trigger.kind).toBe('injection')
  215. expect(agent.session.events.at(-1)!.type).toBe('turn/end') // turn-enclosed
  216. send(agent, 'go')
  217. await waitForIdle(ctx, agent)
  218. const flat = JSON.stringify(adapter.requests[0]!.messages)
  219. expect(flat).toContain('file changed: a.ts')
  220. expect(flat).toContain('<context source=\\"plugin\\">')
  221. })
  222. it('inject() while running appends into the open turn (no extra synthetic turn)', async () => {
  223. const adapter = new MockAdapter([
  224. toolCallResponse('c1', 'noticer', {}, 'calling'),
  225. textResponse('done'),
  226. ])
  227. const ctx = await harness(adapter)
  228. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  229. // A tool that injects mid-execution: at this point the agent is running, so
  230. // inject must append the context/message into the ALREADY-open turn rather
  231. // than wrap it in its own one-shot turn.
  232. ctx.tools.register(defineTool({
  233. name: 'noticer',
  234. description: 'injects a notice',
  235. parameters: {},
  236. async execute() {
  237. agent.inject([{ type: 'text', text: 'mid-turn notice' }], { source: { kind: 'plugin', plugin: 'x' } })
  238. return [{ type: 'text', text: 'ok' }]
  239. },
  240. }))
  241. send(agent, 'go')
  242. await waitForIdle(ctx, agent)
  243. // Exactly ONE turn ran (no synthetic injection turn), and the mid-turn
  244. // context/message sits inside it.
  245. const turnStarts = agent.session.events.filter(e => e.type === 'turn/start')
  246. expect(turnStarts).toHaveLength(1)
  247. const ts0 = turnStarts[0]!
  248. expect(ts0.type === 'turn/start' && ts0.data.trigger.kind).toBe('message')
  249. expect(agent.session.events.some(e => e.type === 'context/message')).toBe(true)
  250. })
  251. it('agent/turn-continuation can force-continue (/loop pattern) and force-stop', async () => {
  252. // force-continue: model never calls tools, but a plugin forces 3 steps
  253. const adapter = new MockAdapter([
  254. textResponse('step 1'),
  255. textResponse('step 2'),
  256. textResponse('step 3'),
  257. ])
  258. const ctx = await harness(adapter)
  259. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  260. let steps = 0
  261. ctx.on('session/event', (_session, event) => { if (event.type === 'step/end') steps++ })
  262. ctx.on('agent/turn-continuation', async (_agent, _turn, _defaultDecision, next) => {
  263. if (steps < 3) return { action: 'continue' as const }
  264. return next()
  265. })
  266. send(agent, 'go')
  267. await waitForIdle(ctx, agent)
  268. expect(steps).toBe(3)
  269. expect(adapter.requests).toHaveLength(3)
  270. })
  271. it('agent/turn-continuation can veto continuation despite tool calls (budget-guard pattern)', async () => {
  272. const adapter = new MockAdapter([toolCallResponse('c1', 'echo', { text: 'x' })])
  273. const ctx = await harness(adapter)
  274. ctx.tools.register(defineTool({
  275. name: 'echo',
  276. description: '',
  277. parameters: { text: { type: 'string' } },
  278. async execute(args) {
  279. return [{ type: 'text', text: String(args.text) }]
  280. },
  281. }))
  282. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  283. ctx.on('agent/turn-continuation', async () => ({ action: 'stop' }) as const)
  284. send(agent, 'go')
  285. await waitForIdle(ctx, agent)
  286. // only one model call despite the tool call requesting a follow-up
  287. expect(adapter.requests).toHaveLength(1)
  288. // tool still executed before the decision
  289. expect(agent.session.events.some(e => e.type === 'tool/result')).toBe(true)
  290. })
  291. it('agent/request waterfall can rewrite the request (model-switch pattern)', async () => {
  292. const adapter = new MockAdapter([textResponse('ok')])
  293. const ctx = await harness(adapter)
  294. ctx.llm.registerAdapter(['other-model'], adapter)
  295. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  296. ctx.on('agent/request', async (_agent, _turn, _step, options, next) => {
  297. options.model = 'other-model'
  298. return next()
  299. })
  300. send(agent, 'hi')
  301. await waitForIdle(ctx, agent)
  302. expect(adapter.requests[0]!.model).toBe('other-model')
  303. })
  304. it('agent/pre-step fires once per step before the step is opened', async () => {
  305. // Two steps (a tool call, then a final text turn) → two model calls → two
  306. // pre-step fires, each carrying the assembled full system prompt, BEFORE
  307. // the step is opened and its request is derived (the request the adapter
  308. // sees reflects any surface state at fire time).
