OpenClaw published version 2026.11.3 on 10 November with fixes for private subagent results appearing in chat, unanswered tool failures being recorded as success, and several credential-handling problems. The release notes explicitly list no new configuration options or features. This is a release about what an agent exposes, how it stops and what it tells you happened.
Our reading is that these repairs deserve attention before another unattended workflow goes into service. A private task result, a completed request and a healthy scheduler are different promises. The implementation records make those differences unusually concrete, while also documenting where testing stopped.
Private work was becoming visible during streaming
The private-completion repair, merged on 30 October, addresses a classification mismatch. Gateway admission registered a private completion as hidden. Later session preparation registered the same run again, but did not apply the same privacy rule. Its text could then stream into the visible conversation.
The change keeps private completion work out of public live chat text in the Control UI and session subscribers. Ordinary public runs still stream. Its scope is narrower than a general privacy guarantee: the PR explicitly says it does not change raw history or run metadata. That distinction matters if you are deciding what private means in your own deployment.
The authors report regression coverage through the real session preparer, Gateway admission and publisher with a session subscriber. They also disclose that there was no Control UI browser end-to-end run, and that the controlled model command did not persist assistant output. We would read that as evidence for the streaming repair, not proof that every possible storage or display route has been checked.
A private task result, a completed request and a healthy scheduler are different promises.
A warning is not a completed answer
Another included repair concerns a denied tool call followed by no authored answer. The underlying PR, merged on 25 October, describes the old run metadata and trajectory recording that turn as successful. The generated blocked-tool warning had counted as answer evidence.
The repair preserves the warning but reports the unanswered run as failed. It also distinguishes pre-tool commentary from a delivered answer. The PR says approval decisions and execution policy are unchanged, so this is an accounting and reporting correction, not permission to get around a denial.
Our recommendation is to make that distinction visible in operational checks. If a scheduled task cannot read its input, a warning may explain the failure, but it should not satisfy the original request. Check the delivered artifact or response alongside the terminal status. Do not treat the presence of any chat text as acceptance.
Authentication gets a stopping point
The authentication deadline change, merged on 31 October, addresses credential lookups or runtime token exchanges that never return. Authentication now stops at the existing run deadline or cancellation, with an error identifying the provider and authentication step. It introduces no new settings or credential storage.
There is an important limit. The implementation record says JavaScript cannot forcibly stop a plugin that ignores cancellation. Instead, a late result is rejected before credential installation. We would therefore check both the user-visible timeout and the behaviour of the particular authentication plugin, rather than assuming a cancelled request proves all associated work has stopped.
Configuration reads and workspace transfers get separate repairs
The configuration redaction PR, merged on 1 November, fixes inline environment credentials being exposed when their names did not match a secret-name heuristic. CLI and Gateway configuration reads now redact inline environment values regardless of their names. Saving the displayed configuration preserves the real values through the existing restoration mechanism.
Raw source is withheld when escaped contents cannot be safely redacted, while structured configuration remains available. The record describes an escaped-string leak found during review and addressed before the final candidate. This is a specific correction to configuration output, not evidence that arbitrary logs or copied diagnostic files contain no secrets.
A separate workspace-transfer repair, also merged on 1 November, handles certificate fingerprints containing only whitespace. Blank fingerprints now use standard certificate-authority verification. Explicit correct pins still accept self-signed peers, while wrong pins reject them before credentials or workspace bytes are sent.
The transfer PR describes synthetic fixtures and credentials, and explicitly avoids claiming a demonstrated CLI interception. Its value as evidence is the identified validation defect and the reported rejection behaviour. We would not turn that into a story about an observed attack or a claim that all network connections received the same repair.
A surviving scheduler does not mean a completed job
The queued-cron repair, merged on 31 October, addresses the Gateway exiting when scheduling resumes and a queued run encounters a non-transient storage error. The changed background path observes and logs the rejection. The request remains queued and the Gateway stays available; callers awaiting a manual cron run still receive the storage error.
The PR reports before-and-after checks using isolated runtime builds and a deliberately injected SQLite storage rejection. It labels that controlled fault injection, not organic corruption or a full-disk incident. Our operational takeaway is equally narrow: check whether the job actually left the queue and produced its result. Gateway availability alone does not answer that question.
Where these distinctions matter in real workflows
Two older directory accounts help explain the stakes without testing this release. The February DevClaw workflow describes independent development and QA workers, with failed reviews returning to development. That is an author's deployment account, not an independent evaluation of November's fixes. For a similar setup, our recommendation is to inspect each worker's result and the parent's final delivery separately.
Simon Roses Femerling's February network-monitoring account describes recurring scans, comparisons with previous state and Telegram anomaly summaries. The directory distinguishes the configured workflow from illustrative commands and sample alerts. Neither historical account proves that 2026.11.3 fixes a particular installation. They give operators concrete places to check: private worker output, unanswered requests and scheduled results, rather than a chat window that merely looks active.