tool-subagent.test.ts 16 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385
  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 { EventV2 } from "@opencode-ai/core/event"
  10. import { Location } from "@opencode-ai/core/location"
  11. import { ModelV2 } from "@opencode-ai/core/model"
  12. import { ProviderV2 } from "@opencode-ai/core/provider"
  13. import { AbsolutePath } from "@opencode-ai/core/schema"
  14. import { AgentV2 } 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 { SessionV2 } 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/subagent"
  27. import { ToolRegistry } from "@opencode-ai/core/tool/registry"
  28. import { ToolOutputStore } from "@opencode-ai/core/tool-output-store"
  29. import { tmpdir } from "./fixture/tmpdir"
  30. import { testEffect } from "./lib/effect"
  31. import { executeTool, settleTool, toolIdentity, waitForTool } from "./lib/tool"
  32. const childText = "child final response"
  33. const childModel = ModelV2.Ref.make({ id: ModelV2.ID.make("child"), providerID: ProviderV2.ID.make("test") })
  34. const parentModel = ModelV2.Ref.make({ id: ModelV2.ID.make("parent"), providerID: ProviderV2.ID.make("test") })
  35. const tokens = { input: 0, output: 0, reasoning: 0, cache: { read: 0, write: 0 } }
  36. const outputSessionID = (value: unknown) => Schema.decodeUnknownSync(SubagentTool.Output)(value).sessionID
  37. const executionNode = makeGlobalNode({
  38. service: SessionExecution.Service,
  39. layer: Layer.effect(
  40. SessionExecution.Service,
  41. Effect.gen(function* () {
  42. const events = yield* EventV2.Service
  43. const store = yield* SessionStore.Service
  44. const completed = new Set<SessionV2.ID>()
  45. const complete = Effect.fn("SubagentTest.complete")(function* (sessionID: SessionV2.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* events.publish(SessionEvent.Step.Started, {
  54. sessionID,
  55. assistantMessageID,
  56. agent: AgentV2.ID.make("reviewer"),
  57. model: childModel,
  58. })
  59. yield* events.publish(SessionEvent.Text.Started, {
  60. sessionID,
  61. assistantMessageID,
  62. ordinal: 0,
  63. })
  64. yield* events.publish(SessionEvent.Text.Ended, {
  65. sessionID,
  66. assistantMessageID,
  67. ordinal: 0,
  68. text: childText,
  69. })
  70. yield* events.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: [EventV2.node, SessionStore.node],
  88. })
  89. const layer = AppNodeBuilder.build(
  90. LayerNode.group([
  91. Database.node,
  92. EventV2.node,
  93. Job.node,
  94. ToolOutputStore.cleanupNode,
  95. SessionV2.node,
  96. SessionExecution.node,
  97. PluginRuntime.providerNode,
  98. LocationServiceMap.node,
  99. ]),
  100. [[SessionExecution.node, executionNode]],
  101. )
  102. const it = testEffect(layer)
  103. const withSubagent = (location: Location.Ref) =>
  104. Effect.gen(function* () {
  105. const locations = yield* LocationServiceMap.Service
  106. yield* PluginSupervisor.Service.use((supervisor) => supervisor.flush).pipe(Effect.provide(locations.get(location)))
  107. yield* AgentV2.Service.use((agents) =>
  108. agents.transform((draft) => {
  109. // The caller identity used by executeTool; subagent permission asserts against it.
  110. draft.update(toolIdentity.agent, (agent) => {
  111. agent.mode = "primary"
  112. agent.permissions.push({ action: "*", resource: "*", effect: "allow" })
  113. })
  114. draft.update(AgentV2.ID.make("reviewer"), (agent) => {
  115. agent.mode = "subagent"
  116. agent.model = childModel
  117. })
  118. draft.update(AgentV2.ID.make("fallback"), (agent) => {
  119. agent.mode = "subagent"
  120. })
  121. draft.update(AgentV2.ID.make("primary"), (agent) => {
  122. agent.mode = "primary"
  123. })
  124. }),
  125. ).pipe(Effect.provide(locations.get(location)))
  126. })
  127. describe("SubagentTool", () => {
  128. it.live("registers globally while resolving agents from the caller location", () =>
  129. Effect.acquireRelease(
  130. Effect.promise(() => tmpdir()),
  131. (dir) => Effect.promise(() => dir[Symbol.asyncDispose]()),
  132. ).pipe(
  133. Effect.flatMap((dir) =>
  134. Effect.gen(function* () {
  135. const location = Location.Ref.make({ directory: AbsolutePath.make(dir.path) })
  136. const session = yield* SessionV2.Service
  137. const parent = yield* session.create({ location })
  138. yield* withSubagent(parent.location)
  139. const locations = yield* LocationServiceMap.Service
  140. const registry = yield* ToolRegistry.Service.pipe(Effect.provide(locations.get(parent.location)))
  141. yield* waitForTool(registry, SubagentTool.name)
  142. expect((yield* registry.materialize()).definitions.map((tool) => tool.name)).toContain(SubagentTool.name)
