← Back to OpenClaw News A layered paper-cut illustration of a library vault at dusk where a sealed access door glows, threads of chat bubbles flow into knowledge shelves, and stage spotlights shine on a row of empty chairs
November 21, 2026Release WatchSecurityEnterpriseSkills

OpenClaw Seals a Claude CLI File-Access Gap, Teaches Knowledge to Read Slack and Spotlights Microsoft Maintainers

No new release on Friday, but main changed who can do what inside a session: permission modes now hold for Claude CLI chats, and invited collaborators can only tighten your guardrails. Knowledge learned to read Slack, Node.js patch levels deserve a check, and the Foundation asked what a free OpenClaw really costs.

No new OpenClaw release shipped on Friday. When we checked at 6:28 AM Saturday, npm still listed 2026.11.6 as latest, 2026.11.10-beta.1 as beta and 2026.10.33 as extended-stable, and the GitHub releases list had nothing newer. Main kept moving, though, and two of Friday's merges change who can do what inside a session. One closes a gap that let Claude CLI chats read files outside their workspace. The other stops invited collaborators from loosening the guardrails you set. Knowledge also learned to read Slack. Outside the code, OpenClaw's Foundation published an essay on what the "free" project really costs, and the official account thanked Microsoft's team for giving back as maintainers. Below we also cover a Node.js patch check, the Slack skill on ClawHub, and smaller main changes. All times are Pacific, and none of these merges is in a release yet.

Permission Modes Now Hold for Claude CLI Chats

Pull request #202941, which enforces the session permission mode on every Gateway tool call, merged at 10:05 PM Friday. A chat set to Read only, Ask me or Auto-review is supposed to keep file tools inside its sessionRoot. On the built-in runtime it did. A Claude CLI chat is different: the CLI calls OpenClaw's own tools, such as mcp__openclaw__read, through the Gateway, and the Gateway applied the session's mode only to restricted senders. The PR states the result plainly: "An owner's Claude CLI chat in workspace mode read files anywhere on the host."

The fix makes the Gateway apply the session's permission policy for every caller. The PR's evidence table shows the before and after on an isolated Gateway. Asked to read a file outside the workspace, the Claude CLI agent on the merge base returned the file's contents, while the patched branch refused with "Path escapes sandbox root." The built-in ChatGPT agent refused in both cases. Sessions with no mode set, such as cron runs, Skill Workshop and subagents without a mode, behave as before.

The rule is now written into the OpenClaw permission modes documentation: "Whoever runs a tool enforces the mode for it." The page adds two limits worth reading twice. A CLI backend's own native tools, such as Claude's own tools in a Claude CLI chat placed on a node, follow that CLI's permission controls, not the session mode. And "Exec itself is not confined to sessionRoot, except on a runtime that sandboxes its commands." The PR notes that exec behavior on Claude CLI chats is unchanged and that the remaining difference is tracked separately. We found no security advisory for this; the newest one in the repository is still the GHSA-pw3x-g8v8-fvj2 delegation advisory we covered on Thursday.

SEN-X Take

If your team runs Claude CLI as an OpenClaw backend and relies on Read only or Auto-review to keep agents out of the rest of the disk, assume that boundary did not hold for OpenClaw's own file tools until this lands in a release you run. Until then, keep those agents on hosts or containers that hold nothing sensitive. Also note that a permission mode was never a sandbox for shell commands.

Invited Collaborators Can Only Tighten Your Guardrails

Earlier Friday, at 5:48 PM, pull request #201615, which keeps invited collaborators from loosening a session's guardrails or approving its commands, merged. The problem it describes: someone invited into a session could change its permission mode to something looser than the owner set, "for example from Ask me to Auto-review," and approve shell commands the owner's own turns were waiting on. In the PR's words, "Inviting a collaborator could quietly remove the guardrails you chose."

Now, for anyone whose access comes from an invitation, permission mode is tighten-only, and they can see and resolve only the approvals raised by their own turns. Session creators, admins, team roles with broader session rights, iOS approvers and older sessions without a recorded creator keep the behavior they had on main. The PR adds a schema 41 database migration for SQLite and PostgreSQL that records which profile raised each approval, and it reports a 137-cell test matrix in which 41 cells failed against an earlier, broader version of the change and all 137 now pass.

