loop.spec.ts 44 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020
  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, persona = '') {
  11. const ctx = new Context()
  12. await ctx.plugin(LlmService)
  13. await ctx.plugin(SessionStore)
  14. await ctx.plugin(SystemPrompt, { persona })
  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'), { provider: 'mock', 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'), { provider: 'mock', 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'), { provider: 'mock', 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('renders harness identity, then the persona, then tool guidance — with {{variables}} resolved', async () => {
  127. const adapter = new MockAdapter([textResponse('ok')])
  128. // The persona is a TEMPLATE: {{model}} is the loop-registered variable
  129. // projecting this agent's configured model, so the model knows its own name.
  130. const ctx = await harness(adapter, 'You are a test agent on {{model}}.')
  131. ctx.systemPrompt.section({ name: 'tool:noop', order: 100, text: 'Use the noop tool wisely.' })
  132. ctx.tools.register(defineTool({
  133. name: 'noop',
  134. description: 'does nothing',
  135. parameters: {},
  136. async execute() {
  137. return []
  138. },
  139. }))
  140. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  141. send(agent, 'hi')
  142. await waitForIdle(ctx, agent)
  143. const request = adapter.requests[0]
  144. expect(request!.system).toBe('You are an AI agent powered by the DeepSeek Harness SDK.\n\nYou are a test agent on mock.\n\nUse the noop tool wisely.')
  145. expect(request!.tools?.map(t => t.name)).toEqual(['noop'])
  146. })
  147. it('resolves {{cwd}} from the agent session workspace (factory create with meta.cwd)', async () => {
  148. const adapter = new MockAdapter([textResponse('ok')])
  149. const ctx = await harness(adapter, 'Working in {{cwd}}.')
  150. const handle = await ctx.agents.create({
  151. agentId: AgentId('a-cwd'),
  152. sessionId: SessionId('s-cwd'),
  153. meta: { cwd: '/work/space' },
  154. agentOptions: { provider: 'mock', model: 'mock' },
  155. })
  156. const agent = handle.agent as ReactLoopAgent
  157. send(agent, 'hi')
  158. await waitForIdle(ctx, agent)
  159. expect(adapter.requests[0]!.system).toBe('You are an AI agent powered by the DeepSeek Harness SDK.\n\nWorking in /work/space.')
  160. })
  161. it('contains a strict-variable render failure: the turn errors, the loop keeps serving turns', async () => {
  162. // A missing cwd variable must fail one turn without preventing a later valid turn.
  163. const adapter = new MockAdapter([textResponse('ok after rescue')])
  164. const ctx = await harness(adapter, 'In {{cwd}}.')
  165. const errors: Error[] = []
  166. ctx.on('agent/error', (_agent, _turn, _step, error) => void errors.push(error))
  167. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  168. send(agent, 'hi')
  169. await waitForIdle(ctx, agent)
  170. expect(adapter.requests).toHaveLength(0) // the request was never sent
  171. expect(errors.some(e => e.message.includes('no value for this assembly'))).toBe(true)
  172. const turnEnd = agent.session.events.find(e => e.type === 'turn/end')
  173. expect(turnEnd?.type === 'turn/end' && turnEnd.data.reason.kind).toBe('error')
  174. // The loop survived: a waterfall listener rescues {{cwd}} and the SAME
  175. // agent completes a real model turn.
  176. ctx.on('system-prompt/assemble', async (assembly, _context, next) => {
  177. assembly.variables['cwd'] = '/rescued'
  178. return next()
  179. })
  180. send(agent, 'again')
  181. await waitForIdle(ctx, agent)
  182. expect(adapter.requests).toHaveLength(1)
  183. expect(adapter.requests[0]!.system).toBe('You are an AI agent powered by the DeepSeek Harness SDK.\n\nIn /rescued.')
  184. const turnEnds = agent.session.events.filter(e => e.type === 'turn/end')
  185. expect(turnEnds).toHaveLength(2)
  186. expect(turnEnds[1]?.type === 'turn/end' && turnEnds[1].data.reason.kind).toBe('completed')
  187. })
  188. it('supports the model-via-agent/request path with a {{model}} persona: the supplier states it via the assemble waterfall', async () => {
  189. // AgentOptions.model unset: the model arrives in the agent/request
  190. // waterfall (the loop's documented fallback — see runStep's no-model
  191. // error). {{model}} renders BEFORE that waterfall, so the SAME plugin
  192. // states the fact early on system-prompt/assemble — the owner of a
  193. // late-bound fact owns stating it wherever it is claimed.
  194. const adapter = new MockAdapter([textResponse('ok')])
  195. const ctx = await harness(adapter, 'You run on {{model}}.')
  196. ctx.on('system-prompt/assemble', async (assembly, _context, next) => {
  197. assembly.variables['provider'] = 'mock'
  198. assembly.variables['model'] = 'mock'
  199. return next()
  200. })
  201. ctx.on('agent/request', async (_agent, _turn, _step, config, _next) => {
  202. return { ...config, provider: 'mock', model: 'mock' }
  203. })
  204. const agent = ctx.agentLoop.create(AgentId('a-late-model'), {})
  205. send(agent, 'hi')
  206. await waitForIdle(ctx, agent)
  207. expect(adapter.requests).toHaveLength(1)
  208. expect(adapter.requests[0]!.model).toBe('mock')
  209. expect(adapter.requests[0]!.system).toBe('You are an AI agent powered by the DeepSeek Harness SDK.\n\nYou run on mock.')
