feat: add experimental background subagents (#27084)

packages/opencode/src/tool/registry.ts

@@ -7,6 +7,7 @@ import { GlobTool } from "./glob"
 import { GrepTool } from "./grep"
 import { ReadTool } from "./read"
 import { TaskTool } from "./task"
+import { TaskStatusTool } from "./task_status"
 import { TodoWriteTool } from "./todo"
 import { WebFetchTool } from "./webfetch"
 import { WriteTool } from "./write"
@@ -49,6 +50,8 @@ import { Git } from "@/git"
 import { Skill } from "../skill"
 import { Permission } from "@/permission"
 import { Reference } from "@/reference/reference"
+import { BackgroundJob } from "@/background/job"
+import { SessionStatus } from "@/session/status"
 import { RuntimeFlags } from "@/effect/runtime-flags"
 
 const log = Log.create({ service: "tool.registry" })
@@ -86,6 +89,8 @@ export const layer: Layer.Layer<
   | Agent.Service
   | Skill.Service
   | Session.Service
+  | SessionStatus.Service
+  | BackgroundJob.Service
   | Provider.Service
   | Git.Service
   | Reference.Service
@@ -111,6 +116,7 @@ export const layer: Layer.Layer<
 
     const invalid = yield* InvalidTool
     const task = yield* TaskTool
+    const taskStatus = yield* TaskStatusTool
     const read = yield* ReadTool
     const question = yield* QuestionTool
     const todo = yield* TodoWriteTool
@@ -219,6 +225,7 @@ export const layer: Layer.Layer<
           edit: Tool.init(edit),
           write: Tool.init(writetool),
           task: Tool.init(task),
+          task_status: Tool.init(taskStatus),
           fetch: Tool.init(webfetch),
           todo: Tool.init(todo),
           search: Tool.init(websearch),
@@ -243,6 +250,7 @@ export const layer: Layer.Layer<
             tool.edit,
             tool.write,
             tool.task,
+            ...(flags.experimentalBackgroundSubagents ? [tool.task_status] : []),
             tool.fetch,
             tool.todo,
             tool.search,
@@ -358,28 +366,30 @@ export const layer: Layer.Layer<
 )
 
 export const defaultLayer = Layer.suspend(() =>
-  layer.pipe(
-    Layer.provide(Config.defaultLayer),
-    Layer.provide(Plugin.defaultLayer),
-    Layer.provide(Question.defaultLayer),
-    Layer.provide(Todo.defaultLayer),
-    Layer.provide(Skill.defaultLayer),
-    Layer.provide(Agent.defaultLayer),
-    Layer.provide(Session.defaultLayer),
-    Layer.provide(Provider.defaultLayer),
-    Layer.provide(Git.defaultLayer),
-    Layer.provide(Reference.defaultLayer),
-    Layer.provide(LSP.defaultLayer),
-    Layer.provide(Instruction.defaultLayer),
-    Layer.provide(AppFileSystem.defaultLayer),
-    Layer.provide(Bus.layer),
-    Layer.provide(FetchHttpClient.layer),
-    Layer.provide(Format.defaultLayer),
-    Layer.provide(CrossSpawnSpawner.defaultLayer),
-    Layer.provide(Ripgrep.defaultLayer),
-    Layer.provide(Truncate.defaultLayer),
-    Layer.provide(RuntimeFlags.defaultLayer),
-  ),
+  layer
+    .pipe(
+      Layer.provide(Config.defaultLayer),
+      Layer.provide(Plugin.defaultLayer),
+      Layer.provide(Question.defaultLayer),
+      Layer.provide(Todo.defaultLayer),
+      Layer.provide(Skill.defaultLayer),
+      Layer.provide(Agent.defaultLayer),
+      Layer.provide(Session.defaultLayer),
+      Layer.provide(Layer.mergeAll(SessionStatus.defaultLayer, BackgroundJob.defaultLayer)),
+      Layer.provide(Provider.defaultLayer),
+      Layer.provide(Git.defaultLayer),
+      Layer.provide(Reference.defaultLayer),
+      Layer.provide(LSP.defaultLayer),
+      Layer.provide(Instruction.defaultLayer),
+      Layer.provide(AppFileSystem.defaultLayer),
+      Layer.provide(Bus.layer),
+      Layer.provide(FetchHttpClient.layer),
+      Layer.provide(Format.defaultLayer),
+      Layer.provide(CrossSpawnSpawner.defaultLayer),
+      Layer.provide(Ripgrep.defaultLayer),
+      Layer.provide(Truncate.defaultLayer),
+    )
+    .pipe(Layer.provide(RuntimeFlags.defaultLayer)),
 )
 
