OpenClaw Tags 2026.11.5-beta.1 With Knowledge Scribes and a PostgreSQL Data Plane as Main Fixes the Way Back to Stable
OpenClaw 2026.11.5-beta.1 is tagged and on npm beta, a 10,222-PR train led by Knowledge Scribes, a PostgreSQL data plane and pull requests from chat. Hours later, main fixed updates off the beta. Also: Particle-1 arrives as a provider plugin, Stop now pauses the queue, a symlinked media path fix and the model-usage skill.
2026.11.5-beta.1 Is Tagged, Signed and on npm's Beta Channel
Yesterday we said we'd watch for a v2026.11.5-beta.1 tag. It's here. The OpenClaw 2026.11.5-beta.1 prerelease on GitHub is marked as a prerelease and points at release commit bc061f9. Through the GitHub API, the annotated tag shows a verified signature from maintainer Dallin Romney, created at 5:52 PM Pacific on Tuesday, November 3. The package reached npm about two hours later. The npm registry metadata for the openclaw package records 2026.11.5-beta.1 at 8:01 PM Pacific and now has the beta tag pointing to it. latest stays on 2026.10.10 and extended-stable on 2026.9.36. The GitHub release page went public at 8:51 PM Pacific.
This is a big first beta, as the jump from 2026.10.10 suggests. The tag-pinned 2026.11.5-beta.1 contribution record says it covers "10,222 in-range PRs," after excluding 993 pull requests that already shipped in 2026.10.10-beta.5. The release notes list eight highlights:
- Knowledge Scribes: "first-class knowledge capture and reusable scribe workflows across agents, sessions, and the Control UI."
- A PostgreSQL data plane: PostgreSQL-backed state, migration tooling and compatibility coverage for existing deployments.
- Control UI: an overhaul of mobile and desktop navigation, layouts and interaction details.
- Pull requests from chat: preparing, publishing, reviewing and checking the status of PRs directly from a conversation.
- Cross-device and cloud execution: wider attached-device, cloud-placement and cyber-access workflows, "including safer capability handoffs."
- Decision and media models: specialized decision routing and media-capable model support, including ClawRouter integrations.
- Plugin experience: better plugin discovery, installation, authoring and runtime integration.
- Backup and upgrade safety: "clearer state preservation and rollback behavior" across backup, recovery and upgrade paths.
The release notes don't add much detail beyond these one-line summaries. Each highlight links to a handful of pull requests, so plan to read those before relying on a feature.
Treat this beta as a preview of where OpenClaw is going next, not as an upgrade. More than ten thousand pull requests in one train means a long tail of bugs, and the first one has already turned up (below). If you want to test Knowledge Scribes or the PostgreSQL data plane, do it on a spare Gateway with copied state. Keep production on 2026.10.10, which only became stable on Monday evening and is still the npm default.
Beyond the Headlines: Durable Automations, Guests and a Bounded Doctor
The Changes section has the pieces most teams will notice day to day. Automations gain "persistent streams and richer execution context for durable scheduled work." Sharing gets wider session sharing, guest access and "visitor lifecycle controls." Learned skills get better capture, reuse and guided authoring. Model routing becomes more provider-aware. Doctor has broader checks and automated repair guidance for migrated installations, and memory gets "safer memory cleanup and maintenance flows."
Two fixes in the notes are worth calling out. One stops hosted Doctor runs from stalling live sessions "by bounding worker lifetime and cancellation." The other stops aborted data: URL attachment previews in the Control UI from triggering a native fetch crash. The notes also list fixes to reply threading and deduplication, Talk reconnects, Codex session lifecycles and provider discovery, including restoring Claude CLI model discovery "when no Anthropic model is pinned."
The release-verification section links the npm package, tarball integrity hash, release SHA and CI evidence. It also records one gap: "npm Telegram beta E2E: not supplied." In other words, the end-to-end Telegram test that usually backs a beta wasn't attached to this one.
Guest access and visitor controls carry the most risk on this list, because they change who can see a session. Before trying them, write down who should reach each shared session and how that access ends, then check the beta against that list. If Telegram is a main channel for you, the missing end-to-end proof is a reason to test Telegram delivery yourself before trusting the beta with real conversations.