  309. const adapter = new MockAdapter([
  310. toolCallResponse('c1', 'echo', {}, 'calling echo'),
  311. textResponse('done'),
  312. ])
  313. const ctx = await harness(adapter)
  314. ctx.tools.register(defineTool({
  315. name: 'echo', description: 'echo', parameters: {},
  316. async execute() { return [{ type: 'text', text: 'echoed' }] },
  317. }))
  318. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  319. const fires: { turn: number; step: number; fullSystemPrompt: string }[] = []
  320. ctx.on('agent/pre-step', (subject, turn, step, fullSystemPrompt) => {
  321. if (subject === agent) fires.push({ turn, step, fullSystemPrompt })
  322. })
  323. send(agent, 'go')
  324. await waitForIdle(ctx, agent)
  325. // One fire per step, in order, each with the assembled system prompt.
  326. expect(fires).toEqual([
  327. { turn: 1, step: 1, fullSystemPrompt: '' },
  328. { turn: 1, step: 2, fullSystemPrompt: '' },
  329. ])
  330. })
  331. it('agent/pre-step fires BEFORE the step it precedes opens (events land outside the step)', async () => {
  332. // A listener appending a surface node in pre-step lands it BEFORE step/start
  333. // in the log — proving the seam fires outside the step. The node is still in
  334. // the derived request for that step (derive happens after step/start).
  335. const adapter = new MockAdapter([textResponse('ok')])
  336. const ctx = await harness(adapter)
  337. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  338. let injected = false
  339. ctx.on('agent/pre-step', (subject) => {
  340. if (subject === agent && !injected) {
  341. injected = true
  342. subject.session.append('context/message', {
  343. content: [{ type: 'text', text: 'INJECTED-IN-PRE-STEP' }],
  344. source: { kind: 'plugin', plugin: 'test' },
  345. }, { surfaceOp: 'append' })
  346. }
  347. })
  348. send(agent, 'go')
  349. await waitForIdle(ctx, agent)
  350. // The adapter's request includes the node injected during pre-step (derive
  351. // reflects it).
  352. const text = JSON.stringify(adapter.requests[0]!.messages)
  353. expect(text).toContain('INJECTED-IN-PRE-STEP')
  354. // And the injected event sits BEFORE the first step/start in the log —
  355. // the seam fired outside the step.
  356. const events = agent.session.events
  357. const injectedSeq = events.find(e => e.type === 'context/message')!.seq
  358. const firstStepStartSeq = events.find(e => e.type === 'step/start')!.seq
  359. expect(injectedSeq).toBeLessThan(firstStepStartSeq)
  360. })
  361. it('a throwing agent/pre-step listener ends the turn (error), not the loop', async () => {
  362. // The seam fires before step/start, so a throw escapes to runTurn's outer
  363. // catch: the not-yet-open step closes as a no-op, the failure surfaces via
  364. // agent/error, and the turn ends `error` (recorded on the durable turn/end).
  365. // The loop survives and a follow-up prompt still runs.
  366. const adapter = new MockAdapter([textResponse('second turn ok')])
  367. const ctx = await harness(adapter)
  368. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  369. let throwOnce = true
  370. ctx.on('agent/pre-step', () => {
  371. if (throwOnce) { throwOnce = false; throw new Error('boom in pre-step') }
  372. })
  373. const errors: Error[] = []
  374. ctx.on('agent/error', (_a, _t, _s, error) => void errors.push(error))
  375. send(agent, 'first')
  376. await waitForIdle(ctx, agent)
  377. // The first turn failed at step 1 (no model call happened), surfaced via
  378. // agent/error, with the durable failure on turn/end.reason.
  379. expect(errors).toHaveLength(1)
  380. expect(errors[0]!.message).toContain('boom in pre-step')
  381. expect(adapter.requests.length).toBe(0)
  382. const firstTurnEnd = agent.session.events.find(e => e.type === 'turn/end')
  383. expect(firstTurnEnd?.type === 'turn/end' && firstTurnEnd.data.reason).toMatchObject({ kind: 'error', step: 1 })
  384. // The step opened-and-closed count stays balanced even though it never ran.
