feat(atlas): add background task session tracking with retry scheduling

- Add background-launch-session-tracking.ts to persist delegated sessions
- Add task-context.ts for task context resolution utilities
- Modify idle-event.ts to schedule retries when background tasks are running
- Update tool-execute-after.ts to integrate session tracking
- Add comprehensive tests for background task retry behavior

🤖 Generated with assistance of OhMyOpenCode

src/hooks/atlas/background-launch-session-tracking.ts

@@ -0,0 +1,63 @@
+import type { PluginInput } from "@opencode-ai/plugin"
+import { appendSessionId, type BoulderState, upsertTaskSessionState } from "../../features/boulder-state"
+import { log } from "../../shared/logger"
+import { HOOK_NAME } from "./hook-name"
+import { extractSessionIdFromOutput, validateSubagentSessionId } from "./subagent-session-id"
+import { resolveTaskContext } from "./task-context"
+import type { PendingTaskRef, ToolExecuteAfterInput, ToolExecuteAfterOutput } from "./types"
+
+export async function syncBackgroundLaunchSessionTracking(input: {
+  ctx: PluginInput
+  boulderState: BoulderState | null
+  toolInput: ToolExecuteAfterInput
+  toolOutput: ToolExecuteAfterOutput
+  pendingTaskRef: PendingTaskRef | undefined
+  metadataSessionId?: string
+}): Promise<void> {
+  const { ctx, boulderState, toolInput, toolOutput, pendingTaskRef, metadataSessionId } = input
+  if (!boulderState) {
+    return
+  }
+
+  if (toolInput.sessionID && !boulderState.session_ids.includes(toolInput.sessionID)) {
+    appendSessionId(ctx.directory, toolInput.sessionID)
+  }
+
+  const extractedSessionId = metadataSessionId ?? extractSessionIdFromOutput(toolOutput.output)
+  const lineageSessionIDs = toolInput.sessionID && !boulderState.session_ids.includes(toolInput.sessionID)
+    ? [...boulderState.session_ids, toolInput.sessionID]
+    : boulderState.session_ids
+  const subagentSessionId = await validateSubagentSessionId({
+    client: ctx.client,
+    sessionID: extractedSessionId,
+    lineageSessionIDs,
+  })
+
+  if (!subagentSessionId) {
+    return
+  }
+
+  appendSessionId(ctx.directory, subagentSessionId)
+
+  const { currentTask, shouldSkipTaskSessionUpdate } = resolveTaskContext(
+    pendingTaskRef,
+    boulderState.active_plan,
+  )
+
+  if (currentTask && !shouldSkipTaskSessionUpdate) {
+    upsertTaskSessionState(ctx.directory, {
+      taskKey: currentTask.key,
+      taskLabel: currentTask.label,
+      taskTitle: currentTask.title,
+      sessionId: subagentSessionId,
+      agent: typeof toolOutput.metadata?.agent === "string" ? toolOutput.metadata.agent : undefined,
+      category: typeof toolOutput.metadata?.category === "string" ? toolOutput.metadata.category : undefined,
+    })
+  }
+
+  log(`[${HOOK_NAME}] Background launch session tracked`, {
+    sessionID: toolInput.sessionID,
+    subagentSessionId,
+    taskKey: currentTask?.key,
+  })
+}

