Summary

• Problem: active requester completion handoffs treated a queued internal completion user message_end as delivered even when the requester ended before consuming it; the direct-failure fallback steering path also used transcript commit alone.
• Solution: completion handoff wakes now opt into waiting for a following assistant response after transcript commit, including fallback steering after a completion direct-send failure.
• What changed: added waitForAssistantResponseAfterTranscriptCommit to the embedded queue wait path and enabled it only for completion handoff steering.
• What did NOT change: non-completion progress steering still resolves at the transcript commit boundary; channel transports, gateway send behavior, message-tool-only delivery, and timeout defaults were not changed.
• Why it matters / User impact: completed subagent results should either be consumed by the active requester or remain eligible for existing fallback delivery, not silently disappear at requester shutdown.
• Fixes [Bug]: Subagent completion silently dropped when announce steers into a requester run that ends before processing it #92433

Real behavior proof

• Behavior or issue addressed: Subagent completion is no longer marked delivered just because the internal completion user message reached the requester transcript. The fixed boundaries are both queued completion user message_end -> requester agent_end -> no assistant response and completion direct send fails -> fallback steering; both now require requester assistant acknowledgement before reporting a steered completion as delivered.
• Real environment tested: Local Linux worktree /media/vdc/code/ai/aispace/openclaw-worktrees/issue-92433, branch feat/issue-92433, head ef5a2f8fd7daedac01fd5e93eecab43111d57b81.
• Exact steps or command run after this patch:

cd /media/vdc/code/ai/aispace/openclaw-worktrees/issue-92433 && OPENCLAW_REPO=/media/vdc/code/ai/aispace/openclaw-worktrees/issue-92433 OPENCLAW_EVIDENCE_DIR=/media/vdc/code/ai/aispace/openclaw-issue-92433-evidence node --import tsx /media/vdc/code/ai/aispace/openclaw-issue-92433-evidence/subagent-completion-active-run-queue-proof.mjs > /media/vdc/code/ai/aispace/openclaw-issue-92433-evidence/subagent-completion-active-run-queue-proof.out.json

cd /media/vdc/code/ai/aispace/openclaw-worktrees/issue-92433 && OPENCLAW_TEST_FAST=1 node scripts/run-vitest.mjs src/agents/subagent-announce-delivery.test.ts -t "requires requester assistant acknowledgement when completion direct failure fallback-steers"

cd /media/vdc/code/ai/aispace/openclaw-worktrees/issue-92433 && OPENCLAW_TEST_FAST=1 node scripts/run-vitest.mjs src/agents/subagent-announce-delivery.test.ts src/agents/subagent-announce-dispatch.test.ts src/agents/embedded-agent-runner/run/attempt.queue-message.test.ts

cd /media/vdc/code/ai/aispace/openclaw-worktrees/issue-92433 && node scripts/run-tsgo.mjs -p tsconfig.core.json --noEmit

cd /media/vdc/code/ai/aispace/openclaw-worktrees/issue-92433 && ./node_modules/.bin/oxfmt --check --threads=1 src/agents/subagent-announce-delivery.ts src/agents/subagent-announce-delivery.test.ts

git -C /media/vdc/code/ai/aispace/openclaw-worktrees/issue-92433 diff --check

• Evidence after fix: The active-run proof runs outside Vitest and does not call testing.setDepsForTest(). It registers a requester in the production ACTIVE_EMBEDDED_RUNS registry, lets the default production deliverAgentHarnessTaskCompletion() path resolve requester activity, and lets queueEmbeddedAgentMessageWithOutcomeAsync() call the real active-run handle. The handle uses the production steerAndWaitForTranscriptCommit() wait logic, so queue/gateway outcomes are not injected.