  210. })
  211. it.each([
  212. ['BigInt', { n: 1n }],
  213. ['Map', new Map([['key', 'value']])],
  214. ['class instance', new (class ResultMeta { x = 1 })()],
  215. ])('normalizes non-JSON tool meta (%s) before the durable result commit', async (_kind, meta) => {
  216. const adapter = new MockAdapter([
  217. toolCallResponse('bad-meta-call', 'bad-meta', {}, 'calling'),
  218. textResponse('recovered'),
  219. ])
  220. const ctx = await harness(adapter)
  221. ctx.tools.register(defineTool({
  222. name: 'bad-meta',
  223. description: 'returns invalid durable metadata',
  224. parameters: {},
  225. execute: () => Promise.resolve({ content: [{ type: 'text' as const, text: 'apparent success' }], meta }),
  226. }))
  227. const agent = ctx.agentLoop.create(AgentId('bad-meta-agent'), { provider: 'mock', model: 'mock' })
  228. send(agent, 'use the tool')
  229. await waitForIdle(ctx, agent)
  230. const result = agent.session.events.find(event => event.type === 'tool/result')
  231. expect(result?.type).toBe('tool/result')
  232. if (result?.type === 'tool/result') {
  233. expect(result.data.callId).toBe('bad-meta-call')
  234. expect(result.data.isError).toBe(true)
  235. expect(result.data.meta).toBeUndefined()
  236. expect(result.data.content).toEqual([{
  237. type: 'text',
  238. text: 'Error: tool result must be losslessly JSON-serializable',
  239. }])
  240. }
  241. // The normalized failure was durably logged and fed back to the model; the
  242. // turn continued normally instead of failing after an apparent success.
  243. expect(adapter.requests).toHaveLength(2)
  244. expect(JSON.stringify(adapter.requests[1]!.messages)).toContain('losslessly JSON-serializable')
  245. })
  246. it('omits the system field when a system-prompt/assemble veto empties the assembly', async () => {
  247. // The documented escape valve: a deployment that must drop the harness
  248. // openers short-circuits the assemble waterfall; the request then carries
  249. // NO system field at all (not an empty string).
  250. const adapter = new MockAdapter([textResponse('ok')])
  251. const ctx = await harness(adapter)
  252. ctx.on('system-prompt/assemble', async () => ({ sections: [], tools: [], variables: {} }))
  253. const agent = ctx.agentLoop.create(AgentId('a-no-system'), { provider: 'mock', model: 'mock' })
  254. send(agent, 'hi')
  255. await waitForIdle(ctx, agent)
  256. expect(adapter.requests).toHaveLength(1)
  257. expect('system' in adapter.requests[0]!).toBe(false)
  258. })
  259. it('records raw chunks for replay as assistant/chunk session events', async () => {
  260. const adapter = new MockAdapter([textResponse('abc')])
  261. const ctx = await harness(adapter)
  262. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  263. send(agent, 'hi')
  264. await waitForIdle(ctx, agent)
  265. const chunkEvents = agent.session.events.filter(e => e.type === 'assistant/chunk')
  266. // textResponse('abc') = block-start + 3 deltas + block-end + usage + finish = 7
  267. expect(chunkEvents).toHaveLength(7)
  268. // replay: chunk events alone re-assemble to the recorded assistant message
  269. const deltaText = chunkEvents
  270. .flatMap(e => e.type === 'assistant/chunk' ? [e.data.chunk] : [])
  271. .filter((c: StreamChunk): c is Extract<StreamChunk, { type: 'text-delta' }> => c.type === 'text-delta')
  272. .map(c => c.text)
  273. .join('')
  274. expect(deltaText).toBe('abc')
  275. })
  276. it('injects steering between steps and continues the turn', async () => {
  277. const adapter = new MockAdapter([
  278. toolCallResponse('c1', 'slow', {}),
  279. textResponse('addressed the steering'),
  280. ])
  281. const ctx = await harness(adapter)
  282. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  283. ctx.tools.register(defineTool({
  284. name: 'slow',
  285. description: '',
  286. parameters: {},
  287. async execute() {
  288. // steer while the turn is running (during tool execution)
  289. agent.steer([{ type: 'text', text: 'change of plans' }])
  290. return [{ type: 'text', text: 'tool done' }]
  291. },
  292. }))
  293. send(agent, 'start')
  294. await waitForIdle(ctx, agent)
  295. const types = agent.session.events.map(e => e.type)
  296. expect(types).toContain('steering/message')
  297. // steering recorded before the second step's request derived its history
  298. const steeringSeq = agent.session.events.find(e => e.type === 'steering/message')!.seq
  299. const secondStepStart = agent.session.events.filter(e => e.type === 'step/start')[1]
  300. expect(secondStepStart).toBeDefined()
  301. expect(steeringSeq).toBeLessThan(secondStepStart!.seq)
  302. // the second model request saw the steering content
  303. const secondRequest = adapter.requests[1]
  304. const flat = JSON.stringify(secondRequest!.messages)
  305. expect(flat).toContain('change of plans')
  306. })
  307. it('steering while idle behaves like send (starts a turn)', async () => {
  308. const adapter = new MockAdapter([textResponse('ok')])
  309. const ctx = await harness(adapter)
  310. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  311. agent.steer([{ type: 'text', text: 'hello' }])
  312. await waitForIdle(ctx, agent)
  313. expect(agent.session.events.some(e => e.type === 'user/message')).toBe(true)
  314. })
  315. it('inject() while idle wraps context in a one-shot turn, visible to the next request', async () => {
  316. const adapter = new MockAdapter([textResponse('ok')])
  317. const ctx = await harness(adapter)
  318. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  319. agent.inject([{ type: 'text', text: 'file changed: a.ts' }], { source: { kind: 'plugin', plugin: 'watcher' } })
  320. // The idle inject records a self-contained turn (turn/start → context/message
  321. // → turn/end) so the event stays turn-enclosed, but does NOT run the model.
