tool-subagent.test.ts 17 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416
  1. import { describe, expect } from "bun:test"
  2. import { DateTime, Effect, Fiber, Layer, Schema, Stream } from "effect"
  3. import path from "path"
  4. import { Money } from "@opencode-ai/schema/money"
  5. import { AppNodeBuilder } from "@opencode-ai/core/effect/app-node-builder"
  6. import { LayerNode } from "@opencode-ai/util/effect/layer-node"
  7. import { makeGlobalNode } from "@opencode-ai/util/effect/app-node"
  8. import { Database } from "@opencode-ai/core/database/database"
  9. import { Bus } from "@opencode-ai/core/bus"
  10. import { Location } from "@opencode-ai/core/location"
  11. import { Model } from "@opencode-ai/core/model"
  12. import { Provider } from "@opencode-ai/core/provider"
  13. import { AbsolutePath } from "@opencode-ai/core/schema"
  14. import { Agent } from "@opencode-ai/core/agent"
  15. import { Job } from "@opencode-ai/core/job"
  16. import { LocationServiceMap } from "@opencode-ai/core/location-service-map"
  17. import { Session } from "@opencode-ai/core/session"
  18. import { SessionEvent } from "@opencode-ai/core/session/event"
  19. import { SessionExecution } from "@opencode-ai/core/session/execution"
  20. import { SessionPending } from "@opencode-ai/core/session/pending"
  21. import { SessionMessage } from "@opencode-ai/core/session/message"
  22. import { SessionRunnerModel } from "@opencode-ai/core/session/runner/model"
  23. import { SessionStore } from "@opencode-ai/core/session/store"
  24. import { PluginRuntime } from "@opencode-ai/core/plugin/runtime"
  25. import { PluginSupervisor } from "@opencode-ai/core/plugin/supervisor"
  26. import { SubagentTool } from "@opencode-ai/core/tool/plugin/subagent"
  27. import { Tool } from "@opencode-ai/core/tool"
  28. import { tmpdir } from "./fixture/tmpdir"
  29. import { testEffect } from "./lib/effect"
  30. import { executeTool, toolIdentity, waitForTool } from "./lib/tool"
  31. const childText = "child final response"
  32. const childModel = Model.Ref.make({ id: Model.ID.make("child"), providerID: Provider.ID.make("test") })
  33. const parentModel = Model.Ref.make({ id: Model.ID.make("parent"), providerID: Provider.ID.make("test") })
  34. const tokens = { input: 0, output: 0, reasoning: 0, cache: { read: 0, write: 0 } }
  35. const outputSessionID = (value: unknown) =>
  36. Schema.decodeUnknownSync(Schema.Struct({ sessionID: Session.ID }))(value).sessionID
  37. const executionNode = makeGlobalNode({
  38. service: SessionExecution.Service,
  39. layer: Layer.effect(
  40. SessionExecution.Service,
  41. Effect.gen(function* () {
  42. const bus = yield* Bus.Service
  43. const store = yield* SessionStore.Service
  44. const completed = new Set<Session.ID>()
  45. const complete = Effect.fn("SubagentTest.complete")(function* (sessionID: Session.ID) {
  46. if (completed.has(sessionID)) return
  47. if ((yield* store.get(sessionID))?.title.includes("fail")) {
  48. yield* new SessionRunnerModel.ModelNotSelectedError({ sessionID })
  49. return
  50. }
  51. completed.add(sessionID)
  52. const assistantMessageID = SessionMessage.ID.create()
  53. yield* bus.publish(SessionEvent.Step.Started, {
  54. sessionID,
  55. assistantMessageID,
  56. agent: Agent.ID.make("reviewer"),
  57. model: childModel,
  58. })
  59. yield* bus.publish(SessionEvent.Text.Started, {
  60. sessionID,
  61. assistantMessageID,
  62. ordinal: 0,
  63. })
  64. yield* bus.publish(SessionEvent.Text.Ended, {
  65. sessionID,
  66. assistantMessageID,
  67. ordinal: 0,
  68. text: childText,
  69. })
  70. yield* bus.publish(SessionEvent.Step.Ended, {
  71. sessionID,
  72. assistantMessageID,
  73. finish: "stop",
  74. cost: Money.USD.zero,
  75. tokens,
  76. })
  77. })
  78. return SessionExecution.Service.of({
  79. active: Effect.succeed(new Set()),
  80. resume: complete,
  81. wake: () => Effect.void,
  82. interrupt: () => Effect.void,
  83. awaitIdle: (sessionID) => complete(sessionID).pipe(Effect.exit, Effect.asVoid),
  84. })
  85. }),
  86. ),
  87. deps: [Bus.node, SessionStore.node],
  88. })
  89. const layer = AppNodeBuilder.build(
  90. LayerNode.group([
  91. Database.node,
  92. Bus.node,
  93. Job.node,
  94. Session.node,
  95. SessionExecution.node,
  96. PluginRuntime.providerNode,
  97. LocationServiceMap.node,
  98. ]),
  99. [[SessionExecution.node, executionNode]],
  100. )
  101. const it = testEffect(layer)
  102. const withSubagent = (location: Location.Ref) =>
  103. Effect.gen(function* () {
  104. const locations = yield* LocationServiceMap.Service
  105. yield* PluginSupervisor.Service.use((supervisor) => supervisor.flush).pipe(Effect.provide(locations.get(location)))
  106. yield* Agent.Service.use((agents) =>
  107. agents.transform((draft) => {
  108. // The caller identity used by executeTool; subagent permission asserts against it.
