session-runner.test.ts 130 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157215821592160216121622163216421652166216721682169217021712172217321742175217621772178217921802181218221832184218521862187218821892190219121922193219421952196219721982199220022012202220322042205220622072208220922102211221222132214221522162217221822192220222122222223222422252226222722282229223022312232223322342235223622372238223922402241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292229322942295229622972298229923002301230223032304230523062307230823092310231123122313231423152316231723182319232023212322232323242325232623272328232923302331233223332334233523362337233823392340234123422343234423452346234723482349235023512352235323542355235623572358235923602361236223632364236523662367236823692370237123722373237423752376237723782379238023812382238323842385238623872388238923902391239223932394239523962397239823992400240124022403240424052406240724082409241024112412241324142415241624172418241924202421242224232424242524262427242824292430243124322433243424352436243724382439244024412442244324442445244624472448244924502451245224532454245524562457245824592460246124622463246424652466246724682469247024712472247324742475247624772478247924802481248224832484248524862487248824892490249124922493249424952496249724982499250025012502250325042505250625072508250925102511251225132514251525162517251825192520252125222523252425252526252725282529253025312532253325342535253625372538253925402541254225432544254525462547254825492550255125522553255425552556255725582559256025612562256325642565256625672568256925702571257225732574257525762577257825792580258125822583258425852586258725882589259025912592259325942595259625972598259926002601260226032604260526062607260826092610261126122613261426152616261726182619262026212622262326242625262626272628262926302631263226332634263526362637263826392640264126422643264426452646264726482649265026512652265326542655265626572658265926602661266226632664266526662667266826692670267126722673267426752676267726782679268026812682268326842685268626872688268926902691269226932694269526962697269826992700270127022703270427052706270727082709271027112712271327142715271627172718271927202721272227232724272527262727272827292730273127322733273427352736273727382739274027412742274327442745274627472748274927502751275227532754275527562757275827592760276127622763276427652766276727682769277027712772277327742775277627772778277927802781278227832784278527862787278827892790279127922793279427952796279727982799280028012802280328042805280628072808280928102811281228132814281528162817281828192820282128222823282428252826282728282829283028312832283328342835283628372838283928402841284228432844284528462847284828492850285128522853285428552856285728582859286028612862286328642865286628672868286928702871287228732874287528762877287828792880288128822883288428852886288728882889289028912892289328942895289628972898289929002901290229032904290529062907290829092910291129122913291429152916291729182919292029212922292329242925292629272928292929302931293229332934293529362937293829392940294129422943294429452946294729482949295029512952295329542955295629572958295929602961296229632964296529662967296829692970297129722973297429752976297729782979298029812982298329842985298629872988298929902991299229932994299529962997299829993000300130023003300430053006300730083009301030113012301330143015301630173018301930203021302230233024302530263027302830293030303130323033303430353036303730383039304030413042304330443045304630473048304930503051305230533054305530563057305830593060306130623063306430653066306730683069307030713072307330743075307630773078307930803081308230833084308530863087308830893090309130923093309430953096309730983099310031013102310331043105310631073108310931103111311231133114311531163117311831193120312131223123312431253126312731283129313031313132313331343135313631373138313931403141314231433144314531463147314831493150315131523153315431553156315731583159316031613162316331643165316631673168316931703171317231733174317531763177317831793180318131823183318431853186318731883189319031913192319331943195319631973198319932003201320232033204320532063207320832093210321132123213321432153216321732183219322032213222322332243225322632273228322932303231323232333234323532363237323832393240324132423243324432453246324732483249325032513252325332543255325632573258325932603261326232633264326532663267326832693270327132723273327432753276327732783279328032813282328332843285328632873288328932903291329232933294329532963297329832993300330133023303330433053306330733083309331033113312331333143315331633173318331933203321332233233324332533263327332833293330333133323333333433353336333733383339334033413342334333443345334633473348334933503351335233533354335533563357335833593360336133623363336433653366336733683369337033713372337333743375337633773378337933803381338233833384338533863387338833893390339133923393339433953396339733983399340034013402340334043405340634073408340934103411341234133414341534163417341834193420342134223423342434253426342734283429343034313432343334343435343634373438343934403441344234433444344534463447344834493450345134523453345434553456345734583459346034613462346334643465346634673468346934703471347234733474347534763477347834793480348134823483348434853486348734883489349034913492349334943495349634973498349935003501350235033504350535063507350835093510351135123513351435153516351735183519352035213522352335243525352635273528352935303531353235333534353535363537353835393540354135423543354435453546354735483549355035513552355335543555
  1. import { describe, expect } from "bun:test"
  2. import {
  3. LLMClient,
  4. LLMError,
  5. LLMEvent,
  6. Model,
  7. TransportReason,
  8. InvalidRequestReason,
  9. type LLMClientShape,
  10. type LLMRequest,
  11. } from "@opencode-ai/llm"
  12. import * as OpenAIChat from "@opencode-ai/llm/protocols/openai-chat"
  13. import { Database } from "@opencode-ai/core/database/database"
  14. import { EventV2 } from "@opencode-ai/core/event"
  15. import { PermissionV2 } from "@opencode-ai/core/permission"
  16. import { EventTable } from "@opencode-ai/core/event/sql"
  17. import { Project } from "@opencode-ai/core/project"
  18. import { ProjectTable } from "@opencode-ai/core/project/sql"
  19. import { QuestionV2 } from "@opencode-ai/core/question"
  20. import { AbsolutePath } from "@opencode-ai/core/schema"
  21. import { SessionV2 } from "@opencode-ai/core/session"
  22. import { ContextSnapshotDecodeError } from "@opencode-ai/core/session/error"
  23. import { SessionEvent } from "@opencode-ai/core/session/event"
  24. import { SessionInput } from "@opencode-ai/core/session/input"
  25. import { SessionMessage } from "@opencode-ai/core/session/message"
  26. import { Prompt } from "@opencode-ai/core/session/prompt"
  27. import { SessionProjector } from "@opencode-ai/core/session/projector"
  28. import { SessionExecution } from "@opencode-ai/core/session/execution"
  29. import { SessionContextEpoch } from "@opencode-ai/core/session/context-epoch"
  30. import { SessionRunCoordinator } from "@opencode-ai/core/session/run-coordinator"
  31. import { SessionRunner } from "@opencode-ai/core/session/runner"
  32. import * as SessionRunnerLLM from "@opencode-ai/core/session/runner/llm"
  33. import { SessionRunnerModel } from "@opencode-ai/core/session/runner/model"
  34. import { ToolRegistry } from "@opencode-ai/core/tool/registry"
  35. import { ToolOutputStore } from "@opencode-ai/core/tool-output-store"
  36. import { ApplicationTools } from "@opencode-ai/core/tool/application-tools"
  37. import { AgentV2 } from "@opencode-ai/core/agent"
  38. import { Config } from "@opencode-ai/core/config"
  39. import { ConfigCompaction } from "@opencode-ai/core/config/compaction"
  40. import { Tool } from "@opencode-ai/core/tool/tool"
  41. import {
  42. SessionContextEpochTable,
  43. SessionInputTable,
  44. SessionMessageTable,
  45. SessionTable,
  46. } from "@opencode-ai/core/session/sql"
  47. import { SessionStore } from "@opencode-ai/core/session/store"
  48. import { SystemContext } from "@opencode-ai/core/system-context"
  49. import { SystemContextRegistry } from "@opencode-ai/core/system-context/registry"
  50. import { SkillGuidance } from "@opencode-ai/core/skill/guidance"
  51. import { ModelV2 } from "@opencode-ai/core/model"
  52. import { Location } from "@opencode-ai/core/location"
  53. import { ProviderV2 } from "@opencode-ai/core/provider"
  54. import { Cause, DateTime, Deferred, Effect, Exit, Fiber, Layer, Schema, Stream } from "effect"
  55. import { asc, eq } from "drizzle-orm"
  56. import { testEffect } from "./lib/effect"
  57. const database = Database.layerFromPath(":memory:")
  58. const events = EventV2.layer.pipe(Layer.provide(database))
  59. const questions = QuestionV2.layer.pipe(Layer.provide(events))
  60. const projector = SessionProjector.layer.pipe(Layer.provide(events), Layer.provide(database))
  61. const store = SessionStore.layer.pipe(Layer.provide(database))
  62. const requests: LLMRequest[] = []
  63. let response: LLMEvent[] = []
  64. let responses: LLMEvent[][] | undefined
  65. let responseStream: Stream.Stream<LLMEvent, LLMError> | undefined
  66. let streamGate: Deferred.Deferred<void> | undefined
  67. let streamStarted: Deferred.Deferred<void> | undefined
  68. let streamFailure: LLMError | undefined
  69. let toolExecutionGate: Deferred.Deferred<void> | undefined
  70. let toolExecutionsStarted: Deferred.Deferred<void> | undefined
  71. let toolExecutionsReady = 5
  72. let activeToolExecutions = 0
  73. let maxActiveToolExecutions = 0
  74. const client = Layer.succeed(
  75. LLMClient.Service,
  76. LLMClient.Service.of({
  77. prepare: () => Effect.die("unused"),
  78. stream: ((request: LLMRequest) => {
  79. requests.push(request)
  80. if (responseStream) {
  81. const stream = responseStream
  82. responseStream = undefined
  83. return stream
  84. }
  85. const events = streamFailure
  86. ? Stream.fail(streamFailure)
  87. : Stream.fromIterable(responses === undefined ? response : (responses.shift() ?? []))
  88. if (!streamGate) return events
  89. return Stream.unwrap(
  90. (streamStarted ? Deferred.succeed(streamStarted, undefined) : Effect.void).pipe(
  91. Effect.andThen(Deferred.await(streamGate)),
  92. Effect.as(events),
  93. ),
  94. )
  95. }) as unknown as LLMClientShape["stream"],
  96. generate: () => Effect.die("unused"),
  97. }),
  98. )
  99. const model = Model.make({ id: "fake-model", provider: "fake", route: OpenAIChat.route })
  100. const replacementModel = Model.make({ id: "replacement", provider: "fake", route: OpenAIChat.route })
  101. const compactModel = Model.make({
  102. id: "compact",
  103. provider: "fake",
  104. route: OpenAIChat.route.with({ limits: { context: 4_000, output: 50 } }),
  105. })
  106. const recoveryModel = Model.make({
  107. id: "recovery",
  108. provider: "fake",
  109. route: OpenAIChat.route.with({ limits: { context: 20_000, output: 1_000 } }),
  110. })
  111. const authorizations: Tool.Context[] = []
  112. const executions: string[] = []
  113. const permission = Layer.succeed(
  114. PermissionV2.Service,
  115. PermissionV2.Service.of({
  116. assert: () => Effect.die("unused"),
  117. ask: () => Effect.die("unused"),
  118. reply: () => Effect.die("unused"),
  119. get: () => Effect.die("unused"),
  120. forSession: () => Effect.die("unused"),
  121. list: () => Effect.die("unused"),
  122. }),
  123. )
  124. const applications = ApplicationTools.layer
  125. const registry = ToolRegistry.layer.pipe(
  126. Layer.provide(permission),
  127. Layer.provide(applications),
  128. Layer.provide(ToolOutputStore.defaultLayer),
  129. )
  130. const agents = AgentV2.layer
  131. const echo = Layer.effectDiscard(
  132. ToolRegistry.Service.use((registry) =>
  133. registry.register({
  134. echo: Tool.make({
  135. description: "Echo text",
  136. input: Schema.Struct({ text: Schema.String }),
  137. output: Schema.Struct({ text: Schema.String }),
  138. toModelOutput: ({ output }) => [{ type: "text", text: output.text }],
  139. execute: ({ text }, context) =>
  140. Effect.gen(function* () {
  141. authorizations.push(context)
  142. executions.push(text)
  143. activeToolExecutions++
  144. maxActiveToolExecutions = Math.max(maxActiveToolExecutions, activeToolExecutions)
  145. if (activeToolExecutions === toolExecutionsReady && toolExecutionsStarted) {
  146. yield* Deferred.succeed(toolExecutionsStarted, undefined)
  147. }
  148. if (toolExecutionGate) yield* Deferred.await(toolExecutionGate)
  149. return { text }
  150. }).pipe(Effect.ensuring(Effect.sync(() => activeToolExecutions--))),
  151. }),
  152. defect: Tool.make({
  153. description: "Fail unexpectedly",
  154. input: Schema.Struct({}),
  155. output: Schema.Struct({}),
  156. execute: () => Effect.die("unexpected tool defect"),
  157. }),
  158. }),
  159. ),
  160. ).pipe(Layer.provide(registry))
  161. let modelResolveHook = Effect.void
  162. let currentModel = model
  163. const models = SessionRunnerModel.layerWith((session) =>
  164. modelResolveHook.pipe(Effect.as(session.model?.id === "replacement" ? replacementModel : currentModel)),
  165. )
  166. const systemContextKey = SystemContext.Key.make("test/context")
  167. let systemBaseline = "Initial context"
  168. let systemRemoved = false
  169. let systemUnavailable = false
  170. let systemLoadHook = Effect.void
  171. const skillBaselines = new Map<AgentV2.ID, string>()
  172. const systemContext = Layer.effectDiscard(
  173. SystemContextRegistry.Service.pipe(
  174. Effect.flatMap((registry) =>
  175. registry.contribute({
  176. key: systemContextKey,
  177. load: Effect.sync(() =>
  178. SystemContext.combine(
  179. systemRemoved
  180. ? []
  181. : [
  182. SystemContext.make({
  183. key: systemContextKey,
  184. codec: Schema.toCodecJson(Schema.String),
  185. load: systemLoadHook.pipe(
  186. Effect.andThen(
  187. Effect.sync(() => (systemUnavailable ? SystemContext.unavailable : systemBaseline)),
  188. ),
  189. ),
  190. baseline: String,
  191. update: (_previous, current) => current,
  192. removed: () => "System context source removed: test/context",
  193. }),
  194. ],
  195. ),
  196. ),
  197. }),
  198. ),
  199. ),
  200. ).pipe(Layer.provideMerge(SystemContextRegistry.layer))
  201. const location = Location.layer({ directory: AbsolutePath.make("/project") }).pipe(Layer.provide(Project.defaultLayer))
  202. const skillGuidance = Layer.mock(SkillGuidance.Service, {
  203. load: (agent) =>
  204. Effect.succeed(
  205. skillBaselines.has(agent.id)
  206. ? SystemContext.make({
  207. key: SystemContext.Key.make("test/skill-guidance"),
  208. codec: Schema.toCodecJson(Schema.String),
  209. load: Effect.succeed(skillBaselines.get(agent.id)!),
  210. baseline: String,
  211. update: (_previous, current) => current,
  212. removed: () => "Skill guidance removed",
  213. })
  214. : SystemContext.empty,
  215. ),
  216. })
  217. const config = Layer.succeed(
  218. Config.Service,
  219. Config.Service.of({
  220. entries: () =>
  221. Effect.succeed([
  222. new Config.Document({
  223. type: "document",
  224. info: new Config.Info({
  225. compaction: new ConfigCompaction.Info({
  226. buffer: 3_000,
  227. keep: new ConfigCompaction.Keep({ tokens: 1_000 }),
  228. }),
  229. }),
  230. }),
  231. ]),
  232. }),
  233. )
  234. const runner = SessionRunnerLLM.layer.pipe(
  235. Layer.provide(database),
  236. Layer.provide(store),
  237. Layer.provide(events),
  238. Layer.provide(client),
  239. Layer.provide(registry),
  240. Layer.provide(models),
  241. Layer.provide(systemContext),
  242. Layer.provide(location),
  243. Layer.provide(agents),
  244. Layer.provide(skillGuidance),
  245. Layer.provide(config),
  246. )
  247. const coordinator = SessionRunCoordinator.layer.pipe(Layer.provide(runner))
  248. const execution = Layer.effect(
  249. SessionExecution.Service,
  250. SessionRunCoordinator.Service.pipe(
  251. Effect.map((coordinator) =>
  252. SessionExecution.Service.of({
  253. resume: coordinator.run,
  254. wake: coordinator.wake,
  255. interrupt: coordinator.interrupt,
  256. }),
  257. ),
  258. ),
  259. ).pipe(Layer.provide(coordinator))
  260. const sessions = SessionV2.layer.pipe(
  261. Layer.provide(events),
  262. Layer.provide(database),
  263. Layer.provide(store),
  264. Layer.provide(Project.defaultLayer),
  265. Layer.provide(execution),
  266. )
  267. const it = testEffect(
  268. Layer.mergeAll(
  269. database,
  270. events,
  271. questions,
  272. projector,
  273. store,
  274. client,
  275. permission,
  276. applications,
  277. agents,
  278. registry,
  279. echo,
  280. models,
  281. systemContext,
  282. location,
  283. skillGuidance,
  284. config,
  285. runner,
  286. coordinator,
  287. execution,
  288. sessions,
  289. ),
  290. )
  291. const sessionID = SessionV2.ID.make("ses_runner_test")
  292. const otherSessionID = SessionV2.ID.make("ses_runner_other")
  293. const insertSession = (id: SessionV2.ID) =>
  294. Effect.gen(function* () {
  295. const { db } = yield* Database.Service
  296. yield* db
  297. .insert(SessionTable)
  298. .values({
  299. id,
  300. project_id: Project.ID.global,
  301. slug: id,
  302. directory: "/project",
  303. title: "test",
  304. version: "test",
  305. })
  306. .onConflictDoNothing()
