A lot of the recent additions to the GitHub Copilot ecosystem add real value for individual developers, yet they also expand the security surface that an enterprise has to reason about. Most of these new entry points let a developer pull executable instructions, configuration, or full processes from any random repository on the internet, with very little or no central control. This post looks at the five places where I think the gap between “useful for one engineer” and “safe to run across a 5,000 person org” is widest right now.

We’ll look at these topis:

• GitHub Copilot CLI plugin marketplace
• Agent Package Manager (APM)
• gh skill now in the GitHub CLI
• MCP servers across editors
• VS Code extensions and the different registries

Photo by Samuel Regan-Asante on Unsplash

GitHub Copilot CLI plugin marketplace

The Copilot CLI lets you register a marketplace of plugins and install plugins from it. The on-ramp is one command:

copilot plugin marketplace add OWNER/REPO
copilot plugin install some-plugin@some-marketplace

A marketplace is just a GitHub repository with a marketplace.json file in .github/plugin/. There is no review, no signing, no central index. Two marketplaces (copilot-plugins and awesome-copilot) are registered by default, but any user can add any other repo, including a personal fork or a newly created account that copies a real plugin name with a small change.

Versioning is the next gap. The CLI plugin reference has a version field in plugin.json, but the copilot plugin install command has no syntax to pin to a version. You install OWNER/REPO, OWNER/REPO:PATH, a Git URL, a local path, or plugin@marketplace, and you get whatever HEAD of the source happens to be at that moment. copilot plugin update NAME pulls latest. There is no lockfile, no SHA pinning of the kind gh skill has, and no provenance attestation. A marketplace can change a plugin’s contents without changing the version field, and the next update ships those changes to every developer who installed it.

Plugins themselves are executable assets.They sit in the directory the marketplace points at, get pulled to the user’s machine, and run in the user’s shell context with whatever permissions the developer has. That is the same context as their git credentials, their cloud CLI sessions, and any local secrets in their environment.

What is missing for an enterprise:

• No setting on a GitHub Enterprise or organization to restrict which marketplaces a Copilot CLI user is allowed to add. The MCP private registry policy that exists for VS Code does not cover this. I’d want at least to restrict this to repos under the organizations control.
• No way to require signed plugins, or plugins from a verified publisher.
• No audit trail on the GitHub side that tells you which plugins your developers installed and from where.
• Clear versioning out of the box. preferably with provenance signing build in.

If you compare this to how npm or PyPI are usually handled in a regulated org (a private proxy, an allowlist, a vulnerability scanner in the pipeline), the Copilot CLI plugin story today is roughly where npm was around 2014.

Agent Package Manager (APM)

Microsoft APM is a dependency manager for AI agent context. You declare an apm.yml, run apm install, and it pulls instructions, skills, prompts, agents, hooks, plugins, and MCP servers from any git host (GitHub, GitLab, Bitbucket, Azure DevOps, GitHub Enterprise) into every detected agent client on the machine.

The manifest lives in the repo itself (apm.yml and apm.lock.yaml are committed alongside the code, like package.json and package-lock.json). It looks like this:

dependencies:
  apm:
    - anthropics/skills/skills/frontend-design
    - github/awesome-copilot/plugins/context-engineering
    - microsoft/apm-sample-package#v1.0.0
  mcp:
    - name: io.github.github/github-mcp-server
      transport: http

APM does ship a governance story, and it is the most thought-through one in this post. There is apm-policy.yml with tighten-only inheritance from enterprise to org to repo, a published bypass contract, hidden-Unicode scanning on every install, lockfile integrity hashes, and an apm audit --ci mode that you can wire into branch protection.

The catch is that all of this governance is opt-in and lives outside of GitHub itself. A fresh install of apm on a developer laptop has no policy file. Policy is also pull-only: the canonical org policy lives at <org>/.github/apm-policy.yml and the CLI fetches it on demand when a developer or CI job runs apm install or apm audit --ci. There is no push, no agent, and no central enrollment. The fetched policy is cached locally for an hour by default in apm_modules/.policy-cache/, and a fetch_failure: warn|block knob decides what happens when the org repo is unreachable. The default is warn, which means an offline laptop with an empty cache resolves with no policy at all. Repo-local apm-policy.yml files can extends: org and only tighten the parent rules, never relax them. See the Policy Reference and Governance Guide for the full mechanics.

Until your security team writes one, publishes it, and gets it picked up on every machine, an apm install will happily resolve transitive dependencies from any reachable git host. And don’t worry: your CI/CD pipeline will do the same (if the tooling is already installed)!

There is also no auto-install. APM is purely a CLI; it has no editor extension that runs apm install when you open a repo in VS Code. The docs frame it explicitly as “same as npm install after cloning a Node project”, which means the install step relies on the developer running it (or a devcontainer postCreateCommand, or a CI job). The flip side is that the deployed files (under .github/, .claude/, .cursor/, .gemini/) are recommended to be committed, so a teammate who clones the repo gets the agent context immediately, before they ever run apm install. That is convenient and it also means the agent is reading APM-deployed content the moment the editor opens the repo, regardless of whether the local CLI was ever invoked.

APM packages can declare scripts (think npm scripts), and the policy reference exposes manifest.scripts: allow|deny precisely because of this risk. Default is allow. So an attacker who lands a package in your dependency tree can also land scripts, unless your org policy denies them outright.

Versioning is fine on the manifest side: dependencies pin with #tag or #sha, the lockfile records resolved commit SHAs and content hashes, and the org policy can require specific versions with a require_resolution of project-wins, policy-wins, or block. Updates happen on apm install --update, not implicitly. Direct and transitive resolution stay the parts I would worry about: a package you trusted six months ago can pull in a new dependency on its next release, and unless your org policy has a tight dependencies.allow pattern, the new source slips through.

The MCP integration is worth a separate paragraph. apm install --mcp NAME adds an entry under dependencies.mcp in apm.yml and writes the resolved server config straight into the native config file of every detected client (Copilot, Claude, Cursor, Codex, OpenCode, Gemini) on the filesystem, bypassing each client’s own registry or policy layer. The full mechanism is documented in the APM MCP Servers guide. Convenient for a developer; also a clean way around whatever per-client policy exists. You are then relying on the runtime side of those clients to apply policy, and only a few of them do, with workarounds.

gh skill now in the GitHub CLI

Note: this is the GitHub CLI, not the GitHub Copilot CLI!

The gh skill command lets you discover, install, manage, and publish Agent Skills from any GitHub repository:

gh skill install github/awesome-copilot documentation-writer
gh skill install some-user/some-repo some-skill --pin v1.0.0

The supply-chain features here are better. Skills can be pinned to a tag (unsafe) or commit SHA, the install records the git tree SHA in the skill’s frontmatter, gh skill update compares local SHAs against the remote, and gh skill publish will offer to enable immutable releases so that even a repo admin cannot rewrite a published version.

Audit tooling is thin. The gh skill manual lists install, preview, publish, search, and update as the only subcommands. There is a gh skill preview to inspect a skill’s content before installing, and gh skill update uses the stored tree SHA to detect drift, but there is no gh skill audit and no org-side audit log of what your developers installed. If you want to know which skills landed on a developer’s laptop, you have to grep the agent host directories yourself.

The thing that is not there is an org-level allowlist.GitHub itself is unusually direct about this in the changelog:

Skills are installed at your own discretion. They are not verified by GitHub and may contain prompt injections, hidden instructions, or malicious scripts. We strongly recommend inspecting the content of skills before installation.

So the tooling around a single skill is solid, the tooling around “which skills is my org allowed to use” is not. A developer can gh skill install from any public repo, and the agent host (Copilot, Claude Code, Cursor, Codex, Gemini, all the CLI options) will pick the skill up the next time it scans the directory. Skills are a first-class extension point for the agent’s behavior, which means a malicious skill is closer to a custom system prompt than to a passive config file.

Dependabot does not help here either: agent skills, MCP servers, APM packages, and Copilot CLI plugins are not on the list of ecosystems Dependabot supports. That means no automatic update PRs, no security advisories wired in, and no scheduled drift detection across these surfaces. You would have to build that yourself.

MCP servers across editors

This is the area where the situation has gotten more confusing rather than less, even though there has been real work on it. Back in March 2025 MCP exploded into the AI world: extensibility from anywhere into anything! Since then, a lot of servers and OSS repos turned out to be playing around with things. The hard part is that a large share of those repos have since been abandoned. Endor Labs covered this in its State of Dependency Management 2025 report (summary on the Endor Labs press release): more than 10,000 MCP servers were created in less than a year, 75% of them by individual developers rather than organizations, around 40% have no license at all, and 82% touch sensitive APIs. Maintenance signals on the long tail are weak, which means the same servers your developers happily installed last year may already be effectively orphaned.

A short summary of where MCP server config lives today:

• VS Code: .vscode/mcp.json (workspace), the user-profile mcp.json opened via MCP: Open User Configuration, or contributed by an installed VS Code extension. The full schema is in the VS Code MCP servers docs. APM sits on top of this: it stores MCP servers in the repo’s apm.yml, then writes them into the same .vscode/mcp.json file the editor reads. By default apm install does not overwrite locally-authored entries (that needs --force, per the CLI reference), so the file you end up with is APM’s set plus anything that was already there. If a developer thinks “I’m only running the APM-managed servers”, they are wrong: they are running APM-managed plus whatever they (or another tool) wrote into mcp.json previously.
• Cursor, Windsurf, Codex, Claude Code, Gemini CLI, Copilot and other CLI’s: each has its own file in its own location, with its own schema variations.
• The remote Copilot agents (Cloud Agent, Spark, Spaces, Review Agent) each have their own configuration surface and can only be configured by a repo admin.

GitHub did ship an MCP private registry policy for Copilot that lets an enterprise restrict which MCP servers Copilot users can connect to. Useful, and a real step forward, but at the time of writing it only applies inside Copilot in VS Code. The same Copilot identity used in JetBrains, Neovim, the CLI, Spark, Spaces, the Cloud Agent, or the Review Agent is not covered by that policy.

Two patterns make the policy easier to bypass than it looks:

1. Local stdio servers. The VS Code MCP docs describe three config paths: the gallery flow (which the Copilot private registry can gate), the workspace .vscode/mcp.json, and the user-profile mcp.json. The registry policy applies to the gallery flow. A developer who edits either JSON file directly gets a one-time “trust this server” prompt and the server starts. There is a separate VS Code device-management policy that can disable MCP entirely, but it is on/off, not allowlist-aware. See extension runtime security for the surrounding policy surface.
2. Extension-contributed servers. A VS Code extension can contribute MCP servers through its manifest. If an extension is allowed to install (and most orgs do not gate extensions tightly, see the next section), the MCP servers it contributes inherit the same trust as the extension itself. That sidesteps the registry policy entirely.

Even worse: clone the extension repo from github.com, build it, and just use the compiled VSIX file in VS Code!

So the practical state is: you can get a meaningful slice of governance for Copilot in VS Code if you set up the registry, and almost no governance for any of the other clients on the same laptop, all of which can reach the same internal systems. So we are not there yet, but at least a step in the right direction.

VS Code extensions and the registry split

The extension story is the oldest of the five, and it is the one that has changed shape most recently because of the Cursor and Windsurf-style forks. A few things to be explicit about:

• The Microsoft Visual Studio Marketplace is closed to non-Microsoft products by its terms of use. Any VS Code fork (Cursor, Windsurf, VSCodium, Kiro, Antigravity, Positron) cannot legally use it.
• Those forks generally point at Open VSX, the Eclipse Foundation registry. Open VSX has a smaller catalog, less aggressive abuse handling historically, and a publish flow that is easier to ride.
• On April 21, 2026 the Eclipse Foundation launched the Open VSX Managed Registry as an SLA-backed paid tier (99.95% uptime, defined support tiers), with AWS, Google, and Cursor as initial adopters. The launch numbers paint the scale: 300M+ downloads per month, 200M+ daily requests at peak, 12,000+ extensions, 8,000+ publishers. The community instance was being asked to do the job of always-on critical infrastructure, and the AI editors are most of the reason.

For an org this means the threat model differs by editor, even when the developer thinks they are installing “the same extension”. A name on the Microsoft Marketplace is not necessarily owned by the same publisher on Open VSX. Typosquats and copy-jobs of popular extensions show up regularly on both registries, and an extension is essentially arbitrary code in your editor process with access to your workspace files, your environment, and any tokens the editor holds.

The MCP angle ties back in here: an extension can contribute MCP servers, settings, and language model providers. So an extension that gets past your install policy can reintroduce all the things you tried to gate at the registry layer.

What helps in practice:

• The VS Code extensions.allowed and related policies, deployed through your endpoint management, so that only an allowlist of extensions can install at all.
• Mirroring Open VSX internally if you support fork editors, with a curated subset rather than a full passthrough.
• Treating new extension installs the same way you treat new npm dependencies: review, scan, and budget for the maintenance.

Endpoint protection is the layer that catches what the registries miss. Even the official VS Code documentation on extension runtime security is direct that an extension runs with the user’s full permissions: it can read and write any file the editor can, spawn processes, and make network calls. The Marketplace does scan packages and verify signatures (see the Microsoft post on security and trust in the Visual Studio Marketplace), but malicious extensions and credential-stealing supply chain incidents keep landing (see the Wiz writeup on supply chain risk in VS Code extension marketplaces and Check Point’s report on 45,000+ downloads of malicious extensions). For an org that means the controls have to live below the editor: managed device policy that blocks unsigned binaries, EDR that watches the editor’s process tree the same way it watches a browser, outbound DNS and TLS inspection that can flag the unusual call patterns an extension makes, and a workstation lifecycle that assumes a compromised editor is one of the realistic incidents you respond to. Third-party scanners like ExtensionTotal can give you a per-extension risk score before you allow it, but treat them as an addition to your endpoint stack rather than a substitute.

State of the plugin governance for GitHub Copilot

If I line up the different surfaces by how much org-level governance is actually possible today:

The pattern across all of them is that the per-developer experience is great, the per-org enforcement is either absent or has to be assembled from policies that live in different places than the feature itself. None of these are unfixable, and APM in particular shows what the right shape looks like, but the gap between “shipped” and “safe to deploy at scale” is wider than the changelog posts suggest.

