← Back to OpenClaw News An abstract editorial illustration of a steady stable release platform beside a smaller beta platform, with a looping path of repeated steps being replaced by a single clear guide rail
November 16, 2026Release WatchSecurityAgentsSkills

OpenClaw 2026.11.4 Becomes the Stable Default as Beta 4 Lands and Main Replaces Its Tool-Loop Detectors

OpenClaw 2026.11.4 is the new stable release, with private Side chats and safer upgrades, and 2026.11.5-beta.4 follows with automation and rollback fixes. Main replaces its tool-loop detectors with one repeat guard, and a new bug report shows Telegram topic filters failing open in Active Memory.

OpenClaw closed the weekend with two releases within about 90 minutes. Version 2026.11.4 became the default stable build on Sunday evening, five days after 2026.11.3, and 2026.11.5-beta.4 reached the beta channel just before it. Overnight, main replaced its tool-loop detectors with one simpler rule, let Advisor review subagent runs, and started saving full background command output to a file. A newly reported Active Memory bug also means a Telegram group denylist does not cover that group's forum topics. All times below are Pacific.

2026.11.4 Becomes the Stable Default

The OpenClaw 2026.11.4 release notes on GitHub were published at 8:17 PM on Sunday, and npm now lists 2026.11.4 under the latest tag. The extended-stable tag is still 2026.9.37. The audited contribution record counts 99 pull requests since 2026.11.3. This is the release we have been tracking since November 13, when it entered qualification under a performance waiver.

The highlights are about staying up under load and keeping data private. Large conversations get bounded transcript trimming, paged memory exports and Doctor repairs that no longer exhaust worker memory. Reliable replies keeps unanswered prompts across Gateway restarts, stops unrelated reloads from swallowing WhatsApp final replies, and stops Signal reconnects from skipping buffered messages. When Telegram recovery cannot confirm a final reply was delivered, the Gateway now sends a notice saying so.

The privacy fixes are the ones admins of shared setups should read. Pull request #185390, which gives each viewer their own Side chat, fixes a case where everyone viewing a shared session saw the same Side chat, and one person's Clear side chat wiped it for all of them. Threads are now keyed to the signed-in requester, which the Gateway derives from the connection. Connections that share a Gateway token still share one thread. The notes also keep paired-node photos and screenshots private by default and keep internal model prompt projections out of display output.

Upgrades get safer too. Pull request #186971, which stops a newer CLI from migrating state an older Gateway still runs, covers a common split-install trap: a bare global npm install under a different Node could leave a newer CLI beside an older managed Gateway, and an ordinary config write could migrate shared state so that Gateway would not start. The CLI now refuses that migration and names both versions. Other fixes reject stale Files editor saves after a workspace is replaced, treat an empty IRC sender list as deny-all, and keep official plugins on the stable channel instead of drifting back to beta.

One fix many self-hosters want is not in this build. The automatic rollback of releases that fail Doctor or startup, which we covered on November 13, is not in the 2026.11.4 notes. It ships in beta 4 instead. The independent tracker ClawStat.us update verdict for OpenClaw 2026.11.4 rates the release "too new to call" as of Monday. It reports no credible blocker so far, lists Android chat nesting and ollama-cloud 400 errors among early complaints, and still names 2026.11.2 as its best version to run today.

SEN-X Take

2026.11.4 is a hardening release, and the Side chat fix alone justifies upgrading any Gateway that several people sign in to. Because automatic rollback is not in this build, take a backup before you update and confirm the Gateway starts cleanly before you walk away. If you use a global npm install, check that your CLI and Gateway now report the same version.

Beta 4 Follows With Automation and Recovery Fixes

The OpenClaw 2026.11.5-beta.4 release notes were published at 7:12 PM on Sunday, about a day after beta 3, and npm lists it under the beta tag. The record covers 55 pull requests since beta 3. Its Update safety group includes the rollback of releases that fail Doctor or startup, so beta testers get it first.

Automations get several fixes. Failed child results are delivered after parent turns end, late scheduled jobs run correctly, and in-flight jobs survive config changes. One privacy fix stands out. Pull request #198697, which rechecks owner status when a queued automation run starts, closes a gap where a run could receive owner-private memory after its requester lost owner status while the run waited for a free cron slot. Each run now rechecks ownership against the live config when it starts, and a former owner's run gets only its parent conversation's memory. Beta 4 also recovers OpenAI conversations from compaction item conflicts, retries newer Claude models with a compatible OAuth client version, and restores Bonjour discovery in npm installations.

Main Replaces Its Tool-Loop Detectors With One Rule

The biggest behavior change on main merged at 8:59 PM Sunday. Pull request #199591, which replaces the tool-loop detectors with an arguments-only repeat guard, swaps about 1,700 lines across 12 files for a single 105-line guard. The PR explains the old problem: the only detector that was always on ended the turn after three identical failing calls, so "the model got no chance to answer, and the user got a stop message in place of an answer."

The updated OpenClaw tool-loop detection documentation describes the new rule. The guard compares only the tool name and arguments, never results. On the third identical call in a row, the call runs and OpenClaw adds a note telling the model to change its approach. On the twentieth, the call does not run and the turn ends with the reason tool_repeat_limit. Tools that poll or wait by design, such as process poll and log, agents_wait and computer screenshots, are exempt. Plugin tools can declare repeatable themselves. The guard is on by default, tools.loopDetection.enabled: false turns it off, and openclaw doctor --fix removes retired keys such as warningThreshold and detectors.