src/hooks/atlas/background-task-retry.test.ts

@@ -0,0 +1,225 @@
+import { afterEach, beforeEach, describe, expect, mock, test } from "bun:test"
+import { existsSync, mkdirSync, rmSync, writeFileSync } from "node:fs"
+import { tmpdir } from "node:os"
+import { join } from "node:path"
+import { randomUUID } from "node:crypto"
+import type { PluginInput } from "@opencode-ai/plugin"
+import { createAtlasHook } from "./atlas-hook"
+import { clearBoulderState, writeBoulderState } from "../../features/boulder-state"
+import { _resetForTesting, registerAgentName } from "../../features/claude-code-session-state"
+
+type LongTimerCallback = (...args: unknown[]) => void | Promise<void>
+
+describe("atlas background task retry", () => {
+  let testDir: string
+  const sessionID = "main-session-123"
+  const capturedTimers = new Map<number, { callback: () => Promise<void> | void; cleared: boolean }>()
+  let nextFakeTimerId = 1000
+  const originalSetTimeout = globalThis.setTimeout
+  const originalClearTimeout = globalThis.clearTimeout
+
+  async function flushMicrotasks(): Promise<void> {
+    await Promise.resolve()
+    await Promise.resolve()
+  }
+
+  async function firePendingTimers(): Promise<void> {
+    const entries = [...capturedTimers.entries()]
+    for (const [id, entry] of entries) {
+      if (entry.cleared) {
+        continue
+      }
+
+      capturedTimers.delete(id)
+      await entry.callback()
+    }
+    await flushMicrotasks()
+  }
+
+  beforeEach(() => {
+    _resetForTesting()
+    registerAgentName("atlas")
+    registerAgentName("sisyphus")
+
+    testDir = join(tmpdir(), `atlas-background-retry-${randomUUID()}`)
+    mkdirSync(testDir, { recursive: true })
+
+    capturedTimers.clear()
+    nextFakeTimerId = 1000
+
+    globalThis.setTimeout = ((callback: Parameters<typeof setTimeout>[0], delay?: number, ...args: unknown[]) => {
+      const normalizedDelay = typeof delay === "number" ? delay : 0
+      if (typeof callback !== "function") {
+        return originalSetTimeout(callback, delay, ...args)
+      }
+
+      if (normalizedDelay >= 5000) {
+        const id = nextFakeTimerId++
+        capturedTimers.set(id, {
+          callback: () => (callback as LongTimerCallback)(...args),
+          cleared: false,
+        })
+        return id as unknown as ReturnType<typeof setTimeout>
+      }
+
+      return originalSetTimeout(callback, delay, ...args)
+    }) as typeof setTimeout
+
+    globalThis.clearTimeout = ((id?: number | ReturnType<typeof setTimeout>) => {
+      if (typeof id === "number" && capturedTimers.has(id)) {
+        capturedTimers.get(id)!.cleared = true
+        capturedTimers.delete(id)
+        return
+      }
+
+      originalClearTimeout(id as Parameters<typeof clearTimeout>[0])
+    }) as typeof clearTimeout
+  })
+
+  afterEach(() => {
+    globalThis.setTimeout = originalSetTimeout
+    globalThis.clearTimeout = originalClearTimeout
+    _resetForTesting()
+    clearBoulderState(testDir)
+    if (existsSync(testDir)) {
+      rmSync(testDir, { recursive: true, force: true })
+    }
+  })
+
+  test("#given background tasks are still running #when retry fires before they finish #then atlas keeps retrying until continuation can resume", async () => {
+    // given
+    const planPath = join(testDir, "test-plan.md")
+    writeFileSync(planPath, "# Plan\n- [ ] Task 1\n- [ ] Task 2")
+    writeBoulderState(testDir, {
+      active_plan: planPath,
+      started_at: "2026-01-02T10:00:00Z",
+      session_ids: [sessionID],
+      plan_name: "test-plan",
+      agent: "atlas",
+    })
+
+    let backgroundRunning = true
+    const promptMock = mock(async () => ({}))
+    const hook = createAtlasHook({
+      directory: testDir,
+      client: {
+        session: {
+          promptAsync: promptMock,
+          messages: async () => ({ data: [] }),
+        },
+      },
+    } as unknown as PluginInput, {
+      directory: testDir,