SEN-X Take

Multiplayer sessions are only safe if the person who set the guardrails stays in charge of them, and this change finally says so in code. Teams that invite contractors or clients into live sessions should plan to move to the first release carrying it, and in the meantime treat an invitation as handing over approval rights.

Security Tip: Check Your Node.js Version Against November's Patches

  • Know the new patch level. The Node.js November 2026 security releases bulletin covers the 26.x, 24.x and 22.x lines, with fixes shipped as Node.js 26.13.1, 24.22.3 and 22.23.5. The bulletin lists 16 CVEs, seven rated high, including HTTP/2 and HTTPS memory bugs that an unauthenticated client can trigger. We have not traced which of them a given Gateway setup can reach, so the safe course is simply to patch.
  • OpenClaw's floor is lower than that. The OpenClaw Node.js runtime documentation says it requires Node 24.16.0+ or 26.1.0+, and it recommends 26.12.1 or newer, or 24.22.1 or newer on 24.x, because those releases "reject the keys compromised in the October 2026 eMudhra incident." Older compatible versions "remain accepted"; startup prints one security warning, and Doctor, updates and security audit also report the risk. Those recommended versions predate the November bulletin, so meeting OpenClaw's recommendation is not the same as being current.
  • Upgrade Node first. Run node -v on every Gateway host and move to 26.13.1 or 24.22.3. The same docs page says to "Upgrade Node before updating OpenClaw," and Node 22, 23 and 25 are unsupported. If a managed service runs the Gateway, check the Node it records, not just the one on your PATH.
  • Re-check modes after the next update. Once a release carries pull request #202941, open your Claude CLI agents and confirm the permission mode you expect. Tasks that used to read files outside the session root will start failing with "Path escapes sandbox root," which is the fix working, not a regression.

Knowledge Learns From Slack

At 11:39 AM Friday, pull request #202411, which lets Knowledge learn from Slack, merged. Knowledge could already learn from conversations, Granola meetings and Memory. The PR adds Learn from Slack to the Scribes page: once a Slack account is connected, the owner can run it once or on a schedule, and each run saves "the substantive discussions (decisions, commitments, plans, project news)" from the channels the Slack bot is in, with sources. It reads only, and "never sends or changes anything in Slack."

Two opt-ins are off by default: using the owner's Slack user token to read every channel that token can access, and including direct messages and group DMs. Each run re-reads a three-day window to pick up edits and revisits threads with replies in the last 90 days. Snapshots are always marked partial, so a message the scribe did not see is never treated as deleted. It depends on pull request #202410, which lets agents list Slack channels and choose which token reads them, merged six minutes earlier. That PR adds a readAs option of bot or user, and leaving it out keeps the old behavior.

SEN-X Take

For most teams, Slack is where decisions actually get made, so this may be the most useful Knowledge source yet. Start with the bot-only default and add the bot to a few project channels on purpose. Turning on the user token or DMs means an agent's knowledge base can include private conversations, which is a policy decision, not a settings tweak.

Skill Card: Slack by Peter Steinberger

If Knowledge will be reading Slack, it is worth knowing the skill that lets an agent act there. The Slack skill page on ClawHub, published by @steipete and marked official, says to use it "when you need to control Slack from Clawdbot via the slack tool, including reacting to messages or pinning/unpinning items in Slack channels or DMs." When we checked, it showed version v1.0.0, about 54,900 downloads, 160 bookmarks, a security audit result of "Pass," an MIT-0 license and a last update two months ago.

  • What it does. The SKILL.md lists five action groups, all enabled by default: reactions, messages (read, send, edit and delete), pins, member info and the custom emoji list. It uses "the bot token configured for Clawdbot," as the page puts it, and its examples include reading the last 20 messages in a channel and pinning key decisions.
  • Install the right one. ClawHub's API reports two skills with the slug slack, so a bare name is ambiguous. Use the full reference from the listing: openclaw skills install @steipete/slack.
  • Our caveats. Unlike the read-only Knowledge scribe, this skill can send, edit and delete messages as your bot. If you only want an agent to learn from Slack, the scribe is the narrower tool. Before installing it, check which channels your bot has joined, since that is the reach you are handing the agent. We read the listing; we did not install it.