The First Beta Fix: Moving Off 2026.11.5-beta.1
As with 2026.10.10 yesterday, the first problem showed up in updating away from the new build. Pull request #192546, which fixes beta updates failing during isolated candidate startup, merged on main at 4:25 AM Pacific on Wednesday. It says an update from the published 2026.11.5-beta.1 to a compatible candidate "fails at isolated Gateway startup because the copied config retains the beta writer stamp." The candidate in that case carries the 2026.10.10 package label, and the PR adds that the original refusal "was metadata-only, with no demonstrated database data-risk boundary."
The fix is narrow. The Gateway accepts a strictly valid config snapshot only during ordinary foreground startup inside the update's private rehearsal namespace. Partial namespaces, invalid config, force and reset still refuse, and database schema protection is unchanged. The OpenClaw updating guide on docs.openclaw.ai now carries the same contract, and says "the candidate Gateway must contain this repair; the installed updater needs no change."
A second update fix landed nearby. Pull request #192699, which lets the updater recover when memory inspection storage is full, merged at 3:34 AM Pacific. After a migration, Memory Core couldn't allocate a temporary SQLite snapshot and blocked update readiness. In the PR's failing test, 16,384 bytes were needed and 1 was available. Now, running out of temporary space produces a warning, while "confirmed provider failures and source corruption still block."
The practical rule: install a beta only on a Gateway you're willing to keep on betas for a while. The way back depends on the next candidate containing #192546, and no published release has it yet. Take a verified backup before you try 2026.11.5-beta.1, since the updating guide says automatic config copies "are not a full-state backup," and check free disk space on the state volume before any update.
Particle-1 Arrives as a Provider Plugin
The day's official announcement wasn't about the beta. The official @openclaw post announcing Particle-1 in OpenClaw, published at 7:18 AM Pacific on Tuesday, says Particle "reports ~45% fewer output tokens than DeepSeek V4.1 Flash with comparable benchmark accuracy," promises "zero data retention for standard inference," and tells users to update to v2026.10.10 to try it.
The support is a plugin, not a core change. Pull request #187135, which adds the Particle model provider, merged on October 30 UTC. It adds @openclaw/particle-provider with PARTICLE_API_KEY authentication and lists an account's models live, "so new Particle-hosted models need no OpenClaw release." The Particle provider page on docs.openclaw.ai gives the setup: install the plugin with openclaw plugins install @openclaw/particle-provider, run onboarding, then select particle/particle-ai/particle-1. It lists a 1,048,576-token context window, up to 131,072 output tokens, text and image input, function tool calling, and high reasoning by default. The page also says Particle "does not advertise schema-guaranteed structured output."
Particle's own Introducing Particle-1 post from the Particle team, dated October 28, says the model is post-trained on DeepSeek V4.1 Flash and that this first release focuses on software engineering. Across four benchmarks, it reports 28.4% to 52.2% fewer output tokens and 34.4% to 51.6% lower average cost per task than the base model, with pass rates "within the confidence intervals of the base model." Prices differ between the two sources. The OpenClaw docs list $0.25 input and $1.00 output per million tokens as catalog estimates checked October 29, while the blog's benchmark footnote uses $0.30 and $1.20 for both Particle-1 and DeepSeek V4.1 Flash.
Particle-1 is worth a trial for coding agents that run long and cost a lot, but test it on your own work. The benchmark savings are Particle's claims, about Particle's tasks. Confirm the zero-retention terms in Particle's own documentation before sending it sensitive code. Avoid it for workflows that depend on strict JSON Schema output, and give the API key to the Gateway process alone, not to every shell on the host.
Overnight on Main: Stop Means Stop, and Confirmations Ask Real Questions
Main kept its usual pace. GitHub's list of openclaw pull requests merged in the past day shows 492 pull requests merged between 6 AM Pacific Tuesday and 6 AM Wednesday, and 130 of them have "improve(ui)" in their titles, part of a broad rewrite of Control UI wording. Three changes stand out.
Pull request #192250, which pauses queued messages after Stop, fixes a surprising behavior: pressing Stop while messages sat in the composer queue used to send the next queued message as soon as the aborted run ended. Now Stop parks those messages behind a notice that reads "Queue paused · You stopped the run," with Clear and Resume buttons. The PR notes the pause lives in memory per tab, so a reload doesn't keep it.