  322. await new Promise(r => setTimeout(r, 20))
  323. expect(agent.status).toBe('idle')
  324. expect(adapter.requests).toHaveLength(0)
  325. const injectedTurn = agent.session.events.filter(e => e.type === 'turn/start')
  326. expect(injectedTurn).toHaveLength(1)
  327. const it0 = injectedTurn[0]!
  328. expect(it0.type === 'turn/start' && it0.data.trigger.kind).toBe('injection')
  329. expect(agent.session.events.at(-1)!.type).toBe('turn/end') // turn-enclosed
  330. send(agent, 'go')
  331. await waitForIdle(ctx, agent)
  332. const flat = JSON.stringify(adapter.requests[0]!.messages)
  333. expect(flat).toContain('file changed: a.ts')
  334. expect(flat).toContain('<context source=\\"plugin\\">')
  335. })
  336. it('inject() can persist raw structured context without the generic context envelope', async () => {
  337. const adapter = new MockAdapter([textResponse('ok')])
  338. const ctx = await harness(adapter)
  339. const agent = ctx.agentLoop.create(AgentId('raw-context'), { provider: 'mock', model: 'mock' })
  340. const text = '<system-reminder>Additional instructions from: pkg/AGENTS.md</system-reminder>'
  341. const meta = {
  342. kind: 'workspace-instructions',
  343. version: 1,
  344. changes: [{ action: 'set', scope: 'pkg', path: 'pkg/AGENTS.md', digest: 'abc123' }],
  345. }
  346. agent.inject([{ type: 'text', text }], {
  347. source: { kind: 'plugin', plugin: 'workspace-context' },
  348. envelope: 'raw',
  349. meta,
  350. })
  351. send(agent, 'go')
  352. await waitForIdle(ctx, agent)
  353. const contextEvent = agent.session.events.find(event => event.type === 'context/message')
  354. expect(contextEvent?.type === 'context/message' && contextEvent.data).toMatchObject({ envelope: 'raw', meta })
  355. const requestText = JSON.stringify(adapter.requests[0]!.messages)
  356. expect(requestText).toContain('Additional instructions from: pkg/AGENTS.md')
  357. expect(requestText).not.toContain('<context source=')
  358. })
  359. it('inject() while running appends into the open turn (no extra synthetic turn)', async () => {
  360. const adapter = new MockAdapter([
  361. toolCallResponse('c1', 'noticer', {}, 'calling'),
  362. textResponse('done'),
  363. ])
  364. const ctx = await harness(adapter)
  365. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  366. // A tool that injects mid-execution: at this point the agent is running, so
  367. // inject must append the context/message into the ALREADY-open turn rather
  368. // than wrap it in its own one-shot turn.
  369. ctx.tools.register(defineTool({
  370. name: 'noticer',
  371. description: 'injects a notice',
  372. parameters: {},
  373. async execute() {
  374. agent.inject([{ type: 'text', text: 'mid-turn notice' }], { source: { kind: 'plugin', plugin: 'x' } })
  375. return [{ type: 'text', text: 'ok' }]
  376. },
  377. }))
  378. send(agent, 'go')
  379. await waitForIdle(ctx, agent)
  380. // Exactly ONE turn ran (no synthetic injection turn), and the mid-turn
  381. // context/message sits inside it.
  382. const turnStarts = agent.session.events.filter(e => e.type === 'turn/start')
  383. expect(turnStarts).toHaveLength(1)
  384. const ts0 = turnStarts[0]!