  307. .run()
  308. .pipe(Effect.orDie)
  309. })
  310. const setup = Effect.gen(function* () {
  311. const { db } = yield* Database.Service
  312. response = []
  313. systemBaseline = "Initial context"
  314. systemRemoved = false
  315. systemUnavailable = false
  316. systemLoadHook = Effect.void
  317. modelResolveHook = Effect.void
  318. currentModel = model
  319. skillBaselines.clear()
  320. responses = undefined
  321. streamFailure = undefined
  322. responseStream = undefined
  323. streamGate = undefined
  324. streamStarted = undefined
  325. toolExecutionGate = undefined
  326. toolExecutionsStarted = undefined
  327. toolExecutionsReady = 5
  328. activeToolExecutions = 0
  329. maxActiveToolExecutions = 0
  330. yield* db
  331. .insert(ProjectTable)
  332. .values({ id: Project.ID.global, worktree: AbsolutePath.make("/project"), sandboxes: [] })
  333. .onConflictDoNothing()
  334. .run()
  335. .pipe(Effect.orDie)
  336. yield* insertSession(sessionID)
  337. })
  338. const providerUnavailable = () =>
  339. new LLMError({
  340. module: "test",
  341. method: "stream",
  342. reason: new TransportReason({ message: "Provider unavailable" }),
  343. })
  344. const setupOverflowRecovery = Effect.gen(function* () {
  345. yield* setup
  346. const session = yield* SessionV2.Service
  347. response = fragmentFixture("text", "text-earlier", ["Earlier answer"]).completeEvents
  348. yield* session.prompt({
  349. sessionID,
  350. prompt: new Prompt({ text: "Earlier question ".repeat(700) }),
  351. resume: false,
  352. })
  353. yield* session.resume(sessionID)
  354. currentModel = recoveryModel
  355. requests.length = 0
  356. return session
  357. })
  358. const userTexts = (request: LLMRequest) =>
  359. request.messages.flatMap((message) =>
  360. message.role === "user"
  361. ? message.content.flatMap((content) => (content.type === "text" ? [content.text] : []))
  362. : [],
  363. )
  364. const replaySessionProjection = (id: SessionV2.ID) =>
  365. Effect.gen(function* () {
  366. const { db } = yield* Database.Service
  367. const events = yield* EventV2.Service
  368. const recorded = yield* db
  369. .select()
  370. .from(EventTable)
  371. .where(eq(EventTable.aggregate_id, id))
  372. .orderBy(asc(EventTable.seq))
  373. .all()
  374. .pipe(Effect.orDie)
  375. yield* events.remove(id)
  376. yield* db.delete(SessionInputTable).where(eq(SessionInputTable.session_id, id)).run().pipe(Effect.orDie)
  377. yield* db.delete(SessionMessageTable).where(eq(SessionMessageTable.session_id, id)).run().pipe(Effect.orDie)
  378. yield* events.replayAll(
  379. recorded.map((event) => ({
  380. id: event.id,
  381. aggregateID: event.aggregate_id,
  382. seq: event.seq,
  383. type: event.type,
  384. data: event.data,
  385. })),
  386. )
  387. })
  388. type FragmentKind = "text" | "reasoning" | "tool input"
  389. type FragmentFixture = {
  390. readonly delta: EventV2.Definition
  391. readonly completeEvents: LLMEvent[]
  392. readonly partialEvents: LLMEvent[]
  393. readonly expectedAssistant: unknown
  394. readonly expectedContent: unknown
  395. }
  396. const fragmentKinds: readonly FragmentKind[] = ["text", "reasoning", "tool input"]
  397. const fragmentID = (kind: FragmentKind, suffix: string) => `${kind === "tool input" ? "call" : kind}-${suffix}`
  398. const fragmentFixture = (kind: FragmentKind, id: string, chunks: readonly string[]): FragmentFixture => {
  399. const text = chunks.join("")
  400. switch (kind) {
  401. case "text": {
  402. const partialEvents = [
  403. LLMEvent.stepStart({ index: 0 }),
  404. LLMEvent.textStart({ id }),
  405. ...chunks.map((text) => LLMEvent.textDelta({ id, text })),
  406. ]
  407. const expectedContent = { type: "text", id, text }
  408. return {
  409. delta: SessionEvent.Text.Delta,
  410. partialEvents,
  411. completeEvents: [
  412. ...partialEvents,
  413. LLMEvent.textEnd({ id }),
  414. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  415. LLMEvent.finish({ reason: "stop" }),
  416. ],
  417. expectedAssistant: { type: "assistant", finish: "stop", content: [expectedContent] },
  418. expectedContent,
  419. }
  420. }
  421. case "reasoning": {
  422. const partialEvents = [
  423. LLMEvent.stepStart({ index: 0 }),
  424. LLMEvent.reasoningStart({ id }),
  425. ...chunks.map((text) => LLMEvent.reasoningDelta({ id, text })),
  426. ]
  427. const expectedContent = { type: "reasoning", id, text }
  428. return {
  429. delta: SessionEvent.Reasoning.Delta,
  430. partialEvents,
  431. completeEvents: [
  432. ...partialEvents,
  433. LLMEvent.reasoningEnd({ id }),
  434. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  435. LLMEvent.finish({ reason: "stop" }),
  436. ],
  437. expectedAssistant: { type: "assistant", finish: "stop", content: [expectedContent] },
  438. expectedContent,
  439. }
  440. }
  441. case "tool input": {
  442. const partialEvents = [
  443. LLMEvent.stepStart({ index: 0 }),
  444. LLMEvent.toolInputStart({ id, name: "echo" }),
  445. ...chunks.map((text) => LLMEvent.toolInputDelta({ id, name: "echo", text })),
  446. ]
  447. const expectedContent = { type: "tool", id, state: { status: "pending", input: text } }
  448. return {
  449. delta: SessionEvent.Tool.Input.Delta,
  450. partialEvents,
  451. completeEvents: [...partialEvents, LLMEvent.toolInputEnd({ id, name: "echo" })],
  452. expectedAssistant: { type: "assistant", content: [expectedContent] },
  453. expectedContent,
  454. }
  455. }
  456. }
  457. }
  458. const verifyEphemeralDeltas = (kind: FragmentKind) =>
  459. Effect.gen(function* () {
  460. yield* setup
  461. const session = yield* SessionV2.Service
  462. const prompt = `Stream ${kind}`
  463. const chunks = Array.from({ length: 32 }, (_, index) => `${index},`)
  464. const fixture = fragmentFixture(kind, fragmentID(kind, "many"), chunks)
  465. const expectedContext = [{ type: "user", text: prompt }, fixture.expectedAssistant]
  466. yield* session.prompt({ sessionID, prompt: new Prompt({ text: prompt }), resume: false })
  467. const events = yield* EventV2.Service
  468. const live = yield* events.subscribe(fixture.delta).pipe(Stream.take(32), Stream.runCollect, Effect.forkScoped)
  469. yield* Effect.yieldNow
  470. response = fixture.completeEvents
  471. yield* session.resume(sessionID)
  472. const { db } = yield* Database.Service
  473. const deltas = yield* db
  474. .select({ type: EventTable.type })
  475. .from(EventTable)
  476. .where(eq(EventTable.type, EventV2.versionedType(fixture.delta.type, 1)))
  477. .all()
  478. .pipe(Effect.orDie)
  479. expect(Array.from(yield* Fiber.join(live))).toHaveLength(32)
  480. expect(deltas).toHaveLength(0)
  481. expect(yield* session.context(sessionID)).toMatchObject(expectedContext)
  482. yield* replaySessionProjection(sessionID)
  483. expect(yield* session.context(sessionID)).toMatchObject(expectedContext)
  484. })
  485. const verifyPartialFlushOnFailure = (kind: FragmentKind) =>
  486. Effect.gen(function* () {
  487. yield* setup
  488. const session = yield* SessionV2.Service
  489. const prompt = `Fail after ${kind}`
  490. const fixture = fragmentFixture(kind, fragmentID(kind, "partial"), ["Partial"])
  491. const failure = providerUnavailable()
  492. yield* session.prompt({ sessionID, prompt: new Prompt({ text: prompt }), resume: false })
  493. responseStream = Stream.concat(Stream.fromIterable(fixture.partialEvents), Stream.fail(failure))
  494. expect(yield* session.resume(sessionID).pipe(Effect.flip)).toBe(failure)
  495. expect(yield* session.context(sessionID)).toMatchObject([
  496. { type: "user", text: prompt },
  497. {
  498. type: "assistant",
  499. finish: "error",
  500. error: { type: "unknown", message: "Provider unavailable" },
  501. content: [fixture.expectedContent],
  502. },
  503. ])
  504. })
  505. const verifyPartialFlushOnInterruption = (kind: FragmentKind) =>
  506. Effect.gen(function* () {
  507. yield* setup
  508. const session = yield* SessionV2.Service
  509. const prompt = `Interrupt after ${kind}`
  510. const fixture = fragmentFixture(kind, fragmentID(kind, "interrupted"), ["Partial"])
  511. const streamed = yield* Deferred.make<void>()
  512. yield* session.prompt({ sessionID, prompt: new Prompt({ text: prompt }), resume: false })
  513. responseStream = Stream.concat(
  514. Stream.fromIterable(fixture.partialEvents),
  515. Stream.fromEffect(Deferred.succeed(streamed, undefined)).pipe(Stream.flatMap(() => Stream.never)),
  516. )
  517. const runner = yield* SessionRunner.Service
  518. const fiber = yield* runner.run({ sessionID, force: true }).pipe(Effect.forkChild)
  519. yield* Deferred.await(streamed)
  520. yield* Fiber.interrupt(fiber)
  521. expect(yield* session.context(sessionID)).toMatchObject([
  522. { type: "user", text: prompt },
  523. {
  524. type: "assistant",
  525. content: [
  526. kind === "tool input"
  527. ? { type: "tool", id: fragmentID(kind, "interrupted"), state: { status: "error" } }
  528. : fixture.expectedContent,
  529. ],
  530. },
  531. ])
  532. })
  533. describe("SessionRunnerLLM", () => {
  534. it.effect("advertises and executes a globally attached application tool", () =>
  535. Effect.gen(function* () {
  536. yield* setup
  537. const applicationTools = yield* ApplicationTools.Service
  538. const session = yield* SessionV2.Service
  539. const contexts: Tool.Context[] = []
  540. yield* applicationTools.register({
  541. application_context: Tool.make({
  542. description: "Read application context",
  543. input: Schema.Struct({ query: Schema.String }),
  544. output: Schema.Struct({ answer: Schema.String }),
  545. execute: ({ query }, context) =>
  546. Effect.sync(() => {
  547. contexts.push(context)
  548. return { answer: query.toUpperCase() }
  549. }),
  550. }),
  551. })
  552. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Use application context" }), resume: false })
  553. responses = [
  554. [
  555. LLMEvent.stepStart({ index: 0 }),
  556. LLMEvent.toolCall({ id: "call-application", name: "application_context", input: { query: "hello" } }),
  557. LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
  558. LLMEvent.finish({ reason: "tool-calls" }),
  559. ],
  560. [],
  561. ]
  562. yield* session.resume(sessionID)
  563. expect(requests[0]?.tools.map((tool) => tool.name)).toContain("application_context")
  564. expect(contexts).toEqual([
  565. {
  566. sessionID,
  567. agent: AgentV2.ID.make("build"),
  568. assistantMessageID: expect.stringMatching(/^msg_/),
  569. toolCallID: "call-application",
  570. },
  571. ])
  572. expect(yield* session.context(sessionID)).toMatchObject([
  573. { type: "user", text: "Use application context" },
  574. {
  575. type: "assistant",
  576. content: [
  577. {
  578. type: "tool",
  579. id: "call-application",
  580. state: { status: "completed", structured: { answer: "HELLO" } },
  581. },
  582. ],
  583. },
  584. ])
  585. }),
  586. )
  587. it.effect("starts a real runner turn after default prompt recording", () =>
  588. Effect.gen(function* () {
  589. yield* setup
  590. const session = yield* SessionV2.Service
  591. requests.length = 0
  592. responses = undefined
  593. streamGate = undefined
  594. streamStarted = undefined
  595. response = []
  596. const message = yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Run automatically" }) })
  597. expect(requests).toHaveLength(1)
  598. expect(yield* session.messages({ sessionID })).toMatchObject([
  599. { id: message.id, type: "user", text: "Run automatically" },
  600. ])
  601. }),
  602. )
  603. it.effect("streams one request with registry definitions from chronological V2 user history", () =>
  604. Effect.gen(function* () {
  605. yield* setup
  606. const session = yield* SessionV2.Service
  607. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  608. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Second" }), resume: false })
  609. requests.length = 0
  610. responses = undefined
  611. streamGate = undefined
  612. streamStarted = undefined
  613. response = []
  614. yield* session.resume(sessionID)
  615. expect(requests).toHaveLength(1)
  616. expect(requests[0]?.model).toBe(model)
  617. expect(requests[0]?.tools.map((tool) => tool.name)).toEqual(["echo", "defect"])
  618. expect(requests[0]?.messages.map((message) => ({ role: message.role, content: message.content }))).toEqual([
  619. { role: "user", content: [{ type: "text", text: "First" }] },
  620. { role: "user", content: [{ type: "text", text: "Second" }] },
  621. ])
  622. expect(yield* session.messages({ sessionID })).toHaveLength(2)
  623. }),
  624. )
  625. it.effect("retries the first provider turn after system context becomes available", () =>
  626. Effect.gen(function* () {
  627. yield* setup
  628. const session = yield* SessionV2.Service
  629. const { db } = yield* Database.Service
  630. const messageID = SessionMessage.ID.create()
  631. systemUnavailable = true
  632. yield* session.prompt({ id: messageID, sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  633. requests.length = 0
  634. const exit = yield* session.resume(sessionID).pipe(Effect.exit)
  635. expect(Exit.isFailure(exit)).toBe(true)
  636. if (Exit.isFailure(exit)) expect(Cause.squash(exit.cause)).toBeInstanceOf(SystemContext.InitializationBlocked)
  637. expect(requests).toHaveLength(0)
  638. expect(yield* SessionInput.hasPending(db, sessionID, "steer")).toBe(true)
  639. expect(
  640. yield* db
  641. .select()
  642. .from(SessionContextEpochTable)
  643. .where(eq(SessionContextEpochTable.session_id, sessionID))
  644. .get(),
  645. ).toBeUndefined()
  646. systemUnavailable = false
  647. yield* session.prompt({ id: messageID, sessionID, prompt: new Prompt({ text: "First" }) })
  648. yield* (yield* SessionRunCoordinator.Service).awaitIdle(sessionID)
  649. expect(requests).toHaveLength(1)
  650. expect(requests[0]?.messages.map((message) => message.role)).toEqual(["user"])
  651. }),
  652. )
  653. it.effect("interrupts a source Location runner after a Session moves", () =>
  654. Effect.gen(function* () {
  655. yield* setup
  656. const session = yield* SessionV2.Service
  657. const events = yield* EventV2.Service
  658. const { db } = yield* Database.Service
  659. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  660. requests.length = 0
  661. response = []
  662. yield* session.resume(sessionID)
  663. yield* events.publish(SessionEvent.Moved, {
  664. sessionID,
  665. timestamp: DateTime.makeUnsafe(1),
  666. location: { directory: AbsolutePath.make("/moved") },
  667. })
  668. expect(
  669. yield* db
  670. .select()
  671. .from(SessionContextEpochTable)
  672. .where(eq(SessionContextEpochTable.session_id, sessionID))
  673. .get(),
  674. ).toBeUndefined()
  675. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Second" }), resume: false })
  676. const exit = yield* session.resume(sessionID).pipe(Effect.exit)
  677. expect(Exit.isFailure(exit) && Cause.hasInterruptsOnly(exit.cause)).toBe(true)
  678. expect(requests).toHaveLength(1)
  679. expect(yield* SessionInput.hasPending(db, sessionID, "steer")).toBe(true)
  680. }),
  681. )
  682. it.effect("fails gracefully when a stored context snapshot cannot be decoded", () =>
  683. Effect.gen(function* () {
  684. yield* setup
  685. const session = yield* SessionV2.Service
  686. const { db } = yield* Database.Service
  687. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  688. response = []
  689. yield* session.resume(sessionID)
  690. yield* db
  691. .update(SessionContextEpochTable)
  692. .set({ snapshot: { invalid: { value: "bad" } } })
  693. .where(eq(SessionContextEpochTable.session_id, sessionID))
  694. .run()
  695. .pipe(Effect.orDie)
  696. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Second" }), resume: false })
  697. requests.length = 0
  698. const exit = yield* session.resume(sessionID).pipe(Effect.exit)
  699. expect(Exit.isFailure(exit)).toBe(true)
  700. if (Exit.isFailure(exit)) expect(Cause.squash(exit.cause)).toBeInstanceOf(ContextSnapshotDecodeError)
  701. expect(requests).toHaveLength(0)
  702. }),
  703. )
  704. it.effect("does not create a source Location epoch after a concurrent Session move", () =>
  705. Effect.gen(function* () {
  706. yield* setup
  707. const session = yield* SessionV2.Service
  708. const events = yield* EventV2.Service
  709. const { db } = yield* Database.Service
  710. let moved = false
  711. systemLoadHook = Effect.suspend(() => {
  712. if (moved) return Effect.void
  713. moved = true
  714. return events
  715. .publish(SessionEvent.Moved, {
  716. sessionID,
  717. timestamp: DateTime.makeUnsafe(1),
  718. location: { directory: AbsolutePath.make("/moved") },
  719. })
  720. .pipe(Effect.asVoid)
  721. })
  722. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  723. expect(Exit.isFailure(yield* session.resume(sessionID).pipe(Effect.exit))).toBe(true)
  724. expect(yield* SessionInput.hasPending(db, sessionID, "steer")).toBe(true)
  725. expect(
  726. yield* db
  727. .select()
  728. .from(SessionContextEpochTable)
  729. .where(eq(SessionContextEpochTable.session_id, sessionID))
  730. .get(),
  731. ).toBeUndefined()
  732. expect((yield* session.get(sessionID)).location.directory).toBe(AbsolutePath.make("/moved"))