+      backgroundManager: {
+        getTasksByParentSession: () => backgroundRunning ? [{ status: "running" }] : [],
+      } as unknown as NonNullable<Parameters<typeof createAtlasHook>[1]>["backgroundManager"] & {
+        getTasksByParentSession: (sessionID: string) => Array<{ status: string }>
+      },
+    })
+
+    // when
+    await hook.handler({ event: { type: "session.idle", properties: { sessionID } } })
+    await firePendingTimers()
+    backgroundRunning = false
+    await firePendingTimers()
+
+    // then
+    expect(promptMock).toHaveBeenCalledTimes(1)
+  })
+
+  test("#given multiple idle events arrive while background retry is already pending #when tasks are still running #then atlas keeps only one retry timer active", async () => {
+    // given
+    const planPath = join(testDir, "test-plan.md")
+    writeFileSync(planPath, "# Plan\n- [ ] Task 1\n- [ ] Task 2")
+    writeBoulderState(testDir, {
+      active_plan: planPath,
+      started_at: "2026-01-02T10:00:00Z",
+      session_ids: [sessionID],
+      plan_name: "test-plan",
+      agent: "atlas",
+    })
+
+    let backgroundRunning = true
+    const promptMock = mock(async () => ({}))
+    const hook = createAtlasHook({
+      directory: testDir,
+      client: {
+        session: {
+          promptAsync: promptMock,
+          messages: async () => ({ data: [] }),
+        },
+      },
+    } as unknown as PluginInput, {
+      directory: testDir,
+      backgroundManager: {
+        getTasksByParentSession: () => backgroundRunning ? [{ status: "running" }] : [],
+      } as unknown as NonNullable<Parameters<typeof createAtlasHook>[1]>["backgroundManager"] & {
+        getTasksByParentSession: (sessionID: string) => Array<{ status: string }>
+      },
+    })
+
+    // when
+    await hook.handler({ event: { type: "session.idle", properties: { sessionID } } })
+    await hook.handler({ event: { type: "session.idle", properties: { sessionID } } })
+    await hook.handler({ event: { type: "session.idle", properties: { sessionID } } })
+
+    // then
+    expect(capturedTimers.size).toBe(1)
+    backgroundRunning = false
+    await firePendingTimers()
+    expect(promptMock).toHaveBeenCalledTimes(1)
+  })
+
+  test("#given background tasks keep running across multiple retries #when they finally finish on a later retry #then atlas resumes exactly once", async () => {
+    // given
+    const planPath = join(testDir, "test-plan.md")
+    writeFileSync(planPath, "# Plan\n- [ ] Task 1\n- [ ] Task 2")
+    writeBoulderState(testDir, {
+      active_plan: planPath,
+      started_at: "2026-01-02T10:00:00Z",
+      session_ids: [sessionID],
+      plan_name: "test-plan",
+      agent: "atlas",
+    })
+
+    let remainingRunningRetries = 2
+    const promptMock = mock(async () => ({}))
+    const hook = createAtlasHook({
+      directory: testDir,
+      client: {
+        session: {
+          promptAsync: promptMock,
+          messages: async () => ({ data: [] }),
+        },
+      },
+    } as unknown as PluginInput, {
+      directory: testDir,
+      backgroundManager: {
+        getTasksByParentSession: () => {
+          if (remainingRunningRetries > 0) {
+            remainingRunningRetries -= 1
+            return [{ status: "running" }]
+          }
+
+          return []
+        },
+      } as unknown as NonNullable<Parameters<typeof createAtlasHook>[1]>["backgroundManager"] & {
+        getTasksByParentSession: (sessionID: string) => Array<{ status: string }>
+      },
+    })
+
+    // when
+    await hook.handler({ event: { type: "session.idle", properties: { sessionID } } })
+    expect(capturedTimers.size).toBe(1)
+
+    await firePendingTimers()
+    expect(promptMock).toHaveBeenCalledTimes(0)
+    expect(capturedTimers.size).toBe(1)
+
+    await firePendingTimers()
+
+    // then
+    expect(promptMock).toHaveBeenCalledTimes(1)
+    expect(capturedTimers.size).toBe(0)
+  })
+})