  109. draft.update(toolIdentity.agent, (agent) => {
  110. agent.mode = "primary"
  111. agent.permissions.push({ action: "*", resource: "*", effect: "allow" })
  112. })
  113. draft.update(Agent.ID.make("reviewer"), (agent) => {
  114. agent.mode = "subagent"
  115. agent.model = childModel
  116. })
  117. draft.update(Agent.ID.make("fallback"), (agent) => {
  118. agent.mode = "subagent"
  119. })
  120. draft.update(Agent.ID.make("primary"), (agent) => {
  121. agent.mode = "primary"
  122. })
  123. }),
  124. ).pipe(Effect.provide(locations.get(location)))
  125. })
  126. describe("SubagentTool", () => {
  127. it.live("registers globally while resolving agents from the caller location", () =>
  128. Effect.acquireRelease(
  129. Effect.promise(() => tmpdir()),
  130. (dir) => Effect.promise(() => dir[Symbol.asyncDispose]()),
  131. ).pipe(
  132. Effect.flatMap((dir) =>
  133. Effect.gen(function* () {
  134. const location = Location.Ref.make({ directory: AbsolutePath.make(dir.path) })
  135. const session = yield* Session.Service
  136. const parent = yield* session.create({ location })
  137. yield* withSubagent(parent.location)
  138. const locations = yield* LocationServiceMap.Service
  139. const registry = yield* Tool.Service.pipe(Effect.provide(locations.get(parent.location)))
  140. yield* waitForTool(registry, SubagentTool.name)
  141. expect((yield* registry.snapshot()).definitions.map((tool) => tool.name)).toContain(SubagentTool.name)
  142. expect(
  143. yield* executeTool(registry, {
  144. sessionID: parent.id,
  145. ...toolIdentity,
  146. call: {
  147. type: "tool-call",
  148. id: "call-primary",
  149. name: SubagentTool.name,
  150. input: { agent: "primary", description: "primary", prompt: "should fail" },
  151. },
  152. }),
  153. ).toEqual({
  154. status: "error",
  155. error: { type: "tool.execution", message: "Agent primary cannot run as a subagent" },
  156. })
  157. }),
  158. ),
  159. ),
  160. )
  161. it.live("prevents subagents from launching subagents by default", () =>
  162. Effect.acquireRelease(
  163. Effect.promise(() => tmpdir()),
  164. (dir) => Effect.promise(() => dir[Symbol.asyncDispose]()),
  165. ).pipe(
  166. Effect.flatMap((dir) =>
  167. Effect.gen(function* () {
  168. const location = Location.Ref.make({ directory: AbsolutePath.make(dir.path) })
  169. const sessions = yield* Session.Service
  170. const root = yield* sessions.create({ location })
  171. const parent = yield* sessions.create({ parentID: root.id, title: "parent" })
  172. yield* withSubagent(parent.location)
  173. const locations = yield* LocationServiceMap.Service
  174. const registry = yield* Tool.Service.pipe(Effect.provide(locations.get(parent.location)))
  175. yield* waitForTool(registry, SubagentTool.name)
  176. expect(
  177. yield* executeTool(registry, {
  178. sessionID: parent.id,
  179. ...toolIdentity,
  180. call: {
  181. type: "tool-call",
  182. id: "call-nested-subagent",
  183. name: SubagentTool.name,
  184. input: { agent: "reviewer", description: "nested", prompt: "should fail" },
  185. },
  186. }),
  187. ).toEqual({
  188. status: "error",
  189. error: {
  190. type: "tool.execution",
  191. message: expect.stringContaining("Subagent depth limit reached (1)"),
  192. },
  193. })
  194. expect((yield* sessions.list({ parentID: parent.id })).data).toHaveLength(0)
  195. }),
  196. ),
  197. ),
  198. )
  199. it.live("allows nested subagents up to the configured depth", () =>
  200. Effect.acquireRelease(
  201. Effect.promise(() => tmpdir()),
  202. (dir) => Effect.promise(() => dir[Symbol.asyncDispose]()),
  203. ).pipe(
  204. Effect.flatMap((dir) =>
  205. Effect.gen(function* () {
  206. yield* Effect.promise(() =>
  207. Bun.write(path.join(dir.path, "opencode.json"), JSON.stringify({ experimental: { subagent_depth: 2 } })),