  385. const types = agent.session.events.map(e => e.type)
  386. expect(types.filter(t => t === 'step/start').length).toBe(types.filter(t => t === 'step/end').length)
  387. // The loop survived: a second prompt runs a normal completed turn.
  388. send(agent, 'second')
  389. await waitForIdle(ctx, agent)
  390. expect(adapter.requests.length).toBe(1)
  391. const lastTurnEnd = agent.session.events.findLast(e => e.type === 'turn/end')
  392. expect(lastTurnEnd?.type === 'turn/end' && lastTurnEnd.data.reason).toEqual({ kind: 'completed' })
  393. })
  394. it('cancel() mid-stream ends the turn with reason aborted', async () => {
  395. const adapter = new MockAdapter(['hang'])
  396. const ctx = await harness(adapter)
  397. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  398. const reasons: TurnEndReason[] = []
  399. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  400. send(agent, 'go')
  401. // wait until the stream is hanging, then cancel
  402. await new Promise(r => setTimeout(r, 30))
  403. expect(agent.status).toBe('running')
  404. agent.cancel('user interrupt')
  405. await waitForIdle(ctx, agent)
  406. expect(reasons).toEqual([{ kind: 'aborted', reason: 'user interrupt' }])
  407. })
  408. it('surfaces max-tokens as the turn-end reason when the last step is cut off', async () => {
  409. // A single step that ends with a max-tokens finish (no tool calls): the
  410. // turn stops by default and ends max-tokens, not completed.
  411. const adapter = new MockAdapter([maxTokensResponse('truncat')])
  412. const ctx = await harness(adapter)
  413. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  414. const reasons: TurnEndReason[] = []
  415. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  416. send(agent, 'go')
  417. await waitForIdle(ctx, agent)
  418. expect(adapter.requests).toHaveLength(1)
  419. expect(reasons).toEqual([{ kind: 'max-tokens' }])
  420. // and the reason is recorded in the log's turn/end event
  421. const turnEnd = agent.session.events.findLast(e => e.type === 'turn/end')
  422. expect(turnEnd!.data.reason).toEqual({ kind: 'max-tokens' })
  423. })
  424. it('a max-tokens step earlier in a turn still surfaces as max-tokens after a later completed step', async () => {
  425. // Step 1 is cut off (max-tokens, no tool calls → would stop by default), so
  426. // continuation must be FORCED to reach step 2 which finishes normally
  427. // (stop). The rule "any max-tokens step surfaces as max-tokens" means the
  428. // turn ends max-tokens even though the LAST step completed cleanly.
  429. const adapter = new MockAdapter([
  430. maxTokensResponse('first half'),
  431. textResponse('second half'),
  432. ])
  433. const ctx = await harness(adapter)
  434. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  435. let steps = 0
  436. ctx.on('session/event', (_session, event) => { if (event.type === 'step/end') steps++ })
  437. // Force exactly one continuation (step 1 → step 2), then defer to default
  438. // (step 2 is a plain stop with no tool calls → stops).
  439. ctx.on('agent/turn-continuation', async (_agent, _turn, _defaultDecision, next) => {
  440. if (steps < 2) return { action: 'continue' as const }
  441. return next()
  442. })
  443. const reasons: TurnEndReason[] = []
  444. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  445. send(agent, 'go')
  446. await waitForIdle(ctx, agent)
  447. expect(steps).toBe(2)
  448. expect(adapter.requests).toHaveLength(2)
  449. expect(adapter.requests[1]!.messages).toEqual([
  450. { role: 'user', content: [{ type: 'text', text: 'go' }] },
  451. { role: 'assistant', content: [{ type: 'text', text: 'first half' }] },
  452. ])
  453. expect(reasons).toEqual([{ kind: 'max-tokens' }])
  454. })
  455. it('a completed step after no max-tokens keeps the turn completed (max-tokens does not leak across turns)', async () => {
  456. // Two consecutive turns: turn 1 is cut off (max-tokens), turn 2 is a clean
  457. // stop. The per-turn reason must be independent — turn 2 ends completed.