  385. expect(ts0.type === 'turn/start' && ts0.data.trigger.kind).toBe('message')
  386. expect(agent.session.events.some(e => e.type === 'context/message')).toBe(true)
  387. })
  388. it('agent/turn-continuation can force-continue (/loop pattern) and force-stop', async () => {
  389. // force-continue: model never calls tools, but a plugin forces 3 steps
  390. const adapter = new MockAdapter([
  391. textResponse('step 1'),
  392. textResponse('step 2'),
  393. textResponse('step 3'),
  394. ])
  395. const ctx = await harness(adapter)
  396. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  397. let steps = 0
  398. ctx.on('session/event', (_session, event) => { if (event.type === 'step/end') steps++ })
  399. ctx.on('agent/turn-continuation', async (_agent, _turn, _defaultDecision, next) => {
  400. if (steps < 3) return { action: 'continue' as const }
  401. return next()
  402. })
  403. send(agent, 'go')
  404. await waitForIdle(ctx, agent)
  405. expect(steps).toBe(3)
  406. expect(adapter.requests).toHaveLength(3)
  407. })
  408. it('agent/turn-continuation can veto continuation despite tool calls (budget-guard pattern)', async () => {
  409. const adapter = new MockAdapter([toolCallResponse('c1', 'echo', { text: 'x' })])
  410. const ctx = await harness(adapter)
  411. ctx.tools.register(defineTool({
  412. name: 'echo',
  413. description: '',
  414. parameters: { text: { type: 'string' } },
  415. async execute(args) {
  416. return [{ type: 'text', text: String(args.text) }]
  417. },
  418. }))
  419. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  420. ctx.on('agent/turn-continuation', async () => ({ action: 'stop' }) as const)
  421. send(agent, 'go')
  422. await waitForIdle(ctx, agent)
  423. // only one model call despite the tool call requesting a follow-up
  424. expect(adapter.requests).toHaveLength(1)
  425. // tool still executed before the decision
  426. expect(agent.session.events.some(e => e.type === 'tool/result')).toBe(true)
  427. })
  428. it('agent/request waterfall switches models by returning a replacement config; the switch is logged', async () => {
  429. const adapter = new MockAdapter([textResponse('ok')])
  430. const ctx = await harness(adapter)
  431. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  432. ctx.on('agent/request', async (_agent, _turn, _step, config, _next) => {
  433. // The seed is frozen — config is not a mutable per-call knob; a switch
  434. // is proposed by returning a replacement, and the loop logs it.
  435. expect(Object.isFrozen(config)).toBe(true)
  436. expect(() => { (config as { model: string }).model = 'other-model' }).toThrow(TypeError)
  437. return { ...config, model: 'other-model' }
  438. })
  439. send(agent, 'hi')
  440. await waitForIdle(ctx, agent)
  441. expect(adapter.requests[0]!.model).toBe('other-model')
  442. // The header event records what the request ACTUALLY used — the switch is
  443. // a reconstructable fact, not silent drift.
  444. const headerEvent = agent.session.events.find(e => e.type === 'request/header')
  445. expect(headerEvent?.type === 'request/header' && headerEvent.data.header.config.model).toBe('other-model')
  446. })
  447. it('agent/pre-step fires once per step before the step is opened', async () => {
  448. // Two steps (a tool call, then a final text turn) → two model calls → two
  449. // pre-step fires, each carrying the assembled full system prompt, BEFORE
  450. // the step is opened and its request is derived (the request the adapter
  451. // sees reflects any surface state at fire time).
  452. const adapter = new MockAdapter([
  453. toolCallResponse('c1', 'echo', {}, 'calling echo'),
  454. textResponse('done'),
  455. ])
  456. const ctx = await harness(adapter)
  457. ctx.tools.register(defineTool({
  458. name: 'echo', description: 'echo', parameters: {},
  459. async execute() { return [{ type: 'text', text: 'echoed' }] },
  460. }))
  461. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  462. const fires: { turn: number; step: number; fullSystemPrompt: string }[] = []
  463. ctx.on('agent/pre-step', (subject, turn, step, fullSystemPrompt) => {
  464. if (subject === agent) fires.push({ turn, step, fullSystemPrompt })
  465. })
  466. send(agent, 'go')
  467. await waitForIdle(ctx, agent)
  468. // One fire per step, in order, each with the assembled system prompt
  469. // (here just the loop's own harness-identity section — no persona set).
  470. const HARNESS = 'You are an AI agent powered by the DeepSeek Harness SDK.'
  471. expect(fires).toEqual([
  472. { turn: 1, step: 1, fullSystemPrompt: HARNESS },
  473. { turn: 1, step: 2, fullSystemPrompt: HARNESS },
  474. ])
  475. })
  476. it('agent/pre-step fires BEFORE the step it precedes opens (events land outside the step)', async () => {
  477. // The append lands before step/start, yet derive happens afterwards and the
  478. // same step's request must include it.
  479. const adapter = new MockAdapter([textResponse('ok')])
  480. const ctx = await harness(adapter)
  481. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  482. let injected = false
  483. ctx.on('agent/pre-step', (subject) => {
  484. if (subject === agent && !injected) {
  485. injected = true
  486. subject.session.append('context/message', {
  487. content: [{ type: 'text', text: 'INJECTED-IN-PRE-STEP' }],
  488. source: { kind: 'plugin', plugin: 'test' },
  489. }, { surfaceOp: 'append' })
  490. }
  491. })
  492. send(agent, 'go')
  493. await waitForIdle(ctx, agent)
  494. // The adapter's request includes the node injected during pre-step (derive
  495. // reflects it).