  733. }),
  734. )
  735. it.effect("reuses one durable baseline after the context producer changes", () =>
  736. Effect.gen(function* () {
  737. yield* setup
  738. const session = yield* SessionV2.Service
  739. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  740. requests.length = 0
  741. response = []
  742. yield* session.resume(sessionID)
  743. systemBaseline = "Changed context"
  744. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Second" }), resume: false })
  745. yield* session.resume(sessionID)
  746. expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([
  747. ["Initial context"],
  748. ["Initial context"],
  749. ])
  750. expect(requests[1]?.messages.map((message) => message.role)).toEqual(["user", "user", "system"])
  751. expect(requests[1]?.messages.at(-1)?.content).toEqual([{ type: "text", text: "Changed context" }])
  752. expect(yield* session.messages({ sessionID })).toHaveLength(3)
  753. const { db } = yield* Database.Service
  754. expect(
  755. yield* db
  756. .select({ id: EventTable.id })
  757. .from(EventTable)
  758. .where(eq(EventTable.type, "session.next.context.updated.1"))
  759. .all()
  760. .pipe(Effect.orDie),
  761. ).toHaveLength(1)
  762. yield* replaySessionProjection(sessionID)
  763. expect(yield* session.messages({ sessionID })).toHaveLength(3)
  764. }),
  765. )
  766. it.effect("includes the effective default agent system before durable context", () =>
  767. Effect.gen(function* () {
  768. yield* setup
  769. const agent = yield* AgentV2.Service
  770. yield* agent.update((editor) =>
  771. editor.update(AgentV2.ID.make("build"), (agent) => {
  772. agent.system = "Build agent instructions"
  773. agent.mode = "primary"
  774. }),
  775. )
  776. const session = yield* SessionV2.Service
  777. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  778. requests.length = 0
  779. response = fragmentFixture("text", "text-build", ["Done"]).completeEvents
  780. yield* session.resume(sessionID)
  781. expect(requests.at(-1)?.system.map((part) => part.text)).toEqual(["Build agent instructions", "Initial context"])
  782. }),
  783. )
  784. it.effect("uses the configured default agent system for omitted-agent sessions", () =>
  785. Effect.gen(function* () {
  786. yield* setup
  787. const agent = yield* AgentV2.Service
  788. yield* agent.update((editor) => {
  789. editor.update(AgentV2.ID.make("build"), (agent) => {
  790. agent.system = "Build agent instructions"
  791. agent.mode = "primary"
  792. })
  793. editor.update(AgentV2.ID.make("reviewer"), (agent) => {
  794. agent.system = "Reviewer instructions"
  795. agent.mode = "primary"
  796. })
  797. editor.default(AgentV2.ID.make("reviewer"))
  798. })
  799. const session = yield* SessionV2.Service
  800. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  801. requests.length = 0
  802. response = fragmentFixture("text", "text-reviewer", ["Done"]).completeEvents
  803. yield* session.resume(sessionID)
  804. expect(requests.at(-1)?.system.map((part) => part.text)).toEqual(["Reviewer instructions", "Initial context"])
  805. expect((yield* session.messages({ sessionID }))[0]).toMatchObject({ type: "assistant", agent: "reviewer" })
  806. }),
  807. )
  808. it.effect("uses an explicitly selected non-build agent system", () =>
  809. Effect.gen(function* () {
  810. yield* setup
  811. const { db } = yield* Database.Service
  812. const agent = yield* AgentV2.Service
  813. yield* agent.update((editor) =>
  814. editor.update(AgentV2.ID.make("reviewer"), (agent) => {
  815. agent.system = "Reviewer instructions"
  816. agent.mode = "primary"
  817. }),
  818. )
  819. yield* db
  820. .update(SessionTable)
  821. .set({ agent: "reviewer" })
  822. .where(eq(SessionTable.id, sessionID))
  823. .run()
  824. .pipe(Effect.orDie)
  825. const session = yield* SessionV2.Service
  826. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  827. requests.length = 0
  828. response = fragmentFixture("text", "text-selected", ["Done"]).completeEvents
  829. yield* session.resume(sessionID)
  830. expect(requests.at(-1)?.system.map((part) => part.text)).toEqual(["Reviewer instructions", "Initial context"])
  831. expect((yield* session.messages({ sessionID }))[0]).toMatchObject({ type: "assistant", agent: "reviewer" })
  832. }),
  833. )
  834. it.effect("composes selected-agent skill guidance and replaces it after an agent switch", () =>
  835. Effect.gen(function* () {
  836. yield* setup
  837. const session = yield* SessionV2.Service
  838. const events = yield* EventV2.Service
  839. skillBaselines.set(AgentV2.ID.make("build"), "Build skills")
  840. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  841. requests.length = 0
  842. response = []
  843. yield* session.resume(sessionID)
  844. skillBaselines.set(AgentV2.ID.make("reviewer"), "Reviewer skills")
  845. yield* events.publish(SessionEvent.AgentSwitched, {
  846. sessionID,
  847. messageID: SessionMessage.ID.create(),
  848. timestamp: DateTime.makeUnsafe(1),
  849. agent: "reviewer",
  850. })
  851. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Second" }), resume: false })
  852. yield* session.resume(sessionID)
  853. expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([
  854. ["Initial context\n\nBuild skills"],
  855. ["Initial context\n\nReviewer skills"],
  856. ])
  857. }),
  858. )
  859. it.effect("retries first-epoch preparation when the selected agent changes during observation", () =>
  860. Effect.gen(function* () {
  861. yield* setup
  862. const session = yield* SessionV2.Service
  863. const events = yield* EventV2.Service
  864. skillBaselines.set(AgentV2.ID.make("build"), "Build skills")
  865. skillBaselines.set(AgentV2.ID.make("reviewer"), "Reviewer skills")
  866. let switched = false
  867. systemLoadHook = Effect.suspend(() => {
  868. if (switched) return Effect.void
  869. switched = true
  870. return events
  871. .publish(SessionEvent.AgentSwitched, {
  872. sessionID,
  873. messageID: SessionMessage.ID.create(),
  874. timestamp: DateTime.makeUnsafe(1),
  875. agent: "reviewer",
  876. })
  877. .pipe(Effect.asVoid)
  878. })
  879. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  880. requests.length = 0
  881. response = []
  882. yield* session.resume(sessionID)
  883. expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([
  884. ["Initial context\n\nReviewer skills"],
  885. ])
  886. }),
  887. )
  888. it.effect("opens a queued activity once when the selected agent changes during observation", () =>
  889. Effect.gen(function* () {
  890. yield* setup
  891. const session = yield* SessionV2.Service
  892. const events = yield* EventV2.Service
  893. skillBaselines.set(AgentV2.ID.make("build"), "Build skills")
  894. skillBaselines.set(AgentV2.ID.make("reviewer"), "Reviewer skills")
  895. let switched = false
  896. systemLoadHook = Effect.suspend(() => {
  897. if (switched) return Effect.void
  898. switched = true
  899. return events
  900. .publish(SessionEvent.AgentSwitched, {
  901. sessionID,
  902. messageID: SessionMessage.ID.create(),
  903. timestamp: DateTime.makeUnsafe(1),
  904. agent: "reviewer",
  905. })
  906. .pipe(Effect.asVoid)
  907. })
  908. yield* session.prompt({
  909. sessionID,
  910. prompt: new Prompt({ text: "Queued" }),
  911. delivery: "queue",
  912. resume: false,
  913. })
  914. requests.length = 0
  915. response = []
  916. yield* session.resume(sessionID)
  917. expect(requests).toHaveLength(1)
  918. expect((yield* session.context(sessionID)).filter((message) => message.type === "user")).toHaveLength(1)
  919. }),
  920. )
  921. it.effect("retries an agent switch before the final provider-dispatch boundary", () =>
  922. Effect.gen(function* () {
  923. yield* setup
  924. const session = yield* SessionV2.Service
  925. const events = yield* EventV2.Service
  926. const { db } = yield* Database.Service
  927. skillBaselines.set(AgentV2.ID.make("build"), "Build skills")
  928. skillBaselines.set(AgentV2.ID.make("reviewer"), "Reviewer skills")
  929. let switched = false
  930. modelResolveHook = Effect.suspend(() => {
  931. if (switched) return Effect.void
  932. switched = true
  933. return events
  934. .publish(SessionEvent.AgentSwitched, {
  935. sessionID,
  936. messageID: SessionMessage.ID.create(),
  937. timestamp: DateTime.makeUnsafe(1),
  938. agent: "reviewer",
  939. })
  940. .pipe(Effect.asVoid)
  941. })
  942. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  943. requests.length = 0
  944. response = []
  945. yield* session.resume(sessionID)
  946. expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([
  947. ["Initial context\n\nReviewer skills"],
  948. ])
  949. expect(
  950. yield* db
  951. .select({ replacementSeq: SessionContextEpochTable.replacement_seq })
  952. .from(SessionContextEpochTable)
  953. .where(eq(SessionContextEpochTable.session_id, sessionID))
  954. .get()
  955. .pipe(Effect.orDie),
  956. ).toEqual({ replacementSeq: null })
  957. }),
  958. )
  959. it.effect("retries a model switch before the final provider-dispatch boundary", () =>
  960. Effect.gen(function* () {
  961. yield* setup
  962. const session = yield* SessionV2.Service
  963. const events = yield* EventV2.Service
  964. let switched = false
  965. modelResolveHook = Effect.suspend(() => {
  966. if (switched) return Effect.void
  967. switched = true
  968. return events
  969. .publish(SessionEvent.ModelSwitched, {
  970. sessionID,
  971. messageID: SessionMessage.ID.create(),
  972. timestamp: DateTime.makeUnsafe(1),
  973. model: { id: ModelV2.ID.make("replacement"), providerID: ProviderV2.ID.make("fake") },
  974. })
  975. .pipe(Effect.asVoid)
  976. })
  977. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  978. requests.length = 0
  979. response = []
  980. yield* session.resume(sessionID)
  981. expect(requests.map((request) => request.model)).toEqual([replacementModel])
  982. expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([["Initial context"]])
  983. }),
  984. )
  985. it.effect("fences an unchanged epoch read across an agent ABA replacement request", () =>
  986. Effect.gen(function* () {
  987. yield* setup
  988. const session = yield* SessionV2.Service
  989. const events = yield* EventV2.Service
  990. const { db } = yield* Database.Service
  991. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  992. response = []
  993. yield* session.resume(sessionID)
  994. let switched = false
  995. systemLoadHook = Effect.suspend(() => {
  996. if (switched) return Effect.void
  997. switched = true
  998. return events
  999. .publish(SessionEvent.AgentSwitched, {
  1000. sessionID,
  1001. messageID: SessionMessage.ID.create(),
  1002. timestamp: DateTime.makeUnsafe(1),
  1003. agent: AgentV2.ID.make("reviewer"),
  1004. })
  1005. .pipe(
  1006. Effect.andThen(
  1007. events.publish(SessionEvent.AgentSwitched, {
  1008. sessionID,
  1009. messageID: SessionMessage.ID.create(),
  1010. timestamp: DateTime.makeUnsafe(2),
  1011. agent: AgentV2.defaultID,
  1012. }),
  1013. ),
  1014. Effect.asVoid,
  1015. )
  1016. })
  1017. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Second" }), resume: false })
  1018. requests.length = 0
  1019. yield* session.resume(sessionID)
  1020. expect(requests).toHaveLength(1)
  1021. expect(
  1022. yield* db
  1023. .select({ replacementSeq: SessionContextEpochTable.replacement_seq })
  1024. .from(SessionContextEpochTable)
  1025. .where(eq(SessionContextEpochTable.session_id, sessionID))
  1026. .get()
  1027. .pipe(Effect.orDie),
  1028. ).toEqual({ replacementSeq: null })
  1029. }),
  1030. )
  1031. it.effect("rejects stale agent guidance when committing an existing-epoch replacement", () =>
  1032. Effect.gen(function* () {
  1033. yield* setup
  1034. const session = yield* SessionV2.Service
  1035. const events = yield* EventV2.Service
  1036. const { db } = yield* Database.Service
  1037. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  1038. response = []
  1039. yield* session.resume(sessionID)
  1040. yield* events.publish(SessionEvent.AgentSwitched, {
  1041. sessionID,
  1042. messageID: SessionMessage.ID.create(),
  1043. timestamp: DateTime.makeUnsafe(1),
  1044. agent: AgentV2.ID.make("reviewer"),
  1045. })
  1046. const context = (text: string) =>
  1047. Effect.succeed(
  1048. SystemContext.make({
  1049. key: systemContextKey,
  1050. codec: Schema.toCodecJson(Schema.String),
  1051. load: Effect.succeed(text),
  1052. baseline: String,
  1053. update: (_previous, current) => current,
  1054. }),
  1055. )
  1056. const location = (yield* session.get(sessionID)).location
  1057. expect(
  1058. yield* SessionContextEpoch.prepare(
  1059. db,
  1060. events,
  1061. context("Stale build context"),
  1062. sessionID,
  1063. location,
  1064. AgentV2.defaultID,
  1065. ).pipe(Effect.catchDefect(Effect.succeed)),
  1066. ).toBeInstanceOf(SessionContextEpoch.AgentMismatch)
  1067. expect(
  1068. yield* SessionContextEpoch.prepare(
  1069. db,
  1070. events,
  1071. context("Reviewer context"),
  1072. sessionID,
  1073. location,
  1074. AgentV2.ID.make("reviewer"),
  1075. ),
  1076. ).toMatchObject({ baseline: "Reviewer context" })
  1077. }),
  1078. )
  1079. it.effect("blocks a cross-agent provider turn while replacement context is unavailable", () =>
  1080. Effect.gen(function* () {
  1081. yield* setup
  1082. const session = yield* SessionV2.Service
  1083. const events = yield* EventV2.Service
  1084. skillBaselines.set(AgentV2.defaultID, "Build skills")
  1085. skillBaselines.set(AgentV2.ID.make("reviewer"), "Reviewer skills")
  1086. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  1087. response = []
  1088. yield* session.resume(sessionID)
  1089. yield* events.publish(SessionEvent.AgentSwitched, {
  1090. sessionID,
  1091. messageID: SessionMessage.ID.create(),
  1092. timestamp: DateTime.makeUnsafe(1),
  1093. agent: AgentV2.ID.make("reviewer"),
  1094. })
  1095. systemUnavailable = true
  1096. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Second" }), resume: false })
  1097. requests.length = 0
  1098. const blocked = yield* session.resume(sessionID).pipe(Effect.exit)
  1099. expect(Exit.isFailure(blocked)).toBe(true)
  1100. if (Exit.isFailure(blocked))
  1101. expect(Cause.squash(blocked.cause)).toBeInstanceOf(SessionContextEpoch.AgentReplacementBlocked)
  1102. expect(requests).toHaveLength(0)
  1103. systemUnavailable = false
  1104. yield* session.resume(sessionID)
  1105. expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([
  1106. ["Initial context\n\nReviewer skills"],
  1107. ])
  1108. }),
  1109. )
  1110. it.effect("admits removed context as a chronological System message", () =>
  1111. Effect.gen(function* () {
  1112. yield* setup
  1113. const session = yield* SessionV2.Service
  1114. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  1115. requests.length = 0
  1116. response = []
  1117. yield* session.resume(sessionID)
  1118. systemRemoved = true
  1119. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Second" }), resume: false })
  1120. yield* session.resume(sessionID)
  1121. expect(requests[1]?.messages.map((message) => message.role)).toEqual(["user", "user", "system"])
  1122. expect(requests[1]?.messages.at(-1)?.content).toEqual([
  1123. { type: "text", text: "System context source removed: test/context" },
  1124. ])
  1125. expect(yield* session.messages({ sessionID })).toHaveLength(3)
  1126. }),
  1127. )
  1128. it.effect("replaces the baseline lazily after a model switch and drops prior System updates", () =>
  1129. Effect.gen(function* () {
  1130. yield* setup
  1131. const session = yield* SessionV2.Service
  1132. const events = yield* EventV2.Service
  1133. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  1134. requests.length = 0
  1135. response = []
  1136. yield* session.resume(sessionID)
  1137. systemBaseline = "Changed context"
  1138. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Second" }), resume: false })
  1139. yield* session.resume(sessionID)
  1140. yield* events.publish(SessionEvent.ModelSwitched, {
  1141. sessionID,
  1142. messageID: SessionMessage.ID.create(),
  1143. timestamp: DateTime.makeUnsafe(1),
  1144. model: { id: ModelV2.ID.make("replacement"), providerID: ProviderV2.ID.make("fake") },
  1145. })
  1146. systemBaseline = "Replacement context"
  1147. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Third" }), resume: false })
  1148. yield* session.resume(sessionID)
  1149. expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([
  1150. ["Initial context"],
  1151. ["Initial context"],
  1152. ["Replacement context"],
  1153. ])
  1154. expect(requests[1]?.messages.map((message) => message.role)).toEqual(["user", "user", "system"])
  1155. expect(requests[2]?.messages.map((message) => message.role)).toEqual(["user", "user", "user"])
  1156. expect((yield* session.context(sessionID)).map((message) => message.type)).toEqual([
  1157. "user",
  1158. "user",
  1159. "model-switched",
  1160. "user",
  1161. ])
  1162. yield* replaySessionProjection(sessionID)
  1163. expect(yield* session.messages({ sessionID })).toHaveLength(5)
  1164. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Fourth" }), resume: false })
  1165. yield* session.resume(sessionID)