  458. const adapter = new MockAdapter([maxTokensResponse('cut'), textResponse('clean')])
  459. const ctx = await harness(adapter)
  460. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  461. const reasons: TurnEndReason[] = []
  462. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  463. send(agent, 'first')
  464. await waitForIdle(ctx, agent)
  465. send(agent, 'second')
  466. await waitForIdle(ctx, agent)
  467. expect(reasons).toEqual([{ kind: 'max-tokens' }, { kind: 'completed' }])
  468. })
  469. it('does not dispatch tool calls from a max-tokens-truncated step', async () => {
  470. const callId = CallId('c1')
  471. const adapter = new MockAdapter([[
  472. { type: 'block-start', index: 0, blockType: 'tool-call' },
  473. { type: 'tool-call-delta', index: 0, id: callId, name: 'echo', argumentsDelta: '{"text":"x"}' },
  474. { type: 'block-end', index: 0, block: { type: 'tool-call', id: callId, name: 'echo', arguments: '{"text":"x"}' } },
  475. { type: 'usage', usage: { inputTokens: 10, outputTokens: 5 } },
  476. { type: 'finish', reason: { kind: 'max-tokens' } },
  477. ]])
  478. const ctx = await harness(adapter)
  479. let executions = 0
  480. ctx.tools.register(defineTool({
  481. name: 'echo',
  482. description: '',
  483. parameters: { text: { type: 'string' } },
  484. async execute() {
  485. executions += 1
  486. return [{ type: 'text', text: 'should not run' }]
  487. },
  488. }))
  489. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  490. const reasons: TurnEndReason[] = []
  491. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  492. send(agent, 'go')
  493. await waitForIdle(ctx, agent)
  494. expect(executions).toBe(0)
  495. expect(agent.session.events.some(e => e.type === 'tool/call')).toBe(false)
  496. expect(agent.session.deriveMessages()).toEqual([{ role: 'user', content: [{ type: 'text', text: 'go' }] }])
  497. expect(reasons).toEqual([{ kind: 'max-tokens' }])
  498. // No-data-loss: a max-tokens step whose only content was a dropped tool call
  499. // has EMPTY assistant content, but its usage must still be represented. It
  500. // rides on an (empty-content) assistant/message — there is no standalone
  501. // usage event — and that empty message is skipped by deriveMessages(), so
  502. // the derived history above is NOT corrupted by a spurious assistant turn.
  503. const assistantMessage = agent.session.events.find(e => e.type === 'assistant/message')
  504. expect(assistantMessage?.type === 'assistant/message' && assistantMessage.data).toEqual({
  505. turn: 1, step: 1, content: [], usage: { inputTokens: 10, outputTokens: 5 },
  506. })
  507. })
  508. it('appends no assistant/message for a max-tokens step with empty content and no usage', async () => {
  509. // A max-tokens step truncated to a dropped tool call AND with no usage chunk
  510. // has nothing to record: empty content and no accounting → no assistant/message
  511. // (the empty-content host exists only to carry usage). The turn still ends
  512. // max-tokens.
  513. const callId = CallId('c1')
  514. const adapter = new MockAdapter([[
  515. { type: 'block-start', index: 0, blockType: 'tool-call' },
  516. { type: 'tool-call-delta', index: 0, id: callId, name: 'echo', argumentsDelta: '{"text":"x"}' },
  517. { type: 'block-end', index: 0, block: { type: 'tool-call', id: callId, name: 'echo', arguments: '{"text":"x"}' } },
  518. { type: 'finish', reason: { kind: 'max-tokens' } },
  519. ]])
  520. const ctx = await harness(adapter)
  521. ctx.tools.register(defineTool({
  522. name: 'echo',
  523. description: '',
  524. parameters: { text: { type: 'string' } },
  525. async execute() { return [{ type: 'text', text: 'should not run' }] },
  526. }))
  527. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  528. const reasons: TurnEndReason[] = []
  529. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  530. send(agent, 'go')
  531. await waitForIdle(ctx, agent)
  532. expect(reasons).toEqual([{ kind: 'max-tokens' }])
  533. expect(agent.session.events.some(e => e.type === 'assistant/message')).toBe(false)
  534. expect(agent.session.deriveMessages()).toEqual([{ role: 'user', content: [{ type: 'text', text: 'go' }] }])
  535. })
  536. it('appends no assistant/message for a normal stop finish with empty content and no usage', async () => {
  537. // A clean `stop` finish that streamed nothing assembled (no blocks) and
  538. // carried no usage chunk has nothing to record: the content-or-usage guard
  539. // on the normal step path suppresses a pure trace-only empty assistant/message.