  496. const text = JSON.stringify(adapter.requests[0]!.messages)
  497. expect(text).toContain('INJECTED-IN-PRE-STEP')
  498. // And the injected event sits BEFORE the first step/start in the log —
  499. // the seam fired outside the step.
  500. const events = agent.session.events
  501. const injectedSeq = events.find(e => e.type === 'context/message')!.seq
  502. const firstStepStartSeq = events.find(e => e.type === 'step/start')!.seq
  503. expect(injectedSeq).toBeLessThan(firstStepStartSeq)
  504. })
  505. it('a throwing agent/pre-step listener ends the turn (error), not the loop', async () => {
  506. // Before step/start, a pre-step throw reaches the turn catch: no step needs
  507. // closing, the turn records error, and the loop remains available.
  508. const adapter = new MockAdapter([textResponse('second turn ok')])
  509. const ctx = await harness(adapter)
  510. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  511. let throwOnce = true
  512. ctx.on('agent/pre-step', () => {
  513. if (throwOnce) { throwOnce = false; throw new Error('boom in pre-step') }
  514. })
  515. const errors: Error[] = []
  516. ctx.on('agent/error', (_a, _t, _s, error) => void errors.push(error))
  517. send(agent, 'first')
  518. await waitForIdle(ctx, agent)
  519. // The first turn failed at step 1 (no model call happened), surfaced via
  520. // agent/error, with the durable failure on turn/end.reason.
  521. expect(errors).toHaveLength(1)
  522. expect(errors[0]!.message).toContain('boom in pre-step')
  523. expect(adapter.requests.length).toBe(0)
  524. const firstTurnEnd = agent.session.events.find(e => e.type === 'turn/end')
  525. expect(firstTurnEnd?.type === 'turn/end' && firstTurnEnd.data.reason).toMatchObject({ kind: 'error', step: 1 })
  526. // The step opened-and-closed count stays balanced even though it never ran.
  527. const types = agent.session.events.map(e => e.type)
  528. expect(types.filter(t => t === 'step/start').length).toBe(types.filter(t => t === 'step/end').length)
  529. // The loop survived: a second prompt runs a normal completed turn.
  530. send(agent, 'second')
  531. await waitForIdle(ctx, agent)
  532. expect(adapter.requests.length).toBe(1)
  533. const lastTurnEnd = agent.session.events.findLast(e => e.type === 'turn/end')
  534. expect(lastTurnEnd?.type === 'turn/end' && lastTurnEnd.data.reason).toEqual({ kind: 'completed' })
  535. })
  536. it('cancel() mid-stream ends the turn with reason aborted', async () => {
  537. const adapter = new MockAdapter(['hang'])
  538. const ctx = await harness(adapter)
  539. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  540. const reasons: TurnEndReason[] = []
  541. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  542. send(agent, 'go')
  543. // wait until the stream is hanging, then cancel
  544. await new Promise(r => setTimeout(r, 30))
  545. expect(agent.status).toBe('running')
  546. agent.cancel('user interrupt')
  547. await waitForIdle(ctx, agent)
  548. expect(reasons).toEqual([{ kind: 'aborted', reason: 'user interrupt' }])
  549. })
  550. it('surfaces max-tokens as the turn-end reason when the last step is cut off', async () => {
  551. // A single step that ends with a max-tokens finish (no tool calls): the
  552. // turn stops by default and ends max-tokens, not completed.
  553. const adapter = new MockAdapter([maxTokensResponse('truncat')])
  554. const ctx = await harness(adapter)
  555. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  556. const reasons: TurnEndReason[] = []
  557. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  558. send(agent, 'go')
  559. await waitForIdle(ctx, agent)
  560. expect(adapter.requests).toHaveLength(1)
  561. expect(reasons).toEqual([{ kind: 'max-tokens' }])
  562. // Assert the durable row, not only the live listener.
  563. const turnEnd = agent.session.events.findLast(e => e.type === 'turn/end')
  564. expect(turnEnd!.data.reason).toEqual({ kind: 'max-tokens' })
  565. })
  566. it('a max-tokens step earlier in a turn still surfaces as max-tokens after a later completed step', async () => {
  567. // Step 1 is cut off (max-tokens, no tool calls → would stop by default), so continuation
  568. // must be FORCED to reach step 2 which finishes normally (stop).
  569. const adapter = new MockAdapter([
  570. maxTokensResponse('first half'),
  571. textResponse('second half'),
  572. ])
  573. const ctx = await harness(adapter)
  574. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  575. let steps = 0
  576. ctx.on('session/event', (_session, event) => { if (event.type === 'step/end') steps++ })
  577. // Force exactly one continuation (step 1 → step 2), then defer to default
  578. // (step 2 is a plain stop with no tool calls → stops).