  1166. }),
  1167. )
  1168. it.effect("defers replacement while admitted context is temporarily unavailable", () =>
  1169. Effect.gen(function* () {
  1170. yield* setup
  1171. const session = yield* SessionV2.Service
  1172. const events = yield* EventV2.Service
  1173. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  1174. requests.length = 0
  1175. response = []
  1176. yield* session.resume(sessionID)
  1177. yield* events.publish(SessionEvent.ModelSwitched, {
  1178. sessionID,
  1179. messageID: SessionMessage.ID.create(),
  1180. timestamp: DateTime.makeUnsafe(1),
  1181. model: { id: ModelV2.ID.make("replacement"), providerID: ProviderV2.ID.make("fake") },
  1182. })
  1183. systemUnavailable = true
  1184. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Second" }), resume: false })
  1185. yield* session.resume(sessionID)
  1186. systemUnavailable = false
  1187. systemBaseline = "Replacement context"
  1188. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Third" }), resume: false })
  1189. yield* session.resume(sessionID)
  1190. expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([
  1191. ["Initial context"],
  1192. ["Initial context"],
  1193. ["Replacement context"],
  1194. ])
  1195. }),
  1196. )
  1197. it.effect("advances a pending replacement to the latest invalidation boundary", () =>
  1198. Effect.gen(function* () {
  1199. yield* setup
  1200. const session = yield* SessionV2.Service
  1201. const events = yield* EventV2.Service
  1202. const { db } = yield* Database.Service
  1203. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  1204. response = []
  1205. yield* session.resume(sessionID)
  1206. yield* events.publish(SessionEvent.ModelSwitched, {
  1207. sessionID,
  1208. messageID: SessionMessage.ID.create(),
  1209. timestamp: DateTime.makeUnsafe(1),
  1210. model: { id: ModelV2.ID.make("replacement-1"), providerID: ProviderV2.ID.make("fake") },
  1211. })
  1212. yield* events.publish(SessionEvent.ModelSwitched, {
  1213. sessionID,
  1214. messageID: SessionMessage.ID.create(),
  1215. timestamp: DateTime.makeUnsafe(2),
  1216. model: { id: ModelV2.ID.make("replacement-2"), providerID: ProviderV2.ID.make("fake") },
  1217. })
  1218. const latest = yield* SessionInput.latestSeq(db, sessionID)
  1219. expect(
  1220. yield* db
  1221. .select({ replacementSeq: SessionContextEpochTable.replacement_seq })
  1222. .from(SessionContextEpochTable)
  1223. .where(eq(SessionContextEpochTable.session_id, sessionID))
  1224. .get()
  1225. .pipe(Effect.orDie),
  1226. ).toEqual({ replacementSeq: latest })
  1227. }),
  1228. )
  1229. it.effect("retries epoch preparation until observation-time invalidations settle", () =>
  1230. Effect.gen(function* () {
  1231. yield* setup
  1232. const session = yield* SessionV2.Service
  1233. const events = yield* EventV2.Service
  1234. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  1235. response = []
  1236. yield* session.resume(sessionID)
  1237. requests.length = 0
  1238. systemBaseline = "Changed context"
  1239. let invalidations = 0
  1240. systemLoadHook = Effect.suspend(() => {
  1241. if (invalidations === 4) return Effect.void
  1242. invalidations++
  1243. return events
  1244. .publish(SessionEvent.ModelSwitched, {
  1245. sessionID,
  1246. messageID: SessionMessage.ID.create(),
  1247. timestamp: DateTime.makeUnsafe(invalidations),
  1248. model: { id: ModelV2.ID.make(`replacement-${invalidations}`), providerID: ProviderV2.ID.make("fake") },
  1249. })
  1250. .pipe(Effect.asVoid)
  1251. })
  1252. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Second" }), resume: false })
  1253. yield* session.resume(sessionID)
  1254. expect(invalidations).toBe(4)
  1255. expect(requests).toHaveLength(1)
  1256. expect(requests[0]?.system.map((part) => part.text)).toEqual(["Changed context"])
  1257. }),
  1258. )
  1259. it.effect("replays retained context projections while replacement is pending", () =>
  1260. Effect.gen(function* () {
  1261. yield* setup
  1262. const session = yield* SessionV2.Service
  1263. const events = yield* EventV2.Service
  1264. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  1265. requests.length = 0
  1266. response = []
  1267. yield* session.resume(sessionID)
  1268. systemBaseline = "Changed context"
  1269. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Second" }), resume: false })
  1270. yield* session.resume(sessionID)
  1271. yield* events.publish(SessionEvent.ModelSwitched, {
  1272. sessionID,
  1273. messageID: SessionMessage.ID.create(),
  1274. timestamp: DateTime.makeUnsafe(1),
  1275. model: { id: ModelV2.ID.make("replacement"), providerID: ProviderV2.ID.make("fake") },
  1276. })
  1277. yield* replaySessionProjection(sessionID)
  1278. systemBaseline = "Replacement context"
  1279. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Third" }), resume: false })
  1280. yield* session.resume(sessionID)
  1281. expect(requests.at(-1)?.system.map((part) => part.text)).toEqual(["Replacement context"])
  1282. }),
  1283. )
  1284. it.effect("replaces the baseline lazily after completed compaction without reopening replacement on replay", () =>
  1285. Effect.gen(function* () {
  1286. yield* setup
  1287. const session = yield* SessionV2.Service
  1288. const events = yield* EventV2.Service
  1289. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  1290. requests.length = 0
  1291. response = []
  1292. yield* session.resume(sessionID)
  1293. const compactionID = SessionMessage.ID.create()
  1294. yield* events.publish(SessionEvent.Compaction.Started, {
  1295. sessionID,
  1296. messageID: compactionID,
  1297. timestamp: DateTime.makeUnsafe(1),
  1298. reason: "manual",
  1299. })
  1300. yield* events.publish(SessionEvent.Compaction.Ended, {
  1301. sessionID,
  1302. messageID: compactionID,
  1303. timestamp: DateTime.makeUnsafe(2),
  1304. reason: "manual",
  1305. text: "summary",
  1306. recent: "",
  1307. })
  1308. systemBaseline = "Replacement context"
  1309. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Second" }), resume: false })
  1310. yield* session.resume(sessionID)
  1311. expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([
  1312. ["Initial context"],
  1313. ["Replacement context"],
  1314. ])
  1315. yield* replaySessionProjection(sessionID)
  1316. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Third" }), resume: false })
  1317. yield* session.resume(sessionID)
  1318. }),
  1319. )
  1320. it.effect("automatically compacts into a completed summary and retained recent turn", () =>
  1321. Effect.gen(function* () {
  1322. yield* setup
  1323. const session = yield* SessionV2.Service
  1324. response = fragmentFixture("text", "text-first", ["Earlier answer"]).completeEvents
  1325. yield* session.prompt({
  1326. sessionID,
  1327. prompt: new Prompt({ text: "Earlier question ".repeat(180) }),
  1328. resume: false,
  1329. })
  1330. yield* session.resume(sessionID)
  1331. currentModel = compactModel
  1332. requests.length = 0
  1333. responses = [
  1334. fragmentFixture("text", "text-summary", ["## Goal\n- Preserve the task"]).completeEvents,
  1335. fragmentFixture("text", "text-final", ["Continued"]).completeEvents,
  1336. ]
  1337. yield* session.prompt({
  1338. sessionID,
  1339. prompt: new Prompt({ text: "Recent exact request ".repeat(180) }),
  1340. resume: false,
  1341. })
  1342. yield* session.resume(sessionID)
  1343. expect(requests).toHaveLength(2)
  1344. expect(userTexts(requests[0])[0]).toContain("## Goal")
  1345. expect(userTexts(requests[1])).toHaveLength(1)
  1346. expect(userTexts(requests[1])[0]).toContain("<summary>\n## Goal\n- Preserve the task\n</summary>")
  1347. expect(userTexts(requests[1])[0]).toContain(`[User]: ${"Recent exact request ".repeat(180)}`)
  1348. const context = yield* (yield* SessionStore.Service).context(sessionID)
  1349. expect(context.map((message) => message.type)).toEqual(["compaction", "assistant"])
  1350. expect(context[0]).toMatchObject({
  1351. type: "compaction",
  1352. summary: "## Goal\n- Preserve the task",
  1353. })
  1354. requests.length = 0
  1355. responses = [
  1356. fragmentFixture("text", "text-summary-2", ["## Goal\n- Preserve the updated task"]).completeEvents,
  1357. fragmentFixture("text", "text-final-2", ["Continued again"]).completeEvents,
  1358. ]
  1359. yield* session.prompt({
  1360. sessionID,
  1361. prompt: new Prompt({ text: "Newest exact request ".repeat(180) }),
  1362. resume: false,
  1363. })
  1364. yield* session.resume(sessionID)
  1365. expect(requests).toHaveLength(2)
  1366. expect(userTexts(requests[0])[0]).toContain(
  1367. "<previous-summary>\n## Goal\n- Preserve the task\n</previous-summary>",
  1368. )
  1369. expect(userTexts(requests[0])[0]).toContain("Recent exact request")
  1370. expect((yield* (yield* SessionStore.Service).context(sessionID))[0]).toMatchObject({
  1371. type: "compaction",
  1372. summary: "## Goal\n- Preserve the updated task",
  1373. })
  1374. }),
  1375. )
  1376. it.effect("forces one compaction and retries after provider context overflow", () =>
  1377. Effect.gen(function* () {
  1378. const session = yield* setupOverflowRecovery
  1379. responses = [
  1380. [
  1381. LLMEvent.stepStart({ index: 0 }),
  1382. LLMEvent.providerError({ message: "prompt too long", classification: "context-overflow" }),
  1383. ],
  1384. fragmentFixture("text", "text-summary", ["## Goal\n- Recover overflow"]).completeEvents,
  1385. fragmentFixture("text", "text-final", ["Recovered"]).completeEvents,
  1386. ]
  1387. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Continue" }), resume: false })
  1388. yield* session.resume(sessionID)
  1389. expect(requests).toHaveLength(3)
  1390. expect(userTexts(requests[1])[0]).toContain("## Goal")
  1391. expect(userTexts(requests[2])[0]).toContain("<summary>\n## Goal\n- Recover overflow\n</summary>")
  1392. expect(yield* session.context(sessionID)).toMatchObject([
  1393. { type: "compaction", summary: "## Goal\n- Recover overflow" },
  1394. { type: "assistant", finish: "stop" },
  1395. ])
  1396. yield* replaySessionProjection(sessionID)
  1397. expect(yield* session.context(sessionID)).toMatchObject([
  1398. { type: "compaction" },
  1399. { type: "assistant", finish: "stop" },
  1400. ])
  1401. }),
  1402. )
  1403. it.effect("persists a second context overflow after one recovery", () =>
  1404. Effect.gen(function* () {
  1405. const session = yield* setupOverflowRecovery
  1406. const overflow = () => [
  1407. LLMEvent.stepStart({ index: 0 }),
  1408. LLMEvent.providerError({ message: "prompt too long", classification: "context-overflow" }),
  1409. ]
  1410. responses = [
  1411. overflow(),
  1412. fragmentFixture("text", "text-summary", ["## Goal\n- Recover once"]).completeEvents,
  1413. overflow(),
  1414. ]
  1415. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Continue" }), resume: false })
  1416. yield* session.resume(sessionID)
  1417. expect(requests).toHaveLength(3)
  1418. expect(yield* session.context(sessionID)).toMatchObject([
  1419. { type: "compaction" },
  1420. { type: "assistant", finish: "error", error: { message: "prompt too long" } },
  1421. ])
  1422. }),
  1423. )
  1424. it.effect("recovers once from a raw context overflow failure", () =>
  1425. Effect.gen(function* () {
  1426. const session = yield* setupOverflowRecovery
  1427. responseStream = Stream.fail(
  1428. new LLMError({
  1429. module: "test",
  1430. method: "stream",
  1431. reason: new InvalidRequestReason({
  1432. message: "prompt too long",
  1433. classification: "context-overflow",
  1434. }),
  1435. }),
  1436. )
  1437. responses = [
  1438. fragmentFixture("text", "text-summary", ["## Goal\n- Recover raw overflow"]).completeEvents,
  1439. fragmentFixture("text", "text-final", ["Recovered"]).completeEvents,
  1440. ]
  1441. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Continue" }), resume: false })
  1442. yield* session.resume(sessionID)
  1443. expect(requests).toHaveLength(3)
  1444. expect(yield* session.context(sessionID)).toMatchObject([
  1445. { type: "compaction", summary: "## Goal\n- Recover raw overflow" },
  1446. { type: "assistant", finish: "stop" },
  1447. ])
  1448. }),
  1449. )
  1450. it.effect("publishes the original overflow when recovery summarization fails", () =>
  1451. Effect.gen(function* () {
  1452. const session = yield* setupOverflowRecovery
  1453. responses = [
  1454. [LLMEvent.providerError({ message: "prompt too long", classification: "context-overflow" })],
  1455. [LLMEvent.providerError({ message: "summary unavailable" })],
  1456. ]
  1457. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Continue" }), resume: false })
  1458. yield* session.resume(sessionID)
  1459. expect(requests).toHaveLength(2)
  1460. const context = yield* session.context(sessionID)
  1461. expect(context.some((message) => message.type === "compaction")).toBe(false)
  1462. expect(context.slice(-2)).toMatchObject([
  1463. { type: "user", text: "Continue" },
  1464. { type: "assistant", finish: "error", error: { message: "prompt too long" } },
  1465. ])
  1466. }),
  1467. )
  1468. it.effect("interrupts overflow recovery while the summary provider is running", () =>
  1469. Effect.gen(function* () {
  1470. const session = yield* setupOverflowRecovery
  1471. responses = [
  1472. [LLMEvent.providerError({ message: "prompt too long", classification: "context-overflow" })],
  1473. fragmentFixture("text", "text-summary", ["## Goal\n- Interrupted"]).completeEvents,
  1474. ]
  1475. const firstGate = yield* Deferred.make<void>()
  1476. const summaryGate = yield* Deferred.make<void>()
  1477. streamGate = firstGate
  1478. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Continue" }), resume: false })
  1479. const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
  1480. while (requests.length < 1) yield* Effect.yieldNow
  1481. streamGate = summaryGate
  1482. yield* Deferred.succeed(firstGate, undefined)
  1483. while (requests.length < 2) yield* Effect.yieldNow
  1484. yield* session.interrupt(sessionID)
  1485. expect(yield* Fiber.await(run)).toMatchObject({ _tag: "Failure" })
  1486. streamGate = undefined
  1487. expect(requests).toHaveLength(2)
  1488. expect((yield* session.context(sessionID)).some((message) => message.type === "compaction")).toBe(false)
  1489. }),
  1490. )
  1491. it.effect("preserves effective System updates while compaction replacement is blocked", () =>
  1492. Effect.gen(function* () {
  1493. yield* setup
  1494. const session = yield* SessionV2.Service
  1495. const events = yield* EventV2.Service
  1496. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First" }), resume: false })
  1497. requests.length = 0
  1498. response = []
  1499. yield* session.resume(sessionID)
  1500. systemBaseline = "Changed context"
  1501. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Second" }), resume: false })
  1502. yield* session.resume(sessionID)
  1503. const compactionID = SessionMessage.ID.create()
  1504. yield* events.publish(SessionEvent.Compaction.Started, {
  1505. sessionID,
  1506. messageID: compactionID,
  1507. timestamp: DateTime.makeUnsafe(1),
  1508. reason: "manual",
  1509. })
  1510. yield* events.publish(SessionEvent.Compaction.Ended, {
  1511. sessionID,
  1512. messageID: compactionID,
  1513. timestamp: DateTime.makeUnsafe(2),
  1514. reason: "manual",
  1515. text: "summary",
  1516. recent: "",
  1517. })
  1518. systemUnavailable = true
  1519. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Third" }), resume: false })
  1520. yield* session.resume(sessionID)
  1521. expect(requests.at(-1)?.system.map((part) => part.text)).toEqual(["Initial context"])
  1522. expect(
  1523. requests
  1524. .at(-1)
  1525. ?.messages.some(
  1526. (message) =>
  1527. message.role === "system" &&
  1528. message.content[0]?.type === "text" &&
  1529. message.content[0].text === "Changed context",
  1530. ),
  1531. ).toBe(true)
  1532. }),
  1533. )
  1534. it.effect("projects reasoning and tool events without executing or continuing tools", () =>
  1535. Effect.gen(function* () {
  1536. yield* setup
  1537. const session = yield* SessionV2.Service
  1538. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Use tools" }), resume: false })
  1539. requests.length = 0
  1540. responses = undefined
  1541. streamGate = undefined
  1542. streamStarted = undefined
  1543. response = [
  1544. LLMEvent.stepStart({ index: 0 }),
  1545. LLMEvent.reasoningStart({ id: "reasoning-1" }),
  1546. LLMEvent.reasoningDelta({ id: "reasoning-1", text: "Think" }),
  1547. LLMEvent.reasoningEnd({ id: "reasoning-1" }),
  1548. LLMEvent.toolInputStart({ id: "call-error", name: "write" }),
  1549. LLMEvent.toolInputDelta({ id: "call-error", name: "write", text: '{"path":"README.md"}' }),
  1550. LLMEvent.toolInputEnd({ id: "call-error", name: "write" }),
  1551. LLMEvent.toolCall({ id: "call-error", name: "write", input: { path: "README.md" }, providerExecuted: true }),
  1552. LLMEvent.toolError({ id: "call-error", name: "write", message: "Denied" }),
  1553. LLMEvent.toolResult({ id: "call-error", name: "write", result: { type: "error", value: "Denied" } }),
  1554. LLMEvent.toolCall({
  1555. id: "call-provider",