 function isZodType(value: unknown): value is z.ZodType {

packages/opencode/src/tool/task.ts

@@ -1,24 +1,31 @@
 import * as Tool from "./tool"
 import DESCRIPTION from "./task.txt"
+import { ToolJsonSchema } from "./json-schema"
+import { BackgroundJob } from "@/background/job"
+import { Bus } from "@/bus"
 import { Session } from "@/session/session"
 import { SessionID, MessageID } from "../session/schema"
 import { MessageV2 } from "../session/message-v2"
 import { Agent } from "../agent/agent"
 import { deriveSubagentSessionPermission } from "../agent/subagent-permissions"
 import type { SessionPrompt } from "../session/prompt"
+import { SessionStatus } from "@/session/status"
 import { Config } from "@/config/config"
-import { Effect, Exit, Schema } from "effect"
+import { TuiEvent } from "@/cli/cmd/tui/event"
+import { Cause, Effect, Exit, Option, Schema, Scope } from "effect"
 import { EffectBridge } from "@/effect/bridge"
+import { RuntimeFlags } from "@/effect/runtime-flags"
 
 export interface TaskPromptOps {
   cancel(sessionID: SessionID): Effect.Effect<void>
   resolvePromptParts(template: string): Effect.Effect<SessionPrompt.PromptInput["parts"]>
   prompt(input: SessionPrompt.PromptInput): Effect.Effect<MessageV2.WithParts>
+  loop(input: SessionPrompt.LoopInput): Effect.Effect<MessageV2.WithParts>
 }
 
 const id = "task"
 
-export const Parameters = Schema.Struct({
+const BaseParameters = Schema.Struct({
   description: Schema.String.annotate({ description: "A short (3-5 words) description of the task" }),
   prompt: Schema.String.annotate({ description: "The task for the agent to perform" }),
   subagent_type: Schema.String.annotate({ description: "The type of specialized agent to use for this task" }),
@@ -29,18 +36,78 @@ export const Parameters = Schema.Struct({
   command: Schema.optional(Schema.String).annotate({ description: "The command that triggered this task" }),
 })
 
+export const Parameters = Schema.Struct({
+  description: Schema.String.annotate({ description: "A short (3-5 words) description of the task" }),
+  prompt: Schema.String.annotate({ description: "The task for the agent to perform" }),
+  subagent_type: Schema.String.annotate({ description: "The type of specialized agent to use for this task" }),
+  task_id: Schema.optional(Schema.String).annotate({
+    description:
+      "This should only be set if you mean to resume a previous task (you can pass a prior task_id and the task will continue the same subagent session as before instead of creating a fresh one)",
+  }),
+  command: Schema.optional(Schema.String).annotate({ description: "The command that triggered this task" }),
+  background: Schema.optional(Schema.Boolean).annotate({
+    description: "When true, launch the subagent in the background and return immediately",
+  }),
+})
+
+function output(sessionID: SessionID, text: string) {
+  return [
+    `task_id: ${sessionID} (for resuming to continue this task if needed)`,
+    "",
+    "<task_result>",
+    text,
+    "</task_result>",
+  ].join("\n")
+}
+
+function backgroundOutput(sessionID: SessionID) {
+  return [
+    `task_id: ${sessionID} (for polling this task with task_status)`,
+    "state: running",
+    "",
+    "<task_result>",
+    "Background task started. Continue your current work and call task_status when you need the result.",
+    "</task_result>",
+  ].join("\n")
+}
+
+function backgroundMessage(input: { sessionID: SessionID; description: string; state: "completed" | "error"; text: string }) {
+  const tag = input.state === "completed" ? "task_result" : "task_error"
+  const title =
+    input.state === "completed"
+      ? `Background task completed: ${input.description}`
+      : `Background task failed: ${input.description}`
+  return [title, `task_id: ${input.sessionID}`, `state: ${input.state}`, "", `<${tag}>`, input.text, `</${tag}>`].join(
+    "\n",
+  )
+}
+
+function errorText(error: unknown) {
+  if (error instanceof Error) return error.message
+  return String(error)
+}
+
 export const TaskTool = Tool.define(
   id,
   Effect.gen(function* () {
     const agent = yield* Agent.Service
+    const background = yield* BackgroundJob.Service
+    const bus = yield* Bus.Service
     const config = yield* Config.Service
     const sessions = yield* Session.Service
+    const scope = yield* Scope.Scope
+    const status = yield* SessionStatus.Service
+    const flags = yield* RuntimeFlags.Service
 