src/hooks/atlas/idle-event.ts

@@ -90,7 +90,10 @@ function scheduleRetry(input: {
     const currentProgress = getPlanProgress(currentBoulder.active_plan)
     if (currentProgress.isComplete) return
     if (options?.isContinuationStopped?.(sessionID)) return
-    if (hasRunningBackgroundTasks(sessionID, options)) return
+    if (hasRunningBackgroundTasks(sessionID, options)) {
+      scheduleRetry({ ctx, sessionID, sessionState, options })
+      return
+    }
 
     await injectContinuation({
       ctx,
@@ -194,6 +197,7 @@ export async function handleAtlasSessionIdle(input: {
   }
 
   if (hasRunningBackgroundTasks(sessionID, options)) {
+    scheduleRetry({ ctx, sessionID, sessionState, options })
     log(`[${HOOK_NAME}] Skipped: background tasks running`, { sessionID })
     return
   }

src/hooks/atlas/task-context.ts

@@ -0,0 +1,45 @@
+import { readCurrentTopLevelTask } from "../../features/boulder-state"
+import type { PendingTaskRef, TrackedTopLevelTaskRef } from "./types"
+
+export function resolvePreferredSessionId(currentSessionId?: string, trackedSessionId?: string): string {
+  return currentSessionId ?? trackedSessionId ?? "<session_id>"
+}
+
+export function resolveTaskContext(
+  pendingTaskRef: PendingTaskRef | undefined,
+  planPath: string,
+): {
+  currentTask: TrackedTopLevelTaskRef | null
+  shouldSkipTaskSessionUpdate: boolean
+  shouldIgnoreCurrentSessionId: boolean
+} {
+  if (!pendingTaskRef) {
+    return {
+      currentTask: readCurrentTopLevelTask(planPath),
+      shouldSkipTaskSessionUpdate: false,
+      shouldIgnoreCurrentSessionId: false,
+    }
+  }
+
+  if (pendingTaskRef.kind === "track") {
+    return {
+      currentTask: pendingTaskRef.task,
+      shouldSkipTaskSessionUpdate: false,
+      shouldIgnoreCurrentSessionId: false,
+    }
+  }
+
+  if (pendingTaskRef.reason === "explicit_resume") {
+    return {
+      currentTask: readCurrentTopLevelTask(planPath),
+      shouldSkipTaskSessionUpdate: true,
+      shouldIgnoreCurrentSessionId: true,
+    }
+  }
+
+  return {
+    currentTask: pendingTaskRef.task,
+    shouldSkipTaskSessionUpdate: true,
+    shouldIgnoreCurrentSessionId: true,
+  }
+}

src/hooks/atlas/tool-execute-after-background-launch.test.ts

@@ -1,11 +1,13 @@
 /// <reference types="bun-types" />
 
-import { afterEach, beforeEach, describe, expect, it, mock, afterAll } from "bun:test"
-import { existsSync, mkdirSync, rmSync } from "node:fs"
+import { afterEach, beforeEach, describe, expect, it, mock, afterAll, spyOn } from "bun:test"
+import { existsSync, mkdirSync, rmSync, writeFileSync } from "node:fs"
 import { tmpdir } from "node:os"
 import { join } from "node:path"
 import type { PluginInput } from "@opencode-ai/plugin"
 import { createOpencodeClient, type Project } from "@opencode-ai/sdk"
+import { readBoulderState, writeBoulderState } from "../../features/boulder-state"
+import { createToolExecuteBeforeHandler } from "./tool-execute-before"
 
 const isCallerOrchestratorMock = mock(async () => true)
 const collectGitDiffStatsMock = mock(() => ({
@@ -27,6 +29,9 @@ afterAll(() => { mock.restore() })
 
 const { createToolExecuteAfterHandler } = await import("./tool-execute-after")
 
+type OpencodeClient = ReturnType<typeof createOpencodeClient>
+type SessionGetResult = Awaited<ReturnType<OpencodeClient["session"]["get"]>>
+
 describe("createToolExecuteAfterHandler background launch detection", () => {
   let testDirectory = ""
 
@@ -47,17 +52,39 @@ describe("createToolExecuteAfterHandler background launch detection", () => {
     }
   })
 