  1556. name: "web_search",
  1557. input: { query: "hello" },
  1558. providerExecuted: true,
  1559. providerMetadata: { fake: { source: "provider" } },
  1560. }),
  1561. LLMEvent.toolResult({
  1562. id: "call-provider",
  1563. name: "web_search",
  1564. result: {
  1565. type: "content",
  1566. value: [
  1567. { type: "text", text: "Hello" },
  1568. { type: "media", mediaType: "image/png", data: "data:image/png;base64,aGVsbG8=", filename: "hello.png" },
  1569. ],
  1570. },
  1571. providerExecuted: true,
  1572. providerMetadata: { fake: { source: "provider" } },
  1573. }),
  1574. LLMEvent.stepFinish({
  1575. index: 0,
  1576. reason: "tool-calls",
  1577. usage: {
  1578. inputTokens: 10,
  1579. nonCachedInputTokens: 8,
  1580. outputTokens: 4,
  1581. reasoningTokens: 1,
  1582. cacheReadInputTokens: 2,
  1583. },
  1584. }),
  1585. LLMEvent.finish({ reason: "tool-calls" }),
  1586. ]
  1587. yield* session.resume(sessionID)
  1588. expect(requests).toHaveLength(1)
  1589. expect(requests[0]?.tools.map((tool) => tool.name)).toEqual(["echo", "defect"])
  1590. expect(yield* session.context(sessionID)).toMatchObject([
  1591. { type: "user", text: "Use tools" },
  1592. {
  1593. type: "assistant",
  1594. finish: "tool-calls",
  1595. tokens: { input: 8, output: 3, reasoning: 1, cache: { read: 2, write: 0 } },
  1596. content: [
  1597. { type: "reasoning", id: "reasoning-1", text: "Think" },
  1598. {
  1599. type: "tool",
  1600. id: "call-error",
  1601. name: "write",
  1602. state: {
  1603. status: "error",
  1604. input: { path: "README.md" },
  1605. error: { type: "unknown", message: "Denied" },
  1606. },
  1607. },
  1608. {
  1609. type: "tool",
  1610. id: "call-provider",
  1611. name: "web_search",
  1612. provider: { executed: true, metadata: { fake: { source: "provider" } } },
  1613. state: {
  1614. status: "completed",
  1615. input: { query: "hello" },
  1616. structured: {},
  1617. content: [
  1618. { type: "text", text: "Hello" },
  1619. { type: "file", mime: "image/png", source: { type: "data", data: "aGVsbG8=" }, name: "hello.png" },
  1620. ],
  1621. },
  1622. },
  1623. ],
  1624. },
  1625. ])
  1626. }),
  1627. )
  1628. it.effect("continues with reloaded history after durably settling one local tool call", () =>
  1629. Effect.gen(function* () {
  1630. yield* setup
  1631. const session = yield* SessionV2.Service
  1632. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Echo this" }), resume: false })
  1633. requests.length = 0
  1634. authorizations.length = 0
  1635. executions.length = 0
  1636. streamGate = undefined
  1637. streamStarted = undefined
  1638. responses = [
  1639. [
  1640. LLMEvent.stepStart({ index: 0 }),
  1641. LLMEvent.toolCall({ id: "call-echo", name: "echo", input: { text: "hello" } }),
  1642. LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
  1643. LLMEvent.finish({ reason: "tool-calls" }),
  1644. ],
  1645. [
  1646. LLMEvent.stepStart({ index: 0 }),
  1647. LLMEvent.textStart({ id: "text-final" }),
  1648. LLMEvent.textDelta({ id: "text-final", text: "Done" }),
  1649. LLMEvent.textEnd({ id: "text-final" }),
  1650. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  1651. LLMEvent.finish({ reason: "stop" }),
  1652. ],
  1653. ]
  1654. yield* session.resume(sessionID)
  1655. expect(requests).toHaveLength(2)
  1656. expect(requests[1]?.messages.map((message) => message.role)).toEqual(["user", "assistant", "tool"])
  1657. expect(authorizations).toMatchObject([{ sessionID, toolCallID: "call-echo" }])
  1658. expect(executions).toEqual(["hello"])
  1659. expect(yield* session.context(sessionID)).toMatchObject([
  1660. { type: "user", text: "Echo this" },
  1661. {
  1662. type: "assistant",
  1663. finish: "tool-calls",
  1664. content: [
  1665. {
  1666. type: "tool",
  1667. id: "call-echo",
  1668. name: "echo",
  1669. state: {
  1670. status: "completed",
  1671. input: { text: "hello" },
  1672. structured: { text: "hello" },
  1673. content: [{ type: "text", text: "hello" }],
  1674. },
  1675. },
  1676. ],
  1677. },
  1678. { type: "assistant", finish: "stop", content: [{ type: "text", id: "text-final", text: "Done" }] },
  1679. ])
  1680. }),
  1681. )
  1682. it.effect("reloads a model switch before a tool-driven continuation turn", () =>
  1683. Effect.gen(function* () {
  1684. yield* setup
  1685. const session = yield* SessionV2.Service
  1686. const events = yield* EventV2.Service
  1687. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Echo this" }), resume: false })
  1688. requests.length = 0
  1689. responses = [
  1690. [
  1691. LLMEvent.stepStart({ index: 0 }),
  1692. LLMEvent.toolCall({ id: "call-echo", name: "echo", input: { text: "hello" } }),
  1693. LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
  1694. LLMEvent.finish({ reason: "tool-calls" }),
  1695. ],
  1696. [
  1697. LLMEvent.stepStart({ index: 0 }),
  1698. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  1699. LLMEvent.finish({ reason: "stop" }),
  1700. ],
  1701. ]
  1702. toolExecutionGate = yield* Deferred.make<void>()
  1703. toolExecutionsStarted = yield* Deferred.make<void>()
  1704. toolExecutionsReady = 1
  1705. const run = yield* Effect.forkChild(session.resume(sessionID))
  1706. yield* Deferred.await(toolExecutionsStarted)
  1707. yield* events.publish(SessionEvent.ModelSwitched, {
  1708. sessionID,
  1709. messageID: SessionMessage.ID.create(),
  1710. timestamp: DateTime.makeUnsafe(1),
  1711. model: { id: ModelV2.ID.make("replacement"), providerID: ProviderV2.ID.make("fake") },
  1712. })
  1713. systemBaseline = "Replacement context"
  1714. yield* Deferred.succeed(toolExecutionGate, undefined)
  1715. yield* Fiber.join(run)
  1716. expect(requests.map((request) => request.model)).toEqual([model, replacementModel])
  1717. expect(requests.map((request) => request.system.map((part) => part.text))).toEqual([
  1718. ["Initial context"],
  1719. ["Replacement context"],
  1720. ])
  1721. }),
  1722. )
  1723. it.effect("restores durable reasoning provider metadata in a second-turn request", () =>
  1724. Effect.gen(function* () {
  1725. yield* setup
  1726. const session = yield* SessionV2.Service
  1727. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Think first" }), resume: false })
  1728. requests.length = 0
  1729. response = [
  1730. LLMEvent.stepStart({ index: 0 }),
  1731. LLMEvent.reasoningStart({ id: "reasoning-anthropic" }),
  1732. LLMEvent.reasoningDelta({ id: "reasoning-anthropic", text: "Signed thought" }),
  1733. LLMEvent.reasoningEnd({ id: "reasoning-anthropic", providerMetadata: { anthropic: { signature: "sig_1" } } }),
  1734. LLMEvent.reasoningStart({
  1735. id: "reasoning-openai",
  1736. providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: null } },
  1737. }),
  1738. LLMEvent.reasoningDelta({ id: "reasoning-openai", text: "Encrypted thought" }),
  1739. LLMEvent.reasoningEnd({
  1740. id: "reasoning-openai",
  1741. providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: "encrypted-state" } },
  1742. }),
  1743. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  1744. LLMEvent.finish({ reason: "stop" }),
  1745. ]
  1746. yield* session.resume(sessionID)
  1747. yield* replaySessionProjection(sessionID)
  1748. expect(yield* session.context(sessionID)).toMatchObject([
  1749. { type: "user", text: "Think first" },
  1750. {
  1751. type: "assistant",
  1752. content: [
  1753. { type: "reasoning", text: "Signed thought", providerMetadata: { anthropic: { signature: "sig_1" } } },
  1754. {
  1755. type: "reasoning",
  1756. text: "Encrypted thought",
  1757. providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: "encrypted-state" } },
  1758. },
  1759. ],
  1760. },
  1761. ])
  1762. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Continue" }), resume: false })
  1763. response = []
  1764. yield* session.resume(sessionID)
  1765. expect(requests[1]?.messages[1]?.content).toEqual([
  1766. { type: "reasoning", text: "Signed thought", providerMetadata: { anthropic: { signature: "sig_1" } } },
  1767. {
  1768. type: "reasoning",
  1769. text: "Encrypted thought",
  1770. providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: "encrypted-state" } },
  1771. },
  1772. ])
  1773. }),
  1774. )
  1775. it.effect("replays durable provider-executed tool results inline in a second-turn request", () =>
  1776. Effect.gen(function* () {
  1777. yield* setup
  1778. const session = yield* SessionV2.Service
  1779. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Search first" }), resume: false })
  1780. requests.length = 0
  1781. response = [
  1782. LLMEvent.stepStart({ index: 0 }),
  1783. LLMEvent.toolCall({
  1784. id: "hosted-search",
  1785. name: "web_search",
  1786. input: { query: "Effect" },
  1787. providerExecuted: true,
  1788. providerMetadata: { openai: { itemId: "hosted-search" } },
  1789. }),
  1790. LLMEvent.toolResult({
  1791. id: "hosted-search",
  1792. name: "web_search",
  1793. result: { type: "json", value: [{ title: "Effect" }] },
  1794. providerExecuted: true,
  1795. providerMetadata: { anthropic: { blockType: "web_search_tool_result" } },
  1796. }),
  1797. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  1798. LLMEvent.finish({ reason: "stop" }),
  1799. ]
  1800. yield* session.resume(sessionID)
  1801. yield* replaySessionProjection(sessionID)
  1802. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Continue" }), resume: false })
  1803. response = []
  1804. yield* session.resume(sessionID)
  1805. expect(requests[1]?.messages.map((message) => message.role)).toEqual(["user", "assistant", "user"])
  1806. expect(requests[1]?.messages[1]?.content).toMatchObject([
  1807. {
  1808. type: "tool-call",
  1809. id: "hosted-search",
  1810. name: "web_search",
  1811. input: { query: "Effect" },
  1812. providerExecuted: true,
  1813. providerMetadata: { openai: { itemId: "hosted-search" } },
  1814. },
  1815. {
  1816. type: "tool-result",
  1817. id: "hosted-search",
  1818. name: "web_search",
  1819. result: { type: "json", value: [{ title: "Effect" }] },
  1820. providerExecuted: true,
  1821. providerMetadata: { anthropic: { blockType: "web_search_tool_result" } },
  1822. },
  1823. ])
  1824. }),
  1825. )
  1826. it.effect("starts recorded local tools eagerly and awaits settlement before continuing", () =>
  1827. Effect.gen(function* () {
  1828. yield* setup
  1829. const session = yield* SessionV2.Service
  1830. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Echo five times" }), resume: false })
  1831. requests.length = 0
  1832. executions.length = 0
  1833. toolExecutionGate = yield* Deferred.make<void>()
  1834. toolExecutionsStarted = yield* Deferred.make<void>()
  1835. const providerGate = yield* Deferred.make<void>()
  1836. response = []
  1837. responses = undefined
  1838. const initial = Stream.fromIterable([
  1839. LLMEvent.stepStart({ index: 0 }),
  1840. ...Array.from({ length: 5 }, (_, index) =>
  1841. LLMEvent.toolCall({ id: `call-echo-${index}`, name: "echo", input: { text: `${index}` } }),
  1842. ),
  1843. ])
  1844. const final = Stream.fromIterable([
  1845. LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
  1846. LLMEvent.finish({ reason: "tool-calls" }),
  1847. ])
  1848. streamGate = undefined
  1849. responseStream = Stream.concat(
  1850. initial,
  1851. Stream.fromEffect(Deferred.await(providerGate)).pipe(Stream.flatMap(() => final)),
  1852. )
  1853. const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
  1854. yield* Deferred.await(toolExecutionsStarted)
  1855. expect(executions).toHaveLength(5)
  1856. expect(maxActiveToolExecutions).toBe(5)
  1857. expect(yield* session.context(sessionID)).toMatchObject([
  1858. { type: "user", text: "Echo five times" },
  1859. {
  1860. type: "assistant",
  1861. content: Array.from({ length: 5 }, (_, index) => ({
  1862. type: "tool",
  1863. id: `call-echo-${index}`,
  1864. state: { status: "running", input: { text: `${index}` } },
  1865. })),
  1866. },
  1867. ])
  1868. yield* Deferred.succeed(providerGate, undefined)
  1869. yield* Effect.yieldNow
  1870. expect(requests).toHaveLength(1)
  1871. yield* Deferred.succeed(toolExecutionGate, undefined)
  1872. yield* Fiber.join(run)
  1873. toolExecutionGate = undefined
  1874. toolExecutionsStarted = undefined
  1875. expect(executions).toHaveLength(5)
  1876. expect(maxActiveToolExecutions).toBe(5)
  1877. expect(requests).toHaveLength(2)
  1878. }),
  1879. )
  1880. it.effect("settles repeated provider-local tool call IDs against their owning assistant messages", () =>
  1881. Effect.gen(function* () {
  1882. yield* setup
  1883. const session = yield* SessionV2.Service
  1884. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Echo twice" }), resume: false })
  1885. requests.length = 0
  1886. executions.length = 0
  1887. responses = [
  1888. [
  1889. LLMEvent.stepStart({ index: 0 }),
  1890. LLMEvent.toolCall({ id: "tool_0", name: "echo", input: { text: "first" } }),
  1891. LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
  1892. LLMEvent.finish({ reason: "tool-calls" }),
  1893. ],
  1894. [
  1895. LLMEvent.stepStart({ index: 0 }),
  1896. LLMEvent.toolCall({ id: "tool_0", name: "echo", input: { text: "second" } }),
  1897. LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
  1898. LLMEvent.finish({ reason: "tool-calls" }),
  1899. ],
  1900. [],
  1901. ]
  1902. yield* session.resume(sessionID)
  1903. expect(executions).toEqual(["first", "second"])
  1904. expect(requests).toHaveLength(3)
  1905. expect(yield* session.context(sessionID)).toMatchObject([
  1906. { type: "user", text: "Echo twice" },
  1907. {
  1908. type: "assistant",
  1909. content: [
  1910. {
  1911. type: "tool",
  1912. id: "tool_0",
  1913. state: { status: "completed", structured: { text: "first" }, content: [{ type: "text", text: "first" }] },
  1914. },
  1915. ],
  1916. },
  1917. {
  1918. type: "assistant",
  1919. content: [
  1920. {
  1921. type: "tool",
  1922. id: "tool_0",
  1923. state: {
  1924. status: "completed",
  1925. structured: { text: "second" },
  1926. content: [{ type: "text", text: "second" }],
  1927. },
  1928. },
  1929. ],
  1930. },
  1931. ])
  1932. yield* replaySessionProjection(sessionID)
  1933. expect(yield* session.context(sessionID)).toMatchObject([
  1934. { type: "user", text: "Echo twice" },
  1935. {
  1936. type: "assistant",
  1937. content: [
  1938. {
  1939. type: "tool",
  1940. id: "tool_0",
  1941. state: { status: "completed", structured: { text: "first" }, content: [{ type: "text", text: "first" }] },
  1942. },
  1943. ],
  1944. },
  1945. {
  1946. type: "assistant",
  1947. content: [
  1948. {
  1949. type: "tool",
  1950. id: "tool_0",
  1951. state: {
  1952. status: "completed",
  1953. structured: { text: "second" },
  1954. content: [{ type: "text", text: "second" }],
  1955. },
  1956. },
  1957. ],
  1958. },
  1959. ])
  1960. }),
  1961. )
  1962. it.effect("joins concurrent resume calls into one active provider run", () =>
  1963. Effect.gen(function* () {
  1964. yield* setup
  1965. const session = yield* SessionV2.Service
  1966. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Run once" }), resume: false })
  1967. requests.length = 0
  1968. responses = undefined
  1969. response = [
  1970. LLMEvent.stepStart({ index: 0 }),
  1971. LLMEvent.textStart({ id: "text-once" }),
  1972. LLMEvent.textDelta({ id: "text-once", text: "Once" }),
  1973. LLMEvent.textEnd({ id: "text-once" }),
  1974. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  1975. LLMEvent.finish({ reason: "stop" }),
  1976. ]
  1977. streamGate = yield* Deferred.make<void>()
  1978. streamStarted = yield* Deferred.make<void>()
  1979. const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
  1980. yield* Deferred.await(streamStarted)
  1981. const second = yield* session.resume(sessionID).pipe(Effect.forkChild)
  1982. yield* Effect.yieldNow
  1983. expect(requests).toHaveLength(1)
  1984. yield* Deferred.succeed(streamGate, undefined)
  1985. yield* Fiber.join(first)
  1986. yield* Fiber.join(second)
  1987. streamGate = undefined
  1988. streamStarted = undefined
  1989. expect(requests).toHaveLength(1)
  1990. expect(yield* session.context(sessionID)).toMatchObject([
  1991. { type: "user", text: "Run once" },
  1992. { type: "assistant", finish: "stop", content: [{ type: "text", id: "text-once", text: "Once" }] },
  1993. ])
  1994. }),
  1995. )
  1996. it.effect("steers an active provider turn with newly recorded prompts", () =>
  1997. Effect.gen(function* () {
  1998. yield* setup
  1999. const session = yield* SessionV2.Service
  2000. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Start working" }), resume: false })
  2001. requests.length = 0
  2002. responses = [
  2003. [
  2004. LLMEvent.stepStart({ index: 0 }),
  2005. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2006. LLMEvent.finish({ reason: "stop" }),
  2007. ],
  2008. [
  2009. LLMEvent.stepStart({ index: 0 }),
  2010. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2011. LLMEvent.finish({ reason: "stop" }),
  2012. ],
  2013. ]
  2014. streamGate = yield* Deferred.make<void>()
  2015. streamStarted = yield* Deferred.make<void>()