     const run = Effect.fn("TaskTool.execute")(function* (
       params: Schema.Schema.Type<typeof Parameters>,
       ctx: Tool.Context,
     ) {
       const cfg = yield* config.get()
+      const runInBackground = params.background === true
+      if (runInBackground && !flags.experimentalBackgroundSubagents) {
+        return yield* Effect.fail(new Error("Background subagents require OPENCODE_EXPERIMENTAL_BACKGROUND_SUBAGENTS=true"))
+      }
 
       if (!ctx.extra?.bypassAgentCheck) {
         yield* ctx.ask({
@@ -93,20 +160,125 @@ export const TaskTool = Tool.define(
         modelID: msg.info.modelID,
         providerID: msg.info.providerID,
       }
+      const metadata = {
+        parentSessionId: ctx.sessionID,
+        sessionId: nextSession.id,
+        model,
+        ...(runInBackground ? { background: true } : {}),
+      }
 
       yield* ctx.metadata({
         title: params.description,
-        metadata: {
-          sessionId: nextSession.id,
-          model,
-        },
+        metadata,
       })
 
       const ops = ctx.extra?.promptOps as TaskPromptOps
       if (!ops) return yield* Effect.fail(new Error("TaskTool requires promptOps in ctx.extra"))
       const runCancel = yield* EffectBridge.make()
 
-      const messageID = MessageID.ascending()
+      const runTask = Effect.fn("TaskTool.runTask")(function* () {
+        const parts = yield* ops.resolvePromptParts(params.prompt)
+        const result = yield* ops.prompt({
+          messageID: MessageID.ascending(),
+          sessionID: nextSession.id,
+          model: {
+            modelID: model.modelID,
+            providerID: model.providerID,
+          },
+          agent: next.name,
+          tools: {
+            ...(next.permission.some((rule) => rule.permission === "todowrite") ? {} : { todowrite: false }),
+            ...(next.permission.some((rule) => rule.permission === id) ? {} : { task: false }),
+            ...Object.fromEntries((cfg.experimental?.primary_tools ?? []).map((item) => [item, false])),
+          },
+          parts,
+        })
+        return result.parts.findLast((item) => item.type === "text")?.text ?? ""
+      })
+
+      const resumeWhenIdle: (input: { userID: MessageID; state: "completed" | "error" }) => Effect.Effect<void> = Effect.fn(
+        "TaskTool.resumeWhenIdle",
+      )(function* (input: { userID: MessageID; state: "completed" | "error" }) {
+        const latest = yield* sessions.findMessage(ctx.sessionID, (item) => item.info.role === "user").pipe(Effect.orDie)
+        if (Option.isNone(latest)) return
+        if (latest.value.info.id !== input.userID) return
+        if ((yield* status.get(ctx.sessionID)).type !== "idle") {
+          yield* Effect.sleep("300 millis")
+          return yield* resumeWhenIdle(input)
+        }
+        yield* bus.publish(TuiEvent.ToastShow, {
+          title: input.state === "completed" ? "Background task complete" : "Background task failed",
+          message:
+            input.state === "completed"
+              ? `Background task "${params.description}" finished. Resuming the main thread.`
+              : `Background task "${params.description}" failed. Resuming the main thread.`,
+          variant: input.state === "completed" ? "success" : "error",
+          duration: 5000,
+        })
+        yield* ops.loop({ sessionID: ctx.sessionID }).pipe(Effect.ignore, Effect.forkIn(scope, { startImmediately: true }))
+      })
+
+      const continueIfIdle = Effect.fn("TaskTool.continueIfIdle")(function* (input: {
+        userID: MessageID
+        state: "completed" | "error"
+      }) {
+        yield* resumeWhenIdle(input).pipe(Effect.ignore, Effect.forkIn(scope, { startImmediately: true }))
+      })
+
+      const inject = Effect.fn("TaskTool.injectBackgroundResult")(function* (state: "completed" | "error", text: string) {
+        const currentParent = yield* sessions.get(ctx.sessionID)
+        const message = yield* ops.prompt({
+          sessionID: ctx.sessionID,
+          noReply: true,
+          agent: currentParent.agent ?? ctx.agent,
+          parts: [
+            {
+              type: "text",
+              synthetic: true,
+              text: backgroundMessage({
+                sessionID: nextSession.id,
+                description: params.description,
+                state,
+                text,
+              }),
+            },
+          ],
+        })
+        yield* continueIfIdle({ userID: message.info.id, state })
+      })
+
+      const existing = yield* background.get(nextSession.id)
+      if (existing?.status === "running") {
+        return yield* Effect.fail(new Error(`Task ${nextSession.id} is already running. Use task_status to check progress.`))
+      }
+
+      if (runInBackground) {
+        const info = yield* background.start({
+          id: nextSession.id,
+          type: id,
+          title: params.description,
+          metadata,
+          run: runTask().pipe(
+            Effect.tap((text) => inject("completed", text).pipe(Effect.ignore)),
+            Effect.catchCause((cause) =>
+              (Cause.hasInterruptsOnly(cause)
+                ? Effect.void
+                : inject("error", errorText(Cause.squash(cause))).pipe(Effect.ignore)
+              ).pipe(Effect.andThen(Effect.failCause(cause))),
+            ),
+          ),
+        })
+
+        return {
+          title: params.description,
+          metadata: {
+            ...metadata,
+            jobId: info.id,
+          },
+          output: backgroundOutput(nextSession.id),
+        }
+      }
+
       const cancel = ops.cancel(nextSession.id)
 