  579. ctx.on('agent/turn-continuation', async (_agent, _turn, _defaultDecision, next) => {
  580. if (steps < 2) return { action: 'continue' as const }
  581. return next()
  582. })
  583. const reasons: TurnEndReason[] = []
  584. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  585. send(agent, 'go')
  586. await waitForIdle(ctx, agent)
  587. expect(steps).toBe(2)
  588. expect(adapter.requests).toHaveLength(2)
  589. expect(adapter.requests[1]!.messages).toEqual([
  590. { role: 'user', content: [{ type: 'text', text: 'go' }] },
  591. { role: 'assistant', content: [{ type: 'text', text: 'first half' }], provenance: { provider: 'mock', model: 'mock' } },
  592. ])
  593. expect(reasons).toEqual([{ kind: 'max-tokens' }])
  594. })
  595. it('a completed step after no max-tokens keeps the turn completed (max-tokens does not leak across turns)', async () => {
  596. // Two consecutive turns: turn 1 is cut off (max-tokens), turn 2 is a clean
  597. // stop. The per-turn reason must be independent — turn 2 ends completed.
  598. const adapter = new MockAdapter([maxTokensResponse('cut'), textResponse('clean')])
  599. const ctx = await harness(adapter)
  600. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  601. const reasons: TurnEndReason[] = []
  602. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  603. send(agent, 'first')
  604. await waitForIdle(ctx, agent)
  605. send(agent, 'second')
  606. await waitForIdle(ctx, agent)
  607. expect(reasons).toEqual([{ kind: 'max-tokens' }, { kind: 'completed' }])
  608. })
  609. it('does not dispatch tool calls from a max-tokens-truncated step', async () => {
  610. const callId = CallId('c1')
  611. const adapter = new MockAdapter([[
  612. { type: 'block-start', index: 0, blockType: 'tool-call' },
  613. { type: 'tool-call-delta', index: 0, id: callId, name: 'echo', argumentsDelta: '{"text":"x"}' },
  614. { type: 'block-end', index: 0, block: { type: 'tool-call', id: callId, name: 'echo', arguments: '{"text":"x"}' } },
  615. { type: 'usage', usage: { inputTokens: 10, outputTokens: 5 } },
  616. { type: 'finish', reason: { kind: 'max-tokens' } },
  617. ]])
  618. const ctx = await harness(adapter)
  619. let executions = 0
  620. ctx.tools.register(defineTool({
  621. name: 'echo',
  622. description: '',
  623. parameters: { text: { type: 'string' } },
  624. async execute() {
  625. executions += 1
  626. return [{ type: 'text', text: 'should not run' }]
  627. },
  628. }))
  629. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  630. const reasons: TurnEndReason[] = []
  631. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  632. send(agent, 'go')
  633. await waitForIdle(ctx, agent)
  634. expect(executions).toBe(0)
  635. expect(agent.session.events.some(e => e.type === 'tool/call')).toBe(false)
  636. expect(agent.session.deriveMessages()).toEqual([{ role: 'user', content: [{ type: 'text', text: 'go' }] }])
  637. expect(reasons).toEqual([{ kind: 'max-tokens' }])
  638. // Empty content still needs an assistant/message to carry usage; derivation
  639. // skips that host so it does not create a spurious assistant turn.
  640. const assistantMessage = agent.session.events.find(e => e.type === 'assistant/message')
  641. expect(assistantMessage?.type === 'assistant/message' && assistantMessage.data).toEqual({
  642. turn: 1, step: 1, content: [], provenance: { provider: 'mock', model: 'mock' }, usage: { inputTokens: 10, outputTokens: 5 },
  643. })
  644. })
  645. it('appends an empty completion anchor for a max-tokens step with no usage', async () => {
  646. // The truncated tool call is dropped from durable content, while the
  647. // successful provider call still needs an exact replay anchor.
  648. const callId = CallId('c1')
  649. const adapter = new MockAdapter([[
  650. { type: 'block-start', index: 0, blockType: 'tool-call' },
  651. { type: 'tool-call-delta', index: 0, id: callId, name: 'echo', argumentsDelta: '{"text":"x"}' },
  652. { type: 'block-end', index: 0, block: { type: 'tool-call', id: callId, name: 'echo', arguments: '{"text":"x"}' } },
  653. { type: 'finish', reason: { kind: 'max-tokens' } },
  654. ]])
  655. const ctx = await harness(adapter)
  656. ctx.tools.register(defineTool({
  657. name: 'echo',
  658. description: '',
  659. parameters: { text: { type: 'string' } },
  660. async execute() { return [{ type: 'text', text: 'should not run' }] },
  661. }))
  662. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  663. const reasons: TurnEndReason[] = []
  664. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  665. send(agent, 'go')
  666. await waitForIdle(ctx, agent)
  667. expect(reasons).toEqual([{ kind: 'max-tokens' }])
  668. const assistant = agent.session.events.find(e => e.type === 'assistant/message')!
  669. expect(assistant.type === 'assistant/message' && assistant.data).toEqual({
  670. turn: 1,
  671. step: 1,
  672. content: [],
  673. provenance: { provider: 'mock', model: 'mock' },
  674. })
  675. expect(assistant.sourceEventSeqs?.length).toBeGreaterThan(0)
  676. expect(agent.session.deriveMessages()).toEqual([{ role: 'user', content: [{ type: 'text', text: 'go' }] }])
  677. })
  678. it('appends an empty completion anchor for a normal stop with no usage', async () => {
  679. // A clean content-less call stays absent from derived messages but remains
  680. // a durable successful-call boundary for replay consumers.
