OpenClaw
OpenClaw proposes showing whether a child is working, waiting or still delivering
OpenClaw’s subagent proposal separates execution from result delivery and names what a child is waiting for. Selected-task waits, explicit incoming-message intent and descendant settlement aim to prevent premature parent replies, without claiming that process-local notifications survive a cold restart.
OpenClaw proposed a more explicit subagent status model late on September 13. The change addresses an operational ambiguity: a child may still be executing, waiting on another child, awaiting input, or finished with computation while its result has not reached the parent. Treating all of those as a single completion flag can produce an early answer or leave a workflow parked without an owner.
The existing documentation already distinguishes a subagent run from its session and describes results returning to the requester for review. This proposal extends the tools and nested inspector around that model. It adds selected-task waiting and explicit notification, steering and follow-up controls within existing tool names, while drawing observations from the current task, approval and delivery owners.
The facts
- The proposed observations distinguish active execution from a ready result still queued for parent delivery. - A childless sessions_yield requires explicit incoming-message intent rather than silently parking completed work. - Cron waits for descendant settlement, and owned TaskFlow/Lobster checkpoints can resume. - The task protocol adds optional version-tagged observations while keeping responses without them supported. - Passive notifications, tool attention and progress context are process-local; Workboard write claims are cooperative, not filesystem locks. - The author reports real-provider fanout, cancellation and timeout exercises, but explicitly does not claim cold process-crash recovery.
Why it matters
Delegation only helps if the parent can tell why it is waiting and whether a result has actually returned. This proposal makes the delivery leg part of the operator-visible story rather than treating it as an afterthought. It also keeps an important authority boundary: native Codex continues to own its execution and permissions. A richer OpenClaw inspector is not permission to reinterpret or override the child runtime’s controls.
Current
Inspected on 2026-09-14. The OpenClaw stable-channel baseline is 2026.9.4 on npm latest; extended-stable 2026.6.35. The primary source was open (not merged) when captured. Mainline merges, open proposals and packaged releases are distinct availability states; the release baseline does not establish that a proposal has shipped.
Evidence
Primary evidence: openclaw/openclaw PR #147571 (https://github.com/openclaw/openclaw/pull/147571). Supporting context: OpenClaw — subagent run, session and delivery documentation (https://docs.openclaw.ai/tools/subagents); OpenClaw — official product README (https://raw.githubusercontent.com/openclaw/openclaw/main/README.md). Public source bodies and linked documentation were inspected. Contributor test results remain attributed reports, not independently reproduced experiments.
Operator take
For a multi-child workflow, ask the inspector to identify a waiting child and its named dependency before allowing the parent to announce completion. Test a result that is ready but undelivered, a cancellation during parent delivery, and an incoming-message wait with no descendants. Separately test process restart if the workflow needs that property. Process-local success in a real-provider harness demonstrates useful behavior, but it is not evidence that a newly started gateway can reconstruct every notification.
Caveat
The PR was open at capture. Its evidence reports an initial live campaign that exposed two probe failures, followed by prompt corrections and passing reruns; those qualifications are retained rather than reduced to a blanket green claim. No new durable store or migration is introduced, and this article does not assert crash recovery, filesystem locking or independently reproduced test results.
The PR was open at capture. Its evidence reports an initial live campaign that exposed two probe failures, followed by prompt corrections and passing reruns; those qualifications are retained rather than reduced to a blanket green claim. No new durable store or migration is introduced, and this article does not assert crash recovery, filesystem locking or independently reproduced test results.