  2016. const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2017. yield* Deferred.await(streamStarted)
  2018. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Change direction" }) })
  2019. yield* Deferred.succeed(streamGate, undefined)
  2020. yield* Fiber.join(first)
  2021. streamGate = undefined
  2022. streamStarted = undefined
  2023. yield* Effect.yieldNow
  2024. expect(requests).toHaveLength(2)
  2025. expect(userTexts(requests[0]!)).toEqual(["Start working"])
  2026. expect(userTexts(requests[1]!)).toEqual(["Start working", "Change direction"])
  2027. expect((yield* session.context(sessionID)).map((message) => message.type)).toEqual([
  2028. "user",
  2029. "assistant",
  2030. "user",
  2031. "assistant",
  2032. ])
  2033. }),
  2034. )
  2035. it.effect("starts queued input after the active activity settles", () =>
  2036. Effect.gen(function* () {
  2037. yield* setup
  2038. const session = yield* SessionV2.Service
  2039. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Start working" }), resume: false })
  2040. requests.length = 0
  2041. responses = [
  2042. [
  2043. LLMEvent.stepStart({ index: 0 }),
  2044. LLMEvent.toolCall({ id: "call-echo", name: "echo", input: { text: "hello" } }),
  2045. LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
  2046. LLMEvent.finish({ reason: "tool-calls" }),
  2047. ],
  2048. [
  2049. LLMEvent.stepStart({ index: 0 }),
  2050. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2051. LLMEvent.finish({ reason: "stop" }),
  2052. ],
  2053. [
  2054. LLMEvent.stepStart({ index: 0 }),
  2055. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2056. LLMEvent.finish({ reason: "stop" }),
  2057. ],
  2058. ]
  2059. streamGate = yield* Deferred.make<void>()
  2060. streamStarted = yield* Deferred.make<void>()
  2061. const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2062. yield* Deferred.await(streamStarted)
  2063. yield* session.prompt({
  2064. sessionID,
  2065. prompt: new Prompt({ text: "Wait until the next activity" }),
  2066. delivery: "queue",
  2067. })
  2068. yield* Deferred.succeed(streamGate, undefined)
  2069. yield* Fiber.join(first)
  2070. streamGate = undefined
  2071. streamStarted = undefined
  2072. expect(requests).toHaveLength(3)
  2073. expect(userTexts(requests[0]!)).toEqual(["Start working"])
  2074. expect(userTexts(requests[1]!)).toEqual(["Start working"])
  2075. expect(userTexts(requests[2]!)).toEqual(["Start working", "Wait until the next activity"])
  2076. }),
  2077. )
  2078. it.effect("preserves durable queued input for a later wake after interruption", () =>
  2079. Effect.gen(function* () {
  2080. yield* setup
  2081. const session = yield* SessionV2.Service
  2082. const { db } = yield* Database.Service
  2083. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Interrupt current work" }), resume: false })
  2084. requests.length = 0
  2085. responses = [
  2086. [],
  2087. [
  2088. LLMEvent.stepStart({ index: 0 }),
  2089. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2090. LLMEvent.finish({ reason: "stop" }),
  2091. ],
  2092. ]
  2093. streamGate = yield* Deferred.make<void>()
  2094. streamStarted = yield* Deferred.make<void>()
  2095. const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2096. yield* Deferred.await(streamStarted)
  2097. yield* session.prompt({
  2098. sessionID,
  2099. prompt: new Prompt({ text: "Run after interrupt" }),
  2100. delivery: "queue",
  2101. })
  2102. yield* session.interrupt(sessionID)
  2103. expect(yield* Fiber.await(run)).toMatchObject({ _tag: "Failure" })
  2104. expect(requests).toHaveLength(1)
  2105. expect(yield* SessionInput.hasPending(db, sessionID, "queue")).toBe(true)
  2106. const resumed = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2107. while (requests.length < 2) yield* Effect.yieldNow
  2108. yield* Deferred.succeed(streamGate, undefined)
  2109. yield* Fiber.join(resumed)
  2110. streamGate = undefined
  2111. streamStarted = undefined
  2112. expect(requests).toHaveLength(2)
  2113. expect(userTexts(requests[0]!)).toEqual(["Interrupt current work"])
  2114. expect(userTexts(requests[1]!)).toEqual(["Interrupt current work", "Run after interrupt"])
  2115. }),
  2116. )
  2117. it.effect("preserves durable steering input for a later resume after interruption", () =>
  2118. Effect.gen(function* () {
  2119. yield* setup
  2120. const session = yield* SessionV2.Service
  2121. const { db } = yield* Database.Service
  2122. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Interrupt current work" }), resume: false })
  2123. requests.length = 0
  2124. responses = [
  2125. [],
  2126. [
  2127. LLMEvent.stepStart({ index: 0 }),
  2128. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2129. LLMEvent.finish({ reason: "stop" }),
  2130. ],
  2131. ]
  2132. streamGate = yield* Deferred.make<void>()
  2133. streamStarted = yield* Deferred.make<void>()
  2134. const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2135. yield* Deferred.await(streamStarted)
  2136. yield* session.prompt({
  2137. sessionID,
  2138. prompt: new Prompt({ text: "Steer after interrupt" }),
  2139. })
  2140. yield* session.interrupt(sessionID)
  2141. expect(yield* Fiber.await(run)).toMatchObject({ _tag: "Failure" })
  2142. expect(requests).toHaveLength(1)
  2143. expect(yield* SessionInput.hasPending(db, sessionID, "steer")).toBe(true)
  2144. const resumed = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2145. while (requests.length < 2) yield* Effect.yieldNow
  2146. yield* Deferred.succeed(streamGate, undefined)
  2147. yield* Fiber.join(resumed)
  2148. streamGate = undefined
  2149. streamStarted = undefined
  2150. expect(requests).toHaveLength(2)
  2151. expect(userTexts(requests[0]!)).toEqual(["Interrupt current work"])
  2152. expect(userTexts(requests[1]!)).toEqual(["Interrupt current work", "Steer after interrupt"])
  2153. }),
  2154. )
  2155. it.effect("runs queued active inputs as separate FIFO activities", () =>
  2156. Effect.gen(function* () {
  2157. yield* setup
  2158. const session = yield* SessionV2.Service
  2159. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Start working" }), resume: false })
  2160. requests.length = 0
  2161. responses = [
  2162. [
  2163. LLMEvent.stepStart({ index: 0 }),
  2164. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2165. LLMEvent.finish({ reason: "stop" }),
  2166. ],
  2167. [
  2168. LLMEvent.stepStart({ index: 0 }),
  2169. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2170. LLMEvent.finish({ reason: "stop" }),
  2171. ],
  2172. [
  2173. LLMEvent.stepStart({ index: 0 }),
  2174. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2175. LLMEvent.finish({ reason: "stop" }),
  2176. ],
  2177. ]
  2178. streamGate = yield* Deferred.make<void>()
  2179. streamStarted = yield* Deferred.make<void>()
  2180. const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2181. yield* Deferred.await(streamStarted)
  2182. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Queue first" }), delivery: "queue" })
  2183. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Queue second" }), delivery: "queue" })
  2184. yield* Deferred.succeed(streamGate, undefined)
  2185. yield* Fiber.join(first)
  2186. streamGate = undefined
  2187. streamStarted = undefined
  2188. expect(requests).toHaveLength(3)
  2189. expect(userTexts(requests[0]!)).toEqual(["Start working"])
  2190. expect(userTexts(requests[1]!)).toEqual(["Start working", "Queue first"])
  2191. expect(userTexts(requests[2]!)).toEqual(["Start working", "Queue first", "Queue second"])
  2192. }),
  2193. )
  2194. it.effect("opens queued input after idle steering activity settles", () =>
  2195. Effect.gen(function* () {
  2196. yield* setup
  2197. const session = yield* SessionV2.Service
  2198. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Start steering activity" }), resume: false })
  2199. yield* session.prompt({
  2200. sessionID,
  2201. prompt: new Prompt({ text: "Queue later activity" }),
  2202. delivery: "queue",
  2203. resume: false,
  2204. })
  2205. requests.length = 0
  2206. responses = [
  2207. [
  2208. LLMEvent.stepStart({ index: 0 }),
  2209. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2210. LLMEvent.finish({ reason: "stop" }),
  2211. ],
  2212. [
  2213. LLMEvent.stepStart({ index: 0 }),
  2214. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2215. LLMEvent.finish({ reason: "stop" }),
  2216. ],
  2217. ]
  2218. yield* session.resume(sessionID)
  2219. expect(requests).toHaveLength(2)
  2220. expect(userTexts(requests[0]!)).toEqual(["Start steering activity"])
  2221. expect(userTexts(requests[1]!)).toEqual(["Start steering activity", "Queue later activity"])
  2222. }),
  2223. )
  2224. it.effect("coalesces steers into the active queued activity before starting the next queued activity", () =>
  2225. Effect.gen(function* () {
  2226. yield* setup
  2227. const session = yield* SessionV2.Service
  2228. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Start working" }), resume: false })
  2229. requests.length = 0
  2230. responses = [
  2231. [
  2232. LLMEvent.stepStart({ index: 0 }),
  2233. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2234. LLMEvent.finish({ reason: "stop" }),
  2235. ],
  2236. [
  2237. LLMEvent.stepStart({ index: 0 }),
  2238. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2239. LLMEvent.finish({ reason: "stop" }),
  2240. ],
  2241. [
  2242. LLMEvent.stepStart({ index: 0 }),
  2243. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2244. LLMEvent.finish({ reason: "stop" }),
  2245. ],
  2246. [
  2247. LLMEvent.stepStart({ index: 0 }),
  2248. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2249. LLMEvent.finish({ reason: "stop" }),
  2250. ],
  2251. ]
  2252. const firstGate = yield* Deferred.make<void>()
  2253. const secondGate = yield* Deferred.make<void>()
  2254. streamGate = firstGate
  2255. const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2256. while (requests.length < 1) yield* Effect.yieldNow
  2257. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Queue first" }), delivery: "queue" })
  2258. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Queue second" }), delivery: "queue" })
  2259. streamGate = secondGate
  2260. yield* Deferred.succeed(firstGate, undefined)
  2261. while (requests.length < 2) yield* Effect.yieldNow
  2262. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Steer first queued activity" }) })
  2263. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Also steer first queued activity" }) })
  2264. yield* Deferred.succeed(secondGate, undefined)
  2265. yield* Fiber.join(first)
  2266. streamGate = undefined
  2267. expect(requests).toHaveLength(4)
  2268. expect(userTexts(requests[0]!)).toEqual(["Start working"])
  2269. expect(userTexts(requests[1]!)).toEqual(["Start working", "Queue first"])
  2270. expect(userTexts(requests[2]!)).toEqual([
  2271. "Start working",
  2272. "Queue first",
  2273. "Steer first queued activity",
  2274. "Also steer first queued activity",
  2275. ])
  2276. expect(userTexts(requests[3]!)).toEqual([
  2277. "Start working",
  2278. "Queue first",
  2279. "Steer first queued activity",
  2280. "Also steer first queued activity",
  2281. "Queue second",
  2282. ])
  2283. }),
  2284. )
  2285. it.effect("coalesces multiple active steering prompts into one continuation turn", () =>
  2286. Effect.gen(function* () {
  2287. yield* setup
  2288. const session = yield* SessionV2.Service
  2289. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Start working" }), resume: false })
  2290. requests.length = 0
  2291. responses = [
  2292. [
  2293. LLMEvent.stepStart({ index: 0 }),
  2294. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2295. LLMEvent.finish({ reason: "stop" }),
  2296. ],
  2297. [
  2298. LLMEvent.stepStart({ index: 0 }),
  2299. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2300. LLMEvent.finish({ reason: "stop" }),
  2301. ],
  2302. ]
  2303. streamGate = yield* Deferred.make<void>()
  2304. streamStarted = yield* Deferred.make<void>()
  2305. const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2306. yield* Deferred.await(streamStarted)
  2307. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "First steer" }) })
  2308. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Second steer" }) })
  2309. yield* Deferred.succeed(streamGate, undefined)
  2310. yield* Fiber.join(first)
  2311. streamGate = undefined
  2312. streamStarted = undefined
  2313. yield* Effect.yieldNow
  2314. expect(requests).toHaveLength(2)
  2315. expect(userTexts(requests[1]!)).toEqual(["Start working", "First steer", "Second steer"])
  2316. yield* (yield* SessionRunCoordinator.Service).wake(sessionID)
  2317. yield* Effect.yieldNow
  2318. expect(requests).toHaveLength(2)
  2319. }),
  2320. )
  2321. it.effect("runs steering input accepted while the active provider turn fails", () =>
  2322. Effect.gen(function* () {
  2323. yield* setup
  2324. const session = yield* SessionV2.Service
  2325. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Start working" }), resume: false })
  2326. requests.length = 0
  2327. responses = undefined
  2328. response = []
  2329. streamFailure = providerUnavailable()
  2330. streamGate = yield* Deferred.make<void>()
  2331. streamStarted = yield* Deferred.make<void>()
  2332. const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2333. yield* Deferred.await(streamStarted)
  2334. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Recover with this" }) })
  2335. yield* Deferred.succeed(streamGate, undefined)
  2336. expect(yield* Fiber.join(first).pipe(Effect.flip)).toBe(streamFailure)
  2337. streamFailure = undefined
  2338. streamGate = undefined
  2339. streamStarted = undefined
  2340. yield* Effect.yieldNow
  2341. expect(requests).toHaveLength(2)
  2342. expect(userTexts(requests[1]!)).toEqual(["Start working", "Recover with this"])
  2343. }),
  2344. )
  2345. it.effect("durably fails local tools left running by a prior process before continuing", () =>
  2346. Effect.gen(function* () {
  2347. yield* setup
  2348. const session = yield* SessionV2.Service
  2349. const events = yield* EventV2.Service
  2350. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Recover interrupted tool" }), resume: false })
  2351. yield* SessionInput.promoteSteers((yield* Database.Service).db, events, sessionID, Number.MAX_SAFE_INTEGER)
  2352. const assistantMessageID = SessionMessage.ID.create()
  2353. yield* events.publish(SessionEvent.Step.Started, {
  2354. sessionID,
  2355. assistantMessageID,
  2356. timestamp: yield* DateTime.now,
  2357. agent: "build",
  2358. model: { id: ModelV2.ID.make("fake-model"), providerID: ProviderV2.ID.make("fake") },
  2359. })
  2360. yield* events.publish(SessionEvent.Tool.Input.Started, {
  2361. sessionID,
  2362. timestamp: yield* DateTime.now,
  2363. assistantMessageID,
  2364. callID: "call-interrupted",
  2365. name: "echo",
  2366. })
  2367. yield* events.publish(SessionEvent.Tool.Input.Ended, {
  2368. sessionID,
  2369. timestamp: yield* DateTime.now,
  2370. assistantMessageID,
  2371. callID: "call-interrupted",
  2372. text: '{"text":"stale"}',
  2373. })
  2374. yield* events.publish(SessionEvent.Tool.Called, {
  2375. sessionID,
  2376. timestamp: yield* DateTime.now,
  2377. assistantMessageID,
  2378. callID: "call-interrupted",
  2379. tool: "echo",
  2380. input: { text: "stale" },
  2381. provider: { executed: false },
  2382. })
  2383. requests.length = 0
  2384. response = []
  2385. yield* session.resume(sessionID)
  2386. expect(requests).toHaveLength(1)
  2387. expect(requests[0]?.messages.map((message) => message.role)).toEqual(["user", "assistant", "tool"])
  2388. expect(yield* session.context(sessionID)).toMatchObject([
  2389. { type: "user", text: "Recover interrupted tool" },
  2390. {
  2391. type: "assistant",
  2392. content: [
  2393. {
  2394. type: "tool",
  2395. id: "call-interrupted",
  2396. state: { status: "error", error: { type: "unknown", message: "Tool execution interrupted" } },
  2397. },
  2398. ],
  2399. },
  2400. ])
  2401. }),
  2402. )
  2403. it.effect("durably fails hosted tools left running by a prior process before continuing inline", () =>
  2404. Effect.gen(function* () {
  2405. yield* setup
  2406. const session = yield* SessionV2.Service
  2407. const events = yield* EventV2.Service
  2408. yield* session.prompt({
  2409. sessionID,
  2410. prompt: new Prompt({ text: "Recover interrupted hosted tool" }),
  2411. resume: false,
  2412. })
  2413. yield* SessionInput.promoteSteers((yield* Database.Service).db, events, sessionID, Number.MAX_SAFE_INTEGER)
  2414. const assistantMessageID = SessionMessage.ID.create()
  2415. yield* events.publish(SessionEvent.Step.Started, {
  2416. sessionID,
  2417. assistantMessageID,
  2418. timestamp: yield* DateTime.now,
  2419. agent: "build",
  2420. model: { id: ModelV2.ID.make("fake-model"), providerID: ProviderV2.ID.make("fake") },
  2421. })
  2422. yield* events.publish(SessionEvent.Tool.Input.Started, {
  2423. sessionID,
  2424. timestamp: yield* DateTime.now,
  2425. assistantMessageID,
  2426. callID: "call-hosted-interrupted",
  2427. name: "web_search",
  2428. })
  2429. yield* events.publish(SessionEvent.Tool.Input.Ended, {
  2430. sessionID,