If you are responsible for any of this in a larger org, the short version of what I would do:

1. Decide which of these surfaces you want your developers to use at all. Default-allow is a choice that has consequences, not a neutral starting point.
2. For the ones you allow, pick the strongest available control today (VS Code extension policy, the Copilot MCP registry, an APM policy file) and ship it.
3. For the ones with no control today (CLI plugins, gh skill), at minimum log and review, and feed back to GitHub and Microsoft that this gap matters. Overall, tighten your grip on endpoint protection and your firewall/proxy configurations.

The features themselves are fine. The missing layer is the one every package ecosystem has had to grow eventually: a place for an org to say which sources it trusts, applied uniformly across every client that can pull from them.

Over the past decade or two, companies have been rolling out data privacy policies that more or less get accepted and forgotten, similar to the way that most people roll past the EULAs that appear in every commercial installation process. With the rise of LLMs, though, and the opportunity to "fine-tune" a model based on data to create a customized model, I think suddenly companies are going to want to (and will) pay very very close attention to what a service provider can do with your data.

Remember when Google told us they were reading your mail? Google claimed that "all users of email must necessarily expect that their emails will be subject to automated processing", and that the process was fully automated, "and no humans read your email". That, by the way, was a decade ago, but the practice hasn't stopped--it turns out that they're doing it to train their AI models, too. As has been said before, "data is the new oil", and companies have been looking for places to drill for oil for quite some time.

But we've left an era where the data was just being extracted into numbers and statistics and stored in a database somewhere, and entered an era where the actual text can be visualized and "understood" by LLM models. What's more, the data can now be used to train new models to act in ways to "help" you (for whatever definition a corporate entity chooses that to mean) without your permission. Well, I mean, you do have to give permission, you have to agree to a blanket all-or-nothing statement that you must agree to if you wish to make use of the particular service. Even if that service is really more of a necessary utility of basic life in the 2020s.

Consider, for example, email. Considering how many corporations and governments have transitioned away from "snail mail" to its electronic counterpart for necessary communication with people, it's very arguably as necessary a utility for modern life as electricity, water, garbage, and postal mail ever was. Which brings us back to GMail again. For millions of people, this is the only steady email address they will ever have in their life. (Very few people ever actually spin up their own mail servers in the cloud, it turns out!)

So do individuals really have the option of opting out? If the privacy policy is essentially an all-or-nothing practice, and you need the service in order to... well, live... how much of a choice is there? This seems like an area ripe for consumer protection laws and pro-consumer activism, but that's a lead-up to my real point.

What happens to that document you uploaded?

Consider, for a moment, the LLM-powered OCR that you're using as-a-service from some service provider. Consider that your company, like so many others, considers your data to be confidential and private. Now consider the privacy policy that your OCR service provider demands you accept in order to make use of that service. Have you read it carefully? Have you had your Legal team review it?

Has everybody considered what the implications are if the service provider decides one day to make use of all those documents you uploaded to them for their own purposes?

Say, for example, that service provider decides to use all those documents you gave them (and I use that word deliberately--you handed them over without any legal right to restrict what the service provider could do with them based on what was in the privacy policy) to start, say, using them as training data on a different model? Or for benchmark purposes? Or as part of their test suite?

I think that, in the very near future, there's going to be a serious falling-out with companies that don't maintain very strict data privacy policies around these services. Companies which don't go "above and beyond" to demonstrate that they adhere to a strict "verbose opt-in" policy for using customer data as part of their training sets will find themselves facing some uncomfortable questions from customers who want to make sure that the images and documents they send as part of their everyday operations aren't being used after-the-fact.

In fact, I would probably suggest CISOs start--today--to begin an audit of every service provider in use by their development teams, in order to understand just how much of the company's confidential data is potentially being used by third parties, and for particularly sensitive industries (financial, legal, medical, to start), require explicit signed statements from those third parties that the documents being uploaded are not being used to train AI models, along with a binding agreement that any exercise that would even remotely constitute doing so in the future will first be accompanied by an explicit request for agreement to do so. (In other words, "Swear to me you're not using them now, and swear that I'll know it if you ever start to do so.")

"But it's all legal!"

Let me make my position here clear: I'm not weighing in on whether or not it's legal for service-providers to use those documents in whatever manner they see fit, for two reasons:

1. The current privacy policies are probably written to allow for such use, and
2. I am not a lawyer

... but the perception has always been that the data would be used for purposes that are quite some distance away from what a particular company might be concerned about. (And even then, lots of people have had concerns/objections about it already.) Perception matters almost as much or more than the legal liabilities, because while it's a ton of work for me to legally prove that you violated your legal obligation, it requires almost no work at all to drop you as a vendor and switch over to one of the (potentially) dozens of other providers, some of whom are likely willing to make those guarantees (and may even be doing so right now).

In a world where we are very likely staring an implosion within the tech industry around all these AI-based startups (and possibly the Magnificent Seven as well) and companies are going to start to seriously struggle with cash flow, the "Let's come up with a new product that does X"="But we'd need to train a new model"-"Oh, we can use customer data to do that, they already agreed to it" conversations in corporate meeting rooms are going to significantly accelerate in temptation. If anybody's Legal team is even remotely worried, now's the time to jump-start this conversation.

And meanwhile, if you're one of those service providers, you might want to give serious thought to putting some systemic restrictions on how your data scientists' or developers' access to customer data, and preemptively make those data privacy statements very loudly. Remove the temptation now.

Because, I think, when public opinion turns against "using my data to train your models", it's going to turn very hard, very fast, and companies that aren't on the right side of this will take some really bad PR blows, and you do not want your customers to have very good reasons to ditch your service when you're already scrambling to survive.

2026.5.2

Highlights

• External plugin installation, update, doctor repair, dependency reporting, and artifact metadata now cover the npm-first cutover, stale configured installs, missing package payloads, and beta-channel plugin fallback. Thanks @vincentkoc.
• Gateway and agent hot paths are leaner across startup, session listing, task maintenance, prompt prep, plugin loading, tool descriptor planning, filesystem guards, and large runtime configs.
• Control UI and WebChat are more resilient across Sessions, Cron, long-running Gateway WebSockets, grouped-message width, slash-command feedback, iOS PWA bounds, selection contrast, and Talk diagnostics.
• Messaging fixes cover WhatsApp Channel/Newsletter targets, Telegram topic commands and networking, Discord delivery/startup edge cases, Slack threads, Signal groups/media, and visible reply routing.
• Provider and media fixes cover OpenAI-compatible TTS/Realtime, OpenRouter/DeepSeek replay, Anthropic-compatible streaming, LM Studio reasoning metadata, Brave/SearXNG/Firecrawl web search, media paths, music, and voice-call routing.

Changes

• Gateway/startup and restart: skip plugin-backed auth-profile overlays during startup secrets preflight, reducing gateway readiness latency while keeping reload and OAuth recovery paths overlay-capable; add openclaw gateway restart --force and --wait <duration>, log active task run IDs before restart deferral timers, and report timeout restarts as explicit forced restarts. (#68327) Thanks @JIRBOY.
• Plugins/ClawHub: make diagnostics, onboarding, doctor repair, and channel setup carry ClawPack metadata through install records while keeping explicit clawhub: installs on ClawHub and bare package installs on npm for the launch cutover. Thanks @vincentkoc.
• Plugins/CLI: include package dependency install state in openclaw plugins list --json so scripts can spot missing plugin dependencies without runtime-loading plugins.
• Plugins/update: on the beta OpenClaw update channel, default-line npm and ClawHub plugin updates try @beta first and fall back to default/latest when no plugin beta release exists.
• Plugins/runtime: scope broad runtime preloads to the effective plugin ids derived from config, startup planning, configured channels, slots, and auto-enable rules instead of importing every discoverable plugin.
• Agents/runtime: reuse the startup-loaded plugin registry for request-time providers, tools, channel actions, web/capability/memory/migration helpers, and memoized provider extra-params, and memoize transcript replay-policy resolution for stable config and process-env runs while preserving model-specific transport hook patches and custom-env provider behavior. Thanks @DmitryPogodaev.
• Infra/path-guards: add a fast path for canonical absolute POSIX containment checks, avoiding repeated path.resolve and path.relative work in hot filesystem walkers. Refs #75895, #75575, and #68782. Thanks @Enderfga.
• Tools/plugins: add a platform-level tool descriptor planner for descriptor-first visibility, generic availability checks, and executor references, and cache plugin tool descriptors captured from api.registerTool(...) so repeated prompt-time planning can skip plugin runtime loading while execution still loads the live plugin tool. (#76079) Thanks @shakkernerd.
• Docs/Codex: clarify that ChatGPT/Codex subscription setups should use openai/gpt-* with agentRuntime.id: "codex" for native Codex runtime, while openai-codex/* remains the PI OAuth route. Thanks @pashpashpash.
• Plugins/source checkout: load bundled plugins from the extensions/* pnpm workspace tree in source checkouts, so plugin-local dependencies and edits are used directly while packaged installs keep using the built runtime tree. Thanks @vincentkoc.
• Plugins/beta: externalize ACPX behind @openclaw/acpx and diagnostics OpenTelemetry behind @openclaw/diagnostics-otel, keeping their heavier runtime stacks out of the core package until installed; prepare Google Chat, LINE, Matrix, Mattermost, BlueBubbles, diagnostics Prometheus, Google Meet, Nextcloud Talk, Nostr, Zalo, Zalo Personal, diagnostics OpenTelemetry, Discord, Diffs, Lobster, Memory LanceDB, Microsoft Teams, QQ Bot, Voice Call, WhatsApp, Brave, Codex, Feishu, Synology Chat, Tlon, and Twitch for 2026.5.1-beta.1/2026.5.1-beta.2 npm and ClawHub publishing, and keep publishable plugin dist trees out of the core npm package. Thanks @vincentkoc.
• Providers/xAI: add Grok 4.3 to the bundled catalog and make it the default xAI chat model.
• Google Meet: let API-created rooms set accessType and entryPointAccess, add googlemeet end-active-conference for closing managed spaces after a call, and add googlemeet test-listen plus the matching google_meet test_listen action so transcribe-mode joins wait for real caption or transcript movement before reporting listen-first health. (#74824; refs #72478) Thanks @BsnizND and @DougButdorf.
• Plugins/ClawHub/onboarding: prefer versioned ClawPack artifacts when ClawHub publishes digest metadata, verify ClawPack response headers and downloaded bytes, persist ClawPack digest/artifact metadata on install/update records and install-on-demand provider setup entries, and allow official bundled-plugin cutovers to record ClawHub artifact metadata while preserving npm as the launch default for bare package specs and retaining npm/local fallback paths. Thanks @vincentkoc.
• Plugins/Crestodian: add ClawHub plugin search plus Crestodian plugin list/search/install/uninstall operations, with approval and audit coverage for install and uninstall.
• Channels/thread bindings: replace split subagent/ACP thread-spawn toggles with threadBindings.spawnSessions, default thread-bound spawns on, and let openclaw doctor --fix migrate the legacy keys. (#75943)
• Providers/OpenAI: add extraBody/extra_body passthrough for OpenAI-compatible TTS endpoints, so custom speech servers can receive fields such as lang in /audio/speech requests. Fixes #39900. Thanks @R3NK0R.
• Channels/WhatsApp: support explicit WhatsApp Channel/Newsletter @newsletter outbound message targets with channel session metadata instead of DM routing. Fixes #13417; carries forward the narrow outbound target idea from #13424. Thanks @vincentkoc and @agentz-manfred.
• Dependencies: refresh workspace, bundled runtime, and plugin dependency pins, including TypeBox 1.1.37, AWS SDK 3.1041.0, Microsoft Teams 2.0.9, Marked 18.0.3, Pi 0.71.1, OpenAI 6.35.0, Codex 0.128.0, Zod 4.4.1, and Matrix 41.4.0. Thanks @mariozechner, @aws, and @microsoft.
• Discord/channels: add reusable message-channel access groups plus Discord channel-audience DM authorization, so allowlists can reference accessGroup:<name> across channel auth paths. (#75813)
• Crabbox/scripts: print the selected Crabbox binary, version, and supported providers before pnpm crabbox:* commands, and reject stale binaries that lack blacksmith-testbox provider support.
• Agents/Codex: add committed happy-path prompt snapshots for Codex/message-tool Telegram direct, Discord group, and heartbeat turns so prompt drift can be reviewed. Thanks @pashpashpash.
• Agents/workspace: add agents.defaults.skipOptionalBootstrapFiles for skipping selected optional workspace files during bootstrap without disabling required workspace setup. (#62110) Thanks @mainstay22.
• Plugins/CLI: add first-class git: plugin installs with ref checkout, commit metadata, normal scanner/staging, and plugins update support for recorded git sources. Thanks @badlogic.
• Google Meet: add live caption health for Chrome transcribe mode, including caption observer state, transcript counters, last caption text, and recent transcript lines in status and doctor output. Refs #72478. Thanks @DougButdorf.
• Voice Call/Google Meet: add Twilio Meet join phase logs around pre-connect DTMF, realtime stream setup, and initial greeting handoff for easier live-call debugging. Thanks @donkeykong91 and @PfanP.
• macOS app: move recent session context rows into a Context submenu while keeping usage and cost details root-level, so the menu bar companion stays compact with many active sessions. Thanks @Guti.
• Gateway/SDK: add SDK-facing tools.invoke RPC with shared HTTP policy, typed approval/refusal results, and SDK helper support. Refs #74705. Thanks @BunsDev and @ai-hpc.
• Discord: keep active buttons, selects, and forms working across Gateway restarts until they expire, so multi-step Discord interactions are less likely to break during upgrades or restarts. Thanks @amknight.
• Messages/docs: clarify that BodyForAgent is the primary inbound model text while Body is the legacy envelope fallback, and add Signal coverage so channel hardening patches target the real prompt path. Refs #66198. Thanks @defonota3box.
• Slack: publish a safe default App Home tab view on app_home_opened, include the Home tab event in setup manifests, and keep track of bot-participated threads across restarts so ongoing threaded conversations can continue auto-replying after the Gateway restarts. Fixes #11655; refs #52020. Thanks @TinyTb and @amknight.
• Control UI/Usage: add UTC quarter-hour token buckets for the Usage Mosaic and reuse them for hour filtering, keeping the legacy session-span fallback for older summaries. (#74337) Thanks @konanok.
• BlueBubbles: add opt-in channels.bluebubbles.replyContextApiFallback that fetches the original message from the BlueBubbles HTTP API when the in-memory reply-context cache misses (multi-instance deployments sharing one BB account, post-restart, after long-lived TTL/LRU eviction). Off by default; channel-level setting propagates to accounts that omit the flag through mergeAccountConfig; routed through the typed BlueBubblesClient so every fetch is SSRF-guarded by the same three-mode policy as every other BB client request; reply-id shape is validated and part-index prefixes (p:0/<guid>) are stripped before the request; concurrent webhooks for the same replyToId coalesce into one fetch and successful responses populate the reply cache for subsequent hits. Also promotes BlueBubbles attachment download failures from verbose to runtime error so silently-dropped inbound images are visible at default log level, and extends sanitizeForLog to redact ?password=…/?token=… query params and Authorization: headers before they reach the log sink (CWE-532). (#71820) Thanks @coletebou and @zqchris.
• CLI/proxy: add openclaw proxy validate so operators can verify effective proxy configuration, proxy reachability, and expected allow/deny destination behavior before deploying proxy-routed OpenClaw commands. (#73438) Thanks @jesse-merhi.
• Agents/Codex: default Codex app-server dynamic tools to native-first, keeping OpenClaw integration tools while leaving file, patch, exec, and process ownership to the Codex harness; default Codex-harness direct source replies to the OpenClaw message tool when visible reply delivery is not explicitly configured, keeping channel-visible output as a deliberate tool call. (#75308, #75765) Thanks @pashpashpash.
• Heartbeats/agents: add a structured heartbeat_respond tool for tool-capable heartbeat runs so agents can record quiet outcomes or explicit notification text without relying only on HEARTBEAT_OK parsing. (#75765) Thanks @pashpashpash.
• Gateway/config: allow $include directives to read files from operator-approved OPENCLAW_INCLUDE_ROOTS directories while preserving default config-directory confinement. Thanks @ificator.