       function onAbort() {
@@ -119,36 +291,11 @@ export const TaskTool = Tool.define(
         }),
         () =>
           Effect.gen(function* () {
-            const parts = yield* ops.resolvePromptParts(params.prompt)
-            const result = yield* ops.prompt({
-              messageID,
-              sessionID: nextSession.id,
-              model: {
-                modelID: model.modelID,
-                providerID: model.providerID,
-              },
-              agent: next.name,
-              tools: {
-                ...(next.permission.some((rule) => rule.permission === "todowrite") ? {} : { todowrite: false }),
-                ...(next.permission.some((rule) => rule.permission === id) ? {} : { task: false }),
-                ...Object.fromEntries((cfg.experimental?.primary_tools ?? []).map((item) => [item, false])),
-              },
-              parts,
-            })
-
+            const text = yield* runTask()
             return {
               title: params.description,
-              metadata: {
-                sessionId: nextSession.id,
-                model,
-              },
-              output: [
-                `task_id: ${nextSession.id} (for resuming to continue this task if needed)`,
-                "",
-                "<task_result>",
-                result.parts.findLast((item) => item.type === "text")?.text ?? "",
-                "</task_result>",
-              ].join("\n"),
+              metadata,
+              output: output(nextSession.id, text),
             }
           }),
         (_, exit) =>
@@ -167,6 +314,7 @@ export const TaskTool = Tool.define(
     return {
       description: DESCRIPTION,
       parameters: Parameters,
+      jsonSchema: flags.experimentalBackgroundSubagents ? undefined : ToolJsonSchema.fromSchema(BaseParameters),
       execute: (params: Schema.Schema.Type<typeof Parameters>, ctx: Tool.Context) =>
         run(params, ctx).pipe(Effect.orDie),
     }

