fix(agents): reconcile child output before lost-context sweep failure

LiuwqGit · cursoragent · LiuwqGit · commit 1604b7471b13 · 2026-06-05T08:45:24.000+08:00
When a subagent run loses live execution context, treat readable child assistant output as terminal evidence and complete as ok instead of always emitting lost-context error. Closes #90299. Co-authored-by: Cursor <cursoragent@cursor.com>
diff --git a/src/agents/subagent-lost-context-completion.test.ts b/src/agents/subagent-lost-context-completion.test.ts
@@ -0,0 +1,33 @@
+import { beforeEach, describe, expect, it, vi } from "vitest";
+import * as announceOutput from "./subagent-announce-output.js";
+import {
+  LOST_ACTIVE_EXECUTION_CONTEXT_ERROR,
+  resolveStaleActiveSubagentOutcome,
+} from "./subagent-lost-context-completion.js";
+
+describe("resolveStaleActiveSubagentOutcome", () => {
+  beforeEach(() => {
+    vi.restoreAllMocks();
+  });
+
+  it("returns ok when child session has readable assistant output", async () => {
+    vi.spyOn(announceOutput, "readSubagentOutput").mockResolvedValue("# ARCHITECTURE.md");
+    await expect(
+      resolveStaleActiveSubagentOutcome({
+        childSessionKey: "agent:main:subagent:child",
+      }),
+    ).resolves.toEqual({ status: "ok" });
+  });
+
+  it("returns lost-context error when no readable output is available", async () => {
+    vi.spyOn(announceOutput, "readSubagentOutput").mockResolvedValue(undefined);
+    await expect(
+      resolveStaleActiveSubagentOutcome({
+        childSessionKey: "agent:main:subagent:child",
+      }),
+    ).resolves.toEqual({
+      status: "error",
+      error: LOST_ACTIVE_EXECUTION_CONTEXT_ERROR,
+    });
+  });
+});
diff --git a/src/agents/subagent-lost-context-completion.ts b/src/agents/subagent-lost-context-completion.ts
@@ -0,0 +1,24 @@
+/**
+ * Reconcile stale active subagent runs that lost live execution context.
+ *
+ * When the sweeper cannot resolve terminal state from the session store, readable child
+ * assistant output is treated as ground truth and completes as ok instead of a plain failure.
+ */
+import type { SubagentRunOutcome } from "./subagent-announce-output.js";
+import { readSubagentOutput } from "./subagent-announce-output.js";
+
+export const LOST_ACTIVE_EXECUTION_CONTEXT_ERROR = "subagent run lost active execution context";
+
+/** Resolve terminal outcome for a stale active run with no live agent.run context. */
+export async function resolveStaleActiveSubagentOutcome(params: {
+  childSessionKey: string;
+}): Promise<SubagentRunOutcome> {
+  const output = await readSubagentOutput(params.childSessionKey);
+  if (output?.trim()) {
+    return { status: "ok" };
+  }
+  return {
+    status: "error",
+    error: LOST_ACTIVE_EXECUTION_CONTEXT_ERROR,
+  };
+}
diff --git a/src/agents/subagent-registry.test.ts b/src/agents/subagent-registry.test.ts
@@ -2218,6 +2218,67 @@ describe("subagent registry seam flow", () => {
     expect(run?.cleanupCompletedAt).toBeTypeOf("number");
   });
 
+  it("completes stale active runs as ok when child session has readable output", async () => {
+    mocks.callGateway.mockImplementation(async (request: { method?: string }) => {
+      if (request.method === "agent.wait") {
+        return { status: "pending" };
+      }
+      if (request.method === "chat.history") {
+        return {
+          messages: [
+            {
+              role: "assistant",
+              content: [{ type: "text", text: "# ARCHITECTURE.md\nrelease readiness design" }],
+            },
+          ],
+        };
+      }
+      return {};
+    });
+    mocks.loadSessionStore.mockReturnValue({
+      "agent:main:subagent:child": {
+        sessionId: "sess-child",
+        updatedAt: 1,
+        status: "running",
+      },
+    });
+
+    const createdAt = Date.parse("2026-03-24T11:00:00Z");
+    vi.setSystemTime(createdAt);
+    mod.registerSubagentRun({
+      runId: "run-lost-context-with-output",
+      childSessionKey: "agent:main:subagent:child",
+      requesterSessionKey: "agent:main:main",
+      requesterDisplayKey: "main",
+      task: "deliver architecture doc",
+      cleanup: "keep",
+      expectsCompletionMessage: true,
+    });
+
+    vi.setSystemTime(createdAt + 65_000);
+    await mod.testing.sweepOnceForTests();
+
+    await waitForFast(() => {
+      const announceParams = findRecordCallArg(
+        mocks.runSubagentAnnounceFlow,
+        0,
+        "lost context recovered announce",
+        (record) => record.childRunId === "run-lost-context-with-output",
+      );
+      expectRecordFields(
+        announceParams.outcome,
+        { status: "ok" },
+        "lost context recovered announce outcome",
+      );
+    });
+
+    const run = mod
+      .listSubagentRunsForRequester("agent:main:main")
+      .find((entry) => entry.runId === "run-lost-context-with-output");
+    expectRecordFields(run?.outcome, { status: "ok" }, "lost context recovered run outcome");
+    expect(run?.outcome?.error).toBeUndefined();
+  });
+
   it("uses session-store start time when sweeping stale explicit-timeout runs", async () => {
     mocks.callGateway.mockImplementation(async (request: { method?: string }) => {
       if (request.method === "agent.wait") {
diff --git a/src/agents/subagent-registry.ts b/src/agents/subagent-registry.ts
@@ -40,6 +40,7 @@ import {
   SUBAGENT_ENDED_REASON_KILLED,
   type SubagentLifecycleEndedReason,
 } from "./subagent-lifecycle-events.js";
+import { resolveStaleActiveSubagentOutcome } from "./subagent-lost-context-completion.js";
 import {
   emitSubagentEndedHookOnce,
   resolveLifecycleOutcomeFromRunOutcome,
@@ -958,20 +959,25 @@ async function sweepSubagentRuns() {
             continue;
           }
 
+          const lostContextOutcome = await resolveStaleActiveSubagentOutcome({
+            childSessionKey: entry.childSessionKey,
+          });
           await completeSubagentRunWithRecovery(
             {
               runId,
               endedAt: now,
-              outcome: {
-                status: "error",
-                error: "subagent run lost active execution context",
-              },
-              reason: SUBAGENT_ENDED_REASON_ERROR,
+              outcome: lostContextOutcome,
+              reason:
+                lostContextOutcome.status === "ok"
+                  ? SUBAGENT_ENDED_REASON_COMPLETE
+                  : SUBAGENT_ENDED_REASON_ERROR,
               sendFarewell: true,
               accountId: entry.requesterOrigin?.accountId,
               triggerCleanup: true,
             },
-            "sweeper-lost-context",
+            lostContextOutcome.status === "ok"
+              ? "sweeper-lost-context-recovered"
+              : "sweeper-lost-context",
           );
           continue;
         }