Fixes

• Agents/OpenAI: default GPT-5 API-key sessions to the SSE Responses transport unless WebSocket is explicitly selected, restoring replies in fresh Control UI and WebChat beta installs where the auto WebSocket path connected but produced no model events.
• Agents/sessions: preserve terminal lifecycle state when final run metadata persists from a stale in-memory snapshot, preventing sessions from staying stuck as running after completed or timed-out turns.
• Gateway/CLI/status: make openclaw gateway start repair stale managed service definitions that point at old OpenClaw versions, missing binaries, or temporary installer paths before starting; add concrete service, config, listener-owner, and log collection next steps when gateway probes fail and Bonjour finds no local gateway; avoid repeated plugin tool descriptor config hashing so large runtime configs do not block reply startup and trigger reconnect/timeouts. Refs #49012. (#75944) Thanks @vincentkoc and @joshavant.
• Plugins/update/config: stop treating the non-plugin auth command root as a bundled plugin id, keep packaged upgrades and beta external plugin installs on stable runtime aliases and matching prerelease npm specs, detect tracked plugin install records whose package directories disappeared during openclaw update, reinstall them before normal plugin updates, fail the update if install records still point at missing disk payloads, and validate configured web-search providers plus statically suppressed model/provider pairs against the active plugin set at config load. Thanks @vincentkoc.
• Codex/app-server: resolve managed binaries from bundled dist chunks and from the @openai/codex package bin when installs do not provide a nearby .bin/codex shim, avoiding false missing-binary startup failures.
• Status: show the openai-codex OAuth profile for openai/gpt-* sessions running through the native Codex runtime instead of reporting auth as unknown. (#76197) Thanks @mbelinky.
• Status/update: resolve beta update-channel checks from the installed version when config still says stable, show configured channels in openclaw status and config-only openclaw channels status output even when the Gateway is unreachable, and let status --deep reuse live gateway channel credential state instead of warning on command-path-only token misses. Thanks @vincentkoc.
• Plugins/externalization: add official npm-first catalogs for externalized channel, provider, and generic plugins; install official external web-search plugins before saving provider config; repair missing configured, selected-search, and env-selected plugin installs from npm by default; keep official install docs, update examples, live Codex checks, diagnostics ClawHub packages, and persisted bundled-plugin relocation on default npm tags; and keep ACPX, Google Chat, and LINE publishable plugin dist trees out of the core package while ClawHub pack files roll out. Thanks @vincentkoc.
• Plugins/ClawHub/source/registry: use the ClawHub artifact resolver response as the install decision before downloading, keep bare plugin package specs on npm for the launch cutover and reserve ClawHub resolution for explicit clawhub: specs until ClawHub pack readiness is deployed, discover source-only plugins such as Codex from extensions/*, install ClawPack artifacts from the explicit npm-pack .tgz resolver path, persist artifact kind, npm integrity, shasum, and tarball metadata for update/diagnostics flows, fall back to version metadata when the artifact resolver route is missing, keep the Docker ClawHub fixture aligned with npm-pack artifact resolution, explain unavailable explicit ClawHub ClawPack artifact downloads with a temporary npm install hint, and hash manifest/package metadata when validating persisted plugin registries so fast same-size rewrites cannot leave stale plugin metadata trusted. Thanks @vincentkoc.
• Control UI: add validated gateway.controlUi.chatMessageMaxWidth instead of patched bundled CSS, ignore malformed persisted cron rows before they enter UI state, guard stale cron render paths, and bound the default Sessions tab query to recent activity and fewer rows while keeping filters editable. Fixes #67935, #55047, #54439, and #76050; supersedes #54550 and #54552. (#76051) Thanks @xiew4589-lang and @Neomail2.
• Gateway/channels: cap startup fanout at four channel/account handoffs and recover from Bonjour ciao self-probe races, reducing Windows startup stalls with many Telegram accounts. Fixes #75687.
• Gateway/sessions: keep sessions.list polling responsive on large session stores by reusing list-safe session cache/indexes and returning a lightweight compaction checkpoint preview instead of heavyweight summaries. Thanks @rolandrscheel.
• Control UI/Gateway: keep long-running dashboard WebSocket sessions alive with protocol pings, keep Stop available after reconnect or reload by recovering session-scoped active-run abort state, contain standalone iOS PWA viewports with safe-area-aware document locking, use high-contrast text selection colors, and show inline feedback when local slash-command dispatch is unavailable or fails unexpectedly. Fixes #70991, #60850, and #52105; supersedes #60854. Thanks @alexandre-leng, @kvncrw, @Badschaff, @efe-arv, and @MooreQiao.
• CLI/update: treat inherited Gateway service markers as origin hints and only block package replacement when the managed Gateway is still live, so self-updates can stop the service and continue safely. (#75729) Thanks @hxy91819.
• Agents/failover: exempt run-level timeouts that fire during tool execution from model fallback, timeout-triggered compaction, and generic timeout payload synthesis, avoiding misleading "LLM request timed out" errors after the primary model has already responded. Fixes #52147. (#75873) Thanks @simonusa.
• Docker: copy Bun 1.3.13 from a digest-pinned image and keep CI on the same version. Fixes #74356. Thanks @fede-kamel and @sallyom.
• Agents/compaction: keep prior context on consecutive turns against z.ai-style providers (z.ai direct, openrouter z-ai/*, in-house GLM gateways), avoiding accidental Pi state reset after successful turns. (#76056) Thanks @openperf.
• Doctor/plugins: run a one-time 2026.5.2 configured-plugin install repair based on meta.lastTouchedVersion, update stale configured plugin manifests that still declare channels without channelConfigs, install actively used downloadable OpenClaw plugins through the configured external source, preserve unmanaged third-party plugin node_modules, and then mark the config touched for the release.
• Sessions/transcripts: use one session.writeLock.acquireTimeoutMs policy for session transcript lock acquisitions and raise the default wait to 60 seconds, avoiding user-visible lock timeouts during legitimate slow prep, cleanup, compaction, and mirror work. Fixes #75894. Thanks @shandutta.
• Agents/restart recovery: match cleaned transcript locks by exact transcript lock paths plus the canonical session fallback, so interrupted main sessions using topic-suffixed transcripts resume after gateway restart. Refs #76052. Thanks @anyech.
• Agents/runtime: cache the stable system-prompt prefix and reuse prompt-report tool schema stats during dispatch prep, reducing repeated CPU work before streaming starts. Fixes #75999; supersedes #76061. Thanks @zackchiutw and @STLI69.
• Telegram/native commands: pass persisted session files into plugin commands for topic-bound sessions, so /codex bind works from Telegram forum topics. Refs #75845 and #76049. Thanks @MatthewSchleder.
• Security audit/plugins: ignore plugin install backup, disabled, and dependency debris directories when enumerating installed plugin roots, avoiding false-positive findings for .openclaw-install-backups after plugin updates. Fixes #75456.
• Telegram: honor runtime conversation bindings for native slash commands in bound top-level groups, so commands like /status@bot route to the active non-main session instead of falling back to the default route. Fixes #75405; supersedes #75558. Thanks @ziptbm and @yfge.
• Gateway/tasks: make task registry maintenance use pass-local backing-session lookups and fresh active child-session indexes, avoiding repeated full task snapshots and session-store clones on large stale registries. Fixes #73517 and #75708; supersedes #74406 and #75709. Thanks @Lightningxxl, @glfruit, and @jared-rebel.
• Auth/sessions: JSON-clone auth-profile cache/runtime snapshots and remaining session cleanup previews instead of using structuredClone, preserving mutation isolation while avoiding native-memory growth on large stores. Fixes #45438. Thanks @markus-lassfolk.
• Models CLI: restore openclaw models list --provider <id> catalog and registry fallback rows for unconfigured providers, so provider-specific verification commands no longer report "No models found." Fixes #75517; supersedes #75615. Thanks @lotsoftick and @koshaji.
• Gateway/macOS: write LaunchAgent services with a canonical system PATH and stop preserving old plist PATH entries, so Volta, asdf, fnm, and pnpm shell paths no longer affect gateway child-process Node resolution. Fixes #75233; supersedes #75246. Thanks @nphyde2.
• Slack/hooks: preserve bot alert attachment text in message-received hook content when command text is blank. Fixes #76035; refs #76036. Thanks @amsminn.
• Sessions/agents: route Gateway session-store writes, CLI cleanup maintenance, and agent-delete session purges through a dedicated in-process writer and borrow the validated mutable cache during the writer slot, avoiding runtime file locks plus repeated sessions.json rereads and JSON clones on hot metadata updates. Refs #68554. Thanks @henkterharmsel.
• Memory/markdown: replace CRLF managed blocks in place and collapse duplicate marker blocks without rewriting unmanaged markdown, so Dreaming and Memory Wiki files self-heal from repeated generated sections. Fixes #75491; supersedes #75495, #75810, and #76008. Thanks @asaenokkostya-coder, @ottodeng, @everettjf, and @lrg913427-dot.
• Agents/tools: return critical tool-loop circuit-breaker stops as blocked tool results instead of thrown tool failures, so models see the guardrail and stop retrying the same call. Thanks @rayraiser.
• Agents/sessions: preserve pre-existing runtime model and context window after heartbeat turns so a per-run heartbeat model override does not bleed into shared-session status. Fixes #75452. Thanks @zhangguiping-xydt.
• Model commands: clarify direct and inline /model acknowledgements for non-default selections as session-scoped. Thanks @addu2612.
• Doctor/gateway: stop warning that non-existent, unconfigured user-bin directories are required in the Gateway service PATH. Fixes #76017. Thanks @xiphis.
• TUI/setup: skip full provider model normalization during context-window warmup and bound Terminal hatch bootstrap provider requests, avoiding cold-start stalls with large model registries and first-run hatching stuck behind the watchdog. (#76241) Thanks @547895019 and @joshavant.
• Agents: enable malformed tool-call argument repair for Codex and Azure OpenAI Responses transports while keeping generic OpenAI Responses paths out of the repair gate. Fixes #75154. Thanks @Nimraakram22.
• Memory Wiki: accept relative Markdown links that include the .md suffix during broken-wikilink validation, avoiding false positives for native render-mode links. Thanks @Kenneth8128.
• OpenAI Codex: show the device-pairing code in the interactive SSH/headless prompt while keeping the short-lived code out of persistent runtime logs. Fixes #74212. Thanks @da22le123.
• QA Lab: stop gateway children when the suite parent disappears, so interrupted local QA runs cannot leave hot orphaned gateways behind.
• Codex/app-server/plugins: tolerate second connection closes during startup recovery, include retry counts plus stringified restart errors, and allow the official npm Codex plugin to install without the unsafe-install override while keeping /codex command ownership and covering the real npm Docker live path through managed .openclaw/npm dependencies plus uninstall failure proof.
• Plugins/CLI: cache plugin CLI registration entries per command program so completion state generation does not repeat the full plugin sweep in one invocation. Thanks @ScientificProgrammer.
• Plugins: reuse gateway-bindable plugin loader cache entries for later default-mode loads without serving default-built registries to gateway-bound requests, reducing repeated plugin registration during dispatch. Refs #61756. Thanks @DmitryPogodaev.
• Gateway/secrets: include the caught error message in secrets.reload and secrets.resolve warning logs while keeping RPC errors generic, so operators can diagnose reload and permission failures. Thanks @davidangularme.
• Providers/OpenRouter/LM Studio/Anthropic: fill DeepSeek V4 reasoning_content replay placeholders for openrouter/deepseek/deepseek-v4-flash and openrouter/deepseek/deepseek-v4-pro, normalize binary LM Studio reasoning metadata from Gemma 4 and other local models, and recover Anthropic-compatible stream text deltas that arrive before their matching content block. Fixes #76018 and #76007. Thanks @cloph-dsp and @vliuyt.
• fix(infra): block workspace state-directory env override [AI]. (#75940) Thanks @pgondhi987.
• MCP/OpenAI and media: normalize parameter-free MCP tool schemas before OpenAI tool submission, honor explicit short [[tts:text]]...[[/tts:text]] blocks while keeping untagged short auto-TTS suppressed, and accept home-relative MEDIA:~/... attachment paths under the existing file-read policy. Fixes #75362, #73758, and #73796. Thanks @tolkonepiu, @SymbolStar, @yfge, and @fabkury.
• Hooks/doctor: warn when hooks.transformsDir points outside the canonical hooks transform directory, so invalid workspace skill paths get a direct recovery hint before the Gateway crash-loops. Fixes #75853. Thanks @midobk.
• Proxy/audio: convert standard FormData bodies before proxy-backed undici fetches, so audio transcription and multipart uploads no longer send [object FormData] when HTTP_PROXY or HTTPS_PROXY is configured. Fixes #48554. Thanks @dco5.
• Discord/setup/startup/native commands: write resolved guild/channel allowlist selections to the selected guild and channel, persist slash-command deploy hashes across process restarts, treat abort-time Carbon reconnect-exhausted events as expected shutdown during stale-socket restarts, allow explicit ack reactions in tool-only guild channels, and warn when slash dispatch or direct plugin execution produces no visible reply. Fixes #74922 and #58986; carries forward #58216; supersedes #47788, #73949, and #62057. Thanks @samvilian, @BlueBirdBack, @Eldersonar, @Perttulands, and @jb510.
• Discord/delivery/media: use session-backed A2A announce target lookup for multi-account sessions_send, keep typing indicators alive during long tool runs and auto-compaction, preserve multipart Content-Type headers for uploads, preserve attachment and sticker filenames, and keep non-ASCII channel names in session labels while preserving ASCII-slug allowlists. Fixes #42652 and #59744; refs #51626 and #44773; supersedes #73975. Thanks @irchelper, @dpalfox, @Lanfei, @Squirbie, @FunJim, @xela92, @rockcent, and @swjeong9.
• Discord/threads/PluralKit: canonicalize proxied webhook turns to the original message id for dedupe, inject thread starter context only on the first effective thread turn, and resolve thread ownerId/parentId from Discord API-style snake_case payload fields so bot-owned autoThreads do not require unnecessary mentions. Fixes #41355; supersedes #44447 and #44449. Thanks @acgh213, @p3nchan, and @mgh3326.
• Gateway/diagnostics: include a bounded redacted startup error message in stability bundles, so crash-loop reports identify the failing plugin or contract without exposing secrets. Refs #75797. Thanks @ymebosma.
• Gateway/pricing: defer optional model pricing catalog refresh until after sidecars and channels reach the ready path, so slow OpenRouter or LiteLLM pricing fetches cannot block Gateway readiness. Fixes #74128; supersedes #73486. Thanks @ctbritt and @alprclbi.
• Gateway/pricing: abort in-flight model pricing catalog fetches when Gateway shutdown stops the refresh loop, and avoid post-stop cache writes or refresh timers. Fixes #72208. Thanks @rzcq.
• Codex/app-server: make startup retry cleanup ownership-aware so concurrent Codex lanes cannot close another lane's freshly restarted shared app-server client. Thanks @vincentkoc.
• Google Meet/Twilio/Voice Call: report missing dial-in details during setup, explain that Twilio needs a phone dial plan for Meet URLs, start the phone leg before Meet PIN DTMF, delay intro speech until after post-connect dialing, log each stage, and accept provider call IDs for gateway speak/continue while reporting ended-call state from history.
• Control UI/Talk: allow the OpenAI Realtime WebRTC offer endpoint through the Control UI CSP, configure browser sessions with explicit VAD/transcription input settings, and surface OpenAI realtime error/lifecycle events instead of leaving Talk stuck as live with no diagnostic. Fixes #73427.
• Plugins: clarify config-selected duplicate plugin override diagnostics and document manifest schema updates for bundled-plugin forks. Fixes #8582. Thanks @sachah.
• CLI backends/Claude: make live-session JSONL turn caps bounded and configurable via reliability.outputLimits, raising the default guard for tool-heavy Claude CLI turns while preserving memory limits. Fixes #75838. Thanks @hcordoba840.