packages/opencode/src/tool/task.txt

@@ -15,10 +15,11 @@ When NOT to use the Task tool:
 Usage notes:
 1. Launch multiple agents concurrently whenever possible, to maximize performance; to do that, use a single message with multiple tool uses
 2. When the agent is done, it will return a single message back to you. The result returned by the agent is not visible to the user. To show the user the result, you should send a text message back to the user with a concise summary of the result. The output includes a task_id you can reuse later to continue the same subagent session.
-3. Each agent invocation starts with a fresh context unless you provide task_id to resume the same subagent session (which continues with its previous messages and tool outputs). When starting fresh, your prompt should contain a highly detailed task description for the agent to perform autonomously and you should specify exactly what information the agent should return back to you in its final and only message to you.
-4. The agent's outputs should generally be trusted
-5. Clearly tell the agent whether you expect it to write code or just to do research (search, file reads, web fetches, etc.), since it is not aware of the user's intent. Tell it how to verify its work if possible (e.g., relevant test commands).
-6. If the agent description mentions that it should be used proactively, then you should try your best to use it without the user having to ask for it first. Use your judgement.
+3. If OPENCODE_EXPERIMENTAL_BACKGROUND_SUBAGENTS is enabled, background=true launches the subagent asynchronously. Use task_status(task_id=..., wait=false) to poll, or wait=true to block until done.
+4. Each agent invocation starts with a fresh context unless you provide task_id to resume the same subagent session (which continues with its previous messages and tool outputs). When starting fresh, your prompt should contain a highly detailed task description for the agent to perform autonomously and you should specify exactly what information the agent should return back to you in its final and only message to you.
+5. The agent's outputs should generally be trusted
+6. Clearly tell the agent whether you expect it to write code or just to do research (search, file reads, web fetches, etc.), since it is not aware of the user's intent. Tell it how to verify its work if possible (e.g., relevant test commands).
+7. If the agent description mentions that it should be used proactively, then you should try your best to use it without the user having to ask for it first. Use your judgement.
 
 Example usage (NOTE: The agents below are fictional examples for illustration only - use the actual agents listed above):
 

