Describe the bug

When spawning sub-agents via sessions_spawn from a Telegram DM session and then sessions_yield, the completion announce back to the parent session is unreliable:

• Pattern A (silent loss): Sub-agent completes successfully, but the parent session never wakes up. The user gets no notification. Only when the user manually sends a follow-up message does the parent realize the child finished.

• Pattern B (duplicate delivery): Multiple sub-agent completion events queue up, then when the parent does wake (user sends a message), ALL queued completions are delivered at once — resulting in 3 replies for 3 sub-agents simultaneously.

Both patterns happen intermittently on the same setup with no config changes between them.

Environment

• OpenClaw: 2026.5.28 (e932160)
• OS: macOS 26.5.1 (arm64) — Mac mini M4
• Node: 26.0.0
• Channel: Telegram (direct DM, not group)
• Agents: 9 configured
• Model: zhipu/glm-5.1 (Coding Plan)
• Sub-agents: spawned with sessions_spawn(mode="run"), waited with sessions_yield

Steps to reproduce

1. From a Telegram DM session, spawn 1-3 sub-agents via sessions_spawn
2. Call sessions_yield to wait for completions
3. Observe: sometimes the parent wakes and delivers the result, sometimes it stays silent until the user sends a follow-up message
4. When multiple sub-agents complete close together, sometimes all results arrive in a burst

Expected behavior

Each sub-agent completion should reliably:

1. Wake the yielded parent session
2. Deliver exactly ONE reply to the Telegram chat
3. No silent loss, no duplicate burst delivery

Actual behavior

• ~50% of the time: completion is silently lost (parent stays yielded)
• ~20% of the time: multiple completions arrive simultaneously in a burst
• ~30% of the time: works as expected (single reliable delivery)

These are rough estimates from ~15 sub-agent runs over the past 2 days.

Logs / Evidence

No error-level logs observed during silent loss. The sub-agent sessions show they completed normally (full transcript, correct output). The issue appears to be in the announce/delivery layer, not the sub-agent execution.

Related Issues

• [Bug]: Sub-agent completion announce fails with "unsupported channel: telegram" #21968 — Telegram announce fails with "unsupported channel" (fixed)
• Sub-agent announce silently fails when message exceeds Telegram character limit #27317 — Announce silently fails when message exceeds Telegram length limit (fixed)
• [Bug]: Subagent completion silently lost — no retry, no notification, no auto-restart on timeout #44925 — Subagent completion silently lost (partially fixed)
• [Bug]: subagent completion spawns a fresh run on the parent's route instead of resuming the yielded session (supersedes #80310) #81490 — Fresh run spawned instead of resuming yielded session (fixed)
• Auto-announce completion events not reliably delivering to parent sessions #84159 — Auto-announce completion events not reliably delivering (open)

Workaround (agent-side)

We've added a rule in our agent workspace: after sessions_yield resumes, the parent agent MUST immediately send an explicit reply with the sub-agent results, regardless of whether auto-announce succeeded. This partially mitigates the silent-loss case, but does not help when the parent never wakes at all.

Suggested fix

It would be great if the announce layer had a more robust delivery guarantee for Telegram:

1. If channel delivery fails, fall back to injecting the completion as an internal event in the parent session (so the parent can self-reply)
2. Add deduplication for queued completion events to prevent burst delivery
3. Log at warn level when announce fails or is retried, so it is observable in openclaw logs