  681. const adapter = new MockAdapter([[{ type: 'finish', reason: { kind: 'stop' } }]])
  682. const ctx = await harness(adapter)
  683. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  684. const reasons: TurnEndReason[] = []
  685. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  686. send(agent, 'go')
  687. await waitForIdle(ctx, agent)
  688. expect(reasons).toEqual([{ kind: 'completed' }])
  689. const assistant = agent.session.events.find(e => e.type === 'assistant/message')!
  690. expect(assistant.type === 'assistant/message' && assistant.data).toEqual({
  691. turn: 1,
  692. step: 1,
  693. content: [],
  694. provenance: { provider: 'mock', model: 'mock' },
  695. })
  696. expect(assistant.sourceEventSeqs?.length).toBe(1)
  697. expect(agent.session.deriveMessages()).toEqual([{ role: 'user', content: [{ type: 'text', text: 'go' }] }])
  698. })
  699. it('keeps safe max-tokens assistant content while dropping truncated tool calls', async () => {
  700. const callId = CallId('c1')
  701. const adapter = new MockAdapter([[
  702. { type: 'block-start', index: 0, blockType: 'text' },
  703. { type: 'text-delta', index: 0, text: 'partial text' },
  704. { type: 'block-end', index: 0, block: { type: 'text', text: 'partial text' } },
  705. { type: 'block-start', index: 1, blockType: 'tool-call' },
  706. { type: 'tool-call-delta', index: 1, id: callId, name: 'echo', argumentsDelta: '{"text"' },
  707. { type: 'finish', reason: { kind: 'max-tokens' } },
  708. ]])
  709. const ctx = await harness(adapter)
  710. let stepResults = 0
  711. ctx.on('agent/step-result', async (_agent, _turn, _step, message, next) => {
  712. stepResults += 1
  713. expect(message.content).toEqual([{ type: 'text', text: 'partial text' }])
  714. return next()
  715. })
  716. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  717. send(agent, 'go')
  718. await waitForIdle(ctx, agent)
  719. expect(stepResults).toBe(1)
  720. expect(agent.session.events.some(e => e.type === 'tool/call')).toBe(false)
  721. expect(agent.session.deriveMessages()).toEqual([
  722. { role: 'user', content: [{ type: 'text', text: 'go' }] },
  723. { role: 'assistant', content: [{ type: 'text', text: 'partial text' }], provenance: { provider: 'mock', model: 'mock' } },
  724. ])
  725. })
  726. it('contains a step/end observer failure without changing continuation', async () => {
  727. const adapter = new MockAdapter([
  728. toolCallResponse('c1', 'echo', { text: 'x' }),
  729. textResponse('continued after tool call'),
  730. ])
  731. const ctx = await harness(adapter)
  732. ctx.tools.register(defineTool({
  733. name: 'echo',
  734. description: '',
  735. parameters: { text: { type: 'string' } },
  736. async execute(args) {
  737. return [{ type: 'text', text: String(args.text) }]
  738. },
  739. }))
  740. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  741. let threw = false
  742. // Post-commit session observers cannot control the loop. The tool call still
  743. // drives the second model request, and the turn completes normally.
  744. ctx.on('session/event', (_session, event) => {
  745. if (event.type === 'step/end' && !threw) { threw = true; throw new Error('bad step/end listener') }
  746. })
  747. send(agent, 'go')
  748. await waitForIdle(ctx, agent)
  749. expect(adapter.requests).toHaveLength(2)
  750. const turnEnd = agent.session.events.findLast(e => e.type === 'turn/end')
  751. expect(turnEnd?.type === 'turn/end' && turnEnd.data.reason.kind).toBe('completed')
  752. })
  753. it('chains queued messages into consecutive turns', async () => {
  754. const adapter = new MockAdapter([textResponse('first'), textResponse('second')])
  755. const ctx = await harness(adapter)
  756. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  757. const turns: number[] = []
  758. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/start') turns.push(event.data.turn) })
  759. // queue two messages while idle — first starts turn 1 immediately;
  760. // queue the second during turn 1 when the first assistant chunk streams
  761. let queued = false
  762. ctx.on('session/event', (_s, event) => {
  763. if (event.type === 'assistant/chunk' && !queued) {
  764. queued = true
  765. send(agent, 'second message')
  766. }
  767. })
  768. send(agent, 'first message')
  769. await waitForIdle(ctx, agent)
  770. expect(turns).toEqual([1, 2])
  771. expect(adapter.requests).toHaveLength(2)
  772. })
  773. it('awaits session/flush at turn end (persistence checkpoint)', async () => {
  774. const adapter = new MockAdapter([textResponse('ok')])
  775. const ctx = await harness(adapter)
  776. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  777. let flushed = 0
  778. let flushedBeforeIdle = false
  779. ctx.on('session/flush', async (session) => {
  780. await new Promise(r => setTimeout(r, 10))
  781. flushed++
  782. flushedBeforeIdle = agent.status !== 'idle'
  783. void session
  784. })
  785. send(agent, 'hi')
  786. await waitForIdle(ctx, agent)
  787. expect(flushed).toBe(1)
  788. expect(flushedBeforeIdle).toBe(true)
  789. })
  790. it('errors from the model surface as agent/error and end the turn', async () => {
  791. const adapter = new MockAdapter([]) // script exhausted → throws
  792. const ctx = await harness(adapter)
  793. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  794. const errors: Error[] = []
  795. const reasons: TurnEndReason[] = []
  796. ctx.on('agent/error', (_agent, _turn, _step, error) => void errors.push(error))
  797. ctx.on('session/event', (_s, event) => { if (event.type === 'turn/end') reasons.push(event.data.reason) })
  798. send(agent, 'hi')
  799. await waitForIdle(ctx, agent)
  800. expect(errors).toHaveLength(1)
  801. expect(errors[0]!.message).toContain('script exhausted')
  802. expect(reasons[0]).toMatchObject({ kind: 'error' })
  803. // The durable failure lives entirely on turn/end.reason (with the failing
  804. // step), not a standalone error event.