SEN-X Take

This is a sensible trade. Stopping a turn after three failed calls was strict enough to hurt real work, while the smarter detectors were off for most installs anyway. A nudge at three and a hard stop at twenty is easy to explain to users. Plugin authors whose tools poll on purpose should add the repeatable flag before this reaches a release.

Advisor Now Reviews Subagents, and Long Output Goes to a File

Pull request #199676, which lets Advisor review subagent runs too, merged at 9:05 AM Sunday. Until now, Advisor skipped subagents. The PR notes they "can loop, poll finished work or drift," and the parent only sees the final report after the time is spent. Each subagent run is now reviewed in its own session with the same triggers and steering, with no new config. The Advisor plugin reference now says heartbeat and cron runs are still not reviewed, but subagents they spawn are. Under the Codex harness, Codex's own child threads remain out of scope.

At 10:29 PM, pull request #199992 started writing background exec output to a file named in the exit wake. A Gateway-host command now writes all its stdout and stderr to an owner-only file in the OS temp directory, up to 64 MiB. When the exit wake leaves output out, it ends with Full output: <path>, so the agent can read just the part it needs. The OpenClaw exec tool documentation says sandbox, node and durable commands keep no such file. With tools.fs.workspaceOnly, the agent has to read it with exec.

Also on Main

NVIDIA SkillEvaluator 0.9 Puts Numbers on Skill Quality

Outside the repository, NVIDIA's SkillEvaluator 0.9.0 release on GitHub, published at 3:03 AM Friday, adds evaluation datasets, version comparison with --compare-to, token and cost budgets, reusable no-skill baselines and opt-in GitHub and GitLab CI setup. The Apache-2.0 tool runs an agent on the same tasks with and without a skill and reports the difference as Skill Lift. NVIDIA's August technical blog post on evaluating agent skills with SkillEvaluator says "OpenClaw is piloting SkillEvaluator for official organizations on ClawHub," with results shown in an Evals tab. A post on X said NVIDIA tested 58 production skills and found about one in four did not help. We could not find that figure in NVIDIA's release notes or blog post, so we leave it out.

SEN-X Take

Security scans tell you a skill is safe to run. They do not tell you it is worth the context it uses. A with-and-without comparison is the right test, and teams with internal skill libraries can run it in CI today instead of waiting for ClawHub's pilot to widen. Start with your most-used skills, since a weak skill there wastes tokens on every turn.

Security Tip: A Telegram Group Denylist Does Not Cover Its Topics Yet

If you use Active Memory's deniedChatIds to keep a Telegram group out of memory recall, check whether that group uses forum topics. GitHub issue #199704 on Active Memory chat filters and Telegram forum topics, opened at 9:50 AM Sunday against 2026.11.3, shows that a topic session resolves to <chatId>:topic:<N>, not the group's chat id. As a result, a group id in allowedChatIds silently skips every topic, and a group id in deniedChatIds blocks only the root chat, so "every topic still runs Active Memory." The reporter calls this a fail-open for a privacy denylist. They saw the allowlist case on a live Gateway and derived the denylist case from the same code. The issue carries impact:security and P1 labels. A community fix PR, #200028, was closed unmerged at 12:36 AM Monday, and a maintainer said they would handle the issue directly. It was still open when we checked.

  • List topics, not just groups. Until a fix ships, add each <chatId>:topic:<N> you want blocked to deniedChatIds, and recheck when someone creates a new topic.
  • Prefer an allowlist for sensitive groups. The Active Memory session policy documentation says a non-empty allowedChatIds narrows every chat type, and that Active Memory skips a turn when it cannot resolve a conversation id. With this bug, an allowlist fails closed: topics are skipped, not exposed.
  • Watch the issue. The docs say the denylist "always wins," which is the behavior you want once topics resolve to their parent group.

Skill Card: OpenClaw Memory Audit

For teams thinking about what ends up in agent memory, the OpenClaw Memory Audit skill on ClawHub, published by shingo0620, scans a workspace and its memory logs for leaked API keys and tokens. Version 1.1.2 is MIT-0 licensed, shows a "Security audit Pass" on ClawHub, and lists about 2,200 downloads. It was first published in February 2026.

  • Local and read-only. Its SKILL.md says it runs a bundled scan_secrets.py over text files with python3, needs no provider credentials, and does not edit configs.
  • What it looks for. OpenAI keys including project keys, Telegram bot tokens, JWTs, AWS credentials and generic 32-plus-character secrets.
  • Schedule check. It reads cron.list() and recommends a weekly audit if none exists. Its advice on a hit is to revoke the secret first, then redact the file and clear the session memory. We read the listing and SKILL.md but did not run the script.

What We Checked and Left Out

An X post said OpenAI now runs more than 500 persistent OpenClaw agents. That figure was already in OpenAI's November 12 post, which we covered on November 13. We did not revisit Hermes Agent's OpenClaw importer, because its migration guide describes an existing feature rather than a new one. We also skipped a vendor's NAS app update that mentions OpenClaw remote access, and a secondhand podcast account of how Meta's Muse began with OpenClaw, since we did not track down a primary source for either. No new GitHub security advisories were published for OpenClaw in the window.

What to Watch Next

First, whether early 2026.11.4 reports, such as the ollama-cloud output-limit errors, turn into a quick patch. Second, when automatic rollback and the queued-automation privacy fix reach stable through 2026.11.5. Third, whether the Telegram topic filter fix lands before more teams rely on denylists for sensitive groups.

Planning your move to OpenClaw 2026.11.4?

SEN-X helps teams stage OpenClaw upgrades with backups and health checks, review Active Memory and chat filters for shared groups, and measure which skills actually earn their place.

Contact SEN-X →