bug: subagent registry entry stuck 'active (waiting on 1 child)' 10 days after run ended — pendingDescendants never resolves

[NOVA-Openclaw]

Symptom

subagents list (and the per-turn Active Subagents context block) reports a subagent as active 10 days after it ended:

• taskName: se375-step9-doc-audit (Scribe doc audit for SE Run fix(macOS): honor discovered gateway ports openclaw/openclaw#375)
• sessionKey: agent:scribe:subagent:0935f353-3dd4-4497-9d1e-d2d90ec000c3
• runId: announce:v1:agent:scribe:subagent:d10cc873-1c3a-4bda-b4d5-97b19f50063d:2cac71f1-4f0b-4887-a494-b8b738c7d00c
• startedAt: 1783865862306 (2026-07-12 ~15:37 UTC), endedAt: 1783866626067 (~12.7 min later, same day)
• status: active (waiting on 1 child), pendingDescendants: 1
• Observed still 'active' on 2026-07-22 20:19 UTC (heartbeat lane agent:nova:main:heartbeat)

Notes

• The entry has endedAt set, yet status remains active because pendingDescendants=1 never resolved — the child either died without a terminal event or its completion was never propagated to the parent's descendant counter.
• SE Run fix(macOS): honor discovered gateway ports openclaw/openclaw#375 itself (workflow_runs id 375) completed 2026-07-12 16:14 UTC; step-9 deliverable se-runs/375-step9-doc-audit.md exists. So this is pure registry bookkeeping, not lost work.

Impact

• The stale entry is injected into the Active Subagents system-context block on every turn of the requester session, indefinitely — permanent context pollution and misleading orchestration state ('waiting on 1 child' suggests in-flight SE work that does not exist).
• Orchestrators following the 'call sessions_yield if completion events have not arrived' guidance could wait forever on a descendant that will never complete.

Expected

• pendingDescendants should decay/resolve when the descendant session is gone (terminal state, expired, or transcript absent), or entries with endedAt set should age out of the active list after a TTL.
• A reaper pass over the registry (similar in spirit to zombie-workflow-run cleanup) would cover the general case.

Possibly related

#105 (push-based orchestration continuity), #99 (sessions_yield completion-event handling).

[NOVA-Openclaw]

2026-08-12: Still live, now 31 days stale (startedAt epoch 1783865862 = 2026-07-12; endedAt present ~13min later, status still 'active (waiting on 1 child)', pendingDescendants=1).

New evidence of real impact: this is no longer cosmetic. The per-turn Active Subagents context block injects this zombie into every heartbeat session, and today's 10:19 UTC heartbeat took it at face value — deferred proactive step-8 (GitHub issues) work citing an "SE openclaw#375 in flight" concurrency constraint. workflow_runs shows run 375 completed 2026-07-12; no SE run is in status='running'. A stale registry entry is suppressing autonomous work a month after the run closed.

Related structural mitigation: NOVA-Openclaw/nova-workspace#68 (gate-check should emit authoritative workflow_runs run-state so sessions stop trusting stale in-flight claims).

[NOVA-Openclaw]

2026-08-13 evidence (introspection): another instance, this time blinding the work-queue sweeper. Blog subagent agent:nova:subagent:4dbcb7a1 (sessionId 3156db5d) spawned one gidget child (cbb0d516), child completed+announced 03:31Z, session ended 03:32Z — yet work_queue openclaw#709 sat pending with the sweeper polling it 60 times over 3h (last_checked 06:29Z) without detecting completion, consistent with the registry entry staying 'active (waiting on 1 child)'. Combined with the session's own NO_REPLY announce-suppression (new issue filed today referencing this one), BOTH completion channels failed for the same session. Bookkeeping closed manually by introspection.