  540. const adapter = new MockAdapter([[{ type: 'finish', reason: { kind: 'stop' } }]])
  541. const ctx = await harness(adapter)
  542. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  543. const reasons: TurnEndReason[] = []
  544. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  545. send(agent, 'go')
  546. await waitForIdle(ctx, agent)
  547. expect(reasons).toEqual([{ kind: 'completed' }])
  548. expect(agent.session.events.some(e => e.type === 'assistant/message')).toBe(false)
  549. expect(agent.session.deriveMessages()).toEqual([{ role: 'user', content: [{ type: 'text', text: 'go' }] }])
  550. })
  551. it('keeps safe max-tokens assistant content while dropping truncated tool calls', async () => {
  552. const callId = CallId('c1')
  553. const adapter = new MockAdapter([[
  554. { type: 'block-start', index: 0, blockType: 'text' },
  555. { type: 'text-delta', index: 0, text: 'partial text' },
  556. { type: 'block-end', index: 0, block: { type: 'text', text: 'partial text' } },
  557. { type: 'block-start', index: 1, blockType: 'tool-call' },
  558. { type: 'tool-call-delta', index: 1, id: callId, name: 'echo', argumentsDelta: '{"text"' },
  559. { type: 'finish', reason: { kind: 'max-tokens' } },
  560. ]])
  561. const ctx = await harness(adapter)
  562. let stepResults = 0
  563. ctx.on('agent/step-result', async (_agent, _turn, _step, message, next) => {
  564. stepResults += 1
  565. expect(message.content).toEqual([{ type: 'text', text: 'partial text' }])
  566. return next()
  567. })
  568. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  569. send(agent, 'go')
  570. await waitForIdle(ctx, agent)
  571. expect(stepResults).toBe(1)
  572. expect(agent.session.events.some(e => e.type === 'tool/call')).toBe(false)
  573. expect(agent.session.deriveMessages()).toEqual([
  574. { role: 'user', content: [{ type: 'text', text: 'go' }] },
  575. { role: 'assistant', content: [{ type: 'text', text: 'partial text' }] },
  576. ])
  577. })
  578. it('stops the turn when a step/end session-event listener failure has recorded an error', async () => {
  579. const adapter = new MockAdapter([
  580. toolCallResponse('c1', 'echo', { text: 'x' }),
  581. textResponse('should not run'),
  582. ])
  583. const ctx = await harness(adapter)
  584. ctx.tools.register(defineTool({
  585. name: 'echo',
  586. description: '',
  587. parameters: { text: { type: 'string' } },
  588. async execute(args) {
  589. return [{ type: 'text', text: String(args.text) }]
  590. },
  591. }))
  592. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  593. let threw = false
  594. // A throwing step/end session-event listener is the surviving boundary-listener
  595. // failure path (step boundaries have no agent/* mirror): closeStep contains it
  596. // and surfaces it as a turn error rather than stranding the turn open.
  597. ctx.on('session/event', (_session, event) => {
  598. if (event.type === 'step/end' && !threw) { threw = true; throw new Error('bad step/end listener') }
  599. })
  600. send(agent, 'go')
  601. await waitForIdle(ctx, agent)
  602. expect(adapter.requests).toHaveLength(1)
  603. const turnEnd = agent.session.events.findLast(e => e.type === 'turn/end')
  604. expect(turnEnd?.type === 'turn/end' && turnEnd.data.reason.kind).toBe('error')
  605. })
  606. it('chains queued messages into consecutive turns', async () => {
  607. const adapter = new MockAdapter([textResponse('first'), textResponse('second')])
  608. const ctx = await harness(adapter)
  609. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  610. const turns: number[] = []
  611. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/start') turns.push(event.data.turn) })
  612. // queue two messages while idle — first starts turn 1 immediately;
  613. // queue the second during turn 1 when the first assistant chunk streams
  614. let queued = false
  615. ctx.on('session/event', (_s, event) => {
  616. if (event.type === 'assistant/chunk' && !queued) {
  617. queued = true
  618. send(agent, 'second message')
  619. }
  620. })
  621. send(agent, 'first message')
  622. await waitForIdle(ctx, agent)
  623. expect(turns).toEqual([1, 2])
  624. expect(adapter.requests).toHaveLength(2)
  625. })
  626. it('awaits session/flush at turn end (persistence checkpoint)', async () => {
  627. const adapter = new MockAdapter([textResponse('ok')])
  628. const ctx = await harness(adapter)
  629. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  630. let flushed = 0
  631. let flushedBeforeIdle = false
  632. ctx.on('session/flush', async (session) => {
  633. await new Promise(r => setTimeout(r, 10))
  634. flushed++
  635. flushedBeforeIdle = agent.status !== 'idle'
  636. void session
  637. })
  638. send(agent, 'hi')
  639. await waitForIdle(ctx, agent)
  640. expect(flushed).toBe(1)
  641. expect(flushedBeforeIdle).toBe(true)
  642. })
  643. it('errors from the model surface as agent/error and end the turn', async () => {
  644. const adapter = new MockAdapter([]) // script exhausted → throws
  645. const ctx = await harness(adapter)
  646. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  647. const errors: Error[] = []
  648. const reasons: TurnEndReason[] = []
  649. ctx.on('agent/error', (_agent, _turn, _step, error) => void errors.push(error))
  650. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  651. send(agent, 'hi')
  652. await waitForIdle(ctx, agent)
  653. expect(errors).toHaveLength(1)
  654. expect(errors[0]!.message).toContain('script exhausted')
  655. expect(reasons[0]).toMatchObject({ kind: 'error' })
  656. // The durable failure lives entirely on turn/end.reason (with the failing
  657. // step), not a standalone error event.
