OpenClaw
OpenClaw drafts supervised tasks that stop when execution becomes uncertain
An experimental OpenClaw TaskFlow draft proposes durable ownership, finite attempts and explicit task endpoints. It would recover work lost before dispatch but request input after an ambiguous side effect, rather than claiming that a surviving process or a running record proves progress.
An OpenClaw contributor opened a supervised TaskFlow proof of concept on September 9 to address a familiar unattended-agent failure: the conversation ends, a command survives, and nobody owns collecting the result. The draft asks for architectural feedback on an opt-in controller, not approval to redefine existing TaskFlow. Its central choice is conservative recovery. When execution might already have changed something, the episode stops for input instead of guessing that retrying is safe.
The facts
- The proposal records structured goals, finite attempt and deadline policies, transactional ownership and durable next-action timestamps. - Its endpoints are success, preaccepted partial success, input required, failure and cancellation; new input creates a new episode without rewriting the old endpoint. - Pre-dispatch work can be reclaimed, while interrupted dispatched attempts preserve uncertainty instead of automatically replaying effects. - The proof of concept uses actual Codex and Claude CLI full-turn adapters with a reasoning and file-tool allowlist. - The author reports 69 focused tests plus a separate process-and-SQLite crash proof; persistence, retention and downgrade design still require maintainer review.
Why it matters
This separates liveness from responsibility. A process can be alive while its coordinator is gone, and a database can say running after all useful supervision stopped. Explicit ownership and observed custody would make those states inspectable. The trade-off is intentional: a task may require a human decision even when an optimistic retry could finish it. For file changes or external actions that are not safely repeatable, admitting uncertainty is more useful than a tidy success badge.
Current
Inspected on 2026-09-09. The OpenClaw stable-release baseline is v2026.9.3 published 2026-09-08T14:15:53Z. The main source was open (not merged) when captured. Release metadata is a version boundary; an open proposal, a merged mainline change and a packaged release remain different states.
Evidence
The primary evidence is openclaw/openclaw PR #142724 (https://github.com/openclaw/openclaw/pull/142724). Supporting context comes from openclaw/openclaw issue #142723 (https://github.com/openclaw/openclaw/issues/142723); OpenClaw — current official README (https://raw.githubusercontent.com/openclaw/openclaw/main/README.md). These pages were inspected directly. Related project records explain the source cluster; they are not independent replications of contributor test claims.
Operator take
Watch the controller and persistence review rather than treating the branch as a ready-made automation service. A meaningful evaluation should kill an owner before dispatch, after an effect, and while a durable timer is waiting, then inspect the recorded next owner and endpoint. The author's proof separates synthetic transition decisions from actual model runs; preserve that distinction when comparing results. Existing controller-managed TaskFlow remains the baseline, so the draft is a proposed stronger contract rather than evidence that today's implementation violated an existing guarantee.
Caveat
The draft does not supervise every chat, detached command or child agent. Semantic success remains model-reported, exactly-once external effects are not implemented, and endpoint publication still depends on recovered supervision, writable storage and clock progress. Local test claims were inspected, not independently reproduced.
The draft does not supervise every chat, detached command or child agent. Semantic success remains model-reported, exactly-once external effects are not implemented, and endpoint publication still depends on recovered supervision, writable storage and clock progress. Local test claims were inspected, not independently reproduced.