  2431. timestamp: yield* DateTime.now,
  2432. assistantMessageID,
  2433. callID: "call-hosted-interrupted",
  2434. text: '{"query":"stale"}',
  2435. })
  2436. yield* events.publish(SessionEvent.Tool.Called, {
  2437. sessionID,
  2438. timestamp: yield* DateTime.now,
  2439. assistantMessageID,
  2440. callID: "call-hosted-interrupted",
  2441. tool: "web_search",
  2442. input: { query: "stale" },
  2443. provider: { executed: true, metadata: { openai: { itemId: "call-hosted-interrupted" } } },
  2444. })
  2445. requests.length = 0
  2446. response = []
  2447. yield* session.resume(sessionID)
  2448. expect(requests).toHaveLength(1)
  2449. expect(requests[0]?.messages.map((message) => message.role)).toEqual(["user", "assistant"])
  2450. expect(requests[0]?.messages[1]?.content).toMatchObject([
  2451. {
  2452. type: "tool-call",
  2453. id: "call-hosted-interrupted",
  2454. providerExecuted: true,
  2455. providerMetadata: { openai: { itemId: "call-hosted-interrupted" } },
  2456. },
  2457. { type: "tool-result", id: "call-hosted-interrupted", providerExecuted: true, result: { type: "error" } },
  2458. ])
  2459. }),
  2460. )
  2461. it.effect("durably fails pending tool input left by a prior process before continuing", () =>
  2462. Effect.gen(function* () {
  2463. yield* setup
  2464. const session = yield* SessionV2.Service
  2465. const events = yield* EventV2.Service
  2466. yield* session.prompt({
  2467. sessionID,
  2468. prompt: new Prompt({ text: "Recover interrupted tool input" }),
  2469. resume: false,
  2470. })
  2471. yield* SessionInput.promoteSteers((yield* Database.Service).db, events, sessionID, Number.MAX_SAFE_INTEGER)
  2472. const assistantMessageID = SessionMessage.ID.create()
  2473. yield* events.publish(SessionEvent.Step.Started, {
  2474. sessionID,
  2475. assistantMessageID,
  2476. timestamp: yield* DateTime.now,
  2477. agent: "build",
  2478. model: { id: ModelV2.ID.make("fake-model"), providerID: ProviderV2.ID.make("fake") },
  2479. })
  2480. yield* events.publish(SessionEvent.Tool.Input.Started, {
  2481. sessionID,
  2482. timestamp: yield* DateTime.now,
  2483. assistantMessageID,
  2484. callID: "call-pending-interrupted",
  2485. name: "echo",
  2486. })
  2487. requests.length = 0
  2488. response = []
  2489. yield* session.resume(sessionID)
  2490. expect(requests).toHaveLength(1)
  2491. expect(requests[0]?.messages.map((message) => message.role)).toEqual(["user", "assistant", "tool"])
  2492. expect(yield* session.context(sessionID)).toMatchObject([
  2493. { type: "user", text: "Recover interrupted tool input" },
  2494. { type: "assistant", content: [{ type: "tool", id: "call-pending-interrupted", state: { status: "error" } }] },
  2495. ])
  2496. }),
  2497. )
  2498. it.effect("starts the first queued activity when woken while idle", () =>
  2499. Effect.gen(function* () {
  2500. yield* setup
  2501. const session = yield* SessionV2.Service
  2502. yield* session.prompt({
  2503. sessionID,
  2504. prompt: new Prompt({ text: "Wait for fresh activity" }),
  2505. delivery: "queue",
  2506. resume: false,
  2507. })
  2508. requests.length = 0
  2509. yield* (yield* SessionRunCoordinator.Service).wake(sessionID)
  2510. yield* Effect.yieldNow
  2511. expect(requests).toHaveLength(1)
  2512. expect(userTexts(requests[0]!)).toEqual(["Wait for fresh activity"])
  2513. }),
  2514. )
  2515. it.effect("does not spend one activity step budget across queued activities", () =>
  2516. Effect.gen(function* () {
  2517. yield* setup
  2518. const session = yield* SessionV2.Service
  2519. const queued = Array.from({ length: 26 }, (_, index) => `Queued activity ${index + 1}`)
  2520. for (const text of queued) {
  2521. yield* session.prompt({ sessionID, prompt: new Prompt({ text }), delivery: "queue", resume: false })
  2522. }
  2523. requests.length = 0
  2524. responses = queued.map(() => [
  2525. LLMEvent.stepStart({ index: 0 }),
  2526. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2527. LLMEvent.finish({ reason: "stop" }),
  2528. ])
  2529. yield* session.resume(sessionID)
  2530. expect(requests).toHaveLength(queued.length)
  2531. expect(userTexts(requests.at(-1)!)).toEqual(queued)
  2532. }),
  2533. )
  2534. it.effect("retries inbox input after prompt projection rolls back", () =>
  2535. Effect.gen(function* () {
  2536. yield* setup
  2537. const session = yield* SessionV2.Service
  2538. const events = yield* EventV2.Service
  2539. const defect = new Error("fail after prompt promotion")
  2540. let fail = true
  2541. yield* events.project(SessionEvent.PromptLifecycle.Promoted, () => (fail ? Effect.die(defect) : Effect.void))
  2542. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Recover promoted input" }), resume: false })
  2543. expect(yield* session.resume(sessionID).pipe(Effect.catchDefect(Effect.succeed))).toBe(defect)
  2544. fail = false
  2545. requests.length = 0
  2546. response = [
  2547. LLMEvent.stepStart({ index: 0 }),
  2548. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2549. LLMEvent.finish({ reason: "stop" }),
  2550. ]
  2551. yield* (yield* SessionRunCoordinator.Service).wake(sessionID)
  2552. while (requests.length === 0) yield* Effect.yieldNow
  2553. expect(userTexts(requests[0]!)).toEqual(["Recover promoted input"])
  2554. }),
  2555. )
  2556. it.effect("does not strand a committed promotion when a post-commit listener defects", () =>
  2557. Effect.gen(function* () {
  2558. yield* setup
  2559. const session = yield* SessionV2.Service
  2560. const events = yield* EventV2.Service
  2561. yield* events.listen((event) =>
  2562. event.type === SessionEvent.PromptLifecycle.Promoted.type
  2563. ? Effect.die("fail after prompt promotion commits")
  2564. : Effect.void,
  2565. )
  2566. yield* session.prompt({
  2567. sessionID,
  2568. prompt: new Prompt({ text: "Run committed promotion" }),
  2569. resume: false,
  2570. })
  2571. requests.length = 0
  2572. yield* session.resume(sessionID)
  2573. expect(requests).toHaveLength(1)
  2574. expect(userTexts(requests[0]!)).toEqual(["Run committed promotion"])
  2575. }),
  2576. )
  2577. it.effect("runs different sessions concurrently", () =>
  2578. Effect.gen(function* () {
  2579. yield* setup
  2580. yield* insertSession(otherSessionID)
  2581. const session = yield* SessionV2.Service
  2582. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Run first" }), resume: false })
  2583. yield* session.prompt({ sessionID: otherSessionID, prompt: new Prompt({ text: "Run second" }), resume: false })
  2584. requests.length = 0
  2585. responses = undefined
  2586. response = []
  2587. streamGate = yield* Deferred.make<void>()
  2588. streamStarted = yield* Deferred.make<void>()
  2589. const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2590. yield* Deferred.await(streamStarted)
  2591. const second = yield* session.resume(otherSessionID).pipe(Effect.forkChild)
  2592. yield* Effect.yieldNow
  2593. expect(requests).toHaveLength(2)
  2594. expect(requests.map((request) => request.providerOptions?.openai?.promptCacheKey)).toEqual([
  2595. sessionID,
  2596. otherSessionID,
  2597. ])
  2598. yield* Deferred.succeed(streamGate, undefined)
  2599. yield* Fiber.join(first)
  2600. yield* Fiber.join(second)
  2601. streamGate = undefined
  2602. streamStarted = undefined
  2603. }),
  2604. )
  2605. it.effect("bounds external session prompt cache keys", () =>
  2606. Effect.gen(function* () {
  2607. yield* setup
  2608. const externalSessionID = SessionV2.ID.fromExternal({
  2609. namespace: "discord",
  2610. key: "thread-one",
  2611. })
  2612. const otherExternalSessionID = SessionV2.ID.fromExternal({
  2613. namespace: "discord",
  2614. key: "thread-two",
  2615. })
  2616. yield* insertSession(externalSessionID)
  2617. yield* insertSession(otherExternalSessionID)
  2618. const session = yield* SessionV2.Service
  2619. yield* session.prompt({
  2620. sessionID: externalSessionID,
  2621. prompt: new Prompt({ text: "Run external session" }),
  2622. resume: false,
  2623. })
  2624. yield* session.prompt({
  2625. sessionID: otherExternalSessionID,
  2626. prompt: new Prompt({ text: "Run other external session" }),
  2627. resume: false,
  2628. })
  2629. requests.length = 0
  2630. yield* session.resume(externalSessionID)
  2631. yield* session.resume(otherExternalSessionID)
  2632. const keys = requests.map((request) => request.providerOptions?.openai?.promptCacheKey)
  2633. expect(keys).toEqual([externalSessionID.slice(4), otherExternalSessionID.slice(4)])
  2634. expect(keys.every((key) => typeof key === "string" && key.length === 64)).toBe(true)
  2635. expect(keys[0]).not.toBe(keys[1])
  2636. }),
  2637. )
  2638. it.effect("fans out one failed run and allows a later retry", () =>
  2639. Effect.gen(function* () {
  2640. yield* setup
  2641. const session = yield* SessionV2.Service
  2642. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Retry after failure" }), resume: false })
  2643. requests.length = 0
  2644. responses = undefined
  2645. response = []
  2646. streamFailure = providerUnavailable()
  2647. streamGate = yield* Deferred.make<void>()
  2648. streamStarted = yield* Deferred.make<void>()
  2649. const first = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2650. yield* Deferred.await(streamStarted)
  2651. const second = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2652. yield* Effect.yieldNow
  2653. expect(requests).toHaveLength(1)
  2654. yield* Deferred.succeed(streamGate, undefined)
  2655. const [firstExit, secondExit] = yield* Effect.all([Fiber.await(first), Fiber.await(second)])
  2656. expect(secondExit).toEqual(firstExit)
  2657. streamFailure = undefined
  2658. streamGate = undefined
  2659. streamStarted = undefined
  2660. yield* session.resume(sessionID)
  2661. expect(requests).toHaveLength(2)
  2662. }),
  2663. )
  2664. it.effect("durably settles local tool failures before continuing", () =>
  2665. Effect.gen(function* () {
  2666. yield* setup
  2667. const session = yield* SessionV2.Service
  2668. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Call missing" }), resume: false })
  2669. requests.length = 0
  2670. responses = [
  2671. [
  2672. LLMEvent.stepStart({ index: 0 }),
  2673. LLMEvent.toolCall({ id: "call-missing", name: "missing", input: {} }),
  2674. LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
  2675. LLMEvent.finish({ reason: "tool-calls" }),
  2676. ],
  2677. [
  2678. LLMEvent.stepStart({ index: 0 }),
  2679. LLMEvent.textStart({ id: "text-after-error" }),
  2680. LLMEvent.textDelta({ id: "text-after-error", text: "Recovered" }),
  2681. LLMEvent.textEnd({ id: "text-after-error" }),
  2682. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2683. LLMEvent.finish({ reason: "stop" }),
  2684. ],
  2685. ]
  2686. streamGate = undefined
  2687. streamStarted = undefined
  2688. yield* session.resume(sessionID)
  2689. expect(requests).toHaveLength(2)
  2690. expect(yield* session.context(sessionID)).toMatchObject([
  2691. { type: "user", text: "Call missing" },
  2692. {
  2693. type: "assistant",
  2694. content: [
  2695. {
  2696. type: "tool",
  2697. id: "call-missing",
  2698. state: { status: "error", error: { message: "Unknown tool: missing" } },
  2699. },
  2700. ],
  2701. },
  2702. { type: "assistant", finish: "stop", content: [{ type: "text", id: "text-after-error", text: "Recovered" }] },
  2703. ])
  2704. }),
  2705. )
  2706. it.effect("propagates unexpected local tool defects operationally", () =>
  2707. Effect.gen(function* () {
  2708. yield* setup
  2709. const session = yield* SessionV2.Service
  2710. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Call defect" }), resume: false })
  2711. requests.length = 0
  2712. responses = [
  2713. [
  2714. LLMEvent.stepStart({ index: 0 }),
  2715. LLMEvent.toolCall({ id: "call-defect", name: "defect", input: {} }),
  2716. LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
  2717. LLMEvent.finish({ reason: "tool-calls" }),
  2718. ],
  2719. ]
  2720. expect(yield* session.resume(sessionID).pipe(Effect.catchDefect(Effect.succeed))).toBe("unexpected tool defect")
  2721. expect(requests).toHaveLength(1)
  2722. }),
  2723. )
  2724. it.effect("interrupts runner continuation when a question is dismissed", () =>
  2725. Effect.gen(function* () {
  2726. yield* setup
  2727. const session = yield* SessionV2.Service
  2728. const registry = yield* ToolRegistry.Service
  2729. const questions = yield* QuestionV2.Service
  2730. yield* registry.register({
  2731. question: Tool.make({
  2732. description: "Ask the user",
  2733. input: Schema.Struct({}),
  2734. output: Schema.Struct({}),
  2735. execute: (_, context) =>
  2736. questions.ask({ sessionID: context.sessionID, questions: [] }).pipe(Effect.as({}), Effect.orDie),
  2737. }),
  2738. })
  2739. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Ask then stop" }), resume: false })
  2740. requests.length = 0
  2741. responses = [
  2742. [
  2743. LLMEvent.stepStart({ index: 0 }),
  2744. LLMEvent.toolCall({ id: "call-question", name: "question", input: {} }),
  2745. LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
  2746. LLMEvent.finish({ reason: "tool-calls" }),
  2747. ],
  2748. [],
  2749. ]
  2750. const run = yield* session.resume(sessionID).pipe(Effect.exit, Effect.forkChild)
  2751. let pending = yield* questions.list()
  2752. while (pending.length === 0) {
  2753. yield* Effect.yieldNow
  2754. pending = yield* questions.list()
  2755. }
  2756. yield* questions.reject(pending[0]!.id)
  2757. const exit = yield* Fiber.join(run)
  2758. expect(exit._tag).toBe("Failure")
  2759. if (exit._tag === "Failure") expect(Cause.hasInterruptsOnly(exit.cause)).toBe(true)
  2760. expect(requests).toHaveLength(1)
  2761. expect(yield* session.context(sessionID)).toMatchObject([
  2762. { type: "user", text: "Ask then stop" },
  2763. {
  2764. type: "assistant",
  2765. content: [
  2766. {
  2767. type: "tool",
  2768. id: "call-question",
  2769. state: { status: "error", error: { type: "unknown", message: "Tool execution interrupted" } },
  2770. },
  2771. ],
  2772. },
  2773. ])
  2774. }),
  2775. )
  2776. it.effect("awaits started local tools before surfacing provider stream failure", () =>
  2777. Effect.gen(function* () {
  2778. yield* setup
  2779. const session = yield* SessionV2.Service
  2780. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Settle before failing" }), resume: false })
  2781. const failure = providerUnavailable()
  2782. toolExecutionGate = yield* Deferred.make<void>()
  2783. responseStream = Stream.concat(
  2784. Stream.fromIterable([
  2785. LLMEvent.stepStart({ index: 0 }),
  2786. LLMEvent.toolCall({ id: "call-before-failure", name: "echo", input: { text: "settle" } }),
  2787. ]),
  2788. Stream.fail(failure),
  2789. )
  2790. const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2791. while (executions.length === 0) yield* Effect.yieldNow
  2792. yield* Effect.yieldNow
  2793. yield* Deferred.succeed(toolExecutionGate, undefined)
  2794. expect(yield* Fiber.join(run).pipe(Effect.flip)).toBe(failure)
  2795. toolExecutionGate = undefined
  2796. expect(yield* session.context(sessionID)).toMatchObject([
  2797. { type: "user", text: "Settle before failing" },
  2798. {
  2799. type: "assistant",
  2800. content: [
  2801. { type: "tool", id: "call-before-failure", state: { status: "completed", structured: { text: "settle" } } },
  2802. ],
  2803. },
  2804. ])
  2805. }),
  2806. )
  2807. it.effect("durably fails blocked local tools when a provider turn is interrupted", () =>
  2808. Effect.gen(function* () {
  2809. yield* setup
  2810. const session = yield* SessionV2.Service
  2811. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Interrupt blocked tool" }), resume: false })
  2812. executions.length = 0
  2813. toolExecutionGate = yield* Deferred.make<void>()
  2814. responseStream = Stream.concat(
  2815. Stream.fromIterable([
  2816. LLMEvent.stepStart({ index: 0 }),
  2817. LLMEvent.toolCall({ id: "call-before-interrupt", name: "echo", input: { text: "blocked" } }),
  2818. ]),
  2819. Stream.never,
  2820. )
  2821. const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2822. while (executions.length === 0) yield* Effect.yieldNow
  2823. yield* session.interrupt(sessionID)
  2824. toolExecutionGate = undefined
  2825. expect(yield* Fiber.await(run)).toMatchObject({ _tag: "Failure" })
  2826. yield* session.interrupt(sessionID)
  2827. expect(yield* session.context(sessionID)).toMatchObject([
  2828. { type: "user", text: "Interrupt blocked tool" },
  2829. {
  2830. type: "assistant",
  2831. content: [
  2832. {
  2833. type: "tool",
  2834. id: "call-before-interrupt",
  2835. state: { status: "error", error: { type: "unknown", message: "Tool execution interrupted" } },
  2836. },
  2837. ],
  2838. },
  2839. ])
  2840. yield* replaySessionProjection(sessionID)
  2841. expect(yield* session.context(sessionID)).toMatchObject([
  2842. { type: "user", text: "Interrupt blocked tool" },
  2843. { type: "assistant", content: [{ type: "tool", id: "call-before-interrupt", state: { status: "error" } }] },
  2844. ])
  2845. requests.length = 0
  2846. responseStream = undefined
  2847. response = []
  2848. yield* session.resume(sessionID)