• Telegram/DMs/network/commands: keep incidental message_thread_id reply-with-quote metadata on flat DM sessions unless topic isolation is configured, raise outbound text and typing Bot API guards to 60 seconds with safe timeout overrides and typing fallback retries, and register/clear command menus in default and group-chat scopes so /status and plugin commands stay available in forum topics. Fixes #75975, #76013, and #74032; updates #6457. Thanks @ProjectEvolutionEVE, @iaki1206, @dae-sun, and @WouldenShyp.
• Providers/OpenAI: resolve keychain:<service>:<account> OPENAI_API_KEY refs before creating OpenAI Realtime browser sessions or voice bridges, with a bounded cached Keychain lookup. Fixes #72120. Thanks @ctbritt.
• Discord/gateway: reconnect when the gateway socket closes while waiting for the shared IDENTIFY concurrency window, instead of silently skipping IDENTIFY and leaving the bot online but unresponsive. Fixes #74617. Thanks @zeeskdr-ai.
• Voice Call: add sessionScope: "per-call" for fresh per-call agent memory while preserving the default per-phone caller history. Fixes #45280. Thanks @pondcountry.
• Music generation: raise too-small tool timeouts to the provider-safe 10-second floor and collapse cascading abort fallback errors into a clearer root-cause summary. Thanks @shakkernerd.
• Memory-core/dreaming: include the primary runtime workspace in multi-agent dreaming sweeps without mixing main-agent session transcripts into configured subagent workspaces. Fixes #70014. Thanks @ttomiczek.
• Control UI: add tab/RPC timing attribution and decouple slow Overview/Cron secondary refreshes so Sessions navigation gets immediate visible feedback. Refs #64004. Thanks @WaMaSeDu.
• Memory: retry transient SQLite index file swaps during atomic reindex on Windows, so brief EBUSY, EPERM, or EACCES locks do not fail memory rebuilds. Fixes #64187. Thanks @kunpeng-ai-lab.
• Telegram/startup/models: use the existing getMe request guard and higher timeoutSeconds configs for slow Bot API paths, and make model picker confirmations say selections are session-scoped. Fixes #75783 and #75965. Thanks @tankotan and @sd1114820.
• Control UI/slash commands: keep fallback command metadata on a browser-safe registry path, so provider thinking runtime imports cannot blank the Web UI with process is not defined. Fixes #75987. Thanks @novkien.
• Heartbeat/Discord: keep async exec completion events out of the generic System (untrusted) prompt block and let the dedicated exec heartbeat prompt handle them, so Discord no longer receives raw exec failure tails as separate system-style messages. Fixes #66366. Thanks @Promee-ThaBossHoss.
• Heartbeat/scheduler: make heartbeat phase scheduling active-hours-aware so the scheduler seeks forward to the first in-window phase slot instead of arming timers for quiet-hours slots and relying solely on the runtime guard. Non-UTC activeHours.timezone values (e.g. Asia/Shanghai) now correctly influence when the next heartbeat timer fires, avoiding wasted quiet-hours ticks and long dormant gaps after gateway restarts. Fixes #75487. Thanks @amknight.
• Channels: strip plain-text MiniMax and XML tool-call scaffolding from shared user-facing reply sanitization, so messaging channels do not deliver raw model tool syntax when a provider emits it as text instead of structured tool calls. Fixes #62820. Thanks @canh0chua.
• Infer/media: report missing image-understanding and audio-transcription provider configuration for image describe, image describe-many, and audio transcribe instead of blaming the input path when no provider is available. Fixes #73569 and supersedes #73593, #74288, and #74495. Thanks @bittoby, @tmimmanuel, @Linux2010, and @vyctorbrzezowski.
• CLI/infer: reject local codex/* one-shot model probes before simple-completion dispatch and point operators at the Codex app-server runtime path instead of ending with an empty-output error.
• Docs/health: clarify that session listing surfaces stored conversation rows rather than Discord/channel socket liveness, and point connectivity checks at channel status and health probes. Fixes #70420. Thanks @ashersoutherncities-art and @martingarramon.
• WhatsApp/Cron: keep DM pairing-store approvals out of implicit cron and heartbeat recipient fallback, so scheduled automation only uses explicit targets, active configured recipients, or configured allowFrom entries. Fixes #62339. Thanks @kelvinisly-collab.
• Google Meet: keep the agent-facing google_meet tool visible on non-macOS hosts but block local Chrome realtime actions with guidance, so Linux agents can still use transcribe, Twilio, chrome-node, and artifact flows without choosing the macOS-only BlackHole path. Refs #75950. Thanks @actual-software-inc.
• macOS/settings: keep opening General from rewriting openclaw.json during Tailscale settings hydration, preserving gateway, auth, meta, and wizard until the user changes a setting. Fixes #59545. Thanks @Tengdw.
• Discord: prioritize interaction callbacks ahead of stale background REST work without polling active REST buckets, validate oversized gateway payloads and member-intent requests before send, and forward explicit component payloads from message actions. (#75363)
• Active Memory: use the configured recall timeout as the blocking prompt-build hook budget by default and move cold-start setup grace behind explicit setupGraceTimeoutMs config, so the plugin no longer silently extends 15000 ms configs to 45000 ms on the main lane. Fixes #75843. Thanks @vishutdhar.
• Plugins/web-provider: reuse the active gateway plugin registry for runtime web provider resolution after deriving the same candidate plugin ids as the loader path, avoiding a redundant loadOpenClawPlugins call on every request while preserving origin and scope filters. Fixes #75513. Thanks @jochen.
• Crestodian/CLI: exit non-zero when interactive Crestodian is invoked without a TTY, so scripts and CI no longer treat the setup error as success. Fixes #73646 and supersedes #73928 and #74059. Thanks @bittoby, @luyao618, and @Linux2010.
• Cron: keep implicit/default isolated cron announce deliveries out of the main session awareness queue, so isolated jobs do not accumulate in the main conversation. Fixes #61426. Thanks @Lihannon.
• Subagents: avoid duplicate parent-visible replies when a parent uses sessions_send on its own persistent native subagent session, while preserving announce delivery for async sends. Fixes #73550. Thanks @sylviazhang2006-design.
• Web search/Brave: add opt-in brave.http diagnostics for Brave request URLs/query params, response status/timing, and cache hit/miss/write events without logging API keys or response bodies. Fixes #55196. Thanks @mecampbellsoup.
• Web search/Brave: add plugins.entries.brave.config.webSearch.baseUrl for Brave-compatible proxies, including endpoint-aware cache keys for both web and LLM Context modes. Fixes #19075. Thanks @jkoprax and @vishnukool.
• Web search/config: validate explicit tools.web.search.provider values against bundled and installed plugin manifests, while warning for stale third-party plugin config. Fixes #53092. Thanks @TinyTb.
• Web search/SearXNG: retry empty non-general category searches once with the general category, so unsupported category engines do not return empty results when general search has matches. Fixes #73552. Thanks @Loukky.
• CLI/message: skip gateway-stop hooks for read-only message read and bound stop-hook shutdown for other message actions, so one-shot Discord reads cannot hang behind plugin lifecycle cleanup.
• Plugins/web-provider: cache repeated bundled web search and web fetch provider registry loads by default while preserving explicit cache opt-outs. Supersedes #75992. Thanks @DmitryPogodaev.
• Agents/sandbox: preserve existing workspace file modes when sandbox edits atomically replace files, so 0644 files do not collapse to 0600 after Write/Edit/apply_patch. Fixes #44077. Thanks @patosullivan.
• Control UI/WebChat: route typed /new through the New Chat dashboard-session creation flow instead of chat.send, while keeping /reset as the explicit current-session reset. Fixes #69599. Thanks @WolvenRA.
• Agents/models: keep legacy CLI runtime model refs such as claude-cli/* in the configured allowlist after canonical runtime migration, so cron payload.model overrides keep working. Fixes #75753. Thanks @RyanSandoval.
• Codex/app-server: restart the shared Codex app-server client once when it closes during startup thread resume, preserving the existing thread binding instead of retrying thread/start on a closed client. Thanks @vincentkoc.
• Gateway/watch: keep colored subsystem log prefixes in the managed tmux pane even when the parent shell exports NO_COLOR, while preserving explicit FORCE_COLOR=0 opt-out. Thanks @vincentkoc.
• Agents/compaction: submit a non-empty runtime-event marker for pre-compaction memory flush turns, so strict Anthropic providers no longer reject the silent flush as an empty user message. Fixes #75305. Thanks @sableassistant3777-source.
• Plugin SDK: re-export isPrivateIpAddress from plugin-sdk/ssrf-runtime, restoring source-checkout builds for SearXNG and Firecrawl private-network guards. Thanks @vincentkoc.
• Discord/message actions: advertise upload-file and route it through Discord's send runtime with agent-scoped media reads, so agents can discover and send file attachments. Fixes #60652 and supersedes #60808, #61087, and #61100. Thanks @claw-io, @efe-arv, @joelnishanth, and @sjhddh.
• Sessions: suppress exact inter-session control replies such as NO_REPLY and keep agent-to-agent announce bookkeeping out of visible transcripts. Fixes #53145. Thanks @TarahAssistant.
• CLI/directory: report unsupported directory operations for installed channel plugins instead of prompting to reinstall the plugin when it lacks a directory adapter. Fixes #75770. Thanks @lawong888.
• Web search/SearXNG/Firecrawl/Kimi: show the SearXNG JSON API search.formats prerequisite, pass through img_src image URLs, fail explicitly when Kimi returns ungrounded answers, keep public provider requests on strict SSRF guards, reject private/loopback/metadata/non-HTTP(S) hosted Firecrawl scrape targets, and allow explicit self-hosted private Firecrawl endpoints. Fixes #52573, #74357, and #63877; supersedes #65592, #61416, #74360, #48133, #59666, #63941, and #74013. Thanks @evanpaul14, @sghael, @wangwllu, @fede-kamel, @kn1ghtc, @jhthompson12, @jzakirov, @Mlightsnow, and @shad0wca7.
• CLI/models: report gateway model fallback attempts in infer model run --json and avoid double-prefixing provider-qualified defaults such as openrouter/auto in models status. Partially fixes #69527. Thanks @alexifra.
• Providers/OpenRouter: strip trailing assistant prefill turns from verified OpenRouter Anthropic model requests when reasoning is enabled, so Claude 4.6 routes no longer fail with Anthropic's prefill rejection through the OpenAI-compatible adapter. Fixes #75395. Thanks @sbmilburn.
• Voice Call: add per-number inbound routing for dialed-number greetings, response agents/models/prompts, and TTS voice overrides. Fixes #56604. Thanks @healthstatus.
• Feishu: preserve Feishu/Lark HTTP error bodies for message sends, media sends, and chat member lookups, so HTTP 400 failures include vendor code, message, log id, and troubleshooter details. Fixes #73860. Thanks @desksk.
• Agents/transcripts: avoid reopening large Pi transcript files through the synchronous session manager for maintenance rewrites, persisted tool-result truncation, manual compaction boundary hardening, and queued compaction rotation. Thanks @mariozechner.
• Web search/Exa/MiniMax: accept Exa webSearch.baseUrl overrides with endpoint-partitioned caches, include MiniMax Search in setup, and let MINIMAX_API_KEY participate in MiniMax Search auto-detection. Fixes #54928; supersedes #54939 and #65828. Thanks @mrpl327, @lyfuci, and @Jah-yee.
• Plugins/ClawHub: preserve official source-linked trust through archive installs, so OpenClaw can install trusted ClawHub plugin packages that trigger the built-in dangerous-pattern scanner. Thanks @vincentkoc.
• Plugins/ClawHub: install package runtime dependencies for archive-backed plugin installs, so ClawHub packages such as WhatsApp load declared dependencies after download. Thanks @vincentkoc.
• Plugins/tools: cache repeated plugin tool factory results only for matching request context, reducing per-turn tool prep without leaking sandbox, session, browser, delivery, or runtime config state. Fixes #75956. Thanks @Linux2010.
• Providers/LM Studio: allow models.providers.lmstudio.params.preload: false to skip OpenClaw's native model-load call so LM Studio JIT loading, idle TTL, and auto-evict can own model lifecycle. Fixes #75921. Thanks @garyd9.
• Agents/transcripts: keep chat history, restart recovery, fork token checks, and stale-token compaction checks on bounded async transcript reads or cached async indexes instead of reparsing large session files. Thanks @mariozechner.
• Telegram: inherit the process DNS result order for Bot API transport and downgrade recovered sticky IPv4 fallback promotions to debug logs, while keeping pinned-IP escalation warnings visible. Fixes #75904. Thanks @highfly-hi and @neeravmakwana.
• Sessions: keep durable external conversation pointers, including group and thread-scoped chat sessions, out of age, count, and disk-budget maintenance eviction while still allowing synthetic runtime entries to age out. Fixes #58088. Thanks @drinkflav.
• Web search/Providers MiniMax: allow MINIMAX_OAUTH_TOKEN to satisfy MiniMax Search credentials and derive Coding Plan usage polling from the configured MiniMax base URL, so OAuth-authorized and global setups use the right endpoint. Fixes #65768 and #65054. Thanks @kikibrian, @zhouhe-xydt, @sixone74, and @Yanhu007.
• Control UI/WebChat: skip assistant-media transcript supplements when stale media refs resolve to no playable media, so text-only final replies are not stored a second time as gateway-injected assistant messages. Fixes #73956. Thanks @HemantSudarshan.
• Sessions: reject sessions_send targets that resolve to thread-scoped chat sessions, so inter-agent coordination cannot be injected into active human-facing Slack or Discord threads. Fixes #52496. Thanks @barry-p5cc.
• Subagents: honor sessions_spawn with expectsCompletionMessage: false by skipping parent completion handoff delivery while still running child cleanup. Fixes #75848. Thanks @alfredjbclaw.
• Media/completions: treat media-only message-tool sends as delivered async completion output, avoiding duplicate raw MEDIA: fallback posts after video or music generation finishes.
• Gateway/logging: keep deferred channel startup logs on the subsystem logger, so Slack, Discord, Telegram, and voice-call startup messages keep timestamped prefixes. Thanks @vincentkoc.
• Codex/app-server: recover JSON-RPC frames split by raw command-output newlines and include a redacted preview when malformed app-server messages still reach the console. Thanks @vincentkoc.
• Replies/typing: keep typing alive for queued follow-up messages that are genuinely waiting behind an active run, instead of making chat surfaces look idle while work is queued. Fixes #65685. Thanks @papag00se.
• ACP/Discord: suppress completion announce delivery for inline thread-bound ACP session runs, so Discord thread-bound ACP replies are not delivered twice. Fixes #60780. Thanks @solavrc.
• Discord/threads: ignore webhook-authored copies in already-bound Discord session threads even when the webhook id differs, preventing PluralKit proxy copies from creating duplicate turn pressure. Fixes #52005. Thanks @acgh213.
• Discord/threads: return the created thread as partial success when the follow-up initial message fails, so agents do not retry thread creation and create empty duplicate threads. Fixes #48450. Thanks @dahifi.
• Discord/components: consume every button or select in a non-reusable component message after the first authorized click, so single-use panels cannot fire sibling callbacks. Fixes #54227. Thanks @fujiwarakasei.
• macOS/config: preserve existing gateway.auth and unrelated config keys during app fallback writes, so dashboard or Talk settings changes cannot strand Control UI clients by dropping persisted auth. Fixes #75631. Thanks @Fuma2013.
• Control UI/TUI: keep reconnecting chat sends bound to the same backing session id and let TUI relaunches resume the last selected session, avoiding silent fresh sessions after refresh, reconnect, or terminal restart. Fixes #63195, #68162, and #73546. Thanks @bond260312-cmyk, @zhong18804784882, and @mtuwei.
• Plugins/tools: let plugin manifests declare static tool availability so reply startup skips unavailable plugin tool runtimes instead of importing factories that only return null. Thanks @shakkernerd.
• Discord/reactions: skip reaction listener registration when DMs and group DMs are disabled and every configured guild has reactionNotifications: "off", avoiding needless reaction-event queue work. Fixes #47516. Thanks @x4v13r1120.
• CLI sessions: preserve explicit manual-attach reuse bindings so trusted CLI sessions are not invalidated on the first turn when auth, prompt, or MCP fingerprints drift. Fixes #75849. Thanks @alfredjbclaw.
• Telegram/streaming: keep partial preview streaming enabled for plain reply-to replies, disabling drafts only for real native quote excerpts that require Telegram quote parameters. Fixes #73505. Thanks @choury.
• Config: log the "newer OpenClaw" version warning once per process instead of once per config snapshot read. (#75927) Thanks @RomneyDa.
