OpenClaw v2026.11.3 stops calling a failed tool turn a success

When a tool failed and nothing answered it, OpenClaw used to end your turn quietly and mark it a success. Now it tells you the tool failed.

Slide text

01 02 03 04 Agent turns

openclaw/openclaw · CHANGELOG/2026.11.3.md · #182458

OpenClaw v2026.11.3 stops calling a failed tool turn a success

An unanswered tool failure is now reported to you.

Subagents marked private now finish privately. Their completions no longer leak into the chat reply while it streams to you.

A parent agent can now cancel a child task and hand off to replacement work. Before, the cancelled child's completion receipt could block that handoff.

Provider authentication that stalls no longer holds a run open past its deadline. And when you cancel a run, the sign-in work under it stops as well.

Part two: secrets that stay where they belong.

A cached plugin result now stays tied to the credentials that fetched it. Firecrawl search keys stay with their endpoints, and memory embedding keys go only where they're meant to.

Reading your config now hides credentials written inline. Exa errors no longer echo your search key back, and managed Signal daemons and bundled language servers stop inheriting the Gateway's secrets.

If OpenClaw can't read a credential, it now says so instead of acting as if you never set one. And picking an auth profile from the wrong provider is refused.

When a sandbox is required, image previews now stay inside the allowed folders, even in Full Access mode. Image allowances an administrator set explicitly still work.

Workspace transfers now verify the certificate before any credentials or data leave your machine. And an empty fingerprint no longer erases a node's saved certificate pin.

Part three: a Gateway that stays up.

The Gateway loads fewer module files at startup, so it comes up faster. And a message with a long run of punctuation no longer stalls its abort-command check.

When storage fails for a queued cron job, the Gateway now handles it instead of exiting. And turning a one-shot job into a recurring one clears its one-shot cleanup.

Incoming Twitch chat stays out of parser error logs. With several Tlon approvals pending, you now name the request. And an exhausted Anthropic usage cap no longer burns retries meant for brief rate limits.

Part four: updates that can roll back.

If an update's first package check times out, OpenClaw now recovers safely and can roll back to the package you had. Profile-backed updates also stop tripping over their own maintenance lock.

Git repair no longer counts an unconfirmed command as a success. A soft reset in a terminal session stays soft, and the memory provider accepts none and auto in any letter case.

This release adds no new settings. The Android app follows once its release jobs finish. Every change is in the release changelog. If your agents call tools or subagents, I'd update today.