-  function createHandler() {
-    const project = {
+  function createProject(): Project {
+    return {
       id: "project-1",
       worktree: testDirectory,
       time: {
         created: Date.now(),
       },
-    } satisfies Project
+    }
+  }
+
+  function createSessionGetResult(parentID: string | undefined): SessionGetResult {
+    return {
+      data: {
+        parentID,
+      },
+      error: undefined,
+      request: new Request("https://example.com/session"),
+      response: new Response(null, { status: 200 }),
+    } as SessionGetResult
+  }
+
+  function createHandler(parentSessionIDs?: Record<string, string | undefined>) {
+    const project = createProject()
+    const client = createOpencodeClient()
+
+    if (parentSessionIDs) {
+      spyOn(client.session, "get").mockImplementation((input) => Promise.resolve(
+        createSessionGetResult(parentSessionIDs[input.path.id]),
+      ) as never)
+    }
 
     const ctx = {
-      client: createOpencodeClient(),
+      client,
       project,
       directory: testDirectory,
       worktree: testDirectory,
@@ -98,5 +125,76 @@ describe("createToolExecuteAfterHandler background launch detection", () => {
         expect(collectGitDiffStatsMock).not.toHaveBeenCalled()
       })
     })
+
+    describe("#when a background task launch belongs to the active boulder task", () => {
+      it("#then it should persist the delegated session without transforming the launch output", async () => {
+        const sessionID = "ses_parent"
+        const childSessionID = "ses_child123"
+        const planPath = join(testDirectory, "background-launch-plan.md")
+        const project = createProject()
+        const client = createOpencodeClient()
+
+        spyOn(client.session, "get").mockImplementation((input) => Promise.resolve(
+          createSessionGetResult(input.path.id === childSessionID ? sessionID : undefined),
+        ) as never)
+
+        writeFileSync(planPath, `# Plan
+
+## TODOs
+- [ ] 1. Implement auth flow
+`)
+
+        writeBoulderState(testDirectory, {
+          active_plan: planPath,
+          started_at: "2026-01-02T10:00:00Z",
+          session_ids: [sessionID],
+          plan_name: "background-launch-plan",
+        })
+
+        const pendingFilePaths = new Map<string, string>()
+        const pendingTaskRefs = new Map()
+        const ctx = {
+          client,
+          project,
+          directory: testDirectory,
+          worktree: testDirectory,
+          serverUrl: new URL("https://example.com"),
+          $: Bun.$,
+        } satisfies PluginInput
+        const beforeHandler = createToolExecuteBeforeHandler({ ctx, pendingFilePaths, pendingTaskRefs })
+        const afterHandler = createToolExecuteAfterHandler({
+          ctx,
+          pendingFilePaths,
+          pendingTaskRefs,
+          autoCommit: true,
+          getState: () => ({ promptFailureCount: 0 }),
+        })
+
+        await beforeHandler(
+          { tool: "task", sessionID, callID: "call-bg-task" },
+          { args: { prompt: "Implement auth flow" } },
+        )
+
+        const output = {
+          title: "Sisyphus Task",
+          output: "Background task launched.\n\nBackground Task ID: bg_123\n\n<task_metadata>\nsession_id: ses_child123\n</task_metadata>",
+          metadata: {
+            sessionId: childSessionID,
+            agent: "sisyphus-junior",
+            category: "deep",
+          },
+        }
+
+        await afterHandler(
+          { tool: "task", sessionID, callID: "call-bg-task" },
+          output,
+        )
+
+        expect(output.output).toContain("Background task launched.")
+        expect(collectGitDiffStatsMock).not.toHaveBeenCalled()
+        expect(readBoulderState(testDirectory)?.session_ids).toContain(childSessionID)
+        expect(readBoulderState(testDirectory)?.task_sessions?.["todo:1"]?.session_id).toBe(childSessionID)
+      })
+    })
   })
 })