• Telegram/message actions: treat benign delete-message 400s as no-op warnings instead of runtime errors, so stale or already-removed messages do not create noisy delete failures. Fixes #73726. Thanks @Avicennasis.
• Telegram: split long default markdown sends and media follow-up text into safe HTML chunks, so outbound messages over Telegram's limit no longer fail as one oversized Bot API request. Fixes #75868. Thanks @zhengsx.
• Gateway/chat history: merge Claude CLI transcript imports for Anthropic-routed sessions that still have a Claude CLI binding, so local chat history does not hide CLI JSONL turns. Fixes #75850. Thanks @alfredjbclaw.
• Media: trim serialized JSON suffixes after local MEDIA: directive file extensions, so generated-image metadata cannot pollute the parsed media path and cause false ENOENT delivery failures. Fixes #75182. Thanks @TnzGit and @hclsys.
• Plugins/runtime: hot-reload Gateway plugin runtime surfaces after plugin enable/disable changes while keeping source-changing plugin install, update, and uninstall operations restart-backed so loaded module code is not reused. Fixes #72097.
• Cron: make scheduler reload schedule comparison tolerate malformed persisted jobs, so one bad cron entry no longer aborts the whole tick. Fixes #75886. Thanks @samfox-ai.
• Doctor/channels: warn after migrations when default Telegram or Discord accounts have no configured token and their env fallback (TELEGRAM_BOT_TOKEN or DISCORD_BOT_TOKEN) is unavailable, with secret-safe migration docs for checking state-dir .env. Fixes #74298. Thanks @lolaopenclaw.
• Gateway/diagnostics: keep idle liveness samples in telemetry instead of visible warning logs unless diagnostic work is active, waiting, or queued. Thanks @vincentkoc.
• Channels/cron: reject provider-prefixed targets for the wrong channel and let prefixed announce targets such as telegram:123 select their channel when delivery falls back to last, so Telegram IDs cannot be coerced into WhatsApp phone numbers. Fixes #56839. Thanks @bencoremans.
• Control UI/chat: keep live replies visible when a raw session alias such as main sends the chat turn but Gateway emits events under the canonical session key for the same run. Fixes #73716. Thanks @teebes.
• CLI/models: reject --agent on openclaw models set and set-image instead of silently writing agent-scoped requests to global model defaults. Fixes #68391. Thanks @derrickabellard.
• CLI: stop treating the legacy singular openclaw tool ... token as a plugin id under restrictive plugins.allow, so it falls through as a normal unknown/reserved command instead of suggesting a stale allowlist entry. Fixes #64732. Thanks @efe-arv, @SweetSophia, and @hashtag1974.
• Media: write inbound media buffers through same-directory temp files before rename, so failed disk writes do not leave zero-byte artifacts for later voice transcription. Fixes #55966. Thanks @OpenCodeEngineer.
• TTS/Telegram: keep trusted local audio generated by the TTS tool queued for voice-note delivery even when the run-level built-in tool list omits the raw tts name. Fixes #74752. Thanks @Loveworld3033 and @andyliu.
• TTS: require explicit user or config audio intent for the agent speech tool so dashboard chats stay text unless audio is requested. Fixes #69777. Thanks @alexandre-leng.
• Plugins/config: keep bundled source-checkout plugins from being runtime-gated by install-only minHostVersion metadata, accept prerelease host floors, trim plugin-service startup failures to one log line, and avoid broad channel-runtime loading during base config parsing. Thanks @vincentkoc.
• Heartbeat: strip legacy [TOOL_CALL]...[/TOOL_CALL] and [TOOL_RESULT]...[/TOOL_RESULT] pseudo-call blocks from heartbeat replies before channel delivery. Fixes #54138. Thanks @Deniable9570.
• macOS/Voice Wake: send wake-word and Push-to-Talk transcripts through the selected macOS session target instead of always falling back to main WebChat. Fixes #51040. Thanks @carl-jeffrolc.
• Providers/xAI: give Grok web_search a 60s default timeout, harden malformed xAI Responses parsing, and return structured timeout errors instead of aborting the tool call. Fixes #58063 and #58733. Thanks @dnishimura, @marvcasasola-svg, and @Nanako0129.
• Providers/configure: preserve the existing default model when adding or reauthing a provider whose plugin returns a default-model config patch. Fixes #50268. Thanks @rixcorp-oc.
• Slack/DMs/routing: honor dmHistoryLimit for fresh 1:1 DMs, keep top-level DMs on stable DM sessions even when replyToMode targets thread replies, send text/block-only proactive DMs directly with chat.postMessage(channel=<user id>), match Slack target route syntax such as channel:C..., user:U..., or <@U...>, and match public-channel allowlists against bare runtime channel IDs. Fixes #64427, #58832, #62042, #41608, and #41264; supersedes #56530. Thanks @brantley-creator, @daye-jjeong, @MarkMolina, @Winnsolutionsadmin, @babutree, and @Realworld404.
• Slack/delivery/capabilities: preserve missing-scope details in outbound errors, read granted scopes from auth.test metadata before legacy APIs, retry Slack writes only for wrapped DNS request failures such as EAI_AGAIN, and prefer the account bound to the outbound target peer in multi-workspace sends. Fixes #62391, #44625, and #68789; supersedes #66807. Thanks @alexey-pelykh, @Qquanwei, @martingarramon, @sonnyb9, and @rijhsinghani.
• Slack/message actions/tools: send media before follow-up Block Kit messages for file sends, forward agent-scoped media roots through the bundled upload-file path, resolve <!subteam^...> user-group mentions before waking mention-gated channels, and let read fetch an exact Slack message timestamp or thread reply. Fixes #51458, #64625, #73827, and #53943. Thanks @HirokiKobayashi-R, @benpchandler, @CG-Intelligence-Agent-Jack, and @zomars.
• PDF/Gemini: send native PDF analysis API keys in the x-goog-api-key header instead of the request URL, keeping secrets out of proxy and access logs. Supersedes #60600. Thanks @garagon.
• Web search/Gemini/DuckDuckGo/Brave/fetch: route abort signals into Gemini provider fetches, late-bind managed agent web_search calls to the current runtime config snapshot, reuse Google provider API key/base URL as lower-priority Gemini search fallbacks, pass Gemini freshness/date filters through grounding, include DuckDuckGo in setup, honor Gemini/Grok/x_search baseUrl overrides, point Brave metadata at canonical docs, support Brave LLM Context freshness/date ranges, resolve external webFetchProviders for non-sandboxed fetches, and point missing-key errors to web_fetch or browser where appropriate. Fixes #72995, #75420, #66498, #65862, #65870, and #74915; supersedes #57496, #65940, #61972, #65892, and #51005. Thanks @RoseKongPS, @richardmqq, @Aoiujz, @ismael-81, @Jah-yee, @Lanfei, @Magicray1217, @remusao, @ultrahighsuper, @mingmingtsao, and @zhaoyang97.
• Slack/directory: make openclaw directory peers/groups list --channel slack prefer token-backed live readers and return the connected Slack account from directory self, so valid Slack tokens no longer produce empty directory CLI results. Fixes #50776. Thanks @pjaillon.
• Slack: keep assistant typing status, temporary typing reactions, and status reactions active for group/channel turns that use message-tool-only visible replies, while still suppressing automatic source replies. Fixes #75877. Thanks @teosborne.
• Slack: recover full inbound DM text from top-level rich-text blocks when Slack sends a shortened message preview, so long direct messages still reach the agent intact. Fixes #55358. Thanks @tonyjwinter.
• Replies: strip legacy [TOOL_CALL]{tool => ..., args => ...}[/TOOL_CALL] pseudo-call text from user-facing replies and flag it in tool-call diagnostics instead of showing raw tool syntax in channels. Fixes #63610. Thanks @canh0chua.
• WhatsApp: close long-lived web sockets through Baileys end(error) before falling back to raw websocket close, so listener teardown runs Baileys cleanup instead of leaving zombie sockets. Fixes #52442. Thanks @essendigitalgroup-cyber.
• Twitch/plugins: emit a flat JSON Schema for Twitch channel config so single-account and multi-account configs validate before runtime load, and add source-checkout diagnostics for missing pnpm workspace dependencies. Thanks @vincentkoc.
• Gateway/sessions: move hot transcript reads and mirror appends onto async bounded IO with serialized parent-linked writes, keeping large session histories from stalling Gateway requests and channel replies. Fixes #75656. Thanks @DerFlash.
• macOS/Talk Mode: downmix multi-channel microphone buffers before handing them to Apple Speech across Push-to-Talk, Talk Mode, Voice Wake, and the wake-word tester, so pro audio interfaces no longer produce empty transcripts. Fixes #42533. Thanks @jbuecker.
• macOS/Talk Mode: subscribe native WebChat to active-session transcript updates and render external spoken user turns in the chat thread instead of only showing assistant replies. Fixes #75155. Thanks @SledderBling.
• macOS/Voice Wake: accept trigger-only phrases in the built-in Voice Wake test, matching the settings UI and runtime trigger-only path instead of requiring extra command text after the wake word. Fixes #64986. Thanks @zoiks65.
• Cron/TTS: run cron announce payloads through the normal TTS directive transform before outbound delivery, so scheduled [[tts]] replies generate voice payloads instead of leaking raw tags. Fixes #52125. Thanks @kenchen3000.
• WhatsApp: save downloadable quoted image media from reply context as inbound media, so agents can inspect an image that a user replied to instead of only seeing <media:image>. Fixes #59174. Thanks @gaffner.
• Sessions/store: stop persisting the runtime-only skillsSnapshot.resolvedSkills array inside each session entry, so sessions.json no longer carries a copy of every parsed SKILL.md body for every active session; ensureSkillSnapshot rehydrates the array from disk on cold resume so the embedded runner, the Claude CLI skills plugin, and the Claude live-session fingerprint all see populated skills, and legacy stores self-heal on the next save. Refs #11950, #6650, #15000. Thanks @amoghasgekar.
• Doctor/WhatsApp: warn when Linux crontabs still run the legacy ensure-whatsapp.sh health check, which can misreport Gateway inactive when cron lacks the systemd user-bus environment. Fixes #60204. Thanks @mySebbe.
• Slack/setup: print the generated app manifest as plain JSON instead of embedding it inside the framed setup note, so it can be copied into Slack without deleting border characters. Fixes #65751. Thanks @theDanielJLewis.
• Channels/WhatsApp: route CLI logout through the live Gateway and stop runtime-backed listeners before channel removal, so removing a WhatsApp account does not leave the old socket replying until restart. Fixes #67746. Thanks @123Mismail.
• Voice Call/Twilio: honor TTS directive text and provider voice/model overrides during telephony synthesis, so [[tts:...]] tags are not spoken literally and voiceId overrides reach OpenAI/ElevenLabs calls. Fixes #58114. Thanks @legonhilltech-jpg.
• Agents/session-locks: reclaim untracked current-process session locks with matching starttime during acquisition and startup cleanup, so Gateway restarts recover from self-owned orphan .jsonl.lock files. Fixes #75805; refs #49603. Thanks @cdznho.
• Agents/subagents: initialize built-in context engines before native sessions_spawn resolves spawn preparation, so cliBackend-only cold starts no longer fail with an unregistered legacy context engine. Fixes #73095. (#73904) Thanks @brokemac79.
• Plugins/Bonjour: ship the ciao runtime dependency with packaged OpenClaw so fresh OCM envs can start default mDNS discovery without a missing-module failure. Thanks @shakkernerd.
• Agents/tools: scope reply plugin-tool discovery to manifest-declared tool owners and already-active matching tool entries, avoiding broad plugin runtime loading for narrow or core-only tool allowlists. Thanks @shakkernerd.
• Agents/replies: defer implicit image model discovery and keep OAuth auth-store adoption on persisted profiles during reply startup, cutting OCM MarCodex warm prep to sub-second in live checks. Thanks @shakkernerd.
• Plugins/tools: enforce contracts.tools as the manifest ownership contract for plugin tool registration, rejecting undeclared runtime tool names and adding bundled plugin drift coverage. Thanks @shakkernerd.
• Agents/Codex: stop prompting message-tool-only source turns to finish with NO_REPLY, so quiet turns are represented by not calling the visible message tool instead of conflicting final-text instructions. Thanks @pashpashpash.
• Gateway/config: report failed backup restores as failed in logs and config observe audit records instead of marking them valid. (#70515) Thanks @davidangularme.
• Compaction: use the active session model fallback chain for implicit summarization failures without persisting fallback model selection, so Azure content-filter 400s can recover. Fixes #64960. (#74470) Thanks @jalehman and @OpenCodeEngineer.
• Gateway/config: allow gateway config.patch to update documented subagent thinking defaults. Fixes #75764. (#75802) Thanks @kAIborg24.
• Plugins/CLI: keep git plugin install paths credential-free, preserve existing git checkouts until replacement succeeds, honor duplicate npm install mode, and remove managed git repos on uninstall. Thanks @vincentkoc.
• Plugins/CLI: redact authenticated git URLs from git install command failure details, so failed clone or checkout output cannot leak credentials during plugin installs. Thanks @vincentkoc.
• Channels/status reactions: remove stale non-terminal lifecycle reactions when a run reaches done or error, so Discord does not leave a permanent thinking emoji after completion. Fixes #75458. Thanks @davelutztx.
• Discord/doctor: migrate unsupported per-channel agentId entries under guild channel config into top-level bindings[] routes, so openclaw doctor --fix preserves the intended agent route instead of stripping it as an unknown key. Fixes #62455. Thanks @lobster-biscuit.
• Discord/DMs: set inbound direct-message ctx.To to the semantic user:<id> target while keeping delivery routed through the DM channel, so mirror and recovery paths do not treat DMs as channel conversations. Fixes #68126. Thanks @illuminate0623.
• Discord/DMs: keep no-guild inbound messages on direct-message routing when Discord channel lookup is temporarily unavailable, preventing degraded DMs from forking into channel sessions. Fixes #59817. Thanks @DooPeePey.
• Discord: retry outbound API calls on HTTP 5xx, request-timeout, and transient transport failures instead of only Discord rate limits, reducing dropped cron and agent replies during short Discord or network outages. Fixes #52396. Thanks @sunshineo.
• Discord: include Components v2 Text Display content from referenced replies and forwarded snapshots, so component-only messages still appear in reply context. Fixes #56228. Thanks @HollandDrive.
• Discord: add configurable gateway READY timeouts for startup and runtime reconnects, so staggered multi-account setups can avoid false restart loops. Fixes #72273. Thanks @sergionsantos.
• Discord: preserve native slash-command description localizations through command reconcile, so localized Discord descriptions no longer get overwritten by English defaults. Fixes #56580. Thanks @mhseo93.
• Discord: add configured outbound mention aliases so known @Name references can be rewritten to real Discord user mentions instead of relying only on the transient directory cache. Fixes #67587. Thanks @McoreD.
• Discord: avoid startup REST amplification by skipping native command deploy retries after Discord rate limits and deriving the bot id from parseable bot tokens instead of requiring a /users/@me lookup. Fixes #75341. Thanks @PrinceOfEgypt.
• Plugins/hooks: derive hook ctx.channelId from the conversation target instead of the provider name, so Discord and other channel plugins can keep per-channel state isolated. Fixes #59881. Thanks @bradfreels.
• Gateway/config: log config health-state write failures instead of silently hiding config observe-recovery write errors. Thanks @sallyom.
• Diagnostics: reset stuck-session timers on reply, tool, status, block, and ACP progress events, and back off repeated session.stuck diagnostics while a session remains unchanged. Supersedes #72010. Thanks @rubencu.
• Gateway/agents: avoid rebuilding core tools for plugin-only allowlists and keep the full plugin registry cache warm across scoped plugin loads, reducing per-turn latency spikes. Fixes #75882, #75907, #75906, #75887, and #75851. (#75922) Thanks @obviyus.
• Agents/failover: classify bare status: internal server error provider messages as retryable server errors so model fallback can rotate instead of stopping. (#73844) Thanks @thesomewhatyou.
• Gateway/startup: return the shared retryable startup-sidecars error for startup-gated control-plane RPCs such as sessions.create, sessions.send, sessions.abort, agent.wait, and tools.effective, so clients can retry early sidecar races. (#76012) Thanks @scoootscooob.
• Providers/Google: fix Gemini 2.5 Flash-Lite reasoning: "minimal" rejections by raising its thinking-budget floor to 512 while preserving the existing Gemini 2.5 Pro and Flash minimal presets. (#70629) Thanks @ericberic.
• Agents/status: resolve session_status(sessionKey="current") for sparse channel-plugin sessions after literal current lookups miss, so Scope, Slack, Discord, and other plugin-driven agents avoid retrying through Unknown sessionKey: current. Fixes #74141. (#72306) Thanks @bittoby.
