fix(atlas): tighten session reuse metadata parsing

src/hooks/atlas/index.test.ts

@@ -33,6 +33,8 @@ mock.module("../../shared/opencode-storage-detection", () => ({
 }))
 
 const { createAtlasHook } = await import("./index")
+const { createToolExecuteAfterHandler } = await import("./tool-execute-after")
+const { createToolExecuteBeforeHandler } = await import("./tool-execute-before")
 const { MESSAGE_STORAGE } = await import("../../features/hook-message-injector")
 
 describe("atlas hook", () => {
@@ -410,6 +412,62 @@ describe("atlas hook", () => {
       cleanupMessageStorage(sessionID)
     })
 
+    test("should clean pending task refs when a task returns background launch output", async () => {
+      // given - direct handlers with shared pending maps
+      const sessionID = "session-bg-launch-cleanup-test"
+      setupMessageStorage(sessionID, "atlas")
+
+      const planPath = join(TEST_DIR, "background-cleanup-plan.md")
+      writeFileSync(planPath, `# Plan
+
+## TODOs
+- [ ] 1. Implement auth flow
+`)
+      writeBoulderState(TEST_DIR, {
+        active_plan: planPath,
+        started_at: "2026-01-02T10:00:00Z",
+        session_ids: ["session-1"],
+        plan_name: "background-cleanup-plan",
+      })
+
+      const pendingFilePaths = new Map<string, string>()
+      const pendingTaskRefs = new Map<string, { key: string; label: string; title: string } | null>()
+      const beforeHandler = createToolExecuteBeforeHandler({
+        ctx: createMockPluginInput(),
+        pendingFilePaths,
+        pendingTaskRefs,
+      })
+      const afterHandler = createToolExecuteAfterHandler({
+        ctx: createMockPluginInput(),
+        pendingFilePaths,
+        pendingTaskRefs,
+        autoCommit: true,
+        getState: () => ({ promptFailureCount: 0 }),
+      })
+
+      // when - the task is captured before execution
+      await beforeHandler(
+        { tool: "task", sessionID, callID: "call-bg-launch" },
+        { args: { prompt: "Implement auth flow" } }
+      )
+      expect(pendingTaskRefs.size).toBe(1)
+
+      // and the task returns a background launch result
+      await afterHandler(
+        { tool: "task", sessionID, callID: "call-bg-launch" },
+        {
+          title: "Sisyphus Task",
+          output: "Background task launched.\n\nSession ID: ses_bg_12345",
+          metadata: {},
+        }
+      )
+
+      // then - the pending task ref is still cleaned up
+      expect(pendingTaskRefs.size).toBe(0)
+
+      cleanupMessageStorage(sessionID)
+    })
+
      test("should persist preferred subagent session for the current top-level task", async () => {
        // given - boulder state with a current top-level task, Atlas caller
        const sessionID = "session-task-session-track-test"

src/hooks/atlas/subagent-session-id.test.ts

@@ -50,4 +50,19 @@ session_id: ses_real_metadata_456
     // then
     expect(result).toBe("ses_real_metadata_456")
   })
+
+  test("does not let task_metadata parsing bleed into incidental body text after the closing tag", () => {
+    // given
+    const output = `<task_metadata>
+session_id: ses_real_metadata_456
+</task_metadata>
+
+debug log: session_id: ses_wrong_body_789`
+
+    // when
+    const result = extractSessionIdFromOutput(output)
+
+    // then
+    expect(result).toBe("ses_real_metadata_456")
+  })
 })

src/hooks/atlas/subagent-session-id.ts

@@ -1,8 +1,11 @@
 export function extractSessionIdFromOutput(output: string): string | undefined {
-  const taskMetadataMatches = [...output.matchAll(/<task_metadata>[\s\S]*?session_id:\s*(ses_[a-zA-Z0-9_]+)[\s\S]*?<\/task_metadata>/gi)]
-  const lastTaskMetadataMatch = taskMetadataMatches.at(-1)
-  if (lastTaskMetadataMatch) {
-    return lastTaskMetadataMatch[1]
+  const taskMetadataBlocks = [...output.matchAll(/<task_metadata>([\s\S]*?)<\/task_metadata>/gi)]
+  const lastTaskMetadataBlock = taskMetadataBlocks.at(-1)?.[1]
+  if (lastTaskMetadataBlock) {
+    const taskMetadataSessionMatch = lastTaskMetadataBlock.match(/session_id:\s*(ses_[a-zA-Z0-9_]+)/i)
+    if (taskMetadataSessionMatch) {
+      return taskMetadataSessionMatch[1]
+    }
   }
 
   const explicitSessionMatches = [...output.matchAll(/Session ID:\s*(ses_[a-zA-Z0-9_]+)/g)]

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

@@ -71,6 +71,10 @@ export function createToolExecuteAfterHandler(input: {
     }
 
     const outputStr = toolOutput.output && typeof toolOutput.output === "string" ? toolOutput.output : ""
+    const pendingTaskRef = toolInput.callID ? pendingTaskRefs.get(toolInput.callID) : undefined
+    if (toolInput.callID) {
+      pendingTaskRefs.delete(toolInput.callID)
+    }
     const isBackgroundLaunch = outputStr.includes("Background task launched") || outputStr.includes("Background task continued")
     if (isBackgroundLaunch) {
       return
@@ -80,10 +84,6 @@ export function createToolExecuteAfterHandler(input: {
       const gitStats = collectGitDiffStats(ctx.directory)
       const fileChanges = formatFileChanges(gitStats)
       const subagentSessionId = extractSessionIdFromOutput(toolOutput.output)
-      const pendingTaskRef = toolInput.callID ? pendingTaskRefs.get(toolInput.callID) : undefined
-      if (toolInput.callID) {
-        pendingTaskRefs.delete(toolInput.callID)
-      }
 
       const boulderState = readBoulderState(ctx.directory)
       if (boulderState) {