{
  "proof": "subagent-completion-active-run-queue-ack",
  "head": "ef5a2f8fd7daedac01fd5e93eecab43111d57b81",
  "runtime": "node --import tsx; production deliverAgentHarnessTaskCompletion; production ACTIVE_EMBEDDED_RUNS registry; production queueEmbeddedAgentMessageWithOutcomeAsync; production steerAndWaitForTranscriptCommit; no testing.setDepsForTest and no injected queue/gateway outcome",
  "ackScenario": {
    "delivery": {
      "delivered": true,
      "path": "steered",
      "phases": [
        {
          "phase": "direct-primary",
          "delivered": true,
          "path": "steered"
        }
      ]
    }
  },
  "teardownScenario": {
    "queueOutcome": {
      "queued": false,
      "sessionId": "active-run-proof-teardown-session",
      "reason": "runtime_rejected",
      "gatewayHealth": "live",
      "errorMessage": "active session ended before queued steering message was consumed"
    }
  },
  "assertions": {
    "completionDeliveredAfterAssistantAck": true,
    "activeQueueUsedAssistantAckOption": true,
    "teardownBeforeAssistantDoesNotReportDelivered": true,
    "teardownFailureMentionsUnconsumed": true,
    "noInjectedQueueOrGatewayOutcome": true
  },
  "events": [
    {
      "event": "active-run-registered",
      "mode": "assistant-ack",
      "resolvedSessionId": "active-run-proof-ack-session",
      "isActive": true
    },
    {
      "event": "queue-handle-called",
      "mode": "assistant-ack",
      "waitForTranscriptCommit": true,
      "waitForAssistantResponseAfterTranscriptCommit": true
    },
    {
      "event": "queued-user-message-end",
      "mode": "assistant-ack"
    },
    {
      "event": "assistant-message-start",
      "mode": "assistant-ack"
    },
    {
      "event": "active-run-registered",
      "mode": "teardown-before-assistant",
      "resolvedSessionId": "active-run-proof-teardown-session",
      "isActive": true
    },
    {
      "event": "queue-handle-called",
      "mode": "teardown-before-assistant",
      "waitForTranscriptCommit": true,
      "waitForAssistantResponseAfterTranscriptCommit": true
    },
    {
      "event": "queued-user-message-end",
      "mode": "teardown-before-assistant"
    },
    {
      "event": "agent-end-before-assistant",
      "mode": "teardown-before-assistant"
    }
  ]
}

Test Files  2 passed (2)
Tests  2 passed | 200 skipped (202)

Test Files  3 passed (3)
Tests  118 passed (118)
[test] passed 1 Vitest shard in 18.79s

node scripts/run-tsgo.mjs -p tsconfig.core.json --noEmit exited 0 with no stdout/stderr output.

All matched files use the correct format.
Finished in 48ms on 2 files using 1 threads.

git diff --check exited 0 with no whitespace errors.

• Observed result after fix: A completion handoff delivered through the production active-run queue returns delivered only after the queued completion is followed by requester assistant activity. The same production queue wait returns runtime_rejected with active session ended before queued steering message was consumed when the requester reaches agent_end after the queued completion but before any assistant response, so transcript durability alone is no longer treated as completion delivery. Ordinary non-completion steering still uses transcript-commit delivery.
• What was not tested: No live Discord, Telegram, Slack, Desktop, Mantis, or other external channel transport proof was run in this environment. The active-run proof uses deterministic in-process requester transcript events against the production active-run registry and queue owner; it is not a full external Gateway/model-provider/channel run. It specifically removes the previous proof gap where queue/gateway delivery outcomes were injected.

Regression Test Plan

• Target test file: src/agents/embedded-agent-runner/run/attempt.queue-message.test.ts, src/agents/subagent-announce-delivery.test.ts, src/agents/subagent-announce-dispatch.test.ts
• Scenario locked in: queued completion user message_end -> requester agent_end -> no assistant response rejects the active completion wake as unconsumed, and completion direct-send failure fallback steering also waits for requester assistant acknowledgement.
• Why this is the smallest reliable guardrail: The new acknowledgement option is opt-in and enabled only for completion handoff steering. It reuses existing queue outcome and fallback paths rather than adding channel-specific delivery logic.

Merge risk

• Risk labels considered: clawsweeper:no-new-fix-pr, clawsweeper:needs-maintainer-review, clawsweeper:needs-product-decision, nonlocal channel/live proof boundary.
• Risk explanation: The patch changes runtime completion acknowledgement semantics, so maintainer review should focus on whether assistant message_start/message_end is the intended completion-consumption boundary. The risk is contained because non-completion steering keeps its previous transcript-commit behavior and unsupported transcript-wait targets still fall back to best-effort wake options.
• Why acceptable: The local proof exercises the production active-run queue owner and converts the false delivered signal into the already-supported fallback flow without changing public channel contracts.

Root Cause

• Root cause: The active requester wake and fallback steering paths used waitForTranscriptCommit as delivery proof for subagent completion handoffs. That proves the internal completion message was committed as a queued user entry, but it does not prove the requester consumed the completion or produced a visible assistant response.
• Why this is root-cause fix: The requester transcript event stream now distinguishes durability from consumption. A matching queued user message_end proves durability; a later assistant message_start/message_end proves requester consumption. If terminal requester state arrives first, the wake is unconfirmed and can fall through to existing fallback delivery.
• Fix classification: Root-cause reliability fix for subagent completion delivery acknowledgement.
• Maintainer-ready confidence: High for the active-run queue owner and source-level race covered here. The fix narrows the delivered boundary only for completion handoff paths, and related local runtime tests pass. External channel/live transport proof remains outside this local environment.
• Patch quality notes: The diff is limited to subagent completion steering options and regression expectations. It does not add sleeps, widen channel transport behavior, or change the completion direct-first dispatch ordering.
• Architecture / source-of-truth check: The active requester transcript stream is the source of truth. The patch changes the source-of-truth delivery boundary before downstream fallback decisions, rather than adding downstream string checks or transport-specific special cases.