• Cron: retry recurring wake-now main-session jobs through temporary heartbeat busy skips before recording success, so queued cron events no longer appear as ok ghost runs while the main lane is still busy. Fixes #75964. (#76083) Thanks @kshetrajna12 and @xuruiray.
• Providers/Google: keep Gemini thinking-signature-only stream chunks active during reasoning, so Gemini 3.1 Pro Preview replies no longer hit idle timeouts before visible text. Fixes #76071. (#76080) Thanks @marcoschierhorn and @zhangguiping-xydt.
• CLI/skills: show per-agent model and command visibility in openclaw skills check --agent, and let doctor report or disable unavailable skills allowed for the default agent. (#75983) Thanks @mbelinky.
• Agents/runtime/tools: keep reply startup on Gateway metadata, manifest catalog rows, auth-store state, and plugin loader cache-key compatibility checks so scoped runtime registries, model allowlists, thinking metadata, media/PDF/generation tools, Comfy workflows, OpenAI Codex OAuth image generation, and image/video/music tool registration avoid broad provider/runtime loads while preserving explicit config and auth-backed providers. Thanks @shakkernerd.
• Discord: document canonical mention formatting in agent prompt hints and channel docs so outbound replies use <@USER_ID>, <#CHANNEL_ID>, and <@&ROLE_ID> instead of legacy nickname mentions. (#75173)
• Heartbeat scheduler: gate exec-event/notification/spawn/retry wakes through a centralized cooldown so backgrounded process.start exit notifications can no longer self-feed runaway heartbeat runs (configured every: "30m" was firing every ~10s in production, pegging the gateway event loop with eventLoopDelayMaxMs >6s spikes that stalled control-UI asset serving and TUI handshakes). Documented wake-now paths (manual, wake, task completion, blocked-task follow-up, /hooks/wake mode=now, and cron --wake now) remain immediate; retryable busy skips no longer poison the cooldown for the next retry; per-agent flood guard caps any unexpected feedback loop at 5 runs/60s. (#64016, refs #17797 and #75436) Thanks @hexsprite.
• fix: block workspace CLOUDSDK_PYTHON override and always set trusted interpreter for gcloud. (#74492) Thanks @pgondhi987.
• Providers/Z.AI: move the bundled GLM catalog and auth env metadata into the plugin manifest, so models list --all --provider zai shows the full known catalog without duplicated runtime seed data. Thanks @shakkernerd.
• Providers/Qianfan and Providers/Stepfun: declare setup auth metadata (api-key method, QIANFAN_API_KEY, STEPFUN_API_KEY) in the plugin manifest so onboarding and models setup surface the expected env var without falling back to legacy providerAuthEnvVars runtime seed data. Thanks @shakkernerd.
• fix(infra): block ambient Homebrew env vars from brew resolution. (#74463) Thanks @pgondhi987.
• Onboarding/configure: avoid staging every default plugin runtime dependency after config writes, so skipped setup flows only prepare config-selected plugin deps instead of pulling broad feature-plugin packages. Thanks @vincentkoc.
• Thinking/providers: resolve bundled provider thinking profiles through lightweight provider policy artifacts when startup-lazy providers are not active, so OpenAI Codex GPT-5.x keeps xhigh available in Gateway session validation. Fixes #74796. Thanks @maxschachere.
• Security/Windows: ignore workspace .env system-path variables and resolve stale-process taskkill.exe from the validated Windows install root, preventing repository-local env files from redirecting cleanup helpers. Thanks @pgondhi987.
• CLI/plugins: refresh persisted plugin registry policy in place for plugins enable and plugins disable, so routine toggles no longer rebuild and hash every plugin source when the target is already indexed. Thanks @vincentkoc.
• Windows/install: run npm from a writable installer temp directory and pin the Bedrock runtime dependency below a Windows ARM Node 24 npm resolver failure, so global OpenClaw installs no longer fail before onboarding. Thanks @mariozechner.
• CLI/plugins: scope install and enable slot selection to the selected plugin manifest/runtime fallback, so plugin installs no longer load every plugin runtime or broad status snapshot just to update memory/context slots. Thanks @vincentkoc.
• Plugins/TTS: keep bundled speech-provider discovery available on cold package Gateway paths and add bundled plugin matrix runtime probes for health, readiness, RPC, TTS discovery, and post-ready runtime-deps watchdog coverage. Refs #75283. Thanks @vincentkoc.
• Google Meet/Twilio: show delegated voice call ID, DTMF, and intro-greeting state in googlemeet doctor, and avoid claiming DTMF was sent when no Meet PIN sequence was configured. Refs #72478. Thanks @DougButdorf.
• Plugins/tools: prefer built bundled plugin code during tool discovery and skip channel runtime hydration while preserving companion provider registrations, reducing per-run plugin-tool prep cost without dropping executable plugin tools. Fixes #75290. Thanks @thanos-openclaw.
• Plugins/loader: scope plugin-tool registry reuse to the enabled plugin plan and stored Gateway method keys, so embedded runner tool lookup can reuse compatible startup registries without hiding enabled non-startup plugin tools. Fixes #75520. Thanks @whtoo.
• Voice Call/Twilio: send notify-mode initial TwiML directly in the outbound create-call request while keeping conversation and pre-connect DTMF calls webhook-driven, so one-shot notify calls do not depend on a first-answer webhook fetch. Supersedes #72758. Thanks @tyshepps.
• Discord/Slack: defer status-reaction cleanup until run finalization so queued, thinking, tool, and terminal reactions no longer flicker during normal progress updates. (#75582)
• Discord/voice: leave voice off for text-only configs unless explicitly configured, rerun configured voice auto-join after gateway RESUMED events, ignore already-destroyed stale voice connections during reconnect cleanup, lengthen the default voice join Ready wait with configurable timeouts, merge configured media-understanding providers such as Deepgram into partial active registries, apply per-channel systemPrompt overrides to voice transcript turns, and run voice-channel turns under a voice-output policy that hides the agent tts tool. Fixes #73753, #40665, #63098, #65687, #47095, and #61536; refs #74044, #39825, and #65039. Thanks @sanchezm86, @SecureCloudProjO, @liz709, @darealgege, @kzicherman, @ayochim, @OneMintJulep, @qearlyao, and @aounakram.
• Plugins/CLI: reuse the cold manifest registry while building plugin status and inspect reports, so large configured plugin sets no longer rediscover the bundled/plugin registry once per inspect row. Thanks @vincentkoc.
• Gateway/health: refresh cached health RPC snapshots when channel runtime state diverges, so Discord and other channel status reads no longer report stale running or connected values until the cache TTL expires. (#75423)
• Gateway/sessions: keep session-store reads from running stale prune and entry-count cap maintenance during startup, so oversized stores no longer block chat history readiness after updates while writes and sessions cleanup --enforce still preserve the cleanup safeguards. Fixes #70050. Thanks @tangda18.
• Security/audit: keep plain security audit on the cold config/filesystem path and reserve plugin runtime security collectors for --deep, so large plugin installs cannot execute every plugin runtime during routine audits. Thanks @vincentkoc.
• WhatsApp: stage qrcode through root mirrored runtime dependencies so packaged QR pairing can render from staged plugin-runtime-deps installs. Fixes #75394. Thanks @FelipeX2001.
• Interactive channel payloads: send Discord component-only interaction replies, Slack block-only slash replies, Telegram button/select fallback labels, and LINE quick-reply fallback option text instead of accepting empty renderable payloads. Thanks @vincentkoc.
• Auto-reply/docking: require /dock-* route switches to start from direct chats, so group or channel participants cannot reroute a shared session's future replies into a linked DM. Thanks @vincentkoc.
• Discord: keep text-DM main-session route updates pinned to the configured DM owner, matching component interactions so another direct-message sender cannot redirect future main-session replies. Thanks @vincentkoc.
• Mattermost/Matrix: keep direct-message main-session route updates pinned to the configured DM owner so paired or temporarily allowed senders cannot redirect future shared-session replies. Thanks @vincentkoc.
• Discord: keep SecretRef-backed bot tokens discoverable for message actions without resolving the token during schema generation, and resolve scoped channel SecretRefs before outbound agent message sends even when the tool is built from a config snapshot. Fixes #75324. Thanks @slideshow-dingo and @Conan-Scott.
• Updates: run package post-install doctor repair with the managed Gateway service profile and state paths when a daemon is installed, so shell/profile mismatches no longer repair the caller state while the restarted Gateway keeps stale config. Thanks @vincentkoc.
• Models/DeepInfra: declare DeepInfra manifest catalog discovery and derive its runtime fallback catalog from the manifest, restoring provider-filtered models list --all --provider deepinfra rows without duplicated static model data. Thanks @shakkernerd.
• CLI/update: verify managed gateway restarts against the installed service port instead of the caller shell port, so package updates do not report a healthy daemon as failed when profiles use different gateway ports. Thanks @vincentkoc.
• Gateway/agent: reject strict openclaw agent --deliver requests with missing delivery targets before starting the agent run, so users do not wait for a completed turn that cannot send anywhere. Thanks @vincentkoc.
• Setup/import: honor non-interactive --import-from onboarding flags by running the migration import path instead of silently completing normal setup without importing anything. Thanks @vincentkoc.
• Doctor/plugins: keep plain doctor --non-interactive from installing bundled plugin runtime dependencies, so headless health checks report missing deps while doctor --fix remains the explicit repair path. Thanks @vincentkoc.
• Doctor/gateway: require an interactive confirmation before installing or rewriting the Gateway service, so doctor --fix --non-interactive can repair plugin/config drift without replacing the operator's launchd/systemd service from a temporary environment. Thanks @vincentkoc.
• Plugins/runtime-deps: include packaged OpenClaw identity in bundled plugin loader cache keys, so same-path package upgrades stop reusing stale versioned runtime-deps mirrors. Fixes #75045. Thanks @sahilsatralkar.
• Plugin SDK: restore reply-prefix and reply-pipeline helpers on the deprecated root/compat SDK surface so external plugins still using openclaw/plugin-sdk do not fail message dispatch after update. Fixes #75171. Thanks @zhangxiliang.
• Plugins/runtime-deps: prune inactive same-package versioned runtime-deps roots after bundled dependency repair, so upgrades do not leave old openclaw-<version>-<hash> package caches behind after doctor runs. Thanks @vincentkoc.
• Plugins/runtime-deps: prune legacy version-scoped plugin runtime-deps roots during bundled dependency repair and cover the path in Package Acceptance's upgrade-survivor matrix, so upgrades from 2026.4.x no longer leave stale per-plugin runtime trees after doctor runs. Thanks @vincentkoc.
• Plugins/runtime-deps: keep Gateway startup plugin imports and runtime plugin fallback loads verify-only after startup/config repair planning, so packaged installs no longer spawn package-manager repair from hot paths after readiness. Refs #75283 and #75069. Thanks @brokemac79 and @xiaohuaxi.
• Plugins/runtime-deps: treat package.json runtime-deps manifests as supersets when generated materialization metadata is absent, so bundled plugin activation stops restaging already-installed dependency subsets on every activation. Fixes #75429. (#75431) Thanks @loyur.
• iMessage: add stdin write callback and error listener to IMessageRpcClient so async EPIPE from a closed child process rejects the pending request instead of crashing the gateway with uncaughtException. Fixes #75438.
• MCP/stdio: settle MCP stdio transport send() from the write callback instead of resolving immediately on buffer acceptance, so async write errors reject the promise instead of being lost. Refs #75438.
• Process/exec: add stdin error listener in runCommandWithTimeout so EPIPE from a prematurely-exited child is swallowed instead of escaping to uncaughtException. Refs #75438.
• Voice Call/realtime: add default-off fast memory/session context for openclaw_agent_consult, giving live calls a bounded answer-or-miss path before the full agent consult. Fixes #71849. Thanks @amzzzzzzz.
• Google Meet: interrupt Realtime provider output when local barge-in clears playback, so command-pair audio stops model speech instead of only restarting Chrome playback. Fixes #73850. (#73834) Thanks @shhtheonlyperson.
• Gateway/config: cap oversized plugin-owned schemas in the full config.schema response so large installed plugin sets cannot balloon Gateway RSS or crash schema clients. Thanks @vincentkoc.
• Plugins/update: skip ClawHub and marketplace plugin updates when the bundled version is newer than the recorded installed version, so openclaw update no longer overwrites working bundled plugins with older external packages. Fixes #75447. Thanks @amknight.
• Gateway/sessions: use bounded tail reads for sessions-list transcript usage fallbacks and cap bulk title/last-message hydration, keeping large session stores responsive when rows request derived previews. Thanks @vincentkoc.
• Gateway/sessions: yield during bulk transcript title/preview hydration and copy compaction checkpoints asynchronously, keeping the Gateway event loop responsive for large session stores and large transcripts. Refs #75330 and #75414. Thanks @amknight.
• Gateway/sessions: stream bounded transcript reads for session detail, history, artifacts, compaction, and send/subscribe sequence paths so small Gateway requests no longer materialize large transcripts or OOM on oversized session logs. Thanks @vincentkoc.
• Gateway/chat: bound chat-history transcript reads to the requested display window so large session logs no longer OOM the Gateway when clients ask for a small history page. Thanks @vincentkoc.
• BlueBubbles: detect audio attachments by Apple UTIs (public.audio, public.mpeg-4-audio, com.apple.m4a-audio, com.apple.coreaudio-format) in addition to audio/* MIME, so iMessage voice notes whose webhook payload only carries the UTI are now classified as audio in the inbound <media:audio> placeholder instead of falling through to the generic <media:attachment> tag. Thanks @omarshahine.
• Voice Call/Twilio: honor stored pre-connect TwiML before realtime webhook shortcuts and reject DTMF sequences outside conversation mode, so Meet PIN entry cannot be skipped or silently dropped. Thanks @donkeykong91 and @PfanP.
• Docs/sandboxing: clarify that sandbox setup scripts (sandbox-setup.sh, sandbox-common-setup.sh, sandbox-browser-setup.sh) are only available from a source checkout, and add inline docker build commands for npm-installed users so sandbox image setup works without cloning the repo. Fixes #75485. Thanks @amknight.
• Google Meet/Voice Call: play Twilio Meet DTMF before opening the realtime media stream and carry the intro as the initial Voice Call message, so the greeting is generated after Meet admits the phone participant instead of racing a live-call TwiML update. Thanks @donkeykong91 and @PfanP.
• Google Meet/Voice Call: make Twilio setup preflight honor explicit --transport twilio and fail local/private Voice Call webhook URLs, including IPv6 loopback and unique-local forms, before joins. Thanks @donkeykong91 and @PfanP.
• Voice Call/Twilio: retry transient 21220 live-call TwiML updates and catch answered-path initial-greeting failures, so a fast answered callback no longer crashes the Gateway or drops the Twilio greeting/listen transition. (#74606) Thanks @Sivan22.
• CLI/startup: preserve OPENCLAW_HIDE_BANNER banner suppression for route-first startup callers that rely on the default process environment while keeping read-only status/channel paths from repairing bundled plugin runtime dependencies. Refs #75183.
• Voice Call/Twilio: register accepted media streams immediately but wait for realtime transcription readiness before speaking the initial greeting, so reconnect grace handling stays live while OpenAI STT startup is no longer starved by TTS. Fixes #75197. (#75257) Thanks @donkeykong91 and @PfanP.
• Voice Call CLI: run gateway-delegated voicecall continue through operation-id polling and protocol-shaped errors, so long conversational turns keep their transcript result without blocking a single Gateway RPC. (#75459) Thanks @serrurco and @DougButdorf.
• Voice Call CLI: delegate operational voicecall commands to the running Gateway runtime and skip webhook startup during CLI-only plugin loading, preventing webhook port conflicts and setup --json hangs. Fixes #72345. Thanks @serrurco and @DougButdorf.
• Agents/pi-embedded-runner: extract the abortable provider-call wrapper from runEmbeddedAttempt to module scope so its promise handlers no longer close over the run lexical context, releasing transcripts, tool buffers, and subscription callbacks when a provider call hangs past abort. (#74182) Thanks @cjboy007.