  658. const turnEnd = agent.session.events.find(e => e.type === 'turn/end')
  659. expect(turnEnd?.type === 'turn/end' && turnEnd.data.reason).toMatchObject({ kind: 'error', step: 1 })
  660. })
  661. it('disposing the loop fiber mid-turn stops the loop (HMR safety)', async () => {
  662. const adapter = new MockAdapter(['hang'])
  663. const ctx = await harness(adapter)
  664. let agent!: ReactLoopAgent
  665. const fiber = await ctx.plugin(Object.assign((inner: Context) => {
  666. agent = inner.agentLoop.create(AgentId('scoped'), { model: 'mock' })
  667. }, { inject: ['agentLoop'] }))
  668. expect(ctx.agents.get(AgentId('scoped'))).toBe(agent)
  669. send(agent, 'go')
  670. await new Promise(r => setTimeout(r, 30))
  671. expect(agent.status).toBe('running')
  672. await fiber.dispose()
  673. await agent.done
  674. expect(agent.status).toBe('disposed')
  675. expect(ctx.agents.get(AgentId('scoped'))).toBeUndefined()
  676. expect(() => { send(agent, 'too late') }).toThrow('disposed')
  677. })
  678. it('creates agents from config on startup', async () => {
  679. const adapter = new MockAdapter([textResponse('from config')])
  680. const ctx = new Context()
  681. await ctx.plugin(LlmService)
  682. await ctx.plugin(SessionStore)
  683. await ctx.plugin(SystemPrompt)
  684. await ctx.plugin(ToolRegistry)
  685. await ctx.plugin(AgentRegistry)
  686. await ctx.plugin(AgentLoop, {
  687. agents: [{ id: AgentId('config-agent'), model: 'mock', systemPrompt: 'Config prompt' }],
  688. })
  689. ctx.llm.registerAdapter(['mock'], adapter)
  690. const agent = ctx.agents.get(AgentId('config-agent'))! as ReactLoopAgent
  691. expect(agent).toBeDefined()
  692. expect(agent.id).toBe('config-agent')
  693. expect(agent.options.model).toBe('mock')
  694. // the agent is alive: send triggers a turn
  695. send(agent, 'hi')
  696. await waitForIdle(ctx, agent)
  697. expect(adapter.requests).toHaveLength(1)
  698. })
  699. it('replays a session log into an identical derived history', async () => {
  700. const adapter = new MockAdapter([
  701. toolCallResponse('c1', 'echo', { text: 'x' }),
  702. textResponse('done'),
  703. ])
  704. const ctx = await harness(adapter)
  705. ctx.tools.register(defineTool({
  706. name: 'echo',
  707. description: '',
  708. parameters: { text: { type: 'string' } },
  709. async execute(args) {
  710. return [{ type: 'text', text: String(args.text) }]
  711. },
  712. }))
  713. const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
  714. send(agent, 'run')
  715. await waitForIdle(ctx, agent)
  716. const replayed = ctx.sessions.create(SessionId('replayed'), { seed: [...agent.session.events] })
  717. expect(replayed.deriveMessages()).toEqual(agent.session.deriveMessages())
  718. // event-by-event identity of types
  719. expect(replayed.events.map(e => e.type)).toEqual(
  720. agent.session.events.map(e => e.type))
  721. })
  722. })