  208. )
  209. const location = Location.Ref.make({ directory: AbsolutePath.make(dir.path) })
  210. const sessions = yield* Session.Service
  211. const root = yield* sessions.create({ location })
  212. const parent = yield* sessions.create({ parentID: root.id, title: "parent", model: parentModel })
  213. yield* withSubagent(parent.location)
  214. const locations = yield* LocationServiceMap.Service
  215. const registry = yield* Tool.Service.pipe(Effect.provide(locations.get(parent.location)))
  216. yield* waitForTool(registry, SubagentTool.name)
  217. const settled = yield* executeTool(registry, {
  218. sessionID: parent.id,
  219. ...toolIdentity,
  220. call: {
  221. type: "tool-call",
  222. id: "call-configured-nested-subagent",
  223. name: SubagentTool.name,
  224. input: { agent: "reviewer", description: "nested", prompt: "should run" },
  225. },
  226. })
  227. expect(settled).toMatchObject({
  228. status: "completed",
  229. metadata: { status: "completed" },
  230. content: [{ type: "text", text: childText }],
  231. })
  232. expect(settled.metadata).toEqual({
  233. sessionID: outputSessionID(settled.metadata),
  234. status: "completed",
  235. })
  236. expect((yield* sessions.get(outputSessionID(settled.metadata))).parentID).toBe(parent.id)
  237. }),
  238. ),
  239. ),
  240. )
  241. it.live("runs a foreground child session and returns the final assistant text", () =>
  242. Effect.acquireRelease(
  243. Effect.promise(() => tmpdir()),
  244. (dir) => Effect.promise(() => dir[Symbol.asyncDispose]()),
  245. ).pipe(
  246. Effect.flatMap((dir) =>
  247. Effect.gen(function* () {
  248. const location = Location.Ref.make({ directory: AbsolutePath.make(dir.path) })
  249. const sessions = yield* Session.Service
  250. const parent = yield* sessions.create({ location, model: parentModel })
  251. yield* withSubagent(parent.location)
  252. const locations = yield* LocationServiceMap.Service
  253. const registry = yield* Tool.Service.pipe(Effect.provide(locations.get(parent.location)))
  254. yield* waitForTool(registry, SubagentTool.name)
  255. const progress: Tool.Metadata[] = []
  256. const settled = yield* executeTool(registry, {
  257. sessionID: parent.id,
  258. ...toolIdentity,
  259. progress: (update) => Effect.sync(() => progress.push(update)),
  260. call: {
  261. type: "tool-call",
  262. id: "call-subagent",
  263. name: SubagentTool.name,
  264. input: { agent: "reviewer", description: "review", prompt: "review this" },
  265. },
  266. })
  267. expect(settled).toMatchObject({
  268. status: "completed",
  269. metadata: { status: "completed" },
  270. content: [{ type: "text", text: childText }],
  271. })
  272. const child = yield* sessions.get(outputSessionID(settled.metadata))
  273. expect(settled.metadata).toEqual({ sessionID: child.id, status: "completed" })
  274. expect(progress[0]?.metadata).toEqual({ sessionID: child.id, status: "running" })
  275. expect(child).toMatchObject({
  276. parentID: parent.id,
  277. location: parent.location,
  278. agent: "reviewer",
  279. model: childModel,
  280. })
  281. expect((yield* sessions.pending(child.id)).find((message) => message.type === "user")?.data.text).toBe(
  282. "You are a subagent spawned by another session.\nreview this",
  283. )
  284. const fallback = yield* executeTool(registry, {
  285. sessionID: parent.id,
  286. ...toolIdentity,
  287. call: {
  288. type: "tool-call",
  289. id: "call-subagent-fallback",
  290. name: SubagentTool.name,
  291. input: { agent: "fallback", description: "fallback", prompt: "fallback" },
  292. },
  293. })
  294. const fallbackChild = yield* sessions.get(outputSessionID(fallback.metadata))
  295. expect(fallbackChild).toMatchObject({ parentID: parent.id, model: parentModel })
  296. }),
  297. ),
  298. ),
  299. )
  300. it.live("returns child runner failures as tool errors", () =>