packages/opencode/src/tool/task_status.ts

@@ -0,0 +1,167 @@
+import * as Tool from "./tool"
+import DESCRIPTION from "./task_status.txt"
+import { BackgroundJob } from "@/background/job"
+import { Session } from "@/session/session"
+import { MessageV2 } from "@/session/message-v2"
+import { SessionID } from "@/session/schema"
+import { SessionStatus } from "@/session/status"
+import { PositiveInt } from "@opencode-ai/core/schema"
+import { RuntimeFlags } from "@/effect/runtime-flags"
+import { Effect, Option, Schema } from "effect"
+
+const DEFAULT_TIMEOUT = 60_000
+const POLL_MS = 300
+
+const Parameters = Schema.Struct({
+  task_id: SessionID.annotate({ description: "The task_id returned by the task tool" }),
+  wait: Schema.optional(Schema.Boolean).annotate({ description: "When true, wait until the task reaches a terminal state or timeout" }),
+  timeout_ms: Schema.optional(PositiveInt).annotate({
+    description: "Maximum milliseconds to wait when wait=true (default: 60000)",
+  }),
+})
+
+type State = BackgroundJob.Status
+type InspectResult = { state: State; text: string }
+
+function format(input: { taskID: SessionID; state: State; text: string }) {
+  const tag = input.state === "completed" || input.state === "running" ? "task_result" : "task_error"
+  return [`task_id: ${input.taskID}`, `state: ${input.state}`, "", `<${tag}>`, input.text, `</${tag}>`].join("\n")
+}
+
+function errorText(error: NonNullable<MessageV2.Assistant["error"]>) {
+  const data = Reflect.get(error, "data")
+  const message = data && typeof data === "object" ? Reflect.get(data, "message") : undefined
+  if (typeof message === "string" && message) return message
+  return error.name
+}
+
+function inspectMessage(message: MessageV2.WithParts): InspectResult | undefined {
+  if (message.info.role !== "assistant") return
+  const text = message.parts.findLast((part) => part.type === "text")?.text ?? ""
+  if (message.info.error) return { state: "error", text: text || errorText(message.info.error) }
+  if (message.info.finish && !["tool-calls", "unknown"].includes(message.info.finish)) return { state: "completed", text }
+  return { state: "running", text: text || "Task is still running." }
+}
+
+export const TaskStatusTool = Tool.define(
+  "task_status",
+  Effect.gen(function* () {
+    const jobs = yield* BackgroundJob.Service
+    const sessions = yield* Session.Service
+    const status = yield* SessionStatus.Service
+    const flags = yield* RuntimeFlags.Service
+
+    const inspect: (taskID: SessionID) => Effect.Effect<InspectResult> = Effect.fn("TaskStatusTool.inspect")(function* (taskID: SessionID) {
+      const job = yield* jobs.get(taskID)
+      if (job) {
+        return {
+          state: job.status,
+          text:
+            job.output ??
+            job.error ??
+            (job.status === "running"
+              ? "Task is still running."
+              : job.status === "cancelled"
+                ? "Task was cancelled."
+                : ""),
+        }
+      }
+
+      const current = yield* status.get(taskID)
+      if (current.type === "busy" || current.type === "retry") {
+        return {
+          state: "running",
+          text: current.type === "retry" ? `Task is retrying: ${current.message}` : "Task is still running.",
+        }
+      }
+
+      const latestAssistant = yield* sessions.findMessage(taskID, (item) => item.info.role === "assistant").pipe(Effect.orDie)
+      if (Option.isSome(latestAssistant)) {
+        const latest = inspectMessage(latestAssistant.value)
+        if (!latest) return { state: "error", text: "Task is not running in this process." }
+        if (latest.state === "running") return { state: "error", text: "Task is not running in this process and has no final output." }
+        return latest
+      }
+      return { state: "error", text: "Task is not running in this process and has not produced output." }
+    })
+
+    const waitForTerminal: (taskID: SessionID, timeout: number) => Effect.Effect<{ result: InspectResult; timedOut: boolean }> = Effect.fn(
+      "TaskStatusTool.waitForTerminal",
+    )(function* (taskID: SessionID, timeout: number) {
+      const result = yield* inspect(taskID)
+      if (result.state !== "running") return { result, timedOut: false }
+      if (timeout <= 0) return { result, timedOut: true }
+      const sleep = Math.min(POLL_MS, timeout)
+      yield* Effect.sleep(`${sleep} millis`)
+      return yield* waitForTerminal(taskID, timeout - sleep)
+    })
+
+    const run = Effect.fn("TaskStatusTool.execute")(function* (
+      params: Schema.Schema.Type<typeof Parameters>,
+      _ctx: Tool.Context,
+    ) {
+      if (!flags.experimentalBackgroundSubagents) {
+        return yield* Effect.fail(new Error("task_status requires OPENCODE_EXPERIMENTAL_BACKGROUND_SUBAGENTS=true"))
+      }
+
+      const session = yield* sessions.get(params.task_id).pipe(Effect.catchCause(() => Effect.succeed(undefined)))
+      if (!session) {
+        return {
+          title: "Task status",
+          metadata: {
+            task_id: params.task_id,
+            state: "error" as const,
+            timed_out: false,
+          },
+          output: format({
+            taskID: params.task_id,
+            state: "error",
+            text: `Task not found: ${params.task_id}`,
+          }),
+        }
+      }
+
+      const waited =
+        params.wait === true
+          ? yield* jobs.wait({ id: params.task_id, timeout: params.timeout_ms ?? DEFAULT_TIMEOUT })
+          : { info: yield* jobs.get(params.task_id), timedOut: false }
+      const inspected = waited.info
+        ? {
+            result: {
+              state: waited.info.status,
+              text:
+                waited.info.output ??
+                waited.info.error ??
+                (waited.info.status === "running" ? "Task is still running." : ""),
+            },
+            timedOut: waited.timedOut,
+          }
+        : params.wait === true
+          ? yield* waitForTerminal(params.task_id, params.timeout_ms ?? DEFAULT_TIMEOUT)
+          : { result: yield* inspect(params.task_id), timedOut: false }
+      const text = inspected.timedOut
+        ? `Timed out after ${params.timeout_ms ?? DEFAULT_TIMEOUT}ms while waiting for task completion.`
+        : inspected.result.text
+
+      return {
+        title: "Task status",
+        metadata: {
+          task_id: params.task_id,
+          state: inspected.result.state,
+          timed_out: inspected.timedOut,
+        },
+        output: format({
+          taskID: params.task_id,
+          state: inspected.result.state,
+          text,
+        }),
+      }
+    })
+
+    return {
+      description: DESCRIPTION,
+      parameters: Parameters,
+      execute: (params: Schema.Schema.Type<typeof Parameters>, ctx: Tool.Context) => run(params, ctx).pipe(Effect.orDie),
+    }
+  }),
+)

packages/opencode/src/tool/task_status.txt

@@ -0,0 +1,13 @@
+Poll the status of a background subagent task launched with the task tool.
+
+Use this for tasks started with `task(background=true)`.
+
+Parameters:
+- `task_id` (required): the task session id returned by the task tool
+- `wait` (optional): when true, wait for completion
+- `timeout_ms` (optional): max wait duration in milliseconds when `wait=true`
+
+Returns compact, parseable output:
+- `task_id`
+- `state` (`running`, `completed`, `error`, or `cancelled`)
+- `<task_result>...</task_result>` or `<task_error>...</task_error>` containing final output, error summary, or current progress text