Also on Main: Skill Workshop, Templates, Steering and OpenCode

Several smaller merges are worth a line each. Pull request #202572, which gives Skill Workshop a start page with Drafts, History and Learned skills, merged at 10:37 AM Friday. The page opens on Drafts when drafts are waiting, and the page notes that applying an update saves the current version first, so it can be undone. Pull request #202773, which lets an agent draft a Template from a description, merged at 3:23 PM. Before it, a request to build a new Template sent the agent off for about thirteen minutes of hand-building, because the bundled skill only knew how to share the current agent. Nothing is published unless the person chooses Publish.

Pull request #202397, which lets host integrations steer an active conversation without starting another run, merged at 8:54 AM. It is strictly off by default, admitted only when gateway.sessionControl.steeringEnabled is explicitly true. At 9:58 PM, pull request #202949, which adds browser sign-in for OpenCode Console and OpenCode Go, added a device-code browser sign-in; until then, both providers offered only an API key field, which still works. Each provider signs in separately because OpenCode refresh tokens are single-use. Earlier, at 7:11 PM, pull request #202815, which stops recursive Git shims from exhausting agent sandboxes, fixed a loop where a user's forwarding shim could re-enter OpenClaw's Git wrapper until a sandbox hit its process and memory limits. And at 6:18 AM Saturday, pull request #203044, which opens shared sessions in their source Control UI, added an opt-in "Open original" menu item for Session Share on paired nodes.

Community: What "Free" Costs, and Microsoft Maintainers

At 1:03 PM Friday, the official OpenClaw post on X introducing Joe Pishgar's essay asked: "If OpenClaw is so powerful, why is it free?" The answer, in the "OpenClaw Isn't Free" essay on the OpenClaw blog, is that the community pays in time. Pishgar, listed on the Foundation's people page as Community and DevRel, cites 3,606 contributors, 83,000 forks and 279 releases, and estimates about 120,000 hours of human labor, "nearly 58 years of full-time work." The methodology note is candid that the activity counts come from public data while "the hours per activity and work outside GitHub are modeled assumptions, not recorded timesheets." It starts from 52,929 merged pull requests over 12 months, as counted by GitRanks.

Later, at 4:20 PM, the official OpenClaw post on X about Microsoft building on OpenClaw said: "Love seeing Microsoft build on OpenClaw, and even more seeing your team give back as maintainers." We checked who that means. OpenClaw's CONTRIBUTING.md points to the OpenClaw Foundation people page for the current team, and its Core Maintainers section lists five people with Microsoft as their affiliation: Eduardo Piva, Galin Iliev, Gio Della-Libera, Omar Shahine and Scott Hanselman. The same list includes maintainers from NVIDIA, OpenAI, Atlassian, Tencent, Red Hat, Hugging Face and Xiaomi. The page does not say when each person joined.

SEN-X Take

The essay's hour count is a model, not a measurement, but the point holds for anyone building a business on OpenClaw: the project runs on volunteer time and vendor engineers. A maintainer list spanning Microsoft, NVIDIA and OpenAI is a good sign for staying power. It also means enterprise priorities will shape the roadmap, so smaller users should show up in issues and Discord to be heard.

What We Left Out

A reply on X from Peter Steinberger said people now "have to update" Node after a recent Node security issue. The Node range in package.json's engines field is identical in 2026.11.5, 2026.11.6 and main, and the docs describe a warning rather than a block, so we reported only what the docs say. We did not cover user posts listing 2026.11.10 beta features, since we reported that beta on Thursday and Friday. We also skipped a claim about GPT-6.2 Sol Ultrafast in Codex, a Zhejiang University social-media agent and a third-party agent skill that signs on-chain transactions, none of which we could tie to an OpenClaw source, and MyClaw 3.0, which is not an OpenClaw product.

What to Watch Next

Watch for the next beta or stable tag that carries pull requests #202941 and #201615, the first Knowledge vaults fed from Slack, and whether OpenClaw's Node.js recommendation moves up to the November patch releases.

Sharing OpenClaw Sessions Across a Team?

SEN-X helps teams set permission modes that actually hold, decide what Knowledge should read from Slack, and keep Gateway hosts patched before each OpenClaw upgrade.

Contact SEN-X →