  805. const turnEnd = agent.session.events.find(e => e.type === 'turn/end')
  806. expect(turnEnd?.type === 'turn/end' && turnEnd.data.reason).toMatchObject({ kind: 'error', step: 1 })
  807. })
  808. it('disposing the loop fiber mid-turn stops the loop (HMR safety)', async () => {
  809. const adapter = new MockAdapter(['hang'])
  810. const ctx = await harness(adapter)
  811. let agent!: ReactLoopAgent
  812. const fiber = await ctx.plugin(Object.assign((inner: Context) => {
  813. agent = inner.agentLoop.create(AgentId('scoped'), { provider: 'mock', model: 'mock' })
  814. }, { inject: ['agentLoop'] }))
  815. expect(ctx.agents.get(AgentId('scoped'))).toBe(agent)
  816. send(agent, 'go')
  817. await new Promise(r => setTimeout(r, 30))
  818. expect(agent.status).toBe('running')
  819. await fiber.dispose()
  820. await agent.done
  821. expect(agent.status).toBe('disposed')
  822. expect(ctx.agents.get(AgentId('scoped'))).toBeUndefined()
  823. expect(() => { send(agent, 'too late') }).toThrow('disposed')
  824. })
  825. it('creates agents from config on startup', async () => {
  826. const adapter = new MockAdapter([textResponse('from config')])
  827. const ctx = new Context()
  828. await ctx.plugin(LlmService)
  829. await ctx.plugin(SessionStore)
  830. await ctx.plugin(SystemPrompt)
  831. await ctx.plugin(ToolRegistry)
  832. await ctx.plugin(AgentRegistry)
  833. await ctx.plugin(AgentLoop, {
  834. agents: [{ id: AgentId('config-agent'), provider: 'mock', model: 'mock' }],
  835. })
  836. ctx.llm.registerAdapter(['mock'], adapter)
  837. const agent = ctx.agents.get(AgentId('config-agent'))! as ReactLoopAgent
  838. expect(agent).toBeDefined()
  839. expect(agent.id).toBe('config-agent')
  840. expect(agent.options.model).toBe('mock')
  841. // the agent is alive: send triggers a turn
  842. send(agent, 'hi')
  843. await waitForIdle(ctx, agent)
  844. expect(adapter.requests).toHaveLength(1)
  845. })
  846. it('attaches config agent cwd to the fresh session header', async () => {
  847. const ctx = new Context()
  848. await ctx.plugin(LlmService)
  849. await ctx.plugin(SessionStore)
  850. await ctx.plugin(SystemPrompt)
  851. await ctx.plugin(ToolRegistry)
  852. await ctx.plugin(AgentRegistry)
  853. await ctx.plugin(AgentLoop, {
  854. agents: [{ id: AgentId('config-agent'), provider: 'mock', model: 'mock', cwd: '/work/project' }],
  855. })
  856. const agent = ctx.agents.get(AgentId('config-agent'))! as ReactLoopAgent
  857. expect(agent.session.header.cwd).toBe('/work/project')
  858. })
  859. it('replays a session log into an identical derived history', async () => {
  860. const adapter = new MockAdapter([
  861. toolCallResponse('c1', 'echo', { text: 'x' }),
  862. textResponse('done'),
  863. ])
  864. const ctx = await harness(adapter)
  865. ctx.tools.register(defineTool({
  866. name: 'echo',
  867. description: '',
  868. parameters: { text: { type: 'string' } },
  869. async execute(args) {
  870. return [{ type: 'text', text: String(args.text) }]
  871. },
  872. }))
  873. const agent = ctx.agentLoop.create(AgentId('a1'), { provider: 'mock', model: 'mock' })
  874. send(agent, 'run')
  875. await waitForIdle(ctx, agent)
  876. const replayed = ctx.sessions.create(SessionId('replayed'), { seed: [...agent.session.events] })
  877. expect(replayed.deriveMessages()).toEqual(agent.session.deriveMessages())
  878. // event-by-event identity of types
  879. expect(replayed.events.map(e => e.type)).toEqual(
  880. agent.session.events.map(e => e.type))
  881. })
  882. })