• Docker: restore python3 in the gateway runtime image after the slim-runtime switch. Fixes #75041.
• Agents/session-repair: fix resumed sessions failing with repeated 400 errors on Anthropic and strict OpenAI-compatible providers (Qwen, mlx-vlm) after an interrupted conversation or blank user input. Fixes #75271 and #75313. Thanks @amknight.
• CLI/Voice Call: scope voicecall command activation to the Voice Call plugin so setup and smoke checks no longer broad-load unrelated plugin runtimes or hang after printing JSON. Thanks @vincentkoc.
• Doctor/plugins: warn when restrictive plugins.allow is paired with wildcard or plugin-owned tool allowlists, making the exclusive plugin allowlist behavior visible before users hit empty callable-tool runs. Refs #58009 and #64982. Thanks @KR-Python and @BKF-Gitty.
• Google Meet/Voice Call: keep Twilio Meet joins in conversation mode and reuse the realtime intro prompt when no voice-call-specific intro is configured, so answered phone bridge calls speak instead of joining silently. Refs #72478. Thanks @DougButdorf.
• Auto-reply/group chats: keep the message tool available for message-tool-only visible replies and apply group-scoped tool policy before deciding fallback delivery, so Discord/Slack-style rooms reply visibly in the correct channel after upgrades. Fixes #74842; refs #75207. Thanks @davelutztx and @aa-on-ai.
• Agents/commitments: keep inferred follow-ups internal when heartbeat target is none, strip raw source text from stored commitments, disable tools during due-commitment heartbeat turns, bound hidden extraction queue growth, expire stale commitments, and add QA/Docker safety coverage. Thanks @vignesh07.
• Telegram/agents: keep typing indicators and optional generation tools off the reply critical path, so fresh Telegram replies no longer stall while provider catalogs and media models load. (#75360) Thanks @obviyus.
• Agents/commitments: run hidden follow-up extraction on the configured agent/default model instead of falling back to direct OpenAI, so OpenAI Codex OAuth-only gateways no longer spam background API-key failures. Fixes #75334. Thanks @sene1337.
• Agents/media: keep async music generation completions on the requester-session wake path even when direct-send completion is enabled, so finished audio stays agent-mediated while video can still opt into direct channel delivery. (#75335) Thanks @vincentkoc.
• Security/config-audit: redact CLI argv and execArgv secrets before persisting config audit records, covering write, observe, and recovery paths. Fixes #60826. Thanks @koshaji.
• Gateway/models: keep default and configured model-list views responsive when provider catalog discovery stalls, without hiding real catalog load failures, while --all still waits for the exact full catalog. Fixes #75297; refs #74404. Thanks @lisandromachado and @najef1979-code.
• Plugins/runtime-deps: accept already materialized package-level runtime-deps supersets as converged, so later lazy plugin activation no longer prunes and relaunches pnpm install after gateway startup pre-staging, reducing event-loop pressure from repeated runtime-deps repair on packaged installs. Fixes #75283; refs #75297 and #72338. Thanks @brokemac79, @lisandromachado, and @midhunmonachan.
• Plugins/runtime-deps: remove OpenClaw-owned legacy runtime-deps symlinks before replacing staged bundled plugin dependencies, so updates can recover from older symlinked installs instead of failing the symlink safety guard. Thanks @goldmar.
• Discord: retry queued REST 429s against learned bucket/global cooldowns and reacquire fresh voice upload URLs after CDN upload rate limits, so outbound sends recover without reusing stale single-use upload URLs. Thanks @discord.
• TTS/providers: keep bundled speech-provider compat fallback available when plugins are globally disabled, so cold gateway and CLI startup can still resolve fallback speech providers instead of leaving explicit TTS provider selection with no registered providers. Refs #75265. Thanks @sliekens.
• Discord: collapse repeated native slash-command deploy rate-limit startup logs into one non-fatal warning while keeping per-request REST timing in verbose output. Thanks @discord.
• Discord: report native slash-command deploy aborts as REST timeouts with method, path, timeout budget, and observed duration, so startup logs explain slow Discord API calls instead of showing a generic aborted operation. Thanks @discord.
• Security/logging: redact payment credential field names such as card number, CVC/CVV, shared payment token, and payment credential across default log and tool-payload redaction patterns so wallet-style MCP tools do not expose raw payment credentials in UI events or transcripts. Thanks @stainlu.
• Providers/OpenAI Codex: preserve existing wrapped Codex streams during OpenAI attribution so PI OAuth bearer injection reaches ChatGPT/Codex Responses, and strip native Codex-only unsupported payload fields without touching custom compatible endpoints. (#75111) Thanks @keshavbotagent.
• Plugins/runtime-deps: materialize newly required bundled plugin packages after local openclaw onboard and openclaw configure config writes, while keeping remote setup read-only, so first Gateway startup no longer discovers missing channel/provider deps after setup claimed success. Fixes #75309; refs #75069. Thanks @scottgl9 and @xiaohuaxi.
• Plugins/runtime-deps: expire stale legacy install locks whose live PID cannot be tied to the current process incarnation, so Docker PID reuse no longer leaves bundled dependency repair stuck behind old .openclaw-runtime-deps.lock directories. Fixes #74948; refs #74950 and #74346. Thanks @dchekmarev.
• Plugins/runtime-deps: recover interrupted bundled runtime-dependency installs whose package sentinels exist but generated materialization is incomplete, forcing npm/pnpm repair in Gateway startup, doctor, and lazy plugin loads instead of leaving channels crash-looping on missing packages. Fixes #75309; refs #75310, #75296, and #75304. Thanks @scottgl9.
• Plugins/runtime-deps: treat no-main and export-map package sentinels without reachable entry files as incomplete, so Gateway startup, doctor, and lazy plugin loads repair interrupted bundled dependency installs instead of accepting package.json-only partial installs. Fixes #75309; refs #75183. Thanks @shakkernerd.
• Plugins/runtime-deps: keep runtime inspection and channel maintenance commands from downloading bundled plugin dependencies, route explicit repairs through openclaw plugins deps --repair, and still allow Gateway/DO paths to repair missing deps before import. Refs #75069. Thanks @xiaohuaxi.
• Updates: force non-deferred, no-cooldown update restarts after package-manager updates requested through the live Gateway control plane and fail release validation on post-swap stale chunk import crashes, so Telegram/Discord imports do not stay pointed at removed dist files. Fixes #75206. Thanks @xonaman and @faux123.
• Agents/tool-result guard: use the resolved runtime context token budget for non-context-engine tool-result overflow checks, so long tool-heavy sessions no longer compact early when contextTokens is larger than native contextWindow. Fixes #74917. Thanks @kAIborg24.
• Gateway/systemd: exit with sysexits 78 for supervised lock and EADDRINUSE conflicts so RestartPreventExitStatus=78 stops Restart=always restart loops instead of repeatedly reloading plugins against an occupied port. Fixes #75115. Thanks @yhyatt.
• Agents/runtime: skip blank visible user prompts at the embedded-runner boundary before provider submission while still allowing internal runtime-only turns and media-only prompts, so Telegram/group sessions no longer leak raw empty-input provider errors when replay history exists. Fixes #74137. Thanks @yelog, @Gracker, and @nhaener.
• Agents/Codex: isolate local Codex app-server CODEX_HOME and HOME per agent and add a deliberate Codex migration path with selectable skill copies, so personal Codex CLI skills, plugins, config, and hooks no longer leak into OpenClaw agents unless the operator migrates them into the workspace. Thanks @pashpashpash.
• Security/Nextcloud Talk: make webhook signature validation use the padded timing-safe compare path even when the supplied signature length is wrong, keep normalized header lookup behavior, and extend regression coverage for tampered bodies, wrong secrets, array-backed headers, and truncated signatures. Carries forward earlier contributor work from #50516 by teddytennant. (#58097) Thanks @gavyngong.
• Plugins/runtime-deps: replace stale symlinked mirror target roots before writing runtime-mirror temp files and skip rewriting already materialized hardlinks, so cross-version container upgrades no longer crash-loop on read-only image-layer paths while warm mirrors do less churn. Fixes #75108; refs #75069. Thanks @coletebou and @xiaohuaxi.
• Auto-reply/group chats: fall back to automatic source delivery when a channel precomputes message-tool-only replies but the message tool is unavailable, so Discord/Slack-style group turns do not silently complete without a visible reply. Fixes #74868. Thanks @kagura-agent.
• Browser/gateway: share one browser control runtime across the HTTP control server and browser.request, and refresh browser profile config from the source snapshot, so CLI status/start honors configured browser.executablePath, headless, and noSandbox instead of falling back to stale auto-detection. Fixes #75087; repairs #73617. Thanks @civiltox and @martingarramon.
• Agents/subagents: bound automatic orphan recovery with persisted recovery attempts and a wedged-session tombstone, and teach task maintenance/doctor to reconcile those sessions so restart loops no longer require manual sessions.json surgery. Fixes #74864. Thanks @solosage1.
• Plugins/runtime-deps: keep bundled provider policy config loading from staging plugin runtime dependencies, so config reads no longer fail on locked-down /var/lib/openclaw/plugin-runtime-deps directories. Fixes #74971. Thanks @eurojojo.
• Memory/runtime-deps: retain the native node-llama-cpp runtime only when local memory search is configured, so packaged installs can repair local embeddings without relying on unreachable global npm installs. Fixes #74777. Thanks @LLagoon3.
• Gateway/startup: skip pre-bind web-fetch provider discovery for credential-free tools.web.fetch config, so Docker/Kubernetes gateways bind even when optional fetch limits are present. Fixes #74896. Thanks @KoykL.
• Signal: match group allowlists against inbound Signal group ids as well as sender ids, and process explicitly configured Signal groups without requiring mentions unless requireMention is set. Fixes #53308. Thanks @minupla and @juan-flores077.
• Signal: bound signal-cli installer release and archive downloads with explicit timeouts, declared and streamed size checks, and partial-file cleanup. Fixes #54153. Thanks @jinduwang1001-max and @juan-flores077.
• Slack: require bot-authored room messages with allowBots=true to come from an explicitly channel-allowlisted bot or from a room where an explicit Slack owner is present, so broad bot relays cannot run unattended. Fixes #59284. Thanks @andrewhong-translucent.
• Signal: derive getAttachment HTTP response caps from channels.signal.mediaMaxMb with base64 headroom, so inbound photos and videos no longer drop behind the 1 MiB RPC default. Fixes #73564. Thanks @heyhudson.
• Signal: keep the long-lived receive SSE monitor open while idle instead of applying the 10s RPC/check deadline, so signal-cli 0.14.3 event streams no longer reconnect before inbound messages arrive. Fixes #74741. Thanks @fgabelmannjr and @k7n4n5t3w4rt.
• CLI/progress: suppress nested progress spinners and line clears while TUI input owns raw stdin, so Crestodian /status no longer disturbs the active input row. (#75003) Thanks @velvet-shark.
• Models/OpenAI Codex: restore openai-codex/gpt-5.4-mini for ChatGPT/Codex OAuth PI runs after live OAuth proof, and align the manifest, forward-compat metadata, docs, and regression tests so stale cron and heartbeat configs resolve again. Fixes #74451. Thanks @0xCyda, @hclsys, and @Marvae.
• Plugins/runtime-deps: always write a dependency map in generated runtime-deps install manifests, so npm does not crash or prune staged bundled-plugin packages when the plan is empty. Fixes #74949. Thanks @hclsys.
• Telegram: use durable message edits for streaming previews instead of native draft state, so generated replies no longer flicker through draft-to-message transitions that look like duplicates. (#75073) Thanks @obviyus.
• Telegram: echo preflighted DM voice-note transcripts back to the originating chat, including Telegram DM topic thread metadata, instead of only echoing later media-understanding transcripts. Fixes #75084. Thanks @M-Lietz.
• Telegram: clamp low long-polling client timeouts so configured timeoutSeconds values below the getUpdates poll window no longer force a fresh HTTPS connection every few seconds. Fixes #75114. Thanks @hpinho77.
• Web search: describe web_search as using the configured provider instead of hard-coding Brave when DuckDuckGo or another provider is active. Fixes #75088. Thanks @sun-rongyang.
• Infra/tmp: tolerate concurrent temp-dir permission repairs by rechecking directories that another process already tightened, so parallel ACP subprocess startup no longer throws Unsafe fallback OpenClaw temp dir. Fixes #66867. Thanks @Kane808-AI and @jarvisz8.
• Agents/compaction: add an opt-in agents.defaults.compaction.midTurnPrecheck mid-turn precheck that detects tool-loop context pressure and triggers compaction before the next tool call instead of waiting for end-of-turn. (#73499) Thanks @marchpure and @haoxingjun.
• Gateway/approvals: let loopback token/password-backed native approval clients resolve exec approvals without attaching stale paired Gateway identities, while remote and unauthenticated approval clients keep normal device identity behavior. (#74472)
• Gateway/config: include rejected validation paths in foreground and service last-known-good recovery logs plus main-agent notices, so unsupported direct edits explain which key caused restore instead of looking like silent reversion. Fixes #75060. Thanks @amknight.
• Plugins/runtime-deps: hash the OS-canonical packageRoot via fs.realpathSync.native (with path.resolve fallback) when computing the bundled runtime-deps stage key, so loader and channel bundled-root callers no longer derive divergent stage directories under ~/.openclaw/plugin-runtime-deps/openclaw-<version>-<hash>/ and bundled channels stop failing with ENOENT on shared dist chunks under Windows npm symlinks, junctions, or PM2 multi-instance worker layouts. Fixes #74963. (#75048) Thanks @openperf and @vincentkoc.
• fix(logging): add redaction patterns for Tencent Cloud, Alibaba Cloud, HuggingFace and Replicate API keys (#58162). Thanks @gavyngong
• Pairing: surface unexpected allowlist filesystem stat errors instead of treating the allowlist as missing, so permission and I/O failures are visible during pairing authorization checks. (#63324) Thanks @franciscomaestre.
• macOS app: reserve layout space for exec approval command details so the allow dialog no longer overlaps the command, context, and action buttons. (#75470) Thanks @ngutman.
• Agents/failover: carry sessionId, lane, provider, model, and profileId attribution through FailoverError and describeFailoverError/coerceToFailoverError so structured error logs (e.g. gateway.err.log ingestion) can attribute exhausted-fallback wrapper errors to the originating session and last-attempted provider instead of dropping the metadata after the per-profile errors. Fixes #42713. (#73506) Thanks @wenxu007.
• Context Engine: treat assembled prompt as the default authority for preemptive overflow prechecks so engines that return a windowed, self-contained context no longer trigger false hard-fail compactions on huge raw history. Engines whose assembled view can hide overflow risk can opt back into the legacy behavior with AssembleResult.promptAuthority: "preassembly_may_overflow". (#74255) Thanks @100yenadmin.
• Mattermost: refresh current native slash command registrations before accepting callbacks so stale tokens from deleted or regenerated commands stop being accepted without a gateway restart while failed validations stay briefly cached and lookup starts are rate-limited per command, gate each callback against the resolved command's own startup token so a token leaked for one slash command cannot poison another command's failure cache, redact slash validation lookup errors, and add a body read timeout to the multi-account routing path so slow callback senders cannot tie up the dispatcher. Thanks @feynman-hou and @eleqtrizit.
• Security/dotenv: block COMSPEC in workspace .env so a malicious repo cannot redirect Windows cmd.exe resolution, and lock in case-insensitive workspace-.env regression coverage for the full Windows shell trust-root family (COMSPEC, PROGRAMFILES, PROGRAMW6432, SYSTEMROOT, WINDIR). (#74460) Thanks @mmaps.
• Gateway/install: drop stale version-manager and package-manager PATH entries preserved from old service files during gateway install --force and doctor repair, so the repair path no longer recreates gateway-path-nonminimal warnings. Fixes #75220. (#75440) Thanks @leonaIee, @renaudcerrato, and @aaajiao.