  143. expect(
  144. yield* executeTool(registry, {
  145. sessionID: parent.id,
  146. ...toolIdentity,
  147. call: {
  148. type: "tool-call",
  149. id: "call-primary",
  150. name: SubagentTool.name,
  151. input: { agent: "primary", description: "primary", prompt: "should fail" },
  152. },
  153. }),
  154. ).toEqual({ type: "error", value: "Agent primary cannot run as a subagent" })
  155. }),
  156. ),
  157. ),
  158. )
  159. it.live("prevents subagents from launching subagents by default", () =>
  160. Effect.acquireRelease(
  161. Effect.promise(() => tmpdir()),
  162. (dir) => Effect.promise(() => dir[Symbol.asyncDispose]()),
  163. ).pipe(
  164. Effect.flatMap((dir) =>
  165. Effect.gen(function* () {
  166. const location = Location.Ref.make({ directory: AbsolutePath.make(dir.path) })
  167. const sessions = yield* SessionV2.Service
  168. const root = yield* sessions.create({ location })
  169. const parent = yield* sessions.create({ parentID: root.id, title: "parent" })
  170. yield* withSubagent(parent.location)
  171. const locations = yield* LocationServiceMap.Service
  172. const registry = yield* ToolRegistry.Service.pipe(Effect.provide(locations.get(parent.location)))
  173. yield* waitForTool(registry, SubagentTool.name)
  174. expect(
  175. yield* executeTool(registry, {
  176. sessionID: parent.id,
  177. ...toolIdentity,
  178. call: {
  179. type: "tool-call",
  180. id: "call-nested-subagent",
  181. name: SubagentTool.name,
  182. input: { agent: "reviewer", description: "nested", prompt: "should fail" },
  183. },
  184. }),
  185. ).toEqual({ type: "error", value: expect.stringContaining("Subagent depth limit reached (1)") })
  186. expect((yield* sessions.list({ parentID: parent.id })).data).toHaveLength(0)
  187. }),
  188. ),
  189. ),
  190. )
  191. it.live("allows nested subagents up to the configured depth", () =>
  192. Effect.acquireRelease(
  193. Effect.promise(() => tmpdir()),
  194. (dir) => Effect.promise(() => dir[Symbol.asyncDispose]()),
  195. ).pipe(
  196. Effect.flatMap((dir) =>
  197. Effect.gen(function* () {
  198. yield* Effect.promise(() =>
  199. Bun.write(path.join(dir.path, "opencode.json"), JSON.stringify({ experimental: { subagent_depth: 2 } })),
  200. )
  201. const location = Location.Ref.make({ directory: AbsolutePath.make(dir.path) })
  202. const sessions = yield* SessionV2.Service
  203. const root = yield* sessions.create({ location })
  204. const parent = yield* sessions.create({ parentID: root.id, title: "parent", model: parentModel })
  205. yield* withSubagent(parent.location)
  206. const locations = yield* LocationServiceMap.Service
  207. const registry = yield* ToolRegistry.Service.pipe(Effect.provide(locations.get(parent.location)))
  208. yield* waitForTool(registry, SubagentTool.name)
  209. const settled = yield* settleTool(registry, {
  210. sessionID: parent.id,
  211. ...toolIdentity,
  212. call: {
  213. type: "tool-call",
  214. id: "call-configured-nested-subagent",
  215. name: SubagentTool.name,
  216. input: { agent: "reviewer", description: "nested", prompt: "should run" },
  217. },
  218. })
  219. expect(settled.output?.structured).toMatchObject({ status: "completed", output: childText })
  220. expect((yield* sessions.get(outputSessionID(settled.output?.structured))).parentID).toBe(parent.id)
  221. }),
  222. ),
  223. ),
  224. )
  225. it.live("runs a foreground child session and returns the final assistant text", () =>
  226. Effect.acquireRelease(
  227. Effect.promise(() => tmpdir()),
  228. (dir) => Effect.promise(() => dir[Symbol.asyncDispose]()),
  229. ).pipe(
  230. Effect.flatMap((dir) =>
  231. Effect.gen(function* () {
  232. const location = Location.Ref.make({ directory: AbsolutePath.make(dir.path) })
  233. const sessions = yield* SessionV2.Service
  234. const parent = yield* sessions.create({ location, model: parentModel })
  235. yield* withSubagent(parent.location)
  236. const locations = yield* LocationServiceMap.Service
  237. const registry = yield* ToolRegistry.Service.pipe(Effect.provide(locations.get(parent.location)))
  238. yield* waitForTool(registry, SubagentTool.name)
  239. const progress: ToolRegistry.Progress[] = []
  240. const settled = yield* settleTool(registry, {
  241. sessionID: parent.id,
  242. ...toolIdentity,
  243. progress: (update) => Effect.sync(() => progress.push(update)),
  244. call: {
  245. type: "tool-call",
  246. id: "call-subagent",
  247. name: SubagentTool.name,
  248. input: { agent: "reviewer", description: "review", prompt: "review this" },
  249. },
  250. })