Pull request #192012, which asks before destructive actions with clear buttons, moves every Control UI confirmation into one in-app dialog. The title asks the question, the body states the consequence, and the buttons name the action. The skill library and personal instructions also stop using the browser's native window.confirm, which the PR says "answers silently in webviews without a dialog bridge."
Pull request #192369, on subagent tool denies for visible spawned sessions, merged at 6:02 AM Pacific. Sessions created with sessions_spawn and visible: true were getting subagent default tool denies, which on the Codex harness could leave them "with no shell or file tools at all." They now behave like the interactive sessions they are. The PR adds that inherited and delegated policy still apply, "so a visible spawn from a restricted turn stays restricted."
The Stop fix matters most for anyone who queues several instructions and then changes their mind. Until it ships, clear the queue before pressing Stop. The visible-spawn change loosens defaults on purpose, so after it ships, check which agents can spawn visible sessions. They'll get the parent's shell and file access, so the parent's tool policy becomes the real boundary.
Security Tip: Don't Trust a Symlinked Path to Pick the File for Media Inference
A bug fixed overnight is a reminder that a path and the file behind it can differ. Issue #192717, on audio and video inference reading the wrong file through a symlink parent, reports that on Linux openclaw infer audio transcribe --file 'link/../memo.wav' "can successfully transcribe a different file from the one the operating system opens for that path." infer video describe had the same problem. The shared CLI adapter collapsed link/.. on paper, while the operating system followed the symlink first. In the reporter's test, the backend received the bytes of a decoy file and the command still exited 0.
Pull request #192718, which stops media inference reading the wrong file through symlink parents, merged at 3:42 AM Pacific. It resolves the real parent directory before dispatch and keeps the existing media-root and safe-open checks. Final-file links outside the allowed parent were already rejected before any backend read, and still are. The PR describes it as a path-selection fix with "no new permission owner" or authorization bypass.
- Until a release carries #192718, pass plain absolute paths to
infer audioandinfer video, with no..after a symlinked directory. - If an agent or script builds these paths from user input or from directories other people can write to, resolve them first, for example with
realpath, and check the result. - Don't let untrusted users create symlinks inside directories your agents read media from.
- If transcripts or descriptions don't match what you expected, compare the hash of the file you meant to send with the one the backend actually read.
Skill Card: model-usage, for Seeing Which Model Spends Your Budget
With a new low-cost model in the picker, this week's skill is about measuring spend. The bundled model-usage SKILL.md in the OpenClaw repository describes it as a way to "summarize CodexBar local cost logs by model for Codex or Claude, including current or full breakdowns." The model-usage listing by Peter Steinberger on ClawHub installs it with openclaw skills install @steipete/model-usage.
- What it needs. The
codexbarcommand, which the skill offers to install from thesteipete/tap/codexbarHomebrew formula on macOS and Linux. Without it, the bundled Python script reads exported CodexBar JSON through--input. - Two modes.
--mode currentpicks the most expensive model in the most recent daily row, and--mode allgives a full per-model breakdown. Add--format json --prettyfor output a script can use. - Local data. Its documented commands read CodexBar's local cost logs, or a JSON file or stdin you pass in, and print a summary.
- The limits. It covers Codex and Claude only, and per the skill, "values are cost-only per model; tokens are not split by model in CodexBar output."
Run model-usage for a week before moving any work to a cheaper model, so you know which model your Codex or Claude spend actually goes to. That baseline is what makes a Particle-1 trial meaningful. Keep in mind that this skill won't see Particle traffic, so track that spend from Particle's own billing. Then compare like for like: the same kind of tasks, with cost per finished task rather than cost per token.
What to Watch Next
First, a 2026.11.5-beta.2, and whether it picks up the #192546 fix for moving off beta.1. Second, the first 2026.10.10 patch release, since several fixes on main, including the symlinked media path fix, haven't shipped in a stable build yet. Third, whether Particle's promised model updates show up in OpenClaw's live model list without a new release, as the provider PR says they should. We'll recheck the tags and npm on our next run.
Testing the 2026.11.5 beta?
SEN-X helps teams trial OpenClaw betas on staging Gateways, plan a safe path back to stable, and measure model costs before switching providers.
Contact SEN-X →