  301. Effect.acquireRelease(
  302. Effect.promise(() => tmpdir()),
  303. (dir) => Effect.promise(() => dir[Symbol.asyncDispose]()),
  304. ).pipe(
  305. Effect.flatMap((dir) =>
  306. Effect.gen(function* () {
  307. const location = Location.Ref.make({ directory: AbsolutePath.make(dir.path) })
  308. const sessions = yield* Session.Service
  309. const parent = yield* sessions.create({ location })
  310. yield* withSubagent(parent.location)
  311. const locations = yield* LocationServiceMap.Service
  312. const registry = yield* Tool.Service.pipe(Effect.provide(locations.get(parent.location)))
  313. yield* waitForTool(registry, SubagentTool.name)
  314. expect(
  315. yield* executeTool(registry, {
  316. sessionID: parent.id,
  317. ...toolIdentity,
  318. call: {
  319. type: "tool-call",
  320. id: "call-subagent-failure",
  321. name: SubagentTool.name,
  322. input: { agent: "reviewer", description: "fail review", prompt: "please fail" },
  323. },
  324. }),
  325. ).toEqual({
  326. status: "error",
  327. error: {
  328. type: "tool.execution",
  329. message: expect.stringContaining("No model is available for session"),
  330. },
  331. })
  332. }),
  333. ),
  334. ),
  335. )
  336. it.live("notifies once when background work completes", () =>
  337. Effect.acquireRelease(
  338. Effect.promise(() => tmpdir()),
  339. (dir) => Effect.promise(() => dir[Symbol.asyncDispose]()),
  340. ).pipe(
  341. Effect.flatMap((dir) =>
  342. Effect.gen(function* () {
  343. const location = Location.Ref.make({ directory: AbsolutePath.make(dir.path) })
  344. const sessions = yield* Session.Service
  345. const parent = yield* sessions.create({ location })
  346. yield* withSubagent(parent.location)
  347. const locations = yield* LocationServiceMap.Service
  348. const registry = yield* Tool.Service.pipe(Effect.provide(locations.get(parent.location)))
  349. yield* waitForTool(registry, SubagentTool.name)
  350. const bus = yield* Bus.Service
  351. const admitted = yield* bus.subscribe(SessionEvent.InputAdmitted).pipe(
  352. Stream.filter((event) => event.data.sessionID === parent.id && event.data.input.type === "synthetic"),
  353. Stream.take(1),
  354. Stream.runCollect,
  355. Effect.forkScoped({ startImmediately: true }),
  356. )
  357. const settled = yield* executeTool(registry, {
  358. sessionID: parent.id,
  359. ...toolIdentity,
  360. call: {
  361. type: "tool-call",
  362. id: "call-background-subagent",
  363. name: SubagentTool.name,
  364. input: { agent: "reviewer", description: "background review", prompt: "review", background: true },
  365. },
  366. })
  367. const childID = outputSessionID(settled.metadata)
  368. expect(settled.metadata).toMatchObject({
  369. status: "running",
  370. })
  371. expect(settled.metadata).toEqual({ sessionID: childID, status: "running" })
  372. expect(settled.content).toEqual([{ type: "text", text: expect.stringContaining(`id: ${childID}`) }])
  373. const admission = Array.from(yield* Fiber.join(admitted))[0]
  374. expect(admission?.data.input.data.text).toContain(`<subagent id="${childID}" state="completed"`)
  375. expect(admission?.data.input.data).toMatchObject({
  376. description: "background review",
  377. metadata: {
  378. source: "subagent",
  379. childID,
  380. agent: "reviewer",
  381. state: "completed",
  382. },
  383. })
  384. const database = yield* Database.Service
  385. yield* SessionPending.promote(database.db, bus, parent.id, "steer")
  386. const synthetic = (yield* sessions.context(parent.id)).filter((message) => message.type === "synthetic")
  387. expect(synthetic).toHaveLength(1)
  388. expect(synthetic[0]?.text).toContain(`<subagent id="${childID}" state="completed"`)
  389. expect(synthetic[0]?.text).toContain(childText)
  390. }),
  391. ),
  392. ),
  393. )
  394. })