Welcome to F# Weekly,

A roundup of F# content from this past week:

Microsoft News

• Welcome to SkiaSharp 4.0 Preview 1
• Building an AI-Powered Conference App with .NET’s Composable AI Stack
• Governing MCP tool calls in .NET with the Agent Governance Toolkit
• Combining API versioning with OpenAPI in .NET 10 applications
• High-Performance Distributed Caching with .NET and Postgres on Azure
• VSTest is Removing its Newtonsoft.Json Dependency
• PowerToys 0.99 is here: new monitor controls, easier window management, and Dock upgrades
• Visual Studio April Update – Cloud Agent Integration – Visual Studio Blog
• Visual Studio Code 1.118

Videos

• Functional programming in F# — Fad or Fabulous | Ken Bonny
• .NET MAUI Engineering Team Live Stream: AI-Powered .NET MAUI Development with MauiDevFlow
• Blazor Community Standup: Blazor virtualization gets flexible
• On .NET Live — Make your first NuGet library with .NET 10 and AI

Blogs

• Nick Kossolapov: I built a Game Boy emulator in F#
• Urs Enzler: Event Sourcing — Read Models to the Rescue
• Roman Deneau: What if your pattern matching could call functions?

Highlighted projects

• nickkossolapov/fame-boy — A Game Boy emulator written entirely in F#
• fable-compiler/Fable.Beam — F# to Erlang/BEAM transpiler, opening the door to F# on the Erlang ecosystem
• vykrum/Hywe — Active design sandbox for weaving spatial topology and flow-based hierarchies via design computation, built with F# and WebAssembly
• Savelenko/FSharp.Control.Fold — Represent folding of data structures and streams as composable values via F# applicative computation expressions
• Azizs2162/fyper — Build type-safe Cypher queries for Neo4j in F# using plain records, query expressions, and automatic parameterization

New Releases

• FSharp.Data 8.1.11
• Fable.Core 5.0.0
• WebSharper 10.1.1.670
• dotnet-repl 0.3.259
• Equinox.EventStore 4.1.3
• Oxpecker 2.0.1
• Oxpecker.ViewEngine 2.0.1
• Oxpecker.OpenApi 2.1.2
• Thoth.Json.Core 0.9.0
• Fable.Python 5.1.0
• XstarS.GuidModule 2.14.0
• FSharp.Control.Fold 2.0.0
• Conjecture.FSharp 0.27.1
• Giraffe.ViewEngine.Style 0.2.0
• Farse 0.11.0
• Parquet.Net 6.0.0-pre.8
• NBomber 6.4.0-beta.3
• DSharpPlus 5.0.0-nightly-02585
• Expecto 11.0.0-alpha9
• Terminal.Gui 2.0.2-develop.16
• SwaggerProvider 4.0.0-beta03
• Orleans.FSharp 2.0.0-alpha.1
• Agent.NET v1.0.0-rc.8

That’s all for now. Have a great week.

If you want to help keep F# Weekly going, click here to jazz me with Coffee!