src/hooks/atlas/tool-execute-after.ts

@@ -4,16 +4,17 @@ import {
   getPlanProgress,
   getTaskSessionState,
   readBoulderState,
-  readCurrentTopLevelTask,
   upsertTaskSessionState,
 } from "../../features/boulder-state"
 import { log } from "../../shared/logger"
 import { isCallerOrchestrator } from "../../shared/session-utils"
+import { syncBackgroundLaunchSessionTracking } from "./background-launch-session-tracking"
 import { collectGitDiffStats, formatFileChanges } from "../../shared/git-worktree"
 import { shouldPauseForFinalWaveApproval } from "./final-wave-approval-gate"
 import { HOOK_NAME } from "./hook-name"
 import { DIRECT_WORK_REMINDER } from "./system-reminder-templates"
 import { isSisyphusPath } from "./sisyphus-path"
+import { resolvePreferredSessionId, resolveTaskContext } from "./task-context"
 import { extractSessionIdFromMetadata, extractSessionIdFromOutput, validateSubagentSessionId } from "./subagent-session-id"
 import {
   buildCompletionGate,
@@ -23,50 +24,7 @@ import {
 } from "./verification-reminders"
 import { isWriteOrEditToolName } from "./write-edit-tool-policy"
 import type { PendingTaskRef, SessionState } from "./types"
-import type { ToolExecuteAfterInput, ToolExecuteAfterOutput, TrackedTopLevelTaskRef } from "./types"
-
-function resolvePreferredSessionId(currentSessionId?: string, trackedSessionId?: string): string {
-  return currentSessionId ?? trackedSessionId ?? "<session_id>"
-}
-
-function resolveTaskContext(
-  pendingTaskRef: PendingTaskRef | undefined,
-  planPath: string,
-): {
-  currentTask: TrackedTopLevelTaskRef | null
-  shouldSkipTaskSessionUpdate: boolean
-  shouldIgnoreCurrentSessionId: boolean
-} {
-  if (!pendingTaskRef) {
-    return {
-      currentTask: readCurrentTopLevelTask(planPath),
-      shouldSkipTaskSessionUpdate: false,
-      shouldIgnoreCurrentSessionId: false,
-    }
-  }
-
-  if (pendingTaskRef.kind === "track") {
-    return {
-      currentTask: pendingTaskRef.task,
-      shouldSkipTaskSessionUpdate: false,
-      shouldIgnoreCurrentSessionId: false,
-    }
-  }
-
-  if (pendingTaskRef.reason === "explicit_resume") {
-    return {
-      currentTask: readCurrentTopLevelTask(planPath),
-      shouldSkipTaskSessionUpdate: true,
-      shouldIgnoreCurrentSessionId: true,
-    }
-  }
-
-  return {
-    currentTask: pendingTaskRef.task,
-    shouldSkipTaskSessionUpdate: true,
-    shouldIgnoreCurrentSessionId: true,
-  }
-}
+import type { ToolExecuteAfterInput, ToolExecuteAfterOutput } from "./types"
 
 export function createToolExecuteAfterHandler(input: {
   ctx: PluginInput
@@ -116,15 +74,23 @@ export function createToolExecuteAfterHandler(input: {
     if (toolInput.callID) {
       pendingTaskRefs.delete(toolInput.callID)
     }
+    const boulderState = readBoulderState(ctx.directory)
     const isBackgroundLaunch = outputStr.includes("Background task launched") || outputStr.includes("Background task continued")
       || outputStr.includes("Background delegate launched")
       || outputStr.includes("Background agent task launched")
     if (isBackgroundLaunch) {
+      await syncBackgroundLaunchSessionTracking({
+        ctx,
+        boulderState,
+        toolInput,
+        toolOutput,
+        pendingTaskRef,
+        metadataSessionId,
+      })
       return
     }
 
     if (toolOutput.output && typeof toolOutput.output === "string") {
-      const boulderState = readBoulderState(ctx.directory)
       const worktreePath = boulderState?.worktree_path?.trim()
       const verificationDirectory = worktreePath ? worktreePath : ctx.directory
       const gitStats = collectGitDiffStats(verificationDirectory)