  251. expect(settled.output?.structured).toMatchObject({ status: "completed", output: childText })
  252. const child = yield* sessions.get(outputSessionID(settled.output?.structured))
  253. expect(progress[0]?.structured).toEqual({ sessionID: child.id, status: "running" })
  254. expect(child).toMatchObject({
  255. parentID: parent.id,
  256. location: parent.location,
  257. agent: "reviewer",
  258. model: childModel,
  259. })
  260. const fallback = yield* settleTool(registry, {
  261. sessionID: parent.id,
  262. ...toolIdentity,
  263. call: {
  264. type: "tool-call",
  265. id: "call-subagent-fallback",
  266. name: SubagentTool.name,
  267. input: { agent: "fallback", description: "fallback", prompt: "fallback" },
  268. },
  269. })
  270. const fallbackChild = yield* sessions.get(outputSessionID(fallback.output?.structured))
  271. expect(fallbackChild).toMatchObject({ parentID: parent.id, model: parentModel })
  272. }),
  273. ),
  274. ),
  275. )
  276. it.live("returns child runner failures as tool errors", () =>
  277. Effect.acquireRelease(
  278. Effect.promise(() => tmpdir()),
  279. (dir) => Effect.promise(() => dir[Symbol.asyncDispose]()),
  280. ).pipe(
  281. Effect.flatMap((dir) =>
  282. Effect.gen(function* () {
  283. const location = Location.Ref.make({ directory: AbsolutePath.make(dir.path) })
  284. const sessions = yield* SessionV2.Service
  285. const parent = yield* sessions.create({ location })
  286. yield* withSubagent(parent.location)
  287. const locations = yield* LocationServiceMap.Service
  288. const registry = yield* ToolRegistry.Service.pipe(Effect.provide(locations.get(parent.location)))
  289. yield* waitForTool(registry, SubagentTool.name)
  290. expect(
  291. yield* executeTool(registry, {
  292. sessionID: parent.id,
  293. ...toolIdentity,
  294. call: {
  295. type: "tool-call",
  296. id: "call-subagent-failure",
  297. name: SubagentTool.name,
  298. input: { agent: "reviewer", description: "fail review", prompt: "please fail" },
  299. },
  300. }),
  301. ).toEqual({ type: "error", value: expect.stringContaining("No model is available for session") })
  302. }),
  303. ),
  304. ),
  305. )
  306. it.live("notifies once when background work completes", () =>
  307. Effect.acquireRelease(
  308. Effect.promise(() => tmpdir()),
  309. (dir) => Effect.promise(() => dir[Symbol.asyncDispose]()),
  310. ).pipe(
  311. Effect.flatMap((dir) =>
  312. Effect.gen(function* () {
  313. const location = Location.Ref.make({ directory: AbsolutePath.make(dir.path) })
  314. const sessions = yield* SessionV2.Service
  315. const parent = yield* sessions.create({ location })
  316. yield* withSubagent(parent.location)
  317. const locations = yield* LocationServiceMap.Service
  318. const registry = yield* ToolRegistry.Service.pipe(Effect.provide(locations.get(parent.location)))
  319. yield* waitForTool(registry, SubagentTool.name)
  320. const events = yield* EventV2.Service
  321. const admitted = yield* events.subscribe(SessionEvent.InputAdmitted).pipe(
  322. Stream.filter((event) => event.data.sessionID === parent.id && event.data.input.type === "synthetic"),
  323. Stream.take(1),
  324. Stream.runCollect,
  325. Effect.forkScoped({ startImmediately: true }),
  326. )
  327. const settled = yield* settleTool(registry, {
  328. sessionID: parent.id,
  329. ...toolIdentity,
  330. call: {
  331. type: "tool-call",
  332. id: "call-background-subagent",
  333. name: SubagentTool.name,
  334. input: { agent: "reviewer", description: "background review", prompt: "review", background: true },
  335. },
  336. })
  337. const childID = outputSessionID(settled.output?.structured)
  338. expect(settled.output?.structured).toMatchObject({
  339. status: "running",
  340. output: expect.stringContaining(`id: ${childID}`),
  341. })
  342. const admission = Array.from(yield* Fiber.join(admitted))[0]
  343. expect(admission?.data.input.data.text).toContain(`<subagent id="${childID}" state="completed"`)
  344. expect(admission?.data.input.data).toMatchObject({
  345. description: "background review",
  346. metadata: {
  347. source: "subagent",
  348. childID,
  349. agent: "reviewer",
  350. state: "completed",
  351. },
  352. })
  353. const database = yield* Database.Service
  354. yield* SessionPending.promoteSteers(database.db, events, parent.id)
  355. const synthetic = (yield* sessions.context(parent.id)).filter((message) => message.type === "synthetic")
  356. expect(synthetic).toHaveLength(1)
  357. expect(synthetic[0]?.text).toContain(`<subagent id="${childID}" state="completed"`)
  358. expect(synthetic[0]?.text).toContain(childText)
  359. }),
  360. ),
  361. ),
  362. )
  363. })