Agents In Use
What’s new todayLast update
3 use cases added today
News / FIELD GUIDE

OpenClaw 2026.11.4 fixes replies that never arrive

OpenClaw fixes missing scheduled replies, Telegram restart notices and subagent handoffs. Read what its November release proves and what still needs checking.

By Agents In Use editorial6 min read

OpenClaw published version 2026.11.4 on 16 November, with fixes for scheduled replies, restart recovery and subagent completion messages. The release notes describe repairs to several points between an agent doing work and someone receiving its answer. For anyone relying on recurring digests or delegated tasks, those are changes worth examining before adding another workflow.

This is a delivery update, not evidence of better reasoning. Our reading of the release and its implementation records is that operators should distinguish an answer in a transcript, a successful send and a message visible to its recipient. The records below show why that distinction matters, but they do not establish a universal delivery guarantee.

Check the recipient's conversation, not just the job's completion record.

Scheduled work could finish without sending its reply

The cron repair, merged on 10 November, addresses jobs targeted at an existing conversation. The agent could finish its turn and leave its answer in the transcript, yet fail to send it to the channel. System-event wakes used the same affected delivery path. The recorded error was "Plugin invocation scope is closed".

The implementation account describes an ordering problem. Dispatch closed the turn's plugin invocation scope before the session-event reply dispatcher drained its queue. The later send still carried that closed context. The correction lets the delivery callback leave the finished turn's scope before routing the reply. That is a specific delivery repair, rather than a change to the prompt or model.

The author tested before and after with a real local Matrix homeserver and a mock OpenAI provider. A separate observer account read the room through homeserver sync. Before the fix, the cron marker existed in the run summary but not in the room. Afterwards, the marker appeared in both. Only Matrix had a real-server test in that comparison; other channels were not run. The system-event wake path was reproduced before the fix, but the before-and-after scenario exercised cron.

Telegram recovery reports uncertainty instead of replaying

A separate Telegram restart fix, merged on 11 November, deals with a completed reply whose delivery remains unconfirmed after a Gateway restart. Recovery already refused to resend an ambiguous answer. The problem was that the explanation could wait until the user sent another message. The repair immediately hands the existing uncertainty notice to lifecycle delivery in the original chat.

That does not mean the missing answer is automatically resent. The implementation explicitly avoids ambiguous-answer replay, and notice delivery can itself fail if the channel remains unavailable. Its cold-process proof used the real Gateway and queue with a scripted OpenAI-compatible model and a loopback Bot API. It did not send real Telegram messages or test Telegram's external flood-control behaviour.

For an operator, our recommendation is to preserve that uncertainty rather than convert it into a success label. If a reply describes an external action, check that action's destination before asking the agent to repeat it. Treat "the answer was generated" and "the answer reached me" as separate questions in the incident record.

Subagent completion had two different blockers

The capability preparation repair, merged on 12 November, targets the 2026.11.4 release branch. Its report identifies direct child-completion announcements failing for Claude CLI parents on official 2026.11.1 to 2026.11.3 installations. The parent could show "Subagent completion delivery is blocked". The affected inline read belonged to that release line; the report says main and 2026.11.5-beta.2 did not have it.

The fix prepares the requester's capability profile before applying the shared synchronous policy. In the reported isolated Linux Gateway tests, patched Claude CLI parents received their children's direct completions. Those observations are narrower than a test of every runtime. The author did not verify an embedded, non-CLI requester live.

Another completion repair, also merged on 12 November, addresses a worker whose own child cannot deliver its result. The suspended descendant could keep the worker's completion waiting. The change excludes suspended deliveries when deciding whether completion is ready, while retaining the suspended child's result for recovery. Its Gateway comparisons used a stand-in model.

Our recommendation is to test the delegation pattern you actually use. Include a worker that finishes after its parent turn ends, and a worker with a child that cannot deliver. Record whether the requester receives the result and whether recovery information remains available. Do not treat a successful ordinary chat as a substitute for those checks.

A saved answer is another separate check

The local-reply persistence fix, merged on 10 November, addresses answers delivered through the local message tool but absent from conversation history. Confirmed source answers now persist within the accepted turn's range. The report separates that storage repair from downstream consumers choosing the correct row, and from native Codex interruption. Delivery, persistence and later capture are not interchangeable claims.

Historical cases explain the operational stakes

Our historical Telegram digest case records Alex Rezvov's February account of apparent success without a useful digest arriving. His reported correction concerned thinking-mode configuration in his setup. It is context for checking the actual conversation, not proof of the November repair or a universal prescription to disable thinking.

The historical family reminder case records DJ Chung's May account of Hamilton coordinating schedules in Telegram. Parents confirmed responsibilities, and closing the host MacBook's lid could stop the assistant. That account is not a test of this release. It illustrates a different dependency to include in an operator's checklist alongside the delivery path.

For an existing deployment, our recommendation is to select one harmless scheduled message and one bounded delegated task as upgrade checks. Write down the expected recipient, the visible result and the evidence needed to close each check. If the system cannot confirm delivery, retain that uncertainty and investigate it before trusting the same workflow with a time-sensitive reminder.

Sources