  2849. expect(requests[0]?.messages.map((message) => message.role)).toEqual(["user", "assistant", "tool"])
  2850. }),
  2851. )
  2852. it.effect("interrupts a blocked provider turn without local tool activity", () =>
  2853. Effect.gen(function* () {
  2854. yield* setup
  2855. const session = yield* SessionV2.Service
  2856. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Interrupt provider" }), resume: false })
  2857. requests.length = 0
  2858. response = []
  2859. streamGate = yield* Deferred.make<void>()
  2860. streamStarted = yield* Deferred.make<void>()
  2861. const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2862. yield* Deferred.await(streamStarted)
  2863. yield* session.interrupt(sessionID)
  2864. const exit = yield* Fiber.await(run)
  2865. streamGate = undefined
  2866. streamStarted = undefined
  2867. expect(Exit.isFailure(exit) && Cause.hasInterruptsOnly(exit.cause)).toBeTrue()
  2868. expect(requests).toHaveLength(1)
  2869. yield* session.interrupt(sessionID)
  2870. }),
  2871. )
  2872. it.effect("durably fails blocked local tools when interrupted while awaiting settlement", () =>
  2873. Effect.gen(function* () {
  2874. yield* setup
  2875. const session = yield* SessionV2.Service
  2876. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Interrupt tool settlement" }), resume: false })
  2877. executions.length = 0
  2878. toolExecutionGate = yield* Deferred.make<void>()
  2879. response = [
  2880. LLMEvent.stepStart({ index: 0 }),
  2881. LLMEvent.toolCall({ id: "call-await-interrupt", name: "echo", input: { text: "blocked" } }),
  2882. LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
  2883. LLMEvent.finish({ reason: "tool-calls" }),
  2884. ]
  2885. const runner = yield* SessionRunner.Service
  2886. const run = yield* runner.run({ sessionID, force: true }).pipe(Effect.forkChild)
  2887. while (executions.length === 0) yield* Effect.yieldNow
  2888. yield* Fiber.interrupt(run)
  2889. toolExecutionGate = undefined
  2890. expect(yield* Fiber.await(run)).toMatchObject({ _tag: "Failure" })
  2891. expect(yield* session.context(sessionID)).toMatchObject([
  2892. { type: "user", text: "Interrupt tool settlement" },
  2893. {
  2894. type: "assistant",
  2895. content: [
  2896. {
  2897. type: "tool",
  2898. id: "call-await-interrupt",
  2899. state: { status: "error", error: { type: "unknown", message: "Tool execution interrupted" } },
  2900. },
  2901. ],
  2902. },
  2903. ])
  2904. }),
  2905. )
  2906. it.effect("fails after the bounded number of local tool continuation steps", () =>
  2907. Effect.gen(function* () {
  2908. yield* setup
  2909. const session = yield* SessionV2.Service
  2910. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Loop forever" }), resume: false })
  2911. requests.length = 0
  2912. authorizations.length = 0
  2913. executions.length = 0
  2914. streamGate = undefined
  2915. streamStarted = undefined
  2916. responses = Array.from({ length: 25 }, (_, index) => [
  2917. LLMEvent.stepStart({ index: 0 }),
  2918. LLMEvent.toolCall({ id: `call-echo-${index}`, name: "echo", input: { text: `${index}` } }),
  2919. LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
  2920. LLMEvent.finish({ reason: "tool-calls" }),
  2921. ])
  2922. const failure = yield* session.resume(sessionID).pipe(Effect.flip)
  2923. expect(failure).toMatchObject({ _tag: "SessionRunner.StepLimitExceededError", sessionID, limit: 25 })
  2924. expect(requests).toHaveLength(25)
  2925. expect(executions).toHaveLength(25)
  2926. }),
  2927. )
  2928. it.effect("does not restart a capped tool loop for a coalesced stale wake", () =>
  2929. Effect.gen(function* () {
  2930. yield* setup
  2931. const session = yield* SessionV2.Service
  2932. const coordinator = yield* SessionRunCoordinator.Service
  2933. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Loop forever" }), resume: false })
  2934. requests.length = 0
  2935. responses = Array.from({ length: 25 }, (_, index) => [
  2936. LLMEvent.stepStart({ index: 0 }),
  2937. LLMEvent.toolCall({ id: `call-capped-${index}`, name: "echo", input: { text: `${index}` } }),
  2938. LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
  2939. LLMEvent.finish({ reason: "tool-calls" }),
  2940. ])
  2941. streamGate = yield* Deferred.make<void>()
  2942. streamStarted = yield* Deferred.make<void>()
  2943. const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
  2944. yield* Deferred.await(streamStarted)
  2945. yield* coordinator.wake(sessionID)
  2946. yield* Deferred.succeed(streamGate, undefined)
  2947. expect(yield* Fiber.join(run).pipe(Effect.flip)).toMatchObject({ _tag: "SessionRunner.StepLimitExceededError" })
  2948. streamGate = undefined
  2949. streamStarted = undefined
  2950. yield* Effect.yieldNow
  2951. expect(requests).toHaveLength(25)
  2952. }),
  2953. )
  2954. it.effect("accepts a terminal response on the final bounded provider turn", () =>
  2955. Effect.gen(function* () {
  2956. yield* setup
  2957. const session = yield* SessionV2.Service
  2958. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Finish at the limit" }), resume: false })
  2959. requests.length = 0
  2960. responses = [
  2961. ...Array.from({ length: 24 }, (_, index) => [
  2962. LLMEvent.stepStart({ index: 0 }),
  2963. LLMEvent.toolCall({ id: `call-terminal-${index}`, name: "echo", input: { text: `${index}` } }),
  2964. LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
  2965. LLMEvent.finish({ reason: "tool-calls" }),
  2966. ]),
  2967. [
  2968. LLMEvent.stepStart({ index: 0 }),
  2969. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  2970. LLMEvent.finish({ reason: "stop" }),
  2971. ],
  2972. ]
  2973. yield* session.resume(sessionID)
  2974. expect(requests).toHaveLength(25)
  2975. }),
  2976. )
  2977. it.effect("projects provider errors as terminal assistant step failures", () =>
  2978. Effect.gen(function* () {
  2979. yield* setup
  2980. const session = yield* SessionV2.Service
  2981. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Fail durably" }), resume: false })
  2982. requests.length = 0
  2983. responses = undefined
  2984. streamGate = undefined
  2985. streamStarted = undefined
  2986. response = [LLMEvent.stepStart({ index: 0 }), LLMEvent.providerError({ message: "Provider unavailable" })]
  2987. yield* session.resume(sessionID)
  2988. expect(requests).toHaveLength(1)
  2989. expect(yield* session.context(sessionID)).toMatchObject([
  2990. { type: "user", text: "Fail durably" },
  2991. { type: "assistant", finish: "error", error: { type: "unknown", message: "Provider unavailable" } },
  2992. ])
  2993. }),
  2994. )
  2995. it.effect("projects provider errors emitted before assistant step start", () =>
  2996. Effect.gen(function* () {
  2997. yield* setup
  2998. const session = yield* SessionV2.Service
  2999. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Fail before step" }), resume: false })
  3000. requests.length = 0
  3001. response = [LLMEvent.providerError({ message: "Provider unavailable" })]
  3002. yield* session.resume(sessionID)
  3003. expect(requests).toHaveLength(1)
  3004. expect(yield* session.context(sessionID)).toMatchObject([
  3005. { type: "user", text: "Fail before step" },
  3006. { type: "assistant", finish: "error", error: { type: "unknown", message: "Provider unavailable" } },
  3007. ])
  3008. }),
  3009. )
  3010. it.effect("does not recover context overflow after durable assistant output", () =>
  3011. Effect.gen(function* () {
  3012. yield* setup
  3013. const session = yield* SessionV2.Service
  3014. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Fail after output" }), resume: false })
  3015. requests.length = 0
  3016. response = [
  3017. LLMEvent.stepStart({ index: 0 }),
  3018. LLMEvent.textStart({ id: "text-partial" }),
  3019. LLMEvent.textDelta({ id: "text-partial", text: "Partial" }),
  3020. LLMEvent.textEnd({ id: "text-partial" }),
  3021. LLMEvent.providerError({ message: "prompt too long", classification: "context-overflow" }),
  3022. ]
  3023. yield* session.resume(sessionID)
  3024. expect(requests).toHaveLength(1)
  3025. expect(yield* session.context(sessionID)).toMatchObject([
  3026. { type: "user", text: "Fail after output" },
  3027. {
  3028. type: "assistant",
  3029. finish: "error",
  3030. error: { message: "prompt too long" },
  3031. content: [{ type: "text", text: "Partial" }],
  3032. },
  3033. ])
  3034. }),
  3035. )
  3036. it.effect("projects raw provider stream failures as terminal assistant step failures", () =>
  3037. Effect.gen(function* () {
  3038. yield* setup
  3039. const session = yield* SessionV2.Service
  3040. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Fail raw stream durably" }), resume: false })
  3041. const failure = providerUnavailable()
  3042. responseStream = Stream.fail(failure)
  3043. expect(yield* session.resume(sessionID).pipe(Effect.flip)).toBe(failure)
  3044. yield* replaySessionProjection(sessionID)
  3045. expect(yield* session.context(sessionID)).toMatchObject([
  3046. { type: "user", text: "Fail raw stream durably" },
  3047. { type: "assistant", finish: "error", error: { type: "unknown", message: "Provider unavailable" } },
  3048. ])
  3049. }),
  3050. )
  3051. it.effect("does not continue automatically after a provider error follows a local tool call", () =>
  3052. Effect.gen(function* () {
  3053. yield* setup
  3054. const session = yield* SessionV2.Service
  3055. yield* session.prompt({
  3056. sessionID,
  3057. prompt: new Prompt({ text: "Do not continue failed provider" }),
  3058. resume: false,
  3059. })
  3060. requests.length = 0
  3061. const executionCount = executions.length
  3062. response = [
  3063. LLMEvent.stepStart({ index: 0 }),
  3064. LLMEvent.toolCall({ id: "call-before-provider-error", name: "echo", input: { text: "settled" } }),
  3065. LLMEvent.providerError({ message: "Provider unavailable" }),
  3066. ]
  3067. yield* session.resume(sessionID)
  3068. expect(requests).toHaveLength(1)
  3069. expect(executions.slice(executionCount)).toEqual(["settled"])
  3070. }),
  3071. )
  3072. it.effect("durably fails a hosted tool when its provider errors before returning a result", () =>
  3073. Effect.gen(function* () {
  3074. yield* setup
  3075. const session = yield* SessionV2.Service
  3076. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Fail hosted tool durably" }), resume: false })
  3077. requests.length = 0
  3078. response = [
  3079. LLMEvent.stepStart({ index: 0 }),
  3080. LLMEvent.toolCall({
  3081. id: "call-hosted-provider-error",
  3082. name: "web_search",
  3083. input: { query: "effect" },
  3084. providerExecuted: true,
  3085. }),
  3086. LLMEvent.providerError({ message: "Provider unavailable" }),
  3087. ]
  3088. yield* session.resume(sessionID)
  3089. expect(requests).toHaveLength(1)
  3090. expect(yield* session.context(sessionID)).toMatchObject([
  3091. { type: "user", text: "Fail hosted tool durably" },
  3092. {
  3093. type: "assistant",
  3094. content: [{ type: "tool", id: "call-hosted-provider-error", state: { status: "error" } }],
  3095. },
  3096. ])
  3097. }),
  3098. )
  3099. it.effect("durably fails a hosted tool left unresolved at normal provider EOF", () =>
  3100. Effect.gen(function* () {
  3101. yield* setup
  3102. const session = yield* SessionV2.Service
  3103. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Fail hosted tool at EOF" }), resume: false })
  3104. response = [
  3105. LLMEvent.stepStart({ index: 0 }),
  3106. LLMEvent.toolCall({
  3107. id: "call-hosted-eof",
  3108. name: "web_search",
  3109. input: { query: "effect" },
  3110. providerExecuted: true,
  3111. }),
  3112. ]
  3113. yield* session.resume(sessionID)
  3114. yield* replaySessionProjection(sessionID)
  3115. expect(yield* session.context(sessionID)).toMatchObject([
  3116. { type: "user", text: "Fail hosted tool at EOF" },
  3117. { type: "assistant", content: [{ type: "tool", id: "call-hosted-eof", state: { status: "error" } }] },
  3118. ])
  3119. }),
  3120. )
  3121. it.effect("durably fails a hosted tool left unresolved by a raw provider stream failure", () =>
  3122. Effect.gen(function* () {
  3123. yield* setup
  3124. const session = yield* SessionV2.Service
  3125. yield* session.prompt({
  3126. sessionID,
  3127. prompt: new Prompt({ text: "Fail hosted tool on raw failure" }),
  3128. resume: false,
  3129. })
  3130. const failure = providerUnavailable()
  3131. responseStream = Stream.concat(
  3132. Stream.fromIterable([
  3133. LLMEvent.stepStart({ index: 0 }),
  3134. LLMEvent.toolCall({
  3135. id: "call-hosted-raw-failure",
  3136. name: "web_search",
  3137. input: { query: "effect" },
  3138. providerExecuted: true,
  3139. }),
  3140. ]),
  3141. Stream.fail(failure),
  3142. )
  3143. expect(yield* session.resume(sessionID).pipe(Effect.flip)).toBe(failure)
  3144. yield* replaySessionProjection(sessionID)
  3145. expect(yield* session.context(sessionID)).toMatchObject([
  3146. { type: "user", text: "Fail hosted tool on raw failure" },
  3147. {
  3148. type: "assistant",
  3149. finish: "error",
  3150. error: { type: "unknown", message: "Provider unavailable" },
  3151. content: [{ type: "tool", id: "call-hosted-raw-failure", state: { status: "error" } }],
  3152. },
  3153. ])
  3154. }),
  3155. )
  3156. it.effect("keeps interleaved assistant text blocks separate", () =>
  3157. Effect.gen(function* () {
  3158. yield* setup
  3159. const session = yield* SessionV2.Service
  3160. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Two blocks" }), resume: false })
  3161. responses = undefined
  3162. streamGate = undefined
  3163. streamStarted = undefined
  3164. response = [
  3165. LLMEvent.stepStart({ index: 0 }),
  3166. LLMEvent.textStart({ id: "text-1" }),
  3167. LLMEvent.textStart({ id: "text-2" }),
  3168. LLMEvent.textDelta({ id: "text-1", text: "First" }),
  3169. LLMEvent.textDelta({ id: "text-2", text: "Second" }),
  3170. LLMEvent.textEnd({ id: "text-1" }),
  3171. LLMEvent.textEnd({ id: "text-2" }),
  3172. LLMEvent.stepFinish({ index: 0, reason: "stop" }),
  3173. LLMEvent.finish({ reason: "stop" }),
  3174. ]
  3175. yield* session.resume(sessionID)
  3176. expect(yield* session.context(sessionID)).toMatchObject([
  3177. { type: "user", text: "Two blocks" },
  3178. {
  3179. type: "assistant",
  3180. content: [
  3181. { type: "text", id: "text-1", text: "First" },
  3182. { type: "text", id: "text-2", text: "Second" },
  3183. ],
  3184. },
  3185. ])
  3186. }),
  3187. )
  3188. for (const kind of fragmentKinds) {
  3189. it.effect(`broadcasts provider ${kind} deltas without storing projection rewrites`, () =>
  3190. verifyEphemeralDeltas(kind),
  3191. )
  3192. it.effect(`durably closes partial ${kind} when the provider stream fails`, () => verifyPartialFlushOnFailure(kind))
  3193. it.effect(`durably closes partial ${kind} when the provider stream is interrupted`, () =>
  3194. verifyPartialFlushOnInterruption(kind),
  3195. )
  3196. }
  3197. it.effect("rejects duplicate streamed text starts", () =>
  3198. Effect.gen(function* () {
  3199. yield* setup
  3200. const session = yield* SessionV2.Service
  3201. responses = undefined
  3202. streamGate = undefined
  3203. streamStarted = undefined
  3204. response = [LLMEvent.textStart({ id: "text-1" }), LLMEvent.textStart({ id: "text-1" })]
  3205. expect(yield* session.resume(sessionID).pipe(Effect.catchDefect(Effect.succeed))).toBe(
  3206. "Duplicate text start: text-1",
  3207. )
  3208. }),
  3209. )
  3210. it.effect("transitions streamed raw tool input to parsed called input", () =>
  3211. Effect.gen(function* () {
  3212. yield* setup
  3213. const session = yield* SessionV2.Service
  3214. yield* session.prompt({ sessionID, prompt: new Prompt({ text: "Call provider tool" }), resume: false })
  3215. responses = undefined
  3216. streamGate = undefined
  3217. streamStarted = undefined
  3218. response = [
  3219. LLMEvent.stepStart({ index: 0 }),
  3220. LLMEvent.toolInputStart({ id: "call-parsed", name: "web_search" }),
  3221. LLMEvent.toolInputDelta({ id: "call-parsed", name: "web_search", text: '{"query":"hello"}' }),
  3222. LLMEvent.toolInputEnd({ id: "call-parsed", name: "web_search" }),
  3223. LLMEvent.toolCall({ id: "call-parsed", name: "web_search", input: { query: "hello" }, providerExecuted: true }),
  3224. ]
  3225. yield* session.resume(sessionID)
  3226. expect(yield* session.context(sessionID)).toMatchObject([
  3227. { type: "user", text: "Call provider tool" },
  3228. {
  3229. type: "assistant",
  3230. content: [{ type: "tool", id: "call-parsed", state: { status: "error", input: { query: "hello" } } }],
  3231. },
  3232. ])
  3233. }),
  3234. )
  3235. it.effect("rejects malformed streamed tool input ordering", () =>
  3236. Effect.gen(function* () {
  3237. yield* setup
  3238. const session = yield* SessionV2.Service
  3239. responses = undefined
  3240. streamGate = undefined
  3241. streamStarted = undefined
  3242. response = [LLMEvent.toolInputDelta({ id: "call-1", name: "read", text: "{}" })]
  3243. expect(yield* session.resume(sessionID).pipe(Effect.catchDefect(Effect.succeed))).toBe(
  3244. "Tool input delta before start: call-1",
  3245. )
  3246. }),
  3247. )
  3248. })